Chapter 3

Link Prediction Techniques,

Applications, and Performance: A
Survey

In this chapter 1 , we exhibit a review of previous methodologies shedding light on link

prediction with the point of convergence mostly on social network graphs. We order
these methodologies into several categories; one category of those calculates a similarity
score between pairs of vertices in which higher scored pairs are assumed to have links
between them. Another category of algorithms is based on probabilistic approaches in
which Bayesian and relational models have been used. Dimensionality reduction
approaches consisting of embedding and factorization-based methods have grouped into
one, and some other approaches also have been studied.

Recently, numerous methodologies of link prediction have been implemented. These

methods can be grouped into several categories, like similarity-based, probabilistic
models, learning-based models, etc.

1 Published in Physica A: Statistical Mechanics and its Applications

21
Chapter 3. Link Prediction Survey 22

Local similarity

Similarity-based Global similarity

Quasi-local
similarity

Hierarchical
random graph
(HRG)

Stochastic block
Link Prediction Probabilistic and model (SBM)
Approaches maximum
likelihood-based
Local probabilistic
model

Probabilistic
relational model

Exponential
random graphs

Embedding-based
Dimensionality
reduction-based
Matrix
factorization-based

Learning-based

Information
theory-based
Other approaches

Clustering-based

F IGURE 3.1: Taxonomy of Link Prediction Approaches

3.1 Similarity-based methods

Similarity-based metrics are the simplest one in link prediction, in which for each pair x
and y, a similarity score S(x, y) is calculated. The score S(x, y) is based on the structural
or node’s properties of the considered pair. The non-observed links (i.e., U − E T ) are
assigned scores according to their similarities. The pair of nodes having a higher score
represents the predicted link between them. The similarity measures between every pair
can be calculated using several properties of the network, one of which is structural
Chapter 3. Link Prediction Survey 23

property. Scores based on this property can be grouped in several categories like local
and global, node-dependent and path-dependent, parameter-dependent and
parameter-free, and so on.

3.1.1 Local similarity indices

Local indices are generally calculated using information about common neighbors and
node degree. These indices consider immediate neighbors of a node. Examples of such
indices contains common neighbor [56], preferential attachment [57], Adamic/Adar [58],
resource allocation [59], etc.

3.1.1.1 Common neighbors (CN)

In a given network or graph, the size of common neighbors for a given pair of nodes x and
y is calculated as the size of the intersection of the two nodes neighborhoods [56].

S(x, y) = |Γ(x) ∩ Γ(y)|, (3.1)

where Γ(x) and Γ(y) are neighbors of the node x and y respectively. The likelihood of the
existence of a link between x and y increases with the number of common neighbors
between them. In a collaboration network, Newman calculated this quantity and
demonstrated that the probability of collaboration between two nodes depends upon the
common neighbors of the selected nodes. Kossinets and Watts [60, 61] investigated a
large social network and recommended that two students are more likely to be friends
who are having numerous common friends. It has been observed that the common
neighbor approach performs well on most real-world networks and beats other complex
methods.
Chapter 3. Link Prediction Survey 24

3.1.1.2 Jaccard coefficient (JC)

This metric [62] is similar to the common neighbor. Additionally, it normalizes the above
score, as given below.
|Γ(x) ∩ Γ(y)|
S(x, y) = . (3.2)
|Γ(x) ∪ Γ(y)|
i.e., the Jaccard coefficient is defined as the probability of selection of common
neighbors of pairwise vertices from all the neighbors of either vertex. The pairwise
Jaccard score increases with the number of common neighbors between the two vertices
considered. Liben-Nowell et al. [17] demonstrated that this similarity metric performs
worse as compared to Common Neighbors.

3.1.1.3 Adamic/Adar index (AA)

Adamic and Adar [58] presented a metric to calculate a similarity score between two
web pages based on shared features, which are further used in link prediction after some
modification by Liben-Nowell et al. [17].

1
S(x, y) = ∑ , (3.3)
z∈Γ(x)∩Γ(y)
log kz

where kz is the degree of the node z. It is clear from the equation that more weights are
assigned to the common neighbors having smaller degrees. This is also intuitive in the
real-world scenario, for example, a person with more number of friends spend less
time/resource with an individual friend as compared to the less number of friends.

3.1.1.4 Preferential attachment (PA)

The idea of preferential attachment is applied to generate a growing scale-free network.

The term growing represents the incremental nature of nodes over time in the network.
Chapter 3. Link Prediction Survey 25

The likelihood incrementing new connection associated with a node x is proportional to

kx , the degree of the node. Preferential attachment score between two nodes x and y can
be computed as [57]
S(x, y) = kx .ky . (3.4)

This index shows the worst performance on most networks, as reported in the result
section. The simplicity (as it requires the least information for the score calculation) and
the computational time of this metric are the main advantages. Also, it can be used in a
non-local context as it requires only degree as information and not the common
neighbors. In assortative networks, the performance of the PA improves, while very bad
for disassortative networks. In other words, PA shows better results if larger degree
nodes are densely connected, and lower degree nodes are rarely connected.

In a supervised learning framework, Hasan et al. [63] showed that aggregate functions
(e.g., sum, multiplication, etc.) over feature values of vertices could be applied to
compute link feature value. In the above equation, summation can also be used instead of
multiplication as an aggregate function, and in fact, it has been proved to be quite useful.
[63] showed the preferential attachment with aggregate function “sum” performs well for
the link prediction in coauthorship network.

3.1.1.5 Resource allocation Index (RA)

The original dynamics of this similarity index is originated from Ou et al. [64] work
published in “Physical Review E” on resource allocation dynamics on complex networks.
Consider two non-adjacent vertices x and y. Suppose node x sends some resources to y
through the common nodes of both x and y then the similarity between the two vertices is
computed in terms of resources sent from x to y. This is expressed mathematically as [59]

1
S(x, y) = ∑ . (3.5)
z∈Γ(x)∩Γ(y)
kz
Chapter 3. Link Prediction Survey 26

This similarity measure and the Adamic/Adar are very similar to each other, as shown by
the equations 3.5 and 3.3, respectively. The difference is that the RA index heavily
penalizes to higher degree nodes compared to the AA index. Prediction results of these
indices become almost the same for smaller average degree networks. This index shows
good performance on heterogeneous networks with a high clustering coefficient,
especially on transportation networks (e.g., USAir97 as reported in the result section).

3.1.1.6 Cosine similarity or Salton index (SI)

In a vector space, document similarities can be computed using the Salton index [65],
also known as Cosine similarity. This similarity index between two records (documents)
is measured by calculating the Cosine of the angle between them. The metric is all about
the orientation and not magnitude. The Cosine similarity [65] can be computed as

|Γ(x) ∩ Γ(y)|
S(x, y) = p . (3.6)
(kx .ky )

3.1.1.7 Sorensen index

This index [66] of similarity was applied mainly to the ecological data samples and given
by Thorvald Sorensen in 1948. It is very similar to the Jaccard index, as we can observe
in the equation 3.7. McCune et al. show that it is more robust than Jaccard against the
outliers [67].
2|Γ(x) ∩ Γ(y)|
S(x, y) = . (3.7)
kx + ky

3.1.1.8 CAR-based common neighbor index (CAR)

CAR-based indices are presented based on the assumption that the link existence between
two nodes is more likely if their common neighbors are members of a local community
(local-community-paradigm (LCP) theory) [68]. In other words, the likelihood existence
Chapter 3. Link Prediction Survey 27

x x x

y y y

CN(x,y) = 3
CN(x,y) = 3 CN(x,y) = 3 LCL(x,y) = 3
LCL(x,y) = 1 LCL(x,y) = 2 CAR(x,y) = 3×3 =9
CAR(x,y) = 3×1 = 3 CAR(x,y) = 3×2 = 6

F IGURE 3.2: CAR Index = (Number of CNs) × (Number of LCLs)

increases with the number of links among the common neighbors (local community links
(LCLs)) of the seed node pair as described in Figure 3.2.

S(x, y) = CN(x, y) × LCL(x, y)

|γ(z)| (3.8)
= CN(x, y) × ∑ ,
z∈Γ(x)∩Γ(y)
2

where CN(x, y) = |Γ(x) ∩ Γ(y)| is number of common neighbors LCL(x, y) refers to the
number of local community links which are defined as the links among the common
neighbors of seed nodes x and y [68]. γ(z) is the subset of neighbors of node z that are
also common neighbors of x and y.

3.1.1.9 Hub promoted index (HPI)

Ravasz et al. [15] published a paper on a cellular organization in metabolic networks.

They show that the metabolic networks are composed of several small and highly
connected topological modules and are combined into larger and less cohesive
hierarchical structures. The number of such modules and their degree of clustering
Chapter 3. Link Prediction Survey 28

follow the power law. This similarity index promotes the formation of links between the
sparsely connected nodes and hubs. It also tries to prevent links formation between the
hub nodes. This similarity metric can be expressed mathematically as

|Γ(x) ∩ Γ(y)|
S(x, y) = . (3.9)
min(kx , ky )

3.1.1.10 Hub depressed index (HDI)

This index is the same as the previous one but with the opposite goal as it avoids the
formation of links between hubs and low degree nodes in the networks. The Hub
depressed index promotes the links evolution between the hubs as well as the low degree
nodes. The mathematical expression for this index [15] is given below.

|Γ(x) ∩ Γ(y)|
S(x, y) = . (3.10)
max(kx , ky )

3.1.1.11 Local naive Bayes-based common neighbors (LNBCN)

The above similarity indices are somehow based on common neighbors of the node pair
where each of the which are equally weighted. This method [69] is based on the Naive
Bayes theory and arguments that different common neighbors play different role in the
network and hence contributes differently to the score function computed for
non-observed node pairs.

C(z) 1−ρ
S(x, y) = ∑ [log( ) + log( )], (3.11)
z∈Γ(x)∩Γ(y)
1 −C(z) ρ
Chapter 3. Link Prediction Survey 29

where C(z) is node clustering coefficient and ρ is the network density expressed as

m
ρ= .
n(n − 1)/2

3.1.1.12 Leicht-Holme-Newman local index (LHNL)

Leicht et al. [70] presented a paper on vertex similarity in networks. Their work is based
on the concept of self-similarity, i.e., two vertices are similar to each other if their
corresponding neighbors are self-similar to themselves. This score is defined by the ratio
of the path of length two that exits between two vertices and the expected path of the
same length between them.
|Γ(x) ∩ Γ(y)|
S(x, y) = . (3.12)
kx .ky

3.1.1.13 Node clustering coefficient (CCLP)

This index [35] is also based on the clustering coefficient property of the network in which
the clustering coefficients of all the common neighbors of a seed node pair are computed
and summed to find the final similarity score of the pair. Mathematically, this index can
be expressed as follows
S(x, y) = ∑ C(z), (3.13)
z∈Γ(x)∩Γ(y)

where
t(z)
C(z) =
kz (kz − 1)
is clustering coefficient of the node z and t(z) is the total triangles passing through the
node z.
Chapter 3. Link Prediction Survey 30

3.1.1.14 Node and link clustering coefficient (NLC)

This similarity index [71] is based on the basic topological feature of a network called
”Clustering Coefficient” [11, 12]. The clustering coefficients of both nodes and links are
incorporated to compute the similarity score.

|Γ(x) ∩ Γ(z)| |Γ(y) ∩ Γ(z)|

S(x, y) = ∑ ×C(z) + ×C(z), (3.14)
z∈Γ(x)∩Γ(y)
kz − 1 kz − 1

3.1.2 Global similarity indices

Global indices are computed using entire topological information of a network. The
computational complexities of such methods are higher and seem to be infeasible for
large networks.

3.1.2.1 Katz index

This index [55] can be considered as a variant of the shortest path metric. It directly
aggregates over all the paths between x and y and dumps exponentially for longer paths to
penalize them. It can be expressed mathematically as

∞ ∞
<l>
S(x, y) = ∑ β l |pathsx,y |= ∑ β l (Al )x,y, (3.15)
l=1 l=1

where, paths<l>
x,y is considered as the set of total l length paths between x and y, β is

a damping factor that controls the path weights and A is the adjacency matrix. For the
convergence of above equation,
1
β< ,
λ1
where λ1 is the maximum eigen value of the matrix A. If 1 is added to each element of
the diagonal of the resulting similarity matrix S, this expression can be written in matrix
Chapter 3. Link Prediction Survey 31

terms as
S = β AS + I, (3.16)

where I is the identity matrix of the proper dimension. The similarity between all pairs of
nodes can be directly computed using the closed-form by rearranging for S in the previous
expression and subtracting the previously added 1 to the elements in the diagonal. Katz
score for each pair of nodes in the network is calculated by finding the similarity matrix
as
S = (I − β A)−1 − I. (3.17)

The computational complexity of the given metric is high, and it can be roughly
estimated to be cubic complexity which is not feasible for a large network.

3.1.2.2 Random walk with restart (RWR)

Let α be a probability that a random walker iteratively moves to an arbitrary neighbor

and returns to the same starting vertex with probability (1 − α). Consider qxy to be the
probability that a random walker who starts walking from vertex x and located at the
vertex y in steady-state. Now, this probability of walker to reach the vertex y is expressed
mathematically as [72]
q~x = αPT q~x + (1 − α)~ex , (3.18)

where ~ex is the seed vector of length |V | (i.e., the total number of vertices in the graph).
This vector consists of zeros for all components except the elements x itself. The transition
matrix P can be expressed as

1

if x and y are connected,
kx
Pxy =
0

otherwise.
Chapter 3. Link Prediction Survey 32

Simplifying the above equation we get,

q~x = (1 − α)(I − αPT )−1~ex . (3.19)

Since this similarity is not symmetric, the final score between the node pair (x, y) can be
computed as
S(x, y) = qxy + qyx . (3.20)

It is clear from the equation 3.19 that matrix inversion is required to solve, which is quite
expensive and prohibitive for large networks. A faster version of this index is implemented
in [72].

3.1.2.3 Shortest path

Lots of algorithms [73–75] are available to compute the shortest path between a vertex
pair in a graph that applies to a different scenario. Liben-Nowell et al. [17] provided the
shortest path with its negation as a metric to link prediction. The inverse relation between
the similarity and length of the shortest path is captured by the following mathematical
equation given below [17].
S(x, y) = −|d(x, y)|, (3.21)

where Dijkstra algorithm [73] is applied to efficiently compute the shortest path d(x, y)
between the node pair (x, y). The prediction accuracy of this index is low compared to
most local indices that make room for the consideration of indirect path in link prediction.

Several paths of different lengths can exist between a vertex pair, the similarity between
such pair is computed by several other methods like Katz index, local path index, etc.
Chapter 3. Link Prediction Survey 33

3.1.2.4 Leicht-Holme-Newman global index (LHNG)

This global index, proposed by Leicht et al. [70], is based on the principle that two nodes
are similar if either of them has an immediate neighbor, which is similar to the other
node. This is a recursive definition of similarity where a termination condition is needed.
The termination condition is introduced in terms of self-similarity, i.e., a node is similar
to itself. Thus, the similarity score equation consists of two terms: first, the neighbor
similarity, and the second, self-similarity, as given below.

S(x, y) = φ ∑ Ax,z Sz,y + ψδx,y . (3.22)

z

Here, the first term is neighborhood similarity and the second term is self-similarity. φ
and ψ are free parameters that make a balance between these two terms. In matrix form
[18, 45]

S = φ AS + ψI = ψ(I − φ A)−1
(3.23)
= ψ(I + φ A + φ 2 A2 + ...)

When the free parameter ψ = 1, this index resembles to the Katz index [55]. Moreover,
we note that A1 (x, y), A2 (x, y), etc, represent number of paths of length 1, 2, and so on
respectively. After some calculation, the final similarity score can be expressed as given
below [70].
α −1 −1
S = 2mλ1 D−1 (I − A) D , (3.24)
λ1
where D is the diagonal matrix, and β is dumping factor that penalizes the longer path
contribution. Dropping the constant term 2mλ1 and rearranging the equation 3.24, it
becomes
β
DSD = A(DSD) + I. (3.25)
λ1
The equation 3.25 solved by iterating this equation repeatedly with the initial value of
(DSD) = 0 and converges normally in 100 iterations as claimed by the authors [70].
Chapter 3. Link Prediction Survey 34

3.1.2.5 Cosine based on L+ (Cos+ )

Laplacian matrix is extensively used as an alternative representation of graphs in spectral

graph theory [76]. This matrix can be defined as L = D − A, where, D is the diagonal
matrix consisting of the degrees of each node of the matrix and A is the adjacency matrix
of the graph. The Pseudoinverse of the matrix defined by Moore-Penrose is represented
as L+ and each entry of this matrix is used to represent the similarity score [77] between
the two corresponding nodes. The most common way to compute this Pseudoinverse
is by computing the singular value decomposition (SVD) of the Laplacian matrix [(L =
U ΣV T ), where U and V are left and right singular vectors of SVD] as follows

L+ = V Σ+ U T , (3.26)

Σ+ is obtained by taking the inverse of each nonzero element of the Σ. Further, the
similarity between two nodes x and y can be computed using any inner product measure
such as Cosine similarity given as

+
Lx,y
S(x, y) = q . (3.27)
+ +
Lx,x Ly,y

3.1.2.6 Average commute time (ACT)

This index is based on the random walk concept. A random walk is a Markov chain
[78, 79] which describes the movements of a walker. ACT is first coined by Gob̈el and
Jagers [80] and applied in link prediction by [81]. It defined as the average number of
movements/steps required by a random walker to reach the destination node y, and come
back to the starting node x. If m(x, y) be the number of steps required by the walker to
reach y from x, then the following expression captures this concept.

n(x, y) = m(x, y) + m(y, x). (3.28)

Chapter 3. Link Prediction Survey 35

This equation can be simplified using the Pseudoinverse of the Laplacian matrix L+ [77,
82] as
+ + +
n(x, y) = |E|(lxx + lyy − 2lxy ), (3.29)

+ denotes the (x, y) entry of the matrix L+ . Pseudoinverse of the Laplacian, L+

where lxy
can be computed as [77]

eeT −1 eeT
L+ = (L − ) + , (3.30)
n n
where e is a column vector consisting of 1’s. The square root of the equation 3.29 is
called Euclidean commute time distance (ECTD) [77], so smaller value of this equation
will represent higher similarity. The final expression representing this similarity index is
thus, given by the squared reciprocal of the equation 3.29 and by ignoring the constant
term |E|.
1
S(x, y) = + + +. (3.31)
lxx + lyy − 2lxy

3.1.2.7 Matrix forest index (MF)

This index is based on the concept of spanning tree which is defined as the subgraph that
spans total nodes without forming any cycle. The spanning tree may contain total or less
number of links as compared to the original graph. Chebotarev and Shamis proposed a
theorem called matrix-forest theorem [83] which states that the number of spanning tree
in a graph is equal to the cofactor of any entry of Laplacian matrix of the graph. Here,
the term forest represents the union of all rooted disjoint spanning trees. The similarity
between two nodes x and y can be computed with the equation 3.32 given below.

S = (I + L)−1 , (3.32)
Chapter 3. Link Prediction Survey 36

where (I + L)(x,y) is the number of spanning rooted forests (x as root) consisting of both
the nodes x and y. Moreover, this quantity is equal to the cofactor of (I + L)(x,y) .

3.1.2.8 SimRank (SR)

SimRank [84] is a measure of structural context similarity and shows object-to-object

relationships. It is not domain-specific and recommends to apply in directed or mixed
networks. The basic assumption of this measure is that two objects are similar if they are
related to similar objects. SimRank computes how soon two random walkers meet each
other, starting from two different positions. The measure is computed recursively using
the equation 3.33.

α kx ky
∑ j=1 S(Γi (x), Γ j (y)) x 6= y


kx ky ∑i=1
S(x, y) = (3.33)
1

x=y

where, α ∈ (0, 1) is a constant. Γi (x) and Γ j (y) are the ith and jth elements in the
neighborhood sets Γ(x) and Γ(y) respectively. Initially, S(x, y) = A(x, y), i.e., S(x, x) = 1
and S(x, y) = 0 for x 6= y. This measure can be represented in matrix form as

S(x, y) = αW T SW + (1 − α)I, (3.34)

where W is the transformation matrix and computed by normalizing each column of

ai j
adjacency matrix A as Wi j = ∑nk=1
.

The computational complexity of this measure is high for a large network, and to reduce
its time, the authors [84] suggest pruning recursive branches after radius l. The time
required to compute this score for each pair is O(k2l+2 ), and total time is O(n2 k2l+2 ).
Chapter 3. Link Prediction Survey 37

3.1.2.9 Rooted PageRank (RPR)

The idea of PageRank [40] was originally proposed to rank the web pages based on the
importance of those pages. The algorithm is based on the assumption that a random walker
randomly goes to a web page with probability α and follows hyper-link embedded in the
page with probability (1 − α). Chung et al. [85] used this concept incorporated with a
random walk in link prediction framework. The importance of web pages, in a random
walk, can be replaced by stationary distribution. The similarity between two vertices x and
y can be measured by the stationary probability of y from x in a random walk where the
walker moves to an arbitrary neighboring vertex with probability α and returns to x with
probability (1 − α). Mathematically, this score can be computed for all pair of vertices as

RPR = (1 − α)(I − α N̂)−1 , (3.35)

where N̂ = D−1 A is the normalized adjacency matrix with the diagonal degree matrix
D[i, i] = ∑ j A[i, j].

3.1.3 Quasi-local indices

Quasi-local indices have been introduced as a trade-off between local and global
approaches or performance and complexity, as shown in Table 3.1. These metrics are as
efficient to compute as local indices. Some of these indices extract the entire topological
information of the network. The time complexities of these indices are still below
compared to the global approaches. Examples of such indices include local path index,
local random walk index [81], local directed path (LDP) [86], etc.
Chapter 3. Link Prediction Survey 38

3.1.3.1 Local path index (LP)

With the intent to furnish a good trade-off between accuracy and computational
complexity, the local path-based metric is considered [53]. The metric is expressed
mathematically as
SLP = A2 + εA3 , (3.36)

where ε represents a free parameter. Clearly, the measurement converges to common

neighbor when ε = 0. If there is no direct connection between x and y, (A3 )xy is equated
to the total different paths of length 3 between x and y. The index can also be expanded to
generalized form
SLP = A2 + εA3 + ε 2 A4 + ... + ε (n−2) An , (3.37)

where n is the maximal order. Computing this index becomes more complicated with the
increasing value of n. The LP index [53] outperforms the proximity-based indices, such
as RA, AA, and CN.

3.1.3.2 Path of length 3 (L3)

Georg Simmel, a German sociologist, first coined the concept “triadic closure” and made
popular by Mark Granovetter in his work [87] “The Strength of Weak Ties”. The authors
[54] proposed a similarity index in protein-protein interaction (PPI) network, called path
of length 3 (or L3) published in the Nature Communication. They experimentally show
that the triadic closure principle (TCP) does not work well with PPI networks. They
showed the paradoxical behavior of the TCP (i.e., the path of length 2), which does not
follow the structural and evolutionary mechanism that governs protein interaction. The
TCP predicts well to the interaction of self-interaction proteins (SIPs), which are very
small (4%) in PPI networks and fails in prediction between SIP and non SIP that
amounts to 96%. They showed that the L3 index performs well in such conditions and
Chapter 3. Link Prediction Survey 39

give mathematical expression to compute this index [54] as

ax,u .au,v .av,y

S(x, y) = ∑ √ . (3.38)
u,v ku .kv

Recently, Pech et al. [88] in Physica A, proposed a work that models the link prediction
as a linear optimization problem. They introduced a theoretical explanation of how direct
count of paths of length 3 significantly improves the prediction accuracy. Meanwhile,
some more studies [89, 90] focusing the length of path have been proposed in the
literature. Muscoloni et al. [89] incorporate the concept of local community paradigm
(LCP) with 2 and 3 length paths and introduced new similarity indices viz.,
Cannistraci-Hebb indices CH2 − L2 and CH2 − L3 corresponding to them. These indices
are based on the common neighbor‘s rewards to internal local community links (iLCL)
and penalization to external local community links (eLCL) [89]. The mathematical
expression to compute these two similarity indices are as follows.

1 + cz
SCH2−L2 (x, y) = ∑ , (3.39)
z∈Γ(x)∩Γ(y)
1 + oz

where cz are total number of neighbors of z which are also members of (Γ(x) ∩ Γ(y)) and
oz are those neighbors counting that are not in (Γ(x) ∩ Γ(y)), also not x or y.
p
CH2−L3 az1,z2 (1 + c˜z1 )(1 + c˜z2 )
S (x, y) = ∑ p . (3.40)
z1∈Γ(x),z2∈Γ(y) (1 + o˜z1 )(1 + o˜z2 )

Here, az1,z2 is 1 when there is link between z1 and z2, 0, otherwise. c˜z1 is the number of
links between z1 and the set of intermediate nodes on all 3 length paths between x and
y. Similarly, o˜z1 is the number of links between z1 and nodes that are not in the set of
intermediate nodes of any 3 length path between x and y, also not x or y.
Chapter 3. Link Prediction Survey 40

3.1.3.3 Similarity based on local random walk and superposed random walk (LRW
and SRW)

Liu and Lü [81] proposed new similarity measures by exploiting the random walk concept
on graphs with limited walk steps. They defined node similarity based on random walks
of lower computational complexity compared to the other random walk based methods
[72, 81]. Given a random walker, starting from the node x, the probability of reaching the
random walker to the node y in t steps is

~x (t) = PT π
π ~x (t − 1), (3.41)

~x (0) is a column vector with xth element as 1 while others are 0’s and PT is
where π
the transpose of the transition probability matrix P. Pxy entry of this matrix defines the
probability of a random walker at node x will move to the next node y. It is expressed as
axy
Pxy = kx , where axy is 1 when there is a link between x and y and 0, otherwise. The authors
computed the similarity score (LRW) between two nodes based on the above concept as

kx ky
SLRW (x, y) = πxy (t) + πyx (t). (3.42)
2|E| 2|E|

This similarity measure focus on only few steps covered by the random walker (hence
quasi-local) and not the stationary state compared to other approaches [72, 81].

Random walk based methods suffer from the situation where a random walker moves far
away with a certain probability from the target node whether the target node is closer or
not. This is an obvious problem in social networks that show a high clustering index i.e.,
clustering property of the social networks. This degrades the similarity score between the
two nodes and results in low prediction accuracy. One way to counter this problem is that
continuously release the walkers at the starting point, which results in a higher similarity
between the target node and the nearby nodes. By superposing the contribution of each
Chapter 3. Link Prediction Survey 41

TABLE 3.1: Comparison of similarity-based approaches

Properties Local Indices Global Indices Quasi-local Indices

Nature Simple Complex Moderate
Features employed Local neighborhood Entire network More local neighborhood
Computational complexity Low high Moderate
Parallelization Easy More complex Moderate
Implementation Feasible for large networks Feasible for small networks Feasible for large networks

walker (walkers move independently), SRW is expressed as

t
SSRW (x, y)(t) = ∑ SLRW (l), (3.43)
l=1

Remarks Similarity-based approaches mostly focus on the structural properties of the

networks to compute the similarity score. Local approaches consider, in general,
neighborhood information (direct neighbors or neighbors of neighbor), which take less
time for computation. This is the property that makes the local approaches feasible for
massive real-world network datasets. Global approaches consider the entire structural
information of the network; that is why time required to capture this information is more
than local and quasi-local approaches. Also, sometimes, entire topological information
may not be available at the time of computation, especially in a decentralized
environment. So, parallelization over the global approaches may not possible or very
complex compared to the local and quasi-local approaches. Quasi-local approaches
extract more structural information than local and somehow less information compared
to the global. Table 3.1 shows a simple comparison among similarity-based approaches
to link prediction.

3.2 Probabilistic and maximum likelihood models

For a given network G(V, E), the probabilistic model optimizes an objective function to
set up a model that is composed of several parameters. Observed data of the given network
can be estimated by this model nicely. At that point, the likelihood of the presence of a
Chapter 3. Link Prediction Survey 42

non-existing link (i, j) is evaluated using conditional probability P(Ai j = 1|θ ). Several
probabilistic models [91–93] and maximum likelihood models [1, 2] have been proposed
in the literature to infer missing links in the networks. The probabilistic models normally
require more information like node or edge attribute knowledge in addition to structural
information. Extracting these attribute information is not easy; moreover, the parameter
tuning is also a big deal in such models that limit their applicability. Maximum likelihood
methods are complex and time-consuming, so these models are not suitable for real large
networks. Some seminal probabilistic and maximum likelihood models are tabulated in
the Table 3.2 [3].

3.2.1 Local probabilistic model for link prediction

Wang et al. [91] proposed a local probabilistic model for link prediction in an undirected
network. They employed three different types of features viz., topological, semantic,
and co-occurrence probability features extracted from different sources of information.
They presented an idea of a central neighborhood set derived from the local topology

Features
Topological features
Semantic features Logistic regression

Local MRF Co-occurrence probability

Predicted output

F IGURE 3.3: Local probabilistic model for link prediction [3]

of the considered node-pair, which is relevant information for the estimation of a link
between them. They computed non-derivable frequent itemsets (i.e., those itemsets whose
occurrence statistics can not be derived from other itemset patterns) from the network
events log data, which is further used as training data for the model. An event corresponds
to a publication of a paper (i.e., authors’ interactions in the paper is a an event, and a set
Chapter 3. Link Prediction Survey 43

of such events is the event log) in the Coauthorship network. The event log consists of
transactional2 . data upon which frequent itemset mining approaches [94–98] are applied.
The model [91] is shown in Figure 3.3, which considers the following approach given
below.

First, the central neighborhood set between x and y is calculated based on local event log
data. One of the usual ways to find the central neighborhood set is to find the shortest
path between two vertices of specified length, and the vertices are lying on this path can
be included in the required set. There can be more than one shortest path between two
vertices, so more neighborhood sets can be possible. Neighborhood sets of shorter lengths
and more frequent (frequency score is used when more shortest paths of the same length
are available) are chosen for the central neighborhood set. The authors considered the
shortest path up to length 4 since the nodes lying on the shorter length path are more
relevant.

In the second step, for a given central neighborhood set, non-derivable frequent itemsets
are used to learn the local probabilistic model. Calders et al. [99] proposed a depth-first
search method to calculate non-derivable itemsets and the same algorithm used by the
authors [91]. [Why non-derivable frequent itemsets? Pavlov et al. [100] first introduced
the concept of frequent itemset to construct an MRF [101]. They argued that a K -itemset
and its support represents a K -way statistics, which can be viewed as a constraint on the
true underlying distribution that generates the data. Given a set of itemset constraints, a
maximum entropy distribution satisfying all these constraints is selected as the estimate
for the true underlying distribution. This maximum entropy distribution is equivalent to
an MRF. Since the number formed links are very few compared to all possible links in
a sparse network, the authors [91] used a support threshold of one to extract all frequent
itemsets. Theses extracted itemsets are large in number that results in expensive learning
for the MRF. To reduce this cost, only non-derivable itemsets are extracted]. They find all

2 Typically, social networks are the results of evolution of chronological sets of events (e.g., authors
participation in the Coauthorship networks). A transaction dataset consists of such events as described by
[91]
Chapter 3. Link Prediction Survey 44

such itemsets that lie entirely within the central neighborhood set. Using these itemsets
[102], a Markov random field is learned.

In the last step, the iterative scaling algorithm [91] is used to learn a local MRF for the
given central neighborhood set. This process continues overall itemset constraints and
continuously updates the model until the model converges. Once the model learning
process is over, one can infer the co-occurrence probability by computing the marginal
probability over the constructed model. The Junction tree inference algorithm [103] is
used to infer co-occurrence probability. The algorithm to induce co-occurrence
probability feature for a pair of vertices can be found in [91].

3.2.2 Probabilistic relational model for link prediction (PRM)

Existing works show that node attributes play a significant role to improve the link
prediction accuracy. However, no generic framework is available to incorporate node and
link attributes and hence, not applicable to all scenarios. To this end, the probabilistic
model is a good and concrete solution that provides a systematic approach to incorporate
both node and link attributes in the link prediction framework. Pioneering works on
PRM include Getoor et al. [47] study on directed networks, Taskar et al. [104] study on
undirected networks, Jennifer Neville work on [92] for both networks, etc. [47]
published in JMLR is based on Relational Bayesian network (RBN) where relation links
are directed and [104] published in NIPS is based on Relational Markov network (RMN)
where relational links are undirected.

PRM was originally designed for attribute prediction in relational data, and it later
extended to link prediction task [47, 92, 104]. The authors employed the attribute
prediction framework to link prediction. This casting can be understood with the
following example [27]. Consider the problem of link prediction in a coauthorship
network. Non-relational frameworks of link prediction consider only one entity type
“person” as node and one relationship; however, relational frameworks (PRMs) include
Chapter 3. Link Prediction Survey 45

more entity types like article, conference venue, institution, etc. Each entity can have
attributes like a person (attributes: name, affiliation institute, status (student, professor)),
article (attributes: publication year, type (regular, review)), etc. Several relational links
may possible among these entities like advisor-advisee/research scholar relation between
two persons, author relationship between person and paper entities, and paper can be
related to the conference venue with publish relationship. Moreover, relationships (links)
among these entities can also have attributes viz., exists (if there is a link between the
two involved entities), or not-exist (no link between the involved entities). This way, the
link prediction can be reduced to an attribute prediction framework/model.

During the model training, a single link graph is constructed that incorporates above
heterogeneous entities and relationships among them. Model parameters are estimated
discriminatively to maximize the probability of the link existence and other parameters
with the given graph attribute information. The learned model is then applied using
probabilistic inference to predict missing links. More details can be explored in
[47, 92, 104].

3.2.3 Hierarchical structure model (HSM) [1]

These models are based on the assumption that the structures of many real networks are
hierarchically organized, where nodes are divided into groups, which are further
subdivided into subgroups and so forth over multiple scales. Some representative work
[1] systematically encodes such structures from network data to build a model that
estimates model parameters using statistical methods. These parameters are then used in
estimating the link formation probability of unobserved links.

Some studies suggest that many real networks, like biochemical networks (protein
interaction networks, metabolic networks, or genetic regulatory networks), Internet
domains, etc. are hierarchically structured. In hierarchical networks, vertices are divided
into groups, which are further sub-divided into subgroups and so forth over multiple
Chapter 3. Link Prediction Survey 46

TABLE 3.2: Probabilistic and maximum likelihood models for link prediction

Model Network types Characteristics References

High accuracy for
Hierarchical structure
Hierarchical networks HSM and low for Clauset et al. [1]
model (HSM)
non-HSM structure
Guimera et al. [2] ,
Good at predicting Natalie Stanley et al.
Stochastic block model
Noisy networks spurious and missing [105], Toni
(SBM)
links Valles-Catala et al.
[106]
Extracts only
topological features Kashima and Abe
Parametric model Dynamic networks
and performs better [107]
than structural methods
Explicitly clusters links Sinead A. Williamson
Non-parametric model Dynamic networks
instead of nodes [108]
Combines
Local probabilistic co-occurrence features
Coauthorship networks Wang et al. [91]
model with topological and
semantic features
Link prediction with
Heterogeneous social
Factor graph model aggregate statistics Kuo et al. [109]
networks
problem
Information and Social soft-block assignment Jaewon Yang et al.
Affiliation model
networks of each node [110]

scales [111]. Clauset et al. [1] proposed a probabilistic model that takes a hierarchical
structure of the network into account. The model infers hierarchical information from
the network data and further applies it to predict missing links.

The hierarchical structures are represented using a tree (binary), or dendrogram, where,
the leaves (i.e., n) represent the number of total vertices in the network and each internal
vertex out of (n − 1) corresponds to the group of vertices descended from it. Each
internal vertex r is associated with a probability pr , then the existing edge probability pxy
between two vertices x and y is given by pxy = pr where, r is their lowest common
ancestor. The hierarchical random graph is then, represented by the dendrogram D∗ with
the set of probability {pr } as (D∗ , {pr }). Now the learning task is to find the hierarchical
random graph(s) that best estimates the observed real-world network data. Assuming all
possible dendrograms to be equally likely, Bayes theorem says that the probability of the
Chapter 3. Link Prediction Survey 47

dendrogram (D∗ , {pr }) that best estimates the data is proportional to the posterior
probability or likelihood, L from which the model generates the observed network and
our goal is to maximize L. The likelihood of a hierarchical random graph (D∗ , {pr }) is
computed using the following equation

L(D∗ , {pr }) = ∏∗ pEr r (1 − pr )Lr Rr −Er , (3.44)

r∈D

where Lr and Rr are the left and right subtree rooted at r, and Er is the number of links in
the network whose endpoints have r as their lowest common ancestor in D∗ . The above
equation assumes the convention 00 = 1. For a given dendrogram D∗ , it is easy to compute
the probability pr that maximizes L(D∗ , {pr }) i.e.

Er
pr = . (3.45)
Lr Rr

This can be understood with the following example illustrated in the Figure 3.4 Now,
this model can be used to estimate the missing links of the network as follows. Sample
a large number of dendrograms with probability proportional to their likelihood. Then,
compute the mean connecting probability pxy of each non-existing pair (x, y) by averaging
the corresponding probability pxy overall sampled dendrograms. Sort these vertices pairs
scores in descending order and selects top-l links to be predicted.

3.2.4 Stochastic block model (SBM) [2]

Hierarchical structures may not represent most networks. A more general approach to
represent these networks is block model [112, 113] where vertices are distributed
(partitioned) into blocks or communities and the connecting probability between two
vertices depends on blocks they belong to. Guimerà et al. [2] presented a novel
framework where stochastic block model representation of a network is employed to find
missing and spurious links. The authors compute the reliability of the existence of links
Chapter 3. Link Prediction Survey 48

a e

c d

b f

1/4

1/9
1/3

1 1 1

1 1 1 1

a b c d e f a b c d e f

F IGURE 3.4: An illustrating example of HSM for a graph of 6 nodes and its two
possible dendrograms as described in the paper [1]. The internal nodes of each
dendrogram are labeled as the maximum likelihood probability pr , defined by the
equation 3.45. The likelihoods of the left and the right dendrograms are L(D1 ) =
(1/3)(2/3)2 .(1/4)2 (3/4)6 = 0.00165, and L(D2 ) = (1/9)(8/9)8 = 0.0433. Thus, the
second (i.e., right) dendrogram is most probable as it divides the network in a balanced
one at the first level.

given an observed network that is further used to find missing links (non-existing links
with higher reliabilities) and spurious links (existing links with lower probabilities).

The link reliability Rxy between the two vertices x and y is [2]

Rxy = pBM (Axy = 1|Ao ).

i.e. probability that the link truly exists given the observed network Ao , the block model
BM.

Generally, complex networks are outcomes of combination of mechanisms, including

Chapter 3. Link Prediction Survey 49

modularity, role structure, and other factors. In SBM, partitioning vertices of network
based on these mechanisms may result in different block models that capture different
correlations (patterns) of the network. Assume that no prior knowledge of suitable
models, the reliability is expressed as

1 lσox σy + 1
Rxy = ∑∗( rσo σ + 2 ) exp[−H(P)], (3.46)
Z P∈P x y

where the sum is over all possible partitions P∗ of the network into groups, σx and σy
are vertices x and y groups in partition P respectively. Moreover, lσoα σβ and rσo α σβ are the
number of links and maximum possible links in the observed network between groups α
and β . The function H(P) is
 
r
H(P) = ∑ [ln(rαβ ) + ln oαβ ], (3.47)
α≤β
lαβ

and Z = ∑P∈P∗ exp[−H(P)]. Practically, solving equation 3.46, i.e., summing over all
possible partitions is too complex even for a small network. However, the Metropolis
algorithm [114] can be used to correctly sample the relevant partitions and obtain link
reliability estimates.

The authors employed the link reliability concept to find missing links and to identify the
spurious link in the networks with the following procedure. (i) Generate the observed
network Ao by removing/adding some random links (for finding missing/spurious links)
from/to the true network At . (ii) Compute the link reliability for non-observed links (i.e.
non-existing + missing/spurious links). (iii) Arrange these links with their reliability score
in decreasing order and decide the top-l links as desired ones (i.e., missing/spurious links).

Probabilistic and maximum likelihood methods extract useful features and valuable
correlation among the data using hierarchical and stochastic block models, which result
in significant improvements in prediction results as compared to some similarity-based
methods. However, these are quite complex and time-consuming even on small datasets
that limit their applicability on large scale real-world network datasets.
Chapter 3. Link Prediction Survey 50

3.2.5 Exponential random graph model (ERGM) or P-star model

Exponential random graphs were first first studied by Holland and Leinhardt [115],
further explored by [101], and practically used by several works [116–118]. ERGM is an
ensemble model where one defines it as consisting of a set of all simple undirected
graphs and specifies a probability corresponding to each graph in the ensemble.
Properties of the ERGM is computed by averaging over the ensemble [117]. Pan et al.
[118] also proposed a similar probabilistic framework (ERGM) to find missing and
spurious links in the network. They employed predefined structural Hamiltonian for the
score computation. The Hamiltonian is selected based on some organizing principle such
that the observed network can have lower Hamiltonian than its randomized one. They
defined the structure Hamiltonian by generalizing the 3-order loop to higher-order as

∞
H(A) = − ∑ βl ln(Tr(Al )), (3.48)
l=3

where A is the adjacency matrix of the network, βl is temperature parameter. Here, the
number of loops of length l starting and ending at the node i is [Al ]ii . For undirected
network, loops are counted several times when counting occurs for each involved node
of the loop, also, for a given node it is counted twice (clockwise and anti-clockwise).
Therefore, Tr(Al ) counts approximated to 2l times the number of loops of length l that
can be taken care of by the parameter βl [118].

For large value of l, increment in Tr(Al ) reaches to the leading eigen value λ1 and small
world phenomenon of a social network ensures to have l to a lower cut-off lc .

lc n
H(A) = − ∑ βl ln( ∑ λ1 ) (3.49)
l=3 i=1

Note that the above equation is result of diagonalization of the adjacency matrix Al as
follows
(Tr(Al ) = Tr(∪T Λl ∪)
Chapter 3. Link Prediction Survey 51

n
= Tr(Λl ∪T ∪) = Tr(Λl ) = ∑ λil
i=1

Once, the structural Hamiltonian is defined to capture different parameters (higher order
loops here), the probability of the appearance of the observed network AO = A − AP in an
ensemble M is
1
p(AO ) = exp[−H(AO )], (3.50)
Z
0
where, Z = ∑P∈M exp[−H(A )] is the partition function. The parameters βl are chosen to
maximize the probability expressed in the above equation.

Now, the score of non-observed links can be computed by the conditional probability of
the appearance of link (x, y)

1
S(x, y) = exp[−H(Ã(x, y))], (3.51)
Zxy

where Ã(x, y) is the observed network by adding the link (x, y), and Zxy is a normalization
factor defined as follows [118]

Zxy = exp[−H[Ã(x, y)]] + exp[−H(AO )].

Here, the prediction is based on the assumption that there is no significant change in the
topological structure after adding the link (x, y) to the observed network and the parameter
βl for Ã(x, y) is almost similar to that of the observed network AO .

3.3 Dimension reduction frameworks for link prediction

The curse of dimensionality is a well-known problem in machine learning. Some

researchers [119, 120] employ dimension reduction techniques to tackle the above
problem and apply it in the link prediction scenario. Recently, many authors are working
Chapter 3. Link Prediction Survey 52

on network embedding and matrix decomposition techniques, which are also considered
as dimension reduction techniques.

3.3.1 Embedding-based link prediction

The network embedding is considered as a dimensionality reduction technique in which

higher D dimensional nodes (vertices) in the graphs are mapped to a lower d (d <<
D) dimensional representation (embedding) space by preserving the node neighborhood
structures. In other words, find the embedding of nodes to a lower d-dimensions such
that similar nodes (in the original network) have similar embedding (in the representation
space). Figure 3.5 shows the structure of Zachary Karate club social network (left) and the

F IGURE 3.5: The Karate club network (left) and its representation in the embedding
space with the DeepWalk [4] algorithm.

representation of nodes in the embedding space using DeepWalk [4] (right). The nodes
are colored based on the membership of their communities.

The main component of the network embedding is the encoding function or encoder fen
that map each node to the embedding space.

fen (x) = zx , (3.52)

where zx is the d-dimensional embedding of the node x. The embedding matrix is Z ∈

Rd×|V | , each column of which represents an embedding vector of a node. Now, a similarity
function is S(x, y) is defined that specifies how to model the vector (embedding) space
Chapter 3. Link Prediction Survey 53

Encode (x)

zx

zy

Encode (y)

Original Network Embedding Space

F IGURE 3.6: Embedding of nodes x and y to the embedding space

relationships equivalent to the relationships in the original network, i.e.,

S(x, y) ≈ zTx zy . (3.53)

Here, S(x, y) is the function that reconstructs pairwise similarity values from the
generated embedding. The term S(x, y) is the one that differ according to the function
used in different factorization-based embedding approaches. For example, graph
∆
factorization [121] directly employ adjacency matrix A i.e. (S(x, y) = A(x,y) ) to capture
∆
first order proximity, GraRep [122] selects (S(x, y) = A2(x,y) ) and HOPE [123] uses other
similarity measures(e.g. Jaccard neighborhood overlap). Most embedding methods
realize the reconstruction objective by minimizing the loss function, L

L= ∑ l(zTx zy , S(x, y)). (3.54)

(x,y)∈{V ×V }

Once the equation 3.54 is converged (i.e. trained), one can use the trained encoder to
generate nodes embedding, which can further be employed to infer missing link and other
downstream machine learning tasks.

Recently, some network embedding techniques [4, 124–127] have been proposed and
applied successfully in link prediction problem. The Laplacian eigenmaps [124],
Chapter 3. Link Prediction Survey 54

Logically linear embedding (LLE) [127], and Isomap [128, 129] are examples based on
the simple notion of embedding. such embedding techniques are having quite complex in
nature and face scalability issues. To tackle the scalability issue, graph embedding
techniques have leveraged the sparsity of real-world networks. For example, DeepWalk
[4] extracts local information of truncated random walk and embeds the nodes in
representation space by considering the walk as a sentence in the language model
[130, 131]. It preserves higher order proximity by maximizing the probability of
co-occurrence of random walk of length 2k + 1 (previous and next k nodes centered at a
given node). Node2vec [125] also uses a random walk to preserves higher order
proximity but it is biased which is a trade-off between the breadth-first search (BFS) and
depth-first search (DFS). The experimental results show that the Node2vec performs
better than the Deepwalk. In next step, Trouillon et al. [132] introduced complex
embedding in which simple matrix and tensor factorization have been used for link
prediction that uses a vector with complex values. Such composition of complex
embedding includes all possible binary relations especially symmetric and
anti-symmetric relations. Recently, some more studies have been published in link
prediction using embedding, for example, Cao et al. subgraph embedding [133], Li et al.
deep dynamic network embedding [134], Kazemi et al. [126], etc. some seminal works
in network embedding are listed in the Table 3.3.

TABLE 3.3: Deep learning models for embedding based link prediction

Model Proximity preserved Embedding type Scalability Learning Reference

With random walk DeepWalk Higher order Shallow Yes Unsupervised [4]
Node2vec Higher order Shallow Yes Semi-supervised [125]
HARP Higher order Shallow Yes Supervised [135]
Walklets Higher order Shallow Yes Unsupervised [136]
Without random walk LINE First and second order Shallow Yes Supervised [137]
SDNE First and second order Deep No Semi-supervised [138]
DNGR Higher order Deep Yes Unsupervised [139]
GCN Higher order Deep Yes Semi-supervised [140]
VGAE Higher order Deep No Unsupervised [141]
SEAL First and second order Deep Yes Supervised [142]
ARGA Higher order Deep No Unsupervised [143]
Chapter 3. Link Prediction Survey 55

3.3.2 Factorization-based frameworks for link prediction

From last decade, matrix factorization has been used in lots of papers based on link
prediction [144–151] and recommendation systems [48]. Typically, the latent features
are extracted and using these features, each vertex is represented in latent space, and
such representations are used in a supervised or unsupervised framework for link
prediction. To further improve the prediction results, some additional node/link or other
attribute information can be used. In most of the works, non-negative matrix
factorization (NMF) has been used. Some authors also applied the singular value
decomposition (SVD) technique [152].

Let the input data matrix is represented by X = (x1 , x2 , ..., xn ) that contains n data vectors
as columns. Now, factorization of this matrix can be expressed as

X ≈ FGT , (3.55)

where X ∈ R p×n , F ∈ R p×k , and G ∈ Rn×k . Here, F contains the bases of the latent space
and is called the basis matrix. G contains combination of coefficients of the bases for
reconstructing the matrix X, and is called the coefficient matrix. k is the dimension of
latent space (k < n). Several well-known matrix factorization techniques are expressed
based on some constraints on either of the three matrices, for example, [153],

SVD:
X± ≈ F± GT± . (3.56)

NMF:
X+ ≈ F+ GT+ . (3.57)

Semi-NMF:
X± ≈ F± GT+ . (3.58)

Convex-NMF:
X± ≈ X±W+ GT± . (3.59)
Chapter 3. Link Prediction Survey 56

In the above four equations, Z± represents the nature of the entries in the matrix Z, i.e.
both positive and negative entries allowed in the matrix Z. In the last equation, F = XW
represents the convex combinations of the columns of F. Generally, such a factorization
problem can be modeled as the following Frobenius norm optimization problem

2
min X − FGT f ro
f ,g
(3.60)
subject to F ≥ 0, G ≥ 0.

Here, kZk2f ro is the frobenius norm of Z and the constraints represent NMF factorization.
However, any of the above four constraints can be used depending on the requirement of
the problem underlying.

After solving the above optimization problem, the similarity between a non-existing pair
(x, y) can be computed by the similarity of the xth and yth row vectors in the coefficient
matrix G.

Acar et al. [144] expressed temporal link prediction as a matrix completion problem and
solve it through the matrix and tensor factorization. They proposed a weighted method to
collapsed the temporal data in a single matrix and factorize it using
CANDECOMP/PARAFAC (CP) [154, 155] tensor decomposition method. Ma et al.
[145] also applied matrix factorization to temporal networks where features of each
network are extracted using graph communicability and then collapsed into a single
feature matrix using weighted collapsing tensor (WCT) [146]. They showed the
equivalence between eigen decomposition of Katz matrix and NMF of the
communicability matrix that serves as the foundation of their framework. Further, a
notable work by Menon et al. [147] is proposed for structural link prediction. Here, the
problem is modeled as matrix completion problem [156], and matrix factorization are
used to solve it. They introduced a supervised matrix decomposition framework that
learns latent (unobserved) structural features of the graph and incorporates it with
additional node/link explicit feature information to make a better prediction.
Additionally, they allowed the factorization model to solve class imbalance problem [24]
Chapter 3. Link Prediction Survey 57

by optimizing ranking loss. Chen et al. [148] proposed somehow similar to work [147],
where the authors extracted topological matrix and attribute matrix and factorized these
matrices using NMF. The final score matrix is obtained by integrating these two matrices
in the latent space. Recently some more works [149–151] have been published in this
area.

3.4 Other approaches

3.4.1 Learning-based frameworks for link prediction

Earlier described approaches (e.g., similarity and probabilistic methods) deal with the
computing a score of each non-observed link either by a similarity or a probabilistic
function. However, the link prediction problem can also be modeled as a learning-based
model to exploit graph topological features and attribute information. The problem is
cast as a supervised classification model where a point (i.e., training data) corresponds to
a vertex-pair in the network, and the label of the point represents the presence or absence
of an edge (link) between the pair. In other words, consider a vertex-pair (x, y) in the
graph G(V, E) and the label of the corresponding data point in the classification model is
l(x,y) . Then,

+1 if (x, y) ∈ E,

l(x,y) = (3.61)
−1

if (x, y) ∈
/ E.

This is typically a binary classification task where several classifiers (e.g., decision tree,
naive Bayes, support vector machine, etc.) can be employed to predict the label of
unknown data points (corresponding to missing links in the network).

One of the major challenges of this model (i.e., machine learning) is the selection of
appropriate feature set [27]. Majority of the existing research works [17, 63, 107] extract
feature sets from the network topology (i.e., topological information of the network).
Chapter 3. Link Prediction Survey 58

These features are generic and domain-independent that are applicable to any network.
Such features are typical, neighborhood, and path-based features. Some other works [63,
157] concentrate on extracting node and edge features that play a crucial role to improve
the performance of link prediction. Hasan et al. [63] employed vertex attribute viz.,
the degree of overlap among research keywords incorporated with other features in the
coauthorship network, and showed that the author-pairs having higher values of these
features are top rankers in the list. The cost of extraction of such features is cheap and
easy, while the main disadvantage is the domain-specific nature of them.

3.4.2 Information theory-based link prediction

Several complex networks have utilized the concept of information theory to compute
their complexity on different scales [158, 159]. They defined several correlation measures
and modeled some networks (e.g., star, tree, lattice, ER graph, etc.). They also showed
that the real networks spanned noise entropy space. Bauer et al. [160] used the maximum
entropy principle to assign a statistical weight to any graph and introduced random graph
construction with arbitrary degree distribution.

Tan et al. [161] posed the link prediction problem in the framework of information
theory. They mainly focus on local assortativity to capture local structural properties of
the network and showed that mutual information (MI) method performs well on both low
and highly correlated networks. Motivated by [161], Zhu, B. and Xia [162] added more
local features (i.e., links information of neighbors of the seed nodes as well as their
common neighbors) in their framework and called it as neighbor set information (NSI)
index. Thus, they showed that the different features could be combined in an
information-theoretic model to improve the link prediction accuracy.

Xu et al. [163] considered path entropy as a similarity metric for the link prediction
problem. The authors assumed that there is no correlation among the degrees of the nodes
0 shows no
in the network. Consider the following notations based on the paper [163]: Lxy
Chapter 3. Link Prediction Survey 59

link exists between two vertices x and y, and the corresponding existence is represented
1 . Probability of existence of a link between the above two vertices is given as
by Lxy

y k
1 0
CM−kx
P(Lxy ) = 1 − P(Lxy ) = 1− k
, (3.62)
CMy

k
where CMy represents the number of candidate link sets for the vertex y with all links
k
y
incident with y and CM−kx
denotes the number of candidate link sets for the vertex y with
all links incident with y but none of them is incident with x.

y k
1 1
CM−kx
I(Lxy ) = − log P(Lxy ) = − log(1 − k
) (3.63)
CMy

They show that the likelihood of occurrence of a path having no loops equates to
multiplication of the occurrence probabilities of the links involved in that path. i.e., given
a simple path D = v0 , v1 , v2 , vγ of length γ, the co occurrence probability of path D is
evaluated to
γ−1
P(D) ≈ ∏ P(Lv1i vi+1 ) (3.64)
i=0

and, the sum of links entropies involved in a path equals to the entropy of the path.

γ−1
I(D) ≈ ∑ I(Lv1i,vi+1). (3.65)
i=0

Further, they calculated similarity based on entropy of the path which is the negative of
conditional entropy
PE 1
Sxy = −I(Lxy | ∪maxlen
i=2 Dixy ), (3.66)

where Dixy represents the set consisting of all simple paths of length i between the two
vertices and maxlen is the maximum length of simple path of the network. Outcome
results on several networks demonstrate that the similarity index based on path entropy
performs better than other indices in terms of prediction accuracy and precision. Xu et al.
[164] extend the previous work [163] to the weighted network by considering the weight
Chapter 3. Link Prediction Survey 60

of the paths. Recently, some more efforts have been applied in this direction based on
different features of the networks like influential nodes [165], combining node attributes
with structural similarity [166], local likelihood [167], and maximal entropy random walk
[168].

3.4.3 Clustering-based link prediction

This paragraph gives an overview of the clustering-based link prediction. Huang [5]
presented a paper on graph topology-based link prediction where a generalized
clustering coefficient is used as a predictive parameter. The author introduces a cycle
formation model that shows the relationship between link occurrence probability and its
ability to form different length cycles. This model suggests that the occurrence
probability of a particular link depends on the number of different lengths cycles formed
by adding this link. The model is based on the assumption of the stationary property of
the degree of clustering of the network [169]. This model captures longer cycles by
extending the higher-order clustering coefficients [170] and defines the generalized
clustering coefficient C(k) as

number of k-length cycles

C(k) = , (3.67)
number of k-length paths

where k is the degree of the cycle formation model.

The author treats the link occurrence probability as governed by t link generation
mechanisms g(1), g(2),...,g(k) of cycle formation model, each described by a single
parameter c1 , c2 ,..., ck . The above mentioned link generation mechanism can be
understood with the help of the Figure 3.7. Consider a cycle formation model (CF(k)) of
degree (k = 3). The Seed link (x, y), here, can be generated by the following three
mechanisms; the random link occurrence g(1), length-2 cycle generation g(2) i.e.
(x − a − y and x − c − y), and length-4 cycle generation g(3) i.e. (x − b − d − y). The
main issue is to combine several generation mechanisms to compute total link
Chapter 3. Link Prediction Survey 61

?? y
x

c
a

F IGURE 3.7: An example illustrating the cycle formation link probability model [5],
where the the probability of the missing link (x − y) is generated by the following three
mechanisms; random link occurrence g(1), length-2 cycle generation g(2) i.e. (x − a −
y, x − c − y), and length-4 cycle generation g(3) i.e. (x − b − d − y).

occurrence probability. The author [5] posits a method to combine both path and cycle
(of different lengths) generation mechanism in the framework. The expected general
clustering coefficient of degree k for this model can be estimated as [5]

E[C(k)] = f (c1 , c2 , ..., ck )

(3.68)
= ∑ |Gi |p(Gi )p((el,k+l ) ∈ E|Gi ),
i

where |Gi | is the number of subgraphs possible corresponding to the graph pattern Gi ,
listed in Table 1 of the paper [5], p(Gi ) is the probability of occurrence of one of such
graphs Gi , and p((el,k+l ) is the probability of edge el,l+1 to occur given the pattern Gi .
Finally, given the coefficients, the probability of existence of link is

|pathix,y |
c1 ∏ki=2 ci
px,y (c1 , ..., ck ) = |pathix,y |
. (3.69)
i
c1 ∏ki=2 ci + (1 − c1 ) ∏ki=2 (1 − ci )|pathx,y |

Liu et al. [171] proposed degree related clustering coefficient to quantify the clustering
Chapter 3. Link Prediction Survey 62

ability of nodes. They applied the same to paths of shorter lengths and introduced a new
index Degree related Clustering ability Path (DCP). They performed the degree of
robustness (DR) test for their index and showed that missing links have a small effect on
the index. Recently Wu et al. [35] extracted triangle structure information in the form of
node clustering coefficient of common neighbors. Their experiments on several real
datasets show comparable results to the CAR index in [68]. The same concept of the
clustering coefficient also introduced in the work presented by Wu et al. [71]. Authors
introduce both node and link clustering information in their work. Their experiments on
different network datasets showed better performance results against existing methods,
especially on middle and large network datasets. Kumar et al. [172] explored the concept
of node clustering coefficient to the next level (level-2) that captures more clustering
information of a network. Meanwhile, Benson et al. [173] studied simplicial closure
events to capture higher-order structures in several temporal networks. The simplicial
closure events are the process of closure of timestamped simplices (simplicial
complexes3 are set of nodes with different sizes) available in a dataset. These structures
are common in several real-time complex systems, for example, communication in a
group, collaboration of authors for a paper, etc. To assess these higher-order structures,
the authors study the simplicial closure events on triples of nodes (for simplicity) and
suggest that the open triangles or triples of nodes with strong ties are more likely to close
in the future.

3.5 Experimental Setup and Results Analysis

There can be two mainly scenarios of link prediction. In the first scenario where static
graph is taken into account, a percentage of links are randomly removed fron the network
and during evaluation, probabilties of these links are predicted along with other
non-existing links. This scenario aims to take the temporal evolution of the graph into
account and only links formed after some point in time, t, are removed. The state of the
3 https://en.wikipedia.org/wiki/Simplicial complex
Chapter 3. Link Prediction Survey 63

graph before t is given to the link predictor and its aim is to predict links formed at a later
time. The first setting is applicable when the current knowledge represented by the graph
is incomplete and link prediction aims to complete it as well as when the temporal data
for the graph is unknown or irrelevant. In the secoend scenario, link prediction finds
future links that may appear in near future. This scenario is more challenging because of
the reason: during the time, new nodes can be introduced in the network that may
contain little or no information as these nodes may not be connected to other nodes in the
network initially.

We have used the first scenario during the evaluation of link prediction algorithm.
Random-slicing has been used with 10-fold cross-validation. In this setting, we randomy
split a given dataset in 10 disjoint subsets with consideration of one of them as test set
and remaining as training sets in each iteration.

3.5.1 Datasets

This chapter used 8 network datasets from various fields to study the performance of
similarity-based algorithms. Karate4 [174]: A friendship network of 34 members of a
Karate club at a US university. Dolphin3 [175]: A social network of dolphins living in
Doubtful Sound in New Zealand. Macaque5 [176]: is a biological network of cerebral
cortex of Rhesus macaque. Football3 [177]: American football games network played
between Division IA colleges during regular season Fall 2000. Jazz6 [178]: A
collaboration network of 115 jazz musician where a link between two musicians denotes
music played by both in a band. C. Elegans3 [11]: A neural network of C. Elegans
compiled by D. Watts and S. Strogatz in which each node refers a neuron and, a link
joins two neurons if they are connected by either a synapse or a gap junction. USAir977
is an airline network of US where a node represents an airport, and a link shows the
4 http://www-personal.umich.edu/ mejn/netdata/
5 https://neurodata.io/project/connectomes/
6 http://deim.urv.cat/ alexandre.arenas/data/welcome.htm
7 http://vlado.fmf.uni-lj.si/pub/networks/data/
Chapter 3. Link Prediction Survey 64

connectivity between two airports. Netscience3 [179] is a Coauthorship network of

researchers in the network theory domain where a node is denoted by a researcher, and
an edge denotes coauthorship of at least one paper between two researchers.
TABLE 3.4: Topological information of real-world network datasets

Datasets |V | |E| hDi hKi hCi

Karate 34 78 2.337 4.588 0.570
Dolphin 62 159 3.302 5.129 0.258
Macaque 91 1401 1.658 30.791 0.742
Football 115 613 2.486 10.661 0.403
Jazz 198 2742 2.235 27.697 0.620
C. Elegans 297 2148 2.447 14.456 0.308
USAir97 332 2126 2.738 12.807 0.749
Netscience 1589 2742 5.823 3.451 0.878

Table 3.4 shows some basic topological properties of the considered network datasets.
|V | and |E| are the total numbers of nodes and links of the networks, respectively. hDi
represents the average shortest distance, hKi, the average degree, and hCi, the average
clustering coefficient of the network.

3.5.2 Accuracy

Four accuracy measures, namely Recall@k [180], area under the Precision-Recall curve
[181], area under the ROC curve [52, 182], and average precision [180], have been used
to evaluate each similarity-based algorithm and some other representative methods. We
report these results in the Tables 3.5, 3.6, 3.7, and Table 3.8 for similarity-based
approaches and Tables 3.9, 3.10, 3.11, and 3.12 for other representative methods. In the
tables of other representative methods, the first method (i.e., HSM) is the maximum
likelihood-based method followed by the next three embedding-based methods followed
by the three clustering methods. The last method of the table belongs to other category.
The best results are highlighted in the table on each dataset. These results are generated
with the help of code implemented by Gregorio Alanis-Lobato8 .
8 https://github.com/galanisl/LinkPrediction
Chapter 3. Link Prediction Survey 65

Recall@k In the context of recommendation systems we are most likely interested in

recommending top-N items to the user. So it makes more sense to compute recall metric
in the first N items instead of all the items. Thus the notion of recall@k where k is a user
definable integer that is set by the user to match the top-N recommendations objective.

The recall results for each similarity-based method on all the datasets have been shown
in Table 3.5. This measure represents the ability to find all positive/relevant samples
by a classifier. We observe that the SPM outperforms against the existing methods on
four datasets (Macaque, C. Elegans, Jazz, and USAir97). CAR method best performs on
Karate and HDI on dolphins. The global version of LHN (i.e., LHNG) works best on
football dataset, and RA is the best performing approach on Netscience. Local similarity
methods extract relevant documents more precisely on 3 datasets, and the global methods
retrieve more accurate on 5 datasets. Quasi-local approaches and CAR-based indices lie
in top-5 ranked algorithms. The quasi-local methods have average good performance
compared to the global approaches.

Table 3.9 shows the recall results for other representative methods where SPM
outperforms on C. Elegans, Jazz, USAir97, and Netscience. HSM is a good indicator for
Dolphin, Node2vec for Macaque, and CCLP2 for Karate and Football networks. On
Karate, both the CCLP2 and The SPM show equally good performance.

Area under the precision-recall curve (AUPR) Area under the precision-recall curve
(AUPR) is proved to be more informative for imbalanced datasets. The real-world
networks are highly imbalanced as the number of positive samples is very less than the
negative samples. Table 3.6 shows the AUPR results on eight datasets. Here, We observe
that the aupr results resemble the recall@k, i.e., SPM best performs on five datasets as
that of recall results and CAR, HDI, LHNG, and RA outperform on karate, dolphin,
football and netscience datasets respectively.

AUPR results corresponding to other representative methods are tabulated in the Table
3.10. Here, The Node2vec shows the best performance on all datasets except the
Chapter 3. Link Prediction Survey 66

TABLE 3.5: Recall Results

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

CN 0.11363 0.30152 0.08918 0.23800 0.50078 0.40944 0.12500 0.52000
JC 0 0.01908 0.02540 0.31800 0.52007 0.07722 0.08750 0.60322
AA 0.10000 0.28320 0.09945 0.22600 0.52125 0.40444 0.11250 0.67419
RA 0.02500 0.27328 0.09513 0.23400 0.52795 0.45833 0.13125 0.70709
PA 0.05000 0.33053 0.05459 0 0.10984 0.32611 0.03125 0.00129
SALTON 0.17500 0.27709 0.09891 0.23600 0.51181 0.39000 0.13750 0.52129
SORENSON 0.15000 0.27633 0.08648 0.25200 0.50039 0.37333 0.13125 0.53870
CAR 0.20000 0.27251 0.09243 0.26200 0.51850 0.38333 0.12500 0.54129
CAA 0.15000 0.28015 0.10594 0.33600 0.52362 0.38611 0.13750 0.58000
CRA 0.07500 0.27709 0.11459 0.31800 0.56732 0.43888 0.13125 0.61806
CPA 0.11111 0.29313 0.10108 0.20400 0.48858 0.38555 0.08750 0.33225
HPI 0.10000 0.28167 0.07783 0.23800 0.51259 0.40444 0.10625 0.50451
HDI 0.15000 0.28473 0.09405 0.24600 0.48897 0.38277 0.18750 0.52129
NLC 0.05000 0.30763 0.07351 0.08800 0.44803 0.39444 0.05000 0.00516
LNBCN 0.10000 0.00305 0.08972 0.26200 0.38897 0.40555 0.11250 0.58322
LHNL 0.12500 0.27251 0.09297 0.24400 0.49409 0.41222 0.09375 0.52451
CCLP 0.05000 0.28015 0.09675 0.29000 0.52244 0.40555 0.10625 0.60000
KATZ 0.05000 0.34122 0.08162 0.20600 0.44212 0.39722 0.10625 0.43741
RWR 0.10000 0.38855 0.10216 0.24600 0.33937 0.08222 0.06000 0.30925
Shortest Path 0 0.09160 0.02702 0.03200 0.02007 0.02111 0.02000 0.13868
LHNG 0 0 0 0.36200 0.10669 0.00388 0.01250 0.05185
ACT 0.02500 0.30076 0.04972 0.03600 0.15748 0.33444 0.02000 0.20740
NACT 0 0 0.00540 0.32800 0.33740 0.00888 0 0.34024
Cos+ 0.02500 0.20610 0.04540 0.30400 0.13464 0.01888 0.02000 0.07037
MF 0.05000 0.19923 0.04324 0.30200 0.15590 0.04111 0.10000 0.41642
SPM 0.10000 0.51297 0.16216 0.28000 0.65000 0.47111 0.15000 0.63161
L3 0.05000 0.38549 0.11189 0.20200 0.34409 0.37777 0.07500 0.34645
LP 0.15000 0.38931 0.10594 0.23000 0.36692 0.40000 0.13125 0.37677

Macaque, where SPM performs well. We also observe that the Laplacian eigenmaps
(Leig) and Isomap are the worst performers on all datasets.

Area under the receiver operating characteristics curve (AUROC) The AUROC
(or AUC) results have been reported in the Table 3.7. Here, we observe that the global
approaches perform best on Macaque, C. Elegans, Football, jazz, and dolphin, while the
RWR is the best-ranking algorithm on C. Elegans and Dolphin, the SPM is the best
ranker on the Macaque and Jazz networks and Cos+ is best on Football. The table shows
the local approaches (i.e., RA and LNBCN) best result on usair97 and Netscience. The
bast performance of the RA index on usair97 is that this network is highly heterogeneous
with a higher clustering coefficient and absence of a strongly assortative linking pattern.
Chapter 3. Link Prediction Survey 67

TABLE 3.6: AUPR Results

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

CN 0.07030 0.28027 0.04234 0.17455 0.51238 0.39009 0.11986 0.58330
JC 0.01397 0.03590 0.01626 0.23470 0.51956 0.04744 0.05666 0.48951
AA 0.05710 0.27442 0.05181 0.16056 0.54088 0.39757 0.08579 0.74908
RA 0.04075 0.26368 0.04786 0.16674 0.56630 0.43537 0.10066 0.76599
PA 0.02765 0.34323 0.02085 0.00506 0.06746 0.28537 0.02237 0.00276
SALTON 0.12384 0.26720 0.04554 0.15954 0.52720 0.36980 0.10740 0.58826
SORENSON 0.06751 0.25750 0.04205 0.17329 0.51526 0.36370 0.10011 0.60575
CAR 0.21733 0.25305 0.04368 0.18561 0.53572 0.36416 0.10058 0.59813
CAA 0.16841 0.27093 0.04799 0.25907 0.55630 0.35679 0.06648 0.63863
CRA 0.10457 0.26834 0.05257 0.22227 0.60831 0.42311 0.08433 0.66808
CPA 0.09365 0.28103 0.04094 0.13220 0.50689 0.36225 0.05561 0.29581
HPI 0.03955 0.26042 0.04161 0.17890 0.52255 0.38245 0.09205 0.57262
HDI 0.10237 0.27818 0.04596 0.16474 0.50829 0.37266 0.17651 0.48951
NLC 0.07886 0.29955 0.03999 0.05210 0.41044 0.35726 0.04364 0.00123
LNBCN 0.07839 0.02712 0.03942 0.17575 0.41752 0.39835 0.05266 0.65339
LHNL 0.07935 0.24733 0.04664 0.17493 0.51164 0.39038 0.09565 0.59308
CCLP 0.05137 0.26904 0.04862 0.21183 0.55269 0.39442 0.07150 0.68109
KATZ 0.07354 0.33081 0.04047 0.16959 0.43602 0.38977 0.08983 0.51171
RWR 0.07874 0.38532 0.06197 0.21580 0.25769 0.09175 0.04901 0.20418
Shortest Path 0.01768 0.06612 0.01401 0.02278 0.02293 0.01120 0.02404 0.09914
LHNG 0.01393 0.02578 0.00781 0.30867 0.08835 0.00593 0.02896 0.04370
ACT 0.02262 0.31290 0.01823 0.01571 0.10025 0.29573 0.03442 0.17097
NACT 0.01379 0.02903 0.00725 0.25823 0.22698 0.00594 0.01488 0.19203
Cos+ 0.03392 0.19253 0.02463 0.23921 0.11242 0.02406 0.02476 0.02591
MF 0.06056 0.18374 0.02706 0.22821 0.14706 0.04203 0.04985 0.37889
SPM 0.08345 0.54004 0.08662 0.20540 0.68789 0.44342 0.08450 0.67826
L3 0.07943 0.38737 0.04389 0.13276 0.29754 0.35378 0.05865 0.32287
LP 0.06738 0.38600 0.04183 0.17927 0.31623 0.37741 0.07059 0.38762

One more thing to note that the PA index works well on networks that follow rich-club
phenomenon but, here we observe that the auroc results (please see Table 3.7) on
netscience dataset (having a member of rich-club phenomenon), the PA and its CAR
version (i.e., CPA) are having lowest values compared to all other methods. The reason
is that this network is disconnected (consists of many connected components), and hence
many nodes are isolated and lower degree values. The Quasi-local method viz., the path
of length 3 (i.e., L3), is the best performer on the karate network. On these datasets, only
quasi-local and global approaches lie in top-5. The quasi-local methods lie among top-5
on almost all datasets considered here.

The auroc results of other representative methods are shown in the Table 3.11, where SPM
is the best performer on Macaque, C. Elegans, and Jazz networks. HSM performs best
Chapter 3. Link Prediction Survey 68

on Football and Dolphin networks, CCLP is the best method on USAir97 and Netscience
networks. On Karate, Isomap is the best performing similarity index.

TABLE 3.7: AUROC Results

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

CN 0.66139 0.78749 0.87663 0.86928 0.95854 0.96328 0.81040 0.99832
JC 0.60817 0.40191 0.81608 0.85834 0.96612 0.93311 0.77303 0.99945
AA 0.65683 0.78724 0.88189 0.84932 0.96488 0.97402 0.74429 0.99932
RA 0.72318 0.79259 0.88736 0.85685 0.97466 0.97728 0.78629 0.99953
PA 0.67080 0.91731 0.76145 0.25727 0.76620 0.91754 0.66843 0.74877
SALTON 0.71231 0.78275 0.86753 0.85063 0.96092 0.96565 0.76276 0.99933
SORENSON 0.66563 0.77183 0.86141 0.85060 0.95879 0.96409 0.75858 0.99919
CAR 0.50134 0.78093 0.84277 0.84644 0.96118 0.95233 0.68292 0.95230
CAA 0.45998 0.79080 0.83838 0.84661 0.96122 0.95184 0.66276 0.94446
CRA 0.51796 0.78359 0.84652 0.83926 0.96944 0.96218 0.66596 0.94907
CPA 0.62594 0.81236 0.77510 0.67895 0.94639 0.91978 0.57540 0.76972
HPI 0.65191 0.78251 0.86798 0.88764 0.96274 0.96466 0.75740 0.99957
HDI 0.74994 0.78996 0.87213 0.86098 0.96041 0.96735 0.80747 0.99887
NLC 0.70911 0.84349 0.86321 0.80929 0.95393 0.90702 0.73127 0.59213
LNBCN 0.41790 0.21923 0.75164 0.78726 0.87271 0.96541 0.60227 0.99964
LHNL 0.70088 0.77854 0.86965 0.85957 0.95841 0.96452 0.75054 0.99937
CCLP 0.67934 0.79209 0.88046 0.86181 0.96510 0.96933 0.80289 0.99849
KATZ 0.73788 0.86200 0.86363 0.86480 0.94763 0.96276 0.80212 0.99934
RWR 0.84177 0.92849 0.90660 0.90110 0.95999 0.97113 0.88020 0.99347
Shortest Path 0.61283 0.61649 0.79575 0.75712 0.68467 0.83586 0.85132 0.95818
LHNG 0.64006 0.13196 0.72447 0.89802 0.90150 0.73105 0.77995 0.98541
ACT 0.51413 0.86292 0.74735 0.56118 0.80132 0.92371 0.81672 0.94354
NACT 0.67774 0.23142 0.65767 0.90085 0.91021 0.69823 0.78073 0.94845
Cos+ 0.80465 0.70811 0.86812 0.90329 0.91351 0.91748 0.82919 0.95781
MF 0.75465 0.71741 0.87700 0.89923 0.92916 0.95041 0.84830 0.95504
SPM 0.74565 0.95551 0.90499 0.84371 0.97807 0.95203 0.75200 0.99148
L3 0.84751 0.91431 0.84857 0.84796 0.91036 0.93500 0.77610 0.97264
LP 0.76661 0.91639 0.84942 0.87907 0.91572 0.94884 0.78649 0.99915

Average precision Table 3.8 shows the average precision results of similarity-based
methods on eight datasets. The global approaches here, also are the best performer on
all datasets except Karate and usair97, and dolphin where the quasi-local index (L3), and
local indices RA and HDI respectively are the best. Here, SPM performs overall best on
Macaque, C. Elegans, and jazz networks. The Resource allocation and HDI methods are
top rankers on usair97 and dolphin networks, respectively.
Chapter 3. Link Prediction Survey 69

Table 3.12 represents the average precision results of the other representative methods.
From the table, it is observed that the Node2vec shows the highest average precision
values against all networks except Macaque and Dolphin, where SPM and CCLP
respectively show the best results.

TABLE 3.8: Average Precision Results

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

CN 0.01821 0.11240 0.01319 0.03005 0.06329 0.01728 0.02566 0.00113
JC 0.01164 0.03315 0.00988 0.03158 0.06432 0.01179 0.02240 0.00112
AA 0.01585 0.11140 0.01381 0.02882 0.06464 0.01764 0.02238 0.00117
RA 0.01775 0.11130 0.01380 0.02942 0.06649 0.01850 0.02538 0.00118
PA 0.01457 0.13441 0.01004 0.00373 0.03275 0.01556 0.01523 0.00032
SALTON 0.01938 0.11038 0.01321 0.02895 0.06384 0.01709 0.02419 0.00112
SORENSON 0.01658 0.10809 0.01288 0.02963 0.06347 0.01705 0.02333 0.00112
CAR 0.01882 0.10872 0.01277 0.02994 0.06429 0.01688 0.02151 0.00104
CAA 0.01644 0.11117 0.01291 0.03153 0.06465 0.01692 0.01889 0.00104
CRA 0.01556 0.11031 0.01335 0.03064 0.06685 0.01782 0.01967 0.00104
CPA 0.01932 0.11484 0.01164 0.02437 0.06280 0.01653 0.01514 0.00064
HPI 0.01461 0.10972 0.01294 0.03078 0.06397 0.01714 0.02288 0.00112
HDI 0.02094 0.11184 0.01334 0.02948 0.06339 0.01722 0.02800 0.00112
NLC 0.01725 0.12147 0.01303 0.02145 0.06099 0.01615 0.02017 0.00020
LNBCN 0.01141 0.01766 0.01185 0.02665 0.05660 0.01760 0.01864 0.00114
LHNL 0.01830 0.10794 0.01317 0.02979 0.06336 0.01733 0.02323 0.00112
CCLP 0.01529 0.11138 0.01371 0.03056 0.06488 0.01754 0.02424 0.00115
KATZ 0.01915 0.12693 0.01284 0.02957 0.06022 0.01723 0.02383 0.00108
RWR 0.02122 0.13976 0.01436 0.03425 0.05601 0.01427 0.01647 0.00511
Shortest Path 0.01252 0.06152 0.00914 0.01551 0.02105 0.00731 0.01347 0.00142
LHNG 0.01212 0.02209 0.00687 0.03401 0.04081 0.00520 0.01889 0.00391
ACT 0.01155 0.12567 0.00950 0.01200 0.03723 0.01554 0.01414 0.00354
NACT 0.01244 0.02610 0.00650 0.03337 0.05158 0.00519 0.01091 0.00359
Cos+ 0.01698 0.09097 0.01154 0.03260 0.04411 0.00970 0.01290 0.00324
MF 0.01748 0.09138 0.01193 0.03243 0.04785 0.01174 0.01561 0.00166
SPM 0.01960 0.15628 0.01565 0.03046 0.06972 0.01798 0.02276 0.00115
L3 0.02324 0.13855 0.01256 0.02746 0.05304 0.01644 0.02174 0.00095
LP 0.01951 0.13865 0.01258 0.03025 0.05424 0.01689 0.02297 0.00103

Concluding remarks The above results of similarity-based methods on several datasets

show that the performance of each technique strongly depends on the structural properties
of the network. This highlights the importance of analyzing the properties of the network
before choosing a particular link prediction technique. As we observe in our results, the
quality of the results is related to the average clustering coefficient of the nodes with
degree above one. This is reasonable since most link prediction techniques are variations
Chapter 3. Link Prediction Survey 70

of counting shared neighbors, and the count of shared neighbors increases as the clustering
coefficient does. Another variable that seems to play an important role is the average
degree. This makes sense since as we know the more neighbors there are of a node, the
more information we have to predict new links for it. However, revealing which specific
properties play such an important role in link prediction is still an unsolved problem that
requires further work.

Parameters settings We conducted 10-fold cross-validation to evaluate each method

on four different evaluation metrics described in the earlier subsection. The disadvantage
with the global approaches is parameters tuning that needs to be done carefully to obtain
good results. The dumping parameter β of the Katz index is set to 0.01, the return
probability (1 − c) = 0.3 in the random walk with restart method. The φ of the global
version of Leicht-Holme-Newman, i.e., LHNG, is set to 0.5 that equally balances both
self and neighborhood similarity terms. The free parameter ε = 0.5 and path up to the
length 3 is considered in the local path index.

TABLE 3.9: Recall results for other representative methods

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

HSM 0.07500 0.34885 0.07405 0.24400 0.29606 0.22666 0.17000 0.17407
Leig 0.02500 0.07022 0.01405 0.06000 0.12007 0.01777 0.02000 0.08333
Isomap 0 0.01221 0.01081 0.04200 0.10433 0.01444 0.02000 0.19814
Node2vec 0.03898 0.67234 0.02167 0.09887 0.09738 0.02741 0.01947 0.08102
CCLP 0.05000 0.28015 0.09675 0.29000 0.52244 0.40555 0.10625 0.60000
CCLP2 0.10000 0.40305 0.09675 0.32600 0.41850 0.38555 0.12500 0.41419
NLC 0.05000 0.30763 0.07351 0.08800 0.44803 0.39444 0.05000 0.00516
SPM 0.10000 0.51297 0.16216 0.28000 0.65000 0.47111 0.15000 0.63161

TABLE 3.10: AUPR results for other representative methods

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

HSM 0.06145 0.33552 0.03442 0.18720 0.23787 0.15064 0.12695 0.14478
Leig 0.03166 0.05281 0.00944 0.04602 0.07086 0.01360 0.01759 0.04326
Isomap 0.03072 0.03123 0.01084 0.03013 0.06972 0.01223 0.01948 0.11020
Node2vec 0.90000 0.08058 0.72930 0.79032 0.91563 0.84788 0.65000 0.85818
CCLP 0.05137 0.26904 0.04862 0.21183 0.55269 0.39442 0.07150 0.68109
CCLP2 0.11750 0.40541 0.04602 0.27284 0.44192 0.36988 0.08118 0.41091
NLC 0.07886 0.29955 0.03999 0.05210 0.41044 0.35726 0.04364 0.00123
SPM 0.08345 0.54004 0.08662 0.20540 0.68789 0.44342 0.08450 0.67826
Chapter 3. Link Prediction Survey 71

TABLE 3.11: AUROC results for other representative methods

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

HSM 0.78390 0.91606 0.84022 0.88921 0.87704 0.92754 0.85570 0.97286
Leig 0.72189 0.50369 0.70811 0.81513 0.81454 0.81281 0.77543 0.91977
Isomap 0.80300 0.32430 0.74469 0.79262 0.85134 0.81146 0.81074 0.96592
Node2vec 0.76850 0.63184 0.80230 0.85278 0.87941 0.85538 0.71474 0.89241
CCLP 0.67934 0.79209 0.88046 0.86181 0.96510 0.96933 0.80289 0.99849
CCLP2 0.74751 0.91921 0.84180 0.87319 0.93687 0.94741 0.77198 0.97351
NLC 0.70911 0.84349 0.86321 0.80929 0.95393 0.90702 0.73127 0.59213
SPM 0.74565 0.95551 0.90499 0.84371 0.97807 0.95203 0.75200 0.99148

TABLE 3.12: Average precision results for other representative methods

Methods Karate Macaque C. Elegans Football Jazz USAir97 Dolphin Netscience

HSM 0.01995 0.13304 0.01201 0.03063 0.04953 0.01455 0.01862 0.00427
Leig 0.01585 0.05008 0.00751 0.02061 0.03569 0.00776 0.01124 0.00325
Isomap 0.01764 0.02514 0.00816 0.01789 0.03630 0.00747 0.01200 0.00382
Node2vec 0.04780 0.08355 0.02206 0.10330 0.09875 0.02816 0.02305 0.08175
CCLP 0.01529 0.11138 0.01371 0.03056 0.06488 0.01754 0.02424 0.00115
CCLP2 0.02145 0.14074 0.01259 0.03241 0.05924 0.01669 0.02390 0.00100
NLC 0.01725 0.12147 0.01303 0.02145 0.06099 0.01615 0.02017 0.00020
SPM 0.01960 0.15628 0.01565 0.03046 0.06972 0.01798 0.02276 0.00115

3.5.3 Efficiency

We have performed our experiment on a 64-bit core i7 Intel system having 8 GB internal
memory and 3.60 GHz speed without a dedicated graphics card. To reduce the
computational time, some optimization strategies can be applied (if possible) for
example, the union and intersection of two sets of sizes m and n can be computed in
O(m + n) using hash tables. The computational complexity of the addition and the
subtraction of two matrices are O(n2 ), however, these operations can be performed in
O(nt) in sparse networks where, t << n. The matrix multiplication of two dense
matrices of sizes m × n and m × p are done in O(mnp), while it is O(mt p), where t << n.
The matrix inversion typically takes O(n3 ) time for a square matrix of size n × n
however, some improvements are available that reduce the time to O(n2.81 ) or even less.
The time complexity of the similarity-based methods have been tabulated in the Table
3.13, in which most complexities are explained in [45]. In the table, computational
Chapter 3. Link Prediction Survey 72

complexity of each method are shown using big O notation where n, m, and K are the
number of nodes, number of links, and average degree of the networks.

TABLE 3.13: The computational Complexity of similarity-based methods and the

corresponding references

Method Time Complexity Reference

Local Similarity Index
CN O(nK 3 ) [17]
JC O(nK 3 ) [62]
AA O(nK 3 ) [58]
RA O(nK 3 ) [59]
PA O(nK 2 ) [57]
Salton O(nK 3 ) [65]
Sorenson O(nK 3 ) [66]
CAR O(nK 4 ) [68]
CAA O(nK 4 ) [68]
CRA O(nK 4 ) [68]
CPA O(nK 3 ) [68]
HPI O(nK 3 ) [15]
HDI O(nK 3 ) [15]
LNBCN O(n.O( f (z) + nK 3 )) [69]
LHNL O(nK 3 ) [70]
CCLP O(n2 K 2 ) [35]
NLC O(nK 3 ) [71]
Global Similarity Index
Katz O(nK 3 ) [55]
RWR O(cn2 K) [72]
Shortest Path O(nmlogn) [17]
LHNG O(cn2 K) [70]
ACT O(n3 ) [81]
NACT O(n3 ) [81]
L+ O(n3 ) [77]
MF O(n3 ) [83]
SPM O(n3 ) [183]
Quasi-local similarity Index
LP O(ln2 K) [53]
L3 O(n3 ) [54]
Chapter 3. Link Prediction Survey 73

3.6 Variations of link prediction problem

As earlier mentioned that the techniques listed in this work mainly focus on a simple
abstract graph (i.e., a graph with no vertex or edge attribute). The networks considered
in this work simple undirected and unweighted. However, some modification needs to be
done to apply on weighted and directed networks. In such networks, links are assigned
with weights that represent the strengths of these links. Two types of link direction can
be possible of a node (i.e., incoming and outgoing). So, a node x can have two types
of neighbors (degrees) viz., in-neighbors Γi (x) and out-neighbors Γo (x). Based on these
modifications, earlier similarity approaches can be redefined as given below.

3.6.1 Link prediction in weighted and directed networks

In a directed network, the common neighbor method based on in-neighbors and

out-neighbors are expressed as

Si (x, y) = |Γi (x) ∩ Γi (y)|, (3.70)

and
So (x, y) = |Γo (x) ∩ Γo (y)|. (3.71)

In weighted directed network, the expression are

w(z, x) + w(z, y)
Siweight (x, y) = ∑ , (3.72)
z∈(Γ (x)∩Γ (y))
2
i i

and
w(x, z) + w(y, z)
Soweight (x, y) = ∑ . (3.73)
z∈(Γo (x)∩Γo (y))
2

In a similar way, other approaches can be modified for directed and weighted networks.
The point to be noted here is that first, define several topological features (e.g., degree,
Chapter 3. Link Prediction Survey 74

path, clustering coefficient, etc.) in weighted directed networks and apply these features
to implement several link prediction algorithms.

Mostly works on link prediction focus mainly on simple undirected networks due to
simplicity. The cost for this simplicity is that it fails to extract rich information available
in most real-world networks, which are not undirected in general. Some notable works
on directed weighted networks are nicely presented in [184–194]. Lichtenwalter et al.
[184] work, published in SIGKDD, extracts 12 topological features on a large directed
weighed network of over 5 million nodes and performs ensembles of classification
algorithms (C4.5, J48, Naive Bayes). The training over such a big network (millions of
edges) is problematic; to mitigate this issue, the authors defined edge features of vertices
of 2 and 4 hops only. They perform a quasi-local training to obtain the final model, and
their results are outperforming compared to the state-of-the-art. Further, Bütün et al.
[189] proposed a new topological similarity metric published in ASONAM that takes
into account temporal and weighted information in directed networks which are useful
for the improvement of the accuracy. They extract all possible triad pattern features and
incorporate them with the weighted version of baseline topological similarity metrics
(CN, JC, AA, RA, and PA). They employed a supervised learning framework using these
metrics as features and predict missing links. Recently, Bütün et al. [191] introduced a
supervised learning model for predicting the citation count of scientists (PCCS). They
formulate the problem of PCCS as a link prediction problem and predict links with their
weights in weighted temporal and directed (citation) networks. Their model incorporated
both local and global topological features and claims the excellence of their proposed
work.

The link prediction problem is being explored in several other types of networks, including
temporal networks, signed social networks, heterogeneous networks, bipartite networks,
etc. Some of these variations are studied in this section.
Chapter 3. Link Prediction Survey 75

3.6.2 Link prediction in temporal networks

Today‘s scenario shows that the relationships among users in social networks are
continuously changing; for example, each time in the Facebook network, some users
join, and some others quit. It results in the networks to be highly complex. Here, time is
an important parameter to consider for the evolution of networks. In temporal link
prediction, time is considered as the third dimension and represented by a third-order
tensor A.

1

if node i is connected with node j at time T,
A(i, j, T ) = (3.74)
0

Otherwise

Thus, for a given sequence of snapshots of a network at different time interval T1 , T2 , ...
Tt , the link prediction finds links that evolves at the next time slot Tt+1 .

Several efforts have been employed by the researchers in this direction in the last decade.
Purnamrita et al. [195] introduced a nonparametric method for temporal network link
prediction where the time dimension is partitioned into subsequences of snapshots of the
graph. This approach predicts links based on topological features and local neighbors.
Dunlavy et al. [196] employ matrix and tensor techniques in a framework where matrix
part collapses sequence of snapshots of networks into a single matrix and computes link
scores using truncated svd and extended Katz methods. The tensor part computes the
scores using heuristics and temporal forecasting. The tensor part captures the temporal
patterns effectively in the network, but it costs heavily also. Moreover, Gao et al. [197]
proposed a model based on latent matrix factorization that employs content values with
the structural information to capture the temporal patterns of links in the networks. Table
3.14 shows some more works in this direction.
Chapter 3. Link Prediction Survey 76

TABLE 3.14: Link Prediction in Temporal Networks

Models Network Types Characteristics References

Vu et al.[198], Pujari et
al.[199], Zeng et
High computational al.[200], He et al.[201],
Learning-based models Coauthorship networks
cost Bao et al.[202],
Madadhain et al.[203],
Bringmann et al. [204]
Twitter, Collaboration
Heuristics-based Fast convergence and Catherine et al.[205],
and Coauothorship
models high precision Sherkat et al. [206]
networks
Characterize the
Hu et al.[207], Barbieri
stochastic and dynamic
et al.[208], Gao et
Nodes-attributed relations. Need prior
Probabilistic model al.[197], Ji Liu et
graphs link distribution so
al.[209], Hanneke et al.
impractical for real
[210]
networks

3.6.3 Link prediction in bipartite networks

Till now, we have reviewed link prediction methods in unipartite networks in which links
may present between any pair of vertices. Now, we review the link prediction problem in
a specific graph where only two sets of vertices are present, and a link can be possible
between a pair of vertices in which one vertex belongs to one set of vertices and the other
vertex to another set of vertices. Such types of networks are called bipartite networks.
Lots of social networks logically can be considered as bipartite such as Term-Document
network [211], Scientists-Papers cooperation network [212], RNA-PI network [213],
IMDb network, and many more.

Kunegis et al. [214] study the link prediction problem in bipartite networks and observed
that most common neighbor-based approaches (e.g., Common Neighbors, Adamic/Adar,
Resource Allocation, etc.) are not applicable to these networks. The reason is that
adjacent nodes belong to different clusters and are connected with the path of odd
lengths only. Also, common neighbor-based approaches are based on the path of length
two. The authors give hyperbolic Sine and Von Neumann kernels of odd order to
compute the similarity between vertices. Only the PA method is applicable to these
networks in its natural form because it considers the degree of the neighbors. Some
Chapter 3. Link Prediction Survey 77

researchers [215–217] have implemented common neighbor-based methods (e.g., CN,

AA, RA, PA, LCP-CN, etc) in bipartite networks. Xia et al. [215] studied the link
prediction problem by exploiting structural holes in bipartite networks. They proposed
two implementations of structural holes viz., absent links (consisting of c-type and s-type
links), and minimum description length [218, 219].

Recently several methodologies of link prediction in bipartite networks have been

addressed. Baltakiene et al. [220] implemented maximum entropy principle, an
extension of the recent one [167]. They used probability of Bipartite Configuration
Model [221] as the score function. Allali et al. [222] presented the term ”internal link”
based on which they proposed a new link prediction algorithm.

3.6.4 Link prediction in heterogeneous networks

Most of the contemporary approaches of link prediction focus on homogeneous

networks where the object and the link are of single (same) types such as author
collaboration networks. These networks comprise less information, like which two
authors have collaborated with a paper that causes less accuracy for the prediction task.
In heterogeneous networks, the underlying assumption of a single type of object and
links does not hold good. Such networks contain different types of objects as well as
links that carry more information compared to homogeneous networks and hence more
fruitful to link prediction, also called multi-relational link prediction (MRLP). Examples
of such networks are DBLP bibliography 9 and Flickr networks 10 . In the bibliography
database, authors, papers, venue, terms are different types of objects/nodes, and
relationships are paper-author, author-author, paper-term, paper-venue, and so on.

Sun et al. [223, 224] coined the concept of heterogeneous information network (HIN)
and subsequently meta path concept [225], since then it becomes popular among

9 https://dblp.uni-trier.de
10 http://www. flickr.com/
Chapter 3. Link Prediction Survey 78

researchers. The key idea to multi-relational link prediction (MRLP) is to employ an

appropriate weighting scheme to combine different link types. The authors predict the
relationship building time between two objects by encoding the target relation and
topological features to meta paths in a supervised framework. Moreover, Yang et al.
[226] proposed a new topological feature, namely multi-relational influence propagation
to capture the correlation between different types of links and further incorporate
temporal features to improve link prediction accuracy. Davis et al. [227] proposed a
novel probabilistic framework, a weighted extension of Adamic/Adar measure. Their
approach is based on the idea that the non-existing node pair forms a partial triad with
their common neighbor, and their probabilistic weight is based on such triad census.
Then the prediction score is computed for each link type by adding such weights.
Meanwhile, Sun et al. [228] a new supervised framework for HIN where meta
path-based topological features (i.e., path count, random walk) are extracted, and then
logistic regression is applied to build the relationship prediction model that learns the
weight associated with these features. Table 3.15 lists some more works on link
prediction in heterogeneous networks.

3.7 Link prediction applications

3.7.1 Network reconstruction

Guimerà et al. [2] proposed a framework that applies link prediction for network
reconstruction. They reconstruct of the true network is done from the observed network
based on missing links (removed one) and the spurious links (added links). Although it is
not obvious because no one knows about the amount of missing and spurious links in the
networks. For this, the authors describe the reliability of networks based on the
reliability of both missing and spurious links by formulating the link prediction problem
Chapter 3. Link Prediction Survey 79

TABLE 3.15: Link Prediction in Heterogeneous Networks

Models Network Types Characteristics Reference

Proposed both
Youtube, Gene, unsupervised and
Davis et al. [227]
Climate supervised approach to
Supervised models
link prediction
Extracts meta
path-based topological
DBLP features and applies Y. Sun et al. [229]
logistic regression as
prediction model
Define social
pattern-based features
(social balance and
Epinions, Slashdot, microscopic
Y. Dong et al. [230]
Wikivote, Twitter mechanism), input to
the inference model
namely (transfer) factor
graph models
Non-parametric
Bayesian model that
MovieLens,
considers the similarity
Book-Crossing, B. Cao et al. [231]
Collective LP models between tasks when
Douban
leveraging all the link
data together
Distance feature
extraction usibg both
network and node
features and for
Flickr, DBLP S. Negi et al. [232]
learning Multi-Task
Structure Preserving
Metric Learning
(MTSPML) is used

as a stochastic block model [113]. The reliability of the network A is [2]

R(A) = ∏ Rxy = ∏ L(Axy = 1/Ao ), (3.75)

Axy =1,x<y Axy =1,x<y

where Rxy is the reliability of the link (x, y) that is defined by the likelihood that the link
(x, y) truly exists given the observed network Ao . This equation can be solved by finding
out the network A that maximizes the reliability given by 3.75.

The computational cost of the equation is high, so the authors [2] give a greedy
algorithm to compute it. The algorithm starts with computing the link reliability of all
pairs of vertices. At each step, the algorithm removes the least reliable link and adds the
Chapter 3. Link Prediction Survey 80

most reliable link (non-existing in the current network). This change in the network is
accepted when the reliability of the network increases and rejected otherwise. In case of
rejection, the next step selects the least reliable existing link and the highest reliable
non-existing links for swapping. The algorithm stops when there are no five consecutive
changes (swaps) in the network. The reliability of the network improves from the initial
observed network, which is the reconstructed ones. Now, the authors compare the six (6)
global properties of both the observed and the reconstructed networks and show that the
reconstruction improves the estimates.

3.7.2 Recommender system

The recommender systems [21, 22, 48, 49] (also called information filtering systems)
have been widely applied in social media (like Facebook, Twitter) and online shopping
websites (e.g., Flipkart, Amazon, etc.). Such systems recommend new friends,
followees, and followers on social networking platforms and new products on online
shopping portals based on users’ previous browsing history (such as interests,
preferences, ratings, etc.). Even though collaborative filtering (CF) is a successful
recommendation paradigm that applies transaction (Transaction/purchase is essentially
an implicit and coarse rating on preferring an item [233]), information to enrich user and
item features for recommendation. Although they have been applied in many
recommender systems, they are greatly limited by data sparsity problem [234]. The
recommender system in bipartite networks can be mapped to link prediction problem as
follows [235]. Consider U ∗ and O be the sets of users (first set of vertices) and
objects/items (second set of vertices). Construct the user-item interaction graph
G = (V, E) from the available transactions T (purchasing patterns), where V = U ∗ ∪ O
and E = {(u, o) : u ∈ U ∗ , o ∈ O, u → o ∈ T }.

Huang et al. [21] and Li et al. [236] proposed approaches, where the recommender system
(user-item recommendation) is represented as a bipartite graph, and employed basic link
prediction approaches for the items recommendation. Sadilek et al. [237] proposed FLAP
Chapter 3. Link Prediction Survey 81

(Friendship + Location Analysis and Prediction) system in which both friendship and
location prediction tasks are implemented. They employed users tweets, their locations,
and their neighboring information as model features and inferred social ties and location
using MRF. More related works can be found in [235, 238, 239].

3.7.3 Network completion problem

In general, the network representation of the real-world problem is incomplete or

partially observed or incremental with both missing links and nodes such as wall posts
on Facebook, tweets in the Tweeter, etc. The problem arises due to several reasons like
security, data aggregation overhead, manual errors, etc. Predicting such nodes and links
is the network completion problem in which some notable works like [240] in SIGKDD,
[241] in ASONAM, and [156] in EPL. Filling missing entries of the adjacency matrix of
a network is link prediction, which can be considered as a subset of network completion
problem. Kim et al. [242] cast network completion problem to the
Expectation-Maximization (EM) framework and proposed KronEM, an EM model based
on Kronecker graphs. They, first, represent the network as Kronecker graph and estimate
the model parameters as well as missing links using KronEM algorithm. The estimated
network is then considered as the complete network and re-estimate the model, and this
process is repeated until the convergence. Further, Pech et al. [156] employed the robust
principle component analysis (Robust PCA) [243] [to recover both low rank and sparse
components of a data adjacency matrix] in link prediction framework and introduce a
novel global prediction method using both the components. They reconstruct the original
network using the robust PCA where these components are extracted by minimizing the
weighted combination of the nuclear norm and of the l1 norm [243].

min kX ∗ k∗ + λ kE k1 , (3.76)
X ∗ ,E
Chapter 3. Link Prediction Survey 82

where k.k∗ is nuclear norm (i.e., sum of singular values) of the matrix and k.k1 is the
l1 norm, E is error or noise matrix (sparse matrix containing spurious links as positive
entries and missing links as negative entries) and λ is a positive weighing parameter that
balances the contribution of both the components (low rank and sparse components). X ∗
[= AO − E , here, AO is the observed network] is the set of patterns (links that are newly
predicted and some links that are eliminated). From which only the newly appeared links
are extracted and added to the observed network AO to recover the original matrix (also
known as reconstructed matrix). Once the reconstructed matrix is obtained, link prediction
can be performed accordingly.

3.7.4 Spam mail detection

Spreading and receiving irrelevant emails is common in today’s world that consumes
network bandwidth, memory, etc. Many email service companies are trying to
implement several filter mechanism to stop such emails known as spam mails. To
implement spam filter mechanism, spam detection is a necessary task. In this context,
Huang and Zeng [244] proposed a model to detect spam emails using link prediction.
They construct an email graph (directed and weighted) based on the email data,
consisting of a sender, recipient, and timestamp of the communication as attributes.
Many email communication links between the sender and the receiver are mapped to a
weight of the link between them. Then, an anomaly score is computed for each distinct
sender-recipient pair using the Adamic/Adar link prediction approach by making it
adaptive based on the spreading activation algorithm. Some more work related to this
can be found in [245, 246].

3.7.5 Privacy control in social networks

Lots of users share personal posts, audios, videos, and other sensitive information to
social networking websites. Trust is an important parameter to evaluate users’
Chapter 3. Link Prediction Survey 83

relationships on such media, i.e., the strength of a relationship between two users can be
determined based on the trust in the form of link weights. Thus, it is important for
companies to maintain the privacy of users from anomalous ones. Oufi et al. [247]
proposed a framework implementing a capacity-based algorithm that employs Advogato
trust metric [248, 249] to compute the level of trust between users. This means that the
framework identifies all possible trustworthy users of a seed user, which results in the
privacy of that user in the network from anomalous users.

3.7.6 Identifying missing references in a publication

A research article may contain some irrelevant references and miss some relevant ones.
Identifying such missing references in a research article is an important task to avoid
plagiarism. It becomes more critical for the point of a novice researcher due to a lack of
literature survey carried out by him. Kc et al. [250] proposed a machine learning
approach to link prediction tackle this problem. They provide a framework for the
generation of links between referenced and otherwise interlinked documents. The nodes
of the graph represent documents, and the links between them show references available
between them. They find new links/references of documents based on this graph using
Probability Measure Graph Self-Organizing Map (PM-GraphSOM).

3.7.7 Routing in networks

In complex network theory, link prediction in social networks resembles link quality
prediction in wireless sensor network [251]. The routing problem in a network finds the
shortest path (optimal) between the sender and the receiver. The strength of signal
frequently varies in mobile and ad hoc networks that results in frequent breaks in routes
and degrades the performance result. Weiss et al. [252] and Yadav et al. [253] proposed
some models to estimate the signal strength-based link availability prediction for optimal
routing. Such link information is beneficial to estimate the link breakage time and hence,
Chapter 3. Link Prediction Survey 84

to repair the existing route or to discover a new route for the packets. This reduces
end-to-end routing delay and packet drops, thereby improving the performance. Once
broken links are identified earlier (using link breakage time), routing management
protocol needs either to repair the broken link or to find an alternate route. Several works
state that link prediction may play a crucial role in this scenario that results in low
latency in packet delivery to the receiver and hence improves reliability. Hu and Hou
[254] presented link prediction-based traffic prediction for the best routing of packets in
a wireless network. Some more works in this area can be found in [255, 256]. Recently,
Zhao et al. [251] proposed a neighborhood-based NMF model to estimate the link
quality in the wireless sensor network. They extend the link prediction model to the
wireless sensor network, where they predict the quality of a link based on NMF
associated with structural (neighborhood) information.

3.7.8 Incorporating user’s influence in link prediction

Lots of works based on individual influence have been proposed in social network
analysis, such as link prediction [257, 258], information diffusion [259–261], influence
maximization [44, 262–266], community detection [267, 268], etc. Particularly, the role
of individual influence in link prediction provides a new perspective/insight into the
problem. Influence maximization (IM) [44] is one of the fundamental problems in social
network analysis where the goal is to find a set of users (seed set) that can be further
utilized to maximize the expected influence spread (defined as the expected number of
influenced users) among others. The influence (social influence here) is propagated
through certain channels (i.e., intermediate nodes), that are captured by diffusion models
[259]. IM and diffusion models are a cooperative and correlated task as for IM, several
Diffusion models are used in the computing framework. Zhang et al. [269] proposed a
new framework of link diffusion to predict more links in the microblogging networks.
They find the triadic structure to be the crucial factor that affects the link diffusion
process and hence, link prediction. Earlier, Cervantes et al. [270] proposed a supervised
Chapter 3. Link Prediction Survey 85

learning model to find an influential collaborative researcher in the collaboration

network. They employ the model to the whole network and compare its result with those
sub-networks generated each time when a distinct vertex is removed from the training
set. Finally, results are ranked and examine the collaborative influential of each
researcher based on the presence or absence of it in the network. Finding influential users
(i.e., the seed set) is useful in many applications like viral marketing, where an influential
user can be used to advertise the product to maximize the profit. Other application areas
may be disease prevention using vaccinate to the most influential patient.