VIETNAM NATIONAL UNIVERSITY - HO CHI MINH CITY

HO CHI MINH CITY UNIVERSITY OF TECHNOLOGY

FACULTY OF COMPUTER SCIENCE AND ENGINEERING

MATHS FOUNDATION for COMPUTER SCIENCE (055263)

Assignment

Community Structure Identification

(Version 0.2 for SEM231)

Instructors: Nguyen An Khuong, CSE-HCMUT

Tran Tuan Anh, CSE-HCMUT

Teaching Assistants: Le Thanh Son, KTH Royal Institute of Technology

Tran Dinh Vinh Thuy, École Centrale de Lyon

Ho Chi Minh, November 2023

Contents
1 Introduction 3

2 Community structure identification problem 3

3 Some use-cases 4
3.1 Social networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Customer networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.3 Collaboration networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4 Transportation networks: Multi-link graphical representation . . . . . . . . . . . . . . 5

4 Traditional approaches to the community structure identification problem 6

4.1 Hierarchical clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.2 Girvan-Newman algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.3 Louvain algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.4 Matrix factorization-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.5 Tensor factorization-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

5 Suggested datasets 7

6 Guidelines 7
6.1 Objectives (Questions) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.2 Requirements and Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

References 8
1 Introduction
The origin of graph theory can be dated back to when Leonard Euler first gave his proof on the
Königsberg seven bridges problem. By viewing the problem abstractly, using the notations of letters
and lines, not only did Euler find the solution, he was able to generalize this problem. The idea of
Euler on solving this problem put the stepping stones for the branch of graph theory. Since then, this
branch has been an important field of mathematics, with much work made to study graphs and their
mathematical properties. During the 20th and 21st century, there has been an explosion within the
ways of using graphs to model real-life problems. Graphs are now becoming more and more useful to
represent a wide variety of systems from different areas. For instance, a graph can be used to describe
the relations of one person to others on a social network by treating the person as a vertex and each of
their connections as an edge of a graph. Subsequently, a rising demand is to analyze the interactions
among elements of a graph.
Besides using the normal representation of graphs, with edges and vertices, to deal with practical
problems, we can take use of the description of graphs through matrices or tensors. Indeed, the duality
between graph and matrix multiplication motivates the use of graph approaches in practice. Assuming
we are performing a breadth-first search on a graph depicted in Figure 1. In the first step, we would
like to find all the adjacent nodes of Alice, which are Bob and Carl on the graph. We recall that the
relationship about adjacency between the nodes can be represented by an adjacency matrix A = (aij )
in which aij = 1 means that we have a connection from node i to node j and aij = 0 otherwise. The
matrix A⊤ (the transpose of A) is shown in Figure 1 where the dots denote the values 1. Then, by
using one-hot encoding vector v which is active at node Alice, we can observe that A⊤ v can be used
to represent the set of adjacent nodes {Bob, Carl}.

Figure 1: Duality between breadth-first search on a graph (left) and matrix multiplication (right) [1]

Graphical analysis has played a crucial part in understanding behaviors in various fields, such as
biology, sociology, and computer science. Several problems are defined for each case of investigating
such graph features.

2 Community structure identification problem

The one problem when using complex graphical networks to model real-life behaviors is that it tends to
have a high level of order and organization within the elements of the graph. These graphs can display
several clusters of density-connected vertices called communities. The vertices in one community
probably share some common properties and/or behave similarly within the graph. One community
is hence usually sparsely connected to other communities. Community structure identification in a
graph aims to detect communities (see [2] for precise definition1 ), in other words, clusters, or possibly
their hierarchy structure, given only the information of the graphical representation. It is one of the
most important problems in graphical analysis (see [3] and in [4, Sections 10.2, 10.3, 10.5] for more
detail), and it goes with many names in the literature: community detection, network clustering,
graph partitioning, etc.
1 See also at https://en.wikipedia.org/wiki/Community_structure and the references therein.

3
 For example, in the social network scenario, consider a guy who works for an American start-up,
goes to the tennis club every Friday and attends Chinese classes at a center. It is understandable to
suppose that everyone in the start-up knows each other well and the same thing can be said for people
from the tennis club and the Chinese classes. It is unlikely, however, to have almost every person in
the start-up goes to the same center to learn some Chinese or goes to the same club to play tennis at
the same time as the considering guy. This one particular instance illustrates that in real life, people
tend to form community groups and behave accordingly in each group with little to no connection
to other groups. Identifying such groups in the graphical representation enables the use of further
applications.

3 Some use-cases
3.1 Social networks
When speaking of social networks, we think of Facebook, Instagram, or Twitter, to name a few. In
these social networks, there are always relations between two entities residing in the network. On
Facebook or Instagram, one account can follow or befriend another. Naturally, we want to represent
the network of Facebook users as a graph, where each vertex represents one user. There will be
an edge between two vertices if the two users represented by the vertices have a relation. This
way of visualization can be enhanced if we want to represent more than just simple relationships
by adding weights to the edges. For example, the “followed” edges will weigh 1 while 2 will be for
the “befriended”. The communities in these networks will represent the groups that share things in
common, such as being members of the same family or sharing a similar hobby. Identifying these
communities is meaningful since we will have ideal targets to perform some specific operations. For
example, we may have higher revenue by promoting hotel discounts to people from a traveling group
than to those who are in a gaming group.
For further illustration, Figure 2 represents a simple example of a social network. There are three
communities, each with a different number of members, represented by the large lines encircling the
vertices. Also, we can observe that two of the communities are separated and two of them overlapped.
In practice, we can have also a community that entirely belongs to another community.

Figure 2: Social network using graphical representation

Besides social networks as discussed, there are numbers of other use cases that can be represented
in the same way such as the followings.

3.2 Customer networks

Suppose that we are running an online bookstore. Customers may be interested in more than one
type of book. They can find their interests in the latest fantasy thriller by Stephen King while sharing
some of their pleasures for mathematical textbooks. Finding which types of books each customer

4
invests in can help us improve our service through investigations in the detected communities. In this
case, each book is represented by a vertex in a graph and an edge between two books is created if
these two are bought by the same customer.

3.3 Collaboration networks

Consider an interdisciplinary research center located in Ho chi Minh City. We can construct a graph
whose vertices represent scientists in residence during any time window, say from 2013 to 2023.
An edge will be created between two vertices if the corresponding researchers appear in at least
one joint work during the given time. The communities will represent the people working on the
same particular topic or with similar methodologies. In this use-case, we can see that there can be
communities within a community and overlap communities. For instance, a research article proposing
a new algorithm can be coauthored by mathematicians for the theoretical parts and computer scientists
for the experimental sections. Besides, we can have various authors working in the same field, say
computer vision, but they can form into more specific tasks such as image denoising, object detecting,
etc. Finding such communities within the constructed graph can help tracing related works less
burden, which is sufficient in the first step of research.

3.4 Transportation networks: Multi-link graphical representation

In cities around the world, there are normally means of transportation to go from one city to another,
such as car, plane, or ship. These are considered to be relationships in a graph where each city is
a node. Hence, we place one edge for each line between the two cities. In this case, the graph is
more challenging to be analyzed compared to the one introduced in Subsection 3.1, which is called
multi-link graph [5, Section 7.1]. We are interested in finding the communities in this graph so we can
determine the similarities between cities, which help us in traveling or shipment services. Figure 3
illustrates a multi-link transportation graph with two detected communities. There is more than one
link between City 1 and City 2, hence, our graph is a multi-link graph. To represent this graph, we
can use the adjacent tensor.

Figure 3: Multi-link graph representation

As in Section 1, we also have the duality between the multi-link graph and the tensor. Thus, besides
some classical methods in [6, 7, 8], we can perform tensor decomposition described in [5, Chapter 7].

5
Tensor decomposition methods such as principal component analysis 2 or low-rank approximation 3
could also give insights about the latent structure.

4 Traditional approaches to the community structure identi-

fication problem
We introduce some approaches that can be applied in the community structure identification problem.
These are to get you the ideas of the approaches. You are encouraged to choose your preference
method.

4.1 Hierarchical clustering

Since community structure identification sometimes goes with the name of graph clustering, we can
apply some traditional graph clustering approaches such as hierarchical clustering (see Chapter 7 in
[4]). In this approach, the communities are created by the “closeness” between the vertices in the
graph. The definition of “closeness” may vary, depending on the particular algorithm.

4.2 Girvan-Newman algorithm

The Girvan-Newman algorithm [9, 10] detects communities by removing the edges iteratively according
to their highest “betweenness” score. This algorithm is based on the intuition we discussed in Section
2 that the communities are sparsely connected by a few edges located in the shortest path between
vertices. Thus, removing these edges will reveal the communities inside the graph. Some other
enhancements of this algorithm can be founded in [11] and [12].

4.3 Louvain algorithm

The idea of the Louvain algorithm [13] is based on the maximization of “modularity”, which is a
value measuring the density of edges inside a community. At each iteration, the algorithm places
the vertices to other communities until to attain maximum modularity and it stops when there is no
further improvement in modularity. There are also many improvements to this algorithm that can be
found on the Internet.

4.4 Matrix factorization-based methods

If we used adjacent matrix to represent the graph, we can apply the non-negative matrix factorization
methods (MNF) on the adjacent matrix to find the communities within the graph. The general idea
of this method is to factorize the adjacent matrix A by two smaller matrices W and H such that
A ≈ WH and W and H both have no negative elements. By examining the structures of W and
H, we can find identify the communities within the graph. Some particular research works on this
approach can be found in [14], [15], or [16].

4.5 Tensor factorization-based methods

When working with a multi-link graph, we can not use an adjacent matrix. Instead, we must use a
tensor to represent our graph. The traditional matrix factorization-based approaches obviously will
not work in this case. Hence, we will consider ways to factorize the tensor representing the graph. The
CP decomposition [17, Section 3] can be considered as a generalization of the SVD for tensor. Indeed,
with the given rank R, we can decompose a tensor X as the sum of R rank-one tensors. Precisely, each
factor represents a “community” within the data and the number of factors R in the approximation
should loosely reflect the number of communities in the data [5, Subsection 7.2.4]. There also exists
other means of tensor decomposition, such as Tucker decomposition [17, Section 4]. Examinating
the structures generated by these decompositions can help finding and understanding the underlying
communities of a graph.
2 https://en.wikipedia.org/wiki/Principal_component_analysis
3 https://en.wikipedia.org/wiki/Low-rank_approximation

6
5 Suggested datasets
We now give some datasets that can be used for the experimental parts of this assignment, note that
we do not state the explicit use case for these datasets and some of these datasets require additional
preprocessing before being able to put into your version of implementations.
• Dolphins online social network: A social network of bottlenose dolphins. The dataset contains
a list of all of links, where a link represents frequent associations between dolphins.
• Political books: A network of books about US politics published around the time of the 2004
presidential election and sold by the online bookseller amazon.com. Edges between books rep-
resent frequent co-purchasing of books by the same buyers.
• Jazz musicians: A network between Jazz musicians. Each node is a Jazz musician and an edge
denotes that two musicians have played together in a band.
• NIPS Conference Papers Vols 0-12: A dataset contain full information of papers published from
volume 0 to 12 of conference on neural information processing systems (NIPS). This dataset
requires preprocessing to create a meaningful graphical representation, for instance, the collab-
oration networks as in Section 3.
Several other datasets can be found here and here, each with corresponding description and download
link. Other datasets can also be found in the content of the papers in the References.
If you find the given suggestions hard to interpret or unsatisfied, you can find from other sources
that are not listed here or created your synthesis data. However, if you opt for this direction, please
indicate specifically where you found the dataset on the internet or how your dataset created , its
description (number of vertices and edges, some other prevelence statistics,...).

6 Guidelines
6.1 Objectives (Questions)
1. Choose a use-case that requires or can apply the idea of community structure identification to
solve (be careful in setting the problem and limiting the scope so that your team can manage
to solve it within 4 weeks). You can refer to some ideas in Section 3, and in fact, you can
find dozens of use cases related to the community structure identification problem through the
references in this assignment. You are encouraged to find or construct one use case of your own
if you find the given suggestions unsatisfied.
2. With the chosen use case, find or create a suitable dataset that best represents your problem,
see Section 5 for more details. There will be a penalty if your choice dataset is not specified.
3. Study and implement your chosen approach for this problem and experiment with your dataset
(see [18] for some of the most popular approaches, and see [19, Section 7.2] for sample codes in
R or Python).
4. Explain in detail the behaviors and the results of your approach, especially the motivation, the
approaches, the metrics, and the dataset you use.

6.2 Requirements and Instructions

1. Form your group of up to six people and at least four on BKel as soon as possible (preferably
the same with your HW team).
2. Write a technical report (advisedly in English if possible) by LaTeX4 justifying the group’s
use-case by modeling the problem using any graphical representation and specifying what com-
munities mean in your case. This report should also explain your data choices and approach
in-depth and give meaningful observations and interpretations of your results. The report should
4 See templates and sample available at https://www.overleaf.com/gallery/tagged/report

7
 also cite all of your references, whether it is an article, a research paper, a textbook, etc. The
structure of the report can be the following
(a) Chapter I. Introduction: Explain your motivation, and introduce your use-case with a
problem statement (what you are trying to do with this use-case) on a specific dataset.
(b) Chapter II. Preliminaries: Recall or present all the definitions and properties with
concrete examples of all foundations for later uses. [Doing this part carefully will be helpful
with your final examination.]
(c) Chapter III. Approaches: Present your choices of algorithm or method that can be used
to solve the problem with a detailed explanation.
(d) Chapter IV. Experiments: Perform numerical experiments using the implemented ap-
proach or approaches on your dataset and give your in-depth analysis.
(e) Chapter V. Conclusion: Summarize what you have done, your limitations, and the
directions for future works.
(f) References.
3. Write the codes (required) and prepare demonstrations (if any) your group made during the
time doing this assignment.
4. Prepare slides by Beamer5 explaining your group work thoroughly. The number of slides should
not be lengthy (at most 30 slides, preferably 20-25 slides or less) with adequate contents of your
work. All groups must give presentations with a reporting time of no more than 45 minutes for
each group, record it with your team’s faces6 , and submit the link together with the assignment
(selected teams will have a chance to give presentations during the two last lectures (scheduled as
make-up lectures and will take place in the last week of December 2023). You will be provided
with the entire content of all other groups before the presentation day. Based on that, each
group must prepare at least two questions and two comments/suggestions for all other groups
to give the presentation on another day.
5. All of the meeting minutes of your group must be combined into one .txt file and consist of the
% of each member’s effort on the whole (total effort is 100%) on the first line.
6. Please compress all materials relating to your work as mentioned above in one .zip file and
only the team leader submits it to the e-learning site of this course.7 You will have 4 weeks
to do this project, starting from December 01, 2023 (hard deadline, there will be no
extension.) Note that plagiarism is strictly prohibited and will be handled accordingly.
7. Some basic technical backgrounds related to this assignment will be asked in the final exam.
Therefore, team members must work together so that all of you understand all aspects of the
project. The team leader should organize the team to meet this requirement.
This assignment is two-fold:
1. You are expected to find at least an applicable use-case that can be modelled by a single graph
for the community structure identification problem, such as the ones introduced in Section 3.1,
and implement at least one algorithm from the suggestions in Section 4 to solve this problem
with your use case.
2. At a more advanced level, you can work a use-case where a multi-graph must be used to model
and implement some appropriate algorithms to find the answers to your problem. You can also
tackle a case where the graph’s adjacent matrix (or tensor) representation is sparse. If you opt
for this direction, you can work with a group of less than 4 (strong) members.
If you are not used to working with graphs or/and having trouble comprehending concepts in
English. In that case, a suggestion is to support with [3], rather well-written in Vietnamese and
provided with several explicit computational examples, to grasp the main idea of this assignment
before continuing working.
5 Seetemplates and samples available at https://www.overleaf.com/gallery/tagged/presentation
6 Seea sample at https://www.youtube.com/watch?v=L263tOAkWC0
7 The assignment link on BKeL

8
References
[1] J. Kepner, D. A. Bader, T. Davis, R. Pearce, and M. M. Wolf, “Graphblas and graphchallenge
advance network frontiers,” SIAM News, vol. 55, no. 08, October 2022, URL:
https://sinews.siam.org/Details-Page/graphblas-and-graphchallenge-advance-netwo
rk-frontiers.

[2] S. Fortunato and C. Castellano, Community Structure in Graphs, pp. 490–512. New York, NY:
Springer New York, 2012, URL: https://doi.org/10.1007/978-1-4614-1800-9_33.

[3] N. T. Duc, “Community detection on social graphs, B. Eng Thesis, HCMUT, VNU-HCM (in
vietnamese),” 2018, URL: https://tinyurl.com/TanDucThesis.

[4] J. Leskovec, A. Rajaraman, and J. D. Ullman, Mining of massive data sets. Cambridge University
Press, 2020, URL: https://www.mmds.org/.

[5] J. Kepner and J. Gilbert, Graph algorithms in the language of linear algebra. SIAM, 2011, URL:
https://doi.org/10.1137/1.9780898719918.

[6] J. Kim and J.-G. Lee, “Community detection in multi-layer graphs: A survey,” ACM SIGMOD
Record, vol. 44, no. 3, pp. 37–48, 2015.

[7] L. Getoor and C. P. Diehl, “Link mining: a survey,” Acm Sigkdd Explorations Newsletter, vol. 7,
no. 2, pp. 3–12, 2005.

[8] L. Getoor, N. Friedman, D. Koller, and B. Taskar, “Learning probabilistic models of link struc-
ture,” Journal of Machine Learning Research, vol. 3, no. Dec, pp. 679–707, 2002.

[9] M. Girvan and M. E. Newman, “Community structure in social and biological networks,”
Proceedings of the national academy of sciences, vol. 99, no. 12, pp. 7821–7826, 2002, URL:
https://www.pnas.org/doi/10.1073/pnas.122653799.

[10] M. E. Newman and M. Girvan, “Finding and evaluating community structure in networks,”
Physical review E, vol. 69, no. 2, p. 026113, 2004, URL: https://arxiv.org/pdf/cond-mat/0
308217.pdf.

[11] M. E. Newman, “Fast algorithm for detecting community structure in networks,” Physical review
E, vol. 69, no. 6, p. 066133, 2004, URL: https://arxiv.org/pdf/cond-mat/0309508.pdf.

[12] F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, “Defining and identifying com-
munities in networks,” Proceedings of the national academy of sciences, vol. 101, no. 9, pp. 2658–
2663, 2004, URL: https://www.pnas.org/doi/10.1073/pnas.0400054101.

[13] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities
in large networks,” Journal of statistical mechanics: theory and experiment, vol. 2008, no. 10,
p. P10008, 2008, URL: https://arxiv.org/pdf/0803.0476.pdf.

[14] F. Wang, T. Li, X. Wang, S. Zhu, and C. Ding, “Community discovery using nonnegative matrix
factorization,” Data Mining and Knowledge Discovery, vol. 22, no. 3, pp. 493–521, 2011, URL:
https://www.researchgate.net/publication/220451825_Community_discovery_using_non
negative_matrix_factorization.

[15] Z.-Y. Zhang, Y. Wang, and Y.-Y. Ahn, “Overlapping community detection in complex networks
using symmetric binary matrix factorization,” Physical Review E, vol. 87, no. 6, p. 062803, 2013,
URL: https://arxiv.org/pdf/1303.5855.pdf.

[16] N. P. Nguyen and M. T. Thai, “Finding overlapped communities in online social networks with
nonnegative matrix factorization,” in MILCOM 2012-2012 IEEE Military Communications Con-
ference, pp. 1–6, IEEE, 2012, URL: https://ieeexplore.ieee.org/abstract/document/6415
744.

9
[17] T. G. Kolda and B. W. Bader, “Tensor decompositions and applications,” SIAM review, vol. 51,
no. 3, pp. 455–500, 2009.

[18] V. L. Dao, C. Bothorel, and P. Lenca, “Community structure: A comparative evaluation of

community detection methods,” Network Science, vol. 8, no. 1, pp. 1–41, 2020, URL: https:
//arxiv.org/pdf/1812.06598.
[19] K. McNulty, Handbook of Graphs and Networks in People Analytics: With Examples in R and
Python. CRC Press, 2022, URL: https://ona-book.org/.

10