C OMMUNICATIONS T HEORIES & S YSTEMS

Information-Defined Networks: A Communication Network

Approach for Network Studies

Wenjie Jia1,2 , Tao Jiang1,2,*

1
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China
2
Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology, Wuhan 430074, China
*
The corresponding author, email: [email protected]

Abstract: The research of complex networks facili- speaking, many complex systems can be modeled as
tates the progress of various disciplines, including bi- complex networks for analyzing their properties in
ology, chemistry, social science, computer, and com- practice [1, 2]. As a special case, the associated re-
munication engineering. Recently, it is popular to uti- searches on communication networks, such as ad-hoc
lize complex networks to study the communication networks [3], wireless sensor networks [4] and internet
networks, such as designing efficient routing strate- of everything [5, 6], can be facilitated by the theories
gies and robust communication networks. However, and techniques from complex networks. Specifically,
exploiting the advantages of communication networks it is found that the peculiar structural characteristics of
to investigate networks in various disciplines beyond many complex networks support efficient communi-
telecommunications is still in infancy. Because of this cation without global knowledge [7]. This finding in-
situation, this paper proposes an information-defined spired a series of routing strategies that do not rely on
network (IDN) framework by which a complex net- the global connectivity information of the communica-
work can be abstracted as a communication network tion network [7–12]. Furthermore, it is shown in [13]
associated with multiple intelligent agents. Specif- that the interdependent feature of complex networks
ically, each component and dynamic process in this would lead to cascading failures, which significantly
framework can be defined by information. We show undermined the robustness of the network. This dis-
that the IDN framework promotes the research of un- covery opened an avenue for investigating the robust-
solved problems in the current complex network field, ness and fragility of communication networks from
especially for detecting new interaction types in real- a complex network perspective [14–18]. Addition-
world networks. ally, complex networks also play an essential role in
Keywords: network data analysis; label detection; studying power networks, biological networks, com-
complex network; communication network; network puter networks, transportation networks as well as so-
evolution cial networks.

I. INTRODUCTION However, in addition to the non-trivial topologi-

cal features, such as small-world and scale-free, re-
A complex network is a network with non-trivial source exchanging through networks is a universal
topological features, such as self-organizing, self- phenomenon in both natural and man-made complex
similarity, small-world, and scale-free [1]. Generally systems. For example, in a brain network, various
neurons communicate with each other through some
Received: Apr. 30, 2021 biological signals. An online social network, as an-
Revised: May 07, 2021 other example, is an information exchange platform
Editor: Shiwen Mao
that relies on the Internet. Regarding the resources

China Communications · July 2021 197

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
transmitted in real-world networks as information, we finding gives a “Yes” answer to the main question.
can view these networks as communication networks. The remainder of this paper is organized as follows.
This observation motivates us to ask a question: can Section II describes the foundation of the framework
some unsolved problems in current complex networks — information-defined network components. Section
be solved from a communication network perspective? III introduces how information flows drive network to
To answer this question, we introduce as follows evolve in the IDN framework. Section IV presents the
an unsolved knowledge discovery problem in complex relation between IDN framework and the DE model.
networks, which has attracted lots of attention from Section V concludes this paper.
network scientists and engineers specialized in the in-
formation and communications technology (ICT).
Network GExt
Big data networks original from the Internet, such
as online social network, citation networks, webpage v s
networks, and peer-to-peer networks, emerge as the o

transition from an industrial society to an informa- w

uu
tion society driven by the continuous developments in Agent in G

ICT. These years, it is very popular to design com- x

Agent in GHid
y z Edge in G
munication networks by fully taking into account the Edge in GHid
users’ experience. This trend motivates researchers in
(a) Information-defined network
ICT to study the users’ features and interactions be-
tween users from these big data networks. As a conse-
Tag set on (u,w):
quence, it has attracted a lot of attention from the ICT { Football (0, 1, 0) }

to data mining and data discovery in these big data Agent o

Tag set on (u,v):
networks. Commonly, these networks involve multi- { COVID-19 (1, 0, 0), Physics (0, 0, 1) }
Agent v
ple interaction types [2]. In practice, these types sug- signal processor

gest essential information on the interactions between database

components, but not all of them have been discov- Agent w
ered yet. Thus, detecting the previously-undiscovered update store
optimizer
interaction-types (PUITs) is essential for deepening tune
signal1

our understanding of the data. Additionally, its solu- (2, 0, 0)

Signal generator of Agent u Physics
tions would benefit researchers in mining new features Sub-generator for COVID-19
signal2

in a wide range of datasets, for instance, to discover information on

topic “COVID-19”
Agent u (0, 0.5, 0)
(1, 0, 0)
new interactions between proteins in protein-protein Sub-generator for Football
signal3

interaction networks. Although previous studies have information on

topic “Football” (0, 1, 0)
(1.5, 0, 1)

discussed the edge label detection problem [19, 20], Sub-generator for Physics Agent x
information on Agent z
detecting PUITs is a new one with new challenges. topic “Physics” (0, 0, 1)
Agent y
The first and unique technique for solving this prob-
lem was developed based on a novel temporal net- (b) Information-defined intelligent agent

work model, namely the degradation-evolution (DE) Figure 1. Information-defined network framework. (a)
network model [21]. Therefore, we arrive at a crucial Information-defined network consists of the observed net-
but unsolved problem in network science, i.e., detect- work G and its hidden propagation network GHid . (b) The
ing PUITs in networks which the DE model can not intelligent agent in the IDN framework consists of six com-
mimic. ponents: the signal generator, signal processor, database,
optimizer, signal input channel (outgoing links represented
Focusing on the above problem, we find that the DE
by thick red arrows) and signal output channel (incoming
model is a special case of the network formation model links represented by thick blue arrows).
deduced by the IDN framework. Finally, we conclude
that the IDN framework has the potential to tackle the
crucial but unsolved problem, i.e., detecting PUITs in
networks beyond the scope of the DE model. This

198 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
II. INFORMATION-DEFINED NETWORK (0, 3, 0), whereas a piece of news about COVID-19 is
COMPONENTS written as (4, 0, 0). Hence, we can use I2 = (0, 1, 0)
and I1 = (1, 0, 0) to denote football and COVID-19,
Our modeling approach builds on the intuition that in which are primitive topics in this scenario. Naturally,
complex systems, interactions (described by directed signal S contains I1 -type (I2 -type) information if and
edges in network science) can often be interpreted as only if S[1] > 0 (S[2] > 0). More generally, for a
suitable supply-demand relations. For example, in on- given topic C, we say that signal S carries C-type in-
line social networks like Twitter.com, a directed edge formation, if S[h] > 0 for all h ∈ {h | C[h] = 1}.
where user vi follows user vj indicates that vi de- The C-type information carried by S can be written
mands “information” (e.g., in the form of texts, pho- as C ◦ S = (C[1]S[1], C[2]S[2], · · · , C[M ]S[M ]),
tos, and videos) from vj . We take citation networks where “ ◦ ” refers to the Hadamard product. It fol-
as another example. Paper vi having an out-edge con- lows that if signal S carries C-type information, then
nected to paper vj implies that vi requires some infor- C ≤ sign(S). Especially, the total information carried
mation provided in vj to support some of its claims. by S can be represented by S ◦ I, where I = M
P
h=1 Ih .
Inspired by these observations, we assume an interac- In IDNs, nodes are viewed as intelligent agents (Fig-
tion represents the supply-demand exchange of some ure 1), who are the sources (or suppliers) of informa-
“resource”. Motivated by the basic idea of recom- tion flows. We assume that agents generate signals at
mender systems [22] and natural language process- random via a signal generator they are endowed with.
ing [23] that vectors can represent products or words, Specifically, agent vi ’s signal generator, denoted by qi ,
we assume that resources exchanged by the agents can be represented by an M -length sequence of prob-
can be also represented by vectors consisting of non- ability distributions (Oi1 , Oi2 , . . . , OiM ), where each
negative real numbers. We refer to this representing Oij is with support in [0, ∞) and bounded expecta-
vector as a piece of information or a (digital) signal. tion, for 1 ≤ j ≤ N . Therefore, signals created by
Specifically, we define a signal S as a finite sequence vi are sampled from qi . For a topic C, vi ’s C-type in-
of M non-negative real numbers, (S[1], . . . , S[M ]), formation sub-generator can be expressed as qi |C (qi
where (without loss of generality) M is a large enough restricted to C).
positive integer and S[h] represents the h-th entry of S Both the signal generator and the sub-generators of
for h = 1, 2, · · · , M . Hence, all possible signals span a given agent vj are unknown to other agents. Agent
a space S = [0, ∞)M . vi seeks to learn which topics another agent vj can
For a real number s ≥ 0, define produce based on all the signals generated by vj that
 she previously received. Specifically, at time t, agent
1, if s > 0, vi deems that vj can produce C-type information, if
sign(s) =
0, otherwise. vi has received C-type information carried by signals
generated by vj (i.e., there exists S ∈ Si,j (t), such that
For signal S = (S[1], S[2], · · · , S[M ]), define C ≤ sign(S), where Si,j (t) denotes the collection (see
sign(S) = (sign(S[1]), sign(S[2]), · · · , sign(S[M ])). the definition of collections in Appendix A) of all the
Two signals S1 and S2 belong to the same signal cate- |Si,j (t)| signals generated by vj and received by vi up
gory if and only if sign(S1 ) = sign(S2 ). S1 is weaker to time t). See a more general case in Appendix B.
than S2 , denoted by S1 ≤ S2 , if S1 [h] ≤ S2 [h] for any As intelligent agents, nodes have signal processors,
h = 1, 2, · · · , M . and react appropriately to the received signals accord-
A signal can be interpreted as a piece of information ing to the processed results (Figure 1). For exam-
(message) that may involve several topics. We employ ple, in online social networks, incoming signals drive
nonzero vectors in C = {0, 1}M to represent such top- nodes toward various reactions, e.g., replying, for-
ics. For h = 1, 2, · · · , M , we define the h-th primi- warding or reporting. For signal S ∈ S and topic
tive topic as the topic corresponding to the unit vector C ∈ C, we denote vi ’s response to the C-type infor-
Ih , whose h-th entry is one, whereas all the other en- mation carried by S as Bi (S | C), where Bi (S | C) is
tries are zeros. For example, suppose that two pieces a non-negative real number that quantifies how use-
of news about football are denoted by (0, 2, 0) and ful the information is perceived by vi . For brevity,

China Communications · July 2021 199

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
we define Bi (S) = Bi (S | I). For a label set C 0 = time t. Specifically,
{C1 , C2 , · · · , Cs }, we further define Bi (S | C 0 ) to be
agent vi ’s composite response to multiple pieces of in- D(Xi,j (t))
formation S ◦ C1 , S ◦ C2 , · · · , S ◦ Cs . ( P
S∈Si,j (t) Bi (S | Xi,j (t))
We require the response model to obey three prin- = |Si,j (t)| , if |Si,j (t)| > 0,
ciples: (1) zero signal, denoted by 0, has zero 0, if |Si,j (t)| = 0.
response to agents; (2) the composite response to
multiple pieces of information in a signal respects ~ j (t) = PM i Aj (h, t)Ih and
Let i A h=1
the superposition principle as many response mod-
els studied in signal processing [24] do — for exam- ( P
Bi (S | Ih )
S∈Si,j (t)
ple, Bi (S | {C1 , C2 }) = Bi (S | C1 ) + Bi (S | C2 ) − , if |Si,j (t)| > 0,
i Aj (h, t) = |Si,j (t)|
Bi (S | C1 ◦ C2 ) (see more details in Appendix C); 0, if |Si,j (t)| = 0.
and (3) a message should provide higher response
than messages weaker than it (i.e., if S1 ≤ S2 , then Note that i Aj (h, t) is vi ’s average response to Ih -type
Bi (S1 | C) ≤ Bi (S2 | C) for any C ∈ C). Many mod- information generated by vj and received by vi up to
els in signal processing respect these properties [24], time t. Thus, i Aj (h, t) is a score of vj ’s Ih -type infor-
such as a linear response model defined by the equa- mation supply ability scored by vi at time t.
tion Bi (S | C) = Fi · (S ◦ C), where Fi is an M - Let “ ∨ ” refer to the element-wise OR operator. For
dimensional non-negative real vector associated with instance, (1, 1, 0) ∨ (1, 0, 0) = (1 ∨ 1, 1 ∨ 0, 0 ∨ 0) =
agent vi . See another example of response models in (1, 1, 0). Supposing Xi,j (t) = {C1 , C2 , · · · , Cs },
Appendix G. where tag Ch is a nonzero vector in C for h =
Information also defines edge-level properties. At 1, 2, · · · , s, we define ∨Xi,j (t) = C1 ∨ C2 ∨ · · · ∨ Cs .
time t, each edge vi → vj is associated with a set of In addition, we define ∨Xi,j (t) = 0 for Xi,j (t) = ∅.
tags, denoted by Xi,j (t). Intuitively, Xi,j (t) represents Denoting ∨Xi,j (t) as Xi,j (t), we can write
the set of topics about which vi demands information
from vj at time t. For example, agent vi , a Twitter user, ~ j (t),
D(Xi,j (t)) = Xi,j (t) · i A (1)
begins to seek information about COVID-19 (repre-
sented by tag I1 ) from the World Health Organization
(represented by agent vj ) at time t. Then vi forms a (see Appendix C). Eq. (1) shows that Xi,j (t) deter-
link vi → vj with a tag set composed of a single ele- mines the utility of Xi,j (t). We call Xi,j (t) vi ’s vector-
ment Xi,j (t) = {I1 }. Naturally, Xi,j (t) = ∅ if a link valued linking strategy to vj at time t. Further, agent
from vi to vj does not exist at time t. In our frame- vi ’s linking strategy at time t is defined to be a matrix,
work, self-edges are forbidden (i.e., Xi,i (t) = ∅ for
any i and t). Xi (t) = (Xi,1 (t), Xi,2 (t), · · · , Xi,N (t)). (2)
Each intelligent agent is in full charge of her/his out-
edges. Specifically, for two different intelligent agents According to the IDN framework, networks emerge
vi and vj , the tunable tag set Xi,j (t) is merely managed out of multilayer interactions among intelligent
by agent vi and called vi ’s linking strategy to vj at time agents, and they are endowed with hidden “multi-
t. On the one hand, agent vi ’s different linking strate- layer” structures. For a tag (nonzero vector) C ∈ C,
gies to vj present vi ’s different information demands we define the C-type layer of G(t) to be the sub-
on supplier vj . On the other hand, agent vj ’s infor- network consisting of all the edges in {vi → vj ∈
mation supply is independent with vi ’s linking strate- G(t) | C ≤ Xi,j (t)} and all the agents involved in
gies to her/him and meets vi ’s different information these edges. Thus, there exist at most 2M − 1 non-
demands on her/him in varying satisfaction degrees. empty layers G(t), since there are 2M − 1 tags in C.
For a given Xi,j (t), its satisfaction degree for agent vi This multilayer structure is a natural consequence of
is quantified by a utility function D(Xi,j (t)). Here, underlying interactions that involve multiple tags.
D(Xi,j (t)) is vi ’s average response to the specified in-
formation in the received signals generated by vj up to

200 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
III. INFORMATION-DRIVEN NETWORK faster than the timescale of the network formation pro-
EVOLUTION cess. That is, the propagation process, denoted by
Υ(t) = (vit , St , GSt ), ends before time t increases by
In our framework, the direction of information flows one. After Υ(t) ends, receivers involved in Υ(t) up-
is opposite to the direction of the directed edges. For date their database of received signals [i.e., Sj,it (t − 1)
a network G, there always exist many hidden roads is replaced by Sj,it (t) = Sj,it (t − 1) + {St } (see the
for signals, which extend the original network. In- definition of the operator “ +” in Appendix A), for any
tuitively, this corresponds to our experience that ev- receiver vj in GSt ]. Then each agent vi chooses the
ery day, we are exposed to information from many optimal linking strategy (see its definition in Eq. (2))
different sources; however, available data might only to maximize their perceived usefulness of the infor-
include information on commitments to a restricted mation they receive. Specifically, the optimal link-
number of sources. Besides, it also naturally accom- ing strategy is a solution to a constrained optimization
modates for the existence of relationships that are not problem where the objective function represents the
included in the data due to missing links [25]. We de- total perceived utility of the received information (see
note as GHid the hidden propagation network of G. Appendix C for its derivation):
We denote as GExt the extended propagation net-
work of G which comprises edges in G or in GHid
X X
U (Xi (t)) = D(Xi,j (t)) = ~ j (t).
Xi,j (t) · i A
and all the nodes involved in these edges. Overall, j6=i j6=i
GExt = G ∪ GHid (see Figure 1). We assume (3)
that the hidden propagation network GHid is time-
invariant (see Appendix G for another example). The This optimization problem has two kinds of con-
signal spreading process in G studied here is actually straints: dynamic supply-demand constraints,
a spreading process over GExt . The spreading graph
GS of signal S is constructed as follows: (1) we ini- ~ j (t)),
Xi,j (t) ≤ sign( i A (4)
tialize GS to be an empty graph; (2) the generator of
S is the first node of GS ; (3) if vj receives S from vi
for any i, j, t (see its derivation in Appendix D), and
(implying vj → vi is an edge in the extended propa-
bounded out-degree constraints,
gation network), we add vj → vi to GS . Obviously,
GS is a subnetwork of GExt . Then this propagation N
process, denoted by Υ, can be represented by a triple
X
Xi,j (t) ≤ Ri , (5)
(vi , S, GS ). Υ’s seed and depth are defined to by vi j=1
and the length of the longest simple paths (a simple
path is a path without repeated nodes) between vi and for any i, t. In Eq. (5), Ri is an M -dimensional non-
another node in GS , respectively. negative real vector associated with agent vi describ-
The network formation in our framework is a net- ing vi total information demand (see details in Ap-
work evolution process driven by information flows. pendix E). We call Ri agent vi ’s maximum demand.
At the beginning (t = 0), fixing an existing strongly After each agent solves its optimization problem (see
connected directed network GHid as the base propaga- Appendix F), time increases by one.
tion network, we construct G(0) by adding N nodes An example of this information-driven network for-
from GHid (GHid usually has more than N nodes). mation dynamics is shown in Figure 2. At time
No edge is added to G(0) at first. Then, re-index t = 0, a strongly connected network GHid is given
nodes in GHid , such that {vi }N i=1 represent all the as the propagation network for the network evolu-
nodes of G(0). tion process. The initial network, G(0), comprises
At each time t > 0, a randomly chosen seed agent three nodes, v1 , v2 , v3 , picked from GExt . Here, we
from {vi }N i=1 generates a signal which subsequently define Bi (S) = I · S = S[1] + S[2] + S[3] and
propagates over the extended propagation network of Ri = I = (1, 1, 1) for i = 1, 2, 3 and a signal
G(t) (i.e., GExt (t) = G(t) ∪ GHid ). We assume that S ∈ [0, +∞)3 . At time t = 1, a random node, v2 ,
the timescale of the spreading process is substantially generates a random signal s1 = (4, 2), and sends it to

China Communications · July 2021 201

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
 t time t increases by one. At t = 2, v3 is activated,
and she supplies v1 with a randomly generated sig-
nal s2 = (5, 1). By solving the updated optimization
4 problem, v1 tunes her/his linking strategy to v2 from
tag1
2 2 X1,2 (1) = tag3 to X1,2 (2) = tag2, and then builds an
G(3) out-edge with a tag set {tag1} to v3 . At time t = 3,
1 S#
1
3
S#
3
(1, 5)
(4, 2) (3, 3) v3 is activated again. This time, v3 directly sends v1
tag2
3 a random signal s3 . After updating the database, v1 ’s
tag2
2 2 optimizer suggests her/him to adopt the reverse opti-
G(2)
mal linking strategy. Finally, the network evolves into
1
3
1
3
S"
(4, 2) (5, 1) G(3).
(5, 1)
tag1 S"
2

tag3
2
S! 2 IV. CONNECTING TO THE DE MODEL
G(1) S!
1
3
1
3
(4, 2)
(0, 0) (4, 2) In the DE model, t refers to the time. Notations t < 0,
1 t = 0 and t > 0 are used to denote the past, the present
and the future, respectively. At the present time ( i.e.
2
GExt t = 0), the DE model initialize G = G(0) to be an
G(0)
1
(0, 0) (0, 0)
arbitrary network with n nodes and m edges with la-
3
0 Initilization
ur ur
bels in C, where C = {C1 , C2 , · · · , Ch } represents all
Evolution process Spreading process 1 A 2 (t ) 1 A3 (t )
the labels that can be observed in the whole course of
G’s changing process. Let G(t) = (V(t), E(t)) be a
agents in the hidden network tag1 tag2 tag3
hidden edges (bi-directed) (1, 0) (0, 1) (1, 1)
temporal network with V(t) = {v1 , v2 , · · · , vn }, for
time t = −∞, · · · , −1, 0, 1, · · · , +∞. For an edge
Figure 2. An illustration of the information-driven network vi → vj in G(t), notation C(i, j, t) is employed to de-
evolution process. note the label set associated with it. It is assumed that
every edge in networks should be assigned at least one
label. Therefore, there exists an edge from vi to vj at
v1 via the hidden propagation network. After receiv-
time t if and only if C(i, j, t) 6= ∅.
ing s1 , v1 updates its estimation on v2 ’s information
~ 2 (0) to 1 A
supply abilities from 1 A ~ 2 (1). Subsequently, Regarding edges with the same label in G as the
her/his optimizer outputs an optimal linking strategy, components of a layer of network G, we can divide G
X1 (1) = (X1,1 (1), X1,2 (1), X1,3 (1)) = (0, tag3, 0) into h different layers. Specifically, the l-th layer of
which is a solution to the following optimization prob- G(t), denoted by Gl (t) = (Vl (t), El (t)), is defined to
lem: be the subnetwork consisting of all the edges in E(t)
with label Cl and all the nodes involved in these edges
(see Figure 1a in [21]). For node vi , define its potential
Maximize :
~ 2 (t) + Y [3] · 1 A
~ 3 (t), energy with respect to the l-th layer at time t to be
U (Y ) = Y [2] · 1 A
Subject to : n
(bounded out-degree constraints)
X
Pl (i, t) = (Amax,l − Aj,l )χEl (t) (vi → vj ), (6)
Y [2] + Y [3] ≤ R1 = I; j=1
(dynamic supply-demand constraints)
~ 2 (t)), Y [3] ≤ sign( 1 A
Y [2] ≤ sign( 1 A ~ 3 (t)),
where Amax,l = maxvi ∈V Ai,l and χEl (t) (vi → vj ) =
1 if vi → vj ∈ El (t); otherwise χEl (t) (vi → vj ) = 0.
where the optimization variable Y = (0, Y [2], Y [3]) The value of Pl (i, t) describes how eager node vi is to
and Y [h] ∈ {0, 1}3 for h = 2, 3. At the end of rewire its out-edges in Gl (t) to connect to nodes with
t = 1, v1 adopts the optimal linking strategy by build- higher l-attractiveness at time t (see Figure 1a). Fur-
ing an out-edge to v2 with a tag set {tag3}. After that, ther, define vi ’s potential energy and the system’s po-

202 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
tential energy at time t to be P (i, t) = hl=1 Pl (i, t)
P
covery, interpreting the observed data in terms of an
P
and P (G, t) = i=1 P (i, t), respectively. A node’s intelligent agents’ communication network may lead
higher potential energy means the stronger desire for to new sets of results in data mining and knowledge
this node to rewire its out-edges, and the higher poten- discovery in network data. Future research in net-
tial energy of a system indicates a more structurally work science may explore additional applications of
unstable state of this system. the proposed communication network approach be-
There exists an evolution mechanism in the DE yond those developed here, including network struc-
model: at each time t > 0, a node rewires one of its ture analysis [26], layer detection [2], link predic-
out-edges in a layer of G(t) and then time t increases tion [19, 25], ranking algorithms [27], and recom-
by 1, such that P (G, t + 1) ≤ P (G, t) (see Figure 1c mender systems [22, 28].
in [21]). In addition, there exists a degradation mech- In conclusion, we hope our proposed IDN-based
anism: at each time t < 0, a node rewires one of its communication network approach would facilitate
out-edges in a layer of G(t) and then time t decreases network data analysis and sciences related to networks
by 1, such that P (G, t − 1) ≥ P (G, t) (see Figure 1b as the well-studied complex network approach does.
in [21]).
For an arbitrary instance, {G(t)}N 1
t=−N1 , of the DE ACKNOWLEDGEMENT
model, we find there exists an instance, {G0 (t)}N 2
t=0 ,
generated by the proposed network formation frame- We would like to thank Lixia Xiao, Daiming Qu,
work (NFF) under some special configurations, such Linyuan Lü and the anonymous referees for their sug-
that the instance of the DE model {G(t)}N 1
t=−N1 , rep-
gestions which greatly improved this article. This
resented by a temporal network sequence, is a sub- work was supported in part by Young Elite Scientists
sequence sampled from {G0 (t)}N 2
t=0 , which is also a
Sponsorship Program by CAST under Grant number
temporal network sequence generated by the NFF. In 2018QNRC001, and National Science Foundation of
other words, the former instance is a “fast-forward” China with Grant number 91738202, 62071194.
version of the latter one. To sum up, we obtain the
following theorem (see its proof in Appendix G). APPENDICES
Theorem 1. The DE model defined above is a partic- Appendix A: Basic Notations
ular case of the proposed NFF.
For a vector X its h-th element is denoted by X[h] or
[X][h]. We always use G to denote a network. No-
V. CONCLUSION
tations V(G) and E(G) are employed to represent the
In this paper, we propose an IDN framework that en- agent set and the edge set of G, respectively. Let e
ables us to embed a complex network into a commu- be a directed edge in E(G). We use notations deesrc
nication network of multiple intelligent agents. We and deetar to denote the source agent and the target
prove that the NFF extends the DE model showing agent of e, respectively. Multisets are containers that
that the IDN framework can tackle a crucial but un- store elements following a specific order, and where
solved problem, i.e., the PUIT detection problem in multiple elements can have equivalent values. Let A
networks outside the scope of the DE model. As a be a multiset, we define χA (x) = 1 if x ∈ A, oth-
practical application, this theoretical finding illumi- erwise χA (x) = 0. Let A = {S1 , S2 , · · · , Sn1 } and
nates that our framework enables engineers to discover B = {S10 , S20 , · · · , Sn0 2 } be two multisets, we define
the hidden features of users of communication net- A + B to be the multiset consisting of n1 + n2 ele-
works and help them to design communication sys- ments, where the h-th element Sh00 is
tems providing users better experience. 
Moreover, the proposed approach suggests exciting Sh , if 1 ≤ h ≤ n1 ,
Sh00 = 0
research avenues for researchers in information and Sh−n 1
, if n1 < h ≤ n1 + n2 .
communications engineering, network scientists, and
data analysts. For future research in knowledge dis- The following table summarizes all the frequently-

China Communications · July 2021 203

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
used notations throughout this paper. Table 1. Frequently-used notations in this paper.

Appendix B: Information Supply of Intelligent Notation Interpretation

Agents (Basic operators)
Let C1 and C2 be two information topics. We as- X ◦ X0 Hadamard product
0
sume that (1) agent vi deems that agent vj can pro- X ∨X Element-wise OR operator
duce Ih -type information at time t if and only if there sign(X) Sign function of vector X
exists S ∈ Si,j (t), such that Ih ≤ sign(S), for X[h] or [X][h] The h-th element of vector X
h = 1, 2, · · · , M ; (2) if agent vi deems that agent A+B Addition of multisets
vj can produce both C1 -type and C2 -type information, (Basic network notations)
then agent vi believes vj can produce information of G Network/system
sign(C1 + C2 )-type (i.e., (C1 ∨ C2 )-type. See the defi- vi Node with index i
nition of operator “ ∨ ” at the end of Section II). Based vi → vj Directed edge from vi to vj
on the first assumption, we have at time t, all the prim- V(G) The set of nodes in G
itive topics deemed by vi that vj can produce compose E(G) The set of edges in G
the following set deesrc The source node of edge e
deetar The target node of edge e
t System time
Ii,j (t) = {Ih | ∃S ∈ Si,j (t), s.t. Ih ≤ sign(S)}.
(Notations used in the IDN framework)
(B.1)
S Signal space [0, +∞)M
Let |Ii,j (t)| = s and Ii,j (t) = {J1 , J2 , · · · , Js }.
C Label/tag space {0, 1}M
By Eq. (B.1), there exists Sg ∈ Si,j (t) such that
Ih Unit vector in C with Ih [h] = 1
Jg ≤ sign(Sg ), for g = 1, 2, · · · , s. On the one hand,
qi Agent vi ’s signal generator
we have sign[∨Ii,j (t)] = sign(J1 ∨ J2 ∨ · · · ∨ Js ) ≤
Ri Agent vi ’s maximum demand
sign(S1 ∨S2 ∨· · ·∨Ss ) ≤ sign[∨Si,j (t)]. On the other
Bi (•) Agent vi ’s signal processor
hand, for w ∈ {1, 2, · · · , M } if Iw ·sign[∨Si,j (t)] = 1
Si,j (t) Agent vi ’s database of signals
(where operator “·” refers to the inner product of vec-
generated by vj up to time t
tors), then there exists Sw0 ∈ Si,j (t) such that Iw ≤
Xi,j (t) Agent vi ’s linking strategy to
sign(Sw0 ). By Eq. (B.1), one has Iw ∈ Ii,j (t). There-
vj at time t
fore, sign[∨Si,j (t)] = M
P
h=1 {Ih · sign[∨Si,j (t)]}Ih ≤
Ps Xi,j (t) Agent vi ’s vector-valued linking
g=1 g J = sign[∨I i,j (t)]. Then we have
strategy to vj at time t
D(•) Utility function
sign[∨Si,j (t)] = sign[∨Ii,j (t)]. (B.2)
U (•) The objective function of the
main optimization problem
It follows the above two assumptions that vi deems vj A (h, t) vi ’s average response to Ih -type
i j
can produce C-type information, if and only if there information generated by vj
exists ag ∈ {0, 1} for g = 1, 2, · · · , s, such that A~ (t)
PM
i j h=1 i Aj (h, t)Ih

s
(Notations used in the DE model)
C=
X
ag Jg . (B.3) C(i, j, t) The label set associated with the
g=1
ordered node-pair (vi , vj ) at t
Pl (i, t) Node vi ’s potential energy with
respect to the l-th layer at t
Let S~i,j (t) = ∨Si,j (t). On the one hand, if vi
P (i, t) Node vi ’s potential energy at
deems vj can produce C-type information, by Eqs.
time t
(B.2) and (B.3), we have C ≤ sign( sg=1 Jg ) =
P
P (G, t) The system’s potential energy
sign[∨Ii,j (t)] = sign [∨Si,j (t)] = sign(S~i,j (t)). On at time t
the other hand, it follows Eq. (B.2) that if C ≤
sign(S~i,j (t)), then C ≤ sign[∨Ii,j (t)]. Therefore, C

204 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
can be written in the form of Eq. (B.3). Thus, vi deems |Si,j (t)| > 0, we have
C-type information can be generated by vj at time t.
Above all, we conclude that vi deems vj can produce 1 X
D(Xi,j (t)) = Bi (S | ∨ Xi,j (t))
C-type information, if and only if C ≤ sign(S~i,j (t)). |Si,j (t)|
S∈Si,j (t)
P PM
S∈Si,j (t) h=1 Bi (S | Ih )Ih
Appendix C: Objective Function of the Opti- = Xi,j (t) ·
|Si,j (t)|
mization Problem "M P #
X S∈Si,j (t) Bi (S | Ih )
We assume the composite response to multiple pieces = Xi,j (t) · · Ih
|Si,j (t)|
of information in a signal respects the superposition h=1

principle. Let S ∈ S, i ∈ {1, 2, · · · , N } and Ch be ~ j (t).

= Xi,j (t) · i A
a nonzero tag in C, for h = 1, 2, · · · , s. According to
the superposition principle, we define When |Si,j (t)| = 0, we have i A ~ j (t) = 0 and
D(Xi,j (t)) = 0. Then D(Xi,j (t)) = Xi,j (t) · i A~ j (t).
Bi (S | {C1 , C2 , · · · , Cs }) = Combining the two cases, we obtain D(Xi,j (t)) =
X X ~ j (t). Then the objective function of the
Xi,j (t) · i A
Bi (S | Ch ) − Bi (S | Ch ◦ Ck )
1≤h≤s 1≤h<k≤s
optimization problem can be rewritten as
X
+ Bi (S | Ch ◦ Ck ◦ Cm ) − · · · X X
U (Xi (t)) = D(Xi,j (t)) = ~ j (t).
Xi,j (t) · i A
1≤h<k<m≤s
j6=i j6=i
+ (−1)s−1 Bi (S | C1 ◦ C2 ◦ · · · ◦ Cs ).
(C.1)
Appendix D: Dynamic Supply-demand Con-
Let C ∈ C. According to the principles for response
straints
models introduced in Section II and Eq. (C.1), we Let vi ’s linking strategy to vj at time t be Xi,j (t) =
have Bi (S | I) = M
P
h=1 Bi (S | Ih ). Denote Bi (S | Ih ) ∨Xi,j (t) and Xi,j (t) = {C1 , C2 , · · · , Cs }, where in-
as Bi (S, h) for h = 1, 2, · · · , M , and B ~ i (S) = teger s is no less than one, and Ch ∈ C for h =
PM
h=1 Bi (S, h)Ih . Then we obtain Bi (S | C) = 1, 2, · · · , s. On the one hand, Ch ∈ Xi,j (t) implies
PM ~
h=1 C[h]Bi (S, h) = C · Bi (S). Therefore, that vi seeks Ch -type information generated by vj at
time t, for h = 1, 2, · · · , s. In other words, vi be-
Bi (S | {C1 , C2 , · · · , Cs }) = lieves vj can generate Ch -type information, for h =
X
~ i (S) −
X
~ i (S) 1, 2, · · · , s. According to Eq. (4), Ch ≤ sign(S~i,j (t)),
Ch · B (Ch ◦ Ck ) · B
for h = 1, 2, · · · , s. Then we have
1≤h≤s 1≤h<k≤s
X
+ ~ i (S) − · · ·
(Ch ◦ Ck ◦ Cm ) · B
Xi,j (t) = C1 ∨ C2 ∨ · · · ∨ Cs ≤ sign(S~i,j (t)).
1≤h<k<m≤s
(D.1)
~ i (S)
+ (−1)s−1 (C1 ◦ C2 ◦ · · · ◦ Cs ) · B
~ i (S)
= (C1 ∨ C2 ∨ · · · ∨ Cs ) · B On the other hand, according to Eq. (3), we require
= Bi (S | C1 ∨ C2 ∨ · · · ∨ Cs ).
(C.2) ~ j (t)),
Xi,j (t) = ∨Xi,j (t) ≤ sign( i A (D.2)

Let agent vi ’s linking strategies at time t com- for i = 1, 2, · · · , N and j = 1, 2, · · · , N , to avoid

pose a vector X~i (t) = (Xi,1 (t), Xi,2 (t), · · · , Xi,N (t)). useless demands. Assume sign( i Aj (h, t)) = 1.
P
Let Xi,j (t) = ∨Xi,j (t), for j = 1, 2, · · · , N , and Then we have i Aj (h, t) = S∈Si,j (t) Bi (S | Ih ) >
Xi (t) = (Xi,1 (t), Xi,2 (t), · · · , Xi,N (t)). It follows 0. Thus, there exists S 0 ∈ Si,j (t) such that
Eq. (C.2) that Bi (S | Xi,j (t)) = Bi (S | ∨ Xi,j (t)) = Bi (S 0 | Ih ) = Bi (S 0 ◦ Ih ) > 0. Accord-
~ i (S) = Xi,j (t) · B
[∨Xi,j (t)] · B ~ i (S). Therefore, when ing to the first principle for response models, we

China Communications · July 2021 205

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
have S 0 ◦ Ih 6= 0. Therefore, sign(S~i,j (t)) ≥ Note that constraint (E.3) implies constraint (E.1)
sign(S 0 ) ≥ Ih = sign( i Aj (h, t))Ih . Consequently, when I · Ri ≤ mi .
we have sign( i A ~ j (t)) = PM sign( i Aj (h, t))Ih ≤
h=1
sign(S~i,j (t)). This inequation shows that (D.2) im- Appendix F: Sovling the Optimization Prob-
plies (D.1). Finally, we call constraint (D.2) the dy- lem
namic supply-demand constraint.
Let i Qj (t) = sign( i A ~ j (t)). Matrix Xi (t) =
Appendix E: Bounded Out-degree Constraints (Xi,1 (t), Xi,2 (t), · · · , Xi,N (t)) is an optimal linking
strategy of agent vi at time t if it is a solution to the
For agent vi , let ki,out (t) denote its out-degree at time following optimization problem (denoted by P1[i, t]):
t. We assume the number of out-edges of every agent
in the framework is bounded. Specifically, P ~ j (t),
Maximize : U (Y ) = j6=i Y [j] · i A
N Subject to : Y [j] ≤ i Qj (t), for any j;
P
Y [j] ≤ Ri ;
X
ki,out (t) = sign(I · Xi,j (t)) ≤ mi , (E.1) Pj6=i
j=1 j6=i sign(I · Y [j]) ≤ mi ,

where the non-negative integer mi is a feature of agent where Y = (Y [1], Y [2], · · · , Y [N ]), Y [i] = 0 and
vi . For a vector X, its h-th entry is denoted by X[h] Y [j] ∈ C for j 6= i. We always set I · Ri = mi for i =
or [X][h]. It follows constraint (E.1) that 1, 2, · · · , N . According to Appendix C in the main
P
text, the last constraint j6=i sign(I · Y [j]) ≤ mi can
N N
X X be neglected. Solutions to P1[i, t] always exist. The
[Xi,j (t)][h] = sign(Ih · Xi,j (t)) following proposition provides a method to construct
j=1 j=1
(E.2) a solution to P1[i, t].
N
X
≤ sign(I · Xi,j (t)) ≤ mi . Proposition 1. For h = 1, 2, · · · , M , define
j=1 U (h) (vi ) = {vj ∈ V(G) | i Aj (h, t) > 0} and
 (h)
deg (vi ) = min Ri [h], U (h) (vi ) . Let Vr (vi )
(h)
Eq. (E.2) shows N
P
j=1 [Xi,j (t)][h] is bounded for h = denote a subset of U (h) (vi ) consisting of deg (h) (vi )
1, 2, · · · , M . Let Ri = (Ri [1], Ri [2], · · · , Ri [N ]), agents with the deg (h) (vi ) largest sores of Ih -type
where Ri [h] is the least upper bound of information supply abilities observed by vi at time t
PN
j=1 [Xi,j (t)][h]. Hence, we have for any t in U (h) (vi ). For j = 1, 2, · · · , N , let Xj ∈ C and

N N
X X Xj [h] = 1 ⇐⇒ vj ∈ V (h) (vi ). (F.1)
Xi,j (t) = Ih [Xi,j (t)][h] ≤ Ri . (E.3)
j=1 j=1
Then X = (X1 , X2 , · · · , XN ) is a solution to P1[i, t]
It follows constraint (E.3) that
Proof. (1) We show that X satisfies all the constraints
N N
in P1[i, t]. Then, we have
X X
ki,out (t) = sign(I · Xi,j (t)) = (I · Xi,j (t)) ∨ 0
j=1 j=1 X M
XX M
X X
N
X
" M
X
# N X
X M X[l] = Xl [h]Ih = Ih Xl [h]
= 0∨ [Xi,j (t)][h] ≤ [Xi,j (t)][h] l6=i l6=i h=1 h=1 l6=i
j=1 h=1 j=1 h=1 M M
X X
(h)
Ih V deg (h) (vi )Ih
   
M
X N
X M
X N
X = (vi ) =
=  [Xi,j (t)][h] =  Xi,j (t) [h] h=1 h=1
h=1 j=1 h=1 j=1 M
X
≤ Ri [h] = Ri .
 
N
X
=I · Xi,j (t) ≤ I · Ri . h=1
j=1 (F.2)

206 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
Assume Xj [h] = 1. By Eq. (F.1), we have vj ∈ ists a non-empty multiset S 00 of signals in S such that
V (h) (vi ). Consequently, i A(h, t) > 0, and then αh0 (S 0 + S 00 ) ∈ (β1 , β2 ) and αh (S 0 + S 00 ) = αh (S 0 ).
sign(i Aj (h, t)) = Xj [h]. For any j, we have
Proof. For a signal S ∈ S and an integer h ∈
M M {1, 2, · · · , M }, define ψh,x (S) to be a vector whose
h-th element is x and h0 -th (h0 6= h) ele-
X X
X[j] = Xj [h]Ih ≤ sign(i Aj (h, t))Ih
h=1 h=1
(F.3) ment is S[h0 ]. Further, for a multiset of signals
= i Qj (t). A = {S1 , S2 , · · · , S|A| }, we define ψh,x (A) to be
{ψh,x (S1 ), ψh,x (S2 ), · · · , ψh,x (S|A| )}. Note that for
Eqs. (F.2) and (F.3) imply that X satisfies all the con- any signal S in S and h ∈ {1, 2, · · · , M }, ψh,0 (S)
straints in P1[i, t]. is non-zero since ψh,0 (S) ≥ IM +1 .
(2) We show U (X 0 ) ≤ U (X) for any X 0 (1) Case 1: αh0 (S 0 ) > β2 > β1 > 0. There
which satisfies the constraints in P1[i, t]. Let exists two positive integers p and q with q > p
0 0 0 0 0(h)
X = (X1 , X2 , · · · , XN ) and V (vi ) = {vj ∈ such that αh0 (S 0 ) · p/q ∈ (β1 , β2 ). Let S 00 =
p−1 q−p
U (h) (vi ) | Xj0 [h] = 1}. By the bounded out- z }| { z }| {
P 0 S 0 + S 0 + · · · + S 0 + ψh0 ,0 (S 0 ) + · · · + ψh0 ,0 (S 0 ).
degree constraint, we have j6=i Xj [h] ≤ Ri [h],
for any h. Then we have for any h, |V 0(h) (vi )| ≤ Then we have
min{Ri [h], |U (h) (vi )|}. Thus, for any h, |V 0(h) (vi )| ≤
S[h0 ] + (p − 1) S∈S 0 S[h0 ]
P P
deg (h) (vi ) = |V (h) (vi )|. Recall the definition of 0 00 S∈S 0
αh0 (S + S ) =
P |S 0 | + (q − 1)|S 0 |
X, we have for any h, vj ∈V (h) (vi ) i Aj (h, t) ≥
p S∈S 0 S[h0 ] pαh0 (S 0 )
P P
A
vj ∈V 0(h) (vi ) i j (h, t) Therefore, for any h = = ,
q|S 0 | q
X X
Xj [h] · i Aj (h, t) = Xj [h] · i Aj (h, t)
and for h ∈ {1, 2, · · · , M } with h 6= h0
j6=i vj ∈U (h) (vi )
X X
= i Aj (h, t) ≥ i Aj (h, t)
P P
0 00 S∈S 0 S[h] + (q − 1) S∈S 0 S[h]
vj ∈V (h) (vi ) vj ∈V 0(h) (vi )
αh (S + S ) =
X X |S 0 | + (q − 1)|S 0 |
Xj0 [h] · i Aj (h, t) = Xj0 [h] · i Aj (h, t).
P
= q S∈S 0 S[h]
= = αh (S 0 ).
vj ∈U (h) (vi ) j6=i q|S 0 |

β1 > αh0 (S 0 ). Let S 00 (k) =

P PM
Consequently, U (X) = h=1 Xi [h] · i Aj (h, t) ≥
(2) Case 2:
j6=i k−1
PM P 0 0
z }| {
h=1 j6=i Xj [h] · i Aj (h, t) = U (X ). ψh0 ,n+1 (S 0 ) + · · · + ψh0 ,n+1 (S 0 ). Then we have
To sum up, we have X is a solution the P1[i, t].
0 0
P
0 00 0 S[h ] + (n + 1)(k − 1)|S |

Appendix G: Proof of Theorem 1 αh0 (S + S (k)) = S∈S

|S 0 | + (k − 1)|S 0 |
0
(n + 1)(k − 1)|S 0 |
P
0 S[h ]
Lemma 1. Let n and M be two positive inte- = S∈S 0 +
gers, Il be an (M + 1)-dimensional vector for l = k|S | k|S 0 |
0
1, 2, · · · , M + 1, and S = {IM +1 + M αh0 (S ) − 1
P
h=1 xh Ih | xh ∈ = + (n + 1) → n + 1 (k → +∞),
{0, 1, n + 1}, h ∈ {1, 2, · · · , M }}. Let A = k
{S1 , S2 , · · · , S|A| } denote a multiset of some non-
zero signals in S and h ∈ {1, 2, · · · , M }. Define and for h ∈ {1, 2, · · · , M } with h 6= h0
0
P
αh (A) = S∈A S[h]/|A|. Let h be a given inte- P P
ger in {1, 2, · · · , M }, h be an arbitrary integer in 0 00 S∈S 0 S[h] + (k − 1) S∈S 0 S[h]
αh (S + S (k)) =
{1, 2, · · · , M } with h0 6= h, and S 0 be a given mul- |S 0 | + (k − 1)|S 0 |
tiset of non-zero signals in S. Then for an interval
P
k S∈S 0 S[h]
(β1 , β2 ) with 0 < β1 < β2 < n + 1, there ex- = = αh (S 0 ).
k|S 0 |

China Communications · July 2021 207

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
Thus, there exists a large enough k 0 such that β1 < (3) For each node vi with 1 ≤ i ≤ n, let Ri [h] =
Pn
αh0 (S 0 + S 00 (k)) < n + 1 and αh (S 0 + S 00 ) = αh (S 0 ) j=1 χCi,j,0 (Ch ) for h = 1, 2, · · · , M .
for h ∈ {1, 2, · · · , M } with h 6= h0 . (4) Let v0 be a node which is not in {vi }ni=1 . Let
(3) According to Case 2, there exists S100 such that G0Hid be a directed fully-connected network consisting
αh0 (S 0 +S100 ) ∈ (β2 , n+1) and αh (S 0 +S100 ) = αh (S 0 ) of n + 1 agents, i.e., v0 , v1 , v2 , · · · , vn . At t0 = 0,
for h ∈ {1, 2, · · · , M } with h 6= h0 . Then by Case we construct G0 (0) by adding v1 , v2 , · · · , vn to it. At
1, there exists S200 such that αh0 (S 0 + S100 + S200 ) ∈ the beginning, it is obvious that E(G0 (0)) = ∅ and
(β1 , β2 ) and αh (S 0 + S100 + S200 ) = αh (S 0 + S100 ) for Si,j (0) = ∅ for 1 ≤ i, j ≤ n.
h ∈ {1, 2, · · · , M } with h 6= h0 . Let S 00 = S100 + S200 . (h) (h) (h)
Finally, αh0 (S 0 + S 00 ) ∈ (β1 , β2 ) and αh (S 0 + S 00 ) = (5) For h = 1, 2, · · · , M , let e1 , e2 , · · · , en(h)
αh (S 0 ) for h ∈ {1, 2, · · · , M } with h 6= h0 . be all the edges carrying label Ch in G(−N1 ), where
n(h) is a function whose value is the number of edges
Let {G(t)}t=N 1
t=−N1 be an instance of the DE model.
with label Ch in G(−N1 ). Let n(0) = 0 and h = 1.
At time t0 with h−1
P 0
Ph
In the following, we construct an instance of the l=0 n(l) < t ≤ l=0 n(l), i.e.,
(h)
NFF {G0 (t0 )}N t=N1
t0 =0 , such that {G(t)}t=−N1 is a sub-
2 0
0 < t ≤ n(1), agent det0 −Ph−1 n(l) etar generates
l=0
sequence of {G0 (t0 )}N 2
t0 =0 . The proof consists of 7 a randomly signal St0 ∈ {Ih , (n + 1)Ih } and sends
parts. (h) (h)
it to det0 −Ph−1 n(l) esrc . Let vit0 = det0 −Ph−1 n(l) esrc
(1) Let v1 , v2 , · · · , vn be all the nodes in G(t) l=0
(h)
l=0

for t = −N1 , −N1 + 1, · · · , N1 . Assume that and vjt0 = det0 −Ph−1 n(l) etar . Then according to Sec-
l=0

C1 , C2 , · · · , CM are all the different labels observed in tion II, we have it0 Ajt0 (h, t0 ) = St0 [h] ≥ 1. Let h
{G(t)}t=N run through 1, 2, · · · , M and t0 = M
P
t=−N1 . Let Il be an (M + 1)-dimensional vec- l=0 n(l). Then
1

tor for l = 1, 2, · · · , M + 1, and Ch correspond to Ih at time t = t0 , we have for each agent vi and h0 ∈
0

for h = 1, 2, · · · , M . According to the DE model, we {1, 2, · · · , M }, {j = 1, 2, · · · , n | i Aj (h0 , t0 ) > 0} =

assume without loss of generality that: a) 1 ≤ Ai,h ≤ {j = 1, 2, · · · , n | Ch ∈ Ci,j,0 }. Then according to
n for any i = 1, 2, · · · , n and h = 1, 2, · · · , M ; b) Proposition 1, we have G0 (t0 ) = G(−N1 ).
Ai,h 6= Aj,h for any 1 ≤ i 6= j ≤ n and 1 ≤ h ≤ M . (6) In G0 (t0 ), let →
i Amin,h,t0 =
For each agent (node) vi in G0 (t0 ), its signal genera- 0
min { i Aj (h, t ) | [Xi,j (t)][h] = 1} and i Amax,h,t0 =6→
tor is defined to be qi = (Oi,1 , Oi,2 , · · · , Oi,M ), where max { i Aj (h, t0 ) | [Xi,j (t)] = 0}, where 1 ≤ i, j ≤ n
Oi,h is a binomial distribution. Set si,h ∼ Oi,h . Then and h = 1, 2 · · · , M . According to the NFF (see
 the main text), we can assume that for any i and
6→
 0,
 P(si,h = 0) = 0.5, t0 , i A→
min,h,t0 > i Amax,h,t0 . In the transition from
n+1−2Ai,h
si,h = Ih , P(si,h = Ih ) = 2n , G(−N1 ) to G(−N1 + 1), we assume without loss
 (n + 1)I , P(s = (n + 1)I ) = 2Ai,h −1 ,

of generality that node v1 removes a label Ch0 on
h i,h h 2n
v1 → v2 and re-allocates Ch0 to another directed
6→
and E(si,h ) = Ai,h Ih . Here, P(•) and E(•) denote edge v1 → v3 . Note that 1 A→ min,h0 ,t0 > 1 Amax,h0 ,t0 .
the probability and expectation. For each node vi with Then by Lemma 1, there exists a sequence of signals
1 ≤ i ≤ n, we assume that each signal it can gener- {St0 ,1 , St0 ,2 , · · · , St0 ,m1 (1) } in S(G0 ), such that
ate belongs to S(G0 ) = {IM +1 + M
P
h=1 xh Ih | xh ∈
{0, 1, n + 1}, h ∈ {1, 2, · · · , M }}. Pm1 (1)
|S1,2 (t0 )| 1 A2 (h0 , t0 ) + k=1 St0 ,k
→
(2) For each node vi with 1 ≤ i ≤ n, we assign it 1 Amin,h0 ,t0 >
|S1,2 (t0 )| + m1 (1)
a signal processor Bi (•). Specifically, for a signal S
generated by a node in {vi }ni=1 and h = 1, 2, · · · , M , > 1 A6→
max,h0 ,t0 ,
define Bi (S | Ih ) = S · Ih = S[h]. Then according to
Section II, we have for agents vi and vj and for any h 6= h0
P
S[h]
(
S∈Si,j (t) Pm1 (1)
, if |Si,j (t)| > 0, |S1,2 (t0 )| 1 A2 (h, t0 ) + k=1 St0 ,k
i Aj (h, t) = |Si,j (t)|
= 1 A2 (h, t0 ).
0, if |Si,j (t)| = 0. |S1,2 (t0 )| + m1 (1)

208 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
For t00 = 1, 2, · · · , m1 (1), let agent v2 gener- References
ate signal St0 ,t00 at t0 = t0 + t00 and send it to
[1] M. Newman, Networks. Oxford university press, 2018.
agent v1 . By the above two equations, we have [2] G. Bianconi, Multilayer networks: structure and function.
0 6→ →
1 A2 (h , t0 + m1 (1)) ∈ ( 1 Amax,h0 ,t0 , 1 Amin,h0 ,t0 ), Oxford university press, 2018.
0
1 A2 (h, t0 + m1 (1)) = 1 A2 (h, t0 ) for any h 6= h , [3] D. Lakew, U. Sa’ad, et al., “Routing in flying ad hoc net-
works: A comprehensive survey,” IEEE Communications
and i Aj (h, t0 + m1 (1)) = i Aj (h, t0 ) for any h ∈
Surveys & Tutorials, vol. 22, no. 2, 2020, pp. 1071–1120.
{1, 2, · · · , M }, i 6= 1 and j 6= 2. According to [4] X. Du and H. Chen, “Security in wireless sensor networks,”
Proposition 1, we know that no changes will hap- IEEE Wireless Communications, vol. 15, no. 4, 2008, pp.
pen in the evolution process from t0 = t0 to t01 = 60–66.
[5] C. Dai and Q. Song, “Heuristic computing methods for con-
t0 + m1 (1). Note that 1 A→ 0 0
min,h0 ,t0 > 1 A2 (h , t1 ). tact plan design in the spatial-node-based internet of every-
By Lemma 1 again, there exists a sequence of signals thing,” China Communications, vol. 16, no. 3, 2019, pp. 53–
{St00 ,1 , St00 ,2 , · · · , St00 ,m2 (1) } in S(G0 ), such that 68.
[6] Z. Niu, W. Ma, et al., “Spatial modulation-based ambi-
Pm2 (1) 0 ent backscatter: Bringing energy self-sustainability to mas-
→
|S1,3 (t01 )| 1 A3 (h0 , t01 ) + k=1 St0 ,k sive internet of everything in 6G,” China Communications,
1 Amin,h0 ,t0 > 0 vol. 17, no. 12, 2020, pp. 52–65.
|S1,3 (t1 )| + m2 (1) [7] M. Boguñá, D. Krioukov, et al., “Navigability of complex
> 1 A2 (h0 , t01 ), networks,” Nature Physics, vol. 5, no. 1, 2009, pp. 74–80.
[8] F. Papadopoulos, D. Krioukov, et al., “Greedy forwarding in
dynamic scale-free networks embedded in hyperbolic met-
and for any h 6= h0 ric spaces,” in Proceedings of IEEE INFOCOM, 2010, pp.
1–9.
Pm1 (1) 0 [9] D. Krioukov, F. Papadopoulos, et al., “Greedy forward-
|S1,3 (t01 )| 1 A3 (h, t01 ) + k=1 St0 ,k
0 = 1 A3 (h, t01 ). ing in scale-free networks embedded in hyperbolic metric
|S1,3 (t1 )| + m2 (1) spaces,” ACM SIGMETRICS Performance Evaluation Re-
view, vol. 37, no. 2, 2009, pp. 15–17.
[10] J. Zhang, “Greedy forwarding for mobile social networks
For t000 = 1, 2, · · · , m2 (1), let agent v3 generate sig-
embedded in hyperbolic spaces,” in Proceedings of the ACM
nal St00 ,t000 at t0 = t01 + t000 and send it to agent SIGCOMM Conference, 2013, pp. 555–556.
v1 . Combining the two aspects above, we have [11] K. Bringmann, R. Keusch, et al., “Greedy routing and the al-
0 0 0 0 → gorithmic small-world phenomenon,” in Proceedings of the
1 A3 (h , t1 + m2 (1)) ∈ ( 1 A2 (h , t1 ), 1 Amin,h0 ,t0 ),
0 0 0 ACM Symposium on Principles of Distributed Computing,
1 A3 (h, t1 + m2 (1)) = 1 A3 (h, t1 ) for any h 6= h , 2017, pp. 371–380.
0 0
and i Aj (h, t1 + m2 (1)) = i Aj (h, t1 ) for any h ∈ [12] I. Voitalov, R. Aldecoa, et al., “Geohyperbolic routing and
{1, 2, · · · , M } and i 6= 1 and j 6= 3. addressing schemes,” ACM Special Interest Group on Data
Let X1 = (X1,1 , X1,2 , . . . , X1,n ) be the optimal Communication, vol. 47, no. 3, 2017, pp. 11–18.
[13] S. Buldyrev, R. Parshani, et al., “Catastrophic cascade of
linking strategy of node v1 at time t0 . Let Y = failures in interdependent networks,” Nature, vol. 464, no.
(Y1 , Y2 , · · · , Yn ) and 7291, 2010, pp. 1025–1028.
[14] Z. Chen, J. Wu, et al., “Robustness of interdependent power
 grids and communication networks: A complex network
 X1,2 − Ih0 , if j = 2, perspective,” IEEE Transactions on Circuits and Systems II:
Yj = X + Ih0 , if j = 3, (G.1)
 1,3 Express Briefs, vol. 65, no. 1, 2017, pp. 115–119.
X1,j , otherwise. [15] Y. Xia and D. Hill, “Attack vulnerability of complex com-
munication networks,” IEEE Transactions on Circuits and
Systems II: Express Briefs, vol. 55, no. 1, 2008, pp. 65–69.
According to Proposition 1, we know that Y is the [16] J. Wu, K. Chi, et al., “Analysis of communication network
optimal linking strategy at time t0 = t01 + m2 (1). performance from a complex network perspective,” IEEE
Therefore, we have G0 (t1 ) = G(−N1 + 1) where Transactions on Circuits and Systems I: Regular Papers,
t1 = t01 + m2 (1). vol. 60, no. 12, 2013, pp. 3303–3316.
[17] W. Ren, J. Wu, et al., “A stochastic model of cascading fail-
(7) In the same way, we get t2 , t3 , · · · , t2N1 with ure dynamics in communication networks,” IEEE Transac-
G0 (tk ) = G(−N1 + k) for k = 1, 2, · · · , 2N1 . tions on Circuits and Systems II: Express Briefs, vol. 65,
Above all, we conclude that any instance of the DE no. 5, 2018, pp. 632–636.
[18] X. Fu, H. Yao, et al., “Modeling and analyzing cascading
model, {G(t)}N 1
t=−N1 , can be seen as a sub-sequence dynamics of the clustered wireless sensor network,” Relia-
of an instance of the proposed NFF. bility Engineering & System Safety, vol. 186, pp. 1–10.
[19] F. Parisi, G. Caldarelli, et al., “Entropy-based approach to

China Communications · July 2021 209

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.
 missing-links prediction,” Applied Network Science, vol. 3, Biographies
no. 1, 2018, pp. 1–15.
[20] J. Leskovec, D. Huttenlocher, et al., “Predicting positive Wenjie Jia, received an M.S. degree
and negative links in online social networks,” in Proceed- from Faculty of Mathematics and Statis-
ings of International Conference on World Wide Web, 2010, tics, Hubei University, Wuhan, China. He
pp. 641–650. is currently pursuing the Ph.D. degree in
[21] W. Jia, M. Mariani, et al., “Detecting new edge the School of Electronic Information and
types in a temporal network model,” arXiv preprint Communications, Huazhong University
arXiv:2104.12540, 2021. of Science and Technology. His research
[22] L. Lü, M. Medo, et al., “Recommender systems,” Physics interests include complex networks, data
Reports, vol. 519, no. 1, 2012, pp. 1–49. science and artificial intelligent. Email: [email protected].
[23] J. Coleman, Introducing speech and language processing.
Cambridge university press, 2005.
[24] J. Proakis, Digital signal processing: principles algorithms Tao Jiang, is currently a Chair Pro-
and applications. Pearson Education India, 2001. fessor with the Wuhan National Lab-
[25] L. Lü and T. Zhou, “Link prediction in complex networks: oratory for Optoelectronics, School of
a survey,” Physica A: Statistical Mechanics and its Applica- Electronics Information and Communica-
tions, vol. 390, no. 6, 2011, pp. 1150–1170.
tions, Huazhong University of Science
[26] S. Dorogovtsev, J. Mendes, et al., “Structure of growing net-
and Technology, Wuhan, China. He has
works with preferential linking,” Physical Review Letters,
authored/co-authored about 400 papers in
vol. 85, no. 21, 2000, p. 4633.
[27] L. Lü, T. Zhou, et al., “The h-index of a network node and its major journals/conferences and 6 books in
relation to degree and coreness,” Nature Communications, wireless communications. He served as an Associate Editor for the
vol. 7, 2016, p. 10168. IEEE Network, IEEE Transactions on Signal Processing, IEEE
[28] J. Herlocker, J. Konstan, et al., “Evaluating collaborative Communications Surveys & Tutorials and IEEE Transactions on
filtering recommender systems,” ACM Transactions on In- Vehicular Technology.
formation Systems, vol. 22, no. 1, 2004, pp. 5–53.

210 China Communications · July 2021

Authorized licensed use limited to: PRASHANT DHEER. Downloaded on August 29,2023 at 15:52:20 UTC from IEEE Xplore. Restrictions apply.