IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

23, DECEMBER 1, 2021 16911

Anchored User Selection for Traffic Offloading

Optimization in D2D-Aided Mobile-Edge
Computing
Chenyang Wang , Student Member, IEEE, Ruibin Li , Student Member, IEEE,
Zheng Di, Student Member, IEEE, Chao Qiu , Member, IEEE,
and Xiaofei Wang , Senior Member, IEEE

Abstract—Recently, integrated with the advanced communi-

cation technologies (e.g., 5G) and artificial intelligence (AI),
mobile-edge intelligence (MEI) is regarded as the promising
method to deal with the emerging challenges. Specifically, Device-
to-Device (D2D) communications have been put forward to
reduce the traffic pressure while extending cellular network
capacity. However, the stability of the social network is important
for the design of efficient and reliable traffic offloading strat-
egy, which is often absent from the related work. Besides, most
existing studies merely model the relation between a node pair
as a binary or continuous value, neglecting the rich information
between users. Moreover, many traditional models are conducted
based on small-scale data sets or online Internet services, severely
confining their applications in the D2D scenario. Thus, it is nec-
essary to understand the network structure and select the key
users to address the aforementioned challenges. In this article, we
first propose a network representation model, named MPPT, to
Fig. 1. MEC-based resource allocation in D2D networks.
regard the multidimensional relations as a probability in a third-
order (3-D) tensor space. Then, a mobile D2D social community is
derived by integrating an edge base station (BS) and the nearby computing (MEC) is envisioned as a promising approach to
D2D users, and develop an anchored user selection algorithm address such a challenge. With MEC, computation and con-
to maintain the stability of multiple D2D social communities by
choosing and retaining critical users adaptively under the limited tent resources are placed at the edge of wireless MSNs, called
network resources. Finally, we devise a probability-based onion cellular base station (BS) [1], [2]. The mobile devices can
layers anchored (k, r)-core (P-OLAK) algorithm to identify the get resources from the edge instead of accessing the central
anchor users. The large-scale data sets-based experimental results cloud. Furthermore, Device-to-Device (D2D) communication
show the superiorities of the proposed methods. technology is widely used in multimedia sharing (e.g., video,
Index Terms—Anchored k-core problem, mobile-edge intelli- text, audio, pictures, etc.) without using cellular networks,
gence (MEI), network representation, traffic offloading. which efficiently reduce the traffic pressure in core networks
and improve the quality of IoT services like healthcare devices
services [3]. It provides a lower latency connection between
I. I NTRODUCTION
user pairs in close proximity, which is also an effective method
ITH the fast development of mobile social networks
W (MSNs), duplicate content downloaded posing the rig-
orous challenge of network resources wasted. Mobile-edge
to reduce mobile traffic pressure. Motivated by this, we inte-
grate D2D with MEC to reduce traffic load by expanding the
share of popular multimedia content in MSNs and improve the
Manuscript received January 24, 2021; revised April 29, 2021; accepted computing performance of the cellular network. We denote a
May 15, 2021. Date of publication June 3, 2021; date of current version BS and the D2D users under its coverage as a mobile D2D
November 19, 2021. This work was supported in part by China NSFC community.
(Youth) under Grant 62002260, and in part by the China Postdoctoral Science
Foundation under Grant 2020M670654. This article was presented in part As shown in Fig. 1, we assume the stability of the network
at the Asia Pacific Web (APWeb) and Web-Age Information Management is related to the behavior of individuals, which is called user
(WAIM) Joint International Conference on Web and Big Data (APWeb- engagement. A user with less than three friends, such as v1 and
WAIM), Tianjin, China, Sep. 2020. (Corresponding author: Chao Qiu.)
This work did not involve human subjects or animals in its research. v10 , will drop out of the network and may lead to a cascade
The authors are with the College of Intelligence and Computing, of departures. Therefore, we retain v1 and v10 with incentives
Tianjin University, Tianjin 300072, China (e-mail: chenyangwang@ to prevent the decay of the network. Then, the computation
tju.edu.cn; [email protected]; [email protected]; [email protected];
[email protected]). and multimedia content resource are transferred to v1 and v10
Digital Object Identifier 10.1109/JIOT.2021.3085778 to shared with all other D2D users.
2327-4662 
c 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
16912 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 23, DECEMBER 1, 2021

Fig. 2. Introduction of the embedding from D2D-based MSNs to a 3-D tensor space by the proposed MPPT.

Presently, a large number of scholars use social network of users by adaptively selecting and retaining a set of individ-
analysis technology to study network structure. Specifically, uals for each D2D network under the limitation of resources.
network representation learning (NRL) can map high- Then, we allocate certain resources to these anchors to ensure
dimensional node attributes in a social network into low- that the resource can be shared to other users in the network
dimensional vectors, so as to exploit the vertex classifica- so that the repeated download can be effectively reduced.
tion [4], clustering [5], and link prediction [6]. Further, we proposed a user-edge-cloud framework to select
However, there are two challenges that have not been fully anchors adaptively through communication and cooperation
considered in the previous work. between BS and central cloud. The cloud formulates alloca-
1) Most of the NRL algorithms are used in the analysis tion strategies of anchors based on the process of the structure
of online network, and the attributes on the edge of the information of D2D networks collected by BSs. The edge
users (such as content types, sharing time, and geograph- receives the resources allocated by the cloud and selects
ical information) are merely considered as a binary or the corresponding anchors to reduce the cloud computing
continuous value. stress.
2) Effective and invalid edge relationships still cannot be Contributions of this research are summarized in the fol-
accurately separated in some of the link prediction lowing.
studies considering the edge relations of users. 1) We have proposed a pioneering work that models the
In this article, we conduct multirelational embeddings for multidimensional relationships among users in large-
D2D NRL, and the third-order (3-D) probability-scored tensor scale D2D sharing data sets in the real world as the 3-D
factorization algorithm, named MPPT, is proposed to improve probability tensor instead of considering binary values.
the content sharing prediction. Specifically, as a decisive char- 2) The introduction of GPS slicing in this article avoids
acteristics of D2D networks, global positioning system (GPS) the loss of implicit relationships between D2D users,
is considered to be a crucial relation slice. As shown in Fig. 2, complements the explicit relationship model, and ana-
the 3-D tensor space is derived from a D2D sharing network lyzes true and false triplets more accurately in offline
by considering the multiple edge relations (e.g., App types and MSNs effectively.
GPS) between senders and receivers. 3) An adaptive (k, r)-core anchor selection algorithm is
Besides, many research studies have been carried out on proposed to highlight the stability of mobile D2D
MEC and D2D communications, the stability of networks networks based on the large scale of BSs, aiming at
is largely ignored when designing an efficient and reliable maximizing the total engaged users by adaptively allo-
resource allocation strategies. The decay of an offline D2D cating anchored vertices of the networks, to improve the
MSN, which is determined by the social relations between efficiency and reliability of traffic offload.
D2D users, can lead to the loss of resources due to the con- 4) We devise a user-edge-cloud framework for resource
nections of users are interrupted and reduce the computing allocation and divert computation tasks to BS for pro-
performance of the whole network. cessing to reduce the computing burden of the central
To tackle the aforementioned issue, we propose the adaptive cloud.
anchored (k, r)-core problem considering the large number of 5) We proposed P-OLAK, a probability-based onion lay-
D2D networks under the coverage of BSs in reality and the ers anchored (P-OLAK) algorithm with (k, r)-core to
relations between D2D users affected by spatial and temporal. achieve the adaptive selection of the anchor users in each
The scope of this research lies in maximizing the total number mobile D2D network.

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ANCHORED USER SELECTION FOR TRAFFIC OFFLOADING OPTIMIZATION IN D2D-AIDED MEC 16913

The remainder of this article is organized as follows. uncompressed data for D2D communications. Tang et al. [17]
Section II summarizes the related work. We introduce the pre- applied D2D to manage mobile-edge servers to improve the
liminaries of the system model and formulation in Section III. offloading efficiency of computation tasks.
The design of the proposed MPPT algorithm is illustrated in
Section IV. The anchored user selection is shown in Section V. B. User Engagement and Anchored k-Core Problem
Trace-driven experimental results reveal the effectiveness of
The user engagement phenomenon has been recently studied
the proposed framework in Sections VI and VII. Finally, we
to simulate the human behaviors that they may decide to stay
summarize this article in Section VIII.
engaged in, or drop out of a social network [18], [19]. It is
influenced by that of his/her neighbors. The k-core model, first
II. R ELATED W ORK introduced by Seidman, is widely used to quantify the user
A. MEC and D2D Sharing in MSNs engagement in the social networks [20]. In a k-core model, a
user with less than k neighbors will be eliminated from the
MEC has been developed by the European
social network, and the drop out of a user may lead a cascade
Telecommunications Standard Institute (ETSI) to offer
of departures. This is called the unraveling of a network.
agile and ubiquitous computation for mobile services in
To prevent unraveling, the k-core problem was first proposed
wireless MSNs [7], [8]. Compared to traditional mobile cloud
by Bhawalkar et al. [21] which aims to maximize the engaged
computing, MEC can provide a faster interactive response
users in a social network by selecting a fixed number of users,
by low-latency connections and offload computation tasks
whose presence is important to the stable of the networks.
of mobile devices to edge cloud for processing due to the
Bhawalkar put forward a polynomial-time algorithm based on
resources are placed at the edge of wireless MSNs in physical
bounded treewidth, but it only works on small-scale networks.
proximity to mobile users.
To address this issue, Zhang et al. [22] proposed the OLAK
Specifically, some studies use edge caching to solve the
strategy. The core of the algorithm is to find a set of candidate
problem of traffic offloading, the key issue is which content
anchor vertices and design an auxiliary structure L called onion
should be cached or replaced in the edge of the network. For
layers such that the search space can be significantly reduced.
instance, Yuan et al. [9] investigated the content precaching
Besides, some work also considers the geographical location
problem considering the D2D communications to provide
to measure the engagement of the user with similarity [23].
quick services for users. Wei et al. [10] proposed a similarity-
based edge caching strategy by exploiting the similarity of
popularity determined by the request history and the exist- III. S YSTEM M ODEL
ing ones. For energy-constrained caching, Vallero et al. [11] In the general social networks, most of the edges of
studied the edge caching optimization in energy-efficiency het- users are shown as the simple relation (e.g., sharing activi-
erogeneous radio access networks by considering a renewable ties) with undirected graph presentation. However, there are
energy generation system, energy batteries, and the power grid. many attributes that existed in real-world large-scale social
However, the attributes and relations are not fully discussed networks. For instance, geographic location information and
in the aforementioned studies. sharing activities can be regarded as the key personal tag
Other researchers study the task caching, resource alloca- in D2D networks, which is always used to model the prob-
tion, and task scheduling problems in MEC. Miao et al. [12] lems of group division, community detection, and influence
proposed an online strategy to learn the task patterns of mobile maximization.
devices and formulated the problem as a nonlinear integer pro-
gramming problem to minimize task latency. Chen et al. [13] A. D2D Knowledge Graph
developed a traffic-flow prediction algorithm to train the
Given a directed graph G = {VH , E, VT } as a D2D knowl-
mobile traffic data by using long short-term memory (LSTM)
edge graph (DKG), where V = VH ∪ VT is the users (or
with attention mechanism in single-site mode, and the exten-
entities), VH indicates the head user, and VT signifies the
sion control architecture is proposed to tackle the dynamical
tail user. Let E ⊆ VH × VT stand for the set of relations of
environment with communication and computing resources.
the user pairs, e.g., file type, sharing time, and geographical
Chen et al. [14] proposed a cloudlet-assisted strategy to mini-
information. Furthermore, the edge set E consists of two parts,
mize the energy consumption by compromising among cloud,
i.e., EL (explicit relations) or EP (implicit relations), having
cloudlet, and co-located clouds. Most of the related work does
E = EL + EP .
not consider the D2D communication, but the participation
of user behavior directly affects the dynamic changes of the
network environment. B. D2D Network Representation Learning
Besides, D2D communications for MSN content sharing are In order to learn the representation of DKG, we have the
attracting more and more research interests from academia. It following D2D network representation learning (D2D-NRL).
is widely used in video and voice communication, gaming, and Suppose a DKG G with n users and m entities, we have obser-
multimedia sharing between users [15]. Some studies also inte- vation with T + = {(h, t, k)}, where each triplet contains a head
grate D2D with edge computing to enhance the performance of user h ∈ VH , a tail user t ∈ VT , and a sharing relation k ∈ E.
D2D communications. Qin et al. [16] proposed an edge com- The D2D NRL process is to construct a learning procedure
puting and relaying scheme to strengthen the throughput of of the continuous vector representation of users and content

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
16914 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 23, DECEMBER 1, 2021

Fig. 3. Slices of a third-order tensor.

Fig. 4. Relation tensor space X decomposition.

sharing relationships in low-dimensional space, whereas users
are usually expressed as vectors in a two-bit space, and con-
IV. MPPT: M ULTIRELATIONAL P REDICTION BASED ON
tent relations are usually expressed as operations, including
P ROBABILITY-S CORED T ENSOR FOR D2D-NRL
matrices, tensors, etc.
Definition 1 (Tensor): A tensor is a multiway space with the In offline social networks, edge relationships depend on user
higher order generalization of matrices and vectors. Formally, behavior preferences in a network. For example, users with
an Nth-order tensor (denoted by A ∈ RI1 ×I2 ···×IN ) is generated the same preferences or close relationships often share various
by the tensor product of N vectors, where the element of A content files (e.g., audio, video, app, etc.)
is signed by ai1 i2 ··· iN , iN ∈ 1, 2, . . . , In with 1 ≤ n ≤ N. The relations (edges) in DKG are generally expressed by
Definition 2 (Slices): Slices are the 2-D parts of a tensor the binary or continuous value problem so that the compre-
space, determined by fixing all but two indices. As presented in hensive relational information cannot be fully revealed. This
Fig. 3, we use Xi , Xj , and Xk to indicate the horizontal, lateral, article outlines the offline interactions of users to derive the
and frontal slices in the 3-D tensor space X , respectively. rich implicit information in the networks.
First, rewrite the VH and VT of the DKG as the sender
(head user) and receiver (tail user) in the D2D networks. Then,
C. 2-Order Tensor Relational Model the relation prediction r(h, t) can be derived from the logistic
1) Definition of Relational Tensors: We define a 3-D tensor function
space X ∈ Rn×n×m with n entities and m relations in a DKG
G, the cube xijk = {−1, 0, 1} in tensor X consists of a triplet Phtk = σ (ϕ(h, t, k)) (4)
(i, j, k), which indicting the relational fact of users is reverse, where σ (·) is the sigmoid function by (1/[1 + e−x ]), and ϕ(·)
false, or direct for ∀i, j, k. is the score function. Phtk ∈ [0, 1] ∀i, j, k indicating the cor-
2) D2D-NRL-Based Relational Reasoning: Given the col- responding probability when Phtk = P(Yhtk = 1), i.e., the
lective multirelational data, we decompose the relation tensor relation r(h, t) is true.
into the combination form of a factor matrix A ∈ Rn×r by Define xijk as the prediction of the probabilities of the
a core tensor W ∈ Rr×r×m , where r is the latency coeffi- implicit and explicit triples for all the users v ∈ V, where
cient. Denote that Xk as the k relations in kth slices among xijk = Phtk . We can rewrite the relational probability as
entities. The decomposition formula can be regarded as the follows:
approximately expression in the following:
P = xijk = σ (ϕ(h, t, k)) = P(Yhtk = 1). (5)
Xk ≈ AWk A , for k = 1, . . . , m
T
(1)
The proposed MPPT realizes the D2D-NRL by modeling
which can be concluded as the regularized minimization the sharing activities in terms of “sender–relation–
problem in the following: receiver,” Fig. 4 illustrates the decomposition process of
the multidimensional relations of user interactions (e.g., APP
  
1  
Xk − AWk AT  + λ A 2 +
2  types and GPS information) in the 3-D tensor space X. the
min F F Wk 2F . user multidimensional interactions (various APP types and
A,Wk 2
k k GPS information) in 3-D tensor spaces. We can obtain the
(2) score function as follows in accordance with the observation
relation k between the sender h and receiver t:
Then, the alternating least squares (ALS) method is used to
elevate the computation efficiency of (2). The factor matrices ϕ(h, t, k) = eTh Wk et . (6)
A and Wk can be obtained by We solve the aforementioned score function by using
 m  −1 AdaGrad and stochastic gradient descent (SGD) to minimize
 
m
A← Xk AWTk + XTk AWk Bk + Ck + λI (3) the negative log-likelihood loss function with L2 regularization

k=1 k=1 min f (eh , Wk , et ) + g(eh , Wk , et ) (7)
eh ,Wk ,et
where Bk = Wk AT AWTk , Ck = WTk AT AWk , existing the r(h,t)∈

Kronecker product form of Rk ← ((A ⊗ A)T (A ⊗ A) + where f (eh , Wk , et ) = log(1 + exp(−Yhtk ϕ(h, t, k))), and
λI)−1 (A ⊗ A)vec(Xk ). g(eh , Wk , et ) = λ( eh 22 + Wk 2F + et 22 ).

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ANCHORED USER SELECTION FOR TRAFFIC OFFLOADING OPTIMIZATION IN D2D-AIDED MEC 16915

Algorithm 1 Multirelational Prediction Algorithm in D2D Algorithm 2 k-Core Selection

Network by SGD Require: a D2D knowledge graph G, degree constraint k;
Require: Train set , embedding dimension k, negative rate factor Ensure: k-core subgraph Ck (G).
η, learning ratio α, batch size b, max iteration m, stopping s, 1: for ∀u ∈ G with deg(u) < k do
regularization parameters λ, validation set v . 2: Ck (G) = G \ {u};
Ensure: The sender-receiver pairs and embeddings eh . 3: end for
1: ei = random select (k), for ∀i in E; 4: return Ck (G)
2: Wi = random select (k, k), for ∀i in V;
3: while i = 1, · · · , m do
4: for j = 1, · · · , ||/b do Theorem 1: Given any ground truth over a set of sender–
5: b ← sample(, b, η);
6: Upgrade the embeddings eh with receiver pairs E and relations V, existing an MPPT scheme

r(h,t)∈b ∇ϒ(r(h, t); );
with the n-dimensional sender and receiver embeddings of
7: Upgrade the learning ratio α with AdaGrad; dimensionality r and r×r relation embeddings, where r = n is
8: end for the number of sender–receiver pairs, representing the ground
9: if i mod s = 0 then truth.
10: MRR on  decreases
11: break; Proof: Denote eh and et as the n-dimensional binary vector
12: end if representations of sender eh and receiver et , respectively. For
(i) (j)
13: end while each sender eh , receiver et , and relation r(k) , let the vec-
tors eh = 1 and et = 1 of the ith and jth element, and the
vectors of other elements are 0. The (i, j)th element in matrix
Bring the regularization coefficient λ and training set γ into W ∈ Rr×r is 1 when the relation of (eh , et ) is true, 0 other-
(7), we have the rewritten expression in the following: wise. Thus, the original tensor space can be represented by
 the tensor embeddings products of sender–receiver pairs using
ϒ() = f (eh , Wk , et ) + g(eh , Wk , et ). (8) the logistic sigmoid.
r(h,t)∈ Exploring the potential user relationships in the D2D
The SGD method is used to deal with f (eh , Wk , et ), and we networks is helpful to realize the discovery of the popular
have the gradients of the score function in the following: content dissemination users, achieve the purpose of efficient
content distribution and traffic offloading, thereby reducing the
∇eh ϕ(h, t, k) = Wk et traffic pressure on the backbone network. We introduce the
∇Wk ϕ(h, t, k) = eh eTt process of the key user selection algorithm in the following
∇et ϕ(h, t, k) = WTk eh . (9) section.

Finally, we have the optimization on the embeddings eh in V. A NCHORED U SER S ELECTION

ϒ() as shown in the following: A. k-Core Selection Model
∇eh ϒ(r(h, t)) = −Yhtk ϕ(h, t, k)σ (Wk et ) + 2λeh Given a DKG G and the binary vector representations eh ,
∇Wk ϒ(r(h, t)) = −Yhtk ϕ(h, t, k)σ eh eTt + 2λWk if |eh | = 1, we consider that there exists a friendship relation
between the two users. Thus, we define deg(u) = |eh |v as the
∇et ϒ(r(h, t)) = −Yhtk ϕ(h, t, k)σ WTk eh + 2λet . (10) degree of user u. Let GN = {u, G}, u ∈ VH denote the set of
We demonstrate the procedures of the multirelational neighbors of user u. Furthermore, given a subgraph J of G,
prediction algorithm in the D2D network by SGD in where J ⊆ G. Define GN (J ) as the total number of neighbor
Algorithm 1. Similar to [24], we obtain the negative triplets η users in set J . Note that the k-core is the largest connected
in each positive train triplet from the sender or the receiver of subgraph of network, representing the most closely part of the
the positive triplets when  merely has positive triplets, and network between users, where each user has no less than k
related to embeddings eh , Wr , et . First, we randomly select friends.
the sender–receiver pairs and generate the relation embeddings In the following, we have the formal definition of k-core in
of D2D networks (lines 1 and 2). We iteratively conduct the a DKG. Given a DKG, if the subgraph J represents the k-core
gradient for parameter and upgrade the learning ratio α with Ck (G) of the DKG G, it holds the conditions as follows.
AdaGrad in lines 3–6 and line 7, respectively. Finally, we can 1) For each user in J , deg(u) ≥ k ∀u ∈ J .
have the relationships among each sender–receiver pairs by 2) The subgraph J is the largest connected subgraph of G,
 
quantifying them as slices and append them to tensor space for ∀J ⊂ J , J is not the k-core of Ck (G).
for a subsequent solution. We can obtain the k-core selection algorithm from its def-
For any real triple, the tensor decomposition model is inition, shown as Algorithm 2, where G is an input DKG, k
considered to be fully representational if there is an embed- is the degree constraint of DKG, the number of k-core can be
ding of the transmission–receive pair sum relation. Therefore, realized by deleting the users with less than k degrees. The
we can accurately distinguish between true triples and false computation complexity of Algorithm 2 is O(n).
triples. Since there are certain contradictions between different
types of relations, we use the embedding dimension of the B. Anchored k-Core Model
sender–receiver pair to establish the rank of the decomposition In our theory, once a user v in G is selected, it is always
(i.e., a boundary), as shown in Theorem 1. retained in k-core regardless if it satisfies the degree constraint.

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
16916 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 23, DECEMBER 1, 2021

We call v as an anchored vertex of G. Formally, given the

set of vertices A ⊆ G, we define the Ck (GA ) as the corre-
sponding anchored k-core of G with users in set A selected.
Besides, we denote the neighbors of the anchor users in set
A by F(A, G) = {Ck (GA )\{Ck (G) ∪ A}. The size of the neigh-
bors reflects the effectiveness of anchored users. In Fig. 1, if
we anchor v1 in C3 (G), that is, A={v1 }, then {v2 , v3 , v4 } could
also retain in C3 (G). Therefore, we obtain the new anchored
3-core C3 (GA )={v1 , v2 , v3 , v4 , v5 , v6 , v7 , v8 }. The neighbors of
v1 , F(v1 , G) = {v2 , v3 , v4 }.
In addition, the study of anchored k-core developed by
Zhang proved that only the users in (k-1)-shell can be neigh-
bors [22]. Thus, we define the set of users that probable
become anchors as candidate anchor users, denoted as L, con-
sisting of the vertices in (k-1)-shell and the neighbors which Fig. 5. User-edge-cloud allocation framework.
are not in the selected k-core. In order to reduce the anchor’s
searching complexity, we only consider the users in L in this
article. allocation process on the cloud, two computation processes on
BS, and three data transmission processes.
C. Problem Formulation Specifically, the three transmission processes transfer
Considering that the connections of users in D2D networks resources and data through two links.
are limited by time and space, we take users’ spatiotemporal 1) Allocation Link: The central cloud transfers the compu-
similarity into account and introduce the similarity (k, r)-core tation and content resource to BSs for anchors selection.
model to measure the users’ behaviors of engaged, where r is 2) Cellular Link: Each BS transfers the content resource to
the threshold of similarity between two users. anchors for sharing to other D2D users.
We define the similarity (k, r)-core model as: given an For each BS, the first computation process is computing
attributed graph G, a degree constraint k, and a similarity the anchor ratio of the covered D2D network and the sec-
threshold r,1 the similarity (k, r)-core, denoted by Ckr (G), is ond computation is selecting the corresponding number of best
a connected subgraph S ⊆ G that the degree of all users in anchors after received the resource allocated from the cloud.
V(Ckr (G)) satisfies the degree constraint k, and the similar- In the central cloud, the allocation process allocates computa-
ity of any user pair u,v in E(Ckr (G)) satisfies the similarity tion and content resource to the D2D networks according to
constraint r, sim(u, v) ≥ r. the anchor ratio. Thus, we allocate computing resource and
In this article, we aim at maximizing the total engaged users remove the computation task from the cloud to BS to process
of all D2D networks under the coverage of large-scale BSs so that the computing burden of the central cloud is effectively
with limited resources. Based on this, the sharing and spread- reduced.
ing of resources are reliably maintained so that the repeated
download can be effectively prevented.
E. Adaptive Selection Algorithm of Anchor User
The problem of adaptive anchored (k, r)-core is proposed to
address the above issues based on (k, r)-core where r denote To effectively search the appropriate anchor users with
the spatiotemporal similarity threshold. Specifically, we give the largest neighbors numbers, the existing solution does not
a group of disjoint attributed graphs J as the large scale of fully consider the order of deleting non-k-core users. In order
D2D networks and a budget B as the total number of anchors to improve the efficiency of calculation of neighbors, they
could retain. For each network, we adaptively find a specific proposed the onion layers anchored k-core (OLAK) algo-
number of anchor vertices and the size of the total anchor rithm [22]. It came up with a hierarchical structure called
users set, |A| = B, so that the resulting anchored (k, r)-cores onion layers, which is transformed from the set of candidate
of all networks, Ckr (JA ), are maximized; that is, F(A, J ) is anchor users. Based on this, we propose the probability-
maximized. The offloaded traffic is indicated as the enlarged based onion layers anchored (k, r)-core (P-OLAK) algorithm
degree of anchored (k, r)-core, that is, the size of neighbors. to select the anchor users for each D2D network. The key
point is computing the onion layer to reduce time complexity
D. User-Edge-Cloud Framework for Allocation by removing the users that not satisfy both degree and the rela-
tion embeddings constraint at the same time from L at each
To reduce the computing burden of the cellular network and iteration until the users in L are all from k-core.
provide a faster interactive response via low-latency connec- Algorithm 3 indicates the process of computation, where
tions, we devise a user-edge-cloud framework for anchor user b is the corresponding number of anchor users for selection
selection and resource allocation in D2D networks. The pro- through the allocation process. After the onion layer is built,
cedure of the whole process is shown in Fig. 5, including one we identify the neighbors of a users layer by layer and we
1 Here, the similarity threshold can be regarded as the embeddings e of
h
select the anchors with the largest neighbors numbers at each
sender–receiver pairs and relation obtained from Algorithm 1. iteration. Algorithm 4 illustrates the details of P-OLAK. The

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ANCHORED USER SELECTION FOR TRAFFIC OFFLOADING OPTIMIZATION IN D2D-AIDED MEC 16917

Algorithm 3 Onion Computing (G, k, r)

Require: the social network G, degree constraint k;
Ensure: the onion layers Ls0
1: N = Ck−1,r (G);
2: i = 0;
3: while N  = Ckr (G) do
4: i = i + 1;
5: Li = { v | deg(v, N) < k & v ∈ N};
6: N = N\Li ;
7: end while
8: L0 = { v | v ∈ NB(Li1 , G) \ { N ∪ Li1 } };
9: return Li0

Algorithm 4 P-OLAK(G, k, r, b)
Fig. 6. APP relations mapping from Google Play Store [1].
Require: the social network G, degree constraint k;
Ensure: the set of selected anchor vertices (S).
1: S = ∅; i = 0;
2: L ← Onion Computing(G, k, r) (Algorithm 1); VII. DATA P REPROCESSING
3: while i < b do
We experimented with a real large-scale offline user transfer
4: for each v in L do
5: Compute F(S ∪ v, G); data set from Xender. It is one of the world’s leading D2D
6: end for content sharing [25]. In the data set, the user behavior con-
7: v∗ ← best anchor users in iteration; tains nine columns (FileType, MD5, SenderID, ReceiverID,
8: S = S ∪ v∗ ; i = i + 1; TimeStamp, UserIP, Country, GPS, and FilesSize). We selected
9: end while
three of the most valuable items <APP name, sender,
10: return S
receiver> in the log on 01/08/2016 for the experimental
simulation.
TABLE I
S TATISTICS OF L ARGE G ROUPS
A. Labels of APP Relations and Metadata Conduct
Aiming to predict sharing activities of users and conduct
interest recommendations, we retrieved all APP names and
categories from Google Play as their labels, noting that each
APP only has a unique label. Then, the captured data table
<APP name and the label> properties for each row will
be matched with the experimental data table, and the APP
time complexity of the anchor’s selection is O(nm) in the name will be converted to the corrected label. For instance,
worst, n and m correspond to the size of candidates and the WhatsApp tagging is an example illustrated in Fig. 6. The
cost of neighbors computation. name of WhatsAPP is replaced by the mapped Communication
Relation label, which generates the metadata of our experiment
named Xender-Filtered.
VI. E XPERIMENTS In the experiment, the data set filtered by Xender consisted
A. Data Set Validation and Settings of 64 028 transmissions of 72 550 users with 48 relationship
labels associated with the APP type. Be noted that MPPT
The large-scale data set used is the traces of a realistic
specifically takes into account the geographic cosine sim-
mobile D2D sharing activities, named Xender. Users can trans-
ilarity [25], which is the 49th label we use. Specifically,
mit multimedia content via mobile D2D communication. In
we describe 48 tags for the APP, of which the top two are
order to improve the experimental efficiency, we preprocess
Communication and Tools shown in Fig. 7.
the raw data, including extracting useful data items and divid-
Since that there only exist positive triples in our data set,
ing users into D2D social groups. We define a D2D social
we generate a negative triplet [26] based on the locally closed
group as the user pairs who have conducted a content sharing
world hypothesis. Thus, we obtain the negative examples by
activity at least once, they will be gathered and appear only
randomly changing the sender or receiver. In the following
once in a group. In our experiments, we simulate a D2D social
experiment, negative sampling was used in the process of
group as a mobile D2D network covered by a BS.
positive example training.
The size of the preprocessed data set is about 96 GB with
seven data dimensions include the sender ID, receiver ID, GPS
record, timestamps, etc. We classify the total groups by their B. Clustering and Geographical Similarity
size and conduct the experiment on large groups, which have Different from online relationships, the content of offline
more than 20 nodes. Table I shows the statistics of large groups MSNs is restricted by geographical location and always fol-
with total of 7179 graphs. The maximum value of k can only lows the principle of homogeneity. Considering the diversity of
be 5 due to the sparse of the offline data set. geographical position in the real world, the actual movement

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
16918 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 23, DECEMBER 1, 2021

geographic similarity. It can be clearly seen from Fig. 8(b)

that the geographical locations of active users are always
similar, with more than 70% of users having cosine sim-
ilarities above 0.7, which also manifests that geographical
factor has a significant impact on the offline transmission.
Therefore, geographic cosine similarity is used as the crit-
ical relation slice between vertices (users) among the 3-D
tensor space in MPPT to improve the accuracy of transmission
prediction.
Considering the number of vertices and the relations
between mobile users in large groups, we set k = 3 and
r = 0.45 and vary the budget of anchor resource, that
is, the total number of anchors, for further evaluation. We
call the proposed framework and P-OLAK together as the
Self-Adaptive method. Furthermore, we implement different
anchor selection algorithms for comparison to validate the
Fig. 7. Forty eight types of APP from Xender [1]. performance of Self-Adaptive methods.
1) OLAK: It is the origin OLAK algorithm with the budget
b is fixed, that is, choosing a fixed number of anchor
vertices for each group [22].
2) Random: It is an adaptive algorithm to randomly choose
the specific number of anchor vertices from the candi-
dates for each network according to its anchor ratio.
3) Degree: It is an adaptive algorithm to choose the specific
number of anchor vertices with the highest degree for
each network according to its anchor ratio.
4) Optimum Method: It is an adaptive method where for
each graph separately, using OLAK to obtain the largest
k-cores by gradually increasing the budget b.

C. Results and Discussion

From Fig. 9(a) and (b), we can observed that the difference
of the result is greater when anchoring more users. And com-
pare the followers between Fig. 9(a) [respect to Fig. 9(c)] and
Fig. 9(b) [respect to Fig. 9(d)], the result is more effective in
the condition of k = 4. This also proved that our algorithm
works better in a bigger number of groups or under a harsher
condition of degree constraint.
The comparison of the Self-Adaptive method between the
original nonadaptive OLAK algorithm with k = 3 based on
top 100 large groups. We consider the uniform anchors selec-
tion for Non-Adaptive, i.e., 100 anchors for a simulation time
while the arbitrary anchor’s selection for the Self-Adaptive.
Fig. 8. Preprocessing results of location relation. (a) Clustering results.
The experimental results shown in Fig. 10(a) and (b) illus-
(b) Geographic similarity [28]. trate that the effectiveness of traffic offloading, revealed by
the number of followers and average followers, can be signif-
of each user is not the scope of a single point, DBSCAN [27] icantly improved by about 40%. In addition, as the number
in this article is used to cluster the geographical information of anchors increases, the growth rate of followers gradually
of users. Then, the clustering is used as the eigenvector of ver- decreases. That is because the total number of engaged users
tices to calculate the geographic cosine similarity between user of all networks has reached the upper bound, due to the con-
pairs. Regarding the influence of latitude, longitude, and sam- nections between mobile users in offline social networks are
ple number in the cluster, we used the scanning radius value of sparse. Therefore, it is also significant for the central cloud to
0.05 and the minimum sample size of 1. Therefore, we group provide an appropriate amount of resources of anchors for the
64 028 transmissions into 5590 packets. The clustering results D2D network.
are shown in Fig. 8(a). We conduct the comparative experiment to evaluate the
Since the geographic similarity of all users is sparse, we performance of Self-Adaptive and another three adaptive
select some active user pairs to observe the distribution of algorithms mentioned above. As demonstrated in Fig. 11,

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ANCHORED USER SELECTION FOR TRAFFIC OFFLOADING OPTIMIZATION IN D2D-AIDED MEC 16919

Fig. 9. Comparison of adaptive algorithms. (a) k = 3 Total Followers. (b) k = 4 Total Followers. (c) k = 3 Avg Followers. (d) k = 4 Avg Followers.

(k, r)-core problem based on the large scale of D2D networks

in consideration of both effectiveness and efficiency.

VIII. C ONCLUSION
In order to deal with the mobile traffic explosion, mobile-
edge intelligence (MEI) is proposed as a promising method,
wherein D2D communication has been promoted to reduce
the traffic pressure and improve the computation capacity.
Fig. 10. Comparison of nonadaptive methods. (a) Total Followers. (b) Avg Moreover, considering the stability of the social network and
Followers. the rich information between users. We first propose MPPT,
a network represented model to map the multidimensional
relations as a probability in a third-order (3-D) tensor space.
Then, the similarity-based onion layers anchored (k, r)-core
(P-OLAK) algorithm is proposed to identify the anchor users
and solve the optimization problem of traffic offloading.

R EFERENCES
[1] Q. Zhang, X. Ren, Y. Cao, H. Zhang, X. Wang, and V. Leung,
“Content sharing prediction for device-to-device (D2D)-based offline
mobile social networks by network representation learning,” in Proc.
Asia–Pac. Web (APWeb) Web-Age Inf. Manage. (WAIM) Joint Int. Conf.
Web Big Data, 2020, pp. 112–126.
Fig. 11. Comparison of adaptive methods. [2] X. Wang, C. Wang, X. Li, V. C. M. Leung, and T. Taleb, “Federated
deep reinforcement learning for Internet of Things with decentralized
cooperative edge caching,” IEEE Internet Things J., vol. 7, no. 10,
pp. 9441–9455, Oct. 2020.
[3] Y. Hao, Y. Miao, L. Hu, M. S. Hossain, G. Muhammad, and S. U. Amin,
“Smart-edge-CoCaCo: AI-enabled smart edge with joint computation,
caching, and communication in heterogeneous IoT,” IEEE Netw., vol. 33,
no. 2, pp. 58–64, Mar./Apr. 2019.
[4] J. Chen, Z. Gong, W. Wang, and W. Liu, “HNS: Hierarchical nega-
tive sampling for network representation learning,” Inf. Sci., vol. 542,
pp. 343–356, Jan. 2021.
[5] A. Çelikkanat and F. D. Malliaros, “Topic-aware latent models for rep-
resentation learning on networks,” Pattern Recognit. Lett., vol. 144,
pp. 89–96, Apr. 2021.
[6] Z. Li, X. Wang, J. Li, and Q. Zhang, “Deep attributed network repre-
sentation learning of complex coupling and interaction,” Knowl. Based
Syst., vol. 212, 2021, Art. no. 106618.
Fig. 12. Comparison of running time. [7] Y. He, F. R. Yu, N. Zhao, and H. Yin, “Secure social networks in
5G systems with mobile edge computing, caching, and device-to-device
communications,” IEEE Wireless Commun., vol. 25, no. 3, pp. 103–109,
the Self-Adaptive method can identify a larger number of Jun. 2018.
followers, which means more data can be offloaded from the [8] C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation
cellular network. Besides, although the optimum algorithm can offloading and resource allocation in wireless cellular networks with
mobile edge computing,” IEEE Trans. Wireless Commun., vol. 16, no. 8,
get a better result, it is computationally expensive. pp. 4924–4938, Aug. 2017.
Further, we run the three comparison methods 20 times and [9] P. Yuan, Y. Cai, Y. Liu, J. Zhang, Y. Wang, and X. Zhao, “ProRec:
computed the average running time. As shown in Fig. 12, A unified content caching and replacement framework for mobile edge
computing,” Wireless Netw., vol. 26, pp. 2929–2941, Jan. 2020.
the consuming time of the Self-Adaptive method is the low- [10] X. Wei, J. Liu, Y. Wang, C. Tang, and Y. Hu, “Wireless edge caching
est. Meanwhile, though the Optimum Method can maintain based on content similarity in dynamic environments,” J. Syst. Archit.,
the largest number of vertices of the graphs, we cannot vol. 115, May 2021, Art. no. 102000.
[11] G. Vallero, M. Deruyck, M. Meo, and W. Joseph, “Base station switching
afford the consuming time. This proves that the Self-Adaptive and edge caching optimisation in high energy-efficiency wireless access
method is the best method to solve the adaptive anchored network,” Comput. Netw., vol. 192, Jun. 2020, Art. no. 108100.

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.
16920 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 23, DECEMBER 1, 2021

[12] Y. Miao, Y. Hao, M. Chen, H. Gharavi, and K. Hwang, “Intelligent task Ruibin Li (Student Member, IEEE) received
caching in edge cloud via bandit learning,” IEEE Trans. Netw. Sci. Eng., the B.S. degree from the School of Computer
vol. 8, no. 1, pp. 625–637, Jan.–Mar. 2021. Science and Technology, College of Intelligence
[13] M. Chen, Y. Miao, H. Gharavi, L. Hu, and I. Humar, “Intelligent traffic and Computing, Tianjin University, Tianjin, China,
adaptive resource allocation for edge computing-based 5G networks,” in 2019, where he is currently pursuing the M.S.
IEEE Trans. Cogn. Commun. Netw., vol. 6, no. 2, pp. 499–508, degree.
Jun. 2020. His current research interests include edge
[14] M. Chen, Y. Hao, C.-F. Lai, D. Wu, Y. Li, and K. Hwang, “Opportunistic caching, distributed federated learning optimization,
task scheduling over co-located clouds in mobile environment,” IEEE reinforcement learning, edge intelligence, and deep
Trans. Services Comput., vol. 11, no. 3, pp. 549–561, Jun. 2018. learning.
[15] C. Wang, R. Li, W. Li, C. Qiu, and X. Wang, “SimEdgeIntel: A
open-source simulation platform for resource management in edge
intelligence,” J. Syst. Archit., vol. 115, May 2021, Art. no. 102016.
[16] M. Qin, L. Chen, N. Zhao, Y. Chen, F. R. Yu, and G. Wei, “Computing
and relaying: Utilizing mobile edge computing for P2P communi-
cations,” IEEE Trans. Veh. Technol., vol. 69, no. 2, pp. 1582–1594,
Feb. 2020.
[17] W. Tang, X. Zhao, W. Rafique, L. Qi, W. Dou, and Q. Ni, “An offloading Zheng Di (Student Member, IEEE) received the
method using decentralized P2P-enabled mobile edge servers in edge B.S. degree in software engineering from Tianjin
computing,” J. Syst. Archit., vol. 94, pp. 1–13, Mar. 2019. Normal University, Tianjin, China, in 2018. She is
[18] R. Chitnis, F. V. Fomin, and P. A. Golovach, “Parameterized complex- currently pursuing the M.S. degree with the School
ity of the anchored k-core problem for directed graphs,” Inf. Comput., of Computer Science and Technology, College
vol. 247, pp. 11–22, Apr. 2016. of Intelligence and Computing, Tianjin University,
[19] R. H. Chitnis, F. V. Fomin, and P. A. Golovach, “Preventing unraveling Tianjin.
in social networks gets harder,” in Proc. 27th AAAI Conf. Artif. Intell., Her current research interests include edge com-
2013, p. 1. puting and computing power network.
[20] S. B. Seidman, “Network structure and minimum degree,” Soc. Netw.,
vol. 5, no. 3, pp. 269–287, 1983.
[21] K. Bhawalkar, J. Kleinberg, K. Lewi, T. Roughgarden, and A. Sharma,
“Preventing unraveling in social networks: The anchored k-core
problem,” SIAM J. Discrete Math., vol. 29, no. 3, pp. 1452–1475, 2015.
[22] F. Zhang, W. Zhang, Y. Zhang, L. Qin, and X. Lin, “OLAK: An effi-
cient algorithm to prevent unraveling in social networks,” Proc. VLDB
Endowment, vol. 10, no. 6, pp. 649–660, 2017.
[23] Q. Zhu, H. Hu, C. Xu, J. Xu, and W.-C. Lee, “Geo-social group queries Chao Qiu (Member, IEEE) received the B.S.
with minimum acquaintance constraints,” VLDB J., vol. 26, no. 5, degree in communication engineering from China
pp. 709–727, 2017. Agricultural University, Beijing, China, in 2013, and
[24] T. Trouillon, J. Welbl, S. Riedel, É. Gaussier, and G. Bouchard, the Ph.D. degree in information and communication
“Complex embeddings for simple link prediction,” in Proc. Int. Conf. engineering from the Beijing University of Posts and
Mach. Learn. (ICML), 2016, pp. 2071–2080. Telecommunications, Beijing, in 2019.
[25] X. Wang, Y. Zhang, V. C. M. Leung, N. Guizani, and T. Jiang, “D2D She is currently a Lecturer with the School
big data: Content deliveries over wireless device-to-device sharing in of Computer Science and Technology, College
large-scale mobile networks,” IEEE Wireless Commun., vol. 25, no. 1, of Intelligence and Computing, Tianjin University,
pp. 32–38, Feb. 2018. Tianjin, China. From September 2017 to September
[26] A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, 2018, she visited Carleton University, Ottawa, ON,
“Translating embeddings for modeling multi-relational data,” in Canada, as a Visiting Scholar. Her current research interests include machine
Advances in Neural Information Processing Systems. Red Hook, NY, learning, software-defined networking, and blockchain.
USA: Curran, 2013, pp. 2787–2795.
[27] Z. Yuan, Y. Jiang, and G. Gidófalvi, “Geographical and temporal sim-
ilarity measurement in location-based social networks,” in Proc. 2nd
ACM SIGSPATIAL Int. Workshop Mobile Geograph. Inf. Syst., 2013,
pp. 30–34.
[28] Y. Zhang, Z. Huang, S. Wang, X. Wang, and T. Jiang, “Spark-based
measurement and analysis on offline mobile application market over
device-to-device sharing in mobile social networks,” in Proc. IEEE Xiaofei Wang (Senior Member, IEEE) received the
23rd Int. Conf. Parallel Distrib. Syst. (ICPADS), Shenzhen, China, 2017, master’s and Doctoral degrees from Seoul National
pp. 545–552. University, Seoul, South Korea, in 2006 and 2013,
respectively.
He is currently a Professor with the Tianjin
Key Laboratory of Advanced Networking, School
of Computer Science and Technology, Tianjin
Chenyang Wang (Student Member, IEEE) received University, Tianjin, China. He was a Postdoctoral
the B.S. and M.S. degrees in computer science Fellow with The University of British Columbia,
and technology from Henan Normal University, Vancouver, BC, Canada, from 2014 to 2016.
Xinxiang, China, in 2013 and 2017, respectively. Focusing on the research of social-aware cloud
He is currently pursuing the Ph.D. degree with computing, cooperative cell caching, and mobile traffic offloading, he has
the School of Computer Science and Technology, authored over 100 technical papers in the IEEE J OURNAL ON S ELECTED
College of Intelligence and Computing, Tianjin A REAS IN C OMMUNICATIONS, the IEEE T RANSACTIONS ON W IRELESS
University, Tianjin, China. C OMMUNICATIONS, the IEEE W IRELESS C OMMUNICATIONS, the IEEE
He has been a visiting Ph.D. student under the C OMMUNICATIONS, the IEEE T RANSACTIONS ON M ULTIMEDIA, the IEEE
support of China Scholarship Council with the INFOCOM, and the IEEE SECON.
School of Electrical Engineering, Aalto University, Prof. Wang was a recipient of the National Thousand Talents Plan (Youth)
Espoo, Finland, since May 15, 2021. His current research interests include of China. He received the Scholarship for Excellent Foreign Students in IT
edge computing, big data analytics, reinforcement learning, and deep learning. Field by NIPA of South Korea from 2008 to 2011, the Global Outstanding
Mr. Wang received the Best Student Paper Award of the 24th International Chinese Ph.D. Student Award by the Ministry of Education of China in 2012,
Conference on Parallel and Distributed Systems by IEEE Computer Society and the Peiyang Scholar from Tianjin University. In 2017, he received the
in 2018. Fred W. Ellersick Prize from the IEEE Communication Society.

Authorized licensed use limited to: CINVESTAV. Downloaded on September 01,2022 at 00:04:07 UTC from IEEE Xplore. Restrictions apply.