Received November 12, 2019, accepted December 16, 2019, date of publication December 20, 2019,

date of current version December 31, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2961243

Opportunistic Networks Link Prediction Method

Based on Bayesian Recurrent Neural Network
YULIANG MA AND JIAN SHU
School of Software, Nanchang Hangkong University, Nanchang 330063, China
Corresponding author: Jian Shu (shujian@nchu.edu.cn)
This work was supported in part by the National Natural Science Foundation of China under Grant 61762065 and Grant 61962037, in part
by the Natural Science Foundation of Jiangxi Province under Grant 20171ACB20018 and Grant 20171BBH80022, and in part by the
Innovation Foundation for Postgraduate Student of Jiangxi Province under Grant YC2018093.

ABSTRACT The goal of link prediction is to estimate the possibility of future links among nodes using
known network information and the attributes of the nodes. According to the time-varying characteristics
and the node’s mobility of opportunistic networks, this paper proposes a novel link prediction method
based on the Bayesian recurrent neural network (BRNN-LP) framework. The time series data of a dynamic
opportunistic networks is sliced into snapshots in which there exist the correlation information and spatial
location information. A vector of a snapshot is constructed based on such information, which represents
the link information. Then, the vectors of multiple network snapshots constitute a spatiotemporal vector
sequence. Benefiting from the BRNN’s ability of extracting the features of time series data, the correlation
between spatiotemporal vector sequence and node connection states is learned, and the law of the link
evolution is captured to predict future links. The results on the MIT Reality dataset show that compared
with methods such as the similarity-based indices, the support vector classifier, linear discriminant analysis
and recurrent neural network, the proposed prediction method is more accurate and stable.

INDEX TERMS Link prediction, opportunistic networks, Bayesian recurrent neural network.

I. INTRODUCTION
Opportunistic network (ON) [1] is an ad hoc network based
on Mobile Ad Hoc Networks (MANETs) and Delay Tolerant
Networks (DTNs). ON utilizes node mobility and the ‘‘store-
carry-forward’’ routing pattern to realize network commu-
nications, and the network topology changes over time [2].
Figure 1 shows the process of sending messages from the
FIGURE 1. ON communication.
source node S to the destination node D. The network topol-
ogy of ON changes constantly. ON does not require the
full connectivity of the network and is more suitable for nodes by analysing the known link information, which
the actual requirements of a self-organizing network, which is helpful to obtain the evolutional law of the network
can meet the needs of network communications under bad structure [9], and provide decisions for ON routing and
conditions [3]. ON is one of the hot fields in network, and forwarding, so as to provide better services for the upper-layer
it is widely applied in many fields such as Location-based application.
Service [4], Media Services [5], Mobile Data Offloading [6], According to the time-varying characteristics and the
and Intelligent Traffic [7] and so on. nodes’ mobility of ON, a link prediction method based on
Among the various aspects, the opportunistic forwarding a Bayesian recurrent neural network (BRNN) [11], [12] is
mechanism is the first and foremost issue in ON field. Link proposed.
prediction can dig out the potential relationships between This paper makes two research contributions. Firstly,
it proposes a representation method for the link information in
The associate editor coordinating the review of this manuscript and ON. We construct a vector to represent the link information,
approving it for publication was Bilal Alatas . which includes the correlation information and the spatial

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
185786 VOLUME 7, 2019
Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
location information of node pairs. Secondly, this paper puts
forward a deep learning model, BRRN-LP, to predict the links
of ON. It converts prediction problems into classification
problems, and a better result of link prediction is achieved
by the proposed model.
The rest of the paper is organized as follows. Section 2 anal-
yses the related work. Section 3 describes the problem of
link prediction. Section 4 explains the representation methods
for link information. Section 5 introduces the construction
of the BRRN-LP model. The experiments and analyses are
described in Section 6. The last part is the conclusion.
II. RELATED WORK
The link prediction methods for ON can be grouped into
the following categories: link prediction methods based on
similarity indices, probabilistic models, machine learning and
time series.
A. SIMILARITY INDICES-BASED PREDICTION METHODS
Similarity indices-based prediction methods predict links by
constructing similarity indices which are used to describe the
similarity between nodes. The similarity between nodes is
characterized by attribute similarity and structure similarity.
Usually, structure similarity is employed to predict links,
because the attributes of node are difficult to obtain and not
accurate. Researchers have proposed around 30 traditional
algorithms [13] that are based on different network structures.
The structure similarity indices-based prediction is further
categorized as local, global, and quasi-local methods [14].
Local similarity methods use local information of an edge
to calculate the similarity matrix, such as the degree of a
node, common neighbour (CN) [15], etc. The CN similarity
index considers the impact of a one-hop neighbour on the
node pairs, and assumes that the more common neighbours
are, the more possibility of connection between nodes is, and
thus deriving a series of new similarity index, such as Salton,
Jaccard, preferential attachment (PA) [16], and Adamic-Adar
(AA) [17]. Shang et al. [18] points out that the performance
of these algorithms on highly sparse or treelike networks is
poor, and proposes heterogeneity index, homogeneity index
and heterogeneity adaption index for their heterogeneity.
Global similarity method uses total information on net-
work to calculate the similarity matrix between nodes.
Katz [19] takes into account all paths and distributes different
weights to paths with different lengths. The similarity indices
based on random walks uses the node random walk to estab-
lish the probability transfer matrix, so as to obtain the prob-
ability from one node going to another node, thus obtaining
the probability that the node pairs establish a connection.
Quasi-Local similarity Methods do not require global topo-
logical information but use more information than local
index. The local path (LP) [20] takes into account the contri-
butions of path information of third-order neighbour. In [21],
SimRank is defined in a self-consistent way, according to the
assumption that two nodes are similar if they are connected
to similar nodes.
VOLUME 7, 2019
A successful similarity indices-based prediction method
depends upon the exact representation of the network’s struc-
tural features. This method can be effective for networks with
static topology or topology changing slowly, but it has a poor
predictive effect on ON with sparse and topological changes
because it cannot depict the changes of ON [22].
B. PROBABILISTIC MODEL-BASED
PREDICTION METHODS
Probabilistic model-based prediction methods constructs a
probabilistic model that can describe the network structure
and the relational features primarily by optimizing of the
parameter. Wang et al. [23] design a local probability model
with two-hop neighbours, which utilize Markov random field
and undirected graphs to solve the link prediction problem in
large graphs. Das and Das [24] considered the impacts of fine-
grained and coarse-grained time on links, and established
a Markov prediction model based on time-varying graphs
of social networks, which can better predict the connection
of the node pairs in a dynamic network. Shalforoushan and
Jalali [25] use a bayesian network to learn the relation-
ships between variables, and propose friend recommendation
method based on Bayesian network for a social network.
Probabilistic model-based prediction methods describe the
entire network (including the node attribute information and
network structural information), which allows the method to
obtain more accurate prediction result. However, this method
is computationally complex. The model parameters need to
be adjusted according to the requirements of different net-
works, causing poor applicability.
C. MACHINE LEARNING AND TIME SERIES-BASED
PREDICTION METHODS
Machine learning based prediction methods predict links by
learning the more essential features of the network. Link
information is the input of the machine learning model.
Output is 0 or 1, where 0 means that the node pairs are not
connected, and 1 means that the node pairs are connected. The
commonly used supervision classification algorithms, such
as the support vector machine (SVM), kernel regression [26],
decision tree and linear discriminant analysis (LDA) [27], can
solve link prediction problems, and the SVM is better than the
most classification algorithms on the link prediction for Sina
Weibo [28]. Yang et al. [29] propose a link prediction method
based on clustering and decision trees. After clustering two
types of objects, the method constructs decision trees to solve
the link prediction problem of heterogeneous networks.
With further studies on link prediction for ON, schol-
ars have paid more attention to the regulation of network
changes, and predict links using time series. There are usually
three steps. First, the similarity indices is chosen. Second,
the time series method is employed to analyse the law of
the similarity indices changes. Finally, the predicted scores
of the future connections of the node pairs is obtained [30].
Güneş et al. [31] used time series prediction models to
weight different similarity indices (such as the CN, PA, AA,
185787
 Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN

and Jaccard), which showed that the traditional similarity

indices are suitable for link prediction in dynamic networks.
Wang et al. [32] proposed a time series link prediction method
based on evolution. In this method, the tensor decomposition
is used to find the time factor in the network, and the time
factor is analysed using time series to obtain the time factor
value of the next time. This allows the network structure of
the next time slice to be obtained and the link prediction
problem to be solved. Ozcan et al. proposed NARX neural
networks [33] and Multivariate Time Series Link Prediction
method [34] to solve the link prediction problem in het-
erogeneous dynamic networks. K-K Shang et al. proposed
the concepts of current or history network and the future
network to analyse the link evolution, then proposed the new FIGURE 2. Snapshot of ON.
link prediction metric and algorithms for evolving networks,
and they found that the direct link algorithm performs better
than the indirect link algorithm for a range of time-varying
networks [35]. Furthermore, they found that geographical
location and link weight both have significant influence on
the transport network [36].
For a dynamic ON with more sparse connections, it is
FIGURE 3. Connection diagram of the node pairs in time slices.
necessary for time series analysis by deep learning method.
Using convolutional neural networks (CNNs), [37] extracted
ON’s structural features from the evolutionary process of
the state diagram, and predicted the links between multi- the correlation information and location information within
ple nodes. Considering the historical connection informa- each slice is defined using the contact information and spatial
tion, [38] constructed a sequence vector, and extracted the information respectively. Finally, we construct the vector with
features of the changes of the historical information using correlation information and location information to represent
a recurrent neural network (RNN). Then, link prediction is the link information.
performed.
In this paper, we turn link prediction problems into clas- A. DATA CONVERSION
sification problems. We construct a vector to represent the This paper employs the time series analysis method to seg-
link information, which includes the correlation information ment the continuous data in the data set and divide ON into
and the spatial location information of node pairs. We also several continuous network snapshots.
construct the spatiotemporal vector sequence to represent Figure 3 shows the connection information of the node
historical link information. We propose a link prediction pairs (A, B) in a single network snapshot [TS , TN ]. T denotes
method based on the Bayesian recurrent neural network the time slice length. TS and TN represent the moments when
model (BRNN-LP) to extract the spatiotemporal features the time slice begin and end, respectively. T1S and T2S rep-
from the evolutionary process of ON. The method achieves resent the moments when the node pairs connect during the
a better link prediction. time slice. T1N and T2N represent the moments when the node
pairs disconnect during the time slice. tcon1 and tcon2 represent
III. PROBLEM DESCRIPTION the duration of connection between two nodes. If the time
The topology of ON changes over time. The network topol- slice length T is too large or too small, the information may
ogy is divided into several continuous network snapshots. be blurry or insufficient. Thus, the link prediction results
As shown in Figure 2, the network snapshot set is G = will be affected according to their different values. Later,
{G1 , G2 , . . . , Gt , . . . , GN }, where Gt = (Vt , Et ), Gt repre- we determine the value of T by experiments.
sents the network snapshot at t − th moment, and Vt and
Et respectively represent the sets of nodes and edges in the B. INFORMATION REPRESENTATION
network snapshot at t − th moment. The link prediction in The generation of links depends on the encounter chance that
ON is to predict connection state in the (t + 1) − th network is determined by two factors: the location and the reason
snapshot according to the historical link information of the for encounter. The location factors are influenced by the
node pairs in the previous t network snapshots [38]. habits of the nodes’ movement. The reason for the encounter
seems to be random, but it is more likely that there is some
IV. LINK INFORMATION REPRESENTATION correlation between the nodes. Thus, this paper represents the
The ON dataset are stored in chronological order. Firstly, ON link information with respect to the correlation information
is sliced into several continuous network snapshots. Then and spatial location information.

185788 VOLUME 7, 2019

Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
1) REPRESENTATION OF THE CORRELATION INFORMATION
In ON, there are often potential correlations between nodes,
such as social relationships among people in social networks
and race relationships among animals in animal tracking
networks. However, reliable data are often difficult to obtain.
The contact information of nodes can reflect their correlation.
The contact information includes the frequency and duration
of the connection [39]. The bigger the frequency and duration
of connections are, the stronger the correlation between the
nodes is. We believe that the correlation of the nodes is
proportional to the frequency and duration of connections.
Thus, we define the correlation of the node pairs at t − th
time slice as wt , as shown in (1).
 tion states of the node pairs at the next moment. We construct
L the spatiotemporal vector sequence, which consists of vectors
L
 X
etconi , i ∈ (1, 2, . . . , L)

wt = T (1)
 i=1

 0, (L = 0)
where T denotes the time slice length, L denotes the number
of links that are generated by node pairs, tconi denotes the
duration corresponding to the i − th connection, and L/T
denotes the connection frequency of the node pairs within
the time slice. The more connections that exist in a time slice,
the bigger the connection frequency is. The exponential func-
tion etconi represents that the duration impacts on correlation.
When wt is 0, there is no connection between nodes in the
time slice.
2) REPRESENTATION OF THE SPATIAL
LOCATION INFORMATION
The spatial location information of the node pairs is the
location information when the nodes meet. This paper marks
the location of the node pairs using the communication area.
If two nodes are in the same communication area within a
certain period of time, it is more likely that two nodes will
connect [40]. M is the set of whole communication area
marks to indicate the locations where the link was generated
in the network, where M = {x|x ∈ Z +}. The spatial location
information metric of the node pairs at t − th time slice is the
location pt , as shown in (2).
X j
 2, m⊂M
pt = j∈m (2)

0, m=∅
where m denotes the subset of M to indicate the locations
where the link was generated in the time slice, j is the element
in the m. If the j is repeated, it is only calculated once. If the
pt value is 0, there is no connection between nodes in the time
slice.
3) REPRESENTATION OF THE LINK INFORMATION
Comprehensively considering the correlation information
and location information, we construct a vector VEt to repre-
sent link information at t − th time slice, as shown in (3).
VEt = (wt , pt ) (3)
VOLUME 7, 2019
FIGURE 4. Input sample set composition process of the BRNN-LP model.
V. THE CONSTRUCTION OF PREDICTION MODEL
To learn the deep relationship between the node connection
states and link information, BRNN is employed to capture the
spatiotemporal features of the connections between nodes in
ON. The trained link prediction model can predict the connec-
in several time slices, and take it as the input of the link
prediction model BRNN-LP. The output of the model is the
connection states of the node pairs at the next moment. The
BRNN-LP model is constructed with the following sections:
network structure, hyper-parameter setting, and training opti-
mization algorithm.
A. INPUT AND OUTPUT OF THE MODEL
We construct the spatiotemporal vector sequence and take
it as the input of the BRNN-LP model. The spatiotemporal
vector sequence consists of vectors of several time slices,
where the length of the spatiotemporal vector sequence is the
prediction window size. The vector sequence is defined in (4).
Seq = [VE1 , VE2 , . . . , VEn ] (4)
In order to reduce computational complexity, we don’t use all
historical link information. It is commonly assumed that the
present connection state is only related to link information
of previous n time slices. Selecting an appropriate input
sequence length is helpful for feature extraction and link
prediction. We determine the optimal value of the length of
the vector sequence by experiments.
Assuming that the length of the space-time vector sequence
is 3, construction process of the input sample set for
BRNN-LP model is shown in Figure 4.
We turn link prediction problems into classification prob-
lems. The input of the BRNN-LP is spatiotemporal vector
sequence in a link prediction window. The model’s output of
the node pairs at the next moment is 0 or 1. 0 means that the
node pairs are not connected, and 1 means that the node pairs
are connected.
B. NETWORK STRUCTURE
Recurrent neural network (RNN) [41] are a kind of neural
network, which can be used to analyse time series data.
BRNN uses Bayesian method to express the uncertainty of
neural network parameters. These parameters are treated as
random variables. Furthermore, the posterior distribution of
the parameters based on the given data can be inferred, so that
the BRNN model does not need much data training which is
suitable for solving classification problems with smaller and
185789

FIGURE 5. Network structure.
sparser samples. We use the model BRNN trained by vector
sequence to extract potential features of the link information
changes of the node pairs over time. The construction process
of the BRNN for the link prediction classification task is
shown in formulas (5), (6) and (7).
θ ∼ N (0, I ) (5)
π = fNN (Seq; θ ) (6)
y ∼ Cat(logistic(π )) (7)
where θ is the weight of the neural network fNN , which is
the potential feature between nodes, π is the non-normalized
log probability for predicting all classes, which indicates the
connection probability of the node pairs, y is a category label
indicating whether the node pairs generates a connection at
the next moment. Cat denotes the classification distribution.
When fNN is a recurrent neural network, the above process
describes a BRNN.
Long short-term memory (LSTM) improves the traditional
RNN through the ‘‘gate’’ limit mechanism, which can memo-
rize useful information and solve the problem of gradient van-
ishing. The goal of Bayesian LSTM is to obtain the posterior
distribution of weights, which reduces the training parameters
and shortens the training time. The BRNN model is shown
in Figure 5 [42].
(t )
We use the current input, vector Vi i , of BRNN and the
(ti −1)
potential spatiotemporal feature hi−1 at the last moment
(t )
to generate the potential spatiotemporal features hi i at the
(t )
current time, and hi i is taken as the input of BRNN for the
next moment. Bayesian LSTM uses input gates to control
the inflow of spatiotemporal features at the last moment and
updates the implicit spatiotemporal features of the previous
moment using the forget gate. The implicit spatiotemporal
features are obtained from the output gate. The input of
Bayesian LSTM is the vector sequence, and the output layer
is a classifier using a Logistic Regression, the result of which
is whether the node pairs’ is connected or not.
C. HYPER-PARAMETER SETTING
We adjust the hyper-parameters of the model when training
the deep learning model. Due to there is no good theoret-
ical support, it is often difficult to determine the optimal
185790
Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
hyper-parameters. The construction of the BRNN in this
paper involves the following hyper-parameters: the activation
function of the hidden layer, the initialization weight and the
length of the input sequence.
1) THE ACTIVATION FUNCTION OF THE HIDDEN LAYER
The performance of the BRNN-LP model is affected by the
activation function. To ensure successful training of the deep
network structures, one must select a reasonable activation
function. For example, the sigmoid function has double-end
saturation characteristics, which makes it subject to the ‘‘van-
ishing gradient’’ problem when training deep networks, and
the convergence rate of the model becomes very slow. In view
of this issue, the rectified linear unit (ReLU) was proposed by
some scholars. The ReLU is a linear activation function with
the characteristics of simple calculations, high efficiency, and
no gradient saturation. Therefore, we use the ReLU activation
function to accelerate network training.
2) THE INITIALIZATION WEIGHT
Too bigger or too small initial weights in the RNN may
lead to gradient exploding or gradient vanishing in back
propagation. The BRNN learns the distribution of weights
from the training samples using Bayesian backpropagation,
which avoids the gradient vanishing and gradient exploding
problems. Thus, the initial weight values are generated by
randomly sampling.
3) THE LENGTH OF THE INPUT SEQUENCE
The input of the BRNN-LP model is a spatiotemporal vector
sequence Seq, and the length of sequence is n which indicates
the link information of the n network snapshots. We use
Seq to predict the connection status of the node pairs in
the (n + 1) − th network snapshot. If the length of the input
sequence is too long, it will make the BRNN-LP model have
too many features at one input, leading to increasing the diffi-
culty of model training. If the length of the input sequence is
too short, the BRNN-LP model cannot learn feature of link
evolution. In this paper, an appropriate length of the input
sequence is determined contrast experiments.
D. OPTIMIZATION ALGORITHM FOR TRAINING
This paper chooses Bayes by backpropagation (BBB) to learn
the weights of the BRNN. BBB is a variational inference
scheme for learning the posterior distribution of the weights
θ of neural networks. The posterior distribution generally
follows the Gaussian distribution with a mean of µ and a
standard deviation of σ , and it is denoted as N(θ |µ, σ 2 ).
p(y|θ, Seq) is the log-likelihood of the model. Then we train
the network by minimizing the variational free energy which
is shown in (8).

q(θ )

8(θ) = Eq(θ) (8)
p(y|θ, Seq)p(θ )
where p(θ ) is the prior probability of the weight.
VOLUME 7, 2019
Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
TABLE 1. Information of dataset.
The variational free energy minimization is equivalent to
the log likelihood p(y|θ, Seq) maximization. Equation (8) can
be converted to equation (9).
8(θ ) = −Eq(θ ) p(y|θ, Seq) + KL [q(θ )||p(θ )]
 
(9)
In the case of a zero θ , the KL term can be seen as a form
of weight decay on the mean parameters, where the rate of
weight decay is automatically tuned by the standard deviation
parameters of the prior and posterior [12].
VI. EXPERIMENTS AND ANALYSIS
We select the MIT Reality dataset for the experiments.
Different comparison experiments are designed to verify the
effectiveness of the proposed link prediction method. The
evaluation indices are AUC, Accuracy and Precision.
A. EXPERIMENTAL DATASET
The parameters of the MIT Reality dataset are shown
in Table 1.
The MIT Reality dataset records the communications of
Bluetooth cell phones that are carried by 97 MIT experi-
menters over 246 days. This dataset includes cell phone com-
munication records and Bluetooth connection records. This
dataset also has recorded the location information and activity
information of the experimenters. The sampling period is
300s.
B. EVALUATION INDICES
In this paper, the area under the receiver operating character-
istic curve (AUC), accuracy and precision are used to evaluate
the effect of the BRNN-LP model.
1) EVALUATION INDICES OF AUC
The AUC evaluates the algorithm’s performance according to
the whole list. Then, at each time we randomly pick a missing
link and a non-existent link to compare their scores. The AUC
takes a test edge from the test set and a non-existent edge
from the non-existing edge set [13]. T he AUC is defined as
equation (10).
n0 + 0.5n"
AUC = (10)
n
where n is the number of times of independent comparisons,
there are n0 times the missing link having a higher score and
n" times they have the same score.
VOLUME 7, 2019
2) EVALUATION INDICES OF ACCURACY
Accuracy assesses whether the prediction results are accurate
using the ratio of the number of correctly predicted linked and
unlinked to the total number of predictions. The accuracy is
defined as equation (11).
TP + TN
Accuracy = (11)
P+N
where TP is the number of correctly predicted links, TN is
the number of correctly predicted disconnections, P is the
total number of predicted links and N is the total number of
predicted disconnections.
3) EVALUATION INDICES OF PRECISION
The Precision only focuses on the L links with the top
ranks or the highest scores [13]. The precision is defined as
equation (12).
m
precision = (12)
L
Here, m is the number of correctly predicted links. When
L does not change, we can use a prediction method with
comparatively better precision. Then, the larger m is, the more
probable that a correct link can be found.
C. EXPERIMENTAL RESULTS AND ANALYSIS
This experiment is divided into two steps. The first step is
to compare the results of BRNN with different parameters
using the validation set and test set, and to choose the optimal
parameters for BRNN. The second step is to compare BRNN
with other link prediction methods, and to verify the validity
of the proposed model by results. The proportion of the
training set to the testing set for all experiments was 4:1.
1) PERFORMANCE OF THE BRNN-LP MODEL UNDER
DIFFERENT PARAMETERS
The experiments in this section verify the influences of dif-
ferent parameters on our method. By experiments of differ-
ent the number of training iterations, time slice length, and
input sequence lengths for BRNN-LP, the optimal number
of iterations, time slice length and sequence length can be
determined.
a: THE EFFECT OF THE TRAINING ITERATIONS
The number of iterations will affect the generation of the
model. Too many or too few iterations will lead to over-
fitting or under-fitting. Hence, we use different numbers of
iterations of 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100 in
experiments to choose optimum number of iterations. The
loss value and accuracy curve of the objective function are
shown in Figure 6 and Figure 7.
Figure 6 and figure 7 show that the connection of MIT
data set is relatively sparse, the accuracy at the beginning
is relatively high. When the number of iterations is less
than 10, the loss value of the BRNN-LP model is reduced
dramatically, and the accuracy is increased slowly. When the
185791
 Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN

FIGURE 6. The loss value curve of the training process. FIGURE 8. The accuracy of different time slice lengths.

FIGURE 7. The accuracy curve of the training process. FIGURE 9. The precision of different time slice lengths.

TABLE 2. Prediction results for different time slice lengths.

number of iterations increased from 10 to 40, the loss value
of the BRNN-LP model is gradually reduced, and the accu-
racy is increased slowly. This phenomenon indicates that
the BRNN-LP model is under-fitting. When the number of
iterations increased from 40 to 90, the loss value does not
change much, and the change of accuracy is not obvious.
When the number of iterations is more than 90, the loss
value of the BRNN-LP model begins to fluctuate up and
down slightly and its accuracy of the model is basically the
same as the number of iterations at 40, which reflect that the
BRNN-LP model is over-fitting. The analysis of the above
experimental results show that the BRNN model has strong
feature extraction ability and can obtain better results when
the number of iterations is approximately 40. Considering the and average precision of the BRNN-LP model are the best,
performance and training time of model, the optimal number comparing to other slice lengths. When the time slice length
of iterations of the BRNN-LP model is 40. is too short, the information in the slice will be less, and the
model cannot further extract the features of the link infor-
b: THE EFFECT OF THE TIME SLICE LENGTH mation. When the slice length is too long, the information in
The length of the time slice has a great impact on the the slice will be too much, and the model feature extraction
BRNN-LP model. We use different time slice lengths includ- ability will be affected. Therefore, 360s is employed as the
ing 180s, 240s, 300s, 360s, 480s and 600s [38], to obtain the time slice length in this paper.
most suitable time slice length for this model on the MIT
dataset. Then, we randomly sample 30 times to assess the c: THE EFFECT OF THE INPUT SEQUENCE LENGTH
accuracy and precision of the prediction model for each time From the selected time slice lengths, it can be seen that the
slice length. The results are shown in Figure 8 and Figure 9, amount of link information will affect the performance of
and the average values are shown in Table 2. the BRNN-LP model. If the length of the input sequence is
It can be seen from TABLE 2 that when the slice length is too long, the amount of information and training complexity
360s with 30 times random sampling, the average accuracy will increase, which will lead to model over-fitting. If the

185792 VOLUME 7, 2019

Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
FIGURE 10. The accuracy of different sequence lengths.
FIGURE 11. The precision of different sequence lengths.
length of the input sequence is too short, the amount of
information will be insufficient, which will lead to under-
fitting. This paper sets different input sequence lengths of 1,
2, 3, 4, 5 and 6. We randomly extract 30 samples to calculate
the accuracy and precision of the prediction of the model
under different sequence lengths. The results are shown in the
Figures 10 and 11. The average values are shown in Table 3.
The experimental results show that the length of the input
sequence has a certain impact on the model predictive per-
formance. When the length of the input sequence is 2, its
predictive performance is best, and the average prediction
accuracy and precision are 99.45% and 99.79%, respectively,
which represents the BRNN-LP model can better capture the
hidden features of links. Thus, the optimal length of the input
sequence of the BRNN-LP model is 2.
2) PERFORMANCE COMPARISON OF DIFFERENT
PREDICTION METHODS
The experiments compare the performance of BRNN-LP with
ones of other prediction models. Firstly, we compare the link
prediction methods of the BRNN-LP model with the similar-
ity indices (CN, AA, RA, PA, Katz and LP). Then, this model
is compared with the link prediction model (LDA) [18],
and the link prediction model based on the support vector
classifier (SVC). Finally, this model is compared with the link
prediction model (RNN) [23].
VOLUME 7, 2019
TABLE 3. Prediction results for different sequence lengths.
FIGURE 12. The comparison experiment with link prediction methods
based on similarity indices.
a: COMPARISON WITH PREDICTION METHODS BASED ON
SIMILARITY INDEX
On the condition that the slice length is 360s, the number of
BRNN-LP iterations is 40, and the input sequence length is
2, we compare with similarity indices based link prediction
methods. The samples are chosen from the test sample set,
and the sample sizes are 50, 100, 150, 200, and 250. Five
total rounds of tests are conducted. The AUC is used to
evaluate the performance of the link prediction methods. The
experimental results are shown in Figure 12.
The experimental results show that the prediction effect
of the similarity indices is affected by the number of test
samples. Since the information that is obtained from the path-
based similarity indices is more comprehensive, the path-
based similarity indices (Katz and LP) perform better than
the structure-based similarity indices (AA, RR and CN).
Furthermore, the prediction performance of the above sim-
ilarity indices is not stable. The performance of the proposed
BRRN-LP model is comparatively stable and better than that
of the above similarity indices.
b: COMPARISON WITH PREDICTION METHODS BASED ON
MACHINE LEARNING
On the condition that the slice length is 360s, the number
of BRNN-LP iterations is 40, and the input sequence length
is 2, we compare with the machine learning based link predic-
tion methods. All the experiments were conducted 30 times.
185793

FIGURE 13. The comparison experiment with link prediction methods
based on machine learning (AUC).
FIGURE 14. The comparison experiment with link prediction methods
based on machine learning (Precision).
The test samples are taken from the test set, and the number
of samples is 200. AUC and precision are employed to com-
prehensively evaluate the performance of the link prediction
methods. The experimental results are shown in Figure 13 and
Figure 14, respectively.
The experimental results show that the SVC-based method
is not ideal because it cannot extract deep features of the links
in ON. The LDA-based approach portray link information,
and its performance is better than those of the SVC-based
methods. However, it can be seen that the SVC and LDA are
not as good as the RNN and BRNN prediction methods which
consider the spatiotemporal features of ON. The AUC and
precision for the RNN and BRNN-LP methods are all above
0.98. It can be seen From Figure 14 that the precision of the
BRNN prediction method is 0.9972, which is better than the
RNN. This results show that the proposed BRNN-LP model
can better characterize the features of ON and then make more
effective link predictions for ON.
The above experiments show that the precision, AUC,
and accuracy of the proposed BRNN-LP model are better
compared with link prediction methods based on similarity
indices. In addition, compared with the machine learning-
based link prediction models (SVC, LDA) and deep learning
method RNN, the prediction performance of the BRNN-LP
model is netter as well.
VII. CONCLUSION
This paper proposes a BRNN-LP method for ON. We con-
struct a vector to represent the link information after
185794
Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN
considering the correlation information and the spatial loca-
tion information of node pairs. Taking BRNN advantages
of extracting the spatiotemporal features in the sparse net-
work, a better prediction performance is achieved. In the MIT
dataset, the experimental results show that the BRNN-LP
model achieves better predictive precision and accuracy com-
pared with existing methods of link prediction including the
CN, AA, RA, LP, Katz, SVC, LDA, and RNN.
The method proposed in this paper is to predict link for
single node pair. We will further study the link prediction
method of multi-node pairs in the future.
REFERENCES
[1] P. Jadhav and R. Satao, ‘‘A survey on opportunistic routing protocols for
wireless sensor networks,’’ Procedia Comput. Sci., vol. 79, pp. 603–609,
2016.
[2] C. Y. Aung, I. W.-H. Ho, and P. H. J. Chong, ‘‘Store-carry-cooperative
forward routing with information epidemics control for data delivery in
opportunistic networks,’’ IEEE Access, vol. 5, pp. 6608–6625, 2017.
[3] E. K. Wang, Y. Li, and Y. Ye, ‘‘A dynamic trust framework for opportunistic
mobile social networks,’’ IEEE Trans. Netw. Service Manag., vol. 15, no. 1,
pp. 319–329, Nov. 2017.
[4] H. T. Cheng, F. T. Sun, and S. Buthpitiya, ‘‘Sensorchestra: Collabo-
rative sensing for symbolic location recognition,’’ in Proc. Int. Conf.
Mobile Comput., Appl., Services. Berlin, Germany: Springer, Oct. 2010,
pp. 195–210.
[5] E. Toledano, D. Sawada, and A. Lippman, ‘‘CoCam: A collaborative
content sharing framework based on opportunistic P2P networking,’’ in
Proc. IEEE 10th Consum. Commun. Netw. Conf. (CCNC), Las vegas,
Nevada, USA, Jan. 2013, pp. 158–163.
[6] B. Han, P. Hui, V. S. A. Kumar, M. V. Marathe, J. Shao, and
A. Srinivasan, ‘‘Mobile data offloading through opportunistic communi-
cations and social participation,’’ IEEE Trans. Mobile Comput., vol. 11,
no. 5, pp. 821–834, May 2012.
[7] E. Koukoumidis, L. S. Peh, and M. R. Martonosi, ‘‘Signal guru: Leveraging
mobile phones for collaborative traffic signal schedule advisory,’’ in Proc.
9th Int. Conf. Mobile Syst., Appl., Services, Bethesda, MA, USA, Jun. 2011,
pp. 127–140.
[8] H. D. Ma, P. Y. Yuan, and D. Zhao, ‘‘Research progress on routing problem
in mobile opportunistic networks,’’ J. Softw., vol. 26, no. 3, pp. 600–634,
Mar. 2015.
[9] L. Lue, L. Pan, T. Zhou, Y.-C. Zhang, and H. E. Stanley, ‘‘Toward link
predictability of complex networks,’’ Proc. Nat. Acad. Sci. USA, vol. 112,
pp. 2325–2330, Feb. 2015.
[10] C. Dai, L. Chen, and B. Li, ‘‘Link prediction based on sampling in complex
networks,’’ Appl. Intell., vol. 47, no. 1, pp. 1–12, Jul. 2017.
[11] H. Wang and D.-Y. Yeung, ‘‘Towards Bayesian deep learning: A frame-
work and some existing methods,’’ IEEE Trans. Knowl. Data Eng., vol. 28,
no. 12, pp. 3395–3408, Dec. 2016.
[12] M. Fortunato, C. Blundell, O. Vinyals, ‘‘Bayesian recurrent neu-
ral networks,’’ Apr. 2017, arXiv:1704.02798. [Online]. Available:
https://arxiv.org/abs/1704.02798
[13] L. Lü and T. Zhou, ‘‘Link prediction in complex networks: A survey,’’
Phys. A, Stat. Mech. Appl., vol. 390, no. 6, pp. 1150–1170, 2011.
[14] B. Pandey, P. K. Bhanodia, and A. Khamparia, ‘‘A comprehensive survey
of edge prediction in social networks: Techniques, parameters and chal-
lenges,’’ Expert Syst. Appl., vol. 124, pp. 164–181, Jun. 2019.
[15] D. Liben Nowell and J. Kleinberg, ‘‘The link prediction problem for social
networks,’’ J. Amer. Soc. Inf. Sci. Technol., vol. 58, no. 7, pp. 1013–1019,
Feb. 2007.
[16] M. E. J. Newman, ‘‘Clustering and preferential attachment in growing
networks,’’ Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat. Interdiscip.
Top., vol. 64, no. 2, 2001, Art. no. 025102.
[17] L. A. Adamic and E. Adar, ‘‘Friends and neighbors on the Web,’’ Soc.
Netw., vol. 25, no. 3, pp. 211–230, 2003.
[18] K.-K. Shang, ‘‘Link prediction for tree-like networks,’’ Chaos, An Inter-
discipl. J. Nonlinear Sci., vol. 29, no. 6, Apr. 2019, Art. no. 061103.
[19] L. Katz, ‘‘A new status index derived from sociometric analysis,’’ Psy-
chometrika, vol. 18, no. 1, pp. 39–43, 1953.
VOLUME 7, 2019
Y. Ma, J. Shu: ONs Link Prediction Method Based on BRNN

[20] T. Zhou, L. Lü, and Y.-C. Zhang, ‘‘Predicting missing links via local [37] J. Shu, X. Zhang, L. Zhiyong, and L. Yang, ‘‘Multi-nodes link predic-
information,’’ Eur. Phys. J. B, vol. 71, no. 4, pp. 623–630, Oct. 2009. tion method based on deep convolution neural networks,’’ Acta Electron.
[21] G. Jeh and J. Widom, ‘‘SimRank: A measure of structural-context similar- Sinica, vol. 46, no. 12, pp. 2970–2977, Dec. 2018.
ity,’’ in Proc. 8th Acm Sigkdd Int. Conf. Knowl. Discovery Data Mining, [38] X. Cai, J. Shu, and M. AlkAli, ‘‘Link prediction approach for opportunis-
New York, NY, USA, Jul. 2002, pp. 271–279. tic networks based on recurrent neural network,’’ IEEE Access, vol. 7,
[22] C. Ma, Z. K. Bao, and H. F. Zhang, ‘‘Improving link prediction in pp. 2017–2025, 2018.
complex networks by adaptively exploiting multiple structural features of [39] E. Bulut and B. K. Szymanski, ‘‘Exploiting friendship relations for effi-
networks,’’ Phys. Lett. A, vol. 381, no. 39, pp. 3369–3376, Oct. 2017. cient routing in mobile social networks,’’ IEEE Trans. Parallel Distrib.
[23] C. Wang, V. Satuluri, and S. Parthasarathy, ‘‘Local probabilistic models for Syst., vol. 23, no. 12, pp. 2254–2265, Dec. 2012.
link prediction,’’ in Proc. 7th IEEE Int. Conf. Data Mining (ICD), Omaha, [40] B. Burns, O. Brock, and B. Levine, ‘‘MV routing and capacity building
NE, USA, Oct. 2007, pp. 322–331. in disruption tolerant networks,’’ in Proc. IEEE INFOCOM, Miami, FL,
[24] S. Das and S. K. Das, ‘‘A probabilistic link prediction model in time- USA, Mar. 2005, pp. 398–408.
varying social networks,’’ in Proc. IEEE Int. Conf. Commun. (ICC), Paris, [41] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521,
France, May 2017, pp. 1–6. no. 7553, p. 436, 2015.
[25] S. H. Shalforoushan and M. Jalali, ‘‘Link prediction in social networks [42] J. Shi, J. Chen, J. Zhu, S. Sun, Y. Luo, Y. Gu, and Y. Zhou, ‘‘Zhusuan:
using Bayesian networks,’’ in Proc. Int. Symp. Artif. Intell. Signal Process. A library for Bayesian deep learning,’’ Sep. 2017, arXiv:1709.05870.
(AISP), Mashhad, Iran, Mar. 2015, pp. 246–250. [Online]. Available: https://arxiv.org/abs/1709.05870
[26] D. Huang, S. Zhang, and P. Hui, ‘‘Link pattern prediction in opportunistic
networks with kernel regression,’’ in Proc. 7th Int. Conf. Commun. Syst.
Netw. (COMSNETS), Chengdu, China, Jun. 2015, pp. 1–8.
[27] F. Aghabozorgi and M. R. Khayyambashi, ‘‘A new similarity measure for
link prediction based on local structures in social networks,’’ Phys. A, Stat.
Mech. Appl., vol. 501, pp. 12–23, Jul. 2018.
[28] Y. Li, K. Niu, and B. Tian, ‘‘Link prediction in sina microblog using YULIANG MA was born in Jiujiang, Jiangxi,
comprehensive features and improved SVM algorithm,’’ in Proc. IEEE 3rd in 1995. He is currently pursuing the master’s
Int. Conf. Cloud Comput. Intell. Syst. (CCIS), Shenzhen, China, Nov. 2014, degree with Nanchang Hangkong University.
pp. 18–22. His main research interest includes opportunity
[29] N. Yang, T. Peng, and L. Liu, ‘‘Link prediction method based on clustering networks.
and decision tree,’’ J. Comput. Res. Develop., vol. 54, no. 8, pp. 1795–1803,
Aug. 2017.
[30] Y. Li, X. Zhao, and H. Tang, ‘‘Dynamics nature and link prediction methods
in opportunistic networks,’’ in Proc. IEEE 17th Int. Conf. Comput. Sci. Eng.
(CSE), Chengdu, China, Dec. 2014, pp. 1362–1368.
[31] I. Güneş, S. Gündüz-Ögüdücü, and Z. Çataltepe, ‘‘Link prediction using
time series of neighborhood-based node similarity scores,’’ Data Mining
Knowl. Discovery, vol. 30, no. 1, pp. 147–180, Jan. 2016.
[32] S.-H. Wang, H.-T. Yu, R.-Y. Huang, and Q.-Q. Ma, ‘‘A temporal link
prediction method based on motif evolution,’’ Acta Automatica Sinica,
vol. 45, no. 5, pp. 735–745, May 2016.
[33] A. Ozcan and S. G. Oguducu, ‘‘Link prediction in evolving heterogeneous
networks using the NARX neural networks,’’ Knowl. Inf. Syst., vol. 55,
JIAN SHU was born in 1964. He received
no. 2, pp. 333–360, 2018.
[34] A. Ozcan and S. G. Oguducu, ‘‘Multivariate time series link prediction the M.Sc. degree in computer networks from
for evolving heterogeneous network,’’ Int. J. Inf. Technol. Decis. Making, Northwestern Polytechnical University. He is cur-
vol. 18, no. 1, pp. 241–286, Nov./Jan. 2019. rently a Professor with Nanchang Hangkong
[35] K.-K. Shang, M. Small, X.-K. Xu, and W.-S. Yan, ‘‘The role of direct University. His main research interests include
links for link prediction in evolving networks,’’ Europhys. Lett., vol. 117, wireless sensor networks, embedded systems, and
Jan. 2017, Art. no. 28002. software engineering. He is a Senior Member of
[36] K. K. Shang, W. S. Yan, and M. Small, ‘‘Evolving networks using past CCF.
structure to predict the future,’’ Phys. A, Stat. Mech. Appl., vol. 455,
pp. 120–135, Aug. 2016.

VOLUME 7, 2019 185795