gcn 与gan1
需积分: 0 194 浏览量
2022-08-03
13:26:04
上传
评论
收藏 408KB PDF 举报
arXiv:1901.09165v1 [cs.SI] 26 Jan 2019
GCN-GAN: A Non-linear Temporal Link Prediction
Model for Weighted Dynamic Networks
Kai Lei
†,§
, Meng Qin
†
, Bo Bai
‡,*
, Gong Zhang
‡
, Min Yang
¶,*
†
ICNLAB, School of Electronics and Computer Engineering (SECE), Peking University, Shenzhen, Chin a
§
PCL Research Cente r of Networks and Communications, Peng Cheng Laboratory, Shenzhe n, China
‡
Future Network Theory Lab, 2012 Labs, Huawei Technolo gies, Co. Ltd., Hong Kong, Ch ina
¶
Shenzhen Institutes of Advanced Technolo gy, Chinese Academy of Sciences, She nzhen, China
†
leik@pkusz.edu.cn,
†
mengqin
az@foxmail.com,
‡
{baibo8, nicholas.zhang}@huawei.co m,
¶
min.yang@siat.ac.cn
*
Correspon ding Autho rs
Abstract—In this paper, we generally formulate the dynam-
ics prediction problem of various network systems (e.g., the
prediction of mobi lity, traffic and topology) as the temporal
link prediction task. Different from conventional techniques of
temporal link prediction that ignore the potential non-li near
characteristics and the informative link weights in the dynamic
network, we int roduce a novel non-linear model GCN-GAN to
tackle the challenging temporal link prediction task of weighted
dynamic networks. The proposed model leverages the benefits of
the graph convolutional network (GCN), long short-term memory
(LSTM) as well as the generative adversarial network (GAN).
Thus, the dynamics, topology st ructure and evolutionary patterns
of weighted dynamic networks can be fully exploited to improve
the temporal link prediction performance. Concretely, we first
utilize GCN to explore the local top ological characteristics of
each single snapshot and then emp loy LSTM to characterize
the evolving features of the dyn amic networks. Moreover, GAN
is used t o enhance the ability of t he model to generate the
next weighted network snapshot, which can effectively tackle the
sparsity and the wide-value-range problem of edge weights in
real-life dynamic networks. To verify the model’s effectiveness,
we conduct extensive experiments on four datasets of different
network systems and application scenarios. The experimental
results demonstrate that our model achieves impressive results
compared to the state-of-the-art competitors.
Index Terms—Temporal Link Prediction, Weighted Dynamic
Networks, Generative Adversarial Networks, Graph Convolu-
tional Networks
I. INTRODUCTION
Dynamics is a significant factor tha t hinders the perfor-
mance of most network systems. The prediction of mobility,
traffic and topology has been co nsidered as an effective
technique to cope with the pro blem. For instance, the dy-
namics of communication links in ad hoc networks makes
the design of routing protocol a challenging problem, where
the p rediction o f the d ynamic topology plays an important
role to achieve a more efficient and reliab le communication
[1]. In data center networks, traffic prediction technique could
be utilized to effectively schedule the highly parallel network
flows while avoiding the performance degradation due to
resource shortages [2]. For cellular networks, the prediction of
users’ locations can help to redu ce the resource consumption
(e.g., bandwidth) and ac hieve better Quality of Services (Qo S)
[3]. In a word, if the dyn amics of the network system can be
accurately predicted, the key resources can be effectively pre-
allocated to ensure the sy stem’s high performance.
Although numerous studies have been developed to deal
with the dynamics of network systems [2]–[5], m ost of them
only fo c us on a spec ific application scenario (e .g., flow pre-
diction in da ta center networks), failing to be generalized to
other different scenarios.
In fact, the dynamics pred iction problem of various network
systems can be generally fo rmulated as the temporal link
prediction task, wh ere the system’s behavior is described by
an abstracted dynamic graph. For example, one can model
each host in a data center as a node (entity), and the dynamic
traffic between a pair of hosts can be regarded as the changed
weighted link (relation) between the corr esponding entity pair.
Given the graph snapshots of previous time slices, the temporal
link pr ediction task tries to construct the graph top ology in the
next time slice [6].
Several recent techniques have been proposed to predict the
temporal links in dynamic graphs from different perspectives
[2]–[5]. Despite th e ir e ffectiveness, we argue that temporal link
prediction remains a challenging task for two primary reasons.
First, to the best of ou r knowledge, most of the existing
approa c hes merely consider the link prediction in unweighted
networks, determining the existence and absence of a link
between a certain no de pair. However, the link weights are
essential in real n etworks, wh ic h bring significant information
about the network’s behavior. For example, the link weights
may contain some useful information about delay, flow, signal
strength or distanc e of the network systems. Under such
circumstance, the temporal link prediction technique should
not only determine the existence of links but also consider the
correspo nding weights, which is a more challen ging problem
that most of the conventional methods cannot tackle.
Second, non- linear transformation s over time are commonly
observed in dynamic networks since the formation process of
most networks is complica ted and highly non-linear [7]. How-
ever, conventional method s are almost based on typical linear
models (e.g., non-negative matrix factorization (NMF) [8]),
ignorin g the potential non-linear characteristics of dynamic
networks. These linear models may have limited perfo rmances
for so me network inference tasks, including the temporal
评论0