Link prediction model based on dynamic network representation

Zhong-Ming Han; Sheng-Nan Li; Chen-Ye Zheng; Da-Gao Duan; Wei-Jie Yang

doi:10.7498/aps.69.20191162

Journals >Acta Physica Sinica >Volume 69 >Issue 16 >Page 168901-1 > Article

Acta Physica Sinica
Vol. 69, Issue 16, 168901-1 (2020)

Link prediction model based on dynamic network representation

Zhong-Ming Han^1、2、*, Sheng-Nan Li¹, Chen-Ye Zheng¹, Da-Gao Duan¹, and Wei-Jie Yang¹

Author Affiliations

¹College of Computer and Information Engineering, Beijing Technology and Business University, Beijing 100048, China

²Beijing Key Laboratory of Food Safety Big Data Technology, Beijing Technology and Business University, Beijing 100048, China

show less

DOI: 10.7498/aps.69.20191162 Cite this Article

Zhong-Ming Han, Sheng-Nan Li, Chen-Ye Zheng, Da-Gao Duan, Wei-Jie Yang. Link prediction model based on dynamic network representation[J]. Acta Physica Sinica, 2020, 69(16): 168901-1 Copy Citation Text

show less

Fig. 1. Schematic diagram of dynamic network.

Download full size | View in the Article

Fig. 2. The architecture of link prediction model based on dynamic network representation.

Download full size | View in the Article

Fig. 3. Time interval based LSTM unit.

Download full size | View in the Article

Fig. 4. Schematic diagram of node neighborhood sampling.

Download full size | View in the Article

Fig. 5. Node neighborhood update unit.

Download full size | View in the Article

Fig. 6.

comparison diagram on each data set. (a) UCI dataset; (b) DNC dataset; (b) Wikipedia dataset; (d) Enron dataset

Download full size | View in the Article

Fig. 7.

comparison diagram on each data set. (a) UCI dataset; (b) DNC dataset; (b) Wikipedia dataset; (d) Enron dataset

Download full size | View in the Article

Fig. 8. Recall@k comparison diagram of the variants of DNRLP. (a) UCI dataset; (b) DNC dataset; (c) Wikipedia dataset; (d) Enron dataset.

Download full size | View in the Article

Fig. 9. MRR results of different training rates. (a) DNC dataset; (b) Enron dataset

Download full size | View in the Article

输入:　新增链接 $ {e}_{ij}\in {E}_{{\rm{new}}} $, 随机游走长度 $ L $
输出:　随机游走序列 $ R $
1) For $ {e}_{ij} $ in $ {E}_{{\rm{new}}} $ do:
2)　For $ v $ in $ {e}_{ij} $ do:
3)　　 $ m=\mathrm{ }0 $
4)　　While $ m < L $ do
5)　　　初始化权重分布 $ P $
6)　　　For $ u $ in $ {N}_{v} $ do
7)　　　　根据(13)式计算 $ {f}_{\rm{s}}\left({u}_{\rm{u}}, {u}_{v}\right) $, 加入 $ P $
8)　　　End for
9)　　　根据 $ P $选择下一个节点 $u^\prime$加入 ${R}_{v}$
10)　　　 $ m=m+1 $
11)　　　 $ v=u' $
12)　　End while
13)　　将 $ {R}_{v} $加入 $ R $
14)　End for
15) End for

Table 1.

Information diffusion algorithm.

信息扩散算法

数据集	节点数	边数	时间/d	聚类系数/%
UCI	1899	59835	194	5.68
DNC	2029	39264	982	8.90
Wikipedia	1219241	2284546	4763	0.000837
Enron	384413	1751463	1140	4.96

Table 2.

Dynamic network data details.

动态网络数据详细信息

项目	设置	数量
操作系统	Ubuntu 16.04	1
CPU	Intel®i7-5280K, 6 核, 12线程	1
硬盘	512GB PLEXTOR®PX-512M6Pro SSD	1
内存	Kingston®8GB DDR4 2400	8
重要程序包	Python 3.7	1
深度学习框架	PyTorch	1

Table 3.

Experimental environment setup information.

实验环境设置信息

方法	UCI	DNC	Wikipedia	Enron
Logistic Regression	0.005 1	0.020 9	0.003 7	0.005 2
SVM	0.003 2	0.018 2	0.002 1	0.002 9
Node2Vec	0.004 7	0.019 7	0.003 5	0.003 9
GCN	0.015 9	0.048 4	0.010 1	0.017 6
GraphSAGE	0.016 3	0.049 7	0.012 0	0.018 3
DynGEM	0.015 7	0.028 4	0.010 8	0.014 7
GCN-GAN	0.020 1	0.050 4	0.014 9	0.021 5
DDNE	0.014 2	0.026 8	0.009 6	0.011 6
DNRLP	0.035 1	0.053 9	0.018 7	0.036 3

Table 4.

Link prediction MRR results comparison.

链接预测MRR结果对比

Zhong-Ming Han, Sheng-Nan Li, Chen-Ye Zheng, Da-Gao Duan, Wei-Jie Yang. Link prediction model based on dynamic network representation[J]. Acta Physica Sinica, 2020, 69(16): 168901-1

Download Citation

Tools

Save the article for my favorites

Paper Information