Dynamic Network Embedding by Time-Relaxed Temporal Random Walk

Yifan Song, Darong Lai, Zhihong Chong, Zeyuan Pan

Abstract: Network embedding, also called network representation learning, aims at mapping high dimensional network information into low dimensional vectors. Previous studies mainly focus on static networks. In recent years, dynamic network embedding attracts much attention and methods specific to dynamic network are emerging. However, previous dynamic network embedding methods, such as CTDNE, still have drawbacks when using random walk to generate node sequences. Temporal random walk strictly requires that the time value of next edge be larger (i.e. later visiting) than that of the previous visited edge, which often leads to insufficient information obtained by random walk. In this article, a novel model named Time-Relaxed Temporal Random Walk(TxTWalk) for dynamic network embedding is proposed. Firstly, a time-relaxed function is designed, which enables random walk to select the next edge in a time interval, not strictly larger than the time of previously visited edge. It can make the walking sequences obtained by TxTWalk contain a wider range of temporal information. Then the node sequences are put into the skip-gram model for training to generate embedding of nodes on dynamic networks. Experimental validations on various networks demonstrate that TxTWalk is more effective than other state-of-the-art methods in link prediction.