SimWalk: Learning network latent representations with social relation similarity

Abstract

In this paper, we present a novel method, namely SimWalk, to learn latent representations of networks. SimWalk maps nodes to a continuous vector space which maximizes the likelihood of node sequences. We design a probability-guided random walk procedure based on relation similarity, which encourages node sequences to preserve context-related neighborhoods. Different with previous work which generates rigid node sequences, we believe that relations in social net­works, especially similarity, can guide the walk to generate a more linguistic sequence. In this perspective, our model learns more meaningful representations. We demonstrate SimWalk on several multi-label real-world network classification tasks over state-of-the-art methods. Our results show that SimWalk outperforms the popular methods in complex networks.