DynamiSE: dynamic signed network embedding for link prediction

Haiting Sun, Peng Tian, Yun Xiong, Yao Zhang, Yali Xiang, Xing Jia, Haofen Wang

Abstract: In real-world scenarios, dynamic signed networks are ubiquitous where edges have positive and negative types and evolve over time. Graph neural networks have achieved impressive performance in node representation learning, facilitating various downstream tasks, i.e., link prediction. However, employing existing methods to learn dynamic signed network embedding faces the following challenges. First, it is hard to encode the dynamics and sign semantics of network simultaneously. Positive and negative edges have different sign semantics and the graph structure is dynamically changing with time. Second, it is non-trivial to alleviate over-smoothing and construct the deeper dynamic signed network due to the learning of network dynamics. In this paper, we propose Dynamic Signed Network Embedding (DynamiSE) to tackle the above problems. DynamiSE effectively integrates the balance theory and ordinary differential equation into node representation learning to encode the dynamics of network and capture sign semantics between neighbors. Specifically, we design the dynamic sign collaboration unit to construct a deeper dynamic signed graph neural network, which models the evolution patterns and simulates the propagation and aggregation of sign semantics. The complex sign influence between nodes formed by different semantics of positive and negative edges is captured by the sign semantics aggregation unit. Extensive experiments on real-world dynamic signed networks show that DynamiSE outperforms most state-of-the-art methods in link prediction task.