PolarDSN: An Inductive Approach to Learning the Evolution of Network Polarization in Dynamic Signed Networks

Abstract

The goal of dynamic signed network embedding (DSNE) is to represent the nodes in a dynamic signed network (DSN) as embeddings that preserve the evolving nature of conflicting relationships between nodes. While existing DSNE methods are useful for understanding polarization between users in diverse domains, they fail to consider the concept of a community boundary that contributes to network-wide polarization and lack inductive ability due to their reliance on homophily bias. To address these limitations, we propose a novel DSNE method, named PolarDSN, which learns the evolution of network POLARization and enhances inductive ability for Dynamic Signed Networks. It leverages node-level community boundaries as well as structural characteristics of nodes such as structural isomorphism and temporal transitivity. Experiments on four real-world DSN datasets demonstrate that PolarDSN consistently and significantly outperforms 12 state-of-the-art methods, achieving up to 31.6% and 21.1% improvement in macro-F1 for transductive and inductive settings, respectively. The code is available at https://github.com/kmj0792/PolarDSN.