DySDGNN: Representation Learning in Dynamic Signed Directed Networks

Dynamic signed directed networks are common in our daily lives so that representation learning in such networks is critical for many downstream tasks. However, most existing works only focus on either temporal or signed information in the graph. Moreover, taking account of both temporal and signed information into representation learning still faces some challenges. Firstly, edge sign between nodes could change (i.e. from positive to negative) over some time and this will bring the conflict in learning process. Secondly, balance theory and status theory that prior models based on are not suitable to learn signed semantics for dynamic signed network anymore. To tackle these challenges, we propose a novel clip-based Dynamic Signed Directed Graph Neural Networks model (DySDGNN). In DySDGNN, we design three components: the Attention-based TLSTM layer (A-TLSTM) to capture the different kinds of information within each clip, the Masked Temporal Self-Attention layer (MT-SA) to aggregate node embeddings across clips and the loss function related to the Signed Triad Transition Matrix (STTM) to measure the impact of different triads. We conduct extensive experiments and comparisons on three real-world datasets. The experimental results prove the superiority of DySDGNN compared with SOTA baselines on the downstream task of link sign prediction.