Path-Aware Co-contrastive Learning for Signed Directed Network Embedding

Signed directed network embedding (a.k.a. representation) aims to generate low-dimensional node representations by considering both edge sign and direction. Previous studies predominantly focused on applying a singular social theory to derive node embeddings. It overlooks network structure and lacks comprehensive methods that integrate different theories, resulting in less thorough node representations. Recently, contrastive learning has gained popularity as a self-supervised method, showing great potential in learning from diverse perspectives. Inspired by this, we introduce Path-Aware Co-Contrastive Learning for Signed Directed Network Embedding (PASD). PASD is specifically crafted to capture the intricate interplay between edge sign and direction in both local and global structures. It intentionally incorporates both long and short paths modeling approaches to fully leverage the benefits of balance and status theories. To refine node embeddings, PASD integrates a dual-path co-contrastive learning mechanism, enabling the exploration of directional influences within the network, encompassing positive and negative relationships. Extensive experimental results demonstrate that PASD outperforms state-of-the-art methods on four widely-used datasets, validating the effectiveness of considering balance theory and status theory based on network structure in the task of signed directed network embedding.