Role-Based One-Hop Neighborhood Representation for Link Sign Prediction in Signed Directed Networks

Link sign prediction refers to predicting whether a pair of nodes in a network is connected by a positive or negative link. Regarding this typical and fundamental problem in signed network analysis, subgraph encoding via linear optimization (SELO) currently provides the state-of-art performance to the best of our knowledge. However, existing sign prediction algorithms commonly ignore the diversified structural roles of neighboring nodes and learn node embeddings by directly mirroring unsigned and undirected aggregators. To address these limitations, we propose a role-based one-hop neighborhood representation (RONR) framework to predict link signs in signed directed networks, which outperforms state-of-the-art methods. The main novelties of our proposed approach lie in that: 1) only one-hop neighborhood information of the target link is required; 2) a node attribute assignment algorithm is given to distinguish the role of each node in the neighborhood; and 3) a specific efficient node feature aggregator is developed to learn subgraph embeddings. We conduct experiments on five real-world signed directed networks and evaluate the performances with AUC, F1, micro-F1, and macro-F1. The experimental results show that our proposed RONR achieves significantly better performances than previous feature-based, embedding-based, and subgraph pattern encoding methods, including the state-of-the-art method SELO.