SBGMN: A Multi-view Sign Prediction Network for Bipartite Graphs

Signed bipartite graphs differ from traditional graphs, which include two sets of nodes and signed edges. With the development of information technology, sign prediction in bipartite graphs has become a hotspot in both research and industrial fields. There are a large number of applications that rely on accurate methods to predict signs on bipartite graphs, such as product recommendation systems, e-commerce platforms, social media sentiment analysis, and political inclination analysis. Nowadays, many works have attempted to propose outstanding methods to achieve this task. Thanks to the balance theory and powerful graph neural network models, existing methods can accomplish prediction tasks well. However, researchers are still facing a challenge in fully utilizing node information in such special graphs. Motivated by the current state of research, we propose a new model named Signed Bipartite Graph Multi-View Network (SBGMN), which can provide a more effective sign prediction result in bipartite graphs. This method is based on multiple views and includes four modules. The first is the information enhancer module, which can provide enhanced information for the subsequent modules. The second module is the multi-view kNN module. A well-designed kNN-based one-mode projection generator builds one-mode graphs with nodes of the same type and connections between them. This module is crucial to analyzing the potential relationships between same-type nodes and helps the following multi-view GNNs capture more valuable information for the task. The third one is named the collaborative GNN module. This module is introduced to extract more comprehensive attributes from the original edges and virtual edges established by balance theory. Finally, the sign prediction module maps the refined features from the above modules to the final sign prediction results. Extensive experimental results indicate that the proposed model outperforms state-of-the-art methods.