MVIB-RGSL: Multi View Information Bottleneck based Robust Graph Structure Learning

Recently, graph neural networks have achieved significant results in various graph structure learning tasks. However, the graph structure in reality often contains noise, which could affect the effectiveness and robustness of the model prediction. In order to improve the robustness of the graph structure learning model, many studies have introduced the idea of information bottleneck. Information bottleneck can improve the robustness of the model. However, information bottleneck requires labeled data in the dataset to remove the information irrelevant to the prediction task. Therefore, for the unsupervised scenario, we propose MVIB-RGSL (Multi View Information Bottleneck based Robust Graph Structure Learning). We use the method of graph data augmentation to generate different views on the basis of the original input graph data, thereby expanding the problem to the multi-view learning scenario. Then we use the information bottleneck under the multi-view setting to remove view-specific redundant information between the two views. Finally, we use the adversarial method to train the graph encoder while optimizing the graph structure generated by graph data augmentation, thereby obtaining more effective and robust graph structures and graph representations.