Graph-Revised Convolutional Network

Graph Convolutional Networks (GCNs) have received increasing attention in the machine learning community for effectively leveraging both content features of nodes and linkage patterns across graphs various applications. As real-world are often incomplete noisy, treating them as ground-truth information, which is a common practice most GCNs, unavoidably leads to sub-optimal solutions. Existing efforts addressing this problem either involve an over-parameterized model difficult scale, or simply re-weight observed edges without dealing with missing-edge issue. This paper proposes novel framework called Graph-Revised Network (GRCN), avoids extremes. Specifically, GCN-based graph revision module introduced predicting missing revising edge weights w.r.t. downstream tasks via joint optimization. A theoretical analysis reveals connection between GRCN previous work on multigraph belief propagation. Experiments six benchmark datasets show that consistently outperforms strong baseline methods, especially when original severely labeled instances training highly sparse. (Our code available at https://github.com/Maysir/GRCN).