Cross-GCN: Enhancing Graph Convolutional Network with k-Order Feature Interactions

Graph Convolutional Network (GCN) is an emerging technique that performs learning and reasoning on graph data. It operates feature the structure, through aggregating features of neighbor nodes to obtain embedding each target node. Owing strong representation power, recent research shows GCN achieves state-of-the-art performance several tasks such as recommendation linked document classification. Despite its effectiveness, we argue existing designs forgo modeling cross features, making less effective for or data where are important. Although neural network can approximate any continuous function, including multiplication operator crosses, it be rather inefficient do so (i.e., wasting many parameters at risk overfitting) if there no explicit design. To this end, design a new named Cross-feature Convolution, which explicitly models arbitrary-order with complexity linear dimension order size. We term our proposed architecture Cross-GCN, conduct experiments three graphs validate effectiveness. Extensive analysis validates utility in GCN, especially lower layers.