Beyond Low-Pass Filtering: Graph Convolutional Networks With Automatic Filtering

Graph convolutional networks are becoming indispensable for deep learning from graph-structured data. Most of the existing graph share two big shortcomings. First, they essentially low-pass filters, thus potentially useful middle and high frequency band signals ignored. Second, bandwidth filters is fixed. Parameters a filter only transform inputs without changing curvature function. In reality, we uncertain about whether should retain or cut off at certain point unless have expert domain knowledge. this paper, propose Automatic Convolutional Networks (AutoGCN) to capture full spectrum automatically update filters. While it based on spectral theory, our AutoGCN also localized in space has spatial form. Experimental results show that achieves significant improvement over baseline methods which work as