Improving Graph Neural Network by Filter Preprocessing

For graph-based semi-supervised learning, Graph Convolutional Networks (GCNs) and their variants has shown outstanding results and gained wide attention. Several works on analyzing GCNs from the perspective of spectral graph theory shows that GCNs has the function of low-pass filtering on certain learning tasks, which enables us to have a deeper understanding of GCN. However, GCN achieves the filtering effect through layers of matrix multiplication, so one cannot flexibly control the filtering process. In this paper, we propose to preprocess the graph-structured data by using low-pass filters explicitly before network training and prediction. We conduct experiments on the citation network datasets. Our preliminary results show that, the filter preprocessing step can effectively improve the predicative accuracy of common neural networks or graph neural networks in graph-based semi-supervised learning.