DFNets: Spectral CNNs for Graphs with Feedback-Looped Filters

W. O. K. Asiri Suranga Wijesinghe, Qing Wang

Abstract: We propose a novel spectral convolutional neural network (CNN) model on graph structured data. This model is incorporated with a powerful class of spectral graph filters, called feedback-looped filters. These filters can enable better localization over vertices in q-hop neighborhood, but are still computationally efficient, i.e., attaining fast convergence and memory requirement linearly in the number of edges in a graph. Theoretically, feedback-looped filters have guranteed convenicence w.r.t. a specified error bound, and can also be applied universally to any graph without knowing its structure. The propagation rule in our model further diversifies features from the preceding layers to make very strong gradient flow. We evaluate our model using two benchmark tasks: semi-supervised document classification on citation networks and semi-supervised entity classification on a knowledge graph. Our experimental results show that our model outperforms the state-of-the-art methods in both benchmark tasks over all datasets.

CMT Num: 3235

Code Link: https://github.com/wokas36/DFNets