Frequency Inception Based Graph Convolutional Neural Networks

EasyChair Preprint 8046

Abstract

Graph convolutional neural networks (GCNs) have demonstrated powerful representing ability of irregular data, e.g., skeletal data and graph-structured data, providing the effective mechanism to fuse the neighbor nodes. The representative are the spectral-based methods, which are designed to obtain the beneficial discriminative information from input signals for learning. However, many works have been shown that the essence of the GCN are low-pass filters, which propagate information and distill the beneficial signals, thus performing the information denoising. Although there are some efforts not only concentrate on the fixed low-pass filters, but also the adaptive frequency filters, which harness the dynamic frequency, they do not go deeply into the intrinsic part of the useful signals of all nodes. To explore the core of signals propagating, we design a novel framework FiGCN that leverages the each of channel signal, which comprises of all the neighbor. Specifically, every channel of a node and its neighborhoods contribute dynamically to the final channel signal, which can capture the inherent difference of different channel and neighbor nodes and even determine whether a node is neighborhood or not. Meanwhile, it can enhance the representation ability of nodes and ameliorate the over-smoothing problem. On the other hand, our model can dynamically adjust the importance of neighborhoods to the central vertex. We empirically validate the effectiveness of the proposed framework FiGCN on various benchmark datasets. Experimental results show that our method achieves substantial improvements and outperforms the state-of-the-art performance significantly.

Keyphrases: Graph Convolutional Neural Networks, Knowledge Graph, deep learning

@booklet{EasyChair:8046,
  author    = {Feifei Wei and Kuizhi Mei},
  title     = {Frequency Inception Based Graph Convolutional Neural Networks},
  howpublished = {EasyChair Preprint 8046},
  year      = {EasyChair, 2022}}