Global View For GCN: Why Go Deep When You Can Be Shallow?

Hong Chul Nam, Ye Sle Cha, Chanwoo Park

Keywords: GCN, GNN, Clustering, Semi-supervised Learning

Abstract: Existing graph convolutional network (GCN) methods attempt to expand the receptive field of its convolution by either stacking up more convolutional layers or accumulating multi-hop adjacency matrices. Either approach increases computation complexity while providing a limited view of the network topology. We propose to extend k-hop adjacency matrices into one generalized exponential matrix to provide GCNs with a global overview of the network topology. This technique allows the GCNs to learn global topology without going deep and with much fewer parameters than most state-of-the-art GCNs, challenging the common assumption that deep GCNs are empirically better for learning global features. We show a significant improvement in performance in semi-supervised learning when this technique is used for common GCNs while maintaining much shallower network architectures ($\leq4$ layers) than the existing ones.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)