Simplifying Graph Convolutional Networks

Graph Convolutional Networks (GCNs) and their variants became the de facto methods for learning graph representations, but by drawing inspiration primarily from deep learning approaches they may inherit unnecessary complexity and redundant computation. The paper reduces this excess complexity by successively removing nonlinearities and collapsing weight matrices between consecutive layers, and theoretically analyzes the resulting linear model, showing it corresponds to a fixed low-pass filter followed by a linear classifier.

The experimental evaluation demonstrates that these simplifications do not negatively impact accuracy in many downstream applications. Moreover, the resulting model scales to larger datasets, is naturally interpretable, and yields up to two orders of magnitude speedup over FastGCN, showing that much of a GCN's apparent complexity is unnecessary for strong performance.