Adaptive Propagation Graph Convolutional Network

Graph convolutional networks (GCNs) are a family of neural network models that perform inference on graph data by interleaving vertexwise operations and message-passing exchanges across nodes. Concerning the latter, two key questions arise: 1) how to design differentiable exchange protocol (e.g., one-hop Laplacian smoothing in original GCN) 2) characterize tradeoff complexity with respect local updates. In this brief, we show state-of-the-art results can be achieved adapting number communication steps independently at every node. particular, endow each node halting unit (inspired Graves’ adaptive computation time [1]) after decides whether continue communicating or not. We proposed propagation GCN (AP-GCN) achieves superior similar best so far benchmarks while requiring small overhead terms additional parameters. also investigate regularization term enforce an explicit between accuracy. The code for AP-GCN experiments is released as open-source library.