RankedDrop: Enhancing Deep Graph Convolutional Networks Training
Graph Neural Networks (GNNs) are playing a more and more important role for analyzing unstructured data from the complex real world. Introducing random edge dropping from the input graph at training epochs could reduce over-fitting and over-smoothing phenomenon and increase the depth of GNNs. However, such method relies strongly on the chosen randomness. It makes the accuracy depend on the initialization of the randomness, which let the selection of hyper-parameters be even more difficult. We propose in this paper RankedDrop a novel method with a spatial-aware dropping-edge selection. The selection takes account of graph global information using PageRank, and graph local neighborhood information with node degree. RankedDrop provides a more stable training results comparing to the state-of-the-art solution, by maintaining the advantages of random edge dropping. Furthermore, RankedDrop is a general method that can be deployed on a deep learning framework for enhancing performance of GNNs.
PDF Abstract