Graph convolutional networks for graphs containing missing features
This approach integrates the processing of missing features and graph learning within the same neural network architecture and demonstrates through extensive experiments that this approach significantly outperforms the imputation based methods in node classification and link prediction tasks.
Abstract
Graph Convolutional Network (GCN) has experienced great success in graph\nanalysis tasks. It works by smoothing the node features across the graph. The\ncurrent GCN models overwhelmingly assume that the node feature information is\ncomplete. However, real-world graph data are often incomplete and containing\nmissing features. Traditionally, people have to estimate and fill in the\nunknown features based on imputation techniques and then apply GCN. However,\nthe process of feature filling and graph learning are separated, resulting in\ndegraded and unstable performance. This problem becomes more serious when a\nlarge number of features are missing. We propose an approach that adapts GCN to\ngraphs containing missing features. In contrast to traditional strategy, our\napproach integrates the processing of missing features and graph learning\nwithin the same neural network architecture. Our idea is to represent the\nmissing data by Gaussian Mixture Model (GMM) and calculate the expected\nactivation of neurons in the first hidden layer of GCN, while keeping the other\nlayers of the network unchanged. This enables us to learn the GMM parameters\nand network weight parameters in an end-to-end manner. Notably, our approach\ndoes not increase the computational complexity of GCN and it is consistent with\nGCN when the features are complete. We demonstrate through extensive\nexperiments that our approach significantly outperforms the imputation-based\nmethods in node classification and link prediction tasks. We show that the\nperformance of our approach for the case with a low level of missing features\nis even superior to GCN for the case with complete features.\n