Semi-Supervised Classification with Adaptive High-Order Graph Embedding

Abstract

The problem of semi-supervised graph node classification is to infer the labels of unlabeled nodes based on a partially labeled graph. Graph embedding is an effective method for this problem, which utilizes the context generated by neighbors' information. Some recent approaches preserve high-order proximity to smooth the features embedded with long-range structure dependency. However, the features generated by high-order proximity may be too smooth to lost individual characteristics. To handle this problem, we propose Adaptive High-Order Graph Embedding (AHOGE), an end-to-end graph neural network that implements embedding and classification in a unified model, to retain individual details when preserving high-order proximity. Inspired by Densely Connected Convolutional Networks (DenseNets), AHOGE adaptively adopts the information of $k^{th}$-order proximity for different $k$, using the techniques of Highway Network. Moreover, we introduce multi-class hinge loss to deal with the hard annotated labels and class overlap. Experiments on three benchmark citation network datasets demonstrate that our approach achieves state-of-the-art performances.