Adaptive Node Embedding Propagation for Semi-Supervised Classification

Abstract

Graph Convolutional Networks (GCNs) are state-of-the-art approaches for semi-supervised node classification task. By increasing the number of layers, GCNs utilize high-order relations between nodes that are more than two hops away from each other. However, GCNs with many layers face three drawbacks: (1) over-fitting due to the increasing number of parameters, (2) over-smoothing in which embeddings converge to similar values, and (3) the difficulty in selecting the appropriate number of propagation hops. In this paper, we propose ANEPN that effectively utilizes high-order relations between nodes by overcoming the above drawbacks of GCNs. First, we introduce Embedding Propagation Loss which increases the number of propagation hops while keeping the number of parameters constant for mitigating over-fitting. Second, we propose Anti-Smoothness Loss (ASL) that prevents embeddings from converging to similar values for avoiding over-smoothing. Third, we introduce a metric for predicted class labels for adaptively controlling the number of propagation hops. We show that ANEPN outperforms ten state-of-the-art approaches on three standard datasets.