SAGCN: Towards Structure-Aware Deep Graph Convolutional Networks on Node Classification

Graph Convolutional Networks (GCNs) have recently achiev-ed impressive performance in different classification tasks. However, over-smoothing remains a fundamental burden to achieve deep GCNs for node classification. This paper proposes Structure-Aware Deep Graph Convolutional Networks (SAGCN), a novel model to overcome this burden. At its core, SAGCN separates the initial node features from propagation and directly maps them to the output at each layer. Furthermore, SAGCN selectively aggregates the information from different propagation layers to generate structure-aware node representations, where the attention mechanism is exploited to adaptively balance the information from local and global neighborhoods for each node. Our experiments verify that the SAGCN model achieves state-of-the-art performance in various semi-supervised and full-supervised node classification tasks. More importantly, it outperforms many other backbone models, by using half the number of layers, or even fewer layers.