Multi-label Graph Node Classification with Label Attentive Neighborhood Convolution

Abstract Learning with graph structured data is of great significance for many practical applications. A crucial and fundamental task in graph learning is node classification. In reality, graph nodes are often encoded with various attributes. In addition, the task is usually multi-labeled in nature. In this paper, we tackle the problem of multi-label graph node classification, by leveraging structure, attribute and label information simultaneously. Specifically, to obtain rational node feature representations, we propose an intuitive yet effective graph convolution module to aggregate local attribute information of a given node. Moreover, the homophily hypothesis motivates us to build a label attention module. By exploiting both input and output contextual representations, we utilize the additive attention mechanism and build a label-aware representation learning framework to measure the compatibility between pairs of node embeddings and label embeddings. The proposed novel neural network-based, multi-label classification method has been verified by extensive experiments conducted on five public-available benchmark datasets, including both attributed and non-attributed networks. The results demonstrate the effectiveness of the proposed model with respect to micro-F1, macro-F1 and Hamming loss, comparing with several state-of-the-art methods, including two relational neighbor classifiers and several popular graph neural network models.