Circulant Tensor Graph Convolutional Network for Text Classification

Graph convolutional network (GCN) has shown promising performance on the text classification tasks via modeling irregular correlations between word and document. There are multiple correlations within a text graph adjacency matrix, including word-word, word-document, and document-document, so we regard it as heterogeneous. While existing graph convolutional filters are constructed based on homogeneous information diffusion processes, which may not be appropriate to the heterogeneous graph. This paper proposes an expressive and efficient circulant tensor graph convolutional network (CTGCN). Specifically, we model a text graph into a multi-dimension tensor, which characterizes three types of homogeneous correlations separately. CTGCN constructs an expressive and efficient tensor filter based on the t-product operation, which designs a t-linear transformation in the tensor space with a block circulant matrix. Tensor operation t-product effectively extracts high-dimension correlation among heterogeneous feature spaces, which is customarily ignored by other GCN-based methods. Furthermore, we introduce a heterogeneity attention mechanism to obtain more discriminative features. Eventually, we evaluate our proposed CTGCN on five publicly used text classification datasets, extensive experiments demonstrate the effectiveness of the proposed model.