Sequence Contained Heterogeneous Graph Neural Network

Abstract

Heterogeneous Information Network (HIN) has been proved to be effective when expressing rich information for realworld data recently. However, the previous HIN embedding methods can't handle well the sequence contained HIN, which consists of both non-sequence and sequence entities simultaneously. Previous methods simply represent sequential entities as non-sequential feature vectors by pre-process, leading to the reduction in sequence-dependency information (e.g., context-dependency, time-dependency) within entities. Due to the widespread of sequence-dependency information, the non-sequential methods limit the application of HIN in real-word. Therefore, we first propose a novel heterogeneous graph neural network, named as Sequence Contained Heterogeneous Graph Neural Network (SC-HGNN), to deal with sequential features in this paper. Specifically, we utilize Transformer Projection to preserve the relation within the sequence. Then we consider sequential features for node embedding by using Intra-type Sequence Attention and Inter-type Sequence Attention to capture the relation between homogeneous and heterogeneous sequence respectively. Finally, another transformer is used to encode the updated features of the center node. Extensive experimental results on a real-world dataset demonstrate that our SC-HGNN method outperforms state-of-the-art baselines in sequence contained HIN on node classification and clustering tasks.