An Attention-Based Adversarial Disentangle Heterogeneous Embedding for Improving Node Classification

Abstract

For many years, within the information network analysis and mining (INAM) area, heterogeneous information network (HIN) is considered as a hot topic which have attracted attentions from many researchers. In fact, HIN is considered as the sufficient reflection of real-world information networks which contains multiple types of nodes and links (e.g., knowledge graph, social networks, etc.). Within HIN-based analysis and mining domain, meta-path is considered as the key concept which is employed to extract rich contextual information from semantic relationships, in forms of meta-paths, between multi-typed nodes. However, the various semantic informative sources which are obtained from different meta-paths between nodes are frequently considered as ambiguous and entangled for primitive tasks in INAM area such as node classification, clustering or link prediction. Recently, there are several well-known studies which have dedicated on the disentangling the meta-path-based representations of nodes in HIN through different techniques such as variational embedding and adversarial learning paradigms. However, the disentangled meta-path-based node representation learning in HIN still have several unresolved problems. One of the most crucial limitations of this research direction is related to the capability of properly interpretation for meta-path-varied disentangled node embeddings in order to achieve better fine-tuned performance for different downstream tasks, such as node classification. In order to overcome this problem, in this paper we proposed an improvement of disentangled HIN representation learning for node classification task with the employment of attention-based mechanism, called as: HAttDE model. A self-attention-based mechanism is placed between the encoding and generating components in order to softly align and interpret multiple fused meta-path-based node representations into better classification-friendly embedding spaces. Extensive experiments in benchmark networked datasets demonstrated the effectiveness of our proposed model in comparing with recent state-of-the-art HIN embedding baselines.