GAN-based Complex Semantics Augmented Heterogeneous Information Network Embedding

Abstract

Heterogeneous information network (HIN) embedding is fundamental but extremely challenging because of the existence of various types of nodes and links. Expressive metagraphs are at the core of HIN embedding methods seeking to match the desired semantics; however, they are typically expensive to obtain and have a high risk of missing meaningful paths among nodes. Instead of heavily relying on domain knowledge to enumerate metagraphs among exponential numbers of potential metagraphs, we propose a novel representation learning model that mines the complex semantics underlying HINs. First, a set of meaningful metagraphs are first constructed to model the core semantic relations. Then, we generate semantic sequences via random walks, which are guided by metagraphs. A complex semantic augmented method is suggested to concatenate sequences with shared nodes for capturing semantic paths between distant nodes, which is hard to extract at the level of metapaths and easily fail to mine without domain knowledge. A pretraining and fine-tuning framework employing adversarial learning that generates semantic-preserving and robust representations, with semantically augmented paths reflecting long-distant correlations of multiple types of nodes, is investigated. We conduct the experiments on a real-life HIN, and the results show that our model performs better than the state-of-the-art alternatives.