PROJECT TITLE :

RHINE: Relation Structure-Aware Heterogeneous Information Network Embedding

ABSTRACT:

The goal of heterogeneous information network (HIN) embedding is to learn the low-dimensional representations of nodes within HINs while simultaneously maintaining the networks' structures and semantics. Although the majority of currently available methods take into account heterogeneous relations and achieve promising performance, they typically only use a single model to represent all relations without making any distinctions between them. This inherently limits the capability of HIN embedding. In this paper, we argue that heterogeneous relations have different structural characteristics, and we propose a novel Relation structure-aware HIN Embedding model, which we will refer to as RHINE. We present two structure-related measures that consistently distinguish heterogeneous relations into two categories: Affiliation Relations (ARs), and Interaction Relations (IRs). This was accomplished by conducting in-depth research on four real-world networks and then analyzing the results (IRs). In RHINE, we propose different models that have been specifically tailored to handle ARs and IRs. These models are able to better capture the structures in HINs, which is important because we want to respect the unique structural characteristics of relations. Last but not least, we unify our approach to optimizing and combining all of these models. In addition, taking into account the fact that nodes connected via heterogeneous relations may have multi-aspect semantics despite the fact that each relation focuses on a single aspect, we introduce relation-specific projection matrices in order to learn node and relation embeddings in separate spaces as opposed to a common space. This can better preserve the semantics in HINs and is referred to as the new model RHINE-M. Our models have been shown to perform significantly better than the methods considered to be state-of-the-art in four different tasks, as demonstrated by experiments conducted on real-world datasets.