PROJECT TITLE :

mg2vec: Learning Relationship-Preserving Heterogeneous Graph Representations via Metagraph Embedding

ABSTRACT:

As a result of the fact that heterogeneous information networks (HIN) contain nodes and edges that belong to a variety of different semantic types, these networks are able to model complex data in situations that take place in the real world. As a result, there has been a rise in interest in HIN embedding, which seeks to learn node representations in a low-dimensional space. This is done with the intention of maintaining the structural and semantic information contained within the HIN. When viewed in this light, metagraphs, which are models of common and recurring patterns on HINs, emerge as a powerful tool for capturing semantically rich and frequently covert relationships on HINs. Although metagraphs have been used to address a few specific data mining tasks, they have not been thoroughly investigated for HIN embedding in its more general form. In this paper, we support a variety of relationship mining tasks by utilizing metagraphs to learn relationship-preserving HIN embedding in a self-supervised setting. In particular, we have noticed that the majority of the existing methods frequently make insufficient use of metagraphs. Metagraphs are only utilized during the pre-processing stage, and they do not actively guide representation learning after this stage. In light of this, we propose the novel framework known as mg2vec, which simultaneously learns the embeddings for metagraphs and nodes. That is to say, metagraphs engage in active participation in the learning process by mapping themselves to the same embedding space as the nodes do. In addition, metagraphs direct the learning process by imposing first- and second-order constraints on the embeddings of nodes. This allows them to model not only the latent relationships that exist between a pair of nodes, but also the preferences that are unique to each node. In conclusion, we run a large number of experiments on three different public datasets. The findings indicate that mg2vec performs significantly better than a suite of state-of-the-art baselines when it comes to relationship mining tasks such as prediction, search, and visualization of relationships.