Machine learning tasks and representations for heterogeneous information networks

Heterogeneous information networks (HINs) are ubiquitous. Social networks, knowledge graphs, and interactions between users and items in search and recommender systems can be modeled as networks with multiple types of nodes and edges. Traditional network representation learning models simply learn network embeddings based on a given network, ignoring various real-life scenarios and limitations.
The work in this thesis provides a series of algorithms that are able to deal with different HIN scenarios. We first focus on dynamic HINs as real-life networks are always evolving. M-DHIN is proposed as a method to learn dynamic embeddings; it is also able to predict the future network.
After that, we study the pre-training problem in HINs. Since most networks are unlabeled, we propose a pre-training and fine-tuning framework PF-HIN that uses two self-supervised pre-training tasks. The pre-trained encoder can quickly be adapted to datasets of different domains and different downstream tasks.
We also investigate the few-shot problem for HINs as in practice, only a handful of nodes are labeled. We propose META-HIN, which uses a meta-learning framework that is able to handle the few-shot learning tasks in multiple-graph scenarios.
In the final research chapter, we study how to leverage meaningful textual information that may be contained in a HIN. We propose a novel prompt learning framework P-HIN that is able to simultaneously employ the textual information and handle the few-shot problems. We align the text representation and node representation using a contrastive learning mechanism, so that textual information is incorporated in an effective manner.