Selected Research Projects

Deep Graph Contrastive Learning

Graph Contrastive Learning with Adaptive Augmentation

Existing deep graph neural networks (GNNs) typically require a large amount of labeled data in order to be trained effectively. This can be a significant challenge, as collecting and labeling data can be time-consuming and labor-intensive, especially for large and complex datasets. Fortunately, unlabeled data is often more readily available in practice, and can provide valuable information about the structure and patterns of the data. The question, then, is whether it is possible to learn from unlabeled data using GNNs, and if so, how can we do it effectively? Our research addresses this question using the technique of graph contrastive learning, which involves augmenting the input graphs in various ways and then maximizing the consistency between the augmented versions of the graphs. Our approach significantly improves previous methods for unsupervised graph learning, and in some cases even outperforms supervised approaches that use labeled data.

Further reading:

  • Continue reading our blog post on deep graph contrastive learning
  • Read our two research papers on generic graph contrastive learning, GRACE (GRL+@ICML 2020) and GCA (WWW 2021)
  • Read our empirical study of graph contrastive learning (NeurIPS Dataset and Benchmark 2021)
  • Read our papers on heterogeneous and multiview graph contrastive learning: STENCIL (SDM 2022) and CREME (CIKM 2022)
  • Check out our maintained collection on self-supervised graph representation learning literature and resources
  • Try our developed PyGCL, a PyTorch library that features modularized components, standardized evaluation, and experiment management for graph contrastive learning

Graph-Based Recommender Systems

Session-based Recommendation with Graph Neural Networks

Nowadays, recommender systems have become an indispensable tool for helping users navigate the overwhelming amount of information available online. Considering that users, items, and contexts within recommender systems are tightly connected, a natural question is whether graph structures can be helpful for mining users’ complex behaviors to enhance recommendations. Our work in this rigorous field provides affirmative answers to this question. Particularly, we develop the very first graph-based recommender systems for sequential session data, which captures the temporal dynamics of users’ interactions with items. In addition, we also apply our graph-based approach to other recommendation settings, such as multimedia recommendation, where users’ interactions with multimedia content can be modeled as a graph and used to improve the quality of recommendations.

Further reading:

  • Continue reading our blog post on session-based recommendation with graph neural networks
  • Read our two session-based recommendation models with graph neural networks, SR-GNN (AAAI 2019) and TAGNN (SIGIR 2020)
  • Read our work on graph structure learning for multimedia recommendation, LATTICE (ACMMM 2021) and its extended model MICRO (TKDE 2022)
  • Read our work on graph-based code repository recommendation CODER (WWW 2023)