Multi-view k-Nearest Neighbor Graph Contrastive Learning on Multi-modal Biomedical Data

Machine learning (ML) and deep learning (DL) on biomedical datasets is an important and ubiquitous task with applications from drug discovery to brain image segmentation to diagnosis and classification tasks in medicine. The primary challenges in this domain are (i) finding a way to encode biomedical data so that it can be easily exploited by ML and DL models and (ii) capacities of ML/DL models largely come from high-quality annotated labels, which are expensive to collect. To learn promising representation of biomedical data and handle unlabeled data, we propose Multi-View k-Nearest Neighbor Graph Contrastive Learning (MV-KGCL) in this paper. Specifically, MV-KGCL extends the graph neural network-based contrastive learning, which is powerful in graph representation learning, but is degraded when structural information is not clear enough or there is non-graph observation. MV-KGCL tackles these direct causes of the problem of graph neural network-based contrastive learning models. That is, we extract various k-nearest neighbor (kNN) graphs from multiple views of entity features, and then conduct contrastive learning over these kNN graphs in the local and the global views. Experiments on two datasets demonstrate that MV-KGCL outperforms the state-of-the-art methods due to better evaluation metrics of clustering, capturing meaningful patterns and heterogeneity within the data, and enabling more interpretable and clinically meaningful clusters that are predictive to cognition conditions.