Learning Representations for Bipartite Graphs Using Multi-task Self-supervised Learning

Akshay Sethi, Sonia Gupta, Aakarsh Malhotra, Siddhartha Asthana

Abstract: Representation learning for bipartite graphs is a challenging problem due to its unique structure and characteristics. The primary challenge is the lack of extensive supervised data and the bipartite graph structure, where two distinct types of nodes exist with no direct connections between the nodes of the same kind. Hence, recent algorithms utilize Self Supervised Learning (SSL) to learn effective node embeddings without needing costly labeled data. However, conventional SSL methods learn through a single pretext task, making the trained model specific to the downstream task. This paper proposes a novel approach for learning generalized representations of bipartite graphs using multi-task SSL. The proposed method utilizes multiple self-supervised tasks to learn improved embeddings that capture different aspects of the bipartite graphs, such as graph structure, node features, and local-global information. We utilize deep multi-task learning (MTL) to further assist in learning generalizable self-supervised solution. To mitigate negative transfer when related and unrelated tasks are trained in MTL, we propose a novel DST++ algorithm. The proposed DST++ optimization algorithm improves existing DST by considering task affinities and groupings for better initialization and training. The end-to-end proposed method with complimentary SSL tasks and DST++ multi-task optimization is evaluated on three tasks: node classification, link prediction, and node regression, using four publicly available benchmark datasets. The results demonstrate that our proposed method outperforms state-of-the-art methods for representation learning in bipartite graphs. Specifically, our method achieves up to 12% improvement in accuracy for node classification and up to 9% improvement in AUC for link prediction tasks compared to the baseline methods.