BTR-MCL: a Bioinformatics Tool Recommendation Method with Multi-view Contrastive Learning

Digital informatization in bioinformatics research has developed rapidly, resulting in the emergence of tens of thousands of bioinformatics tools in recent decades. However, few researchers systematically study how to manage and recommend bioinformatics tools. Meanwhile, existing recommendation methods suffer from the incompatibility with bioinformatics tool recommendations, specifically, i) the inability to mine a large amount of descriptive information and potential correlations of bioinformatics tools to make up for the sparse problem of operation-tool history interaction; ii) the inability to accommodate recommendations dominated by tools (i.e., items). To address these problems, we collected quantitative bioinformatics operations, tools, and context (i.e., topics and papers) from the authoritative portal (https://bio.tools) and constructed a knowledge graph dataset for the bioinformatics tools. Based on this, we propose the Bioinformatics Tool Recommendation Method with Multi-view Contrastive Learning (BTR-MCL). The method utilizes BERT to enrich the semantic information of entities and constructs three fine-grained views for multi-view contrastive learning, namely the tool-tool graph named TT view, the operation-tool graph named OT view, and the bioinformatics tools knowledge graph as global structure named KG view. By enhancing the importance of tool embedding in local and global contrastive learning and learning rich semantic information, BTR-MCL achieves significant improvements over state-of-theart baselines on the top-K recommendation, outperforming the higher 82.71% of the best baseline on the Recall@10 metric