SWARM: A Scientific Workflow Fragments Recommendation Approach via Contrastive Learning and Semantic Matching

Yang Gu, Jian Cao, Jinghua Tang, Shiyou Qian, Wei Guan

Abstract: Discovering and recommending scientific workflow fragments that can be reused or repurposed from public repositories is becoming increasingly significant in the scientific domain. Although popular fragment discovery strategies can identify frequent and similar fragments, they lack the ability to further mine pattern semantics. Moreover, current recommendation approaches are primarily based on text matching between natural language queries and the descriptions of candidate fragments, neglecting crucial structural information that conveys their functions. To address these challenges, this paper designs SWARM, a scientific workflow fragments recommendation approach via contrastive learning and semantic matching. SWARM consists of two phases: fragment discovery based on frequent subgraph mining and a contrastive semantics extraction model, and fragment recommendation based on a matching degree prediction model incorporating a pre-trained fragment encoder, which is used to predict and rank the degree of semantic matching between the user query and candidate fragments. SWARM aims to extract and integrate textual and structural semantics from fragments to discover and recommend them. The experimental results on commonly-used real-world datasets show that SWARM outperforms state-of-the-art methods with statistical significance.