One longstanding challenge in cloud API recommender systems is the Mashup cold-start problem, i.e., to recommend suitable cloud APIs for new Mashups without any historical invocation records. To address this challenge...
详细信息
One longstanding challenge in cloud API recommender systems is the Mashup cold-start problem, i.e., to recommend suitable cloud APIs for new Mashups without any historical invocation records. To address this challenge, researchers typically integrate various auxiliary information and feed them into a supervised learning model, or provide contextual information beyond interaction behaviors via explicit graph connections. However, these researches fail to 1) explore the functionally complementary relationship between Mashups and cloud APIs, and 2) model their semantic associations. To this end, we propose a Multi-level Graph Contrastive Learning, namely MGCL. Unlike traditional single-view methods, we comprehensively model Mashup-API relationships from three views: interaction, functionally complementary, and semantics. The interaction view captures dynamic behaviors, while the other two reveal associations based on static attributes. Additionally, we design a multi-level graph contrastive learning mechanism fuse information from these views in a self-supervised manner, enabling both cold-and warm start Mashup representations to benefit from dynamic interaction behaviors and static attributes. Extensive experiments on real-world datasets demonstrate that MGCL outperforms state-of-the-art methods in both Mashup cold-and warm-start scenarios. The implementations are available at: https://***/MengMeng3399/MGCL.
暂无评论