LLM4TAP: LLM-Enhanced TAP Rule Recommendation

Trigger-action programming (TAP) is an Internet of Things (IoT) paradigm that enables nonprofessional end-users to automate smart devices by formulating rules, such as “IF you leave home, THEN turn off lights.” As the number of possible rules increases, manually browsing these rules becomes increasingly time-consuming for users. Recently, graph-based recommendation systems have shown promise in automatically suggesting rules, yet they face two issues. First, these studies struggle to identify and differentiate users’ demands (e.g., turning off lights) and intentions (e.g., energy saving). Second, they overlook the issue of sparse user-rule interactions. In this article, we propose LLM4TAP, a large language model (LLM) enhanced TAP rule recommendation framework, to address these issues. Prior to LLM4TAP, a user-rule graph is constructed to represent the interactions between users and rules. Within LLM4TAP, singular value decomposition is first employed to generate an augmented graph, strengthening global structural relationships between users and rules. Next, the reasoning capabilities of LLMs are utilized to infer users’ demands and intentions from the textual descriptions of rules and user-rule interactions, producing representations of these inferred demands and intentions. Finally, a dual representation alignment method is introduced, integrating user demands and intentions derived from LLMs with the global structural information from the augmentation graph within a contrastive learning framework to enhance representation performance. Extensive experiments demonstrate the effectiveness of LLM4TAP, achieving the maximum improvements of 8.96% and 4.72% over the strongest compared methods on the IFTTT and Wyze datasets, respectively.