IoT Rule Generation With Cross-View Contrastive Learning and Perplexity-Based Ranking

Trigger-Action Platforms (TAPs) streamline task automation in Internet of Things (IoT) ecosystems through intuitive IF-THEN rules. However, the rapid expansion of TAP devices, combined with the diversity and overlap of their functionalities, presents significant challenges for users in formulating rules that accurately capture their intentions and effectively map these intents to the appropriate device actions. This paper presents TARGE (Trigger-Action Rule GEneration), a novel framework for generating IoT automation rules directly from natural language user intents. TARGE leverages Large Language Models (LLMs) to interpret user intents and employs cross-view contrastive learning to generate rule embeddings that capture TAP functionality and device relationships. Its ranking mechanism combines semantic consistency with LLM-derived perplexity to prioritize contextually coherent rules. Evaluated on a dataset of IFTTT rules, TARGE demonstrates robust performance across scenarios involving both well-defined and ambiguous user intents, consistently outperforming state-of-the-art methods by at least 46% in exact match accuracy and 35% in multi-rule recommendations.