A Recommendation System for Trigger-Action Programming Rules via Graph Contrastive Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Trigger-Action Programming (TAP) automates Internet of Things (IoT) devices through user-defined rules.
• The increasing number of IoT devices makes manual rule creation complex and time-consuming.
• Existing TAP recommendation systems overlook user-rule associations and collaborative user information.

Purpose of the Study:

• To propose GCL4TAP, a novel graph contrastive learning-based recommendation system for TAP rules.
• To address the limitations of existing systems by incorporating cross-user rule relationships and user similarities.
• To enhance the efficiency and accuracy of automated rule discovery for IoT devices.

Main Methods:

• Developed DATA2DIV, a data partitioning method to represent cross-user rule relationships in a user-rule bipartite graph.
• Constructed a user-user graph to capture user similarities based on owned device categories and quantities.
• Utilized graph contrastive learning to generate low-dimensional vector representations for users and rules.

Main Results:

• GCL4TAP demonstrated superior performance compared to state-of-the-art methods in extensive experiments.
• The system effectively models collaborative information among users for improved rule recommendations.
• Experimental results on a real-world smart home dataset validate the proposed approach.

Conclusions:

• GCL4TAP offers a significant advancement in recommendation systems for trigger-action programming.
• The graph contrastive learning approach effectively captures complex relationships in IoT automation scenarios.
• The proposed system enhances the user experience by simplifying the automation of diverse IoT devices.