TAP-AHGNN: An Attention-Based Heterogeneous Graph Neural Network for Service Recommendation on Trigger-Action Programming Platform

Trigger-Action Programming (TAP) is a popular IoT programming paradigm that enables users to connect IoT services and automate IoT workflows by creating if-trigger-then-action rules. However, with the increasing number of IoT services, specifying trigger and action services to compose TAP rules becomes progressively challenging for users due to the vast search space. To facilitate users in programming, a novel method named TAP-AHGNN is proposed to recommend feasible action services to auto-complete the rule based on the user-specified trigger service. Firstly, a heterogeneous TAP knowledge graph is designed, from which five meta-paths can be extracted to construct services' neighborhoods. Then, the model incorporates a multi-level attention-based heterogeneous graph convolution module that selectively aggregates neighbor information, and a transformer-based fusion module that enables the integration of multiple types of features. With the two modules mentioned before, the final representations of services can capture both semantic and structural information, which helps generate better recommendation results. Experiments on the real-world dataset demonstrate that TAP-AHGNN outperforms the most advanced baselines at HR@k, NDCG@k and MRR@k. To the best of our knowledge, TAP-AHGNN is the first method for service recommendation on TAP platforms using the heterogeneous graph neural network technique.