An Explainable Recommendation Method for Artificial Intelligence of Things Based on Reinforcement Learning With Knowledge Graph Inference

Abstract

In the realm of consumer electronics, the integration of knowledge graphs with causal inference significantly advances recommendation systems within the Artificial Intelligence of Things (AIoT). This paper introduces a novel method that addresses the limitations of traditional AIoT-based systems, which tend to prioritize correlation over causality and demonstrate limitations in navigating inference paths across knowledge graphs. A reinforcement learning based knowledge graph model is designed to enhance interpretability and trustworthiness in recommendation processes. A soft reward strategy is employed within a Markov decision process, utilizing a multi-hop scoring function to ensure rational outcome assessments. Additionally, a graph search algorithm with user-conditional action pruning is incorporated to facilitate efficient and accurate sampling of inference paths. Experimental results indicate significant improvements in key performance metrics such as normalized discounted cumulative gain, recall, hit ratio, and precision ratio over existing methods. Furthermore, interpretable reasoning paths are provided, establishing a new benchmark in the AIoT-driven recommendation landscape for consumer electronics.