Explainable Recommendation for Hazard Inspection Reasoning Through Knowledge Graph

In the process of hazards inspecting in safety production, the number and types of hazard entities are often vast and varied. Introducing of existing recommendation algorithms may lead to low recommendation quality and lack of explain ability. To address these challenges, achieving dual improvement in accuracy and reliability of hazard inspection recommendations. We propose three metrics that represent the explanation quality of recommendation results: Inspection Recency of Hazard, Risk Level of Entity and Utilization Rate of Graph. And then we propose two optimization methods on model inner-training and post-training: Pruning-Strategy Optimization and Re-Rank Optimization. In hazards inspection dataset of safety production, hazard inspection recommendation NDCG@10 reaches 0.433(5.1 % higher than baseline) and overall path explanation metric scores reaches 1.928(74.8% higher than baseline). Compared with previous algorithms, our methods achieve higher recommendation performance and explainability quality.