Hazardous Entity Recommendation for Safety Production Inspection Based on Multi-task Learning

Abstract

The large number and wide variety of hazardous entities is contradicted with the limited law enforcement strength of safety production, resulting in duplicate or missed inspections. In order to realize the key entity recommendation for safety production inspection, we introduce recommendation algorithms into this field. Data sparsity and cold start problems are inevitable in traditional recommendations, while knowledge graphs can be added as side information to solve the problems. Due to the strong sparsity of safety inspection data and the severe overfitting of existing models, we adaptively improve the multi-task learning algorithm by dividing the model into high layers and low layers and designing the structures respectively. A recommendation model based on multi-task learning and convolutional structures (CMKR) is proposed in this paper to provide better hazardous entity recommendations for safety production inspection. To solve the serious problem of over-fitting of the original multi-task learning algorithm, the convolutional neural network with the characteristics of sparse connection and weight sharing displaces a fully-connected multi-layer perceptron (MLP). ConvKB, an embedding model using CNN for the knowledge graph completion task is used at the high layers to improve the generalization ability of the model. In click-through rate prediction, ACC reaches 0.7061 and AUC reaches 0.7112 on hazardous entity recommendations of key sites. Compared with previous algorithms, the proposed method effectively controls the overfitting problem and improves the overall performance.