Database Alarm Reasoning with Event Knowledge Graph Based on Graph Attention Network and Dynamic Pattern Matching

Temporal knowledge graph prediction has become one of the main methods for event knowledge graph prediction. However, existing methods face difficulties in distinguishing the importance of sequential information, making it challenging to locate key reasoning history information for specific prediction problems. Moreover, in database alarm root cause analysis, it is impossible to infer the cause of faults that do not have alarm phenomena in their historical information. Therefore, this paper proposes a database alarm root cause analysis method for Event Knowledge Graph based on Dynamic Pattern Matching and Graph Attention Network. Graph attention network is used as a neighborhood aggregator to simplify the temporal event graph, obtaining the corresponding features and information of the events and their involved entities at a fixed timestamp as its own vector representation, which serves as the basis for differentiating the importance of sequential information. Dynamic matching method and Gate Recurrent Unit (GRU) are applied to enhance the discrimination weight of sequential information and can infer the cause of faults that do not have alarm phenomena in their historical information. Experimental results on the Oracle dataset show that our model's MRR, Hit@3, and Hit@10 are 81.5%,83.7% and 83.4%, respectively, which outperform state-of-the-art models.