Variational Autoencoder and Graph Attention Root Cause Localization Model Based on Log Data and Graph Structure

When conducting root cause localization, converting data into graph structures for feature extraction can represent complex dependency relationships among data more comprehensively and deeply. A root cause localization model (Log-based Graph Root Cause Localization, LGRCL) based on log data and graph structure is proposed. Process and group multi-source, unstructured log data according to root cause location problem scenarios, and build a log dependency graph. Secondly, in order to more accurately capture the dependencies between data, for feature extraction of log dependency graphs, Variational Auto Encoder (VAE) is used to capture the potential features of the graph structure, and graph attention is introduced to focus on key nodes in the graph. The proposed LGRCL model is compared with the baseline model and better adapts to the complexity of log data and graph structure. The accuracy of the LGRCL model in locating the top five root causes is as high as 94.71%. The accuracy is 35.36% higher than the baseline model on average. The root cause location efficiency is 13.71% higher on average than the baseline model.