Locating Faulty Applications via Semantic and Topology Estimation
Shuyi Niu, Jiawei Jin, Xiutian Huang, Yonggeng Wang, Wenhao Xu, Youyong Kong · 2023
With the explosion of Internet product users, how to locate the faulty ones from numerous back-end applications after a customer complaint has become an essential issue in improving user experience. However, existing solutions mostly rely on manual testing to infer the fault, severely limiting their efficiency. In this paper, we transform the problem of locating faulty applications into two subproblems and propose a fully automated framework. We design a scorecard model in one stage to evaluate the semantic relevance between applications and customer complaints. Then in the other stage, topology graphs that reflect the actual calling relationship and engineering connection relationship between applications are utilized to evaluate the topology relevance between applications. Specifically, we employ a multi-graph co-learning framework constrained by consistency-independence loss and an engineering-theory-driven clustering strategy for the unsupervised learning of graphs. With semantic and topology relevance, we can comprehensively locate relevant faulty applications. Experiments on the Alipay dataset show that our method gains significant improvements in both model performance and efficiency.