Identifying Outcome-Oriented Root Causes via Cross Regression

Haotian Wang, Hao Zou, Haoxuan Li, Yang Shi, Yuanxing Zhang, Kun Kuang, Wenjing Yang, Xinwang Liu, Peng Cui

Keywords: Root Causes, Regression Theory

TL;DR: This paper presents a regression-based approach for identifying all outcome-oriented,necessary root-causes.

Abstract: Root Cause Analysis (RCA) in complex and interconnected systems exhibits significant importance in fields such as microservice maintenance, and supply-chain management. By identifying every intervened variable, existing RCA methods have achieved remarkable progress in localizing and fixing anomalies. However, people may be more interested and focused on those intervened variables that produced effects on a specific outcome, rather than the intervened variables that do not necessarily affect that outcome. This raises concerns on redundant localizing and extra efforts in fixing the anomalies. To fill this gap, we study a novel and challenging problem, termed as Outcome-Oriented Root-Cause Analysis (OORCA), aiming to identify all intervened ancestor variables of the outcome variable. To handle the proposed OORCA problem, we then propose the Cross-Regressing-based Root Cause (CRRC) framework by cross-regressing observational (normal) and interventional (abnormal) data on the outcome variable. Theoretically, our identifiability analysis prove that the proposed CRRC can capture all outcome-oriented root-causes, and our asymptotic analysis offers tractable and informative criteria in the finite-sample regime. Extensive experiments across three benchmarks with 13 competitive baselines highlight the superiority of CRRC in both accuracy and running efficiency.

Primary Area: causal reasoning

Submission Number: 8165