Temporal Causal Inference for Web Application Monitoring: Structure Learning from Sequential Performance Data with Delayed Effects

Abstract

Web application monitoring systems collect vast amounts of sequential performance data, yet traditional approaches struggle to identify complex temporal causal relationships, particularly when effects manifest with delays. This paper introduces a novel framework for temporal causal inference in web application monitoring that addresses the challenge of learning causal structures from sequential performance data exhibiting delayed effects. Our approach integrates Granger causality principles with modern structure learning techniques to construct Directed Acyclic Graphs (DAGs) that capture both instantaneous and lagged causal relationships among performance metrics. We propose a hybrid methodology combining constraint-based and score-based methods specifically designed to handle the non-stationary nature of web performance data and the presence of time-varying confounders. The framework employs a sliding window approach for dynamic causal structure discovery, enabling real-time adaptation to changing system behaviors. Experimental validation using both synthetic datasets and real-world web application traces demonstrates that our method achieves superior performance in identifying true causal relationships compared to baseline approaches, with particular improvements in detecting delayed causal effects. The proposed framework reduces false discovery rates by approximately 35% while maintaining high sensitivity for genuine causal links, even under conditions of high-dimensional data and limited sample sizes. These findings have significant implications for automated root cause analysis, predictive maintenance, and performance optimization in modern web infrastructure.