Causal Aware Anomaly Detection for Modeling the Evolution of Backend Operational States in Ultra Large Scale Online Services

Authors
  • I-Hsien Lai

    University of Wisconsin-Madison

    Author

Abstract

To address the challenges of highly dynamic operational states, highly heterogeneous multi-source observations, and hidden anomaly causes that are susceptible to surface disturbances in ultra-large-scale online service backend systems, a causal perception anomaly detection method based on operational state evolution modeling is proposed. This method is grounded in multi-dimensional operational information such as logs, metrics, links, and topology. First, it constructs a state representation capable of characterizing the temporal correlations and structural dependencies of the backend system. Based on this, a state evolution coding mechanism characterizes the potential changes in the system's continuous operation, enhancing the perception of anomaly accumulation effects and implicit propagation behaviors. Subsequently, causal component extraction and intervention consistency constraints are introduced to distinguish between key driving information and background disturbance information in complex observations, reducing the interference of non-critical fluctuations on the anomaly identification process and thus improving the model's stability in identifying anomaly patterns in complex backend scenarios. Furthermore, evolutionary bias modeling is used to characterize the difference between the predicted state and the actual state, extending anomaly detection from static observation discrimination to structured identification oriented towards dynamic operational processes. The proposed method can collaboratively model state continuity, dependency propagation, and causal relationships within a unified framework, providing a solution for complex online service backend anomaly detection that combines expressive and discriminative capabilities. Related results demonstrate that this method exhibits good overall performance and application value in ultra-large-scale online service backend anomaly detection tasks.

Cover Image
Downloads
Published
2025-07-01
Section
Articles