[論文レビュー] Causal explanations of outliers in systems with lagged time-dependencies

Philipp Schwarz, Johannes Oberpriller+1 more|arXiv (Cornell University)|Feb 4, 2026

Model Reduction and Neural Networks被引用数 0

ひとこと要約

tldr: Adapts causal root-cause analysis (CRCA) to lagged time-dependent systems to localize root-causes of anomalies using truncated and non-truncated lag models, benchmarked on a manufacturing-energy DGP.

ABSTRACT

Root-cause analysis in controlled time dependent systems poses a major challenge in applications. Especially energy systems are difficult to handle as they exhibit instantaneous as well as delayed effects and if equipped with storage, do have a memory. In this paper we adapt the causal root-cause analysis method of Budhathoki et al. [2022] to general time-dependent systems, as it can be regarded as a strictly causal definition of the term "root-cause". Particularly, we discuss two truncation approaches to handle the infinite dependency graphs present in time-dependent systems. While one leaves the causal mechanisms intact, the other approximates the mechanisms at the start nodes. The effectiveness of the different approaches is benchmarked using a challenging data generation process inspired by a problem in factory energy management: the avoidance of peaks in the power consumption. We show that given enough lags our extension is able to localize the root-causes in the feature and time domain. Further the effect of mechanism approximation is discussed.

研究の動機と目的

• Extend CRCA to general time-dependent systems with lagged effects and memory.
• Investigate truncation strategies to manage infinite lag dependencies while preserving causal mechanisms.
• Benchmark the approach on a data-generating process modeling energy peaks in an industrial plant.
• Assess time localization capability and the impact of mechanism approximation on attribution.
• Provide guidance on when to use truncated vs. non-truncated lag CRCA in practice.

提案手法

• Model the system as a causal structural model with lagged dependencies and a target variable.
• Unfold the time-lagged graph up to lag L and truncate dangling parents to obtain a tractable model.
• Apply CRCA attribution using information-theoretic calibration (IT-Score) and Shapley value decomposition to assign attributions to (node, lag) pairs.
• Compare truncated lag-L CRCA, non-truncated lag-L CRCA, and a simple heuristic attribution method.
• Inject controlled Peak scenarios in a DGP that mimics factory energy management to evaluate localization and time accuracy of attributions.
• Discuss computational considerations and memory scaling for the attribution computation.

実験結果

リサーチクエスチョン

• RQ1 How can CRCA be extended to handle lagged time-dependencies and memory in dynamical systems?
• RQ2 What is the impact of truncating the infinite lag dependency graph on the accuracy and locality of root-cause attributions?
• RQ3 Do truncated and non-truncated lag CRCA methods localize root-causes in time and feature space for various injection scenarios?
• RQ4 How does mechanism approximation at dangling nodes affect attribution results compared to preserving the true mechanisms?
• RQ5 What guidance can be provided on selecting lag L for effective root-cause analysis in energy-system-like settings?

主な発見

• CRCA with lag extension can localize root-causes in both feature and time domains given sufficient lag depth.
• Truncated lag models preserve the overall mechanism up to lag L and can attribute peaks to fault sources within the considered window.
• Non-truncated models sometimes outperform truncated ones at larger L due to longer effective lag reasoning, but can suffer from mechanism approximation at dangling nodes.
• Hit-rate for correct attribution (HR@3) improves with increasing L for certain injections that have delayed effects (e.g., soc loss, temperature surge, work arrival).
• Some injections yield immediate effects (e.g., grid noise) with limited or no gains from longer L; across injections, the choice of aggregation (max vs. sum) impacts locality slightly.
• A simple heuristic baseline underperforms compared with CRCA variants.