Causal-INSIGHT: Probing Temporal Models to Extract Causal Structure

2026-03-26

AI Summary

This work investigates how to extract the implicit directed, time-lagged causal dependency structure embedded within pretrained time series forecasting models to elucidate their decision-making rationale. To this end, the authors propose a model-agnostic post-hoc interpretability framework that, during inference, probes model responses through interventional input clamping to construct directed temporal influence signals. They further introduce Qbic, a sparsity-aware graph selection criterion that operates without requiring ground-truth graph labels, effectively balancing predictive fidelity with structural complexity. The approach is compatible with diverse time series model architectures and demonstrates strong generality across synthetic, simulated, and real-world benchmarks. Empirical evaluations show that the method achieves competitive structural accuracy while significantly improving the precision of temporal lag localization.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Understanding directed temporal interactions in multivariate time series is essential for interpreting complex dynamical systems and the predictive models trained on them. We present Causal-INSIGHT, a model-agnostic, post-hoc interpretation framework for extracting model-implied (predictor-dependent), directed, time-lagged influence structure from trained temporal predictors. Rather than inferring causal structure at the level of the data-generating process, Causal-INSIGHT analyzes how a fixed, pre-trained predictor responds to systematic, intervention-inspired input clamping applied at inference time. From these responses, we construct directed temporal influence signals that reflect the dependencies the predictor relies on for prediction, and introduce Qbic, a sparsity-aware graph selection criterion that balances predictive fidelity and structural complexity without requiring ground-truth graph labels. Experiments across synthetic, simulated, and realistic benchmarks show that Causal-INSIGHT generalizes across diverse backbone architectures, maintains competitive structural accuracy, and yields significant improvements in temporal delay localization when applied to existing predictors.

Problem

• causal structure
• temporal models
• time series
• model interpretation
• directed influence

Innovation

• Causal-INSIGHT
• model-agnostic interpretation
• temporal causal structure
• intervention-inspired input clamping
• Qbic