Causal Discovery in Multivariate Time Series through Mutual Information Featurization

2025-08-03

AI Summary

Conventional multivariate time-series causal discovery relies heavily on restrictive linear assumptions and fails under nonlinearity due to the breakdown of conditional independence tests. Method: This paper proposes a paradigm shift—reformulating causal inference as a pattern recognition task. We introduce TD2C, a supervised learning framework that leverages information-flow asymmetry as a learnable causal signal, eliminating reliance on conditional independence assumptions. TD2C integrates information-theoretic (e.g., mutual information) and statistical features (e.g., temporal dependence) to capture complex nonlinear dynamics. Contribution/Results: TD2C achieves zero-shot generalization to high-dimensional nonlinear systems. Evaluated on diverse synthetic and real-world benchmarks, it consistently outperforms state-of-the-art methods, demonstrating superior robustness, scalability, and generalization—particularly in high-dimensional and strongly nonlinear regimes.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Discovering causal relationships in complex multivariate time series is a fundamental scientific challenge. Traditional methods often falter, either by relying on restrictive linear assumptions or on conditional independence tests that become uninformative in the presence of intricate, non-linear dynamics. This paper proposes a new paradigm, shifting from statistical testing to pattern recognition. We hypothesize that a causal link creates a persistent and learnable asymmetry in the flow of information through a system's temporal graph, even when clear conditional independencies are obscured. We introduce Temporal Dependency to Causality (TD2C), a supervised learning framework that operationalizes this hypothesis. TD2C learns to recognize these complex causal signatures from a rich set of information-theoretic and statistical descriptors. Trained exclusively on a diverse collection of synthetic time series, TD2C demonstrates remarkable zero-shot generalization to unseen dynamics and established, realistic benchmarks. Our results show that TD2C achieves state-of-the-art performance, consistently outperforming established methods, particularly in high-dimensional and non-linear settings. By reframing the discovery problem, our work provides a robust and scalable new tool for uncovering causal structures in complex systems.

Problem

• Discover causal relationships in multivariate time series
• Overcome limitations of traditional linear and independence-test methods
• Recognize complex causal patterns via information-theoretic descriptors

Innovation

• Mutual information featurization for causal discovery
• Supervised learning framework TD2C
• Zero-shot generalization with synthetic training