Dynamic Sparse Causal-Attention Temporal Networks for Interpretable Causality Discovery in Multivariate Time Series

2025-07-12

AI Summary

Addressing the challenges of identifying lagged causal relationships, spurious connections, and poor interpretability in multivariate time series (MTS), this paper proposes an interpretable causal discovery framework. Methodologically, it integrates dilated temporal convolution with a dynamic sparse causal attention mechanism, incorporates RMSNorm and causal masking to enhance training stability, and employs statistical shuffle testing to improve robustness. Crucially, we design an adaptive thresholding strategy to accurately detect latent mediators and lagged causal factors. Experiments on financial and marketing datasets demonstrate that our method significantly outperforms TCDF, GCFormer, and CausalFormer—achieving superior accuracy in causal lag estimation, enhanced robustness to noise, and reduced false detection rates. Moreover, it successfully uncovers the mediating effect of advertising on consumer behavior and identifies macroeconomic variables’ lagged impacts on financial markets.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Action, Change, and Causality

Application Category

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

Abstract

Understanding causal relationships in multivariate time series (MTS) is essential for effective decision-making in fields such as finance and marketing, where complex dependencies and lagged effects challenge conventional analytical approaches. We introduce Dynamic Sparse Causal-Attention Temporal Networks for Interpretable Causality Discovery in MTS (DyCAST-Net), a novel architecture designed to enhance causal discovery by integrating dilated temporal convolutions and dynamic sparse attention mechanisms. DyCAST-Net effectively captures multiscale temporal dependencies through dilated convolutions while leveraging an adaptive thresholding strategy in its attention mechanism to eliminate spurious connections, ensuring both accuracy and interpretability. A statistical shuffle test validation further strengthens robustness by filtering false positives and improving causal inference reliability. Extensive evaluations on financial and marketing datasets demonstrate that DyCAST-Net consistently outperforms existing models such as TCDF, GCFormer, and CausalFormer. The model provides a more precise estimation of causal delays and significantly reduces false discoveries, particularly in noisy environments. Moreover, attention heatmaps offer interpretable insights, uncovering hidden causal patterns such as the mediated effects of advertising on consumer behavior and the influence of macroeconomic indicators on financial markets. Case studies illustrate DyCAST-Net's ability to detect latent mediators and lagged causal factors, making it particularly effective in high-dimensional, dynamic settings. The model's architecture enhanced by RMSNorm stabilization and causal masking ensures scalability and adaptability across diverse application domains

Problem

• Discover causal relationships in multivariate time series data
• Improve accuracy and interpretability of causal discovery
• Handle noisy, high-dimensional dynamic environments effectively

Innovation

• Dynamic sparse attention for causality discovery
• Dilated convolutions capture multiscale dependencies
• Adaptive thresholding reduces spurious connections