Coarse-to-Fine Learning of Dynamic Causal Structures
Dezhi Yang, Qiaoyu Tan, Carlotta Domeniconi, Jun Wang, Lizhen Cui, Guoxian Yu
TL;DR
This paper introduces DyCausal, a novel framework for learning fully dynamic causal structures in time series data, effectively capturing evolving causal relationships with improved efficiency and stability.
Contribution
DyCausal employs convolutional networks and linear interpolation to recover fine-grained, time-varying causal graphs, addressing limitations of stationary assumptions in causal discovery.
Findings
Outperforms existing methods on synthetic datasets
Demonstrates stability and efficiency on real-world data
Effectively captures complex dynamic causal relationships
Abstract
Learning the dynamic causal structure of time series is a challenging problem. Most existing approaches rely on distributional or structural invariance to uncover underlying causal dynamics, assuming stationary or partially stationary causality. However, these assumptions often conflict with the complex, time-varying causal relationships observed in real-world systems. This motivates the need for methods that address fully dynamic causality, where both instantaneous and lagged dependencies evolve over time. Such a setting poses significant challenges for the efficiency and stability of causal discovery. To address these challenges, we introduce DyCausal, a dynamic causal structure learning framework. DyCausal leverages convolutional networks to capture causal patterns within coarse-grained time windows, and then applies linear interpolation to refine causal structures at each time step,…
Peer Reviews
Decision·ICLR 2026 Poster
- The paper is well-written and easy to follow. - The paper motivates and undertakes an important and under-discussed problem of time-varying causal structure learning. - Relative to some prior work (e.g., DyCast), which only learns time-varying instantaneous graphs, DyCausal simultaneously learns time-varying lagged relationships. This seems to significantly improve performance. - The experiments clearly demonstrate strong performance relative to relatively recent and strong baselines.
I'm not totally convinced by the analysis of the revised log-determinant penalty. One reason why is that the benefits of DAGMA lie partially in nice optimization properties in a barrier method approach -- it's a priori possible but not obvious to me that the proposed version would perform better in the general structure learning setting. The ablation in C.10 is nice in theory, but the graph in Figure A9 seems to moreso suggest some catastrophic failure than poor optimization.
1. The paper extends causal structure learning to a general setting of fully dynamic causality, where both instantaneous and lagged dependencies evolve over time. This broadens the applicability of causal discovery methods to more realistic time-varying systems. 2. The proposed approach appears technically sound.
1. The paper lacks significant technical innovation. Most components of the proposed framework appear to be incremental adaptations of existing techniques rather than conceptually new contributions. The use of convolutional networks for coarse-grained temporal modeling and the acyclicity constraint are engineering extensions that do not substantially advance the methodological frontier. Moreover, the theoretical results are mostly straightforward derivations of [1], without introducing new analy
1. The authors addressed the limitations of previous methods by considering a fully dynamic causality. 2. They offered clear motivation for their DyCausal algorithm and provided a new form of acyclic constraint. 3. The authors tested DyCausal's performance and compared it with previous methods.
1. The proposed refined dynamic causal graph interpolation is based on the assumption that the causal relationship changes smoothly and follows a linear law, which is a very strong prior assumption. 2. The acyclicity constraint in the question appears to be a variant of the DAGMA constraint, and I am skeptical as to whether it is sufficient to be considered an innovative point. 3. Learning causal relationships requires theoretical guarantees of identifiability. The proposed CNN method employs
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsBayesian Modeling and Causal Inference · Machine Learning in Healthcare · Time Series Analysis and Forecasting