CauFR-TS: Causal Time-Series Identifiability via Factorized Representations

Abstract

Causal discovery from multivariate time series is a fundamental problem for interpretable modelling, causality-aware downstream analysis, and intervention-driven simulation. Recent neural approaches commonly rely on shared latent embeddings to capture temporal dynamics and utilize them for causal structure estimation and downstream prediction. We formally establish that such shared encoders entangle distinct causal mechanisms into a unified latent manifold, which exhibits fundamental theoretical limitations of structural non-identifiability and conditional independence assumptions required for Granger causality. To address these issues, we propose CauFR-TS, a recurrent variational framework that enforces mechanism modularity through dimension-wise encoders and ensures mediation of all cross-variable dependencies through structured latent aggregation. Furthermore, we address the instability of heuristic thresholding in continuous relaxation methods by proposing an adaptive, data-driven unsupervised link selection strategy based on decoder weight distribution. Empirical evaluation on synthetic and in silico biological benchmarks demonstrates that CauFR-TS outperforms recent baselines in graph recovery metrics while preserving competitive probabilistic forecasting performance.