Flow-Based Non-stationary Temporal Regime Causal Structure Learning
Abdellah Rahmani, Pascal Frossard · 2025
Read Flow-Based Non-stationary Temporal Regime Causal Structure Learning by Abdellah Rahmani, Pascal Frossard (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-17065.txt.
Introduction
Abdellah Rahmani and Pascal Frossard learn cause webs in time data that shift by regime in a 2025 paper. Many data runs split into blocks, each with its own cause web and blend rule. Block edges are not known ahead, which makes the learn task hard on two counts.
Noise can also bend or spread in rich ways that break old tools. Past cause hunts tend to posit calm runs or bell-shaped noise with fixed spread. Their flow-based tool is built to face both shifts and rich noise at once, the team notes.
Catalog record
| CREATOR | Abdellah Rahmani, Pascal Frossard |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 137 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-17065
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Flow-Based Non-stationary Temporal Regime Causal Structure Learning Abdellah Rahmani, Pascal Frossard Abstract Understanding causal relationships in multivariate time series is crucial in many scenarios, such as those dealing with financial or neurological data. Many such time series exhibit multiple regimes, i.e., consecutive temporal segments with a priori unknown boundaries, with each regime having its own causal structure. Inferring causal dependencies and regime shifts is critical for analyzing the underlying processes. However, causal structure learning in this setting is challenging due to (1) non-stationarity, i.e., each regime can have its own causal graph and mixing function, and (2) complex noise distributions, which may be non-Gaussian or heteroscedastic. Existing causal discovery approaches cannot address these challenges, since generally assume stationarity or Gaussian noise with constant variance. Hence, we introduce FANTOM, a unified framework for causal discovery that handles non-stationary processes along with non-Gaussian and heteroscedastic noises. FANTOM simultaneously infers the number of regimes and their corresponding indices and learns each regime’s Directed Acyclic Graph. It uses a Bayesian Expectation Maximization algorithm that maximizes the evidence lower bound of the data log likelihood. On the theoretical side, we prove, under mild assumptions, that temporal heteroscedastic causal models, introduced in FANTOM’s formulation, are identifiable in both stationary and non-stationary settings. In addition, extensive experiments on synthetic and real data show that FANTOM outperforms existing methods. 1 Introduction Causal structure learning from multivariate time series (MTS) is a fundamental problem with diverse applications in traffic modeling ( cheng2023causaltime ) , biology ( sachs2005causal ) , climate science ( runge2019detecting ) , or healthcare ( shen2020challenges ) . However, identifying causal relationships in MTS poses several challenges. First, real-world time series are often non-stationary, exhibiting multiple unknown regimes, each potentially governed by different causal relationships. Examples include changing dependencies across climate conditions ( karmouche2023regime ) , financial markets ( huang2020causal ) , and epileptic seizure stages ( wang2024ictal ) . Second, many MTS display complex noises, e.g., non-Gaussian or even heteroscedastic noise, whose variance depends on both instantaneous and lagged causes. This occurs in fMRI data ( seiler2017multivariate ) , EEG measurements ( huizenga1995equivalent ) , or financial data ( hamilton1994autoregressive ) . Recent causal discovery methods capture linear and nonlinear interactions with instantaneous and lagged effects ( runge2018causal ; pamfil2020dynotears ; sun2021nts ) . More recently, Gong et al. ( gong2022rhino ) and Wang et al. ( wang2023neural ) explored structural equation models for a single stationary regime governed by a one causal graph with historically dependent noise, where noise variance depends solely on time-lagged variables, neglecting heteroscedasticity. Existing multi-regime methods include RPCMCI ( saggioro2020reconstructing ) , which identifies only linear, time-lagged interactions and requires prior knowledge of regime numbers and transitions; and CD-NOD ( huang2020causal ) , which handles causal discovery from non stationary MTS, but is limited to homoscedastic noise, cannot infer individual causal graph for each regime, and is incapable of identifying recurring regimes. CASTOR ( rahmanicausal ) , infers both regime indices and separate causal graphs per regime, accommodating instantaneous and lagged causal relationships, yet it is still restricted to normal noise. Consequently, the previous methods cannot jointly infer the number of regimes, their indices, their causal graphs, nor effectively manage either non-Gaussian or heteroscedastic noises. Figure 1: Illustration of FANTOM processing a MTS with two ground truth regimes ( K=2 ). The algorithm recovers the regime indices \mathcal{I}^{*}_{1} and \mathcal{I}^{*}_{2} and learns a temporal causal graph for each regime (dashed edges represent time-lagged links; solid arrows indicate instantaneous links). In the E-step, posterior probabilities p\bigl(z_{t}=r\mid\boldsymbol{x}_{t},\boldsymbol{x}_{<t}\bigr) are estimated, where z_{t}=r means \boldsymbol{x}_{t} belongs to regime r . The M-step then infers causal graphs within each regime. Here, N_{w} is the number of regimes that converges to K=2 . To address these limitations, we propose FANTOM, a new framework for Structural Equation Models (SEMs) in multi-regime MTS under either non-Gaussian or heteroscedastic noises. FANTOM is, to the best of our knowledge, the first method to handle heteroscedasticity in both stationary and non-stationary MTS, as well as non-Gaussianity in non-stationary settings. Given a MTS with multiple regimes, FANTOM simultaneously learns each regime’s causal graph and mixing function, determines the number of regimes, and infers their indices (Figure 1 ). It uses a Bayesian Expectation Maximization (BEM) ( friedman2013bayesian ) procedure to optimize the evidence lower bound (ELBO), alternatively assigning regime indices (Expectation step) and inferring causal relationships in each regime (Maximization step). Unlike Gaussian-based approaches, FANTOM employs conditional normalizing flows ( friedman2013bayesian ) to handle complex distributions and compute regime membership probabilities. It uses Bayesian structure learning that averages across all plausible graphs and naturally filters out spurious edges. Under mild assumptions, we prove that temporal heteroscedastic causal models are identifiable for both stationary and multi-regime MTS. Across extensive comparisons with existing multi-regime causal discovery methods, we show that FANTOM consistently achieves superior performance in structure learning and regime detection. Moreover, it outperforms stationary models, even when they are provided with ground-truth regime partitions, on synthetic and two real-world datasets. The main contributions of this paper can be summarized as follows: • We introduce FANTOM, a unified framework for causal discovery in multi-regime MTS that handles both homoscedastic non-Gaussian and heteroscedastic noises while simultaneously discovering the number of regimes, their indices, and their corresponding causal graphs. • Under mild assumptions in causal discovery, we prove identifiability of the temporal heteroscedastic causal models in the stationary case and show that the number of regimes, their indices, and their graphs are identifiable (up to permutation) in the non-stationary setting. • We demonstrate, via extensive experiments, that FANTOM outperforms state-of-the-art methods on both synthetic and real-world datasets. Related work. Many works tackle causal structure learning from stationary MTS , Granger causality is the primary approach used for this purpose ( lowe2022amortized ; bussmann2021neural ) . However, it is unable to accommodate instantaneous effects. DYNOTEARS ( pamfil2020dynotears ) , learns instantaneous and time lagged structures and leverages the acyclicity constraint, introduced by Zheng et al. in ( zheng2018dags ) , to turn the DAG learning problem to a purely continuous optimization problem. However, DYNOTEARS is limited to linear SEMs. Runge et al. ( runge2020discovering ) proposed a two-stage algorithm PCMCI+ that can scale to large time series, PCMCI+ is able also to handle non linear relationships. Neverthless, DYNOTEARS and PCMCI+ are restricted to homoscedastic noises where variance is a constant over time. For this reason, Rhino ( gong2022rhino ) and SCOTCH ( wang2023neural ) introduced models that tackle stationary MTS with historical noise, where the noise variance is a function of solely time lagged parents. However Rhino, DYNOTEARS and PCMCI+, can not handle heteroscedastic noise and are limited to stationary MTS. Several studies have sought to tackle the challenge of causal discovery in non-stationary MTS ( huang2020causal ; gunther2024causal ; saggioro2020reconstructing ; mameche2025spacetime ) . Remarkably, Huang et al. ( huang2020causal ) address the setting of time series composed of different regimes by modulating causal relationships through a regime index. CD-NOD detects change points and outputs a single summary causal graph, but it overlooks recurring regimes and provides neither regime-specific graphs nor their count. RPCMCI ( saggioro2020reconstructing ) provides a graph for each regime, yet it assumes prior knowledge of the number of regimes, restricts edges to