Discovering causal structures in multivariate time series (MTS) is critical in domains such as finance and neuroscience, yet remains challenging in practice due to three distinct obstacles. First, MTS are routinely corrupted by missing values, where naive impute-then-discover pipelines distort the underlying data distribution and yield spurious causal graphs. Second, identifying instantaneous effects is difficult even on complete data, and missingness further entangles imputation with graph recovery. Third, real-world data frequently exhibits heteroscedastic noise with variance driven by both instantaneous and lagged causes, violating the core assumptions of standard discovery methods. Existing approaches fail to address these challenges concurrently, typically presuming complete data, homoscedastic noise, or strictly time-lagged relationships. Crucially, we show through ablation studies that resolving these issues sequentially, rather than jointly, leads to a substantial degradation in performance. To overcome this, we introduce Cheesefill, an optimal transport framework for causal structure learning under missingness and heteroscedastic noise. Cheesefill parameterizes causal mechanisms using conditional normalizing flows and jointly learns a stochastic correction map that refines naive imputations into trajectories consistent with the inferred causal mechanisms. Theoretically, we prove that optimizing over stochastic correction maps is equivalent to a conditional kernel reformulation of the Kantorovich problem, from which our neural parameterization yields a tractable objective. Extensive experiments on synthetic and real-world benchmarks demonstrate that Cheesefill consistently outperforms existing baselines across diverse missingness regimes.
Figures & tables
Figure 1: Cheesefill overview. The incomplete MTS dataset X is first processed by a parametric imputer with parameters ζ . A stochastic correction map Tψ refines X~ into a sample Y consistent with the structural equation model marginal QΘ . TGNN and CNF, models causal mechanisms and heteroscedastic noise. Joint optimisation over (Θ,ψ,ζ) recovers the temporal causal graph G∗ .
Figure 2: Temporal graph neural network (TGNN) used by Cheesefill.
F1 (%) ↑
Ratio ↓
Precision (%) ↑
Recall (%) ↑
Map
CNF
Inst.
Lag
Inst.
Lag
Inst.
Lag
Inst.
Lag
Off
Off
0.0
47.0
1.00
0.52
0.0
32.0
0.0
89.0
Off
On
0.0
0.0
1.00
1.00
0.0
0.0
0.0
0.0
On
Off
8.0
45.0
0.91
0.54
5.0
32.0
17.0
77.0
On
On
94.2
94.5
0.05
0.05
89.0
91.0
100.0
98.0
Table 1: Ablation of the stochastic correction map and CNF. Inst. and Lag denote instantaneous and lagged edges, respectively. F1, precision, and recall are expressed as percentages. The final row corresponds to full Cheesefill.
Figure 3: MCAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 4: F1 score on NetSim across missingness levels.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Figure taken from Gamella et al. [ 12 ] . Diagrams of the wind tunel causal chamber and its main components, including the amplification circuit that drives the speaker of the wind tunnel. The variables measured by the chamber are displayed in black math print. Sensor measurements are denoted by a tilde. Manipulable variables, that is, actuators and sensor parameters, are shown in bold symbols.
Figure 7
Figure 8: MCAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 9: MCAR MTS with 40 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 10: MAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 11: MAR MTS with 40 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 12: MNAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 2 and 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 13: MCAR and MAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 14: MCAR MTS with 20 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 15: MAR MTS with 20 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 16: MCAR and MAR MTS with 40 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 17: MCAR MTS with 40 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 18: MAR MTS with 40 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 19: MNAR MTS with 20 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 20: MNAR MTS with 20 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 21: MNAR MTS with 40 nodes. F1 score (top, higher is better) and NHD ratio (bottom, lower is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Figure 22: MNAR MTS with 40 nodes. SHD (top, lower is better), Precision (middle, higher is better) and Recall (bottom, higher is better) for instantaneous and time lagged edges, evaluated under 3 maximum lags as the missing rate increases from 10% to 50% .
Structural causal models for time series recover contemporaneous and lagged effects, but most methods require complete observation windows and become misspecified when samples are missing. We introduce ITSY, the first continuous-optimization method for causal discovery from irregular time series under a linear model. ITSY reformulates the structural equation so that prediction uses the nearest available history rather than the possibly missing current slice, and jointly imputes missing values while learning both graphs. A weighted reconstruction objective corrects the noise transformation induced by this reformulation. Across synthetic regimes varying missingness, scale, graph density, and noise, and on a real world benchmark, ITSY consistently improves graph recovery over representative SCM-based baselines, demonstrating the effectiveness of the proposed method. The results establish a focused solution for irregular linear first-order dynamics and clarify the assumptions required for nonlinear or higher-order extensions.
Wenbo Xu, Yue He, Yunhai Wang +2
Renmin University of China, China · Zhejiang University, China
Causal discovery with instantaneous effects in multivariate time series is challenging, as the instantaneous structure must be acyclic. Prior methods enforce this by either separating instantaneous and lagged estimation into multi-stage pipelines or imposing algebraic acyclicity constraints via complex augmented Lagrangian optimization, both of which incur high computational cost. In this work, we propose a different approach: we learn a differentiable permutation of variables using the Gumbel--Sinkhorn operator and triangularize the instantaneous coefficient matrix of a Structural Vector Autoregressive (SVAR) model in the learned order. This converts acyclicity from a hard constraint into a parameterization and keeps it valid throughout optimization. In doing so, our method enables unified, continuous optimization with gradient-based learning, leading to improved efficiency in time--series causal discovery. Across three real-world benchmarks, our method achieves the best overall performance compared with 12 baselines in both discovery accuracy and efficiency. On the large-scale benchmark, it further demonstrates strong scalability, achieving more than a 6x speedup over competing methods.
Heteroscedasticity -- where the variance of a variable changes with other variables -- is pervasive in real data, and elucidating why it arises from the perspective of statistical moments is crucial in scientific knowledge discovery and decision-making. However, standard causal discovery does not reveal which causes act on the mean versus the variance, as it returns a single moment-agnostic graph, limiting interpretability and downstream intervention design. We propose a Bayesian, moment-driven causal discovery framework that infers separate mean and variance causal graphs from observational heteroscedastic data. We first derive the identification results by establishing sufficient conditions under which these two graphs are separately identifiable. Building on this theory, we develop a variational inference method that learns a posterior distribution over both graphs, enabling principled uncertainty quantification of structural features (e.g., edges, paths, and subgraphs). To address the challenges of parameter optimization in heteroscedastic models with two graph structures, we take a curvature-aware optimization approach and develop a prior incorporation technique that leverages domain knowledge on node orderings, improving sample efficiency. Experiments on synthetic, semi-synthetic, and real data show that our approach accurately recovers mean and variance structures and outperforms state-of-the-art baselines.
Yoichi Chikahara
Communication Science Laboratories, NTT, Inc., Kyoto, Japan