Abstract
Predicting causal structure from time series data is crucial forunderstanding complex phenomena in physiology, brain connectivity, climatedynamics, and socio-economic behaviour. Causal discovery in time series ishindered by the combinatorial complexity of identifying true causalrelationships, especially as the number of variables and time points grow. Acommon approach to simplify the task is the so-called ordering-based methods.Traditional ordering methods inherently limit the representational capacity ofthe resulting model. In this work, we fix this issue by leveraging multiplevalid causal orderings, instead of a single one as standard practice. Wepropose DOTS (Diffusion Ordered Temporal Structure), using diffusion-basedcausal discovery for temporal data. By integrating multiple orderings, DOTSeffectively recovers the transitive closure of the underlying directed acyclicgraph, mitigating spurious artifacts inherent in single-ordering approaches. Weformalise the problem under standard assumptions such as stationarity and theadditive noise model, and leverage score matching with diffusion processes toenable efficient Hessian estimation. Extensive experiments validate theapproach. Empirical evaluations on synthetic and real-world datasetsdemonstrate that DOTS outperforms state-of-the-art baselines, offering ascalable and robust approach to temporal causal discovery. On syntheticbenchmarks ($d{=}\!3-\!6$ variables, $T{=}200\!-\!5{,}000$ samples), DOTSimproves mean window-graph $F1$ from $0.63$ (best baseline) to $0.81$. On theCausalTime real-world benchmark ($d{=}20\!-\!36$), while baselines remain thebest on individual datasets, DOTS attains the highest average summary-graph$F1$ while halving runtime relative to graph-optimisation methods. Theseresults establish DOTS as a scalable and accurate solution for temporal causaldiscovery.