Sortability of Time Series Data
Authors: Christopher Lohse, Jonas Wahl
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we show that certain characteristics of datasets, such as varsortability Reisach et al. (2021) and R2sortability Reisach et al. (2023), also occur in datasets for autocorrelated stationary time series. We illustrate this empirically using four types of data: simulated data based on SVAR models and Erdős-Rényi graphs, the data used in the 2019 causality-for-climate challenge (Runge et al., 2019), real-world river stream datasets, and real-world data generated by the Causal Chamber of Gamella et al. (2024). To do this, we adapt varand R2-sortability to time series data. We also investigate the extent to which the performance of continuous score-based causal discovery methods goes hand in hand with high sortability. Arguably, our most surprising finding is that the investigated real-world datasets exhibit high varsortability and low R2-sortability indicating that scales may carry a significant amount of causal information. |
| Researcher Affiliation | Collaboration | Christopher Lohse EMAIL School of Computer Science and Statistics University of Dublin Trinity College IBM Research Europe, Dublin Jonas Wahl EMAIL Deutsches Forschungszentrum für künstliche Intelligenz (DFKI) |
| Pseudocode | No | The paper describes the 'sortnregress' algorithm in Section 4 using narrative text and numbered steps, but it does not present it in a formally structured pseudocode block or algorithm environment. |
| Open Source Code | No | We conduct our experiments in Python, using the TIGRAMITE library2 for simulating data with SVAR models. 2https://github.com/jakobrunge/tigramite. The paper mentions a third-party library used for data simulation but does not provide concrete access to the authors' own implementation code for the methodology described in the paper. |
| Open Datasets | Yes | We illustrate this empirically using four types of data: simulated data based on SVAR models and Erdős-Rényi graphs, the data used in the 2019 causality-for-climate challenge (Runge et al., 2019)1, real-world river stream datasets, and real-world data generated by the Causal Chamber of Gamella et al. (2024)3. 1https://causeme.uv.es/neurips2019/ 3https://github.com/juangamella/causal-chamber |
| Dataset Splits | No | For each number of nodes we randomly generate 500 different graphs and generate n = 500 samples per graph. The dataset includes simulated and partially simulated datasets of varying complexity, including high-dimensional datasets and non-linear dependencies, with 100, 150, 600 or 1000 realisations for each dataset specification. The paper mentions the number of samples or realisations for data generation but does not specify how these datasets are split into training, validation, or test sets. |
| Hardware Specification | No | No specific hardware details (such as GPU or CPU models, or memory specifications) used for running the experiments are mentioned in the paper. |
| Software Dependencies | No | We conduct our experiments in Python, using the TIGRAMITE library2 for simulating data with SVAR models. When assessing the performance of different algorithms across a range of sortability values, the hyperparameters of the DYNOTEARS algorithm are set to λ1 = λ2 = 0.05. The weight threshold is set to 0.1. As a constraint-based comparison algorithm, PCMCI+(Runge, 2020) is run with α = 0.01 and the Par Corr conditional independence test. While Python and the TIGRAMITE library are mentioned, no specific version numbers are provided for Python or any software libraries/tools. |
| Experiment Setup | Yes | When assessing the performance of different algorithms across a range of sortability values, the hyperparameters of the DYNOTEARS algorithm are set to λ1 = λ2 = 0.05. The weight threshold is set to 0.1. As a constraint-based comparison algorithm, PCMCI+(Runge, 2020) is run with α = 0.01 and the Par Corr conditional independence test. We further use the varsortnregress, R2-sortnregress and random regress algorithms as described in Section 4. |