CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery

Authors: Yuxiao Cheng, Ziqian Wang, Tingxiong Xiao, Qin Zhong, Jinli Suo, Kunlun He

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the experiments, we validate the fidelity of the generated data through qualitative and quantitative experiments, followed by a benchmarking of existing TSCD algorithms using these generated datasets.
Researcher Affiliation Collaboration 1Department of Automation, Tsinghua University 2Institute for Brain and Cognitive Science, Tsinghua University (THUIBCS) 3Chinese PLA General Hospital
Pseudocode Yes A.4 ALGORITHMIC REPRESENTATION FOR CAUSALTIME PIPELINE We show the detailed algorithmic representation of our proposed data generation pipeline in Algorithm A.4, where we exclude quality control and TSCD evaluation steps. Algorithm 1 Pipeline for Causal Time Generation (Excluding quality control and TSCD evaluation) ...
Open Source Code Yes For the purpose of reproducibility, we include the source code on Git Hub (https://github. com/jarrycyx/UNN).
Open Datasets Yes Air Quality Index (AQI) is a subset of several air quality features from 36 monitoring stations spread across Chinese cities2... Traffic subset is built from the time-series collected by traffic sensors in the San Francisco Bay Area3. Medical subset is from MIMIC-4, which is a database that provides critical care data for over 40,000 patients admitted to intensive care units (Johnson et al., 2023).
Dataset Splits No To ensure fairness, we searched for the best set of hyperparameters for these baseline algorithms on the validation dataset, and tested performances on testing sets for 5 random seeds per experiment.
Hardware Specification Yes All experiments are deployed on a server with Intel Core CPU and NVIDIA RTX3090 GPU.
Software Dependencies No The paper mentions software like 'scikit-learn package' for dimension reduction and references implementations for Normalizing Flow and Deep SHAP, but does not provide specific version numbers for these or other software dependencies like Python or PyTorch.
Experiment Setup Yes Table 4: Hyper parameters for time-series fitting. Table 5: Hyperparameters settings of the baseline causal discovery and data imputation algorithms. We show the detailed algorithmic representation of our proposed data generation pipeline in Algorithm A.4, where we exclude quality control and TSCD evaluation steps.