Discovering Mixtures of Structural Causal Models from Time Series Data

Authors: Sumanth Varambally, Yian Ma, Rose Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that our method surpasses state-of-the-art benchmarks in causal discovery tasks through extensive experimentation on synthetic and real-world datasets, particularly when the data emanates from diverse underlying causal graphs. Experimentally, we demonstrate the strong performance of our method on both synthetic and real-world datasets. Notably, MCD can accurately infer the mixture causal graphs and mixture membership information, even when the number of SCMs is misspecified.
Researcher Affiliation Academia 1Halıcıo glu Data Science Institute, University of California, San Diego, La Jolla, USA 2Department of Computer Science and Engineering, University of California, San Diego, La Jolla, USA. Correspondence to: Sumanth Varambally <svarambally@ucsd.edu>.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Implementation is available at https: //github.com/Rose-STL-Lab/MCD.
Open Datasets Yes The Netsim benchmark dataset (Smith et al., 2011) and The DREAM3 dataset (Prill et al., 2010) and We use the yahoofinancials package to retrieve the daily closing prices of D = 100 stocks from January 1, 2016 to July 1, 2023.
Dataset Splits Yes We train the model on 80% of the data and validate on the remaining 20%.
Hardware Specification No The paper does not explicitly describe the specific hardware used to run its experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions using 'Py Torch vectorization' and the 'yahoofinancials package', but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We train the model on 80% of the data and validate on the remaining 20%. The time series length T is 100, and the time lag L is set to 2 for all the methods... Table 3 summarizes the hyperparameters used for training. This includes: Matrix LR 10e-2, Likelihood LR 10e-3, Batch Size 128, # Outer auglag steps 100, # Max inner auglag steps 6000, Embedding dim e = D, Sparsity factor λ 5, Spline type Quadratic/Linear.