Dynamical Wasserstein Barycenters for Time-series Modeling

Authors: Kevin Cheng, Shuchin Aeron, Michael C. Hughes, Eric L Miller

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on several human activity datasets show that our proposed DWB model accurately learns the generating distribution of pure states while improving state estimation for transition periods compared to the commonly used linear interpolation mixture models.
Researcher Affiliation Academia Kevin C. Cheng Tufts University kevin.cheng@tufts.edu Shuchin Aeron Tufts University shuchin.aeron@tufts.edu Michael C. Hughes Tufts University michael.hughes@tufts.edu Eric L. Miller Tufts University eric.miller@tufts.edu
Pseudocode Yes Algorithm 1: Dynamical Wasserstein Barycenter (DWB) Time-Series Estimation
Open Source Code Yes Code available at https://github.com/kevin-c-cheng/DynamicalWassBarycenters_Gaussian
Open Datasets Yes Microsoft Research Human Activity (MSR, Morris et al. (2014))
Dataset Splits No The paper does not specify distinct training, validation, and test dataset splits with percentages or counts. It mentions training 'in unsupervised fashion' and evaluating performance based on 'how well a given model s predicted emission distribution fits the observed data over the whole time series'.
Hardware Specification No The paper does not specify the hardware used for experiments, such as particular CPU or GPU models, or detailed cloud computing resources.
Software Dependencies No The paper mentions using PyTorch for auto-differentiation but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes Hyperparameter choices are documented in Tab. 1, with window size n chosen to capture 2-5 seconds of real-time activity in each dataset. We set s = 1.0 and λ = 100 according to the parameter selection study later in Sec. 6.2. We set m0 to the mean of the observed data, and σ0 to the average eigenvalue of the set of covariance matrices obtained from a K-component GMM fit to the data using expectation-maximization via code from Pedregosa et al. (2011).