PrimeNet: Pre-training for Irregular Multivariate Time Series

Authors: Ranak Roy Chowdhury, Jiacheng Li, Xiyuan Zhang, Dezhi Hong, Rajesh K. Gupta, Jingbo Shang

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiment results show that Prime Net significantly outperforms state-of-the-art methods on naturally irregular and asynchronous data from Healthcare and Io T applications for several downstream tasks, including classification, interpolation, and regression. Experiment results show that Prime Net significantly outperforms all baselines on all datasets for all downstream tasks, under both few-shot and full training data settings.
Researcher Affiliation Collaboration Ranak Roy Chowdhury1, Jiacheng Li1, Xiyuan Zhang1, Dezhi Hong2, Rajesh K. Gupta1, Jingbo Shang1 1 University of California, San Diego 2 Amazon {rrchowdh, j9li, gupta, jshang}@eng.ucsd.edu, {xiyuanzh}@ucsd.edu, hondezhi@amazon.com
Pseudocode Yes Algorithm 1: Time CL Data Augmentation; Algorithm 2: Time Reco Data Augmentation
Open Source Code Yes Reproducibility Code is publicly available at https://github.com/ranakroychowdhury/Prime Net
Open Datasets Yes Datasets Physio Net Challenge 2012 (Silva et al. 2012) and MIMIC-III (Johnson et al. 2016) are multivariate time series datasets... Activity (Kaluˇza et al. 2010) dataset has 3-D positions... Appliances Energy (Tan et al. 2021) dataset contains...
Dataset Splits No During pretraining, we measure contrastive learning classification (i.e. how many samples are predicted correctly among the 2B sub-samples) and use the validation accuracy for early stopping. While validation is mentioned, specific details about dataset splits (e.g., percentages or exact counts for training, validation, and testing) are not provided.
Hardware Specification No The paper mentions 'efficient GPU implementation' but does not specify any particular GPU models, CPU models, or other hardware specifications used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., specific libraries, frameworks, or programming language versions like 'PyTorch 1.9' or 'Python 3.8').
Experiment Setup Yes We compute Cross-Entropy Loss for classification and Root Mean Squared Error (RMSE) for regression and interpolation. We conduct grid search on hyper-parameters, η = (0.3, 0.4, 0.5, 0.6, 0.7), α = (0.15, 0.05, 0.03), J = (1, 3, 5), µl, λl = (0.3, 0.4) and µu, λu = (0.7, 0.6) to report test results based on the best held-out validation performance. Best values for η = 0.5, 0.6, 0.5, 0.5 for Physio Net, MIMIC-III, Activity, Appliances Energy, respectively.