Contrastive Learning Is Not Optimal for Quasiperiodic Time Series

Authors: Adrian Atienza, Jakob Bardram, Sadasivan Puthusserypady

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

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed model has undergone extensive simulation studies to evaluate its performance.
Researcher Affiliation Academia Adrian Atienza , Jakob Bardram , Sadasivan Puthusserypady Department of Health Technology, Technical University of Denmark {adar, jakba, sapu}@dtu.dk
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or a link to open-source code for the methodology described.
Open Datasets Yes The model is trained with 10 second-length signals belonging to the Sleep Heart Health Study (SHHS) dataset [Zhang et al., 2018], [Quan et al., 1998]. ... All used databases are publicly available in Physionet [Goldberger et al., 2000] and National Sleep Research Resource (NSRR).
Dataset Splits Yes We conducted a five-fold cross-validation to evaluate the performance of the downstream tasks." and "We used the dataset s predefined partitioning of train and validation sets for evaluating the SVC model fitted on top of the representations." and "In the second, we have conducted a Leave-One Out (LOO) validation across the 23 MIT-AFIB subjects.
Hardware Specification Yes The training procedure and the evaluations are performed on a local computer, with a Nvidia Ge Force RTX 3070 GPU.
Software Dependencies No The paper mentions using Adam optimizer but does not specify versions for other software dependencies like programming languages or libraries.
Experiment Setup Yes The input data is a time series of 1000 samples, which correspond to 10 seconds-length signal sampled at 100Hz. This input is split into segments of a length of 20 samples. The model counts with 6 regular transformer blocks with 4 heads each. The model dimension is set to 128... The projectors and predictors in our approach are implemented as a two-layer Multilayer Perceptron (MLP) with a dimensionality of 512 and 256... The EMA updating factor (τ) is set to 0.995. The window size is set to 2 minutes. We weigh the covariance loss with a factor of 0.1. We optimize the most important 32 features during the selective optimization. The training procedure consists of 30,000 iterations. We use a batch size of 256, and Adam... with a learning rate of 3e 4 and a weight decay of 1.5e 6 as the optimizer.