Unsupervised Scalable Representation Learning for Multivariate Time Series
Authors: Jean-Yves Franceschi, Aymeric Dieuleveut, Martin Jaggi
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We assess the quality of the learned representations on various datasets to ensure their universality. In particular, we test how our representations can be used for classification tasks on the standard datasets in the time series literature, compiled in the UCR repository (Dau et al., 2018). |
| Researcher Affiliation | Academia | Jean-Yves Franceschi Sorbonne Université, CNRS, LIP6, F-75005 Paris, France jean-yves.franceschi@lip6.fr Aymeric Dieuleveut MLO, EPFL, Lausanne CH-1015, Switzerland CMAP, Ecole Polytechnique, Palaiseau, France aymeric.dieuleveut@polytechnique.edu Martin Jaggi MLO, EPFL, Lausanne CH-1015, Switzerland martin.jaggi@epfl.ch |
| Pseudocode | Yes | Algorithm 1: Choices of xref, xpos and (xneg k )k J1,KK for an epoch over the set (yi)i J1,NK. 1 for i J1, NK with si = size(yi) do 2 pick spos = size(xpos) in J1, si K and sref = size xref in Jspos, si K uniformly at random; 3 pick xref uniformly at random among subseries of yi of length sref; 4 pick xpos uniformly at random among subseries of xref of length spos; 5 pick uniformly at random ik J1, NK, then sneg k = size(xneg k ) in J1, size(yk)K and finally xneg k among subseries of yk of length sneg k , for k J1, KK. |
| Open Source Code | Yes | The code corresponding to these experiments is attached in the supplementary material and is publicly available.4 https://github.com/White-Link/Unsupervised Scalable Representation Learning Time Series. |
| Open Datasets | Yes | We assess the quality of the learned representations on various datasets to ensure their universality. In particular, we test how our representations can be used for classification tasks on the standard datasets in the time series literature, compiled in the UCR repository (Dau et al., 2018)... we also evaluate our representations on the recent UEA multivariate time series repository (Bagnall et al., 2018), as well as on a real-life dataset including very long time series... The Individual Household Electric Power Consumption (IHEPC) dataset from the UCI Machine Learning Repository (Dheeru & Karra Taniskidou, 2017) is a single time series of length 2 075 259... |
| Dataset Splits | No | The paper explicitly mentions 'train / test split' for datasets and provides details for the IHEPC dataset ('train (first 5 105 measurements, approximately a year) and test (remaining measurements)'). However, it does not explicitly mention or detail a 'validation' split. |
| Hardware Specification | Yes | Each encoder was trained using the Adam optimizer (Kingma & Ba, 2015) on a single Nvidia Titan Xp GPU with CUDA 9.0, unless stated otherwise. ... The encoder is trained over the train time series on a single Nvidia Tesla P100 GPU in no more than a few hours... Results and execution times on an Nvidia Titan Xp GPU are presented in Table 2. |
| Software Dependencies | No | We used Python 3 for implementation, with Py Torch 0.4.1 (Paszke et al., 2017) for neural networks and scikit-learn (Pedregosa et al., 2011) for SVMs. While PyTorch has a specific version (0.4.1), Python 3 lacks a minor version (e.g., 3.x) and scikit-learn's version is not specified, thus failing to meet the 'multiple key software components with their versions' criterion. |
| Experiment Setup | No | The paper states: 'The full training process and hyperparameter choices are detailed in the supplementary material, Sections S1 and S2.' While it mentions the Adam optimizer, it does not provide specific hyperparameter values or detailed training configurations within the main text. |