Time Weaver: A Conditional Time Series Generation Model
Authors: Sai Shankar Narasimhan, Shubhankar Agarwal, Oguzhan Akcin, Sujay Sanghavi, Sandeep P. Chinchali
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that TIME WEAVER outperforms state-of-the-art benchmarks, such as Generative Adversarial Networks (GANs), by up to 30% in downstream classification tasks on real-world energy, medical, air quality, and traffic datasets. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, USA. Correspondence to: Sai Shankar Narasimhan <nsaishankar@utexas.edu>. |
| Pseudocode | Yes | Algorithm 1 One iteration for training time series ϕtime and metadata ϕmeta feature extractors. |
| Open Source Code | No | No explicit statement or direct link to the open-source code for the paper's methodology was found. The text mentions 'We used the Py Torch implementation (Link to the repo)' but no functional link or clear statement of release for their own code. |
| Open Datasets | Yes | The Electricity dataset consists of power consumption recordings for 370 users over four years from 2011 to 2015. For traffic volume synthesis, we use the metro interstate traffic volume dataset. This dataset contains hourly air pollutants data from 12 air quality monitoring stations in Beijing. The PTB-XL ECG dataset is a 12-channel (1000 time steps long) time series dataset with 17651 train, 2203 validation, and 2167 test samples. Table 1: AIR QUALITY (CHEN, 2019), TRAFFIC (HOGUE, 2019), ELECTRICITY (TRINDADE, 2015), ECG (WAGNER ET AL., 2020). |
| Dataset Splits | Yes | We establish a data split comprising training, validation, and test sets distributed in an 80-10-10 ratio. To obtain the split, we randomly pick 80% of the 434781 samples and assign them to the training set. The same is repeated for the validation and the test sets. For each month, we slide a window of length 96 with a stride of 24, and this provides a total of 12166 train time series samples, 1537 validation time series samples, and 1525 test time series samples. |
| Hardware Specification | Yes | The inference experiments were performed on a single NVIDIA RTX A5000 GPU. |
| Software Dependencies | No | The paper mentions 'PyTorch implementation' but does not specify software dependencies with version numbers. |
| Experiment Setup | Yes | Now, we list the hyperparameter choices used for training the feature extractors in Table 5. These include the number of patches from a single time series sample Npatch, learning rate, etc, and the design choices in terms of the number of self-attention layers, number of transformer heads, etc. |