GT-GAN: General Purpose Time Series Synthesis with Generative Adversarial Networks

Authors: Jinsung Jeon, JEONGHAK KIM, Haryong Song, Seunghyeon Cho, Noseong Park

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments with 4 datasets and 7 baselines. Our method outperforms all existing methods. Our experimental results and visualization prove the efficacy of the proposed model.
Researcher Affiliation Collaboration Jinsung Jeon Yonsei University jjsjjs0902@yonsei.ac.kr Jeonghak Kim Kakao Corp. haggie.pro@kakaocorp.com Haryong Song Linger Studio Corp. harong@lingercorp.com Seunghyeon Cho Yonsei University seunghyeoncho@yonsei.ac.kr Noseong Park Yonsei University noseong@yonsei.ac.kr
Pseudocode No The paper describes the training method in text (Section 3.4) but does not include any explicitly labeled pseudocode or algorithm blocks in the main body. It refers to 'Our training algorithm refer to Appendix 1' but Appendix 1 is not part of the provided text.
Open Source Code Yes We release our model. See our supplementary material. Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See our supplemental material.
Open Datasets Yes Datasets We conduct experiments with 2 simulated and 2 real-world datasets. Sines has 5 features where each feature is created with different frequencies and phases independently. For each feature, i {1, ..., 5}, xi(t) = sin(2πfit + θi), wherefi U[0, 1] and θi U[ π, π]. Mu Jo Co is multivariate physics simulation time series data with 14 features. Stocks is the Google stock price data from 2004 to 2019. Each observation represents one day and has 6 features. Energy is a UCI appliance energy prediction dataset with 28 values.
Dataset Splits No The paper mentions training and testing data implicitly through the 'train-synthesis-and-test-real (TSTR)' method and 'test data', but it does not provide explicit details about validation dataset splits (e.g., percentages or sample counts) within the provided text.
Hardware Specification Yes Our software and hardware environments are as follows: UBUNTU 18.04 LTS, PYTHON 3.8.10, PYTORCH 1.8.1, TENSORFLOW 2.5.0, CUDA 11.2, and NVIDIA Driver 417.22, i9 CPU, and NVIDIA RTX 3090.
Software Dependencies Yes Our software and hardware environments are as follows: UBUNTU 18.04 LTS, PYTHON 3.8.10, PYTORCH 1.8.1, TENSORFLOW 2.5.0, CUDA 11.2, and NVIDIA Driver 417.22, i9 CPU, and NVIDIA RTX 3090.
Experiment Setup Yes The hyperparameters that significantly affect model performance are the absolute tolerance (atol), the relative tolerance (rtol), and the period of the MLE training (PMLE) for the generator. The atol and rtol determine the error control performed by the ODE solvers in CTFPs. We test with various options of the hyperparameters in Appendix F. We found that there is an appropriate error tolerance (atol, rtol) depending on the data input size. For example, the datasets with small input sizes (i.e., Sines, Stocks) have good discriminator scores with (1e-2, 1e-3), and the datasets with large input sizes (i.e., Energy, Mu Jo Co) show good results with (1e-3, 1e-2).