Are Synthetic Time-series Data Really not as Good as Real Data?

Fanzhe Fu, Junru Chen, Jing Zhang, Carl Yang, Lvbin Ma, Yang Yang

Keywords: Time-Series Data, Data Synthesis, Non-Deep-Learning Data Synthesis, InfoBoost, Prediction, Imputation, Feature Decomposition

TL;DR: InfoBoost, a non-deep-learning method for synthesizing time-series data, which enables superior performance for tasks like prediction and imputation, and facilitates feature decomposition without requiring real data & info.

Abstract: To alleviate the commonly encountered inadequate time-series data problem in DL (DL), we develop a non-DL generic data synthesis method. When current methods require real data or data statistics to train generators or synthesize data, our method InfoBoost enables zero-shot training of models without the need for real data or data statistics. Additionally, as an application of our synthetic data, we train an unconditional feature (rhythm, noise, trend) decomposer based on our synthetic data, which is applicable to real time-series data. Through experiments, our non-DL synthetic data enables models to achieve superior performance on unsupervised tasks and self-supervised prediction \& imputation compared models using real data. Visualized case studies further demonstrate the effectiveness of our novel unconditional feature decomposer trained with our synthetic data.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2068