TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation

• DOI: 10.48550/arxiv.2111.08095
• Publication Date: 2021-01-01
• Authors: Abhyuday Desai; Cynthia Freeman; Zuhui Wang; Ian Beaver
• Keywords: Interpretability; Similarity (geometry); Predictability

Recent work in synthetic data generation the time-series domain has focused on use of Generative Adversarial Networks. We propose a novel architecture for synthetically generating with Variational Auto-Encoders (VAEs). The proposed several distinct properties: interpretability, ability to encode knowledge, and reduced training times. evaluate quality by similarity predictability against four multivariate datasets. experiment varying sizes measure impact availability our VAE method as well state-of-the-art methods. Our results tests show that approach is able accurately represent temporal attributes original data. On next-step prediction tasks using generated data, consistently meets or exceeds performance While noise reduction may cause deviate from we demonstrate resulting de-noised can significantly improve Finally, incorporate domain-specific time-patterns such polynomial trends seasonalities provide interpretable outputs. Such interpretability be highly advantageous applications requiring transparency model outputs where users desire inject prior knowledge patterns into generative model.