Cumulative Difference Learning VAE for Time-Series with Temporally Correlated Inflow-Outflow

• DOI: 10.1609/aaai.v38i12.29266
• Publication Date: 2024-03-25T11:08:59Z
• Authors: Tianchun Li; Chengxiang Wu; Pengyi Shi; Xiaoqian Wang
• Keywords: Inflow; Outflow

Time-series generation has crucial practical significance for decision-making under uncertainty. Existing methods have various limitations like accumulating errors over time, significantly impacting downstream tasks. We develop a novel method, DT-VAE, that incorporates generalizable domain knowledge, is mathematically justified, and outperforms existing by mitigating error accumulation through cumulative difference learning mechanism. evaluate the performance of DT-VAE on several tasks using both semi-synthetic real time-series datasets, including benchmark datasets our newly curated COVID-19 hospitalization datasets. The enrich resources analysis. Additionally, we introduce Diverse Trend Preserving (DTP), clustering-based evaluation direct interpretable assessments generated samples, serving as valuable tool evaluating generative models.