Conceptual

Population-Aware Diffusion for Distribution-Preserving Time Series Generation

A diffusion-based generator for synthetic multivariate time series (PaD-TS) that preserves population-level statistics of the source dataset, not just the realism of individual samples. It adds a training objective penalizing drift in per-dimension value distributions and in the distribution of cross-correlations between dimensions, plus a dual-channel encoder that better captures temporal structure, sharply reducing cross-correlation distribution shift while matching state-of-the-art individual-level fidelity.