Principle:Sdv dev SDV PAR Synthesis
| Knowledge Sources | |
|---|---|
| Domains | Time_Series, Synthetic_Data, Autoregressive_Models |
| Last Updated | 2026-02-14 00:00 GMT |
Overview
An autoregressive modeling technique that generates synthetic sequential data by learning temporal patterns and entity-level context from multi-sequence datasets.
Description
PAR (Probabilistic Auto-Regressive) synthesis handles sequential data where each entity (identified by a sequence key) has a time-ordered series of observations. The approach separates context columns (attributes constant within a sequence, like entity demographics) from sequential columns (values that change over time). Context columns are modeled with a GaussianCopulaSynthesizer, while sequential columns are modeled with DeepEcho's PARModel, an autoregressive neural network that generates one timestep at a time conditioned on the context and previous timesteps.
Usage
Use PAR synthesis when generating synthetic time-series or sequential data where each sequence belongs to a distinct entity. The metadata must include a sequence key column (sdtype='id') and optionally a sequence index (for ordered timestamps).
Theoretical Basis
1. Context modeling: Entity-level attributes are modeled using a Gaussian copula, capturing correlations between static properties (e.g., age, region, account type).
2. Sequence assembly: Data is grouped by sequence key and ordered by sequence index to form discrete sequences.
3. Autoregressive generation: The PARModel generates each timestep conditioned on:
- The entity's context vector
- All previously generated timesteps in the sequence
4. Sampling: New contexts are sampled from the copula model, then sequences are generated autoregressively for each context.