SSDataAgent
Agent-based generation and evaluation of synthetic social-survey data.
Updated
Contents
Research question
SSDataAgent studies the use of an LLM agent to generate synthetic survey data. The agent inspects source data, writes and runs modeling code, and evaluates the generated population. Evaluation covers distributional fidelity and the risk of copying source records.
Method
The public repository includes statistical baselines and empirical-copula transfer. The latter resamples shared source rows across variables to retain their joint structure, then maps values to the specified marginals. For categorical variables, the implementation uses the sampled row’s category interval.
The transfer report separates a source-only setting from an oracle setting supplied with target marginals. These settings have different information access and should be compared separately.
Status and contribution
Research code and experiment reports are public. The empirical-copula implementation commit is attributed to houx15.
The reported transfer evaluation contains two scored time pairs and a reduced-precision setting. Its results have not been reproduced for this website; no general performance or privacy guarantee is stated here.