超越边际分布:多表合成数据生成的多维度评估框架
Beyond Marginals: A Multi-Dimensional Evaluation Framework for Multi-Table Synthetic Data Generation
Aparana Gupta · Anurup Dey · Suyash Dwivedi
中文摘要
合成数据生成对隐私合规、机器学习增强与软件测试至关重要。单表场景下的评估方法已较为成熟,但以企业级应用为主的多表(关系型)合成数据仍缺乏统一的评估框架。现有方法仅孤立地评估各列的边际分布,忽略了联合分布、跨表结构完整性、下游任务可用性,以及生产环境中的边界情况。本文提出 SynEval,一个面向多表合成数据库的六维度评估框架。SynEval 联合评估逐列保真度、多元结构保持(包括一种新颖的条件分布检验)、跨表完整性、机器学习效用、隐私保护以及边界用例的稳健性。该框架输出统一的加权质量分数,支持按表与按维度下钻;且与生成器无关,可在任意一对真实与合成 CSV 文件夹上运行,并自动推断表结构。SynEval 将条件分布检验 P(Y|X)、跨表基数校验以及生产就绪边界用例测试整合到同一关系型合成数据评估流水线中。
关键要点
- 01多表关系型合成数据评估尚缺统一框架,现有方法仅看列边际分布,忽略联合分布与跨表结构
- 02SynEval 提出六维度评估:逐列保真度、多元结构(含条件分布 P(Y|X) 检验)、跨表完整性、ML 效用、隐私保护、边界用例稳健性
- 03框架与生成器无关,接受任意 CSV 文件夹并自动推断 schema,输出统一加权分数并支持按表/按维度下钻
- 04整合条件分布、跨表基数校验与生产边界测试于单一流水线
- 05面向企业级关系型合成数据的端到端可复现评估
解读
尚无解读。
原始英文摘要
arXiv:2610.06854v1 Announce Type: cross Abstract: Synthetic data generation is critical for privacy compliance, machine learning augmentation, and software testing. While single-table evaluation is well established, multi-table (relational) synthesis, the dominant enterprise use case, lacks a unified evaluation framework. Existing approaches assess marginal column distributions in isolation, overlooking joint distributions, cross-table structural integrity, downstream utility, and production-readiness edge cases. We present SynEval, a six-dimensional evaluation framework for multi-table synthetic databases. SynEval jointly assesses per-column fidelity, multivariate structure preservation including a novel conditional distribution check, cross-table integrity, ML utility, privacy protection, and edge-case robustness. The framework produces a unified weighted quality score with per-table and per-dimension drill-down, and is generator-agnostic, operating on any pair of real and synthetic CSV folders with automatic schema inference. SynEval is a framework to combine conditional distribution checks P(Y|X), cross-table cardinality validation, and production-readiness edge-case testing within a single evaluation pipeline for relational synthetic data.