Do Generative Models Keep Time? A Time-Aware Evaluation of Synthetic Sequential Tabular Data

2026-07-16

AI Summary

Existing generative models for synthetic temporal tabular data often produce temporally inconsistent outputs, such as time reversals, repetitions, or implausible trajectories, yet conventional evaluation methods fail to detect these issues due to their neglect of the temporal dimension. This work proposes the first systematic evaluation framework tailored for temporal tabular data, which dynamically selects assessment criteria based on four key properties: temporal representation, sampling regularity, trajectory dependency, and pattern structure. The framework holistically evaluates timestamp validity, cross-sectional structural coherence, entity-level dynamic evolution, and time-varying relationships. By elevating both utility and privacy assessments from static records to the trajectory level, it reveals—across 13 real-world datasets—that traditional evaluations significantly diverge from temporally aware results, with failure modes closely tied to model architecture, thereby underscoring the necessity of explicitly modeling the temporal axis.

Abstract

Synthetic sequential tabular data are increasingly used for privacy-preserving data sharing, yet a generator can reproduce every marginal and every foreign-key relationship while emitting timestamps that run backwards or repeat, and while sending entities along paths that no real entity followed. Conventional tabular evaluation, which pools records into static distributions, is blind to such failures. We present a taxonomy-guided evaluation protocol for temporal fidelity, in which the applicable measurements are determined by the data rather than fixed in advance. Each dataset is first characterized along four properties: how time is represented, whether observations are regularly sampled, whether trajectories are mutually dependent, and how the schema links entities to their histories. These properties determine which evaluation dimensions are meaningful. The protocol then measures timestamp validity, cross-sectional structure at aligned time points, within-entity dynamics, and time-varying relational structure, and recasts utility and privacy evaluation over trajectories rather than isolated rows. We apply the protocol to eight generative models across thirteen datasets spanning six domains. Rankings under conventional evaluation disagree substantially with those obtained under temporal evaluation, and the resulting failures are architecture-coherent rather than random. Temporal fidelity must therefore be measured on the time axis itself, rather than inferred from pooled record distributions.

Problem

• temporal fidelity
• synthetic sequential tabular data
• time-aware evaluation
• generative models
• timestamp validity

Innovation

• temporal fidelity
• synthetic sequential data
• time-aware evaluation
• trajectory-based assessment
• generative models