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

2026-07-20 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Seq2Synth, a new taxonomy-guided benchmark, evaluates the temporal fidelity of synthetic sequential tabular data, addressing critical gaps in conventional evaluation methods. Existing benchmarks, which pool records into static distributions, fail to detect issues like backward timestamps, repeated entries, or unrealistic entity trajectories. Seq2Synth characterizes datasets by time representation, sampling regularity, trajectory dependence, and schema structure, then applies four temporal fidelity dimensions: timestamp validity, cross-sectional structure, within-entity dynamics (longitudinal), and time-varying relational structure. It also extends utility and privacy evaluations to trajectory-aware settings. Applied to eight generative models across thirteen datasets, Seq2Synth reveals that rankings under conventional evaluation differ substantially from temporal evaluations, with failures being architecture-coherent rather than random. For instance, autoregressive models preserve within-trajectory order but struggle with relational cardinality, while joint generators better preserve schema structure but often fail on timestamp validity.

Key takeaway

For Machine Learning Engineers developing or deploying generative models for sequential tabular data, you must move beyond static-distribution metrics. Your model selection should incorporate Seq2Synth's temporal fidelity dimensions, as conventional evaluations can mask critical failures like invalid timestamps or distorted trajectories. Prioritize generators that explicitly model timestamps as ordered objects and condition on sequence length, ensuring your synthetic data accurately reflects real-world temporal dynamics and relational structures for robust downstream applications.

Key insights

Temporal fidelity in synthetic sequential tabular data requires direct, time-aware evaluation, not inference from static distributions.

Principles

• Static-distribution evaluation is a poor proxy for temporal fidelity.
• Temporal fidelity failures are architecture-coherent, not random.
• Timestamps should be modeled as ordered objects, not continuous attributes.

Method

Seq2Synth classifies datasets by four properties, selects applicable temporal fidelity dimensions (timestamp, cross-sectional, longitudinal, structural), normalizes inputs, and reports dimension-level scores.

In practice

• Incorporate calendar embeddings or auxiliary timestamp losses into generators.
• Condition sequence length and relational cardinality directly in multi-table schemas.

Topics

• Synthetic Data
• Sequential Tabular Data
• Temporal Fidelity
• Generative Models
• Data Evaluation
• Privacy-Preserving Data

Code references

• sdv-dev/SDMetrics

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Do Generative Models Keep Time? A Time-Aware Evaluation of Synthetic Sequential Tabular Data — Conventional evaluation fails to capture temporal fidelity in synthetic sequential tabular data, necessitating time-aware protocols.
• Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records — Evaluating synthetic EMR data requires a multi-dimensional framework beyond statistical similarity to ensure clinical validity.
• Temporal Difference Learning for Diffusion Models — Temporal difference learning improves diffusion model consistency and sample quality, especially for few-step sampling.
• How Can We Synthesize High-Quality Pretraining Data? A Systematic Study of Prompt Design, Generator Model, and Source Data — Structured output formats and careful source data selection are key to high-quality synthetic pretraining data.
• Temporally-Aligned Evaluation for Audio-Driven Talking Head Generation — Evaluating dynamic generative models requires sequence alignment, not just frame-wise comparison, to handle natural temporal variations.
• The Sequence Knowledge # 780: Synthetic Data for Image Models — Synthetic data generation is crucial for robust vision systems, overcoming real-world data limitations through controlled pixel creation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.