New protocol evaluates temporal fidelity in generative models for sequential data

A new evaluation protocol has been developed to assess the temporal fidelity of generative models when creating synthetic sequential tabular data. This protocol, guided by a taxonomy of data properties, measures timestamp validity, cross-sectional structure, within-entity dynamics, and time-varying relational structure. Applying this protocol to eight generative models across thirteen datasets revealed that rankings differ significantly from those obtained through conventional evaluation methods, indicating that temporal fidelity must be measured directly on the time axis rather than inferred from pooled record distributions. AI

IMPACT This research highlights critical limitations in current generative model evaluations for sequential data, suggesting a need for more robust temporal fidelity assessments.

RANK_REASON The item is an academic paper detailing a new evaluation protocol for generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• arXiv
• CatalyzeX
• computer science
• DagsHub
• Generative Models
• Gotit.pub
• Hugging Face
• IArxiv
• Influence Flower
• machine learning
• ScienceCast
• Synthetic Sequential Tabular Data

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →