Reproduce “Beyond the Norm: A Survey of Synthetic Data Generation for Rare Events” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingyi Gu, Xuan Zhang, Guiling Wang
A comprehensive survey reviews synthetic data generation methods for rare events, proposing a structured evaluation framework that prioritizes "extremeness performance." It categorizes generative models, benchmark datasets, and identifies underexplored high-impact domains, offering a foundational resource for modeling low-frequency, high-impact phenomena.
Key insights
- Integrating Extreme Value Theory (EVT) into generative models (e.g., through guided sampling or loss functions) is a dominant and effective strategy for synthesizing heavy-tailed data and capturing rare-event dynamics.
- Generative modeling for extreme events is advancing beyond univariate analysis, incorporating structured, spatial, and multivariate tail modeling, often with domain-specific constraints, to enhance realism.
- A specialized evaluation framework focusing on metrics like extreme coverage, extremal dependence, and downstream performance is critical, as traditional metrics centered on general data similarity or privacy are insufficient for assessing rare-event fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.