Stat Colloquium: Dr. Hyunjoong Kim

Title:

SURF: A Synthesizer Utilizing Randomized Forests for Class-Conditional Tabular Data Generation with Cross-Tree Validation 

Abstract:

Synthetic tabular data generation remains challenging when datasets contain mixed feature types, severe class imbalance, and complex local dependencies. We propose SURF (Synthesizer Utilizing Randomized Forests), a non-neural, tree-based framework for class-conditional tabular data generation. SURF builds a randomized forest in a single training pass, directly modeling the local class-conditional distribution of $\mathbf{X}$ given $Y=c$ within each partition.
Each candidate record is proposed by one tree and validated by the rest, so weakly supported candidates are rejected without adversarial retraining or a separate discriminator.
We evaluated SURF on five real-world datasets against probabilistic, copula-based, adversarial-based, autoencoder-based, diffusion-based, transformer-based, and tree-based synthesizers. Across a domain-level summary spanning fidelity, utility, anomaly behavior, privacy, and diversity, SURF achieved the most balanced profile among the evaluated methods, at a computational cost well below that of the transformer- and diffusion-based alternatives. These findings demonstrate that SURF provides a practical and balanced approach to synthetic tabular data generation.