FUSE: Feature-Wise Unified Specialization with Cross-Column Exchange for Mixed-Type Tabular Flow Matching
The episode discusses the FUSE paper, which introduces a neural architecture for generating synthetic tabular data. FUSE uses type-specific adaptive mixtures and cross-column attention to handle mixed numerical and categorical features. The hosts highlight its strong results across eight datasets, particularly on marginal fidelity and classifier-based tests, and explain the theoretical bounds on generation error.