Flat-Consensus Diffusion for Robust Data Reshaping under Noisy Evaluator

2026-09-26

AI Summary

This study addresses the susceptibility of noisy evaluators to local reward distortion and unstable trajectory generation in data reshaping. To this end, it proposes FCDiff, a framework that formulates reshaping search as a diffusion-based generative process for the first time. FCDiff employs a micro-macro decomposition architecture integrating Gaussian-smoothed Monte Carlo gradients with weighted Fréchet mean aggregation. Furthermore, it introduces a flat consensus mechanism that shifts the optimization objective from isolated optima toward exploring flat regions, thereby enabling robust feature transformation search. Extensive experiments across eight datasets demonstrate that FCDiff significantly outperforms AutoFE and existing generative baselines, yielding substantial improvements in lower-tail reliability and statistically significant precision gains. Ultimately, the framework ensures the generation of stable and reusable reshaping sequences even under noisy evaluation conditions.

Abstract

Data shape determines how features are structured, how patterns are separated, and how distributions cover the underlying domain. Poor data shape can make models learn noise rather than generalizable structure. This paper studies robust feature-centric data reshaping: generating feature transformations that remain useful, stable, and reproducible under noisy evaluation and imperfect data conditions. We view reshaping operation sequence search as reward-guided diffusion generation, and robust reshaping as searching for regions in the latent reward landscape rather than isolated high-reward transformations. The key challenge is dual instability: noisy evaluators distort local reward guidance, while stochastic generative trajectories can converge to inconsistent solutions. We propose FCDiff, a flat-consensus diffusion framework that addresses both failures through a micro-macro decomposition. The micro layer replaces point-estimate reward guidance with Gaussian-smoothed, Monte Carlo averaged gradients, steering generation toward locally flat reward regions. The macro layer aggregates independently guided trajectories with a weighted Frechet-mean barycenter, selecting consensus-supported basins and filtering stochastic outliers. Across an 8-dataset headline cohort under heavy-tailed evaluator noise, FCDiff attains the best aggregate rank on lower-tail reliability and robustness against both search-based AutoFE and robustness-oriented generative baselines, with statistically significant accuracy gains over every generative baseline. Our results show that robust data reshaping requires searching for flat, consensus-supported regions rather than sharp single-trajectory optima.

Problem

• Data Reshaping
• Noisy Evaluator
• Robustness
• Feature Transformation
• Dual Instability

Innovation

• Flat-Consensus Diffusion
• Data Reshaping
• Noisy Evaluator
• Monte Carlo Averaged Gradients
• Frechet-mean Barycenter