Trajectory-Constrained Feature Transformation Optimization Framework for Stable Tabular Data Reprogramming

Abstract

Tabular data reprogramming is a cornerstone of Data-Centric AI, which seeks to mathematically reconfigure raw features into more expressive and model-efficient representations. This process has traditionally relied on extensive manual engineering. Further, recent automated feature transformation approaches treat each transformation as conditionally independent, preventing the reuse of successful exploration experiences and learning from failures. This myopic view results in redundant evaluations, unstable exploration, and limited generalizability. To address the limitations, we propose the Trajectory-Constrained Transformation Optimization ( TCTO ), a novel data reprogramming framework that reconceptualizes feature transformation as a coordinated planning problem rather than isolated optimization steps. TCTO introduces a cascading multi-agent reinforcement learning (MARL) system that maintains an evolving roadmap where features become nodes and transformations form edges. This framework supports three critical capabilities: (1) trajectory-aware planning with intelligent backtracking from local optima, (2) reuse of previously generated intermediate features, and (3) localized contribution tracking that stabilizes the exploration process. During agent learning, experience replay reuses stored transitions and rewards, avoiding repeated downstream-model evaluations for the sampled experiences. Extensive experiments and case studies across diverse benchmark datasets and tasks demonstrate that TCTO achieves superior performance, stability, and traceability, advancing the DCAI paradigm toward intelligent systems that learn not only from data, but to reprogram data itself. 1