Permutation-Invariant Hierarchical Representation Learning for Reinforcement-Guided Feature Transformation

ICLR 2026 Conference Submission14560 Authors

Keywords: Automated Feature Transformation, Hierarchical Representation Learning, Reinforcement Learning

Abstract: Feature transformation aims to refine tabular feature spaces by mathematically transforming existing features into more predictive representations. Recent advances leverage generative intelligence to encode transformation knowledge into continuous embedding spaces, facilitating the exploration of superior feature transformation sequences. However, such methods face three critical limitations: 1) Neglecting hierarchical relationships between low-level features, mathematical operations and the resulting high-level feature abstractions, causing incomplete representations of the transformation process; 2) Incorrectly encoding transformation sequences as order-sensitive, introducing unnecessary biases into the learned continuous embedding space; 3) Relying on gradient-based search methods under the assumption of embedding space convexity, making these methods susceptible to being trapped in local optima. To address these limitations, we propose a novel framework consisting of two key components. First, we introduce a permutation-invariant hierarchical modeling module that explicitly captures hierarchical interactions from low-level features and operations to high-level feature abstractions. Within this module, an self-attention pooling mechanism ensures permutation invariance of the learned embedding space, aligning generated feature abstractions directly with predictive performance. Second, we develop a policy-guided multi-objective search strategy using reinforcement learning (RL) to effectively explore the embedding space. We select locally optimal search seeds from empirical data based on model performance, then simultaneously optimize predictive accuracy and minimize transformation sequence length starting from these seeds. Finally, extensive experiments are conducted to evaluate the effectiveness, efficiency and robustness of our framework. Our code and data are publicly accessible https://anonymous.4open.science/r/PHER-32A6.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 14560