Reinforcement Feature Transformation for Polymer Property Performance Prediction

TL;DR

This paper introduces a novel reinforcement learning-based method to automatically generate and select meaningful, explainable descriptors for polymer property prediction, addressing dataset quality issues and improving model interpretability.

Contribution

It proposes a Traceable Group-wise Reinforcement Generation framework that automates descriptor creation and selection for better polymer property prediction.

Findings

Effective descriptor generation improves prediction accuracy.

Reinforcement learning automates feature engineering.

Enhanced explainability of polymer representations.

Abstract

Polymer property performance prediction aims to forecast specific features or attributes of polymers, which has become an efficient approach to measuring their performance. However, existing machine learning models face challenges in effectively learning polymer representations due to low-quality polymer datasets, which consequently impact their overall performance. This study focuses on improving polymer property performance prediction tasks by reconstructing an optimal and explainable descriptor representation space. Nevertheless, prior research such as feature engineering and representation learning can only partially solve this task since they are either labor-incentive or unexplainable. This raises two issues: 1) automatic transformation and 2) explainable enhancement. To tackle these issues, we propose our unique Traceable Group-wise Reinforcement Generation Perspective. Specifically, we redefine the reconstruction of the representation space into an interactive process, combining nested generation and selection. Generation creates meaningful descriptors, and selection eliminates redundancies to control descriptor sizes. Our approach employs cascading reinforcement learning with three Markov Decision Processes, automating descriptor and operation selection, and descriptor crossing. We utilize a group-wise generation strategy to explore and enhance reward signals for cascading agents. Ultimately, we conduct experiments to indicate the effectiveness of our proposed framework.