Automatic feature selection and weighting in molecular systems using Differentiable Information Imbalance

2024-10-30

Nature Communications

AI Summary

To address the challenges of quantifying feature selection, ambiguous weight assignment, and limited interpretability in molecular system modeling, this paper proposes the first end-to-end differentiable feature selection framework grounded in Differentiable Information Imbalance (DII). The method jointly optimizes feature selection and importance weighting via gradient-based learning, integrating information-theoretic differentiable mutual information estimation, molecular graph neural network representation learning, and a continuous relaxation-based feature gating mechanism. Evaluated on benchmark datasets including QM9 and MD17, the approach achieves significant improvements in energy and force prediction accuracy. Selected features exhibit strong alignment with domain-specific physicochemical priors, enhancing model interpretability. Moreover, computational overhead is lower than that of conventional embedded methods. This work establishes a novel paradigm for interpretable molecular machine learning, bridging principled information-theoretic foundations with practical deep learning architectures for quantum-chemical property prediction.

Technology Category

• Machine Learning: Dimensionality Reduction/Feature Selection
• Search and Optimization: Learning to Search
• Knowledge Representation and Reasoning: Description Logics

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Problem

• Feature Selection
• Model Complexity Optimization
• Quantitative Comparison

Innovation

• Differential Information Imbalance (DII)
• Automatic Feature Selection
• Molecular System Analysis