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

Area of Science:

• Computational chemistry
• Machine learning
• Data science

Background:

• Feature selection is crucial for simplifying complex datasets.
• Challenges include determining optimal feature subsets and weighting feature importance.

Purpose of the Study:

• To introduce Differentiable Information Imbalance (DII), an automated method for feature selection.
• To address uncertainties in optimal feature number, unit alignment, and relative importance weighting.

Main Methods:

• DII ranks feature information content using distances in a ground truth feature space.
• It identifies a low-dimensional feature subset that preserves relationships.
• Gradient descent optimizes feature weights by minimizing DII, enabling simultaneous unit alignment and importance scaling.

Main Results:

• DII successfully identifies optimal feature subsets and determines reduced dimensionality.
• Demonstrated effectiveness on biomolecular conformation analysis and machine learning force field development.
• The method supports sparse solutions and preserves model interpretability.

Conclusions:

• DII offers a robust solution for feature selection and dimensionality optimization.
• It has broad applicability in molecular systems and other data-driven fields.
• The DII method is accessible via the Python library DADApy.