A review of machine learning methods for imbalanced data challenges in chemistry

Area of Science:

• Chemistry
• Machine Learning
• Data Science

Background:

• Imbalanced datasets are a common challenge in chemistry, leading to biased machine learning (ML) and deep learning (DL) models.
• This bias compromises the accuracy and reliability of predictions for underrepresented chemical classes, limiting model applicability.

Purpose of the Study:

• To provide a comprehensive review of prominent ML methodologies for addressing imbalanced data in chemistry.
• To evaluate these techniques across diverse chemical applications like drug discovery, materials science, cheminformatics, and catalysis.

Main Methods:

• Review of prominent ML approaches including resampling, data augmentation, algorithmic strategies, and feature engineering.
• Evaluation of methods in the context of chemical applications.
• Exploration of future directions such as data augmentation via physical models, large language models (LLMs), and advanced mathematics.

Main Results:

• Identified and categorized key ML techniques for handling imbalanced chemical data.
• Assessed the applicability and limitations of these methods across various chemistry subfields.
• Highlighted the benefits of balanced data for material design and production.

Conclusions:

• Machine learning techniques offer viable solutions to the imbalanced data problem in chemistry.
• Future research should focus on advanced data augmentation strategies and leveraging LLMs for improved model performance.
• Achieving data balance is crucial for advancing new material design and production.