Resampling Techniques for Materials Informatics: Limitations in Crystal Point Groups Classification.

Area of Science:

• Materials Informatics
• Computational Materials Science
• Crystallography

Background:

• Imbalanced datasets are a significant challenge in developing accurate classification models for materials informatics.
• Predicting crystal point groups serves as a representative imbalanced classification problem within this field.

Purpose of the Study:

• To investigate the impact of various resampling and classification techniques on imbalanced datasets in materials informatics.
• To determine optimal data balancing strategies for improving crystal point group prediction accuracy.

Main Methods:

• Evaluation of multiple resampling techniques (undersampling and oversampling) and their parameters.
• Application of various classification algorithms, including k-nearest neighbors and random forest.
• Analysis of classification performance metrics, focusing on minority class enhancement and majority class trade-offs.

Main Results:

• The number of samples omitted (undersampled) or generated (oversampled) is the most critical parameter in resampling algorithms.
• Balancing imbalanced data enhances minority class prediction performance but can decrease majority class accuracy.
• Ideal class balancing (1:1 ratio) is suboptimal; partial balancing, with a minority-to-majority class ratio around two-thirds, yielded the best results.
• Random undersampling combined with k-nearest neighbors and random forest demonstrated the most substantial improvements in classification accuracy.

Conclusions:

• Partial data balancing is a more effective strategy than complete balancing for imbalanced datasets in materials informatics.
• The optimal ratio for partial balancing in this context is approximately two-thirds.
• Random undersampling, particularly with k-nearest neighbors and random forest, offers a promising approach for improving crystal point group prediction on imbalanced datasets.