Classification ensembles for unbalanced class sizes in predictive toxicology.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Imbalanced datasets in binary classification lead to biased models favoring the majority class.
• This bias results in poor sensitivity for minority class prediction, impacting model performance.
• Existing classification rules struggle with differential class sizes, necessitating robust adjustment methods.

Purpose of the Study:

• To investigate the impact of positive-to-negative sample ratios on classification sensitivity, specificity, and concordance.
• To propose and evaluate an ensemble classification approach for adjusting imbalanced training data.
• To introduce and compare two re-sampling techniques: augmentation and abatement.

Main Methods:

• Developed an ensemble classification approach using multiple base classifiers.
• Implemented augmentation: oversampling minority class via bootstrapping.
• Implemented abatement: undersampling majority class via bootstrapping.
• Applied methods to predict estrogen receptor binding and animal liver carcinogenicity using SAR models.

Main Results:

• Ensemble classification effectively adjusts for differential class sizes.
• Both augmentation and abatement generate equal-sized bootstrap samples for base classifiers.
• The abatement method demonstrated superior performance in balancing sensitivity and specificity.

Conclusions:

• Ensemble methods with re-sampling are effective for imbalanced binary classification.
• The abatement re-sampling technique shows promise for improving model fairness and predictive accuracy.
• Findings are applicable to QSAR modeling and other fields with imbalanced datasets.