SMOTE-CD: SMOTE for compositional data.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Compositional data, representing relative proportions, are prevalent but lack solutions for imbalanced classes.
• Existing methods do not adequately handle the unique characteristics of imbalanced compositional data.

Purpose of the Study:

• To propose an adaptation of Synthetic Minority Oversampling TEchnique (SMOTE) for imbalanced compositional data.
• To introduce SMOTE for Compositional Data (SMOTE-CD) and evaluate its effectiveness.

Main Methods:

• Developed SMOTE-CD by adapting the original SMOTE algorithm using compositional data operations.
• Generated synthetic data points through linear combinations of existing data.
• Tested SMOTE-CD with Gradient Boosting, Neural Networks, and Dirichlet regressors on real and synthetic datasets.

Main Results:

• SMOTE-CD demonstrated performance improvements across accuracy, cross-entropy, F1-score, R2 score, and RMSE.
• Oversampling consistently increased the F1-score, especially for real datasets.
• The impact of oversampling varied by model and data; it sometimes decreased majority class performance but yielded best overall results on real data.

Conclusions:

• SMOTE-CD is an effective technique for handling imbalanced compositional data.
• The method shows promise for improving machine learning model performance in scenarios with skewed compositional data.
• A Python package, smote-cd, is available for implementing the SMOTE-CD method.