Optimizing imbalanced learning with genetic algorithm

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Imbalanced datasets with skewed class distributions present a significant challenge in AI model training, leading to bias towards majority classes.
• Existing synthetic data generation techniques such as SMOTE, ADASYN, GANs, and VAEs often fail to significantly enhance model performance, particularly in cases of extreme class imbalance.

Purpose of the Study:

• To introduce a novel approach for synthetic data generation using Genetic Algorithms (GAs) to address extreme class imbalance in AI datasets.
• To demonstrate that GAs can outperform state-of-the-art methods like SMOTE, ADASYN, GANs, and VAEs in improving AI model performance on imbalanced data.

Main Methods:

• Proposed a novel synthetic data generation method utilizing Genetic Algorithms (GAs), analyzing both Simple and Elitist GA variants.
• Evaluated the effectiveness of population initialization and fitness functions within the GA framework.
• Integrated Logistic Regression and Support Vector Machines for performance evaluation.

Main Results:

• The proposed GA-based synthetic data generation method significantly outperformed existing techniques (SMOTE, ADASYN, GAN, VAE) across three diverse datasets (Credit Card Fraud Detection, PIMA Indian Diabetes, PHONEME).
• Demonstrated superior performance based on key metrics: accuracy, precision, recall, F1-score, ROC-AUC, and AP (Accuracy-Precision) curve.
• The method showed effectiveness without requiring large sample sizes.

Conclusions:

• Genetic Algorithms show significant potential for developing accurate and reliable AI models when trained on imbalanced datasets.
• The proposed GA approach offers a promising alternative for enhancing AI model performance in scenarios with extreme class imbalance.
• This research highlights the versatility of GAs beyond traditional optimization tasks for data generation in machine learning.