Addressing data imbalance in collision risk prediction with active generative oversampling.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Data imbalance significantly compromises predictive accuracy in collision risk assessment.
• Existing methods often struggle to effectively address minority class representation in imbalanced datasets.

Purpose of the Study:

• To propose an advanced active generative oversampling method for imbalanced data in collision risk assessment.
• To enhance the diversity of generated samples and improve fault classification performance.
• To dynamically balance model training for optimal results.

Main Methods:

• Integration of Query by Committee (QBC), Auxiliary Classifier Generative Adversarial Network (ACGAN), and Wasserstein Generative Adversarial Network (WGAN).
• Selective enrichment of minority class samples using QBC and diversity metrics.
• Dynamic adjustment of generator and discriminator training epochs based on loss differences.

Main Results:

• The proposed method significantly outperforms existing techniques on four imbalanced datasets.
• Achieved precision, recall, F-measure, and G-mean above 0.92 across all indicators.
• Demonstrated an average improvement of 23-28.3% compared to the ENN method.

Conclusions:

• The novel method effectively handles data imbalance, leading to more accurate collision risk assessment.
• Improved identification of collision samples and reduced misclassification rates for non-collision samples.
• Offers a robust solution for improving fault classification in imbalanced data scenarios.