Neural Network With a Preference Sampling Paradigm for Imbalanced Data Classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Real-world data frequently exhibit imbalance, posing challenges for neural network models.
• Imbalanced data can lead to negative class preference in neural networks.
• Existing undersampling methods often overlook gradient inundation and positive sample representation issues.

Purpose of the Study:

• To propose a new paradigm for effectively handling data imbalance in neural networks.
• To address gradient inundation and insufficient positive sample representation.
• To enhance neural network performance on imbalanced datasets.

Main Methods:

• An informative undersampling strategy derived from performance degradation is employed to combat gradient inundation.
• A boundary expansion strategy incorporating linear interpolation and prediction consistency constraints is utilized to improve positive sample representation.
• The proposed paradigm was tested on 34 imbalanced datasets with varying imbalance ratios.

Main Results:

• The novel paradigm demonstrated superior performance in addressing data imbalance.
• The method effectively restored neural network functionality under imbalanced conditions.
• The proposed approach achieved the best Area Under the ROC Curve (AUC) on 26 out of 34 tested datasets.

Conclusions:

• The developed paradigm offers a robust solution for data imbalance problems in neural networks.
• The strategy effectively mitigates common issues like gradient inundation and inadequate positive sample representation.
• This approach significantly enhances the reliability and accuracy of neural networks when dealing with imbalanced data.