Feedforward neural network models for handling class overlap and class imbalance

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Neural network training is often challenged by class overlap and imbalance in datasets.
• Existing methods like conventional feedforward neural networks (FFNNs), quantum neural networks (QNNs), and resampling techniques have limitations in addressing these issues.

Purpose of the Study:

• To propose a general framework for training feedforward neural network models that can effectively handle class overlap and imbalance.
• To introduce specific models, variance-controlled neural networks (VCNNs) and class-balancing neural networks (CBNNs), as special cases of the proposed framework.

Main Methods:

• Development of a novel error function designed to compensate for class overlap and imbalance during training.
• Implementation of VCNNs that minimize an error function involving class-specific variance (CSV).
• Implementation of CBNNs that utilize class-specific correction (CSC) to address class imbalance.

Main Results:

• Experiments on artificial data illustrate the properties of VCNNs and CBNNs.
• Comparative experiments with FFNNs, QNNs, and resampling techniques highlight the effectiveness of the proposed models.
• Real-world data experiments demonstrate the advantages of VCNNs and CBNNs in scenarios with class overlap and imbalance.

Conclusions:

• The proposed framework and its specializations (VCNNs and CBNNs) offer a robust solution for training neural networks with imperfect datasets.
• VCNNs and CBNNs show significant advantages over conventional methods when dealing with class overlap and imbalance in practical applications.