Novel cost-sensitive approach to improve the multilayer perceptron performance on imbalanced data

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Statistics

Background:

• Traditional learning algorithms struggle with imbalanced datasets, leading to biased models favoring the majority class.
• The equal cost assumption in statistical learning theory exacerbates performance issues on imbalanced data.
• Multilayer perceptron (MLP) neural networks are susceptible to bias when trained on imbalanced datasets.

Purpose of the Study:

• To investigate the class imbalance problem in multilayer perceptron (MLP) neural networks.
• To develop a novel cost-sensitive algorithm to enhance the discrimination ability of MLPs on imbalanced data.
• To analyze the theoretical implications of incorporating prior information via a cost parameter for balanced decision boundaries.

Main Methods:

• A new cost-sensitive algorithm, CSMLP, is proposed for two-class MLPs.
• CSMLP utilizes a joint objective function with a single cost parameter to weigh class errors.
• The learning rule extends Levenberg-Marquardt's rule for computational efficiency.

Main Results:

• Theoretical demonstration that the cost parameter can lead to balanced decision boundaries.
• Empirical validation on real data shows significant improvements in Area Under the ROC Curve (AUC) and G-mean measures.
• The CSMLP approach effectively mitigates the bias towards the majority class.

Conclusions:

• The proposed CSMLP algorithm offers a robust solution for class imbalance in MLP neural networks.
• Cost-sensitive learning with CSMLP enhances model performance on imbalanced datasets.
• This method provides a statistically sound and computationally efficient approach for improved classification accuracy.