A Cost-Sensitive Deep Belief Network for Imbalanced Classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Imbalanced data classification is a common challenge in real-world applications.
• Conventional Deep Belief Networks (DBN) struggle with imbalanced data due to assumptions of equal class costs.
• Cost-sensitive learning offers a solution by assigning differential misclassification costs, but practical implementation is hindered by unknown costs.

Purpose of the Study:

• To propose an effective cost-sensitive Deep Belief Network (DBN) for imbalanced data classification.
• To address the challenge of unknown misclassification costs in imbalanced learning scenarios.
• To enhance DBN performance on imbalanced datasets through an automated cost optimization method.

Main Methods:

• Developed an evolutionary cost-sensitive Deep Belief Network (ECS-DBN).
• Employed adaptive differential evolution to optimize misclassification costs based on training data.
• Integrated the G-mean evaluation measure into the objective function for cost optimization.
• Applied optimized costs to the DBN for classification tasks.

Main Results:

• The proposed ECS-DBN approach demonstrated consistent outperformance against state-of-the-art methods.
• Experiments were conducted on both benchmark and real-world datasets.
• The method showed significant improvements in imbalanced classification tasks, including fault diagnosis in tool condition monitoring.

Conclusions:

• The ECS-DBN provides an effective solution for imbalanced data classification by automatically optimizing misclassification costs.
• Adaptive differential evolution enables cost optimization without requiring prior domain knowledge.
• The proposed method offers a robust and superior alternative for handling skewed class distributions in machine learning.