Variance-constrained multi-view ensemble broad network for imbalanced data.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Classic broad learning systems (BLS) struggle with imbalanced data, exhibiting bias towards majority classes.
• Disparities in data distributions negatively impact the performance of conventional classification techniques.

Purpose of the Study:

• To introduce a novel method, variance-constrained multi-view ensemble broad network (VMEB), to address limitations in imbalanced data classification.
• To enhance the expressive ability for minority classes and improve feature representation robustness in broad learning systems.

Main Methods:

• VMEB employs minimum class variance and class-specific regularization to construct scatter matrices, improving minority class handling.
• An enhanced kernel mapping technique using Gaussian and arc-cosine functions creates deterministic hidden nodes for robust feature representation.
• A multi-view ensemble framework with feature rotation generates diverse data perspectives for base classifiers, enhancing generalization.

Main Results:

• VMEB demonstrated an average improvement of 4.35% in Area Under the Curve (AUC) compared to baseline methods.
• Extensive experiments on 25 imbalanced datasets validated the effectiveness of VMEB in imbalanced classification tasks.

Conclusions:

• VMEB effectively overcomes the bias towards majority classes in imbalanced datasets.
• The proposed VMEB model offers a robust and effective solution for imbalanced data classification challenges.