Cascade interpolation learning with double subspaces and confidence disturbance for imbalanced problems.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Imbalanced classification presents significant challenges in various machine learning applications.
• Existing ensemble methods like Cascade Forest offer solutions for big data but can be generalized.
• The need for robust frameworks that handle skewed class distributions is critical.

Purpose of the Study:

• To introduce a novel ensemble framework, Cascade Interpolation Learning with Double subspaces and Confidence disturbance (CILDC), for imbalanced classification.
• To enhance the generalization capability of cascade models by accommodating a wider range of base classifiers.
• To improve the adaptive threshold learning for imbalanced samples.

Main Methods:

• Developed CILDC, an ensemble framework extending the Cascade Forest concept.
• Integrated base classifiers using a double subspaces strategy and random under-sampling preprocessing.
• Introduced a confidence disturbance technique to adaptively tune thresholds for imbalanced data.

Main Results:

• CILDC demonstrated effectiveness in handling imbalanced classification problems.
• The framework showed strong generalization capabilities across typical imbalanced datasets.
• Experiments confirmed the suitability of Random Forest and Naive Bayes as base classifiers within CILDC.

Conclusions:

• CILDC provides a robust and generalizable solution for imbalanced classification tasks.
• The proposed confidence disturbance and double subspaces strategies enhance model performance.
• The framework offers an effective approach for adaptive threshold learning in skewed datasets.