Affinity and class probability-based fuzzy support vector machine for imbalanced data sets.

Area of Science:

• Data Mining
• Machine Learning
• Pattern Recognition

Background:

• Imbalanced datasets present a significant challenge in data mining.
• Conventional Support Vector Machines (SVM) can be biased towards the majority class, especially with outliers or noise.
• Existing methods struggle to effectively handle the skewed distribution of data points.

Purpose of the Study:

• To propose a novel Affinity and Class Probability-based Fuzzy Support Vector Machine (ACFSVM) technique.
• To enhance the classification performance on imbalanced datasets by addressing the limitations of conventional SVM.
• To improve the decision boundary accuracy by reducing the influence of noisy and outlier majority class samples.

Main Methods:

• Calculating sample affinity using a Support Vector Description Domain (SVDD) model on majority class samples.
• Employing kernel k-nearest neighbor (k-NN) to determine class probabilities for majority class samples.
• Constructing sample memberships by combining affinities and class probabilities to adjust learning contributions.

Main Results:

• ACFSVM effectively identifies and down-weights noisy or outlier majority class samples.
• The proposed method demonstrates improved generalization performance across various imbalanced datasets.
• Experimental results show superior performance in terms of G-Mean, F-Measure, and AUC compared to existing techniques.

Conclusions:

• ACFSVM offers a robust solution for imbalanced data classification by adaptively adjusting sample contributions.
• The technique successfully mitigates the bias towards the majority class, leading to more balanced and accurate models.
• ACFSVM provides a significant advancement in handling challenging imbalanced learning problems.