Adaptive Learning-Based -Nearest Neighbor Classifiers With Resilience to Class Imbalance.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• The k-Nearest Neighbor (k-NN) classifier's accuracy depends heavily on selecting the optimal number of neighbors (k).
• k-NN performance degrades significantly with imbalanced datasets, where class representation varies widely.
• Optimizing k for k-NN is computationally intensive and challenging.

Purpose of the Study:

• To propose Adaptive k-Nearest Neighbor (Ada-NN) and Ada-NN2, variants of k-NN designed to address the challenges of k-selection and class imbalance.
• To develop a method that dynamically determines a point-specific k based on local data characteristics.
• To introduce a Global Imbalance Handling Scheme (GIHS) for mitigating the effects of class imbalance.

Main Methods:

• Ada-NN utilizes artificial neural networks to learn a point-specific k based on neighborhood density and distribution.
• Ada-NN2 employs a heuristic learning method, guided by local density and neighboring training points, to determine k, preserving k-NN simplicity.
• A class-based global weighting scheme (GIHS) is proposed to address class imbalance.

Main Results:

• Ada-NN and Ada-NN2 demonstrate competitive performance compared to standard k-NN and its five state-of-the-art variants.
• Extensive experiments show Ada-NN and Ada-NN2, enhanced with GIHS, outperform k-NN and its 12 imbalanced classification variants across diverse datasets.
• The proposed methods effectively handle class imbalance without significant computational overhead.

Conclusions:

• Ada-NN and Ada-NN2 offer robust solutions for optimizing k-NN classification, particularly in the presence of class imbalance.
• The integration of point-specific k selection and global imbalance handling provides a significant improvement over traditional k-NN approaches.
• The heuristic approach in Ada-NN2 maintains computational efficiency while delivering strong classification performance.