Fuzzy KNN Method With Adaptive Nearest Neighbors.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• The fuzzy k-nearest-neighbor (FKNN) method excels with uncertain data but suffers when the number of neighbors (k) is fixed.
• Suboptimal fixed k values significantly degrade FKNN classification performance.

Purpose of the Study:

• To develop a novel FKNN-based classification method that adaptively determines the optimal k value for each testing sample.
• To enhance classification accuracy and efficiency in fuzzy k-nearest-neighbor algorithms.

Main Methods:

• A novel fuzzy KNN method with adaptive nearest neighbors (A-FKNN) was developed.
• A-FKNN learns optimal k values during training, builds a decision tree (A-FKNN tree), and uses it to find the optimal k for testing samples.
• A faster version, FA-FKNN, was created using a decision tree with subsets of training samples.

Main Results:

• Both A-FKNN and FA-FKNN demonstrated superior classification accuracy compared to existing methods on 32 UCI datasets.
• FA-FKNN achieved a shorter running time, indicating improved computational efficiency.
• The adaptive approach effectively addresses the limitations of fixed k values in FKNN.

Conclusions:

• A-FKNN and FA-FKNN offer significant improvements over traditional FKNN by adapting the k value per sample.
• The proposed methods enhance classification performance and efficiency, making them valuable for handling uncertain and ambiguous data.
• FA-FKNN provides a computationally efficient alternative without compromising accuracy.