The nearest neighbor algorithm of local probability centers.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Conventional nearest neighbor classifiers struggle with nonseparable or overlapping classes.
• Misclassified samples in local neighborhoods can lead to incorrect predictions.
• Existing methods may not effectively handle ambiguous data points.

Purpose of the Study:

• To propose a new classification method addressing limitations of conventional nearest neighbor algorithms.
• To improve classification accuracy in scenarios with overlapping or nonseparable classes.
• To reduce the impact of mislabeled or ambiguous training data.

Main Methods:

• Developed a classification method based on local probabilistic centers of each class.
• Reduced the influence of 'negative contributing points' (misplaced samples) in the training set.
• Classified query samples using distance to centers and computed posterior probability.

Main Results:

• The proposed method significantly enhances classification performance compared to standard nearest neighbor algorithms.
• Using posterior probability for classification yielded smaller classification errors in experiments.
• Theoretical analyses and experiments on constructed and real datasets validated the method's effectiveness.

Conclusions:

• The new classification method offers a substantial improvement for nearest neighbor algorithms, especially with complex datasets.
• Effective handling of local probabilistic centers and negative contributing points is key to improved accuracy.
• The approach provides a robust solution for classification tasks involving overlapping data distributions.