ADAPTIVE_CLASSIFIER: A NOVEL APPROACH FOR KNN PARAMETERIZATION

Abstract

The k-Nearest Neighbors (kNN) algorithm is a widely used method in machine learning, particularly for pattern recognition and classification tasks. Despite its popularity, the determination of the k parameter, which influences the number of nearest neighbors to consider, remains a significant challenge. This research introduces a novel enhancement to the kNN algorithm that specifically targets the calculation of the k parameter, aiming to improve the algorithm's performance across various data mining tasks. Building on a rich body of work, the study proposes a mathematical technique for intelligent clustering of points for classification. This method involves creating an n-dimensional sphere, where n equals the number of features, effectively eliminating the manual trial and error approach traditionally associated with the kNN algorithm. The equation of the sphere is formulated based on the dimension and feature values, with the radius determined through a specific mathematical equation. The proposed algorithm, named Adaptive classifier, was tested against six existing unsupervised algorithms using the diabetes dataset. The results demonstrated superior performance of the Adaptive classifier, with statistical proof provided through a paired t-test. The t-test results showed that the Adaptive Classifier outperforms the other algorithms, reinforcing its effectiveness. This research contributes significantly to the field of machine learning by offering a more efficient and accurate method for k parameter calculation in the kNN algorithm. By improving the algorithm's performance, the proposed method has the potential to enhance results in various data mining tasks, thus broadening the applicability of the kNN algorithm