A new KNN classifier based on the harmonic mean of the majority votes and average distances of the neighbors combined with adaptive K-value selection

This study aims to improve the classification accuracy of traditional KNN by presenting an improved version of the KNN algorithm. This thesis proposes HMAKNN, an improved KNN classifier based on the harmonic mean of the majority voting and average distance phenomena with adaptive and incremental k-value selection. Within the purview of this study, two variants of HMAKNN, regular and weighted, were designed based on the presence or absence of a weighting mechanism. The names of these variants are HMAKNNR and HMAKNNW, respectively. These HMAKNN classifiers were evaluated on a total of thirty-four data sets, of which eight were synthetic obtained from PRTools and twenty-six were real benchmark data sets from UCI and Kaggle repositories. To determine the classification effectiveness and success of the proposed approaches, they were compared with traditional KNN and four well-known weighted KNN models. In contrast to other weighting methods, both HMAKNN classifiers utilize the constructive interaction between majority voting and average distance, as well as the ability to adaptively modify the k-value, thereby substantially enhancing classification accuracy. Based on the results of the classification of real benchmark data sets, HMAKNNW produced the highest ACC value with 83.02%, and HMAKNNR reached the second-highest value with 82.83%. HMAKNNW and HMAKNNR provided an advantage of 2.69% and 2.50% over the other five methods compared, respectively. Furthermore, the classification results of real benchmark data sets indicate that both HMAKNN methods statistically outperform other weighted KNN methods.