An improved KNN classifier based on a novel weighted voting function and adaptive k-value selection

This paper presents a modified KNN classifier (HMAKNN) based on the harmonic mean of the vote and average distance of the neighbors of each class label combined with adaptive k-value selection. Within the scope of this study, two different versions of HMAKNN, regular and weighted, HMAKNN(R) and HMAKNN(W), were developed depending on whether there is a weighting mechanism or not. These proposed HMAKNN classifiers were tested eight syntetic and twenty-six real benchmark data sets. In order to reveal the effectiveness and the performance of the proposed methods on classification, they were compared with its constituent KNN and four other well-known distance-weighted KNN methods. Unlike other weighting methods, both HMAKNN classifiers use the synergy between majority voting and average distance together, along with the ability to adaptively adjust the k-value, helping to significantly improve classification accuracy. The results on twenty-six real benchmark data sets suggest that both HMAKNN methods produce more accurate results in terms of average ACC and FScore metrics and statistically outperform all competing methods.