New fuzzy K-nearest neighbor algorithms for classification performance improvement

In fuzzy k-nearest neighbor, smooth class boundaries are provided by each instance’s fuzzy degree of membership. However, there are additional costs associated with calculating the memberships due to memory limitations and runtime overhead. Furthermore, in the presence of class imbalance and outliers, the effectiveness and efficiency of the most advanced fuzzy kNNs continue to decline. Thus, new fuzzy kNNs with straightforward designs are developed in this study to substantially lessen the influence of these problems and improve overall performance. The local mean vectors with the single linkage and the cumulative means of neighbors are combined, establishing these models, which are referred to as LMSL-FkNN and CMDW-FkNN, respectively. A comprehensive evaluation study spanning five experimental stages is carried out against six cutting-edge kNN competitors utilizing fifty-four real-world (balanced, imbalanced, noisy, and time series) datasets in order to illustrate the competitiveness of the established models. With CMDW-FkNN comfortably dominating the competition across the vast majority of datasets (specifically UCI, highly-Imbalanced, and Time Series datasets), the results supported by statistical tests, across three assessment metrics-accuracy, F-measure, and ROC-show that both models have significantly more promise than their rivals.