On enhancing data classification using local mean-based fuzzy K-nearest neighbor algorithms

Abstract The fuzzy k-nearest neighbor comes to tackle the voting problem of the standard kNN, in which the same weight is assigned to each labeled sample, leading to a detrimental impact on the kNN performance. In fuzzy kNN, each instance’s fuzzy membership is found to produce seamless boundaries between classes. However, the computation of the memberships comes at additional costs due to the runtime overhead and memory requirements. Further, state-of-the-art fuzzy kNNs’ performance still suffers degradation because of class imbalance and outliers. To mitigate the impact of both problems, this study therefore develops two novel fuzzy models: Local Mean and Adaptive Learning fuzzy kNN (LMAL-FkNN) and Local Mean and Global Learning Fuzzy kNN (LMGL-FkNN). This is done by combining local mean vectors with class-based means and calculating the average local and global linkages in LMAl-FkNN and LMGL-FkNN, respectively. By calculating the local and class-based means of the global and local neighbors and using these means to compute the distance to the query, the impact of class imbalance is significantly lessened. Further, by incorporating class-based neighbors and using their means to find the final membership degrees, the outlier effects are substantially reduced. This dual approach enhances the robustness of the proposed models and thus improves the overall performance. To demonstrate the models’ competitiveness, a thorough evaluation study in five experimental phases is conducted against five state-of-the-art kNN rivals using forty-seven datasets. The results show that the LMGL-FkNN, in particular, has far more potential than its competitors over the vast majority of datasets.