An Enhancement of Fuzzy K-Nearest Neighbor Classifier Using Multi-Local Power Means

This study introduces a new method to the family of fuzzy k-nearest neighbor (FKNN) classifiers that is based on the use of power means in the calculation of multi-local means that are used in classification of samples. The proposed new classifier is called multi-local power means fuzzy k-nearest neighbor classifier (MLPM-FKNN). The proposed method can be adapted to the context (of different data sets), due to the power mean being parametric and thus allowing for testing to find the parameter value that can be optimized for the classification accuracy. Furthermore, we can find optimal value for the number of local observations used in calculation of the multi-local mean. The proposed method is usable for example in situations, where class distribution is significantly different and there is only few observations in some classes. The performance of the MLPM-FKNN classifier is studied by testing it with four datasets. The performance is benchmarked against that of the original k-nearest neighbor and the fuzzy k-nearest neighbor classifiers. We find that MLPM-FKNN classifier is able to reach a statistically significantly higher classification accuracy than the benchmarks used and has reasonable performance metrics.