Cluster-Based Nearest-Neighbour Classifier and Its Application on the Lightning Classification

The problem addressed in this paper concerns the prototype generationfor a cluster-based nearest-neighbour classifier. It considers, toclassify a test pattern, the lines that link the patterns of thetraining set and a set of prototypes. An efficient method based onclustering is here used for finding subgroups of similar patterns withcentroid being used as prototype. A learning method is used for iterativelyadjusting both position and local-metric of the prototypes. Finally, weshow that a simple adaptive distance measure improves the performanceof our nearest-neighbour-based classifier.The performance improvement with respect to othernearest-neighbour-based classifiers is validated by testing our methodon a lightning classification task using data acquired from the Fast On-orbitRecording of Transient Events (FORTE) satellite, moreover theperformance improvement is validated through experiments with severalbenchmark datasets.The performance of the proposed methods are also validated using theWilcoxon Signed-Rank test.