Comparative analysis between the traditional K-Nearest Neighbor and Modifications with Weight-Calculation

Tsvetelina Mladenova, Irena Valova

20222022 International Symposium on Multidisciplinary Studies and Innovative Technologies (ISMSIT)23 citationsDOI

Abstract

This paper describes our experimental study and comparison of the standard k-nearest neighbor (KNN) method with modifications of the standard KNN, based on the calculated weights for every neighbor with some well-known and researched methods for weight-calculation, such as Weighted K- Nearest Neighbor, Distance- Weighted K-Nearest Neighbor, Dual Weighted K-Nearest Neighbor, etc. Experiments were conducted with 18 different datasets, in which the results were compared and shown graphically for further analyzation. It can be seen that the nearest neighbor method, KNN, without additional calculation of neighbor weights, performs the worst in the majority of the experiments. Depending on the used dataset and the classification problem, the most accurate method varies. A drop in the accuracy rate is also observed as the number of neighbors increases.

Topics & Concepts

k-nearest neighbors algorithmNearest-neighbor chain algorithmBest bin firstNearest neighbor searchPattern recognition (psychology)Large margin nearest neighborFixed-radius near neighborsNearest neighbor graphComputer scienceNearest neighbourMathematicsAlgorithmArtificial intelligenceData miningCluster analysisCanopy clustering algorithmCorrelation clusteringImbalanced Data Classification TechniquesAnomaly Detection Techniques and ApplicationsMachine Learning and Data Classification