Comparative Analysis Between the Traditional and Nearest Centroid Methods in the K-Nearest Neighbor Algorithm

One of the most commonly used algorithms in machine learning is the method of the nearest neighbors. The key points in this algorithm are determining the parameter k or the number of nearest neighbors and how to select these neighbors. Whether this choice should be based on the distance between objects in the set or use modifications such as the nearest centroid method. This publication describes a study and comparison of the two types of KNN, conducted using 18 different public data sets that were selected to correspond to the data that a small business would have - small learning sets, heterogeneous data, data with anomalies, unbalanced data. A more unusual category of data has been added to the sets used - those that have the problem n>>p, ie. the number of attributes is many times greater than the number of objects. Experimental studies have shown that there are statistically significant improvements in the effectiveness of nearest centroid neighborhood classification in two of the training sets.