Improved KNN Algorithm Based on Local K Value Fitting

Abstract: Aiming at the problem that the k value of the dependent sample is difficult to determine due to the absence of the training process in the k-nearest neighbor algorithm (KNN), an improved KNN algorithm based on local k value fitting (CF-LKNN) is proposed, in which the local k value of the sample is determined by adding the learning process of the sample distribution. Firstly, a curve fitting method is used to solve the equation to obtain an overall k value through similarity measurement. Secondly, the local k-values of neighboring samples within the neighborhood of the classification sample with the overall k value as the radius are calculated. Finally, the class labels are determined by voting based on the local k values of different samples to be classified. The experiment showed that the lowest classification error rate of CF-LKNN was reduced by an average of 3.53% compared to 6 improved KNN algorithms on 8 UCI datasets, and the accuracy improvement was more significant on the balanced dataset compared to the classical KNN algorithm on three indicators on 12 UCI datasets.