A Classifier Based on K-Nearest Neighbors Using Weighted Summation of Reconstruction Errors

Abstract

In this paper, a classifier is introduced based on the nearest neighbor classifier and the reconstruction error for data classification. In the proposed method, first, K nearest data points (neighbors) from each category in the training data are calculated for the test data point. Then, the reconstruction of the test data is performed based on different numbers of nearest neighbors (from one to K) in each category, and the reconstruction error is calculated separately for each number of neighbors. In the next step, for each category, the error is calculated as the weighted sum of the errors obtained from all reconstructions. The weight of the reconstruction error is proportional to the number of neighbors involved in it, so the reconstruction error is multiplied by the number of neighbors. Finally, the test data belongs to the category with the lowest overall error. This process allows a combination of K nearest neighbor classifiers to play a role in data classification. In this paper, 10 datasets from the UCR time series database and five datasets from the UCI classification database are used to evaluate the proposed method. The results of these evaluations show that the proposed method significantly improves the performance of the minimum reconstruction error based KNN classifiers, achieving approximately 5% better recognition rate for some K values and an average recognition rate improvement of about 1.6% for all K values (from 2 to 15)..

Keywords

• classifier
• recognition rate
• k-nearest neighbors
• linear reconstruction
• weighted combination