K-nearest neighbor (KNN) is a lazy supervised learning algorithm, which depends on computing the similarity between the target and the closest neighbor(s). On the other hand, min-max normalization has been reported as a useful method for eliminating the impact of inconsistent ranges among attributes on the efficiency of some machine learning models. The impact of min-max normalization on the performance of KNN models is still not clear, and it needs more investigation. Therefore, this research examines the impacts of the min-max normalization method on the regression performance of KNN models utilizing eight different similarity measures, which are City block, Euclidean, Chebychev, Cosine, Correlation, Hamming, Jaccard, and Mahalanobis. Five benchmark datasets have been used to test the accuracy of the KNN models with the original dataset and the normalized dataset. Mean squared error (MSE) has been utilized as a performance indicator to compare the results. It’s been concluded that the impact of min-max normalization on the KNN models utilizing City block, Euclidean, Chebychev, Cosine, and Correlation depends on the nature of the dataset itself, therefore, testing models on both original and normalized datasets are recommended. The performance of KNN models utilizing Hamming, Jaccard, and Mahalanobis makes no difference by adopting min-max normalization because of their ratio nature, and dataset covariance involvement in the similarity calculations. Results showed that Mahalanobis outperformed the other seven similarity measures. This research is better than its peers in terms of reliability, and quality because it depended on testing different datasets from different application fields.
Read full abstract