A Novel Density-Based KNN in Pattern Recognition
Sajad Haghzad Klidbary, Abazar Arabameri · 2023
The k-nearest neighbor learning algorithm is indeed one of the most commonly used supervised algorithms in machine learning. KNN is a lazy learner, and one of the most important similarity features used in this algorithm is the Euclidean distance. Although, in this paper, to achieve high accuracy, other criteria besides the Euclidean distance are considered for classification as well. In the proposed algorithm (DBKNN), first, a Gaussian function (GF) is applied to each training data in the dimensional space of the dataset. After adding all the applied Gaussians, a weight is assigned to each training data. This weight shows that the data belongs to the related class, and is used in calculating and influencing the Euclidean distance. Several simulations have been performed on the standard datasets, and the results of the simulation have confirmed proposed KNN accuracy related to the conventional KNN. It also performs very well in noisy conditions and outliers. The mathematical analysis of this algorithm is furnished to validate our assertion.