Paper Info
Title
An Enhanced K-Nearest Neighbor Classification Method Based on Maximal Coherence and Validity Ratings.
Abstract
Traditional k-nearest neighbor methods couldn't be able to correctly classify objects when their k nearest neighbors are dominated by other classes. This paper formulates a two-class classification problem, and applies a modified k-nearest neighbors (KNN) classifier algorithm based on maximal coherence, validity ratings, and k-fold cross validation to classify the test samples. We build a validity score for the pairs of sample and their surroundings according to their labels. The k nearest neighbors (including the unknown test object) of each sample in the training set as well as the unknown test object itself will be determined. The unknown test object will be tentatively assigned to a class membership. Then we use the validity scores to quantify the degree to which a pre-determined group of samples resemble their k nearest neighbors. A classifier is designed which take into account the coherence and validity ratings. A numerical example demonstrates the effectiveness of the algorithm in detail. The enhanced KNN method is compared with the conventional KNN and the modified KNN method on both real world wine data and photo-thermal infrared imaging spectroscopy (PT-IRIS) data for up to 20 different k values. Classification accuracy of KNN method and our method in terms of various combinations of k-value and k-fold cross validation are compared. The experimental results show that the proposed enhanced KNN method outperforms the conventional KNN and the modified KNN method on both real world wine data and PT-IRIS data. In addition, the classification accuracy of both the conventional KNN and our method increase drastically when k = 5. The average classification accuracy of our method on the PT-IRIS data featuring small sample size and high overlap is 97.87%.
Year
DOI
Venue
2017
10.1007/978-3-319-59072-1_25
ADVANCES IN NEURAL NETWORKS, PT I
Keywords
Field
DocType
K-nearest neighbors (KNN) classification,Modified k-nearest neighbors (MKNN),Supervised classification,K-fold cross validation
k-nearest neighbors algorithm,Training set,Pattern recognition,Computer science,Test object,Coherence (physics),Artificial intelligence,Classifier (linguistics),Cross-validation,Sample size determination,Machine learning
Conference
Volume
ISSN
Citations 
10261
0302-9743
0
PageRank 
References 
Authors
0.34
4
5
Name
Order
Citations
PageRank
Nian Zhang17112.03
Jiang Xiong202.03
Jing Zhong3476.21
Lara A. Thompson401.35
Hong Ying501.35