An Improvement To The k-Nearest Neighbor Classifier For ECG Database

Malaysian Technical Universities Conference on Engineering and Technology 2017

Jaafar, Haryati^1 ; Ramli, Nur Hidayah^1 ; Abdul Nasir, Aimi Salihah^1

Faculty of Engineering Technology, Universiti Malaysia Perlis, UniCITI Alam Campus, Sg. Chuchuh, Padang Besar, Perlis

02100, Malaysia^1

关键词: Classification performance;    Electrocardiogram signal;    Fuzzy membership function;    K nearest centroid neighbors;    K nearest neighbor (KNN);    K-nearest neighbor classifier;    Mahalanobis distances;    Non-parametric classifiers;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/318/1/012046/pdf
DOI  :  10.1088/1757-899X/318/1/012046

来源: IOP

【 摘 要 】

The k nearest neighbor (kNN) is a non-parametric classifier and has been widely used for pattern classification. However, in practice, the performance of kNN often tends to fail due to the lack of information on how the samples are distributed among them. Moreover, kNN is no longer optimal when the training samples are limited. Another problem observed in kNN is regarding the weighting issues in assigning the class label before classification. Thus, to solve these limitations, a new classifier called Mahalanobis fuzzy k-nearest centroid neighbor (MFkNCN) is proposed in this study. Here, a Mahalanobis distance is applied to avoid the imbalance of samples distribition. Then, a surrounding rule is employed to obtain the nearest centroid neighbor based on the distributions of training samples and its distance to the query point. Consequently, the fuzzy membership function is employed to assign the query point to the class label which is frequently represented by the nearest centroid neighbor Experimental studies from electrocardiogram (ECG) signal is applied in this study. The classification performances are evaluated in two experimental steps i.e. different values of k and different sizes of feature dimensions. Subsequently, a comparative study of kNN, kNCN, FkNN and MFkCNN classifier is conducted to evaluate the performances of the proposed classifier. The results show that the performance of MFkNCN consistently exceeds the kNN, kNCN and FkNN with the best classification rates of 96.5%.