A vector quantization method for nearest neighbor classifier
A vector quantization method for nearest neighbor classifier design Source: Pattern Recognition Letters, Vol. 25, 2004, pp. 725 -731 Author: Chen-Wen Yen, Chieh-Neng Young and Mark L. Nagurka Speaker: Guey-Tzu Chang Date: May 17, 2004/5/17 國立中正大學資訊 程所 1
Nearest neighbor Class 1 data Classifier : Class 2 Class n Condensed Nearest Neighbor (CNN), Hart, 1968 : 2004/5/17 VQ Nearest Neighbor (VQ-NN), Xie, 1993 Adaptive VQ Nearest Neighbor (AVQ-NN), Yen et al. , 2004 國立中正大學資訊 程所 2
Nearest neighbor (Cont. ) Factors: 1. Accuracy 2. Number of prototypes 2004/5/17 國立中正大學資訊 程所 3
VQ-NN Class 1 Class 2 : : Class n Code book (prototype set) 2004/5/17 國立中正大學資訊 程所 5
VQ-NN (Cont. ) • Drawback: – It does not consider the interaction among different classes of samples – Difficult to design an NN classifier that has an optimal number of prototypes 2004/5/17 國立中正大學資訊 程所 6
AVQ-NN Validation set Training set Class 1 Cluster NN Cluster-NN Class 2 : Class n prototype set (initial) Class 1 2 1 Class 2 0 Class n : prototype set N retraining 2004/5/17 國立中正大學資訊 程所 Error < th Y terminate 7
Experimental results Smaller data sets: Wisconsin breast cancer Australian credit card 2004/5/17 國立中正大學資訊 程所 8
Experiment results (Cont. ) Large data sets: Phoneme Kr-vs-Kp 2004/5/17 國立中正大學資訊 程所 9
Conclusion • The proposed approach can achieve high classification accuracy with a relatively small number of prototypes. • Another possible future direction is to investigate the sensitivity of the proposed method to the training set size. 2004/5/17 國立中正大學資訊 程所 10
- Slides: 10