Authors:
Omid Dehzangi
1
;
Ehsan Younessian
1
and
Fariborz Hosseini Fard
2
Affiliations:
1
Nanyang Technological University, Singapore
;
2
SoundBuzz PTE LTD, Subsidiary of Motorola Inc., Singapore
Keyword(s):
Nearest neighbor, Linear discriminant analysis, Adaptive distance measure, Weight learning algorithm.
Related
Ontology
Subjects/Areas/Topics:
Decision Support Systems
;
Informatics in Control, Automation and Robotics
;
Intelligent Control Systems and Optimization
;
Modeling, Simulation and Architectures
;
Optimization Algorithms
;
Robotics and Automation
Abstract:
In this paper, an adaptive approach to designing accurate classifiers using Nearest Neighbor (NN) and Linear Discriminant Analysis (LDA) is proposed. A novel NN rule with an adaptive distance measure is proposed to classify input patterns. An iterative learning algorithm is employed to incorporate a local weight to the Euclidean distance measure that attempts to minimize the number of misclassified patterns in the training set. In case of data sets with highly overlapped classes, this may cause the classifier to increase its complexity and overfit. As a solution, LDA is considered as a popular feature extraction technique that aims at creating a feature space that best discriminates the data distributions and reduces overlaps between different classes of data. In this paper, an improved variation of LDA (im-LDA) is investigated which aims to moderate the effect of outlier classes. The proposed classifier design is evaluated by 6 standard data sets from UCI ML repository and eventuall
y by TIMIT data set for framewise classification of speech data. The results show the effectiveness of the designed classifier using im-LDA with the proposed ad-NN method.
(More)