Hybridization of Learning Vector Quantization (LVQ) and Adaptive Coordinates (AC) for data classification and visualization

• Main Authors: Md. Sarwar, Zahan Tapan, Chee, Siong Teh
• Format: E-Article
• Language: English
• Published: IEEE 2008
• Subjects: T Technology (General)
• Online Access: http://ir.unimas.my/id/eprint/16655/1/Hybridization%20of%20Learning%20Vector%20Quantization%20%28abstract%29.pdf; http://ir.unimas.my/id/eprint/16655/; http://ieeexplore.ieee.org/document/4658440/

Most of the artificial neural network (ANN) methods do not support data classification and visualization simultaneously. Some ANN methods such as learning vector quantization (LVQ), multi-layer perceptrons (MLP) and radial basis function (RBF) perform classification without any visualization. Excellent data visualization on the other hand has been prominently supported by various unsupervised methods such as self-organizing maps (SOM) and its recent variants of visualization induced SOM (ViSOM) and probabilistic regularized SOM (PRSOM). However, being unsupervised these methods do not optimize classification accuracy compared with the supervised classification methods such as LVQ. Thus, the scope of a novel supervised method is felt necessary to facilitate applications requiring good data visualization and intensive classification. LVQ demonstrates classification performance at least as high as other supervised ANN classifiers. Adaptive coordinate (AC) on the other hand, has demonstrated the ability of mirroring weight vectorspsila movements in N-dimensional input space to low dimensional output space to reveal the clustering tendency of data learned by SOM. This mirroring concept motivates this work to hybridize a modified AC with LVQ (LVQwihAC) to support data visualization and classification simultaneously. Empirical studies on benchmark data sets demonstrated that, the LVQwihAC method provides better classification accuracy than the unsupervised methods of SOM, ViSOM and PRSOM besides its promising data visualization with higher computational efficiency. The classification performance is also found at least as good as other supervised classifiers with additional data visualization abilities over them.