An Improved Lvq Algorithm With Data-Structure Preserving Visualization

Sarwar Tapan,Dianhui Wang
2012-01-01
Abstract:Data-structure preserved visualization of high-dimensional data reveals the dataset borders and the spread and overlapping tendency of the class borders in a more informative manner than the usual data-topology preserved mapping produced by Self-Organizing Maps (SOMs). Hence, an extension of SOM called Probabilistic Regularized SOM (PRSOM) is proposed for the data-structure preservation in the visualization; however, PRSOM is less suitable for the classification task due to its regularized positioning of the prototypes. In many practical applications, a good classification rate and data-structure informative visualization of high-dimensional data are simultaneously required from an employed method. However, it is difficult to find a method in the current literature that can perform these two tasks effectively. This paper proposes a variant of the Learning Vector Quantization (LVQ) algorithm as Data-Structure Preserving LVQ (LVQ(dsp)) by combining the classical LVQ1 algorithm with a proposed visualization mechanism, to support these two tasks on high-dimensional datasets. Simulations on several benchmark datasets demonstrated LVQ(dsp)'s promising capability of producing data-structure preserving visualizations in addition to offering excellent classification rates.
What problem does this paper attempt to address?