A single pairwise model for classi cation using online learning with kernels

Abstract

Any binary or multi-class classiffcation problem can be transformedinto a pairwise prediction problem. This expands the data and bringsan advantage of learning from a richer set of examples, in the expenseof increasing costs when the data is in higher dimensions. Therefore,this study proposes to adopt an online support vector machine to workwith pairs of examples. This modiffed algorithm is suitable for largedata sets due to its online nature and it can also handle the sparsitystructure existing in the data. Performances of the pairwise setting andthe direct setting are compared in two problems from different domains.Results indicate that the pairwise setting outperforms the direct settingsigniffcantly. Furthermore, a general framework is designed to use thispairwise approach in a multi-class classiffcation task. Result indicatethat this single pairwise model achieved competitive classiffcation rateseven in large-scaled datasets with higher dimensionality.