Perceptron-Like Large Margin Classifiers

Download

PDF Download (1185Kb)

Description/Abstract

We address the problem of binary linear classification with emphasis on algorithms that lead to separation of the data with large margins. We motivate large margin classification from statistical learning theory and review two broad categories of large margin classifiers, namely Support Vector Machines which operate in a batch setting and Perceptron-like algorithms which operate in an incremental setting and are driven by their mistakes. We subsequently examine in detail the class of Perceptron-like large margin classifiers. The algorithms belonging to this category are further classified on the basis of criteria such as the type of the misclassification condition or the behaviour of the effective learning rate, i.e. the ratio of the learning rate to the length of the weight vector, as a function of the number of mistakes. Moreover, their convergence is examined with a prominent role in such an investigation played by the notion of stepwise convergence which offers the possibility of a rather unified approach. Whenever possible, mistake bounds implying convergence in a finite number of steps are derived and discussed. Two novel families of approximate maximum margin algorithms called CRAMMA and MICRA are introduced and analysed theoretically. In addition, in order to deal with linearly inseparable data a soft margin approach for Perceptron-like large margin classifiers is discussed. Finally, a series of experiments on artificial as well as real-world data employing the newly introduced algorithms are conducted allowing a detailed comparative assessment of their performance with respect to other well-known Perceptron-like large margin classifiers and state-of-the-art Support Vector Machines.