Selecting features in origin analysis

Abstract

When applying a machine-learning approach to develop classifiers in a new domain, an important question is what measurements to take and how they will be used to construct informative features. This paper develops a novel set of machine-learning classifiers for the domain of classifying files taken from software projects; the target classifications are based on origin analysis. Our approach adapts the output of four copy-analysis tools, generating a number of different measurements. By combining the measures and the files on which they operate, a large set of features is generated in a semi-automatic manner. After which, standard attribute selection and classifier training techniques yield a pool of high quality classifiers (accuracy in the range of 90%), and information on the most relevant features. © 2011 Springer-Verlag London Limited.