Comparison of local classifiers for cross-project defect prediction

There is a connection between static source code metrics, for example, lines of code or cyclomatic complexity and potential defects in the source code. Obviously, there is no closed formula, but with the field of machine learning and its techniques we have a tool at our disposal that has the ability to infer rules from large amounts of data. In this thesis, we use machine learning techniques to predict defects in software components from source code metrics. We use a subset of a publicly available dataset from open source projects to train and evaluate our machine learning models. As we do cross-project defect prediction, we predict defects in a software project using classifiers trained with the source code metrics of other projects. In this thesis, we evaluate the cross-project defect prediction performance of two local classifier implementations. Local classifiers partition the training data first, then train a classifier for each partition. This leads to more homogeneous data and should yield better prediction results than training one global classifier with the complete data. We run six experiments which compare our two local classifier implementations with an existing global classifier, which already yields a high success rate in cross-project defect prediction. These experiments are conducted using the CrossPare framework for ease of implementation and comparability of results. The experiments make use of two different methods of bias handling: equal weighting and undersampling. This will additionally show how these techniques influence our local classifiers and the prediction performance.