An investigation on the effect of cross project data for prediction accuracy

Abstract

Quality assurance activities include software testing, verification, validation, fault proneness and fault prediction. Most of existing model predicts faults based on historical data (Within-project) through intellectual analysis. This paper instanced on defect prediction using cross-projects and mixed-project prediction because of the unpredictability in selection of software attributes by analogy based approach that deliver imprecise and ambiguous solution. Cross project delivers momentous results by focussing on selection of training data. Combination of cross project data with regression techniques improves effectiveness of prediction by extracting similar features impacted by all datasets. Feature extraction with similarity based approach creates training model that constitute more accurate prediction. We proposed two experiments with method-level and class level datasets on 23 open source projects with different scenario that shows improvement in accuracy, probability of prediction and probability of false alarm with the involvement of cross project and mixed project combinations. Major outcomes from our experiment conclude that cross project prediction provide better prediction by selecting dataset either by iterative versions of historical information or addition of limited within project data to cross project data.