Software metrics for predicting fault-prone elements of a system at the design stage

A poorly designed software system is more difficult to code, test, and maintain, causing financial loss and time delays. In this thesis, we develop and evaluate a set of design metrics to predict software design quality—in particular, to predict which classes in an object-oriented (OO) design will be fault-prone at later stages of the software life cycle. Before forming prediction models, we need to decide which metrics to use as predictors. Before that, we need to select criteria with which to make that decision. Therefore, stage 1 of the thesis studies how and by what criteria to decide a good metric. We conduct a second-generation systematic literature review on metric evaluation criteria, resulting in an up to date list of 85 criteria. A taxonomy is proposed to organise these criteria under three broad headings: theoretical validity, practical feasibility, empirical usefulness. An order is suggested in which to consider these criteria. Stage 2 identifies a new set of design metrics for fault proneness prediction. We select a set of evaluation criteria from stage 1, and evaluate four existing sets of OO design metrics against them. Based on identified strengths and weaknesses of those metrics, we propose a consolidated set of metrics that are theoretically valid, practically feasible at the design stage, and cover a full range of OO design concepts. Stage 3 evaluates the consolidated metric set empirically for fault proneness prediction. Within a single large system (Eclipse), and across versions of that system, the consolidated set is effective at identifying fault prone classes, outperforming the other metric sets (though at a cost of more false alarms). Some insights are found regarding which types of added and changed classes are harder to predict, and on the value of older and newer data. Stages 1 and 2 significantly extend the literature on metric evaluation criteria, and the evaluation of design metrics. Stage 3 demonstrates the value of our new proposed metric set for fault proneness prediction. This can help to focus attention on classes that need redesign and more extensive testing, improving the quality of software designs