BUG PREDICTION: Log approach – version approach

Bug prediction has generated widespread interest for a long time. The general scenario have been many applications with different approaches. The raising of properties of programming languages made it necessary a study about the correlation between classes, packages and files. In order to improve the quality of code and the developers satisfaction. In the last 20 years there are studied many metrics to evaluate a coding and to limit bugs. In further detail for instance, the Chidamber and Kemerer (CK) object-oriented metrics suite based on coupling between objects, numbers of children or depth of an inheritance tree. We will focus on different levels of granularity (class, method or package), paying attention on a featuring selection like to reduce or to combine metrics. In this work using the last Data analysis approach (neural network, decision tree, linear regression), it was possible to obtain models in order to predict bugs based on five relevant software system (Eclipse JDT Core - Eclipse PDE UI etc.). There models based on the purpose to have an exact prediction of numbers of bugs or to have the knowledge of a present / absence of bugs during the execution or programming. This is useful essentially for the developers to understand the relationship between bugs and software because the modern laugages programming we have a lot dependencies, just think about library, operative system and the several versions of programs. Using the CK metrics have prove that correlation between metrics is more reliable with some software system (eclipse and equinox). To obtain the same level of correlation with other software we need to combine in a specific approach the metrics (with also some interpolations). The hybrid metrics developed, correlated with the bug accurence return a level about 0.8 and 0.9 of correlation using the neural network and decision tree developed in this discussion. While the classification approach, (bug present or absence) developed for the first time with these programs, using the combinated metrics, the result will show very good values of precision, recall and accuration.