An Empirical Investigation Into Fault-proneness Ranking Recommendation Models Based On Feedback

Previous studies have shown that traditional supervised machine learning methods based fault-proneness prediction models in general have a good performance. However, these models are based on a large number of historical fault information on modules. In the area of software engineering, the collection of the fault information from modules is expensive and time-consuming. In this context, it is very valuable to explore how to build an effective prediction model using only a small number of modules with fault information.In order to deal with this problem, this paper adopts a feedback-based image retrieval framework to obtain a fault-proneness ranking recommendation model. In order to select an appropriate learner for this framework, we empirically investigate the following learners:binary logistic regression, artificial neural networks, NaiveBayes, J48, and a similarity based semi-supervised co-training learner Ssadp. In particular, we use "manualUp"(sorting modules in ascending order according to their lines of code) as the baseline model. Based on12datasets from three open-source systems, our results show that:when50%to60%of the minimum modules are used, logistic regression and ANN have a better performance than manualUp, while NaiveBayes and J48have a similar performance to manualUp. Compared with other classifiers, Ssadp has a better performance if the data are properly preprocessed. In practice, in order to obtain an effective ranking model, we suggest that:first, sort the top50%to60%minimum modules in ascending order according to their source lines of code; then, if necessary, use their label information obtained from inspection/testing combined with Logistic, ANN, or Ssadp to build a fault-proneness prediction model and apply this prediction model to ranking the remaining modules in the system.The main contributions of this thesis are summarized as follows:1) We introduced a fault-proneness ranking recommendation model. This model is based the structural metrics from source code and employ users’ feedback to improve the prediction performance;2) We use seven open-source software systems to compare the performances of the above-mentioned model when difference leaners are used. This provides valuable guidance for selecting an appropriate learner for the model.