Machine Learning for Reducing the Effort of Conducting Systematic Reviews in SE

Sammanfattning: Objective : To investigate whether machine learning and text-based data mining can be used to support the primary studies selection process and decrease the needed efforts in systematic reviews conducted in the context of SE. Research Design : A test collection was built from 3 systematic reviews used in previous work in the context of SE. The proposed probabilistic classifier based on Bayes’ Theorem was constructed to predict and classify each article as containing high-quality evidence to warrant inclusion in study selection process or not. Feature engineering techniques were applied to the abstract-based features. Cross-validation experiments were performed to evaluate the efficiency of the document classifier. Three metrics - precision, recall and specificity were used together to measure the classification performance. We assume that a recall rate of 0.9 or higher is required for the classifier to identify an sufficient quantity of relevant papers. As long as recall is at least 0.9, the Precision and Specificity should be as high as possible,. Results : From the hold-out cross validation experiment, the precision achieved with the classifier for two systematic review topics, was 93%, while 79% for another systematic review topic. The results of leave-one-out cross validation experiment were presented in three Confusion Matrix, which in detail indicated that the precision achieved with the classifier for the three systematic review topics was promising in terms of predicting relevant abstracts while relatively poor in terms of excluding irrelevant articles. Conclusion : The classifier based on Bayes’ Theorem has strong potential for performing the systematic review classification tasks in software engineering. The approach presented in this paper could be considered as a possible technique for assisting labor-intensive primary studies’ selection process in an SLR.