Rank-based univariate feature selection methods on machine learning classifiers for code smell detection

TL;DR: In this paper, the authors have implemented 32 machine learning algorithms after performing feature selection through six variations of the filter method, including Mutual Information, fisher score, and univariate ROC-AUC feature selection techniques with brute force and random forest correlation strategies.

Abstract: Detecting code smells and treating them with refactoring are trivial part of maintaining vast and sophisticated software. There is an urgent need for automatic system to treat code smells. Tools provide variable results, based on threshold values and subjective interpretation of smells. Machine learning is one of the best approaches that provides effective solution to this problem. Practitioners do not need expert knowledge on smell’s characteristics for detection, which makes this approach accessible. In this paper, we have implemented 32 machine learning algorithms after performing feature selection through six variations of the filter method. We have used multiple correlation methodologies to discard similar features. Mutual information, fisher score, and univariate ROC–AUC feature selection techniques were used with brute force and random forest correlation strategies. Feature selection eliminates dimensionality curse and improves performance measures drastically. It is the selection of relevant feature subset based on the relation between dependent and independent variables. We have compared performance of classifiers implemented with and without performing feature selection. Results show that accuracy of machine learning models has increased up to 26.5%, f-measure by 70.9%, area under ROC curve has surged up to 26.74%, and average training time has reduced up to 62 s as compared to performance measures of machine learning models executed without feature selection. Mutual information feature selection strategy with random forest correlation methodology has the highest impact on performance measures among all the filter methods. Among 32 classifiers, boosted decision trees (J48) and Naive Bayes algorithms gave best performance after dimensionality reduction.