Crowdsmelling: A preliminary study on using collective knowledge in code smells detection

Code smells are seen as major source of technical debt and, such, should be detected and removed. However, researchers argue that the subjectiveness code detection process is a hindrance to mitigate problem smells-infected code. This paper presents results validation experiment for Crowdsmelling approach proposed earlier. The latter based on supervised machine learning techniques, where wisdom crowd (of software developers) used collectively calibrate algorithms, thereby lessening subjectivity issue. In context three consecutive years Software Engineering course, total “crowd” around hundred teams, with an average members each, classified presence 3 (Long Method, God Class, Feature Envy) in Java These classifications were basis oracles training six algorithms. Over one models generated evaluated determine which algorithms had best performance detecting each aforementioned smells. Good performances obtained Class (ROC= 0.896 Naive Bayes) Long Method 0.870 AdaBoostM1), but much lower Envy 0.570 Random Forrest). suggest feasible Further experiments dynamic required comprehensive coverage increase external validity.