Code Smell Severity Classification using Machine Learning Techniques

Code Smell Severity Classification using Machine Learning Techniques

1. The aim is to assess the performances of code smells severity classification, and select the best models to carry out this task

2. Recommending to fix only the smells with highest severity

3. code smell detection

   • Data modeling: supervised learning. The subjects of the classification are classes or methods
   • Data collection: Qualitas Corpus. 76 systemas were choosen and completed
     • dependent variables: Class level: God Class and Data Class; method level: Long Method and Feature Envy
   • Build Dataset: sampling through Advisors (tools and rules availables)
   • Code smells severity classification: 0 No smell; 1 Non-severe smell; 2 Smell; 3 Severe smell

4. Result: The result is that RandomForests regression is the best model for God Class and Data Class, while RandomForests and decision trees, combined with boosting and the experimented ordinal classification method, have generally the best performances on all datasets.