Recommending Move Method Refactorings Using Dependency Sets (WCRE 2013)

Methods implemented in incorrect classes are common bad smells in object-oriented systems, especially in the case of systems maintained and evolved for years. To tackle this design flaw, we propose a novel approach that recommends Move Method refactorings based on the set of static dependencies established by a method. More specifically, our approach compares the similarity of the dependencies established by a source method with the dependencies established by the methods in possible target classes. We evaluated our approach using systems from a compiled version of the Qualitas Corpus. We report that our approach provides an average precision of 60.63% and an average recall of 81.07%. Such results are, respectively, 129% and 49% better than the results achieved by JDeodorant, a well known move method recommendation system.

Transcript

1. Recommending Move Method
   Refactorings using Dependency Sets
   Vitor Sales, Ricardo Terra,
   Luis Miranda, Marco Tulio Valente
   1
   WCRE, Koblenz, Germany, October 2013
   

   View Slide

2. Introduction
    “Move method” recommendation system
    Recommendations based on dependency sets
   2
   

   View Slide

3. Recommendation Algorithm
    Main algorithm:
   Input: method m, and class C
   for each class C’
   if similarity (m, C’) > similarity(m,C)
   then C’ is a candidate to receive M
   3
   

   View Slide

4. Recommendation Algorithm
    Main algorithm:
   Input: method m, and class C
   for each class C’
   if similarity (m, C’) > similarity(m,C)
   then C’ is a candidate to receive M
    How to calculate similarity (m, C)?
   4
   

   View Slide

5. Method-Class Similarity
   
   5
   

   View Slide

6. Method-Class Similarity
   
   6
   

   View Slide

7. Method-Method Similarity
    Similarity of the dependency sets of each method
    Deps(m):
    classes of the methods called by m
    classes of the fields acessed by m
    classes of the objects created by m
    return type of m
    annotations used by m
    etc
   7
   

   View Slide

8. Method Similarity Function
    meth_sim(m1
   ,m2
   )=
   similarity (Deps(m1
   ), Deps(m2
   ))
   8
   

   View Slide

9. Method Similarity Function
    meth_sim(m1
   ,m2
   )=
   similarity (Deps(m1
   ), Deps(m2
   ))
    Which set similarity coefficient should we use?
    Jaccard, Simple Matching, Yule etc
   9
   

   View Slide

10. Exploratory Study
     System: JHotDraw
     Assumption:
     all methods implemented in the correct class
     Best coefficient:
     The one that generates few recommendations
    10
    

    View Slide

11. # Recommendations, JHotDraw
    11
    

    View Slide

12. # Recommendations, JHotDraw
    12
    

    View Slide

13. Best Coefficient
     Soakal and Sneath 2:
    13
    

    View Slide

14. Tool Support: JMove
    14
    

    View Slide

15. Evaluation
     Comparison with JDeodorant:
     N. Tsantalis and A. Chatzigeorgiou, “Identification of move
    method refactoring opportunities” IEEE TSE, 2009
     How does JMove compares with JDeodorant in terms of
     Precision
     Recall
    15
    

    View Slide

16. Target Systems
     15 systems (qualitas.class)
    16
    

    View Slide

17. Gold Sets
     We manually moved 475 methods to new classes
     Source methdods and target classes randonly selected
    17
    

    View Slide

18. Gold Sets
     We manually moved 475 methods to new classes
     Source methdods and target classes randonly selected
     High chances they are located in the wrong classes
    18
    

    View Slide

19. Gold Sets
     We manually moved 475 methods to new classes
     Source methdods and target classes randonly selected
     High chances they are located in the wrong classes
    19
    

    View Slide

20. Precision
    20
     Only for JHotDraw (5 instances)
     Recs not in GoldSets → False positives
     In the other systems, we do not know
    

    View Slide

21. Precision
    21
     Only for JHotDraw (5 instances)
     Recs not in GoldSets → False positives
     In the other systems, we do not know
    

    View Slide

22. Precision
    22
     Only for JHotDraw (5 instances)
     Recs not in GoldSets → False positives
     In the other systems, we do not know
    

    View Slide

23. Precision
    23
     Only for JHotDraw (5 instances)
     Recs not in GoldSets → False positives
     In the other systems, we do not know
    

    View Slide

24. Recall
    24
     recall1
    : method and target class recommended correctly
     recall2
    : only method recommended correctly
    

    View Slide

25. Recall
    25
     recall1
    : method and target class recommended correctly
     recall2
    : only method recommended correctly
    

    View Slide

26. Recall
    26
     recall1
    : method and target class recommended correctly
     recall2
    : only method recommended correctly
    

    View Slide

27. Conclusion
     Approach for move method recommendations
     Based on the similarity of dependency sets
    27
    

    View Slide

28. Conclusion
     Approach for move method recommendations
     Based on the similarity of dependency sets
     Evaluation with 475 feature-envy instances:
     precision: 60%, recall: 81%
    28
    

    View Slide

29. Conclusion
     Approach for move method recommendations
     Based on the similarity of dependency sets
     Evaluation with 475 feature-envy instances:
     precision: 60%, recall: 81%
     Dowload: http://aserg.labsoft.dcc.ufmg.br/jmove
    29
    

    View Slide

30. Conclusion
     Approach for move method recommendations
     Based on the similarity of dependency sets
     Evaluation with 475 feature-envy instances:
     precision: 60%, recall: 81%
     Dowload: http://aserg.labsoft.dcc.ufmg.br/jmove
     Ongoing work:
     Comparison with other tools (metrics-based)
     Evaluation with real systems
    30
    

    View Slide

31. Thanks!
    Recommending Move Method Refactorings
    using Dependency Sets
    Vitor Sales, Ricardo Terra, Luis Miranda, Marco Tulio Valente
    31
    WCRE, Koblenz, Germany, October 2013
    

    View Slide