Remodularization Analysis Using Semantic Clustering (CSMR-WCRE 2014)

In this paper, we report an experience on using and adapting Semantic Clustering to evaluate software remodularizations. Semantic Clustering is an approach that relies on information retrieval and clustering techniques to extract sets of similar classes in a system, according to their vocabularies. We adapted Semantic Clustering to support remodularization analysis. We evaluate our adaptation using six real-world remodularizations of four software systems. We report that Semantic Clustering and conceptual metrics can be used to express and explain the intention of the architects when performing common modularization operators, such as module decomposition.

Transcript

1. Remodularization Analysis using
   Semantic Clustering
   APPLIED SOFTWARE ENGINEERING
   RESEARCH GROUP
   /
   Gustavo Santos
   Marco Tulio Valente
   Federal University of Minas Gerais
   /
   Nicolas Anquetil
   RMoD Team – INRIA Lille
   

   View Slide

2. Semantic Clustering
   Groups source artifacts that use similar vocabulary in
   semantic clusters, which are likely to reveal the
   intention of the code
   Similarity considers text in common
   Remodularization Analysis using
   Semantic Clustering
   February 4th, 2014 2
   

   View Slide

3. Semantic Clustering
   Remodularization Analysis using
   Semantic Clustering
   3
   Extracted from Kuhn et al., 2007
   February 4th, 2014
   

   View Slide

4. Remodularization Analysis
   Remodularization Analysis using
   Semantic Clustering
   February 4th, 2014
   

   View Slide

5. Improvements to Semantic Clustering
   #1: Text filtering in comments
   Remove JavaDoc metadata, HTML code, etc.
   All public methods from Object class
   #2: Clustering Stop Criterion
   Remodularization Analysis using
   Semantic Clustering
   5
   Our Approach: Similarity Threshold
   February 4th, 2014
   

   View Slide

6. Improvement #3: Remodularization Analysis
   Challenge: Comparison between versions
   Our Approach:
   Execute Semantic Clustering on the earlier version
   Map new classes to previously calculated cluster
   Remodularization Analysis using
   Semantic Clustering
   6
   Same number of clusters
   February 4th, 2014
   

   View Slide

7. Evaluation
   Remodularization Analysis using
   Semantic Clustering
   February 4th, 2014
   

   View Slide

8. Metrics Assessment
   Goal: Do conceptual aspects express architectural
   improvements?
   Methodology
   Isolate the remodularization
   Execute Semantic Clustering
   Remodularization Analysis using
   Semantic Clustering
   8
   Execute Semantic Clustering
   Cluster comparison algorithm
   Calculate the conceptual metrics
   February 4th, 2014
   

   View Slide

9. Dataset
   Remodularization Analysis using
   Semantic Clustering
   9
   February 4th, 2014
   

   View Slide

10. Conceptual Metrics
    Conceptual Cohesion [Marcus and Poshyvanyk, 2005]
    Average cosine similarity of all pairs of classes
    Spread [Ducasse et al., 2006]
    Number of packages a cluster touches
    Remodularization Analysis using
    Semantic Clustering
    10
    Focus [Ducasse et al., 2006]
    Concentration of a cluster
    February 4th, 2014
    

    View Slide

11. RQ1 – Impact on Clusters
    What is the impact of remodularizations in the clusters
    generated by Semantic Clustering?
    Global remodularizations
    Wilcoxon test
    Remodularization Analysis using
    Semantic Clustering
    11
    Wilcoxon test
    February 4th, 2014
    

    View Slide

12. Spread and Focus Results
    Eclipse-2.0.1 – 2.1
    Remodularization Analysis using
    Semantic Clustering
    12
    JHotDraw-7.3.1 – 7.4.1
    Focus
    Spread
    February 4th, 2014
    

    View Slide

13. Wilcoxon Results
    Remodularization Analysis using
    Semantic Clustering
    13
    Summary: Remodularizations tend to consistently increase the spread of
    the existing semantic clusters among the new package structure. Focus is
    not an appropriate quality metric to support remodularization analysis.
    February 4th, 2014
    

    View Slide

14. RQ2 – Operators With Highest Impact
    What are the modularization operators that have more
    impact in the clusters generated by Semantic
    Clustering?
    Top-3 and Bottom-3 of Spread and Focus
    Remodularization Analysis using
    Semantic Clustering
    14
    Using common modularization operators proposed by
    [Rama and Patel, 2010]
    Chi-square test
    February 4th, 2014
    

    View Slide

15. RQ2 – Operators With Highest Impact
    Remodularization Analysis using
    Semantic Clustering
    15
    Summary: Module decomposition is commonly the operator behind the
    increasing in spread and focus. This fact means that the semantic clusters
    cover more packages, but they are also more concentrated inside these
    packages.
    February 4th, 2014
    

    View Slide

16. RQ3 – Impact on Conceptual Cohesion
    What is the impact of module decomposition in terms
    of conceptual cohesion?
    Analyze the impact considering:
    Restructured packages
    New packages
    Remodularization Analysis using
    Semantic Clustering
    16
    New packages
    February 4th, 2014
    

    View Slide

17. RQ3 – Impact on Conceptual Cohesion
    Original Package New Packages
    Restructured Package
    Remodularization Analysis using
    Semantic Clustering
    17
    February 4th, 2014
    Summary: After module decompositions, the new packages have better
    conceptual cohesion than the original packages. CCP is an adequate
    metric to express a quality improvement.
    

    View Slide

18. Conclusions
    Remodularization Analysis using
    Semantic Clustering
    February 4th, 2014
    

    View Slide

19. Conclusions
    Metrics assessment approach with conceptual metrics
    Improvements to Semantic Clustering
    Module Decomposition is the most common operator
    Remodularization Analysis using
    Semantic Clustering
    19
    Top 3 increase of Spread and Focus
    Increase in Conceptual Cohesion
    February 4th, 2014
    

    View Slide

20. Future Work
    More conceptual metrics
    Mapping big refactorings to modularization operators
    Recommend operators according to conceptual
    metrics
    Remodularization Analysis using
    Semantic Clustering
    20
    metrics
    February 4th, 2014
    

    View Slide

21. Thank You!
    21
    

    View Slide