1. The Conceptual Coupling Metrics for Object-Oriented Systems
Denys Poshyvanyk and Andrian Marcus
SEVERE
22nd IEEE International Conference on Software Maintenance
Philadelphia, Pennsylvania
September 27, 2006

2. Motivation Concepts and classes
Implementation and representation of concepts
Semantic information

3. Example
Methods from MySecMan class in Mozilla

4. Approach Latent Semantic Indexing Advantages:
captures essential semantic info via dimensionality reduction
overcomes problems with polysemy and synonymy
easy to apply on the source code

5. Related Work Coupling measures Previously solved problems:
Traceability link recovery
Managing software artifacts
Conceptual cohesion
Software clustering
Concept/feature location
Requirements traceability
Isolating concerns in requirements

6. Extracting Semantic Info
Source code -> Corpus (doc = method)
Preprocessing: split_identifiers & SplitIdentifiers
Vector space = term-by-document matrix
Singular Value Decomposition -> LSI subspace

7. Computing Conceptual Similarity
Cosine between vectors

8. Conceptual Coupling between Classes
Method - Class conceptual similarity
Class - Class conceptual similarity
Conceptual coupling between A and B = 0.4
Class A
Class B
0.5
method1
0.6
method1
0.5
0.2
0.7
method2
0.3
0.4
method2
0.4
0.3
0.2
0.4
method3
method3
0.3

9. Maximal Conceptual Coupling
Conceptual coupling based on the strongest conceptual coupling link
Conceptual coupling between A and B = 0.56
Class A
Class B
0.5
method1
0.6
method1
0.7
0.2
0.7
method2
0.3
0.4
method2
0.6
0.3
0.2
0.4
method3
method3
0.4

10. Are We Measuring Anything New?
Compare with other coupling measures:
Coupling between classes (CBO) [Chidamber’04]
Response for class (RFC) [Chidamber’04]
Message passing coupling (MPC) [Li’93]
Data abstraction coupling (DAC) [Li’93]
Information-flow based coupling (IPC) [Lee’95]
A suite of coupling measures by Briand et al: ACAIC, OCAIC, ACMIC and OCMIC
Tools:
Columbus [Ferenc’04]
IRC2M

11. Software Systems
Ten open-source systems from different domains

12. Principal Component Analysis
Identifying groups of metrics (variables) which measure the same underlying mechanism that defines coupling (dimension)
PCA procedure:
collect data
identify outliers
perform PCA

13. PCA Results: Rotated Components
CoCC and CoCCm define new dimensions (PC2 and PC6)

14. Discussion of the Results
Conceptual similarities between all pairs of classes
Selected classes with highest values of conceptual coupling
No direct structural dependencies

15. Discussion of the Results Cont.
Concepts
TortoiseCVS: merge and update CVS operations
WinMerge: checking out a revision of the file
Related concepts and history of common changes

16. Current & Future Work Connection to change/fault proneness
Impact analysis
Hidden dependencies/indirect coupling
Aspect mining
Refining canonical feature sets
Concept location and clustering