Your SlideShare is downloading. ×
0
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
QSIC09.ppt
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×
Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

QSIC09.ppt

91

Published on

Published in: Technology, Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
91
On Slideshare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
1
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  1. A Bayesian Approach for the Detection of Code and Design Smells Foutse Khomh, Stéphane Vaucher,Yann-Gaël Guéhéneuc, and Houari Sahraoui QSIC’09 Jeju-do 2009/08/25 Ptidej Team – Pattern-based Quality Evaluation and Enhancement Department of Computer Engineering and Software Engineering École Polytechnique de Montréal, Québec, Canada © Khomh, 2009
  2. Outline Motivation Context Problems and Objectives Previous Work Our Solution Bayesian Belief Networks Experiments Discussions2/27 Conclusion
  3. Motivation Independent IV&V team reviews of object-oriented software programs – Assess the quality of programs – Identify potential problems • Defects • Anomalies • “Bad Smells”3/27
  4. Context Code smells and antipatterns Blob: also called God class [23], is a class that centralises functionality and has too many responsibilities. Brown et al. [4] characterise its structure as a large controller class that depends on data stored in several surrounding data4/27 classes
  5. Problems and Objectives Problem 1: Sizes of programs can be large Automate the evaluation process Problem 2: Resources are limited Prioritise manual reviews Problem 3: Quality assessment is subjective Consider uncertainty in the process5/27
  6. Previous Work (1/2) Webster’s book on “antipatterns” [31] Riel’s 61 heuristics [23] Beck’s 22 code smells [11] Brown et al. 40 antipatterns [4] Travassos et al. manual approach [28] Marinescu’s detection strategies [17] (also Munro [19])6/27 Moha et al. rules cards in DECOR [18]
  7. Previous Work (2/2) Either no full automation – Problem 1 is not solved Or strict rules with classes participating or not (binary value) to an antipattern – Problem 2 and 3 are not solved7/27
  8. Our Solution (1/2) Assign a probability that a class is considered an antipattern by a stakeholder Guide a manual inspection according to this probability Use Bayesian belief network to build a model of an antipattern and assign a probability to all classes8/27
  9. Our Solution (2/2) Problem 1: Sizes of programs can be large Assign a probability to all classes of interest automatically Problem 2: Resources are limited Prioritise manual reviews using probabilities Problem 3: Quality assessment is subjective Uncertainty is naturally considered using9/27 Bayesian beliefs networks
  10. BBN (1/3) A Bayesian Belief Network is a directed acyclic graph with probability distribution Graph structure – Nodes = random variables – Edges = probabilities dependencies Each node depends only on its parents Embody the experts’ knowledge10/27
  11. BBN (2/3) Classifier – C = {smell, not smell} Input vector describing a class – <a1, …, an> – P(A|B) = P(B|A) P(A) / P(B)11/27
  12. BBN (2/3) Building a BBN 1. Define its structure 2. Assign/learn its probability tables12/27
  13. 1. Defining the BBN Structure Based on Moha et al.’s rule cards – Detection relies on metrics, binary class relations, and linguistic analysis – Values are combined using set operators – Thresholds are statistically defined (using a box plot)13/27
  14. Rule Card of BlobRULE_CARD : Blob { RULE : Blob { ASSOC : a s s o c i a t e d FROM : m ai n Cl a ss ONE TO : D a t a C l a s s MANY } ; RULE : MainClass { UNION LargeCla ssLowCohesion C o n t r o l l e r C l a s s } ; RULE : LargeCla ssLo wCohesion { UNION L a r g e C l a s s LowCohesion } ; RULE : L a r g e C l a s s { ( METRIC : NMD + NAD, VERY_HIGH , 0 ) } ; RULE : LowCohesion { ( METRIC : LCOM5, VERY_HIGH , 2 0 ) } ; RULE : C o n t r o l l e r C l a s s { UNION ( SEMANTIC : METHODNAME , { P r o c e s s , C o n t r o l , C t r l , Command , Cmd, Proc, UI, Manage, D r i v e } ) ( SEMANTIC : CLASSNAME , { P r o c e s s , C o n t r o l , C t r l , Command , Cmd, Proc , UI, Manage, Drive , System, Subsystem } ) } ; RULE : D a t a C l a s s { ( STRUCT : METHOD_ACCESSOR , 9 0 ) } ;}; 14/27
  15. Conversion into BBN Set operators (OR, AND) are learned using conditional probabilities Hard thresholds and rules are smoothed by interpolating values Binary class relations are counted and15/27 treated as metrics
  16. BBN Structure for the Blob16/27
  17. 2. Resulting BBN for the Blob17/27
  18. Experiments RQ1: To what extent a model built with a BBN based on an existing rule-based model is able detect smells in a program? RQ2: Is a model built with a BBN better than a state-of-the-art approach, DECOR?18/27
  19. Settings Programs – Xerces v2.7.0 – Gantt Project v1.10.2 Oracle: Vote of three groups (of students) on both programs19/27
  20. Settings Calibration (probability tables) – Local (same-system calibration) – External (calibrated on one system, applied to the other)20/27
  21. Local calibration On Xerces only Three-fold cross validation – 3 groups with 5 Blobs in each – Combination of 2 groups for calibration, tested on the other Inspected Classes Group1 6 Waste of 8% to 44% Group2 7 of effort (average 32%)21/27 Group3 9
  22. External calibration Probabilities learned on one, applied on other Precision vs. Recall (per investigation sizes) Xerces (left), Gantt (right)22/27
  23. Experiments RQ1: To what extent a model built with a BBN based on an existing rule-based model is able detect smells in a program? RQ2: Is a model built with a BBN better than a state-of-the-art approach, DECOR? Better than DECOR23/27
  24. Discussions (1/2) Results of BBN are superior to those of DECOR External calibration yields good results even though tweaking is recommended A well calibrated model should be able to estimate the number of smells in a program, but our models overestimate by a factor of 224/27
  25. Discussions (2/2) Used simple techniques to convert rules into probability distributions Worked with discrete probabilities, might need to be adapted to continuous cases25/27
  26. Conclusion Bayesian models support uncertainty in detection process Their performance is better than state- of-the-art rules External calibration seems to indicate that this could be used in an industrial context with minimal costs Support for automated conversion of26/27 rules to support more design smells
  27. Acknowledgements Naouel Moha for providing her data and code to compare with DECOR Natural Sciences and Engineering Research Council of Canada Canada Research Chair in Software Patterns and Patterns of Software27/27
  28. Advertisement IEEE Software Special issue on Software Evolution Deadline: 1st of November, 2009 Publication: July/August 2010 Please do not hesitate to contact me! ☺28/27

×