Fuzzy Set and Cache-based Approach for Bug Triaging

Fuzzy Set and Cache-based Approach for Bug Triaging

Ahmed Y. Tamrawi

Electrical and Computer Engineering Department

Iowa State University

2011

2

Software Bugs1 2 3 4 5

{ Introduction }

Iowa State University Fuzzy Set and Cache-based Approach for Bug Triaging

A common term used to describe a flaw, mistake, or failure in a computer system that produces an incorrect or unexpected result, or causes it to behave in unintended ways.

Definition: (Software Bug)

• Bugs can occur in any software.• Ranging from operating systems, flight auto-

pilot software, to a simple arithmetic program!

• Software bugs are costing ~60 bln US$/Y.

The term “Bug”

(September 9, 1947)

3

More Bugs1 2 3 4 5

{ Introduction }


4

Bug Repository• Software users and developers report bugs,

to allow software developers to fix them.• Bugs are reported using bug reports which

are added to an issue tracking system or bug repository.

1 2 3 4 5

{ Introduction }


reported storedAn interface for Bugs Repository

Bugs Repository

5

• Manual bug triaging is a difficult, expensive, and lengthy process, since it needs the bug triager to manually read, analyze, and assign bug fixers for each newly reported bug.

Bug Triaging1 2 3 4 5

{ Introduction }


Assigning a bug to the most appropriate/capable developer who will fix it.

Definition: (Bug Triaging)

6

Bug Triaging1 2 3 4 5

{ Introduction }


New Bug Reports

Bugs Repository

Software Developers

Bug AssignmentBug Triager

7

Bug Triaging• Bug triager challenges:– Knowledge about the system/project;– Descriptiveness of bug report;– Rate of reporting bugs;– Many developers, different projects, and various

expertise!• Why not to automate the bug triaging

process?– Improve software quality;– Reduce cost and time.

1 2 3 4 5

{ Introduction }


Eclipse – Feb 2011

8

Example


1 2 3 4 5

{ Motivation }

Assigned to: James MoodySummary: New Repository wizard follows implementation model, not user model.Description: The new CVS Repository Connection wizard's layout is confusing. This is because it follows the implementation model of the order of elds in the full CVS location path rather than the user model...

Assigned to: James MoodySummary: Opening repository resources doesn't honor type.Description: Opening repository resource always open the default text editor and doesn'thonor any mapping between resource types and editors. As a result it is not possible to viewthe contents of an image (*.gif le) in a sensible way....

Version Control Management

(VCM)

Technical Aspect

James Moody

This aspect is concerned about various Concurrent Versions System (CVS) repository features and operations within Eclipse project.

9

Technical Aspects & Terms

• A software system has many technical aspects.

• Technical aspects are described via the technical terms extracted from software artifacts.

• A bug report describes issues related to technical aspects via its terms.


1 2 3 4 5

{ Motivation }

10

Automatic Bug Triaging


1 2 3 4 5

{ Motivation }

Who have the most bug-fixing capability/expertise with respect to the reported technical aspect(s) in a give bug report should be the fixer(s)

Key Philosophy for Automatic Bug Triaging

11

Problem Definition


1 2 3 4 5

{Bugzie Model }

In a software system, given a bug report B, and a set of developers D who have past fixing activity.Find the developers(s) with the most fixing expertise with respect to the reported technical aspect(s) in B.

Problem: (Automatic Bug Assignment)

Bugs RepositoryNew Bug

Report B

Software

Developers

12

Bugzie Overview• Bugzie considers the problem as a ranking

problem.– State-of-the-art approaches view the problem as

a classification problem.• For a bug report, Bugzie determines a

ranked list of developers most capable toward the reported issue(s).


1 2 3 4 5

{Bugzie Model }

13

Bugzie Overview• Bugzie utilizes the fuzzy set theory to rank

the fixing expertise of developers toward the technical aspects.

• Bugzie models the association of a developer and technical aspects.

• If a developer has higher fixing association with a technical aspect, he will have higher expertise and rank for that aspect.


1 2 3 4 5

{Bugzie Model }

14

Association of Fixer & Term

• is more capable than in the issues related to t.


1 2 3 4 5

{Bugzie Model }

For a technical term t, a fuzzy set Ct, with associated membership function , represents the set of developers who have the bug-fixing expertise relevant to technical aspects(s) described by t

Definition: (Capable Fixer toward a Term)

Ct

𝝁𝒕

𝟎

𝟏

15

Association of Fixer & Term• The membership score of a developer d

toward a term t is:

• Dd: Bug reports d has fixed.

• Dt: Bug reports containing t.


1 2 3 4 5

{Bugzie Model }

𝜇𝑡 (𝑑 )=|𝐷𝑑∩𝐷𝑡||𝐷𝑑∪𝐷𝑡|

𝜇𝑡 (𝑑)∈[0 ,1]

𝐷𝑡D( )

𝜇𝑡 ( )=0

D( )

𝜇𝑡 ( )=1

𝐷𝑡𝐷𝑡

D( )

D( )

16

Association of Fixer & Bug Report


1 2 3 4 5

{Bugzie Model }

𝜇𝐵 (𝑑 )=1−∏𝑡∈ 𝐵

(1−𝜇𝑡 (𝑑))

Bug Report

(B)t1 t2 tn

∪

𝝁𝑩

𝟎

𝟏CB

¿𝑡∈ 𝐵𝐶𝑡

∪

17

Association of Fixer & Bug Report

• In fuzzy set, union is a flexible combination.• The strong membership to a sub-fuzzy set(s)

implies the strong membership to the combined fuzzy set.

• After calculating for the developers, Bugzie recommends the top-scored ones as fixers for the bug report.


1 2 3 4 5

{Bugzie Model }

18

Bugzie Model


1 2 3 4 5

{Bugzie Model }

4

Bug Report

(B)Pre-processing

2t1 t2 tn

𝑡𝑖∈𝐵𝑢𝑔𝑠𝑅𝑒𝑝 . 𝑡𝑒𝑟𝑚𝑠

Des

cend

ing

on

Reco

mm

enda

tion

List

Reco

mm

enda

tion

3

∀ term𝑡

Bugs Repository

Initial Training

∀ d

evel

oper

s

1

Updating5

Bug Report

(B)

19

Bugzie Caching• Fixer candidates selection (Developers Caching).• Significant terms selection (Terms Caching).


1 2 3 4 5

{Bugzie Model }

∀ term𝑡

Bugs Repository

Initial Training

∀ d

evel

oper

s

∀ term𝑡∈𝑇 (𝑘)

Developers Cache F(x)

Terms Cache T(k)

∀∈𝐹

(𝑥)

20

Data Collection• Collected all fixed bug reports from 7 bug

repositories.• For each bug report, we extracted and

merged the summary and description.• For each system, we pre-processed these

reports: stemming, stop words removal, etc.


System History Range #Bug Reports #Fixers Eclipse 10-10-2001 to 10-28-2010 177,637 2,144 Firefox 04-07-1998 to 10-28-2010 188,139 3,014 Jazz 06-01-2005 to 06-01-2008 34,228 156 Gcc 08-03-1999 to 10-28-2010 19,430 293 Apache 05-10-2002 to 01-01-2011 43,162 1,695 FreeDesktop 01-09-2003 to 12-05-2010 17,084 374 NetBeans 01-01-2008 to 11-01-2010 23,522 380

1 2 3 4 5

{Bugzie Model }

System #Terms Eclipse 193,862 Firefox 177,028 Jazz 39,771 Gcc 63,013 Apache 110,231 FreeDesktop 61,773 NetBeans 42,797

21

Locality of Fixing Activity


Timeline

Bug Report

20102009200820072006

1 2 3 4 5

{Bugzie Model }

22


• If d belongs to the F(x), we count this as a hit.


Bug Report BFixed by d

Fixing Timeline

20102009200820072006

All Developers that have been fixing before B


Recent x%

1 2 3 4 5

{Bugzie Model }

The recent fixing developers are likely to fix bug reports in the near future.

Hypothesis: (Locality of Fixing Activity)

23



1 2 3 4 5

{Bugzie Model }

94% - 98%

96% - 99%

24

Selection of Fixer Candidates

• The locality of fixing activity suggests the actual fixer for a given bug report is likely the one having recent fixing activity.

• For each bug report, Bugzie chooses the top x% of developers sorted by their fixing time as the fixer candidates F(x).


1 2 3 4 5

{Bugzie Model }


Fixing Timeline

20102009200820072006



Recent x%

25

Bug Report

(B)

Developers Caching


1 2 3 4 5

{Bugzie Model }

5

∀ term𝑡

Initial Training


1

Bug Report

(B)Pre-processing

3t1 t2 tn


Des

cend

ing

on

Reco

mm

enda

tion

List

Reco

mm

enda

tion

Updating

4

Updating6 Bugs Repository

∀∈𝐹

(𝑥)

2

26

Selection of Descriptive Terms


1 2 3 4 5

{Bugzie Model }

System #Terms Eclipse 193,862 Firefox 177,028 Jazz 39,771 Gcc 63,013 Apache 110,231 FreeDesktop 61,773 NetBeans 42,797

RECALL :For a developer d and a term t, the higher their association score , the higher significance of t in describing the technical aspects that d has fixing expertise.

27

Selection of Descriptive Terms


1 2 3 4 5

{Bugzie Model }

Descending on

𝑘 (All Terms)

𝑇(𝐴𝑙𝑙𝑇𝑒𝑟𝑚𝑠)

𝑇 (𝑘)

28

Terms Caching


1 2 3 4 5

{Bugzie Model }

Bugs Repository

Initial Training

Terms Cache T(k)

Bug Report

(B)Pre-processing t1 t2 tn

𝑡𝑖∈𝑇 (𝑘)

Des

cend

ing

on

Reco

mm

enda

tion

List

Reco

mm

enda

tion

Updating


∀ d

evel

oper

s

Bug Report

(B)Updating

29

Empirical Evaluation• We evaluated Bugzie on our collected

datasets.• Experiments:– Selection of fixer candidates;– Selection of terms;– Selection of developers and terms;– Comparison with state-of-the-art approaches.


1 2 3 4 5

{ Empirical Evaluation }

System History Range #Bug Reports #Fixers Eclipse 10-10-2001 to 10-28-2010 177,637 2,144 Firefox 04-07-1998 to 10-28-2010 188,139 3,014 Jazz 06-01-2005 to 06-01-2008 34,228 156 Gcc 08-03-1999 to 10-28-2010 19,430 293 Apache 05-10-2002 to 01-01-2011 43,162 1,695 FreeDesktop 01-09-2003 to 12-05-2010 17,084 374 NetBeans 01-01-2008 to 11-01-2010 23,522 380

30

Experiment Setup


1 2 3 4 5


Bug Report B

Creation Timeline

0 1 2 3 4 5 6 7 8 9 10

Bugzie uses frame 0 for initial training1

Using training data, Bugzie recommends a top-n developers to fix bug report B

2

Bugzie updates the training data with the tested bug report B3

Move to next Bug Report

Bug Report B

Des

cend

ing

on

Reco

mm

enda

tion

List

for B

Bugzie repeats steps 2 and 3 till it consumes all bug reports

31

Prediction Accuracy• If the recommendation list for a bug report

contains its actual fixer, we count this as a hit (i.e. a correct recommendation).

• For each frame under test, we calculated Prediction Accuracy (PA).

• If we have 100 bugs and for 60 of those bugs, we could recommend the actual fixing developer is in our Top-2 list, then Top-2 prediction accuracy is 60%.


1 2 3 4 5


𝑃𝐴 (% )= ¿𝐻𝑖𝑡𝑠¿𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝐶𝑎𝑠𝑒𝑠

×100 %

32



1 2 3 4 5


Bug Report

(B)

5

∀ term𝑡

Initial Training


1

Bug Report

(B)Pre-processing

3t1 t2 tn


Des

cend

ing

on

Reco

mm

enda

tion

List

Reco

mm

enda

tion

Updating

4

Updating6 Bugs Repository

∀∈𝐹

(𝑥)

2


Fixing Timeline

20102009200820072006



Recent x%

33



1 2 3 4 5


Top-1 Prediction Accuracy


Firefox ( ):At x = 10%, PA = 72.4%At x = 100%, PA = 70.7%

34

• Selecting a suitable portion of recent fixers does not lessen much the accuracy, and sometimes improves it as in the cases of Firefox, Eclipse, etc.

• Selecting only a portion of available developers as candidates also improves time efficiency.



1 2 3 4 5


35

Selection of Terms


1 2 3 4 5


5

Bugs Repository

Initial Training

Terms Cache T(k)

1

2

Bug Report

(B)Pre-processing

3t1 t2 tn

𝑡𝑖∈𝑇 (𝑘)

Des

cend

ing

on

Reco

mm

enda

tion

List

Reco

mm

enda

tion

Updating

4


∀ d

evel

oper

s

Bug Report

(B)Updating 6

36

Selection of Terms


1 2 3 4 5




Peak Range Peak Range

Eclipse( ):At k = 16, PA = 80%At k = All Terms, PA = 72%

37

• Selection of terms could improve much the prediction accuracy.

• The results suggest that one just needs a small yet significant set of terms for each developer to describe his bug-fixing expertise.

Selection of Terms


1 2 3 4 5


38

Selection of Developers & Terms

• To study the impact of both developers selection (x) and terms selection (k).


1 2 3 4 5


Eclipse

Firefox

39

Selection of Developers & Terms


1 2 3 4 5


Base: Base model with all developers and all termsC.S.: Candidate SelectionT.S.: Terms SelectionBoth: The best PA when applying both C.S. and T.S.

40

Comparison• We compared Bugzie Results with state-of-

the-art approaches.

• Used Weka to re-implement those approaches


1 2 3 4 5


Approach Papers

Naïve Bayes (NB) Cubranic & Murphy[1] Anvik et al.[2] Bhattacharya & Neamtiu[3]

Bayesian Networks (BN) Bhattacharya & Neamtiu Inc. Naïve Bayes (InB) Bhattacharya & Neamtiu Inc. Bayesian Networks (InBN) Bhattacharya & Neamtiu Support Vector Machine (SVM) Anvik et al. Vector Space Model (VSM) Matter et al.[4] C4.5 (Decision Trees) Anvik et al.

41

Comparison• Some of the approaches (C4.5 - Decision

Trees) can not scale up well to our dataset.• We prepared smaller dataset:


1 2 3 4 5


System History Range #Bug Reports #Fixers #Terms Eclipse 01-01-2008 to 10-28-2010 69,829 1,510 103,690 Firefox 01-01-2008 to 10-28-2010 77,236 1,682 85,951 Jazz 06-01-2005 to 06-01-2008 34,228 156 39,771 Gcc 01-01-2008 to 10-28-2010 6,865 161 20,279 Apache 01-01-2008 to 01-01-2011 28,682 1,354 80,757 FreeDesktop 01-01-2008 to 12-05-2010 10,624 161 37,596 NetBeans 01-01-2008 to 11-01-2010 23,522 380 42,797

3-Year Histories of the full dataset

42

Comparison Results


1 2 3 4 5


(d) days, (h) hours, (m) minutes, (s) seconds

43

Conclusions• Bugzie achieves higher accuracy and

efficiency than state-of-the-art approaches.• Bugzie can accommodate the locality of

fixing activity and software evolution with flexible caching of developers and terms.


1 2 3 4 5

{ Conclusions}

44

Thesis Contributions • Bugzie, a scalable, fuzzy set and cache-based

automatic bug triaging approach, which is significantly more efficient and accurate than existing state-of-the-art approaches.

• The finding of the locality of fixing activity.• A comprehensive evaluation on the efficiency and

correctness of Bugzie in comparison with state-of-the-art approaches.

• An observation/method to capture a small and significant set of terms describing developers’ bug-fixing expertise.


1 2 3 4 5

{ Conclusions}

45

Future Work• Use different caching mechanisms for

developers and terms.• Explore the usage of other textual and non-

textual contents of bug reports for bug triaging.

• Use other software artifacts to accurately measure the developer’s expertise.


1 2 3 4 5

{ Conclusions}

47

Thank You!


Fuzzy Set and Cache-based Approach for Bug Triaging

Documents

Transcript of Fuzzy Set and Cache-based Approach for Bug Triaging