The document discusses the impact of noise on the performance of bug prediction models. It introduces three research questions about the resistance of prediction models to noise, the ability to detect and remove noise, and whether removing noise can improve performance. The study approach involves training prediction models on software data with artificially inserted noise, and evaluating the models' performance on clean test data. The results show that the models are reasonably resistant to noise for some projects but performance decreases for others. This indicates noise could negatively impact models if not addressed.

1. Dealing with
Noise
in bug prediction
Sunghun Kim, Hongyu Zhang,
Rongxin Wu and Liang Gong
The Hong Kong University of Science & Technology
Tsinghua University

2. Where are the bugs?
2

3. Where are the bugs?
Complex files!
[Menzies et al.]
2

4. Where are the bugs?
Modified files!
Complex files! [Nagappan et al.]
[Menzies et al.]
2

5. Where are the bugs?
Modified files!
Complex files! [Nagappan et al.]
[Menzies et al.]
Nearby other bugs!
[Zimmermann et al.]
2

6. Where are the bugs?
Modified files!
Complex files! [Nagappan et al.]
[Menzies et al.]
Nearby other bugs! Previously fixed files
[Zimmermann et al.] [Hassan et al.]
2

7. Prediction model
training instances
(features+ labels)
3

8. Prediction model
training instances
(features+ labels)
Learner
3

9. Prediction model
training instances
(features+ labels)
?
Learner
3

10. Prediction model
training instances
(features+ labels)
?
Learner
3

11. Prediction model
training instances
(features+ labels)
?
Learner Prediction
3

12. Prediction model
training instances
(features+ labels)
?
Learner Prediction
3

13. Training on software evolution is key
• Software features can be used to predict bugs
• Defect labels obtained from software evolution
• Supervised learning algorithms
Version Bug
Archive Database
4

14. Change classification
5
Kim, Whitehead Jr., Zhang: Classifying Software Changes: Clean or Buggy? (TSE 2008)

15. Change classification
bug-introducing (“bad”)
X X X X
5
Kim, Whitehead Jr., Zhang: Classifying Software Changes: Clean or Buggy? (TSE 2008)

16. Change classification
BUILD A LEARNER
bug-introducing (“bad”)
X X X X
5
Kim, Whitehead Jr., Zhang: Classifying Software Changes: Clean or Buggy? (TSE 2008)

17. Change classification
BUILD A LEARNER
bug-introducing (“bad”)
X X X X
new change
5
Kim, Whitehead Jr., Zhang: Classifying Software Changes: Clean or Buggy? (TSE 2008)

18. Change classification
BUILD A LEARNER
bug-introducing (“bad”)
X X X X
new change
PREDICT QUALITY
5
Kim, Whitehead Jr., Zhang: Classifying Software Changes: Clean or Buggy? (TSE 2008)

19. Training Classifiers
0 1 0 1 0 1 0 1 … 0 1
Historical
changes
0 0 0 1 0 1 0 1 … 0 0
0 1 1 1 0 1 1 1 … 0 0
0 1 0 3 0 0 0 1 … 0 1
0 1 0 1 0 1 0 1 … 0 0
§ Machine learning techniques
• Bayesian Network, SVM

20. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
commit
fixed bugs Bf
commit
commit
commit
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

21. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
commit
fixed bugs Bf
commit
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

22. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
commit
fixed bugs Bf
commit
linked fixed bugs Bfl
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

23. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
commit
fixed bugs Bf
commit
linked fixes Cfl linked fixed bugs Bfl
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

24. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
related,
commit
but not linked fixed bugs Bf
commit
linked fixes Cfl linked fixed bugs Bfl
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

25. Source Repository Bug Database
all commits C
commit all bugs B
commit commit
commit commit
bug fixes Cf related,
commit
but not linked fixed bugs Bf
commit
linked fixes Cfl linked fixed bugs Bfl
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

26. Source Repository Bug Database
all commits C
oise!
N all bugs B
commit
commit commit
commit commit
bug fixes Cf related,
commit
but not linked fixed bugs Bf
commit
linked fixes Cfl linked fixed bugs Bfl
commit
commit
linked via log messages
7 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

27. Effect of training on
superbiased data (Severity)
Trained on all bugs
Trained on biased data1
Trained on biased data2
0% 20% 40% 60% 80% 100%
Bug Recall
8 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

28. Effect of training on
superbiased data (Severity)
Trained on all bugs
Trained on biased data1
Trained on biased data2
0% 20% 40% 60% 80% 100%
Bug Recall
9 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

29. Effect of training on
superbiased data (Severity)
Trained on all bugs
Trained on biased data1
Trained on biased data2
0% 20% 40% 60% 80% 100%
Bug Recall
10 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

30. Effect of training on
superbiased data (Severity)
Bias in bug severityon all bugs
Trained
affects BugCache on biased data1
Trained
Trained on biased data2
0% 20% 40% 60% 80% 100%
Bug Recall
10 Bird et al. “Fair and Balanced? Bias in Bug-Fix Datasets,” FSE2009

31. Are defect prediction
models learned from
noisy data reliable?
11

32. Study questions
• Q1: How resistant a defect prediction
model is to noise?
• Q2: How much noise could be detected/
removed?
• Q3: Could we remove noise to improve
defect prediction performance?
12

33. Study approach
13

34. Study approach
13

35. Study approach
13

36. Study approach
Training
Testing
13

37. Study approach
Training
Bayes Net Testing
13

38. Making noisy training instances
Training Testing
14

39. Making noisy training instances
1 Removing
buggy labels False negative noise
Training Testing
14

40. Making noisy training instances
1 Removing
buggy labels False negative noise
Training Testing
2 Adding
buggy labels
False positive noise
14

41. Prediction models
buggy
Rev n Rev n+1
... ...
... ...
change
clean
change classification
15

42. Prediction models
buggy
Rev n Rev n+1
... ...
... ...
change
clean
change classification
buggy
File
...
...
...
File
clean
file-level defect prediction
15

43. Performance evaluation
§ 4 possible outcomes from prediction models
§ Classifying a buggy change as buggy (nb->b)
§ Classifying a buggy change as clean (nb->c)
§ Classifying a clean change as clean (nc->c)
§ Classifying a clean change as buggy (nc->b)
nb->b nb->b
§ Precision = , Recall=
nb->b + nc->b nb->b + nb->c
precision ! recall
§ F-measure = 2 !
precision + recall
16

44. Subjects
change classification
subject # instances % buggy # features
Columba 1,800 29.4% 17,411
Eclipse (JDT) 659 10.1% 16,192
Scarab 1,090 50.6% 5,710
file-level defect prediction
subject # instances % buggy # features
SWT 1,485 44% 18
Debug 1,065 24.7% 18
17

45. Experimental results
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
!#)"
!#("
-./0123"
!#'"
!#&" 40115"6.7"-./0123"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
18

46. Columba
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
!#)"
!#("
-./0123"
!#'"
!#&" 40115"6.7"-./0123"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
19

47. Columba
$"
!#,"
!#+" 1. Random guess (50% buggy, 50% clean)
2. Columba’s defect rate is about 30%
!#*"
!"##$%&'()*+",)
3. Precision = 0.3 and Recall =0.5
!#)" 4. F-measure = 0.375 (2*0.5*0.3)/(0.3+0.5)
!#("
-./0123"
!#'"
!#&" 40115"6.7"-./0123"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
19

48. Columba
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
!#)"
!#("
-./0123"
!#'"
!#&" 40115"6.7"-./0123"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
20

49. Eclipse (JDT)
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
!#)"
!#("
-./0123"
!#'"
!#&" 45667"89:"-./0123"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
21

50. Scarab
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
!#)"
!#("
-./0/1"
!#'"
!#&" 23445"670"-./0/1"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
22

51. Eclipse (Debug)
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)%
!#)"
!#("
-./01"
!#'"
!#&" -0223"456"-./01"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)%
23

52. Eclipse (SWT)
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)%
!#)"
!#("
-./"
!#'"
!#&" 01223"456"-./"
!#%"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)%
24

53. Q1: How resistant a defect
prediction model is to noise?
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
-./"
!#)"
!#(" 01234"
!#'" 5673829"
!#&"
:;7<=>1"
!#%"
-;9?92"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
25

54. Q1: How resistant a defect
prediction model is to noise?
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
-./"
!#)"
!#(" 01234"
!#'" 5673829"
!#&"
:;7<=>1"
!#%"
-;9?92"
!#$"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
26

55. Q1: How resistant a defect
prediction model is to noise?
$"
!#,"
!#+"
!#*"
!"##$%&'()*+",)
-./"
!#)"
!#(" 01234"
!#'" 5673829"
!#&"
!#%"
!#$"
20~30% :;7<=>1"
-;9?92"
!"
!" !#$" !#%" !#&" !#'" !#(" !#)"
-./%0,*1212#%+)3%4*5+)%2)#*67)%-&8/%9%4*5+)%:;+167)%-&</%,*3)
26

56. Study questions
• Q1: How resistant a defect prediction
model is to noise?
• Q2: How much noise could be detected/
removed?
• Q3: Could we remove noise to improve
defect prediction performance?
27

57. Detecting noise
1 Removing
buggy labels False negative noise
Original training
2 Adding
buggy labels
False positive noise
28

58. Detecting noise
False negative noise
Original training
False positive noise
29

59. 30
ts. However, it is very hard to get a golden set. In our approach,
e carefully select high quality datasets and assume them the
lden sets. We then add FPs and FNs intentionally to create a
False positive noise
ise set. To add FPs and FNs, we randomly selects instances in a
lden set and artificially change their labels from buggy to clean
from clean to buggy, inspired by experiments in [4].
Original training
?
noise
Clean False negative noise
Detecting noise
F igure 4. C reating biased training set
make FN data sets (for RQ1), we randomly select n% buggy

60. return Aj
Closest 9. Thenoise identification algorit
F igure
list pseudo-code of the C LN I
A
31

61. Noise detection performance
Precision Recall F-measure
Debug 0.681 0.871 0.764
SWT 0.624 0.830 0.712
(noise level =20%)
32

62. Noise detection performance
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3 Precision
0.2 Recall
0.1
F-measure
0
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
FP & FN noise level
Noise Rate
33

63. Study questions
• Q1: How resistant a defect prediction
model is to noise?
• Q2: How much noise could be detected/
removed?
Q3: Could we remove noise to improve
defect prediction performance?
34

64. Bug prediction using cleaned data
Noisey Cleaned
100
75
SWT F-measure
50
25
0
0% 15% 30% 45%
35 Noise level

65. Bug prediction using cleaned data
Noisey Cleaned
100
75
SWT F-measure
50
25
0
0% 15% 30% 45%
36 Noise level

66. Bug prediction using cleaned data
Noisey Cleaned
100
75
SWT F-measure
50
25
76%
F-measure
with 45% noise
0
0% 15% 30% 45%
36 Noise level

67. Study limitations
• All datasets are collected from open source
projects
• The golden set used in this paper may not be
perfect
• The noisy data simulations may not reflect
the actual noise patterns in practice
37

68. Summary
• Prediction models (used in our experiments)
are resistant (up to 20~30%) of noise
• Noise detection is promising
• Future work
- Building oracle defect sets
- Improving noise detection algorithms
- Applying to more defect prediction models
(regression, bugcache)
38