The document proposes novel "Micro Interaction Metrics" (MIMs) that capture developers' interactions by analyzing Mylyn task log data to improve defect prediction. MIMs are extracted from Mylyn data at the file and task levels. An experiment applies MIMs and existing metrics to defect prediction using machine learning on Eclipse project data, finding MIMs improve prediction accuracy over baselines for different projects, algorithms and training periods, as shown by statistical tests.

1. Micro Interaction Metrics
for Defect Prediction
Taek Lee, Jaechang Nam, Dongyun Han, Sunghun Kim, Hoh Peter In
FSE 2011, Hungary, Sep. 5-9

2. Outline
• Research motivation
• The existing metrics
• The proposed metrics
• Experiment results
• Threats to validity
• Conclusion

3. Defect Prediction?
why is it necessary?

4. Software quality assurance
is inherently a resource
constrained activity!

5. Predicting defect-prone
software entities* is
to put the best labor effort
on the entities
* functions or code files

6. Indicators of defects
• Complexity of source codes (Chidamber and Kemerer 1994)
• Frequent code changes (Moser et al. 2008)
• Previous defect information (Kim et al. 2007)
• Code dependencies (Zimmermann 2007)

7. Indeed,
where do defects
come from?

8. Humans Error!
Programmers make mistakes,
consequently defects are
injected, and software fails
Human Bugs Software
Errors Injected fails

9. Programmer Interaction
and Software Quality

10. Programmer Interaction
and Software Quality
“Errors are from cognitive breakdown
while understanding and implementing
requirements”
- Ko et al. 2005

11. Programmer Interaction
and Software Quality
“Errors are from cognitive breakdown
while understanding and implementing
requirements”
- Ko et al. 2005
“Work interruptions or task switching
may affect programmer productivity”
- DeLine et al. 2006

12. Don’t we need to also consider
developers’ interactions
as defect indicators?

13. …, but the existing indicators
can NOT directly capture
developers’ interactions

14. Using Mylyn data, we propose novel
“Micro Interaction Metrics (MIMs)”
capturing developers’ interactions

15. The Mylyn* data is stored
as an attachment to the
corresponding bug reports in
the XML format
* Eclipse plug-in storing and recovering task contexts

17. <InteractionEvent … Kind=“ ” … StartDate=“ ” EndDate=“ ”
… StructureHandle=“ ” … Interest=“ ” … >

18. <InteractionEvent … Kind=“ ” … StartDate=“ ” EndDate=“ ”
… StructureHandle=“ ” … Interest=“ ” … >

19. <InteractionEvent … Kind=“ ” … StartDate=“ ” EndDate=“ ”
… StructureHandle=“ ” … Interest=“ ” … >

20. Two levels of MIMs Design

21. Two levels of MIMs Design
File-level MIMs
specific interactions for a
file in a task
(e.g., AvgTimeIntervalEditEdit)

22. Two levels of MIMs Design
File-level MIMs
specific interactions for a
file in a task
(e.g., AvgTimeIntervalEditEdit)
Task-level MIMs
property values shared
over the whole task
(e.g., TimeSpent)

23. Two levels of MIMs Design
File-level MIMs Mylyn Task Logs
specific interactions for a
file in a task 10:30 Selection file A
(e.g., AvgTimeIntervalEditEdit)
11:00 Edit file B
12:30 Edit file B

24. Two levels of MIMs Design
Mylyn Task Logs
10:30 Selection file A
Task-level MIMs 11:00 Edit file B
property values shared 12:30 Edit file B
over the whole task
(e.g., TimeSpent)

25. The Proposed Micro Interaction Metrics

26. The Proposed Micro Interaction Metrics

27. The Proposed Micro Interaction Metrics

28. For example,
NumPatternSXEY is to capture
this interaction:

29. For example,
NumPatternSXEY is to capture
this interaction:
“How much times did a programmer
Select a file of group X
and then Edit a file of group Y
in a task activity?”

30. X if a file shows defect
locality* properties
group X or Y
Y otherwise
H if a file has
group H or L high** DOI value
L otherwise
* hinted by the paper [Kim et al. 2007]
** threshold: median of degree of interest (DOI) values in a task

31. Bug Prediction Process

32. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010

33. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
All the Mylyn task data collectable from
Eclipse subprojects (Dec 05 ~Sep 10)

34. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010

35. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
Post-defect counting period

36. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
Post-defect counting period

37. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
Post-defect counting period

38. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
Post-defect counting period

39. STEP1: Counting & Labeling
Instances
Task 1 Task 2 Task 3 Task i Task i+1 Task i+2 Task i+3
f1.java
f3.java f2.java
… f2.java …
f1.java f3.java
f3.java
Dec 2005 Time P Sep 2010
Post-defect counting period

40. STEP1: Counting & Labeling
Instances
The number of counted post defects
(edited files only within bug fixing tasks)
Task 1
f1.java Task 3
Task 2
=1 Task i Task i+1 Task i+2 Task i+3
f2.java = 1
f3.java = 2 f1.java
f3.java f2.java
… … f1.java f3.java
f2.java …
f3.java
Labeling rule for the file instance
“buggy” (if # of post-defects > 0)
Dec 2005 “clean” (if # of post-defects = 0) Time P Sep 2010
Post-defect counting period

41. STEP2: Extraction of MIMs
Dec 2005 Time P Sep 2010

42. STEP2: Extraction of MIMs
Task 1 Task 2 Task 3 Task 4
…
Dec 2005 Time P Sep 2010
Metrics extraction period

43. STEP2: Extraction of MIMs
Task 1
f3.java
...
edit
…
edit
…
Dec 2005 Time P Sep 2010
Metrics extraction period

44. STEP2: Extraction of MIMs
Metrics Computation
Task 1
f3.java
...
edit
…
edit
…
Dec 2005 Time P Sep 2010
Metrics extraction period

45. STEP2: Extraction of MIMs
Metrics Computation
Task 1 MIMf3.java  valueTask1
f3.java
...
edit
…
edit
…
Dec 2005 Time P Sep 2010
Metrics extraction period

46. STEP2: Extraction of MIMs
Metrics Computation
Task 1 Task 2 MIMf3.java  valueTask1
f3.java
...
f1.java
...
MIMf1.java  valueTask2
edit edit
… …
edit edit
… …
Dec 2005 Time P Sep 2010
Metrics extraction period

47. STEP2: Extraction of MIMs
Metrics Computation
Task 1 Task 2 Task 3 MIMf3.java  valueTask1
f3.java
...
f1.java
...
f2.java
...
MIMf1.java  valueTask2
edit edit edit
…
edit
…
edit
…
edit
MIMf2.java  valueTask3
… … …
Dec 2005 Time P Sep 2010
Metrics extraction period

48. STEP2: Extraction of MIMs
Metrics Computation
Task 1 Task 2 Task 3 Task 4 MIMf3.java  valueTask1
f3.java
...
f1.java
...
f2.java
...
f1.java MIMf1.java  valueTask2
…edit
edit edit edit
MIMf2.java  valueTask3
…edit..
… … … …
edit edit edit f2.java
… … … …edit…
Dec 2005 Time P Sep 2010
Metrics extraction period

49. STEP2: Extraction of MIMs
Metrics Computation
Task 1 Task 2 Task 3 Task 4 MIMf3.java  valueTask1
f3.java f1.java f2.java f1.java MIMf1.java  (valueTask2+valueTask4)
... ... ... …edit
edit
…
edit
…
edit
…
…edit..
f2.java
… MIMf2.java  (valueTask3+valueTask4)
edit edit edit
… … … …edit…
Dec 2005 Time P Sep 2010
Metrics extraction period

50. STEP2: Extraction of MIMs
Metrics Computation
Task 1 Task 2 Task 3 Task 4 MIMf3.java  valueTask1
f3.java f1.java f2.java f1.java MIMf1.java  (valueTask2+valueTask4)/2
)
... ... ... …edit
edit
…
edit
…
edit
…
…edit..
f2.java
… MIMf2.java  (valueTask3+valueTask4)/2
)
edit edit edit
… … … …edit…
Dec 2005 Time P Sep 2010
Metrics extraction period

51. Understand JAVA tool was used
for extracting 32 source Code
Metrics (CMs)*
* Chidamber and Kemerer, and OO metrics

52. Understand JAVA tool was used
for extracting 32 source Code
Metrics (CMs)*
List of selected source code metrics
CVS
last revision
…
Dec 2005 Time P Sep 2010
* Chidamber and Kemerer, and OO metrics

53. Fifteen History Metrics (HMs)* were
collected from the corresponding
CVS repository
* Moser et al.

54. Fifteen History Metrics (HMs)* were
collected from the corresponding
CVS repository
List of history metrics (HMs)
CVS revisions
…
Dec 2005 Time P Sep 2010
* Moser et al.

55. STEP3: Creating a training corpus
Instance
Extracted MIMs Label
Name
Training
… Classifier
Instance # of post
Extracted MIMs
Name defects
Training
… Regression

56. STEP4: Building prediction models
Classification and Regression
modeling with different machine
learning algorithms using the
WEKA* tool
* an open source data mining tool

57. STEP5: Prediction Evaluation
Classification
Measures

58. STEP5: Prediction Evaluation
How many instances
are really buggy among
the buggy-predicted
outcomes?
Classification
Measures

59. STEP5: Prediction Evaluation
How many instances
are really buggy among
the buggy-predicted
outcomes?
How many instances
are correctly predicted as
‘buggy’ among the real
buggy ones
Classification
Measures

60. STEP5: Prediction Evaluation
Regression
Measures
correlation coefficient (-1~1)
mean absolute error (0~1)
root square error (0~1)

61. STEP5: Prediction Evaluation
Regression
between # of real buggy
instances and # of instances Measures
predicted as buggy
correlation coefficient (-1~1)
mean absolute error (0~1)
root square error (0~1)

62. T-test with 100 times of
10-fold cross validation
Reject H0* and accept H1*
if p-value < 0.05
(at the 95% confidence level)
* H0: no difference in average performance, H1: different (better!)

63. Result Summary
MIMs improve prediction accuracy for
1 different Eclipse project subjects
2 different machine learning algorithms
3 different model training periods

64. Prediction for different project subjects
File instances and % of defects

65. Prediction for different project subjects
MIM: the proposed metrics CM: source code metrics HM: history metrics

66. Prediction for different project subjects
BASELINE: Dummy Classifier
predicts in a purely random manner
e.g., for 12.5% of buggy instances
Precision(B)=12.5%, Recall(B)=50%
F-measure(B)=20%
MIM: the proposed metrics CM: source code metrics HM: history metrics

67. Prediction for different project subjects
MIM: the proposed metrics CM: source code metrics HM: history metrics

68. Prediction for different project subjects
T-test results (significant figures are in bold, p-value < 0.05)

69. Prediction with different algorithms

70. Prediction with different algorithms
T-test results (significant figures are in bold, p-value < 0.05)

71. Prediction in different training periods
Model training period Model testing period
Dec 2005 Sep 2010
Time P

72. Prediction in different training periods
50% : 50%
70% : 30%
80% : 20%
Model training period Model testing period
Dec 2005 Sep 2010
Time P

73. Prediction in different training periods

74. Prediction in different training periods
T-test results (significant figures are in bold, p-value < 0.05)

76. Top 42 (37%) from MIMs
among total 113 metrics
(MIMs+CMs+HMs)

77. Possible Insight
TOP 1: NumLowDOIEdit
TOP 2: NumPatternEXSX
TOP 3: TimeSpentOnEdit

78. Possible Insight
TOP 1: NumLowDOIEdit
TOP 2: NumPatternEXSX
TOP 3: TimeSpentOnEdit
Chances are that more defects might be generated
if a programmer TOP2 repeatedly edit and browse a
file especially related to the previous defects TOP3
with putting more weight on editing time, and
especially TOP1 when editing such the files less
frequently or less recently accessed ever …

79. Performance comparison
with regression modeling
for predicting # of post-defects

80. Predicting Post-Defect Numbers

81. Predicting Post-Defect Numbers
T-test results (significant figures are in bold, p-value < 0.05)

82. Threats to Validity
• Systems examined might not be representative
• Systems are all open source projects
• Defect information might be biased

83. Conclusion
Our findings exemplify that developer’s
interaction can affect software quality
Our proposed micro interaction metrics
improve defect prediction accuracy
significantly
…

84. We believe future defect prediction
models will use more developers’ direct and
micro level interaction information
MIMs are a first step towards it

85. Thank you! Any Question?
• Problem
– Developer’s interaction information can affect
software quality (defects)?
• Approach
– We proposed novel micro interaction metrics
(MIMs) overcoming the popular static metrics
• Result
– MIMs significantly improve prediction accuracy
compared to source code metrics (CMs) and
history metrics (HMs)

86. Backup Slides

88. One possible ARGUMENT:
Some developers may not
have used Mylyn to fix bugs

89. Error chance in counting post-defects
as a result getting biased labels
(i.e., incorrect % of buggy instances)

90. Repeated experiment using
same instances but with a
different heuristics of defect
counting, CVS-log-based
approach*
* with keywords: “fix”, “bug”, “bug report ID” in change logs

91. Prediction with CVS-log-based approach
CVS-log-based

92. Prediction with CVS-log-based approach
CVS-log-based
T-test results (significant figures are in bold, p-value < 0.05)

93. CVS-log-based approach reported more
additional post-defects
(more % of buggy-labeled instances)

94. CVS-log-based approach reported more
additional post-defects
(more % of buggy-labeled instances)
MIMs failed to feature them due to
lack of the corresponding Mylyn data

95. Note that it is difficult to
100% guarantee the quality of
CVS change logs
(e.g., no explicit bug ID, missing logs)