This document provides an overview of software defect prediction approaches from the 1970s to the present. It discusses early approaches using simple metrics like lines of code and complexity metrics. It then covers the development of prediction models using machine learning techniques like regression and classification. More recent topics discussed include just-in-time prediction models, practical applications in industry, using historical metrics from software repositories, addressing noise in data, and the feasibility of cross-project prediction. The document outlines challenges and opportunities for future work in the field of software defect prediction.
1 of 97
More Related Content
Survey on Software Defect Prediction
1. Survey on
Software Defect Prediction
- PhD Qualifying Examination -
July 3, 2014
Jaechang Nam
Department of Computer Science and Engineering
HKUST
2. Outline
• Background
• Software Defect Prediction Approaches
– Simple metric and defect estimation models
– Complexity metrics and Fitting models
– Prediction models
– Just-In-Time Prediction Models
– Practical Prediction Models and Applications
– History Metrics from Software Repositories
– Cross-Project Defect Prediction and Feasibility
• Summary and Challenging Issues
2
3. Motivation
• General question of software defect
prediction
– Can we identify defect-prone entities (source
code file, binary, module, change,...) in advance?
• # of defects
• buggy or clean
• Why?
– Quality assurance for large software
(Akiyama@IFIP’71)
– Effective resource allocation
• Testing (Menzies@TSE`07)
• Code review (Rahman@FSE’11)
3
5. Two Focuses on Defect
Prediction
• How much complex is software and its
process?
– Metrics
• How can we predict whether software has
defects?
– Models based on the metrics
5
9. Identifying Defect-prone Entities
• Akiyama’s equation (Ajiyama@IFIP`71)
– # of defects = 4.86 + 0.018 * LOC (=Lines Of Code)
• 23 defects in 1 KLOC
• Derived from actual systems
• Limitation
– Only LOC is not enough to capture software
complexity
9
10. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Cyclomati
c Metric
Halstea
d
Metrics
MetricsModelsOthers
11. Complexity Metrics and Fitting
Models
• Cyclomatic complexity metrics (McCabe`76)
– “Logical complexity” of a program represented in
control flow graph
– V(G) = #edge – #node + 2
• Halstead complexity metrics (Halsted`77)
– Metrics based on # of operators and operands
– Volume = N * log2n
– # of defects = Volume / 3000
11
12. Complexity Metrics and Fitting
Models
• Limitation
– Do not capture complexity (amount) of change.
– Just fitting models but not prediction models in
most of studies conducted in 1970s and early
1980s
• Correlation analysis between metrics and # of defects
– By linear regression models
• Models were not validated for new entities (modules).
12
13. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Cyclomati
c Metric
Halstea
d
Metrics
Process
Metrics
MetricsModelsOthers
Prediction Model (Classification)
14. Regression Model
• Shen et al.’s empirical study (Shen@TSE`85)
– Linear regression model
– Validated on actual new modules
– Metrics
• Halstead, # of conditional statements
• Process metrics
– Delta of complexity metrics between two successive system versions
– Measures
• Between actual and predicted # of defects on new modules
– MRE (Mean magnitude of relative error)
» average of (D-D’)/D for all modules
• D: actual # of defects
• D’: predicted # of defects
» MRE = 0.48
14
15. Classification Model
• Discriminative analysis by Munson et al.
(Munson@TSE`92)
• Logistic regression
• High risk vs. low risk modules
• Metrics
– Halstead and Cyclomatic complexity metrics
• Measure
– Type I error: False positive rate
– Type II error: False negative rate
• Result
– Accuracy: 92% (6 misclassification out of 78 modules)
– Precision: 85%
– Recall: 73%
– F-measure: 88%
15
16. ?
Defect Prediction Process
(Based on Machine Learning)
16
Classification /
Regression
Software
Archives
B
C
C
B
...
2
5
0
1
...
Instances with
metrics (features)
and labels
B
C
B
...
2
0
1
...
Training
Instances
(Preprocessing
)
Model
?
New instances
Generate
Instances
Build
a
model
17. Defect Prediction
(Based on Machine Learning)
• Limitations
– Limited resources for process metrics
• Error fix in unit testing phase was conducted
informally by an individual developer (no error
information available in this phase). (Shen@TSE`85)
– Existing metrics were not enough to capture
complexity of object-oriented (OO) programs.
– Helpful for quality assurance team but not for
individual developers
17
18. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
Process
Metrics
MetricsModelsOthers
Just-In-Time Prediction Model
Practical Model and
Applications
History
Metrics
CK Metrics
19. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
Just-In-Time Prediction Model
Practical Model and
Applications
Process
Metrics
MetricsModelsOthers
History
Metrics
CK Metrics
20. Risk Prediction of Software
Changes
(Mockus@BLTJ`00)
• Logistic regression
• Change metrics
– LOC added/deleted/modified
– Diffusion of change
– Developer experience
• Result
– Both false positive and false negative rate: 20% in
the best case
20
21. Risk Prediction of Software
Changes
(Mockus@BLTJ`00)
• Advantage
– Show the feasible model in practice
• Limitation
– Conducted 3 times per week
• Not fully Just-In-Time
– Validated on one commercial system (5ESS
switching system software)
21
22. BugCache (Kim@ICSE`07)
• Maintain defect-prone entities in a cache
• Approach
• Result
– Top 10% files account for 73-95% of defects on 7
systems
22
23. BugCache (Kim@ICSE`07)
• Advantages
– Cache can be updated quickly with less cost. (c.f. static
models based on machine learning)
– Just-In-Time: always available whenever QA teams want
to get the list of defect-prone entities
• Limitations
– Cache is not reusable for other software projects.
– Designed for QA teams
• Applicable only in a certain time point after a bunch of changes
(e.g., end of a sprint)
• Still limited for individual developers in development phase
23
24. Change Classification (Kim@TSE`08)
• Classification model based on SVM
• About 11,500 features
– Change metadata such as changed LOC, change count
– Complexity metrics
– Text features from change log messages, source code,
and file names
• Results
– 78% accuracy and 60% recall on average from 12 open-
source projects
24
26. Follow-up Studies
• Studies addressing limitations
– “Reducing Features to Improve Code Change-Based Bug
Prediction” (Shivaji@TSE`13)
• With less than 10% of all features, buggy F-measure is 21%
improved.
– “Software Change Classification using Hunk Metrics”
(Ferzund@ICSM`09)
• 27 hunk-level metrics for change classification
• 81% accuracy, 77% buggy hunk precision, and 67% buggy hunk
recall
– “A large-scale empirical study of just-in-time quality
assurance” (Kamei@TSE`13)
• 14 process metrics (mostly from Mockus`00)
• 68% accuracy, 64% recall on 11open-source and commercial
projects
– “An Empirical Study of Just-In-Time Defect Prediction
Using Cross-Project Models” (Fukushima@MSR`14)
• Median AUC: 0.72 26
27. Challenges of JIT model
• Practical validation is difficult
– Just 10-fold cross validation in current literature
– No validation on real scenario
• e.g., online machine learning
• Still difficult to review huge change
– Fine-grained prediction within a change
• e.g., Line-level prediction
27
28. Next Steps of Defect Prediction
1980s 1990s 2000s 2010s 2020s
Online Learning JIT Model
Prediction Model (Regression)
Prediction Model (Classification)
Just-In-Time Prediction Model
Process
Metrics
MetricsModelsOthers
Fine-grained
Prediction
29. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
Just-In-Time Prediction Model
Practical Model and
Applications
Process
Metrics
MetricsModelsOthers
History
Metrics
CK Metrics
30. Defect Prediction in Industry
• “Predicting the location and number of faults in
large software systems” (Ostrand@TSE`05)
– Two industrial systems
– Recall 86%
– 20% most fault-prone modules account for 62% faults
30
31. Case Study for Practical Model
• “Does Bug Prediction Support Human Developers?
Findings From a Google Case Study” (Lewis@ICSE`13)
– No identifiable change in developer behaviors after using
defect prediction model
• Required characteristics but very challenging
– Actionable messages / obvious reasoning
31
32. Next Steps of Defect Prediction
1980s 1990s 2000s 2010s 2020s
Actionable
Defect
Prediction
Prediction Model (Regression)
Prediction Model (Classification)
Just-In-Time Prediction Model
Practical Model and
Applications
Process
Metrics
MetricsModelsOthers
33. Evaluation Measure for Practical
Model
• Measure prediction performance based on
code review effort
• AUCEC (Area Under Cost Effectiveness Curve)
33
Percent of LOC
Percentofbugsfound
0
100%
100%
50%10%
M1
M2
Rahman@FSE`11, Bugcache for inspections: Hit or miss?
34. Practical Application
• What else can we do more with defect
prediction models?
– Test case selection on regression testing
(Engstrom@ICST`10)
– Prioritizing warnings from FindBugs
(Rahman@ICSE`14)
34
35. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Process
Metrics
MetricsModelsOthers
Practical Model and
Applications
Just-In-Time Prediction Model
History
Metrics
36. Representative OO Metrics
Metric Description
WMC Weighted Methods per Class (# of methods)
DIT Depth of Inheritance Tree ( # of ancestor classes)
NOC Number of Children
CBO Coupling between Objects (# of coupled classes)
RFC
Response for a class: WMC + # of methods called by the
class)
LCOM
Lack of Cohesion in Methods (# of "connected
components”)
36
• CK metrics (Chidamber&Kemerer@TSE`94)
• Prediction Performance of CK vs. code
(Basili@TSE`96)
– F-measure: 70% vs. 60%
37. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Process
Metrics
MetricsModelsOthers
Practical Model and
Applications
Just-In-Time Prediction Model
History
Metrics
38. Representative History Metrics
38
Name
# of
metrics
Metric
source
Citation
Relative code change churn 8 SW Repo.* Nagappan@ICSE`05
Change 17 SW Repo. Moser@ICSE`08
Change Entropy 1 SW Repo. Hassan@ICSE`09
Code metric churn
Code Entropy
2 SW Repo. D’Ambros@MSR`10
Popularity 5
Email
archive
Bacchelli@FASE`10
Ownership 4 SW Repo. Bird@FSE`11
Micro Interaction Metrics (MIM) 56 Mylyn Lee@FSE`11
* SW Repo. = version control system + issue tracking system
39. Representative History Metrics
• Advantage
– Better prediction performance than code metrics
39
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
Moser`08 Hassan`09 D'Ambros`10 Bachille`10 Bird`11 Lee`11
Performance Improvement
(all metrics vs. code complexity metrics)
(F-measure) (F-measure)(Absolute
prediction
error)
(Spearman
correlation)
(Spearman
correlation)
(Spearman
correlation*)
(*Bird`10’s results are from two metrics vs. code metrics, No comparison data in Nagappan`05)
Performance
Improvement
(%)
40. History Metrics
• Limitations
– History metrics do not extract particular program
characteristics such as developer social network,
component network, and anti-pattern.
– Noise data
• Bias in Bug-Fix Dataset(Bird@FSE`09)
– Not applicable for new projects and projects lacking in
historical data
40
41. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Noise
Reduction
Semi-
supervised/active
42. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Noise
Reduction
Semi-
supervised/active
43. Other Metrics
43
Name
# of
metrics
Metric
source
Citation
Component network 28
Binaries
(Windows
Server
2003)
Zimmermann@ICSE`0
8
Developer-Module network 9
SW Repo. +
Binaries
Pinzger@FSE`08
Developer social network 4 SW Repo. Meenely@FSE`08
Anti-pattern 4
SW Repo. +
Design-
pattern
Taba@ICSM`13
* SW Repo. = version control system + issue tracking system
44. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Noise
Reduction
Semi-
supervised/active
45. Noise Reduction
• Noise detection and elimination algorithm
(Kim@ICSE`11)
– Closest List Noise Identification (CLNI)
• Based on Euclidean distance between instances
– Average F-measure improvement
• 0.504 0.621
• Relink (Wo@FSE`11)
– Recover missing links between bugs and
changes
– 60% 78% recall for missing links
– F-measure improvement
• e.g. 0.698 (traditional) 0.731 (ReLink)
45
46. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Semi-
supervised/active
47. Defect Prediction for New Software
Projects
• Universal Defect Prediction Model
• Simi-supervised / active learning
• Cross-Project Defect Prediction
47
48. Universal Defect Prediction Model
(Zhang@MSR`14)
• Context-aware rank transformation
– Transform metric values ranged from 1 to 10 across all
projects.
• Model built by 1398 projects collected from
SourceForge and Google code
48
49. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Semi-
supervised/active
50. Other approaches for CDDP
• Semi-supervised learning with dimension
reduction for defect prediction (Lu@ASE`12)
– Training a model by a small set of labeled
instances together with many unlabeled
instances
– AUC improvement
• 0.83 0.88 with 2% labeled instances
• Sample-based semi-supervised/active
learning for defect prediction (Li@AESEJ`12)
– Average F-measure
• 0.628 0.685 with 10% sampled instances
50
51. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Semi-
supervised/active
52. Cross-Project Defect Prediction
(CPDP)
• For a new project or a project lacking
in the historical data
52
?
?
?
Training
Test
Model
Project A Project B
Only 2% out of 622 prediction combinations worked. (Zimmermann@FSE`09)
53. Transfer Learning (TL)
27
Traditional Machine Learning
(ML)
Learnin
g
System
Learnin
g
System
Transfer Learning
Learnin
g
System
Learnin
g
System
Knowledge
Transfer
Pan et al.@TNN`10, Domain Adaptation via Transfer Component Analysis
54. CPDP
54
• Adopting transfer learning
Transfer learning
Metric
Compensation
NN Filter TNB TCA+
Preprocessing N/A
Feature selection,
Log-filter
Log-filter Normalization
Machine learner C4.5 Naive Bayes TNB Logistic Regression
# of Subjects 2 10 10 8
# of predictions 2 10 10 26
Avg. f-measure
0.67
(W:0.79, C:0.58)
0.35
(W:0.37, C:0.26)
0.39
(NN: 0.35, C:0.33)
0.46
(W:0.46, C:0.36)
Citation
Watanabe@PROMISE
`08
Turhan@ESEJ`09 Ma@IST`12 Nam@ICSE`13
* NN = Nearest neighbor, W = Within, C = Cross
55. Metric Compensation
(Watanabe@PROMISE`08)
• Key idea
• New target metric value =
target metric value * average source metric value
average target metric value
55
s
Source Target New Target
57. NN filter
(Turhan@ESEJ`09)
• Key idea
• Nearest neighbor filter
– Select 10 nearest source instances of
each target instance
57
New Source Target
Hey, you look like me! Could you be my model?
Source
58. NN filter (cont.)
(Turhan@ESEJ`09)
58
Transfer learning
Metric
Compensation
NN Filter TNB TCA+
Preprocessing N/A
Feature selection,
Log-filter
Log-filter Normalization
Machine learner C4.5 Naive Bayes TNB Logistic Regression
# of Subjects 2 10 10 8
# of predictions 2 10 10 26
Avg. f-measure
0.67
(W:0.79, C:0.58)
0.35
(W:0.37, C:0.26)
0.39
(NN: 0.35, C:0.33)
0.46
(W:0.46, C:0.36)
Citation
Watanabe@PROMISE
`08
Turhan@ESEJ`09 Ma@IST`12 Nam@ICSE`13
* NN = Nearest neighbor, W = Within, C = Cross
59. Transfer Naive Bayes
(Ma@IST`12)
• Key idea
59
Target
Hey, you look like me! You will get more chance to be my best
model!
Source
Provide more weight to similar source instances to build a Naive Bayes Model
Build a model
Please, consider me more important than other
instances
60. Transfer Naive Bayes (cont.)
(Ma@IST`12)
• Transfer Naive Bayes
– New prior probability
– New conditional probability
60
61. Transfer Naive Bayes (cont.)
(Ma@IST`12)
• How to find similar source instances for target
– A similarity score
– A weight value
61
F1 F2 F3 F4 Score (si)
Max of target 7 3 2 5 -
src. inst 1 5 4 2 2 3
src. inst 2 0 2 5 9 1
Min of target 1 2 0 1 -
k=# of features, si=score of instance i
62. Transfer Naive Bayes (cont.)
(Ma@IST`12)
62
Transfer learning
Metric
Compensation
NN Filter TNB TCA+
Preprocessing N/A
Feature selection,
Log-filter
Log-filter Normalization
Machine learner C4.5 Naive Bayes TNB Logistic Regression
# of Subjects 2 10 10 8
# of predictions 2 10 10 26
Avg. f-measure
0.67
(W:0.79, C:0.58)
0.35
(W:0.37, C:0.26)
0.39
(NN: 0.35, C:0.33)
0.46
(W:0.46, C:0.36)
Citation
Watanabe@PROMISE
`08
Turhan@ESEJ`09 Ma@IST`12 Nam@ICSE`13
* NN = Nearest neighbor, W = Within, C = Cross
63. TCA+
(Nam@ICSE`13)
• Key idea
– TCA (Transfer Component Analysis)
63
Source Target
Oops, we are different! Let’s meet in another world!
New Source New Target
64. Transfer Component Analysis (cont.)
• Feature extraction approach
– Dimensionality reduction
– Projection
• Map original data
in a lower-dimensional feature space
64
1-dimensional feature
space
2-dimensional feature
space
65. TCA (cont.)
65
Pan et al.@TNN`10, Domain Adaptation via Transfer Component Analysis
Target domain data
Source domain data
67. TCA+
(Nam@ICSE`13)
67
Source Target
Oops, we are different! Let’s meet at another world!
New Source New Target
But, we are still a bit different!
Source Target
Oops, we are different! Let’s meet at another world!
New Source New Target
Normalize US together!
TCA
TCA+
68. Normalization Options
• NoN: No normalization applied
• N1: Min-max normalization (max=1, min=0)
• N2: Z-score normalization (mean=0, std=1)
• N3: Z-score normalization only using source mean
and standard deviation
• N4: Z-score normalization only using target mean
and standard deviation
13
69. Preliminary Results using TCA
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
F-measure
69*Baseline: Cross-project defect prediction without TCA and normalization
Prediction performance of TCA
varies according to different
normalization options!
Baseline NoN N1 N2 N3
N4
Baseline NoN N1 N2 N3
N4
Project A Project B Project B Project A
F-measure
70. TCA+: Decision Rules
• Find a suitable normalization for TCA
• Steps
– #1: Characterize a dataset
– #2: Measure similarity
between source and target datasets
– #3: Decision rules
70
71. TCA+: #1. Characterize a
Dataset
71
3
1
…
Dataset A Dataset B
2
4
5
8
9
6
11
d1,2
d1,5
d1,3
d3,11
3
1
…
2
4
5
8
9
6
11
d2,6
d1,2
d1,3
d3,11
DIST={dij : i,j, 1 ≤ i < n, 1 < j ≤ n, i
< j}
A
72. TCA+: #2. Measure Similarity
between Source and Target
• Minimum (min) and maximum (max) values of
DIST
• Mean and standard deviation (std) of DIST
• The number of instances
72
73. TCA+: #3. Decision Rules
• Rule #1
– Mean and Std are same NoN
• Rule #2
– Max and Min are different N1 (max=1, min=0)
• Rule #3,#4
– Std and # of instances are different
N3 or N4 (src/tgt mean=0, std=1)
• Rule #5
– Default N2 (mean=0, std=1)
73
75. Current CPDP using TL
• Advantages
– Comparable prediction performance to within-prediction
models
– Benefit from the state-of-the-art TL approaches
• Limitation
– Performance of some cross-prediction pairs is still poor.
(Negative Transfer)
75
Source Target
76. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Semi-
supervised/active
77. Feasibility Evaluation for CPDP
• Solution for negative transfer
– Decision tree using project characteristic metrics
(Zimmermann@FSE`09)
• E.g. programming language, # developers, etc.
77
78. Follow-up Studies
• “An investigation on the feasibility of cross-project
defect prediction.” (He@ASEJ`12)
– Decision tree using distributional characteristics of a
dataset E.g. mean, skewness, peakedness, etc.
78
79. Feasibility for CPDP
• Challenges on current studies
– Decision trees were not evaluated properly.
• Just fitting model
– Low target prediction coverage
• 5 out of 34 target projects were feasible for cross-
predictions (He@ASEJ`12)
79
80. Next Steps of Defect Prediction
1980s 1990s 2000s 2010s 2020s
Cross-Prediction
Feasibility Model
Prediction Model (Regression)
Prediction Model (Classification)
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
History
Metrics
Other Metrics
Semi-
supervised/active
81. Semi-
supervised/active
Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
History
Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Other Metrics
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
Personalized
Model
82. Cross-prediction Model
• Common challenge
– Current cross-prediction models are limited to datasets
with same number of metrics
– Not applicable on projects with different feature spaces
(different domains)
• NASA Dataset: Halstead, LOC
• Apache Dataset: LOC, Cyclomatic, CK metrics
82
Source Target
83. Next Steps of Defect Prediction
1980s 1990s 2000s 2010s 2020s
Prediction Model (Regression)
Prediction Model (Classification)
CK Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
Cross-Domain
Prediction
History Metrics
Other Metrics
Noise
Reduction
Semi-
supervised/activePersonalized
Model
85. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
History Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Other Metrics
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
Data Privacy
Noise
Reduction
Semi-
supervised/activePersonalized
Model
86. Other Topics
• Privacy issue on defect datasets
– MORPH (Peters@ICSE`12)
• Mutate defect datasets while keeping prediction accuracy
• Can accelerate cross-project defect prediction with
industrial datasets
• Personalized defect prediction model (Jiang@ASE`13)
– “Different developers have different coding styles,
commit frequencies, and experience levels, all of which
cause different defect patterns.”
– Results
• Average F-measure: 0.62 (personalized models) vs. 0.59 (non-
personalized models)
86
87. Outline
• Background
• Software Defect Prediction Approaches
– Simple metric and defect estimation models
– Complexity metrics and Fitting models
– Prediction models
– Just-In-Time Prediction Models
– Practical Prediction Models and Applications
– History Metrics from Software Repositories
– Cross-Project Defect Prediction and Feasibility
• Summary and Challenging Issues
87
88. Defect Prediction Approaches
1970s 1980s 1990s 2000s 2010s
LOC
Simple Model
Fitting Model
Prediction Model (Regression)
Prediction Model (Classification)
Cyclomati
c Metric
Halstea
d
Metrics
CK Metrics
History Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Other Metrics
Practical Model and
Applications
Data Privacy
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
Noise
Reduction
Semi-
supervised/activePersonalized
Model
89. Next Steps of Defect Prediction
1980s 1990s 2000s 2010s 2020s
Online Learning JIT Model
Actionable
Defect
Prediction
Cross-Prediction
Feasibility Model
Prediction Model (Regression)
Prediction Model (Classification)
CK Metrics
History Metrics
Just-In-Time Prediction Model
Cross-Project Prediction
Other Metrics
Practical Model and
Applications
Universa
l Model
Process
Metrics
Cross-Project
Feasibility
MetricsModelsOthers
Cross-Domain
Prediction
Fine-grained
Prediction
Data Privacy
Noise
Reduction
Semi-
supervised/activePersonalized
Model
95. Evaluation Measures
(classification)
• AUCEC (Area Under Cost Effectiveness Curve)
95
Percent of LOC
Percentofbugsfound
0
100%
100%
50%10%
M1
M2
Rahman@FSE`11, Bugcache for inspections: Hit or miss?
96. Evaluation Measures
(Regression)
• Target
– Metric values vs. the number of bugs
– Actual vs. predicted number of bugs
• Correlation coefficient
– Spearman / Pearson /R2
• Mean squared error
96
97. CK metrics
Metric Description
WMC Weighted Methods per Class (# of methods)
DIT Depth of Inheritance Tree ( # of ancestor classes)
NOC Number of Children
CBO Coupling between Objects (# of coupled classes)
RFC
Response for a class: WMC + # of methods called by the
class)
LCOM
Lack of Cohesion in Methods (# of "connected
components”)
97
Editor's Notes
Cross-project change classification
Feasibility evaluation on cross-project defect prediction
Predicting software quality
Akiyama’s model is the earliest prediction model that predicts the number defects by using size of software such as LOC, # of subroutine calls. IFIC=International Federation of Information Processing
Testing a entire system is not feasible. (Menzies`07)
Inspecting source code is costly as well. (Rahman`11)
complex software itself, complex development process, even developers solving complex problems can introdue bugs on software
Depending on software, high recall might be more important than precision and vice versa.
Cross-project change classification
Feasibility evaluation on cross-project defect prediction
V = N * log_2 n
n = total # of distinct operands and operators
N = total # of distinct operands and operators
Correlation analysis using linear regression
V = N * log_2 n
n = total # of distinct operands and operators
N = total # of distinct operands and operators
Correlation analysis using linear regression
Result from Command and Control Commutation System implanted in Ada
Considered different thresholds for discriminative probability
Diffusion of change: How many file/modules/subsystems were touched together?
Developer experience: # of previous changes by the same developer. Weighted by considering contributions of the set of developers.
Diffusion of change: How many file/modules/subsystems were touched together?
Developer experience: # of previous changes by the same developer. Weighted by considering contributions of the set of developers.
When a bug is found but not in the BugCache, this is a cache miss.
Then, cache is updated with source code files based on locality.
When the cache is full, the cache is replaced based on Least recently used policy that is used for common cache policy in operation system
Based on cache hit or miss, update the cache
Cache miss: entities fixed are not in cache
Load the entities and nearby entities (locality) to the cache
Locality
Files/functions changed together with defects
Recently added files/functions
Recently changed files/functions
Cache replacement policy:
Least Recently Used (LRU) weighted by # of previous defects.
complexity metrics of old and new revision files and then compute delta between the old and new.
complexity metrics of old and new revision files and then compute delta between the old and new.
10 metrics are from Mockus`10
Fukushima: cross-prediction performance can be improved
As there was a transition from fitting model to prediction model, we need another transition from JIT prediction mode to Online learning based JIT prediction models.
Are defect prediction models practical in industry?
As there was a transition from fitting model to prediction model, we need another transition from JIT prediction mode to Online learning based JIT prediction models.
Engstrom: found more defect selected by defect prediction results
WMC: A class with more member methods than its peers is considered to be more complex and therefore more error prone.
DIT: # of ancestor classes
NOC: the number of direct descendants (subclasses) for each class
CBO:
RFC:
LCOM: the number of "connected components" in a class
Cohesion – 연관있는 메서드들은 한 클래안에 다 모아 넣기 high cohesion
Coupling –
Relative code change churn: e.g. churned LOC (the accumulative number of deleted and added lines between a base version and a new version of a source file) divided by Total LOC
Change: e.g., # of revisions, # of authors editing a file
Change Entropy: quantify complexity of changes by using Entropy theory. How many times changed in the same period.
Code metric churn: churned metric value, collected biweekly basis.
Code Entropy: how many lines of code changed in the same period?
Popularity: source code files discussed a lot in emails
Ownership: % of commits by a developer for a source code file
MIM: How long does the source code file is edited.
10% recall improvement for Zimmermann’s approach (All vs. CM)
20% decrease of AIC, 200% increase of D^2 (Taba`10)
Tested on 5 external projects
Average Within F-measure: 0.42
Average Universal F-measure: 0.39
Average Within AUC: 0.72
Average Universal AUC: 0.72
In traditional machine learning, we build a learning system by using instances in a same domain. In our case, the same domain means the same project.
This is same as within-project defect prediction in this research.
However, transfer learning is reusing knowledge from a certain domain which has enough data.
Cross prediction is simply reusing learning system of a source project for a target project.However, in transfer learning, we just extract proper knowledge, which will be really helpful for the target domain.
So, transfer learning algorithms play a role for smart knowledge transfer for the target domain.
This is more than just a simple cross prediction.
How to transfer knowledge from a source is a transfer learning algorithm!
mass of a source instance = sM
mass of test data = kmM
kmM^2은 constant
i = index of instance
k = # of feautres
s_i = similarity score of instance i
m = # of instances
M = mass of one feature in one instance
source instance: s_i * M
m * k * M
mass of a source instance = sM
mass of test data = kmM
kmM^2은 constant
Why cross results are different between NN and TNB? TNB doesn’t apply feature selection.
TNB didn’t report within-results
In machine leaning, there is a feature extraction approach to reduce feature space of data set.
Feature extraction is achieved by a technique called projection. Projection technique maps original data in a low dimensional feature space.
Here is an example of 2-dimensional feature space of a data set. There are four instances labeled.
We project a light on this space to 1-dimensional space, and then four instances are mapped in the one-dimensional space.
PCA is just for reducing feature space dimensionality. However, Transfer component analysis, TCA, try to find a new feature space where the distribution of source and target data sets are similar by projection.
The representative technique is PCA.
I’d like to show how PCA is different from TCA by an example.
Here is an example showing how PCA and TCA works.
In two-dimensional space, there are source and target data sets and we can see distributions are clearly different.
If we apply PCA and TCA , and then we can get the following results in one-dimensional space.
Probability density function
Probability mass function
In PCA, instances are projected into one dimensional space, however, distribution between source and target are still different.
In TCA, all instances are also projected in one-dimensional space, where distribution between source and target is similar.
Positive and negative instance of both training and test domains have discriminative power as shown in this figure.
You can check detailed equations about this algorithm in this paper
[add labels]
Based on these normalization techniques, we defined several normalization options for defect prediction data sets.
NI is min-max normalization which makes maximum and minimum value as 1 and 0 respectively.
N2 is z-score normalization which makes mean and standard deviation as 0 and 1 respectively.
We assume that some data sets may not have enough statistical information. So we defined variations of z-score normalization.
To normalize both source and target data sets, N3 is only using mean and standard deviation from source data (when target data does not have enough statistical information. For example, lack of instances in a data set.
N4 is only using target information for normalizing both source and target data sets.
This is the preliminary results of some prediction combinations.
Baseline means cross-project prediction without and normalization
In Safe to Apache, all TCA results with or without normalization are better than baseline.
However, in Apache to safe, N1, N3 didn’t outperform Baseline. This could be observed in other prediction combinations.
So, we could conclude prediction performance of TCA varies according to different normalization options.
TCA+ provides decision rules to select suitable normalization option.
For the decision rules, we first characterize both source and target data sets to identify their difference.
In the second step, we measure similarity between source and target data sets.
With degree of similarity, we created decision rules!
Then, how could we characterize data set?
Here are two data sets.
Intuitively, Data set A’s distribution is more sparser than data set B.
To quantify this difference, we compute Euclidean distance of all pairs of instances in each data set.
We defined DIST set for distances of all pairs.
Likewise, we can get DIST set from Data set B.
To measure similarity, we compute statistical parameters from DIST set such as minimum, maximum, mean, standard deviation, and the # of instances.
With these information, we creted decision rules
These are decision rules.
If mean and std is same, we assume that distributions bewteen source and target is same. So we applied no normalization.
For Rule2, if max and min values are different, we used N1(min-max normalization)
for Rule3 and 4, we considered std and # of instances. If target information is not enough, then we used source mean and std to normalize both datasets.
In case of Rule 5, if there are no rules are applicable, we applied N2 option, which make mean and std as 0 and 1 respectively.
This decision tree shows precision in advance.
Successful criteria
Precision > 0.5 and Recall > 0.7
As there was a transition from fitting model to prediction model, we need another transition from JIT prediction mode to Online learning based JIT prediction models.
Assumed projects in the same group have the similar distribution
Tested on 5 external projects
Average Within F-measure: 0.42
Average Universal F-measure: 0.39
Average Within AUC: 0.72
Average Universal F-measure: 0.72
As there was a transition from fitting model to prediction model, we need another transition from JIT prediction mode to Online learning based JIT prediction models.
Cross-project change classification
Feasibility evaluation on cross-project defect prediction
As there was a transition from fitting model to prediction model, we need another transition from JIT prediction mode to Online learning based JIT prediction models.
WMC: A class with more member methods than its peers is considered to be more complex and therefore more error prone.
DIT: # of ancestor classes
NOC: the number of direct descendants (subclasses) for each class
CBO:
RFC:
LCOM: the number of "connected components" in a class
Cohesion – 연관있는 메서드들은 한 클래안에 다 모아 넣기 high cohesion
Coupling –