DRONE: Predicting Priority of Reported Bugs by Multi-Factor Analysis
0 likes369 views
The document describes a machine learning approach called DRONE that uses multi-factor analysis to predict the priority of reported bugs. DRONE extracts over 30 features related to temporal, textual, author, severity, product, and related reports factors. It then applies linear regression and thresholding on a dataset of 178,609 bug reports to classify bugs into five priority levels. An evaluation shows DRONE improves the average F-measure over baselines from 18.75% to 29.47%, and identifies product and related report features as most discriminative.
![Priority Vs Severity
3
“Severity is assigned by customers [users] while
priority is provided by developers . . . customer
[user] reported severity does impact the developer
when they assign a priority level to a bug report,
but it’s not the only consideration. For example, it
may be a critical issue for a particular reporter that
a bug is fixed but it still may not be the right thing
for the eclipse team to fix.”
Eclipse PMC Member
Importance: P5 (lowest priority level), major (high severity)](https://image.slidesharecdn.com/dronetianlosun-131011003249-phpapp02/85/DRONE-Predicting-Priority-of-Reported-Bugs-by-Multi-Factor-Analysis-3-320.jpg)
![Author Factor
AUT1 Mean priority of all bugs reports made by the author of BR prior to the
reporting of BR.
AUT2 Median priority of all bugs reports made by the author of BR prior to the
reporting of BR.
AUT3 The number of bug reports made by the author of BR prior to the reporting
of BR.
Related Reports Factor [REP-, Sun et al.]
REP1 Mean priority of the top-20 most similar bug reports to BR as measured
using REP- prior to the reporting of BR.
REP2 Median priority of the top-20 most similar bug reports to BR as measured
using REP prior to the reporting of BR.
REP3-4 The same as REP1-2 except only the top 10 bug reports are considered.
REP5-6 The same as REP1-2 except only the top 5 bug reports are considered.
REP7-8 The same as REP1-2 except only the top 3 bug reports are considered.
REP9-10 The same as REP1-2 except only the top 1 bug reports are considered.
10](https://image.slidesharecdn.com/dronetianlosun-131011003249-phpapp02/85/DRONE-Predicting-Priority-of-Reported-Bugs-by-Multi-Factor-Analysis-10-320.jpg)
![? Accuracy (Precision, Recall, F-measure)
Compare with SEVERISprio [Menzies & Marcus], SEVERISprio+
? Efficiency (Run time)
? Top features (Fisher score)
Research Questions & Measurements
20](https://image.slidesharecdn.com/dronetianlosun-131011003249-phpapp02/85/DRONE-Predicting-Priority-of-Reported-Bugs-by-Multi-Factor-Analysis-20-320.jpg)
DRONE: Predicting Priority of Reported Bugs by Multi-Factor Analysis
- 1. DRONE: Predicting Priority of Reported Bugs by Multi-Factor Analysis Yuan Tian1, David Lo1, Chengnian Sun2 1Singapore Management University 2National University of Singapore
- 2. Bug tracking systems allow developers to prioritize which bugs are to be fixed first ¡ Manual process ¡ Depend on other bugs ¡ Time consuming What is priority and when it is assigned? New Assigned 300 reports to triage daily! Validity Check, Duplicate Check, Priority Assignment Developer Assignment Bug Triager 2
- 3. Priority Vs Severity 3 “Severity is assigned by customers [users] while priority is provided by developers . . . customer [user] reported severity does impact the developer when they assign a priority level to a bug report, but it’s not the only consideration. For example, it may be a critical issue for a particular reporter that a bug is fixed but it still may not be the right thing for the eclipse team to fix.” Eclipse PMC Member Importance: P5 (lowest priority level), major (high severity)
- 4. q Background q Approach Overall Framework Features Classification Module q Experiment Dataset Research Questions Results q Conclusion Outline 4
- 5. q Background q Approach Overall Framework Features Classification Module q Experiment Dataset Research Questions Results q Conclusion Outline 5
- 6. (1) summary (2) description (3) product (4) component (5) author (6) severity (7) priority. 1 2 3 4 5 6 Time-related Info. 7 6 Bug Report
- 7. q Background q Approach Overall Framework Features Classification Module q Experiment Dataset Research Questions Results q Conclusion Outline 7
- 8. Training Reports Related Reports Model Predicted Priority Testing Reports Temporal Textual Author Severity Product Model Builder Model Application Feature Extraction Module Classifier Module Training Phase Testing Phase Overall Framework 8
- 9. Temporal Factor TMP1 Number of bugs reported within 7 days before the reporting of BR. TMP2 Number of bugs reported with the same severity within 7 days before the reporting of BR. TMP3 Number of bugs reported with the same or higher severity within 7 days before the reporting of BR. TMP4-6 The same as TMP1-3 except the time duration is 1 day. TMP7-9 The same as TMP1-3 except the time duration is 3 days. TMP10-12 The same as TMP1-3 except the time duration is 30 days. Textual Factor TXT1-n Stemmed words from the description field of BR excluding stop words. Severity Factor SEV BR’s severity field. 9
- 10. Author Factor AUT1 Mean priority of all bugs reports made by the author of BR prior to the reporting of BR. AUT2 Median priority of all bugs reports made by the author of BR prior to the reporting of BR. AUT3 The number of bug reports made by the author of BR prior to the reporting of BR. Related Reports Factor [REP-, Sun et al.] REP1 Mean priority of the top-20 most similar bug reports to BR as measured using REP- prior to the reporting of BR. REP2 Median priority of the top-20 most similar bug reports to BR as measured using REP prior to the reporting of BR. REP3-4 The same as REP1-2 except only the top 10 bug reports are considered. REP5-6 The same as REP1-2 except only the top 5 bug reports are considered. REP7-8 The same as REP1-2 except only the top 3 bug reports are considered. REP9-10 The same as REP1-2 except only the top 1 bug reports are considered. 10
- 11. Product Factor PRO1 BR’s product field Note: categorical feature PRO2 Number of bug reports made for the same product as that of BR prior to the reporting of BR. PRO3 Number of bug reports made for the same product of the same severity as that of BR prior to the reporting of BR. PRO4 Number of bug reports made for the same product of the same or higher severity as those of BR prior to the reporting of BR. PRO5 Proportion of bug reports made for the same product as that of BR prior to the reporting of BR that are assigned priority P1. PRO6-9 The same as PRO5 except they are for priority P2-P5 respectively. PRO10 Mean priority of bug reports made for the same product as that of BR prior to the reporting of BR. PRO11 Median priority of bug reports mad for the same product as that of BR prior to the reporting of BR. PRO12-22 The same as PRO1-11 except they are for the component field of BR. 11
- 12. Training Reports Related Reports Model Predicted Priority Testing Reports Temporal Textual Author Severity Product Model Builder Model Application Feature Extraction Module Classifier Module Training Phase Testing Phase Overall Framework 12
- 13. Model Building Data Training Features Linear Regression Model GRAY: Thresholding and Linear Regression to Classify Imbalanced Data. 13 Map feature values to real numbers Training Phase
- 14. Model Building Data Training Features Linear Regression Model Application Model Validation Data Thresholding Thresholds GRAY: Thresholding and Linear Regression to Classify Imbalanced Data. 14 Training Phase
- 15. Model Building Data Training Features Linear Regression Model Application Model Validation Data Thresholding Thresholds GRAY: Thresholding and Linear Regression to Classify Imbalanced Data. 15 • Thresholding process maps real numbers to priority levels. Training Phase
- 16. Thresholding Process 16 BR1 1.2 BR2 1.4 BR3 3.1 BR4 3.5 BR5 2.1 BR6 3.2 BR7 3.4 BR8 3.7 BR9 1.3 BR10 4.5 BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 Sort BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 P1 P2 P3 P4 P5 1.2 1.4 3.4 3.7
- 17. Model Building Data Training Features Linear Regression Model Application Testing Features Model Predicted Priority Validation Data Thresholding Thresholds GRAY: Thresholding and Linear Regression to Classify Imbalanced Data. 17 Training Phase Testing Phase
- 18. q Background q Approach Overall Framework Features Classification Module q Experiment Dataset Research Questions Results q Conclusion Outline 18
- 19. q Eclipse Project § 2001-10-10 to 2007-12-14, § 178,609 bug reports. Dataset DRONE TestingDRONE Training Model Building Validation REP- 4.50% 6.89% 85.45% 1.95% 1.21% P1 P2 P3 P4 P5 19
- 20. ? Accuracy (Precision, Recall, F-measure) Compare with SEVERISprio [Menzies & Marcus], SEVERISprio+ ? Efficiency (Run time) ? Top features (Fisher score) Research Questions & Measurements 20
- 21. 0.00% 20.00% 40.00% 60.00% 80.00% 100.00% P1 P2 P3 P4 P5 F-measure DRONE SEVERISprio SEVERISprio+ RQ1: How accurate? 21 29.47% 18.75% 1. Baselines predict everything as P3 ! 2. Average F-measure improves from 18.75% to 29.47% 3. A relative improvement of 57.17%.
- 22. Approach Run Time (in seconds) Feature Extraction(train) Model Building Feature Extraction(test) Model Application SEVERISprio <0.01 812.18 <0.01 <0.01 SEVERISprio+ <0.01 773.62 <0.01 <0.01 DRONE 0.01 69.25 <0.01 <0.01 RQ2: How efficient? 22 Our approach is much faster in Model Building!
- 23. Feature PRO5 PRO16 REP1 REP3 PRO18 PRO10 PRO21 PRO7 REP5 Text “1663” RQ3: What are the top-features? 23
- 24. Feature PRO5 PRO16 REP1 REP3 PRO18 PRO10 PRO21 PRO7 REP5 Text “1663” RQ3: What are the top-features? 6 out of the top-10 features belong to the product factor family. 24
- 25. Feature PRO5 PRO16 REP1 REP3 PRO18 PRO10 PRO21 PRO7 REP5 Text “1663” RQ3: What are the top-features? 3 out of the top-10 features come from the related-report factor family. 25
- 26. Feature PRO5 PRO16 REP1 REP3 PRO18 PRO10 PRO21 PRO7 REP5 Text “1663” RQ3: What are the top-features? 1) org.eclipse.ui.internal.Wor kbench.run(Workbech.jav a:1663) 2) Appears in 15% P5 reports. 26
- 27. Conclusion yuan.tian.2012@smu.edu.sg q Priority prediction is an ordinal + imbalance classification problem - >linear regression + thresholding is one option. q DRONE can improve the average F- measure of baselines from 18.75% to 29.47%, a relative improvement of 57.17%. q Product factor features are the most discriminative features, followed by related-reports factor features.
- 28. Conclusion yuan.tian.2012@smu.edu.sg q Priority prediction is an ordinal + imbalance classification problem - >linear regression + thresholding is one option. q DRONE can improve the average F- measure of baselines from 18.75% to 29.47%, a relative improvement of 57.17%. q Product factor features are the most discriminative features, followed by related-reports factor features.
- 29. Conclusion yuan.tian.2012@smu.edu.sg q Priority prediction is an ordinal + imbalance classification problem - >linear regression + thresholding is one option. q DRONE can improve the average F- measure of baselines from 18.75% to 29.47%, a relative improvement of 57.17%. q Product factor features are the most discriminative features, followed by related-reports factor features.
- 30. 30 I acknowledge the support of Google and the ICSM organizers in the form of a Female Student Travel Grant, which enabled me to attend this conference. Thank you!
- 31. Conclusion yuan.tian.2012@smu.edu.sg q Priority prediction is an ordinal + imbalance classification problem - >linear regression + thresholding is one option. q DRONE can improve the average F- measure of baselines from 18.75% to 29.47%, a relative improvement of 57.17%. q Product factor features are the most discriminative features, followed by related-reports factor features.
- 32. APPENDIX
- 33. 33
- 34. P1:10% P2:20% P3:40% P4:20% P5:10% Proportions of each priority levels in Validation Data: After applying Linear Regression Model on Validation Data: BR1 1.2 BR2 1.4 BR3 3.1 BR4 3.5 BR5 2.1 BR6 3.2 BR7 3.4 BR8 3.7 BR9 1.3 BR10 4.5 BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 Sort BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 Initial P1 P2 P3 P4 P5 1.2 1.4 3.4 3.7 Predicted priority level 34
- 35. BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 P1 P2 P3 P4 P5 1.2 1.4 3.4 3.7 Initialized Thresholds: BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 P3 P4 P5 Tune one threshold Compute F-measures 1.2 1.4 3.4 3.7 1.1 1.3 Compute F-measures 35
- 36. 36 BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 1.2 1.4 3.4 3.7 Tune one threshold 1.1 1.3 Compute F-measures BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 P3 P4 P5 1.4 3.4 3.7 1.3 Update threshold value P2 P1 Higher
- 37. BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 1.4 3.4 3.7 1.3 Threshold 1 is fixed BR1 1.2 BR9 1.3 BR2 1.4 BR5 2.1 BR3 3.1 BR6 3.2 BR7 3.4 BR4 3.5 BR8 3.7 BR10 4.5 Tune for next threshold P3 P4 P5 1.4 3.4 3.7 1.3 P2 P1 37
- 38. q Menzies and Marcus (ICSM 2008) ¡ Analyze reports in NASA ¡ Textual features +feature selection+ RIPPER q Lamkanfi et al. (MSR 2010, CSMR 2011) ¡ Predict coarse-grained severity labels ¡ Severe vs. non-severe ¡ Analyze reports in open-source systems ¡ Compare and contrast various algorithms q Tian et al.(WCRE 2012) ¡ Information retrieval + k nearest neighbour Previous Research Work: Severity Prediciton 38
- 39. q Tokenization Spliting document into tokens according to delimiters. q Stop-word Removal eg: are, is, I, he q Stemming eg: woking, works, worked->work Text Pre-processing 39
- 40. 40 q Textual Features § Compute BM25Fext scores § Feature1: Extract unigram § Feature2: Extract bigrams q Non-Textual Features § Feature3: Product field § Feature4: Component Field Appendix: Similarity Between Bug Reports (REP-) Note: Weights are learned from duplicate bug reports.
- 41. Feature PRO5 Proportion of bug reports made for the same product as that of BR prior to the reporting of BR that are assigned priority P1. PRO16 Proportion of bug reports made for the same component as that of BR prior to the reporting of BR that are assigned priority P1. REP1 Mean priority of the top-20 most similar bug reports to BR as measured using REP- prior to the reporting of BR. REP3 Mean priority of the top-10 most similar bug reports to BR as measured using REP prior to the reporting of BR. PRO18 Proportion of bug reports made for the same component as that of BR prior to the reporting of BR that are assigned priority P3. PRO10 Mean priority of bug reports made for the same product as that of BR prior to the reporting of BR. PRO21 Mean priority of bug reports made for the same component as that of BR prior to the reporting of BR. PRO7 Proportion of bug reports made for the same product as that of BR prior to the reporting of BR that are assigned priority P3. REP5 Mean priority of the top-5 most similar bug reports to BR as measured using REP prior to the reporting of BR. Text “1663” 41