Bug prediction based on your code history
summary
Pipeline links Git commits to Jira tickets and Sonar metrics, extracts ~15 features (time of day, lines changed, package touched, author history) from 18,000 commits over 15 years (≈25 engineers), labels 4,000 historical buggy commits, and trains classifiers (decision trees, SVM, naive Bayes) using Veka for experiments. The deployed detector posts high-risk commits to HipChat; a Naive Bayes model reaches about 77% accuracy on held-out data, while current production sees ~12 weekly user-reported bugs.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 4:23Bugs in production
- 7:58Data sources
- 11:15Metrics and prep
- 13:46Weka tool
- 15:38Classifier selection
- 19:12Naive Bayes classifier
- 23:09Support vector machines
- 26:20Analysis results
- 30:36Conclusions and use
- 35:00Deployment and testing
- 40:07Q&A
related talks
35mBetter C++ using Machine Learning on Large Projects
53mBug prediction based on your code history
38mApplying machine learning to software development to reduce bugs
07mInterview
48mUsing Machine Learning to find Linux bugs
24mUncovering Community and Project Insights Through Data Driven Methods
27mMachine Learning Practical Example
12mUse ML to Predict Delivery Time and Improve Code Reliability