4. Motivation
► Machine Learning (ML) has been widely adopted in most industries.
► But not fully exploited in Software industry, especially Development&Test.
► Huge amount of data collected and available in SDLC.
► Significant correlation between the attributes of Change and Risk.
► Frequently used Risk-based test technique can be time consuming,
subjective assessments and limited by human memory.
4
To build ML models to predict Product Risk with real training
dataset from a development team in SB1
5. Problem: Risk Prediction
5
User
Change
Risk
Score
ML Prediction Model
Training
Dataset
Manual Risk
Assessment
Optionally
combined with
Test
Operation
Defect
Requirement
Development
Improved Test Planning
and Prioritization
9. Risk Score Calculation
9
DEFECT
For a given
Change found
in test phase
INCIDENT
for a given
Component
reported from Prod
i = 1 &
y in (critical, major,
normal, trivial)
Max #
of defects
Defect(i) * Severity(i) * Weight(y)
Incident(i) * Severity(i) * Weight(y)
i = 1 &
y in (critical, major,
normal, trivial)
Max #
of incidents
RISK SCORE
For a given
Change
10. 10
Risk Prediction with Regression
Change
Multiple Linear Regression
Risk Score
123 observations in labeled training dataset
ML-Classified
ML-Predicted
11. Measure success of prediction
How to demonstrate improvement of following metrics:
1. Test Effectiveness = No. of defect found / No. of Test Execution
2. MTBI – Mean time between Incident for a given Component
3. Test Productivity = Number of testing hours per Change
11
12. Choice of ML Technology
12
• Easy-to-use and quick enabler.
• Simulation with ML-algorithms.
• Lisence based – expensive.
• Cloud-based – «GDPR».
• Development (convenient packages).
• Flexibility and better understanding.
• Freeware.
• Local data governance.