1. REMI: Defect Prediction for
Efficient API Testing
ESEC/FSE 2015, Industrial track
September 3, 2015
Mijung Kim*, Jaechang Nam*, Jaehyuk Yeon+, Soonhwang Choi+, and Sunghun
Kim*
*Department of Computer Science and Engineering, HKUST
+Software R&D Center, Samsung Electronics CO., LTD

2. 2
...

3. Motivation
• Reliability of APIs
• Cost-intensive software quality assurance
(QA) tasks at Samsung
– Creating test cases for APIs
– Testing APIs
3

4. Testing Phase at Samsung
Test
Cases
(TCs) Run
TCs
Passin
g Tests
Failing
Tests
Fix
APIs
4
APIs
• Test Cases
– 1 or 2 test cases for each API
• Uniformly distributed across all APIs
• Tedious task and not efficient
Modification and Enhancement

5. Goal
• Apply software defect prediction for the
efficient API testing.
5

6. Proposed Approach: REMI
Risk Evaluation Method for Interface testing
6
Test
Cases
(TCs) Run
TCs
Passin
g Tests
Failing
Tests
Fix
APIs
APIs
REMI
(API Rank)
Modification and Enhancement

7. 7
Predict
Training
?
?
Model
API Project A
: Metric value
: Buggy-labeled instance
: Clean-labeled instance
?: Unlabeled instance
Software Defect Prediction
Related Work
Munson@TSE`92, Basili@TSE`95, Menzies@TSE`07,
Hassan@ICSE`09, Bird@FSE`11,D’ambros@EMSE112
Lee@FSE`11,...

8. Approach
REMI: Risk Evaluation Method for Interface
testing
8
Collect
Metrics
Aggregate
Metrics
Label
Buggy/Clean
APIs
Build
Prediction
Model
Rank
APIs
API
Ranks
SW Repository
Defect
Histor
y
Funtion-level metrics
Code
Metrics
(28)
AltCountLineBlank, AltCountLineCode,
AltCountLineComment, CountInput,
CountLine, CountLineBlank,
CountLineCode, CountLineCodeDecl,
CountLineCodeExe, CountLineComment,
CountLineInactive,
CountLinePreprocessor, CountOutput,
CountPath, CountSemicolon, CountStmt,
CountStmtDecl, CountStmtEmpty,
CountStmtExe, Cyclomatic,
CyclomaticModified, CyclomaticStrict,
Essential, Knots, MaxEssentialKnots,
MaxNesting, MinEssentialKnots,
RatioCommentToCode
Process
Metrics
(12)
(Moser@ICSE`08)
numCommits, totalLocDeleted,
avgLocDeleted, maxLocDeleted,
totalLocAdded, avgLocAdded,
maxLocAdded,
totalLocChurn, avgLocChurn,
Rank APIs in Package A
Probabilit
y
1 api_a 98%
2 api_b 90%
3 api_c 79%
... ... ...
25 api_y 40%
26 api_z 30%
Test
Failure
s
Depth 2
(LoC=10)
Depth 1
(LoC=7)
Depth 0
(LoC=3)
my_api()
func_1()
func_3()
func_4()
func_2() func_5()

9. Research Questions
• Test Development Phase
– Can writing test cases based on REMI
ranking identify additional/distinct
defects?
• Test Execution Phase
– Can test execution with REMI identify
defects more quickly than that without
REMI?
9

10. Experimental Setup
• Random Forest
• Subject
– Tizen-wearable APIs
• Applied REMI for 36 functional packages with
about 1100 APIs
– Release Candidates (RC)
• RC2 to RC3
10
RCn-1 RCn
Build Rank

11. RESULT
11

12. Results for Test Development
Phase
Detect additional defects after applying REMI
12
0
1
2
RC2 before REMI RC2 after REMI
The number of additional defects by modifying and enhancing test cases

13. Results for Test Development
Phase
Detect additional defects after applying REMI
13
RC3 before REMI RC3 after REMI
The number of additional defects by modifying and enhancing test cases

14. Results for Test Execution Phase
Quickly detect defects with the same number of test cases
14
Version REMI
Bug Detection Ability
Test Run
Defects
Detected
Detection
Rate
RC2
without
REMI
873 6.5 0.74%
with REMI 873 18 2.06%
RC3
without
REMI
845 8.1 0.96%
with REMI 845 9 1.07%
Test cases from the selected 30% of APIs

15. Developer Feedbacks
• “The list of risky APIs provided before conducting
QA activities is helpful for testers to allocate their
testing effort efficiently, especially with tight time
constraints.”
• “In the process of applying REMI, overheads arise
during the tool configuration and executions
(approximately 1 to 1.5 hours).”
• “It is difficult to collect the defect information to
label buggy/clean APIs without noise.”
15

16. Conclusion
• REMI
– Defect Prediction is helpful in industry
• Test Case Development
– Could identify additional defects by developing
new test cases for risky APIs.
• Test Execution Phase
– Identify defects more quickly and efficiently when
with the tight time constraints.
16