Recommending Source Code for Rapid Software Prototypes

Transcription

1 Recommending Source Code for Rapid Software Prototypes Collin McMillan 1 Nikki Hariri 2 Denys Poshyvanyk 1 Jane Cleland Huang 2 Bamshad Mobasher 2 1 College of William & Mary 2 DePaul University

2 You own a Start Up In two weeks, you are meeting with this venture capitalist. He might invest >$1 million in you.

3 Your Product A new web enabled device is on the market. You can build the first and only web browser for it. What do we do?

4 Requirements Design Implementation Verification Maintenance

5 Strategize Agile Design Build Test Software

6

7 Study Existing Solutions Feature 1 browser tabs Feature 2 search bar Feature 3 image downloading

8 How can we do that, too? Feature 1 browser tabs APIs? Existing classes? Suitable algorithms?

9 Step One: Build a Prototype Initial version may be buggy, slow, etc. Prototypes often thrown away Known as opportunistic programming * J. Brandt, P. Guo, J. Lewenstein, M. Dontcheva, S. Klemmer. Opportunistic Programming: Writing Code to Prototype, Ideate, and Discover. IEEE Software, Volume 26, Issue 5 (Sept/Oct 2009).

10 Types of Prototypes Horizontal: Locate all features to implement. Vertical: Implement and demonstrate select features.

11 Which Features to Implement? Recommend the Common Features

12 Extracting Features from a Software Repository Tabbed Browsing Print to PDF Search Bar Mip Map Dithering using Outlook Browse using Tabs Windows Integration Pop-up Blocker * H. Dumitru, M. Gibiec, N. Hariri, J. Cleland-Huang, B. Mobasher, C. Castro-Herrera, M. Mirakhorli, "On-demand Feature Recommendations derived from Mining Public Product Descriptions", ICSE 11.

13 Next: Finding Reusable Source Code Features Tabbed Browsing Print to PDF Search Bar Mip Map Dithering using Outlook Windows Integration Pop-up Blocker Search Engine Source Code Module A Module B Module C

14 Building Prototypes Feature 1 download manager Feature 2 search bar Feature 3 tabbed browsing * C. McMillan, M. Grechanik, D. Poshyvanyk, Q. Xie, C. Fu, "Portfolio: Finding Relevant Functions And Their Usages", ICSE 11.

15 Feature 1 download manager for web browser

16

17

18 Feature 2 web browser search bar

19

20 How do we create a prototype out of all this code?

21 Source Code of Mozilla Firefox 8.0: ~/firefox/src/browser/components/ drwxr-xr-x.. drwxr-xr-x aboutfiles drwxr-xr-x buildfiles drwxr-xr-x certerrorfiles drwxr-xr-x dirproviderfiles drwxr-xr-x downloadsfiles drwxr-xr-x feedsfiles drwxr-xr-x migrationfiles drwxr-xr-x placesfiles drwxr-xr-x preferencesfiles drwxr-xr-x privatebrowsingfiles drwxr-xr-x safebrowsingfiles drwxr-xr-x searchfiles drwxr-xr-x sessionstorefiles drwxr-xr-x shellfiles drwxr-xr-x sidebarfiles drwxr-xr-x tabviewfiles drwxr-xr-x testfiles drwxr-xr-x thumbnailsfiles Feature 1 download manager Feature 2 search bar Feature 3 tabbed browsing

22 browser/components/downloads/src/downloadscommon.jsm Highly coupled to other code.

23 Two Competing Goals Domain Relevance Low Coupling

24

25 Our Solution: Set Coverage of Features Recommendations Tabbed Browsing Search Bar Download Manager Set Cover Algorithm Module A Module B Tabbed Browsing Search Bar Windows Integration Print to PDF Download Manager Modules Repositories Problem: Computing the Set Cover is NP complete

26 Set Coverage: Our Approximation Greedy Approximation to Set Coverage Choose the module that is least expensive to add Based on Coupling Between Objects (CBO) Module A is less coupled Module than A or Module Module B? B. Module B Module D Module A Module C

27 A B Use PageRank to Estimate Cost 0.3 C of Adding the Modules. D E

28 A B C 0.3 D E

29 0.26 A B 0.12 Hardest to reuse 0.23 C D 0.07 Easiest to reuse E 0.32

30 A B C 0.3 D E

31 A B C: F 3 D: F E: F 1, F 2

32 A B C: F 3 D: F E: F 1, F 2

33 A B C: F 3 D: F E: F 1, F 2

34 A B C: F 3 D: F E: F 1, F 2

35 A B D CE: F 1, F 2, F 3 0.2

36 A B D CE: F 1, F 2, F 3 0.5

37 0.38 A B D CE: F 1, F 2, F Continue grouping nodes until all features are covered

38 Putting it all together Module Extractor Modules Dependency Extractor Module Dependencies Software Repositories Source Code Search Engine Module x Feature Matrix Module Recommender Module Recommendations Feature Extractor Features Feature Recommender Feature Recommendations Product Description

39 Software Archive Modules from Sourceforge / Merobase 1 241,655 Java packages 13,701 Java projects 440 Million Total Lines of Code Features from Softpedia 1,135 Features 493,347 Feature Descriptions 1 O. Hummel, W. Janjic, C. Atkinson, Code Conjurer: Pulling Reusable Software out of Thin Air", IEEE Software, 2008.

40 Evaluation User Study To compare Prefab with Portfolio, a state ofthe art source code search engine. 31 Java Programmers participated Students at William & Mary 14 had professional experience Average 4.8 years of programming knowledge

41 Remember our Goals Domain Relevance Low Coupling

42 State-of-the-art Search Engines Domain Relevance Low Coupling

43 State-of-the-art Search Engines Feature 1 download manager Feature 2 search bar Feature 3 tabbed browsing What do we type here?

44 Step 1 Enter a description of the product.

45 Step 2 Select features relevant to the query.

46 Step 3 Determine which features are implemented in each module. Participants evaluate which packages implement which features.

47 Cross Validation Design Experiment Group Approach Task Set 1 A B Portfolio Prefab T1 T2 2 A B Prefab Portfolio T3 T4

48 Features Covered Per Package p F 13.4 F crit 3.85 Portfolio Prefab

49 Time Per Query (in Minutes) p F 46.5 F crit 3.90 Portfolio Prefab

50 Conclusions Recommend source code modules that implement a set of features Combines two state of the art approaches Evaluated in a user study with 31 programmers All data online:

51 Prefab Portfolio Features Implemented Per Result Percent of Total Results