170 likes | 292 Views
Using Correlation and Accuracy for Identifying Good Estimators. Gary D. Boetticher Nazim Lokhandwala Univ. of Houston - Clear Lake, Houston, TX, USA boetticher@uhcl.edu Lokhandwala@uhcl.edu. 61. 62. 63. http://nas.cl.uh.edu/boetticher/publications.html.
E N D
Using Correlation and Accuracy for Identifying Good Estimators Gary D. Boetticher Nazim Lokhandwala Univ. of Houston - Clear Lake, Houston, TX, USA boetticher@uhcl.eduLokhandwala@uhcl.edu 61 62 63 http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
82% Human 12% Human 20% ML 68% Algorithm 18% Formal Research vs. Reality according to Jörgensen JSS ’04: Compendium of expert estimation studies TSE ’07: 300+ software est. papers, 76 journals, 15+ Years http://nas.cl.uh.edu/boetticher/publications.html The 3rd International Predictor Models in Software Engineering (PROMISE) Workshop
((Log (TechGradCourses + (TechGradCourses ^ ((Log TotWShops)/(Cos (TechGradCourses ^ ((ProcIndExp + (Cos (TechGradCourses ^ ((ProcIndExp + (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Cos (Log (Log (Log SWProjEstExp))))))))))))) / (TechGradCourses ^ (Log SWProjEstExp)))))) / (((Cos (TechGradCourses ^ ((ProcIndExp + (Cos (TechGradCourses ^ ((ProcIndExp + (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Cos (TechGradCourses ^ ((ProcIndExp + (((ProcIndExp + (Log (Sin MgmtGradCourses)))/(Sin SWPMExp)) + (Sin ((Cos (TechGradCourses ^ ((ProcIndExp + (Cos (TechGradCourses ^ ((ProcIndExp + (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Sin SWPMExp)))))))))) / (TechGradCourses ^ (Log SWProjEstExp)))))) / (((Cos (TechGradCourses ^ ((Log SWProjEstExp) / (((Log (ProcIndExp + (Log (TechGradCourses ^ ((Log SWProjEstExp) / (Log SWProjEstExp)))))) - 3) / (ProcIndExp + (TechGradCourses ^ (Cos (TechGradCourses ^ ((ProcIndExp + (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Cos ((((Log SWProjEstExp) / ((ProcIndExp + (Log (TechGradCourses ^ (TechGradCourses ^ (Log SWProjEstExp))))) / (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Cos (Log (Log (Log SWProjEstExp)))))))))))))) / (Sin SWPMExp)) / (Sin SWPMExp)))))))))))) / (TechGradCourses ^ (Log SWProjEstExp))))))))))) - 3) / (TechGradCourses ^ (Log SWProjEstExp)))))) + ((Log SWProjEstExp) / (Log SWProjEstExp)))))) / (Log (Log (Log (TechGradCourses + (Cos (Log (Log (TechGradCourses ^ (Cos (((((Log SWProjEstExp) / (TechGradCourses ^ (Log SWProjEstExp))) / ((ProcIndExp + (Log (Sin MgmtGradCourses))) / ((Log SWProjEstExp) / (Log SWProjEstExp)))) / (Sin SWPMExp)) / (Sin SWPMExp))))))))))))))))))))))) / (TechGradCourses ^ (Log SWProjEstExp)))))) / (((Log ((((Log TotLangExp) / (Log SWProjEstExp)) / (Log SWProjEstExp)) / (Sin SWPMExp))) - 3) / (TechGradCourses ^ (Log SWProjEstExp)))))) - 3) / (TechGradCourses ^ (Log SWProjEstExp)))))))))) + (((((ProcIndExp + (Log (TechGradCourses ^ (Log (TechGradCourses + ((TechGradCourses ^ (TechGradCourses ^ (Cos (TechGradCourses ^ ((ProcIndExp + (Log (Log (TechGradCourses ^ (TechGradCourses ^ (Cos (Log (Log (TechGradCourses ^ (Cos ((((Log SWProjEstExp) / ((ProcIndExp + (Log (TechGradCourses ^ (Log (TechGradCourses + (Cos (Log (Log (TechGradCourses ^ (Cos (((((Log SWProjEstExp) / (TechGradCourses ^ (Log SWProjEstExp))) / ((ProcIndExp + (Log (Sin MgmtGradCourses))) / ((Log SWProjEstExp) / (Log SWProjEstExp)))) / (Sin SWPMExp)) / (Sin SWPMExp)))))))))))) / ((Log SWProjEstExp) / (Log SWProjEstExp)))) / (Sin SWPMExp)) / (Sin SWPMExp)))))))))))) / (TechGradCourses ^ (Log SWProjEstExp))))))) / (Sin SWPMExp))))))) / (TechGradCourses ^ (Log SWProjEstExp))) / (TechGradCourses ^ (Log SWProjEstExp))) / (TechGradCourses ^ (Log SWProjEstExp))) / (Sin SWPMExp))) Statement of Problem Some Background 2006 http://www.starwarscrawl.com/?id=232 http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
TechUGCourses < 45.5 | Hardware Proj Mgmt Exp < 6 | | No Of Hardware Proj Estimated < 4.5 | | | No Of Hardware Proj Estimated < 3 | | | | TechUGCourses < 23 | | | | | Hardware Proj Mgmt Exp < 0.75 | | | | | | TechUGCourses < 18 | | | | | | | Hardware Proj Mgmt Exp < 0.13 | | | | | | | | TechUGCourses < 0.5 | | | | | | | | | TechUGCourses < -1 : F (1/0) | | | | | | | | | TechUGCourses >= -1 | | | | | | | | | | Degree < 3.5 : A (4/0) | | | | | | | | | | Degree >= 3.5 : A (5/2) | | | | | | | | TechUGCourses >= 0.5 | | | | | | | | | TechUGCourses < 5.5 | | | | | | | | | | Degree < 3.5 : F (5/0) | | | | | | | | | | Degree >= 3.5 | | | | | | | | | | | TechUGCrses < 2 : A (1/0) | | | | | | | | | | | TechUGCrses >= 2 : F (1/0) | | | | | | | | | TechUGCrses >= 5.5 | | | | | | | | | | Degree < 3.5 | | | | | | | | | | | TechUGCrs < 10.5 : A (3/0) | | | | | | | | | | | TechUGCrses >= 10.5 | | | | | | | | | | | | TechUGCrs<12.5 : F (3/0) | | | | | | | | | | | | TechUGCrses >= 12.5 | | | | | | | | | | | | | TechUGCrs<16: A (2/0) | | | | | | | | | | | | | TechUGCrs>15 : A (2/1) | | | | | | | | | | Degree >= 3.5 : F (1/0) | | | | | | | HardProjMgmt Exp >= 0.13 : A (2/0) | | | | | | TechUGCourses >= 18 : A (2/0) | | | | | Hard Proj Mgmt Exp >= 0.75 : F (1/0) | | | | TechUGCourses >= 23 : F (5/0) | | | No Of Hardware Proj Est >= 3 : F (1/0) | | No Of Hardware Proj Est >= 4.5 : A (5/0) | Hardware Proj Mgmt Exp >= 6 : F (4/0) TechUGCrses >= 45.5 : A (2/0) Some Background2007 Statement of Problem How to build human-based estimation models that are accurate, intuitive, and easy to understand? http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
PROMISE 2008 versus 2007 • Sample set: 178 Samples • One learner Accuracy and Intuitive Results • Attribute reduction Analysis. • Relatively Simple models. http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Supplier Software Buyer Software Distribution Server Supplier1 Buyer Admin Supplier2 ... Buyer1 Buyern : Suppliern The Approach • Personal Demographics • Age, Gender, Nationality, etc. • Academic • Courses Undergrad/Grad: • CS, HW, SE, Proj. Mgmt, MIS • Workshops/Conferences: • CS, HW, SE, Proj. Mgmt, MIS • Work • Programming:Ada, ASP, Assembly, C, C++, • COBOL, DBMS, FORTRAN, Java, PASCAL, • Perl, PHP, SAP, TCL, VB, Other • Work Experience (HW/SW) • Project Management Exp. (HW/SW) • # Projects Estimated (HW/SW) • Average Project Size • Domain Experience • Procurement Industry Experience Estimate 28 Components Scale Factor And Correlation Apply Machine Learners http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
User’s Estimates How user compares to other respondents Actual Estimates Feedback to Users http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
S c a l e S c a l e Correlation Correlation S c a l e S c a l e Correlation Correlation Experiments: Data Original Data set Experiment 1 82.8 -29.4 Experiment 2 Experiment 3 29X 0.008 http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Experiments: Tools, Configuration • Outliers Removed • WEKA Toolset • C4.5 (J48) • 1000 Trials • 10-Fold Cross Validation http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Results: Correlation Only 2-Class Problem: 10 Best (A), 10 Worst (F) 1000 Trials, Accuracy=41.6% Attribute Reduction using WRAPPER 1000 Trials, Accuracy=78.6% http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Results: Scale Factor Only 2-Class Problem: 10 Best (A), 10 Worst (F) 1000 Trials, Accuracy=65.0% Attribute Reduction using WRAPPER 1000 Trials, Accuracy=78.2% http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Results: Correlation & Scale Factor 2-Class Problem: 10 Best (A), 10 Worst (F) 1000 Trials, Accuracy=82.2% Attribute Reduction using WRAPPER 1000 Trials, Accuracy=93.3% http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Discussion - 1 How well does the decision tree from the third experiment apply to all the respondents minus outliers? http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Discussion - 2 Challenges in component effort estimation • Scope of effort • Amortization of effort • Reuse can skew estimates (esp. Design for Reuse) • Respondent’s estimates = Boetticher’s estimates http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Conclusions • Good accuracy rates, especially after attribute reduction • Correlation + Scale Factor Intuitive Model • Bridges expert and model groups http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
Thank You! http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop
References • Jorgensen, M., “A review of studies on Expert Estimation of Software Development Effort,” Journal of Systems and Software, 2004. • Jørgensen, Shepperd, A Systematic Review of Software Development Cost Estimation Studies, IEEE Transactions on Software Engineering, 33, 1, January, 2007, Pp. 33-53. http://nas.cl.uh.edu/boetticher/publications.html The 4th International Predictor Models in Software Engineering (PROMISE) Workshop