Alexander used these slides during for his presentation at BeNeVol 2011 in Brussels, Belgium. That is after he blew the fuses in the entire building.
Paper:
Serebrenik A, Vasilescu B and van den Brand M (2011), "Similar tasks, different effort: Why the same amount of functionality requires different development effort?", In Proceedings of the 10th Belgian-Netherlands Software Evolution Seminar, pp. 4-5.
4.16.24 21st Century Movements for Black Lives.pptx
Benevol 2011
1. Similar Tasks, Different Effort:
Why the Same Amount of
Functionality Requires
Different Development Effort?
Alexander Serebrenik
Bogdan Vasilescu
Mark van den Brand
2. Why do some systems require more effort?
• Empirical study
• ISBSG version 11
• largest publically available collection: 5052 projects
• 118 project attributes, including
− amount of functionality
− work effort
• Not all projects are suited for the study
• self-reporting different data quality
• different ways of measuring project attributes
/ W&I / MDSE 23-4-2012 PAGE 1
3. Project selection
ISBSG v.11 5052
Effort Staff hours (recorded) 3537
Full development lifecycle 2261
Project-specific activities only 2079
Functionality IFPUG 1661
Data quality “A” or “B” 1609
/ W&I / MDSE 23-4-2012 PAGE 2
4. Effort and Functionality Distributions
• Effort: • Adjusted FP or unadjusted FP
• skewed, outliers • Adjusted is more reliable
[Kitchenham et al. JSS, 2002]
• skewed, outliers
/W&I / MDSE 23-4-2012 PAGE 3
5. More functionality more effort required
• Log-transformation
for the skewness /
outliers problem
• Adequate
• p-value for the F-
stat ≤ 2.2*10-16,
• p-values intercept
and coefficient ≤
log(SWE) = 2.2*10-16,
2.92717 + • residuals show a
0.84617 * log(AFP) chaotic pattern
/ W&I / MDSE 23-4-2012 PAGE 4
6. Why do some systems require more effort?
• Closer look at the residuals
• technical aspects:
− primary programming language, language type,
development type, platform, and architecture
• organization type
• intended market
• year of project
• Problem of ISBSG
• missing values due to self-reporting
/ W&I / MDSE 23-4-2012 PAGE 5
7. What attributes impact the development effort?
• Goal: compare different project attributes
• ISBSG – 118 attributes
• Remove projects with missing values
• More attributes less projects
• Keep projects with missing values
• NA-category becomes too important
• We choose
• primary programming language, language type, organization
type, intended market, year of project, development type,
platform, architecture
/ W&I / MDSE 23-4-2012 PAGE 6
8. Explanation of impact
• Partition individuals in groups
• Partition = explanation [Cowell, Jenkins 1995]
• Inequality within the groups and between the groups
− Inequality indices
• Better explanation: more inequality between the groups
− Lila is better than red
− Partition refinement doesn’t deteriorate the explanation
/ SET / W&I / TU/e PAGE 7
9. Which inequality index?
• We need a decomposable index applicable to
negative values
/ W&I / MDSE 23-4-2012 PAGE 8
10. Results
Indonesia:
Project attribute Explanation %
expenditure by
educ.level 32.6% missing values
No Missing values
N = 151 N = 1609
Primary Indonesia: 25,37% 16,11%
programming
expenditure by Linux: LOC by
language province 18.9% package 17.4%
Organisation type 17,59% 18,36%
Year of the project 10,88% 5,41%
Architecture 8,68% Linux: LOC by 3,35%
Development 5,43% impl lang 5.32% 5,05%
PlatformIndonesia:
Intended Market by
expenditure 4,61% 1,57%
Linux: LOC by
Language type2.6%
gender 2,45% maintainer 4.45% 1,28%
Development Type
/ W&I / MDSE 23-4-2012 PAGE 9 0,05% 0,07%
11. Conclusions
• Three groups of attributes
• High-impact: primary programming language, organization type
• Middle-impact
− year of the project [cf. Kitchenham et al. 2002]
− architecture, development platform
• Low impact: intended market, language type, devel’t type
• A new technique for analysis of effort fp
/ W&I / MDSE 23-4-2012 PAGE 10
12. Future work
• Partition should be MECE
• “Wholesale & Retail Trade” and “Financial, Property &
Business Services”
• New aggregation/explanation techniques
• Conjecture: relative importance of attributes will be
the same for other datasets
• Models based on data from multiple companies are not
applicable when one company data is considered [Ruhe
1999]
• Both multi-company and company-specific studies are
needed
/ W&I / MDSE 23-4-2012 PAGE 11
Notas del editor
B. Kitchenham, S. L. Pfleeger, B. McColl, and S. Eagan, “An empirical study of maintenance and development estimation accuracy,” Journal of Systems and Software, vol. 64, no. 1, pp. 57–77, 2002.
> lm2 <- lm(log(SWE)~log(AFP))> summary(lm2)Call:lm(formula = log(SWE) ~ log(AFP))Residuals: Min 1Q Median 3Q Max -4.3960 -0.6584 0.0272 0.6760 3.3857 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.92717 0.09386 31.19 <2e-16 ***log(AFP) 0.84617 0.01891 44.75 <2e-16 ***---Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 1.024 on 1607 degrees of freedomMultiple R-squared: 0.5548, Adjusted R-squared: 0.5545 F-statistic: 2003 on 1 and 1607 DF, p-value: < 2.2e-16 df6$residuals <- lm2$residualsdiffEff_ineqMeasures_on_df(df6)