Software Metrics Data Analysis—Exploring the RelativePerformance of Some Commonly Used Modeling Techniques

  • Authors:

  • Andrew R. Gray;Stephen G. Macdonell

  • Affiliations:

  • Software Metrics Research Laboratory, Department of Information Science, University of Otago, Dunedin, New Zealand;Software Metrics Research Laboratory, Department of Information Science, University of Otago, Dunedin, New Zealand

  • Venue:
  • Empirical Software Engineering
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

Whilstsome software measurement research has been unquestionably successful,other research has struggled to enable expected advances in projectand process management. Contributing to this lack of advancementhas been the incidence of inappropriate or non-optimal applicationof various model-building procedures. This obviously raises questionsover the validity and reliability of any results obtained aswell as the conclusions that may have been drawn regarding theappropriateness of the techniques in question. In this paperwe investigate the influence of various data set characteristicsand the purpose of analysis on the effectiveness of four model-buildingtechniques—three statistical methods and one neural networkmethod. In order to illustrate the impact of data set characteristics,three separate data sets, drawn from the literature, are usedin this analysis. In terms of predictive accuracy, it is shownthat no one modeling method is best in every case. Some considerationof the characteristics of data sets should therefore occur beforeanalysis begins, so that the most appropriate modeling methodis then used. Moreover, issues other than predictive accuracymay have a significant influence on the selection of model-buildingmethods. These issues are also addressed here and a series ofguidelines for selecting among and implementing these and othermodeling techniques is discussed.