Using Locally Weighted Regression to Estimate the Functional Size of Software: an Empirical Study

In software engineering, measuring software functional size via the IFPUG (International Function Point Users Group) Function Point Analysis using the standard manual process can be a long and expensive activity, which is possible only when functional user requirements are known completely and in detail. To solve this problem, several early estimation methods have been proposed and have become de facto standard processes. Among these, a prominent one is High-level Function Point Analysis. Recently, the Simple Function Point method has been released by IFPUG; although it is a proper measurement method, it has a great level of convertibility to traditional Function Points and may be used as an estimation method. Both High-level Function Point Analysis and Simple Function Point skip the activities needed to weight data and transaction functions, thus enabling lightweight measurement based on coarse-grained requirements specifications. This makes the process faster and cheaper, but yields approximate measures. The accuracy of the mentioned method has been evaluated, also via large-scale empirical studies, showing that the yielded approximate measures are sufficiently accurate for practical usage. In this paper, locally weighted regression is applied to the problem outlined above. This empirical study shows that estimates obtained via locally weighted regression are more accurate than those obtained via High-level Function Point Analysis, but are not substantially better than those yielded by alternative estimation methods using linear regression. The Simple Function Point method appears to yield measures that are well correlated with those obtained via standard measurement. In conclusion, locally weighted regression appears to be effective and accurate enough for estimating software functional size.