Empirical evaluation of similarity-based missing data imputation for effort estimation

Multivariate regression models have been commonly used to estimate the software development effort to assist project planning and/or management. Since project data sets for model construction often contain missing values, we need to build a complete data set that has no missing values either by using imputation methods or by removing projects and metrics having missing values (removing method). However, while there are several ways to build the complete data set, it is unclear which method is the most suitable for the project data set. In this paper, using project data of 706 cases (47% missing value rate) collected from several companies, we applied four imputation methods (mean imputation, pair-wise deletion, k-nn method and applied CF method) and the removing method to build regression models. Then, using project data of 143 cases (having no missing values), we evaluated the estimation performance of models after applying each imputation and removing method. The result showed that the similarity-based imputation methods (k-nn method and applied CF method) showed the best performance.