A k-Nearest Neighbors approach for COCOMO calibration
Phu Le, Vu Nguyen · 2017
Model-based software estimation uses algorithms and past project data to make predictions for new projects. This paper presents a comparative assessment of four modeling approaches, including the original COCOMO, COCOMO calibration, k-Nearest Neighbors, and a combination of COCOMO calibration and k-Nearest Neighbors. Our results indicate that using kNN to select the nearest projects and calibrating them produce favorable estimation accuracy. The best estimates can be produced using a small number of nearest neighboring projects to calibrate. We recommend that organizations use between 8 and 10 nearest projects to calibrate the COCOMO model.