Cluster-based fuzzy regression trees for software cost prediction

Abstract

The current paper proposes a novel type of decision tree, which is never used for software development cost prediction (SDCP) purposes, the cluster-based fuzzy regression tree (CFRT). This model uses the fuzzy k-means (FKM), which deals with data uncertainty and imprecision. The tree expansion is based on the variability measure by choosing the node with the highest value of granulation diversity. This paper outlined an experimental study comparing CFRT with four SDCP methods, notably linear regression, multi-layer perceptron, K-nearest-neighbors, and classification and regression trees (CART), employing eight datasets and the leave-one-out cross-validation (LOOCV). The results show that CFRT is among the best, ranked first in 3 datasets according to four accuracy measures. Also, according to the Pred(25%) values, the proposed CFRT model outperformed all the twelve compared techniques in four datasets: Albrecht, constructive cost model (COCOMO), Desharnais, and The International Software Benchmarking Standards Group (ISBSG) using LOOCV and 30-fold cross-validation technique.