Alternative methods using similarities in software effort estimation

A large variety of methods has been proposed in the literature about Software Cost Estimation, in order to increase accuracy when predicting the effort of developing new projects. Estimation by Analogy is one of the most studied techniques in this area the last 20 years. The popularity of the methodology can be explained by its accordance to human problem thinking and solving, the straightforward interpretation and the usually comparable accuracy to other methodologies. Furthermore, the methodology is essentially a special case of non-parametric regression, easily implementable and free of theoretical assumptions, based on the notion of "similarity" which is used to define "neighbors". All of these reasons led us to study the technique in more depth, considering alternative ways to exploit similarities, in order to assign weights to neighbors. In this paper, our aim is to review the existing weighting practices and explore some new iterative procedures from matrix algebra, which transform a similarity matrix to a bi-stochastic matrix (a matrix with row and column summing to 1). Specifically, we apply algorithms such as the Sinkhorn--Knopp and the Bregmanian Bi-Stochastication to similarity matrices of well-known software cost datasets in order to derive matrices that assign weights to the neighbors used for effort estimates. We investigate the sensitivity of the results with respect to the similarity function, focusing on a Gaussian kernel matrix with a tuning parameter. The promising results show that the new methods deserve a more thorough investigation and can be considered as generalization of the Estimation by Analogy method.