Analysis of cluster center initialization of 2FA‐kprototypes analogy‐based software effort estimation

Abstract

Analogy‐based estimation is one of the most widely used techniques for effort prediction in software engineering. However, existing analogy‐based techniques suffer from an inability to correctly handle nonquantitative data. To deal with this limitation, a new technique called 2FA‐kprototypes was proposed and evaluated. 2FA‐kprototypes is based on the use of the fuzzy k‐prototypes clustering technique. Although fuzzy k‐prototypes algorithms are well known for their efficiency in clustering numerical and categorical data, they are sensitive to the selection of initial cluster centers. In this paper, the impact of cluster center initialization on improving the prediction accuracy of 2FA‐kprototypes was analyzed and discussed using two cluster initialization techniques: centrality‐based initialization and density‐based initialization. The performance of 2FA‐kprototypes using these two initialization techniques was evaluated and compared with that of 2FA‐kprototypes using random initialization over four datasets: ISBSG, COCOMO81, USP05‐FT, and USP05‐RQ. The results showed an improvement in the performance of 2FA‐kprototypes in terms of estimation accuracy when the all‐in method is used.

Polynomial analogy‐based software development effort estimation using combined particle swarm optimization and simulated annealing

“…There are various similarity functions in the ABE methods, for example, Euclidean (EUC) similarity function, 5,13,14, Manhattan (MHT) similarity function, 7,18,57,58,[60][61][62][63][64][65][66] maximum distance similarity function, 8,67,68 Minkowski (MKS) similarity function, 46,57 fuzzy, 69,70 and optimized induced learning (OIL). 71 The performance of the different similarity functions may vary based on the type of features (numerical, ordinal, or nominal) and the distribution of data samples in N-dimensional feature space.…”

“…7,14,20,37,40,50 In the ABE methods, the effort of a new project can be estimated based on the efforts of K similar projects utilizing an adaptation function. In the majority of the previous ABE methods, mean, 5,14,[36][37][38][39][40][41][42][43][44][45][46][47][49][50][51][52][53][54]56,57,64,68 median, 8,67 and inverse rank weighted mean 7,18,48,51,55,62,63,66 are the mostly used adaptation functions.…”

Locally weighted regression with different kernel smoothers for software effort estimation

“…The solutions obtained by global methods cannot provide a sufficiently good approximation. In contrast, Local learning methods divide the global learning problem into multiple simpler local learning models which reasonably increase predictive performance [1], [3], [8]. This can be achieved by dividing the cost function into multiple independent local cost functions.…”

“…This can be achieved by dividing the cost function into multiple independent local cost functions. In fact, the use of data locality to build prediction models has shown great interest within the research community as it helps to minimize model complexity, reduce bias and enhance accuracy [8][9][10] [11]. Since effort estimation datasets tend to be rather small and heterogeneous, the locality approaches are likely to be more adequate and produce better accuracy than models that do not use locality [12] [13].…”