The Use of Bayesian Networks for Web Effort Estimation: Further Investigation

The objective of this paper is to further investigate the use of Bayesian Networks (BN) for Web effort estimation when using a cross-company dataset. Four BNs were built; two automatically using the Hugin tool with two training sets; two using a structure elicited by a domain expert, with parameters obtained from automatically fitting the network to the same training sets used in the automated elicitation (hybrid models). The accuracy of all four models was measured using two validation sets, and point estimates. As a benchmark, the BN-based predictions were also compared to predictions obtained using Manual StepWise Regression (MSWR), and Case-Based Reasoning (CBR). The BN model generated using Hugin presented similar accuracy to CBR and Mean effort-based predictions. Our results suggest that Hybrid BN models can provide significantly superior prediction accuracy. However, good results also seem to depend on characteristics of the training and validation sets used.