The potential benefit of relevance vector machine to software effort estimation

Three key challenges faced by the task of software effort estimation (SEE) when using predictive models are: (1) in order to support decision-making, software managers should have access not only to the effort estimation given by the predictive model, but also how confident this model is in estimating a given project and how likely other effort values could be the real efforts required to develop this project, (2) SEE data is likely to contain noise, due to the participation of humans in the data collection, and this noise can hinder predictions if not catered, and (3) data collection is an expensive task, and guidelines on when new data need to be collected would be helpful for reducing the cost associated with data collection. However, even though SEE has been studied for decades and many predictors have been proposed, few methods focus on these issues. In this work, we show that relevance vector machine (RVM) is a promising predictive method for addressing these three challenges. More specifically, it explicitly handles noise, it provides probabilistic predictions of effort, and can be used to identify when the required efforts of new projects should be collected for using them as training examples. With that in mind, this work provides the first step in exploiting RVM's potential for SEE by validating both its point prediction and prediction intervals. It then explains in detail future directions in terms of how RVMs can be further exploited for addressing the above mentioned challenges. Our systematic experiments show that RVM is very competitive compared with state-of-the-art SEE approaches, being usually ranked the first or second in 7 across 11 data sets in terms of mean absolute error. We also demonstrate how RVM can be used to judge the amount of noise present in the data. In summary, we show that RVM is a very promising predictor for SEE and should be further exploited.