Abstract
2. RELEVANCY FILTERING Collecting the data required for quality prediction within a development team is time-consuming and expensive. An alternative to make predictions using data that crosses from other projects or even other companies. We show that with/without relevancy filtering, imported data performs the same/worse (respectively) than using local data. Therefore, we recommend the use of relevancy filtering whenever generating estimates using data from another project. Our relevancy filter extends standard analogy-based estimation methods (which we call ABE0). ABE0 generates estimates for a test project by gathering evidence from the effort values of similar projects in some training set. By analyzing the previous research of experts like Shepperd et. al. [9], Mendes et. al. [8] and Li et. al. [7] on the field of analogy-based estimation, we can come up with a baseline technique: Categories and Subject Descriptors: D.2.9 [Software Engineering]: Effort Estimation Build a training data from rows of past projects; The columns of this set are composed of independent variables (the features that define projects) and a dependent variable (the recorded effort value). Decide on how many similar projects (analogies) to use when examining a new test instance , i.e. k-values. For each test instance, select k analogies from training. General Terms: Economics, Measurement Keywords: Effort estimation, data mining, cross, within 1. INTRODUCTION When data is scarce within one project, it is tempting to use data imported from other projects. Such cross-project data exist; for example the PROMISE repository [2] offers a dozen effort estimation data sets for public access. A recent survey paper has evaluated within or cross data for effort estimation [5]. They concluded that they could not make a conclusion; that the current findings are contradictory about the relative merits of within or cross data. In other work [10], we have shown that it is acceptable to use cross data sources for defect prediction, providing that data has been pre-processed by some sort of relevancy filtering. Given a large training set, such relevancy filters select a small subset relevant to t