Essential Content of Software Effort Estimation using Active Learning

Abstract

Do we always need to estimate software effort (SEE) using complex methods? The objective is to characterize the core elements of SEE data, that is, the minimal number of traits and examples required to fully encapsulate the data's meaning. If the amount of important information is minimal, then: 1) the content must be brief; and 2) the value added by complex learning methods must be minimal. Technique: Our QUI Does estimate software effort (SEE) necessarily require the use of sophisticated techniques? The goal is to define the essential components of SEE data, i.e., the bare minimum of characteristics and instances needed to completely capture the meaning of the data. If there is little to no important information, then: 1) the content needs to be concise; and 2) there shouldn't be any benefit from using sophisticated learning techniques. Method The CK approach first determines the Euclidean distance between the SEE data's rows (instances) and columns (features), after which it eliminates synonyms (similar features) and outliers (far instances). Finally, it evaluates the reduced data by comparing the predictions of 1) a state-of-the-art learner (CART) using all the data, and 2) a simple learner using the reduced data. Hold-out studies are used to measure performance, which is then expressed as mean and median MRE, MAR, PRED (25), MBRE, MIBRE, or MMER. Regarding eighteen datasets, QUICK reduced the training data from 69 to 96 percent (median = 89 percent). K 14 1 closest neighbour predictions performed as well in the reduced data as did CART's predictions (using complete data). In summary, certain SEE datasets provide comparatively little essential information. Complex estimation algorithms should be simplified for such datasets as they may be unduly complex. See QUICK as an illustration of a less complex SEE strategy.