Discovery of patterns in software metrics using clustering techniques

Abstract

One mechanism for estimating software quality is through the use of metrics, which are functions that evaluates certain characteristics of the product quality development. A software product can be evaluated from different points of view, and in that sense, the results of the evaluations are numeric vectors, which together describe the quality of the software. This research uses data from NASA's open access which undergo a process of reducing the dimensionality by principal component analysis (PCA), then applied three clustering techniques and evaluates the best grouping using Rand Index. Finally, the top clusters are tested with regression to find the metrics that are related to the error of the Software. The results suggest that groups consisting of software modules whose code source have a higher average of blank lines, show a higher density of error. This could be interpreted as an indication of the order of implementation. On the other hand, shows the presence of a direct relationship between the number of errors in a module with the number of calls functions to other modules. The contribution of this work is related to the use of assessment techniques of clustering, dimensionality reduction, clustering algorithms and regression to discover patterns in software metrics a rigorous manner.