Multilevel Readability Interpretation Against Software Properties: A Data-Centric Approach

Given the wide adoption of the agile software development paradigm, where efficient collaboration as well as effective maintenance are of utmost importance, the need to produce readable source code is evident. To that end, several research efforts aspire to assess the extent to which a software component is readable. Several metrics and evaluation criteria have been proposed; however, they are mostly empirical or rely on experts who are responsible for determining the ground truth and/or set custom thresholds, leading to results that are context-dependent and subjective. In this work, we employ a large set of static analysis metrics along with various coding violations towards interpreting readability as perceived by developers. Unlike already existing approaches, we refrain from using experts and we provide a fully automated and extendible methodology built upon data residing in online code hosting facilities. We perform static analysis at two levels (method and class) and construct a benchmark dataset that includes more than one million methods and classes covering diverse development scenarios. After performing clustering based on source code size, we employ Support Vector Regression in order to interpret the extent to which a software component is readable against the source code properties: cohesion, inheritance, complexity, coupling, and documentation. The evaluation of our methodology indicates that our models effectively interpret readability as perceived by developers against the above mentioned source code properties.