Assessing the user-perceived quality of source code components using static analysis metrics

Abstract

Nowadays, developers tend to adopt a component-based software engineering approach, reusing own implementations and/or resorting to third-party source code. This practice is in principle cost-effective, however it may also lead to low quality software products, if the components to be reused exhibit low quality. Thus, several approaches have been developed to measure the quality of software components. Most of them, however, rely on the aid of experts for defining target quality scores and deriving metric thresholds, leading to results that are context-dependent and subjective. In this work, we build a mechanism that employs static analysis metrics extracted from GitHub projects and defines a target quality score based on repositories’ stars and forks, which indicate their adoption/acceptance by developers. Upon removing outliers with a one-class classifier, we employ Principal Feature Analysis and examine the semantics among metrics to provide an analysis on five axes for source code components (classes or packages): complexity, coupling, size, degree of inheritance, and quality of documentation. Neural networks are thus applied to estimate the final quality score given metrics from these axes. Preliminary evaluation indicates that our approach effectively estimates software quality at both class and package levels.