New Conceptual Cohesion Metrics: Assessment for Software Defect Prediction

Abstract

High cohesion and low coupling is a desirable software design principle as it significantly and positively impacts software quality. While the majority of existing cohesion metrics are structural, this paper proposes a set of new object-oriented metrics capturing the conceptual cohesion of classes. They are derived from the semantic information contained in the source code and involve using Doc2Vec, an artificial neural network based unsupervised model. Two case studies that compare the proposed metrics with an extensive set of relevant existing ones are presented. The empirical validation has been performed on three open source software systems. It includes the assessment of the proposed metrics versus preexisting ones from the perspective of their utility for software defect prediction. The experimental results confirm that the metrics we propose capture additional aspects of cohesion when compared to existing cohesion metrics, while also providing better software defect prediction performance than preexisting related conceptual cohesion metrics.