Temporal convolutional networks for just-in-time design smells prediction using fine-grained software metrics

Abstract

Software systems quality is strongly influenced by the presence of design problems and wrong constructs introduced during software evolution. Design smells are recognized to be indicators of such problems since their presence can indicate violations of the fundamental software design principles. This explains the interest of software researchers and developers in new approaches performing design smell early detection. In this paper, the adoption of a just-in-time design smell prediction approach is proposed. The approach uses a variant of Temporal Convolutional Networks (TCN) to predict the presence of design smells before they are introduced in the model. In comparison to other studies, in the proposed approach we focus on commit-level fine-grained product metrics and process metrics (i.e., ownership and seniority) to make deep neural networks training suitable. The validation is performed on a large dataset composed of six open source systems. The results show good prediction performance for some design smells and for all the considered design smells categories. The experiment performed shows that, with respect to baseline prediction models (i.e., LSTM,CNN and RF), the proposed TCN variant trained on the identified metrics, gives better results.