احسان اسماعيلي

Code smell detection and measuring its effect on software quality attributes; has great importance because it significantly saves time, resources, and software maintenance costs. Code smells have abstract and imprecise definitions that make it difficult to detect them by defining manual rules and predetermined thresholds. Machine learning is a suitable solution for solving the challenges of code smell detection and its effect on quality attributes; Subject to the availability of sufficient and high-quality data. Current methods are based on limited and unreliable datasets; Therefore, they are imprecise, unstable, and non-generalizable. Also, a comprehensive experimental study on the relationship between code smells and quality attributes such as coverageability and testability have not been done. In this thesis, a smart hierarchical method is presented for identifying code smells and subsequently determining their effect on quality attributes. In the first step, supervised and unsupervised machine learning strategies are proposed to detect code smells, and a new combined dataset is presented to increase accuracy. Then, in the second step, a new dataset of quality attributes, consisting of Java projects, is created. Finally, the effect of code smells on each quality attribute is calculated by building and interpreting different machine learning models on the proposed dataset. eva‎luations show that models based on supervised learning strategy perform better than unsupervised models in all cases. Compared to the existing tools, the proposed models have an average of 40% higher precision and recall. The eva‎luation of code quality prediction models shows that the voting model of regressor models gives the highest accuracy in predicting quality attributes based on code smells. Interpretation of this model proves that refused bequest, class data should be private, and Swiss army knife are the most effective code smells on quality attributes.

بوي كد , كيفيت نرم افزار , سنجه هاي كد منبع , يادگيري ماشين , مجموعه داده

code smell , software quality , source code metrics , machine learning , dataset

Ehsan Esmaili

Dr. Saeed Parsa