Predicting Test Smell Categories based on Issue Descriptions using Information Retrieval and Machine Learning Techniques

Effective software testing is essential for achieving high-quality software products. Various strategies and methods are used to ensure the software meets requirements and functionality, but test smell is one of the challenges to achieving effective software testing. These smells contain various issues, such as execution behavior, semantic inconsistencies, dependency management, inefficiencies, and code-related problems. This study investigates applying information retrieval (IR) and machine learning techniques to predict test smell categories by analyzing natural language issue descriptions and shifting from code-based to issue-based approaches. The result reveals that preprocessing techniques impact model performance. Techniques such as Spacy demonstrated better performance due to their sophisticated language models, while TF-IDF proved effective in distinguishing semantic inconsistencies in textual issue descriptions. Oversampling appeared as a crucial method for handling data imbalance and enhancing model performance across all test smell categories. However, N-gram requires further investigation to capture the relationship in issue descriptions. These findings show the possibility of combining natural language processing with machine learning to predict and address test smell categories, contributing to more reliable software testing practices and reducing the mistakes that will happen. This study can potentially explore additional preprocessing strategies, oversampling techniques, and machine learning models to optimize predictive performance.