Feature Envy Detection with Deep Learning and Snapshot Ensemble

Abstract

Code Smell is a code symptom of deep-seated quality problems caused by design defects or improper coding habits in software. It may not directly affect the operation of program, but it affects the readability, understandability and maintainability of code. Therefore, the identification, location and reconstruction of code smell are becoming increasingly significant. Combined with the deep learning methods, based on the existing model for detecting Feature Envy, this paper optimizes it from three aspects: introducing the attention mechanism, expanding and modifying the model structure, and applying the snapshot ensemble. The experimental results show that compared with the standard results, the model proposed in this paper gets a better performance on four evaluation metrics: precision, recall, F1 measure and AUC. Based on the research results of this experiment, we can see the effectiveness of deep learning in the field of code smell detection and the prospect of theories in natural language processing to be utilized in code smell detection, which provides a practical cornerstone for the research of deep learning based smell detection methods in the future.