Treffer: Optimizing Pre-Trained Code Embeddings With Triplet Loss for Code Smell Detection

Code embedding represents code semantics in vector form. Although code embedding-based systems have been successfully applied to various source code analysis tasks, further research is required to enhance code embedding for better code analysis capabilities, aiming to surpass the performance and functionality of static code analysis tools. In addition, standard methods for improving code embedding are essential to develop more effective embedding-based systems, similar to augmentation techniques in the image processing domain. This study aims to create a contrastive learning-based system to explore the potential of a generic method for enhancing code embedding for code classification tasks. A triplet loss-based deep learning network is designed to optimize in-class similarity and increase the distance between classes. An experimental dataset that contains code from Java, Python, and PHP programming languages and 4 different code smells is created by collecting code from open-source repositories on GitHub. We evaluate the proposed system’s effectiveness with widely used BERT, CodeBERT, and GraphCodeBERT pretrained models to create code embedding for the code classification task of code smell detection. Our findings indicate that the proposed system may offer improvements in accuracy, an average of 8% and a maximum of 13% for models. These results suggest that incorporating contrastive learning techniques into the generation process of code representation as a preprocessing step can enhance performance in code analysis.