Automatic detection of Long Method and God Class code smells through neural source code embeddings

Code smells are structures in code that often harm its quality. Manually detecting is challenging, so researchers proposed many automatic detectors. Traditional smell detectors employ metric-based heuristics, but have recently adopted a Machine-Learning (ML) based approach. This paper compares the performance of multiple ML-based detection models against heuristics for God Class and Long Method smells. We assess effectiveness different source representations ML: we evaluate traditionally used metrics embeddings (code2vec, code2seq, CuBERT). study first to pre-trained neural best our knowledge. approach helped us leverage power transfer learning – explore whether knowledge mined from understanding can be transferred detection. A secondary contribution research systematic evaluation approaches on same large-scale, manually labeled MLCQ dataset. Almost every proposes tests this dataset unique study. Consequently, cannot directly compare reported performances derive best-performing