Penelusuran Trace Link Issue-Commit Menggunakan CodeBERT Dan GraphCodeBERT Dan Interpretasi XAI Global

Sadipta, Rafhi (2026) Penelusuran Trace Link Issue-Commit Menggunakan CodeBERT Dan GraphCodeBERT Dan Interpretasi XAI Global. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221014-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Penelusuran trace link antara issue dan commit penting untuk menjaga konsistensi artefak perangkat lunak. Penelitian sebelumnya telah berhasil melakukan penelusuran trace link dengan memanfaatkan transfer learning berbasis model transformer yang diintegrasikan dengan XAI, namun model yang digunakan belum spesifik dirancang untuk memahami representasi kode, dan interpretasi yang dihasilkan hanya bersifat lokal. Tugas akhir ini membangun model penelusuran trace link menggunakan CodeBERT dan GraphCodeBERT yang di fine-tuning pada delapan skenario. Dataset yang digunakan merupakan gabungan dari LinkFormer, 20-MAD, dan CariKado serta OSS. Hasil pengujian menunjukkan performa terbaik dicapai pada konfigurasi fitur lengkap dengan random split, dengan F1-score sebesar 0,971649, sedangkan performa terendah terjadi pada skenario ablasi dengan temporal split, dengan F1-score 0,6015 dan MCC 0,0607. Interpretasi model dilakukan secara global menggunakan empat metode Explainable Artificial Intelligence, yaitu PDP, Global SHAP, LRP, dan H-XAI. Hasil interpretasi menunjukkan bahwa komponen issue, khususnya summary dan description, memberikan kontribusi terbesar terhadap keputusan klasifikasi, diikuti komponen commit, sedangkan komponen diff secara konsisten memberikan kontribusi terkecil meskipun mendapat alokasi token terbesar. Didapatkan juga efek kausal fitur token_overlap terhadap probabilitas prediksi pada config 1 bersifat konsisten dan tidak terdistorsi signifikan oleh diff_len.
=================================================================================================================================
Tracing traceability links between issues and commits is essential for maintaining software artifact consistency. Previous research has successfully performed trace link recovery by leveraging transformer-based transfer learning integrated with explainable artificial intelligence (XAI), however, the models used were not specifically designed to understand code representations, and the resulting interpretations were only local in nature. This study builds a trace link recovery model using CodeBERT and GraphCodeBERT, fine-tuned across eight scenarios. The dataset used is a combination of LinkFormer, 20-MAD, and CariKado, as well as OSS. The test results show that the best performance was achieved with the full feature configuration under random splitting, reaching an F1-score of 0,971649, while the lowest performance occurred in the ablation scenario under temporal splitting, with an F1-score of 0.6015 and an MCC of 0.0607. Model interpretation was conducted globally using four Explainable Artificial Intelligence methods, namely PDP, Global SHAP, LRP, and H-XAI. The interpretation results show that the issue component, particularly the summary and description sub-components, contributes the most to the classification decision, followed by the commit component, while the diff component consistently contributes the least despite receiving the largest token allocation. It was also found that the causal effect of the token_overlap feature on the prediction probability in config 1 is consistent and not significantly distorted by diff_len.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CodeBERT, Explainable Artificial Intelligence, GraphCodeBERT, Issue-Commit, Traceability CodeBERT, Explainable Artificial Intelligence, GraphCodeBERT, Issue-Commit, Traceability
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > T Technology (General) > T11 Technical writing. Scientific Writing T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Rafhi Sadipta
Date Deposited:	24 Jul 2026 22:37
Last Modified:	24 Jul 2026 22:37
URI:	http://repository.its.ac.id/id/eprint/137331

Actions (login required)

View Item