jorgemarcc/graphcodebert-interpretability

Code Similarity Visualization with GraphCodeBERT

This interactive application visualizes token-level embeddings generated by GraphCodeBERT for classical sorting algorithms. It supports pairwise comparison of algorithms based on their representation in the model’s embedding space, using PCA for dimensionality reduction.

✒️ Reference

Martinez-Gil, J. (2025). Augmenting the Interpretability of GraphCodeBERT for Code Similarity Tasks. International Journal of Software Engineering and Knowledge Engineering, 35(05), 657–678.

🚀 Features

• —Select two classical sorting algorithms.
• —Automatic tokenization and embedding via GraphCodeBERT.
• —PCA-based projection into 2D space for visualization.
• —Clear matplotlib plots showing token-level distribution differences.

🧠 Technical Overview

• —Model: `microsoft/graphcodebert-base`
• —Embedding Layer: Last hidden state
• —Reduction: Principal Component Analysis (PCA)
• —Interface: Gradio
• —Languages: Python 3.10+

🛠 Dependencies

All required libraries are listed in requirements.txt:

transformers
torch
scikit-learn
numpy
matplotlib
gradio
Pillow

🖥️ Intended Use

• —Academic teaching and demonstration of code embeddings
• —Qualitative evaluation of pretrained models for source code
• —Supplementary visualization for software engineering publications

📬 Contact

Jorge Martinez-Gil Senior Research Scientist in Computer Science