BUGLEX: Semantic–Lexical Fusion for Performance Bug Classification

This repository provides the code and framework used to evaluate machine learning models for performance bug report classification, including feature engineering (TF-IDF and embeddings) and hybrid model training.

Setup

This project uses uv for fast dependency management.

macOS / Linux (Recommended)

# Install uv (if not already installed):
curl -LsSf https://astral.sh/uv/install.sh | sh

# Sync the local virtual environment:
uv sync

# Activate the virtual environment
source .venv/bin/activate  # bash/zsh

Windows

Please see instructions at the Installing uv page.

Without uv (pip)

python3.13 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Then replace uv run python with python in any command below.

Running Experiments

Commands can be run directly using uv:

# Run the baseline model on a single dataset
uv run python -m src.run_baseline --dataset caffe

# Run the full experiment suite across all datasets
uv run python -m src.run_experiments

# Run all experiments and generate comparison plots automatically
uv run python -m src.run_experiments --with-plots

# Run experiments with a specific preprocessing mode (eg: lemmatize)
uv run python -m src.run_experiments --preprocessing-mode lemmatize 

# Run all preprocessing ablations
uv run python -m src.run_experiments --all-preprocessing

Results

Mean macro-F1 ± std across 30 stratified runs (70/30 split). Source: results/main_table_macro_f1.csv.

Dataset	Baseline NB + TF-IDF	TF-IDF + LogReg	Embedding LogReg	Hybrid LogReg
Caffe	0.623 ± 0.073	0.758 ± 0.061	0.744 ± 0.046	0.788 ± 0.063
Incubator-MXNet	0.503 ± 0.037	0.779 ± 0.031	0.832 ± 0.026	0.844 ± 0.030
Keras	0.530 ± 0.039	0.805 ± 0.034	0.845 ± 0.025	0.856 ± 0.023
PyTorch	0.529 ± 0.039	0.751 ± 0.037	0.796 ± 0.028	0.814 ± 0.027
TensorFlow	0.628 ± 0.027	0.820 ± 0.026	0.875 ± 0.016	0.882 ± 0.017

The hybrid model outperforms the baseline on all five datasets, with statistically significant improvements over the baseline.

Generating Documentation & Reports

To compile the results into the final LaTeX PDF report:

uv run python -m src.tools.build_docs

This will automatically generate the figures and tables before compiling the PDF.

Repository Layout

.
├── datasets/              # Raw data used for models
├── docs/                  # Documentation and report source files
├── main.py                # Local runner
├── pyproject.toml         # Dependencies
├── results/               # Generated results
├── src/                   # Main source code
├── README.md              # This file