Generalized FastDBM

Research software accompanying the paper Computing Fast and Accurate Maps for Explaining Classification Models.

Field	Details
Paper	Computing Fast and Accurate Maps for Explaining Classification Models
Publication	Computers & Graphics, 2025
Authors	Yu Wang, Cristian Grosu, Alexandru Telea
Related paper	Extended version of our EuroVA paper

What This Repository Contains

This repository implements the generalized FastDBM workflow for constructing decision maps that explain classifier behavior over a 2D projection. The code supports experiments that compare map-building strategies, evaluate runtime and accuracy trade-offs, and reproduce the examples shown in the paper.

The work is relevant to explainable AI, visual analytics, model inspection, high-dimensional data analysis, and reproducible research software.

Repository Contents

• demo.ipynb: notebook entry point for checking the examples and paper workflow.
• mapbuilder/: decision-map construction, classifier wrappers, and neighborhood-based map-building utilities.
• invprojection/: inverse projection methods used by the map-building workflow.
• expiriments/: experiment scripts for threshold search, timing, and distance/gradient comparisons.
• requirements.txt: tested Python dependencies for the research workflow.
• illustration.png: overview figure used in the README.

Environment

The code was tested with Python >= 3.10 and < 3.12. Some experiments require a CUDA-capable GPU with CUDA >= 12.1.

Create and activate a virtual environment, then install the dependencies:

python -m pip install -r requirements.txt

Reproducing The Demo

Run the notebook from the repository root:

jupyter notebook demo.ipynb

The notebook demonstrates the generalized FastDBM workflow and should be the first place to check the implementation. The scripts under expiriments/ are kept as research experiment drivers for the timing and accuracy studies used while developing the method.

Notes For Reuse

• Run notebooks and scripts from the repository root so relative paths resolve correctly.
• GPU availability, CUDA version, and library versions can affect runtime measurements.
• The repository is organized as paper-supporting research software rather than a packaged Python library.
• The expiriments/ directory name is retained for compatibility with the existing project layout.

Citation

GitHub can read the repository citation metadata from CITATION.cff. If you use this implementation, please cite:

@misc{softwareGfastDBM,
	title = {Generalized {FastDBM} implementation source code},
	url = {https://github.com/yuwang-vis/generalized_fastDBM},
	author = {Wang, Yu and Grosu, Cristian and Telea, Alexandru},
	year = {2025},
}