Attribution Benchmarks for Geoscience AI Models
The purpose of this repository is to develop synthetic benchmarks for evaluating XAI attribution methods. The synthetic benchmarks are designed such that a ground truth attribution can be derived. This enables quantitative evaluation of XAI methods since we can compare XAI-based attributions to the ground truth.
Mamalakis et al. (2022) proposed this benchmark framework for quantitative comparison of XAI methods. An implementation of this benchmark is provided in this repo (benchmarks/sstanom/).
Krell et al. (in progress) developed a suite of benchmarks to investigate how correlations in the input domain affect XAI-based attributions. An implementation of this benchmark is provided in this repo (benchmarks/varicov/globalcov/).
This repository contains tools for developing synthetic benchmarks for XAI evaluation.
The tools in the src directory are building blocks for developing benchmarks. For example, src/synthetic/benchmark_from_covariance.py uses a covariance to generate a set of synthetic data samples. The scripts in the pipelines directory use those tools to build benchmarks given input data (e.g. a covariance matrix).
In addition to the tools, there are specific benchmarks in the benchmarks directory. Each subfolder should be a specific benchmark application containing (1) input data, (2) application-specific code, and (3) the output benchmark files and visualizations.
Note: All code (e.g. Python and Bash scripts) should be run from repo's root.
python -m venv venv
source venv/bin/activate
pip install numpy pandas scipy matplotlib netCDF4 cmocean statsmodels captum scikit-learn pyarrow seaborn tensorflow cloudpickle imageio innvestigate lime shap pillow xarray zipp
- 'SSTANOM': SST anomaly benchmark
- Benchmark for comparing XAI methods
- See its README to run it
- See related publication for details
- 'GLOBALCOV': correlation strength benchmark
- Benchmark for comparing XAI with varying correlation strength
- See its README to run it
- Publication coming soon