SWXAI Documentation

SWXAI (Space Weather Explainable AI Toolkit) is a Python library that provides a single, consistent API for running explainable-AI methods on image, time-series, and tabular models for space-weather and heliophysics tasks. One SpaceWeatherExplainer class dispatches to the right backend (PyTorch + Captum, SHAP, LIME, AEON, Grad-CAM, …), validates inputs, and writes uniformly-named outputs (attributions.npy, summary.json, *.png).

The toolkit is developed at the Data Mining Lab, Georgia State University.

Quick links

• Installation: see Installation

• First explanation in 10 lines: see Getting started

• What method names can I pass? see Methods

• Source on Bitbucket: bitbucket.org/gsudmlab/swxai

---

Contents

Getting started

• Getting started
  • A 10-line first explanation
  • Three modalities, one API
  • When target is required
  • Where to go next
• Installation
  • Core install (editable)
  • PyTorch (image and gradient-based time-series methods)
  • Optional extras
  • Full reproducibility install
  • Sanity check

User guide

• Methods
  • Image (PyTorch)
  • Time-series
  • Tabular
  • How methods are registered
  • See also
• Images
  • Supported method strings
  • Inputs and shapes
  • Minimal example
  • Grad-CAM
  • Required extras
  • Runnable example script
• Time series
  • Supported method strings
  • Input shape
  • Model-agnostic occlusion
  • Time-Series Transformer + Integrated Gradients
  • Runnable example scripts
• Tabular
  • Supported method strings
  • Input shape
  • Minimal example - SHAP
  • ICE / Ceteris-Paribus extras
  • Runnable example scripts
  • See also
• Time-series features (AEON)
  • Available extractors
  • Quick use
  • Pipeline pattern
  • Install
  • Runnable example scripts
• Slim-TSF Time-Series Forest
  • Two equivalent entry points
  • Swapping the final classifier
  • Explaining a Slim-TSF model
  • Runnable example scripts
• Outputs and shapes
  • Input shapes
  • The Explanation dataclass
  • What Explanation.save(out_dir) writes
  • Logging to Weights & Biases
  • ICE & Ceteris-Paribus shapes
  • Image / time-series attribution shapes

Models

• Models
  • Quick chooser

Tools and reproducibility

• Demos
  • Image
  • Time-series
  • Tabular
  • See also
• Notebooks
  • Running a notebook
  • Headless execution
• Datasets
  • Geomagnetic Storms (Dst <= −100 nT)
  • HMI AR-Patches
  • SWAN-SF
  • Where they’re used
• Command-line interface
  • Arguments
  • Loading saved models
  • ICE / Ceteris-Paribus extra flags
  • Errors
• Metrics
  • Top-k masking faithfulness
  • Pattern: pair this with any explainer

Development

• Testing
  • Running the tests
  • What’s covered
  • Markers
  • Continuous integration tip
• Contributing
  • Setup
  • Adding a new explainer
  • Adding a new feature extractor
  • Adding a model
  • Code style
  • Building the docs locally
  • Running the test suite
  • Releasing

---

At a glance

Modality	Methods	Backends
Image	IG, Saliency, SmoothGrad, Occlusion, Grad-CAM/++	PyTorch + Captum, pytorch-grad-cam
Time-series	IG, Saliency, SmoothGrad, Occlusion, model-agnostic occlusion	PyTorch + Captum, scikit-learn, AEON
Tabular	SHAP (Tree/Kernel), LIME, ICE, Ceteris-Paribus	scikit-learn, SHAP, LIME

Modality	Input shape
Image	[N, C, H, W]
Time-series	[N, T, C] (or [N, T] for univariate)
Tabular	[N, F]