Using Multiple Instance Learning for Explainable Solar Flare Prediction

This github repository provides data, code and some results for the paper:

• Official version on Astronomy & Computing: https://authors.elsevier.com/sd/article/S2213-1337(22)00082-8
• Pre-print on arXiv: https://arxiv.org/abs/2203.13896

Getting the data

The dataset is available via Zenodo:

IRIS Multiple Instance Learning Dataset (9.4 Gb, Numpy npz-Format)

Running the code

The code was run on Python 3.6.9 and with the package versions listed in the requirements.txt file. To run it, adhere to the following steps:

1. Create a virtual environment and install the required packages, e.g. with virtualenv:

virtualenv -p python3 irismil_env
source irismil_venv/bin/activate
pip install -r requirements.txt

2. Run the model_runner.py script with

python model_runner.py <model_name> <parameter_value> <runs_per_fold>

e.g. to run an ibMIL model with r=3 and ten runs for each of the three CV-folds:

python model_runner.py ibMIL 3 10

Videos (10 September 2014 Flare)

Pre-flare phase only:

pf0910_sji_overplot.mp4

Whole observation:

fl0910_sji_overplot.mp4