IntervalSHAP

A Python project (soon to be, package) for time-dependent SHAP (SHapley Additive exPlanations) explanations in survival models. This repository provides scripts and utilities for generating synthetic survival data, running time-dependent SHAP explanations for Random Survival Forest models, and visualizing the generated explanations.

Features

• Generate synthetic survival data for benchmarking
• Compute and visualize time-dependent SHAP values
• Support for scikit-survival models
• Example scripts and figures for reproducibility

Setup

1. Environment Creation

Create a new conda environment named intervalshap with e.g. Python 3.10:

conda create -n intervalshap python=3.10
conda activate intervalshap

If you are not using Anaconda, you can use pip to install the required packages (see below).

2. Install Required Packages

With Anaconda, run the following commands in your environment:

conda install -c sebp scikit-survival>=0.22
conda install shap>=0.47
conda install matplotlib>=3.10
conda install ipython>=8.30

These packages will also install the following dependencies:

• bottleneck
• numpy
• scikit-learn
• pillow

If using pip, install the equivalent versions:

pip install "scikit-survival>=0.22" "shap==0.47" "matplotlib==3.10.0" "ipython==8.30.0"

Repository Structure

• main_script_run.py — Main script to run intervalSHAP analysis
• generate_surv_data.py — Script to generate synthetic survival data (being tested)
• environment.yml — Conda environment specification (alternative to manual setup)
• FLChain-single-event-imputed/ — Example dataset (csv files)
• saved_data_*.pkl — Example saved datasets
• draft-figures/, figures/ — Output figures and plots when draft setting is on (with smaller data)
• other-material-and-slides/ — Additional figures and slides

Usage

1. Activate the environment:

   conda activate intervalshap

2. Run the main script:

   python main_script_run.py

   This will perform the intervalSHAP analysis and generate output figures in the figures/ directory.

3. Generate synthetic data (optional):

   python generate_surv_data.py

Output

• Figures and plots are saved in the figures/ and draft-figures/ directories.
• Example datasets are provided in FLChain-single-event-imputed/.

Notes

• The code is tested with the package versions listed above. Other package and Python versions should work but are not guaranteed.
• For questions or issues, please open an issue on GitHub.