SurvSHAP-IQ [ICML 2026]

This repository is a code supplement to the following paper:

S. H. Langbein, H. Baniecki, F. Fumagalli, N. Koenen, M. N. Wright, J. Herbinger. Functional Decomposition and Shapley Interactions for Interpreting Survival Models. ICML 2026

TL;DR: We introduce a principled framework based on functional decomposition and Shapley values to explain time-dependent feature interactions in machine learning survival models.

Section 4.1: Experiments with simulated data

Create environment with the modified shapiq package (based on version 1.3.1).

conda env create -f env.yml

Simulate data with simulation_experiments/simulate_simsurv_final.r and simulation_experiments/simulate_simsurv_corr.r (data in simulation_experiments/data).

Run experiments on randomly selected observation and plot results with simulation_experiments/sim_survshapiq.py, simulation_experiments/sim_survshapiq_corr.py and simulation_experiments/sim_survshapiq_marg_cond.py.

Obtain full explanations for simualted datasets with simulation_experiments/sim_survshapiq_global.py (results in simulation_experiments/experiments) and compute local accuracy results with simulation_experiments/local_accuracy.py.

Section 4.2: Experiments with real world data

Create environment with the modified shapiq package (based on version 1.3.1).

conda env create -f env.yml

Explain the models with experiments/explain_actg.ipynb and experiments/explain_uvealmelanoma.ipynb.

Run approximator benchmark with sbatch experiments/run_approximators_benchmark.sh (see experiments/run_approximators_benchmark.py).

Plot the benchmark results with experiments/plot_approximators_benchmark.ipynb.

Section 4.2: Multi-modal TCGA-BRCA

See the experiments_tcga-brca directory for code and instructions to reproduce the experiments on the TCGA-BRCA dataset.