Federated Survival Forests

This repo contains the source code of the experiments presented in the paper "Federated Survival Forests" [1]. Federated Survival Forest (FedSurF) is a novel federated learning algorithm for survival analysis. It relies on Random Survival Forests (RSF) [2] to build a federated ensemble of survival trees. Tree selection occurs randomly with a probability proportional to the local performance of each tree. The algorithm is implemented using a custom version of scikit-survival that supports the averaging of survival functions. Baseline methods relying on neural networks are implemented using PyCox.

Experiments are run on standard datasets for survival analysis with federated splits induced by the algorithms described in [3], which allow testing on heterogeneous federations with different label splits for each client.

For more information, please take a look at

• Federated Survival Forests [1] for the FedSurF algorithm, code, experiments, and results description.
• Heterogeneous Datasets for Federated Survival Analysis Simulation [3] for the federated dataset generation algorithms.

🛠️ Usage

Clone the repository:

git clone https://github.com/archettialberto/federated_survival_forests

Run docker compose to start the experiments:

docker compose run fedsurf-experiments

The results will be saved in the logs directory. Please note that docker compose mounts the current directory in the /exp directory of the container, so you can edit the code in exps.py and run without rebuilding the image.

If you want to build a new Docker image, run

docker build -t aarchetti/fedsurf:latest .

⚙️ Installation Outside Docker (Not Recommended)

• Install Python 3.10
• Clone the repo with git clone https://github.com/archettialberto/federated_survival_forests
• Move to the repo directory with cd federated_survival_forests
• Create a new virtual environment with python -m venv venv
• Activate the environment with source venv/bin/activate
• Install the dependencies with pip install -r requirements.txt
• Clone the helper repo with git clone https://github.com/archettialberto/scikit-survival
• Move to the helper repo directory with cd scikit-survival
• Update git submodules with git submodule update --init
• Move back to the federated_survival_forests directory with cd ..
• Install scikit-survival with pip install ./scikit-survival
• At this point, you can run python exps.py without Docker

📕 Bibtex Citation

@inproceedings{archetti2023federated,
    author={Archetti, Alberto and Matteucci, Matteo},
    booktitle={2023 International Joint Conference on Neural Networks (IJCNN)}, 
    title={Federated Survival Forests}, 
    year={2023},
    pages={1-9},
    doi={10.1109/IJCNN54540.2023.10190999}
}

📚 References

[1] A. Archetti and M. Matteucci, “Federated Survival Forests.” In press at 2023 International Joint Conference on Neural Networks (IJCNN2023), Feb. 2023.

[2] H. Ishwaran, U. B. Kogalur, E. H. Blackstone, and M. S. Lauer, “Random survival forests,” Annals of Applied Statistics, vol. 2, no. 3, Sep. 2008, doi: 10.1214/08-AOAS169.

[3] A. Archetti, E. Lomurno, F. Lattari, A. Martin, and M. Matteucci, “Heterogeneous Datasets for Federated Survival Analysis Simulation,” in Companion of the 2023 ACM/SPEC International Conference on Performance Engineering, Coimbra Portugal: ACM, Apr. 2023, pp. 173–180. doi: 10.1145/3578245.3584935.