Neural Monotone Set Functions

A framework for training and evaluating set neural models — including MasNet, ReLUMasNet, DeepSets, and SetTransformer — on benchmark tasks such as FacilityLocation and SubSet. Built on PyTorch Lightning, the framework provides reproducible training, automatic checkpointing, and optional out-of-distribution (OOD) evaluation.

Installation

git clone https://github.com/yonatansverdlov/Monotone-Embedding.git
cd <your_repo>
conda env create -f dependencies.yml
conda activate <environment_name_from_yml>

(Make sure torch, pytorch-lightning, and easydict are included in your requirements.txt.)

Create Real Datasets

MSWEB or MSNBC

cd MonotoneModel_real/data
python msweb_data_generator.py --DATASET_NAME MSWEB
python msweb_data_generator.py --DATASET_NAME MSWEB

Amazon

cd MonotoneModel_real/data
python amazon.py --dataset_name bedding/feeding

PointCloud

d MonotoneModel_real/data
python pointcloud.py --s1_size 128 & python pointcloud.py --s1_size 256 & python pointcloud.py --s1_size 128

Train and evaluate on Text datasets

cd MonotoneModel_real
./run_datasets.sh

Train and evaluate on PointCloud Datasets

cd MonotoneModel_real
./run_pointcloud.sh

Train a model on Synthetic Facility Location

cd MonotoneModel_synthetic
python main.py --task_type FacilityLocation --model_type MasNet

Train a model on Synthetic Set Containment

cd MonotoneModel_synthetic
python main.py --task_type SubSet --model_type SetTransformer

Run out-of-distribution (OOD) test

cd MonotoneModel_synthetic
python main.py --model_type ReLUMasNet --test_dist

Command-Line Arguments

Argument	Type	Default	Description
--task_type	str	FacilityLocation	Task name (FacilityLocation or SubSet)
--model_type	str	MasNet	Model architecture (ReLUMasNet, DeepSets, MasNet, SetTransformer)

Project Structure

.
├── MonotoneModel_real
├── MonotoneModel_synthetic
├── ddependencies.yaml
├── README.md

Citation

If you use this repository in your research, please cite:

@misc{sarangi2025monotoneseparablesetfunctions,
      title={Monotone and Separable Set Functions: Characterizations and Neural Models}, 
      author={Soutrik Sarangi and Yonatan Sverdlov and Nadav Dym and Abir De},
      year={2025},
      eprint={2510.23634},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.23634}, 
}

License

MIT License © 2025 Yonatan Sverdlov