SRPFN: One Sequential Recommendation Model Pretrained from Synthetic Priors Predicts Multiple Datasets

Woosung Kang, Jiwon Jeong, Jonghyeok Shin, Jeongwhan Choi, Noseong Park,

Korea Advanced Institute of Science and Technology (KAIST)

---

Official implementation of SRPFN, accepted at KDD 2026.

TL;DR: A single model pretrained on synthetic priors predicts multiple sequential recommendation datasets without dataset-specific training.

---

Environment

You can create the environment as follows:

conda create -n srpfn python=3.10 pip -y
conda activate srpfn
pip install -r requirements.txt
conda install -n srpfn -c conda-forge graph-tool

All experiments were conducted on a single NVIDIA RTX A6000 GPU.

---

Training

Edit train_config.json if needed, then run:

bash shell/train.sh

shell/train.sh uses the Python executable from the srpfn conda environment by default. Training logs are written under logs/train/, and checkpoints are saved to the environment.save_path value in train_config.json.

---

Inference

To evaluate datasets in the data/ folder, specify the dataset name and evaluation protocol in eval_config.json.

Run inference with:

bash shell/eval.sh