A Temporal(Dynamic) GNN Training framework for recommendation systems, link prediction and self-supervised embedding learning.
Python 3.12 · Linux x86-64 (glibc 2.31+)
python -m pip install frost-gnnCreate events.csv: src is the user ID, dst is the item ID, and time is optional.
src,dst,time
42,46638,1
7,70286,2
42,38083,3
7,95034,4Train TGN and score candidate items for user 42:
frost train events.csv --src 42 --candidates 46638,70286,38083,95034Saves frost-model.pt and prints src,dst,time,score. Scores are in [0, 1]; each is
a ranking score, not a calibrated probability. Query nodes must have appeared
in training. Query time is the latest training time unless you pass --time.
Reuse the model at a chosen time:
frost predict frost-model.pt --src 42 --candidates 46638,70286 --time 10Training uses CPU; add --device cuda:0 to use a GPU.
Pass a numeric NPY matrix, or convert tabular attributes (n_id,age,category,...):
frost features nodes.csv --id-col n_id --output nodes.npy
frost train events.csv --node-features nodes.npy --output featured-model.ptNPY row n maps to node ID n, or use --feature-ids ids.json for other IDs.
The converter writes this mapping automatically.
- CSV prediction: data rules, test CSVs, feature mapping and installation options.
- Model training: configuration, support matrix and multi-GPU training.
Please cite our paper. Metadata: CITATION.cff.
@article{zhao2026frost,
author = {Zhao, Lei and Nke, Joseph and Kim, Jongryool and Lee, Chul-Ho},
title = {{FROST}: Accelerating Temporal Graph Neural Network Training via a Unified {Multi-GPU} Framework},
journal = {Proceedings of the ACM on Management of Data},
volume = {4},
number = {6},
articleno = {350},
year = {2026},
month = dec,
publisher = {Association for Computing Machinery},
doi = {10.1145/3856345},
url = {https://doi.org/10.1145/3856345}
}