TGB Dataset Overview

The Temporal Graph Benchmark~(TGB) aims to provide datasets and evaluation protocols for realistic, reproducible, and robust evaluation for machine learning on temporal graphs.

Edge and Node-level tasks: We include both the dynamic link property prediction task and the dynamic node property prediction task

Rich domains: TGB datasets come from interaction networks, rating networks, trasanction networks, traffic networks, social networks and trade networks.

Diverse in scale: TGB datasets includes small (< 5 million edges), medium (< 25 million edges) and large (> 25 million edges)scale datasets


Leaderboard Submission

To submit to TGB datasets, please fill in the following google form and reach out to shenyang.huang@mail.mcgill.ca if you have any questions. All results should be reported across 5 runs for both validation and test performance. Rules for the Leaderboard is found here.

Dataset splits

All datasets are split chronologically into the train, validation and test set with 70%, 15% and 15% of the edges respectively

Contributing datasets

TGB welcomes community feedback and contributions, if you would like to contribute a datasets or raise an issue, please reach out by email.