A dynamic network based on 100 years of Division I FBS American college football games (1924-2024), with evolving ground-truth community memberships captured in terms of conference realignment.
Head over to this Colab notebook and make a copy. There, the data may be filtered according to start/end year, snapshot aggregation length, etc. The notebook will output a tailored dataset accordingly, together with network statistics, a time-varying adjacency matrix, and a readme. The notebook should also make it clear how to custom-filter the data if desired.
Alternatively, the raw data queried by the Colab notebook may be found in the file all_cfb_games_1924_2024.csv. The script that was used to generate the raw data via the collegefootballdata.com API is main_csv_generation.py.
- Data from 2020 is inconsisent due to Covid-19.
- Likewise for 1939-1945 due to WWII. Indeed, expect to exercise general care when handling data pre-1970.
@inproceedings{humeSpectral2024,
author = {Hume, Jacob and Balzano, Laura},
title = {A Spectral Framework for Tracking Communities in Evolving Networks},
booktitle = {Proceedings of the Third Learning on Graphs Conference},
year = {2024},
organization = {PLMR},
}
College American football is the most-attended sport in the United States. It is also a favorite sport of network scientists, having given rise to the popular dataset in [1] which builds a network by placing edges between teams who faced one another during the 2000 season.
The majority of college football teams belong to a specific conference. Roughly speaking, a given team will play ~2/3 of their games against opposition from within their conference and 1/3 of their games out-of-conference. Conference schedules are not all created equal: historically, teams within closer geographic proximity were more likely to play one another. Sometimes this was encoded in the form of subdivisions (e.g. from 2014-2023 'B1G' contained 'B1G East' and 'B1G West'), yielding hierarchical network structure. Most matchups do not occur every season, but specific rivalries (e.g. Michigan-Ohio State, Clemson-South Carolina) will, including inter-conference ones.
Conferences lend natural ground-truth community structure to the network in [1], and it has therefore been a popular benchmark for static community detection tasks. This dataset expands upon it by capturing all Division I FBS games played between the 1924 and 2024 seasons as well as when each game was played. Dynamics take place in the form of conference realignment: most years witness a small but nontrivial number of teams change conference. Over time, conferences expand, contract, birth, die, rebrand, and resurge. In some cases the Ship of Theseus effect described in [2] occurs. Thus, this dynamic network may be of interest for the evaluation of dynamic community detection methods.
[1] Michelle Girvan and Mark EJ Newman. Community structure in social and biological networks. Proceedings of the national academy of sciences, 99(12):7821–7826, 2002.
[2] Remy Cazabet, Souâad Boudebza, and Giulio Rossetti. Evaluating community detection algorithms for progressively evolving graphs. Journal of Complex Networks, 8(6):cnaa027, 2020.