Repository for the bachelor thesis "Identifying toxic behaviour in online games". This thesis introduces a way to represent a given video game match as an event graph and using Graph Neural Networks (GNNs) to train a model to detect toxic behaviour in a given match.
More specifically we achieve this by projecting a video game match, which itself can be understood as a temporal network, into an event graph.
This graph we can then enhance using other information such as a graph connecting players that frequently play with eachother.
We can then apply various GNNs on this graph to train a model. More specifically we chose a simple GNN based on Principal Neighbourhood Aggregation.
| Type | Dataset | ROC-AUC |
|---|---|---|
| Multiclass | Detoxify | 0.6134 |
| Multiclass | Annotation | 0.6957 |
| Multiclass | Annotation-Enhanced | 0.7237 |
Datasets
Detoxify: Dataset including 10.000 matches labeling matches as toxic based on the NLP tool Detoxify.
Annotation: Dataset based on roughly 1000 human annotated matches.
Annotation-Enhanced: Dataset based on human annotated matches enhanced with a player graph with weights representing the amount of times they play with eachother.
@misc{Schrottenbacher2024,
author = {Patrick Schrottenbacher},
title = {Identifying toxic behaviour in online games},
institution = {Informatik und Mathematik},
type = {bachelorthesis},
pages = {35},
year = {2024},
url = {https://publikationen.ub.uni-frankfurt.de/files/81676/Toxic_video_game_classification.pdf}
repository = {https://github.com/TheBv/toxic-video-games-gnn}
}