EUVIMEDEUVIMED entdecken
Crossref

Toxic Steam: A Large-Scale Characterization and Prediction of Toxic Users in a Gaming Platform

Paula T. M. Gibrim, Vinicius K. Lodi, Gabriel S. Andrade, Marcus V. G. Ribeiro, Maurício Gruppi, Daniel M. Barbosa, Philipe F. Melo, Julio C. S. Reis

Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 86-96

DOI: 10.1609/aiide.v22i1.43567 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, online erschienen am 08.11.2026. Verfasst von Paula T. M. Gibrim und 7 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Steam is a major digital distribution platform for PC games, and its public game reviews provide a setting for studying toxic discourse. Although message-level toxicity detection has been widely studied, user-level analysis has received less attention, particularly on gaming platforms. This study characterized toxicity across 36.8 million English-language reviews from 14.2 million users. Detector-labeled toxicity was rare (1.86% of reviews, 3.95% of users), with higher rates in competitive game contexts; 57.5% of flagged reviews nonetheless recommended the reviewed game. Building on this characterization, we introduced a prevalence-aware benchmark spanning classical models, fine-tuned transformers, and Large Language Models (LLMs), with primary evaluations at approximately 4% positive prevalence. Text-derived representations achieved higher AUC-PR than profile metadata (up to 0.643 versus 0.070). A leave-toxic-out control retained a weaker signal after removing the reviews that defined each user's label, suggesting that performance depended substantially on direct access to the label-defining text. LLM classifiers obtained lower AUC-PR on their smaller evaluation sample, while narrative generators produced evidence-linked explanations for inspection. Their utility for human-review triage remained unvalidated. These results supported further study of user-level prediction as a triage aid, with explicit attention to label circularity and uncertainty. Warning! This paper contains examples of offensive and hateful language.

Zitieren

APA

Gibrim, P. T. M., Lodi, V. K., Andrade, G. S., Ribeiro, M. V. G., Gruppi, M., Barbosa, D. M., Melo, P. F., & Reis, J. C. S. (2026). Toxic Steam: A Large-Scale Characterization and Prediction of Toxic Users in a Gaming Platform. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 86-96. https://doi.org/10.1609/aiide.v22i1.43567

Vancouver

Gibrim PTM, Lodi VK, Andrade GS, Ribeiro MVG, Gruppi M, Barbosa DM, et al. Toxic Steam: A Large-Scale Characterization and Prediction of Toxic Users in a Gaming Platform. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):86-96. doi:10.1609/aiide.v22i1.43567

BibTeX
@article{Gibrim2026,
  author = {Gibrim, Paula T. M. and Lodi, Vinicius K. and Andrade, Gabriel S. and Ribeiro, Marcus V. G. and Gruppi, Maurício and Barbosa, Daniel M. and Melo, Philipe F. and Reis, Julio C. S.},
  title = {Toxic Steam: A Large-Scale Characterization and Prediction of Toxic Users in a Gaming Platform},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
  year = {2026},
  volume = {22},
  number = {1},
  pages = {86--96},
  doi = {10.1609/aiide.v22i1.43567}
}
Kostenlos starten →