TL;DR
This paper tackles the pervasive problem of toxic behavior in online games like League of Legends by proposing a machine learning approach to predict the outcome of the game's crowdsourced "Tribunal" system. By analyzing millions of player reports and decisions, they demonstrate that automated systems can effectively identify and flag toxic behavior, potentially saving significant human effort and improving player experience.
What Problem Does This Paper Address?
Online multiplayer games, especially competitive ones like League of Legends (LoL), are plagued by "toxic behavior" – harassment, verbal abuse, and intentional sabotage. This not only ruins the experience for other players, leading them to quit, but also poses a significant threat to the sustainability of the gaming industry. LoL uses a crowdsourced system called the Tribunal, where human players review reports of toxic behavior and vote on whether to punish offenders. While effective, this system is costly, time-consuming, and relies heavily on human effort. The paper aims to automate this decision-making process by using machine learning to predict the Tribunal's verdict, thereby reducing the burden on human reviewers and potentially speeding up the process.
Methods
The authors employ a supervised machine learning approach, specifically using a Random Forest classifier. They extract a rich set of features from three main categories:
- In-game Performance: This includes metrics like kills, deaths, assists (KDA), gold earned, damage dealt/received, and time played. They normalize these to account for team-based gameplay and the impact of toxic players on their teammates' performance.
- User Reports: Features derived from the reports submitted by players who witnessed the toxic behavior, such as the number of allied/enemy reports, the type of reported offense, and the presence of written comments.
- Chat Logs: Linguistic analysis of in-game chat messages, focusing on the "valence" of words (a measure of positivity/negativity) to quantify the tone of communication. They use the Affective Norms for English Words (ANEW) dataset for this.
They train models using these features and evaluate their ability to predict whether a player will be punished or pardoned by the Tribunal.
Datasets
The primary dataset consists of over 10 million user reports from approximately 1.46 million toxic players, along with their corresponding crowdsourced decisions from the League of Legends Tribunal. This data was collected from three different regions: North America (NA), Western Europe (EUW), and South Korea (KR), allowing for analysis of regional differences and classifier portability.
Key Insights and Ideas
- Crowdsourcing is noisy but valuable: While human reviewers are the "gold standard," their decisions can be influenced by various factors. Machine learning can learn from these aggregated decisions to identify patterns.
- Context is crucial: Toxic behavior isn't just about individual performance; it's heavily influenced by in-game events, player interactions, and the communication within the game.
- Chat valence is a strong indicator: The emotional tone of chat messages, quantified by word valence, is a significant predictor of toxic behavior, even for non-communication-based offenses.
- Agreement matters: Decisions made with overwhelming agreement among human reviewers are more reliable for training machine learning models.
- Portability is possible, with caveats: A classifier trained on data from one region (e.g., NA) can perform reasonably well on data from another (e.g., EUW), though regional differences in language and culture can impact performance.
Key Results
- The proposed Random Forest classifier achieved an AUC of 0.7991 for predicting Tribunal decisions (punish/pardon) using all features, demonstrating good performance.
- Models trained on cases with higher reviewer agreement (strong majority, overwhelming majority) yielded better accuracy.
- The "full model" (combining all feature types) generally outperformed individual models (in-game performance, user reports, chat).
- The chat model, in particular, showed strong performance, with the valence score of offender's chat being a top predictor.
- The classifier demonstrated reasonable portability across regions, with an AUC of 0.7157 for EUW decisions when trained on NA data.
- A cost analysis suggests that automating Tribunal decisions could save Riot Games significant money (estimated at over $470,000 in the first year) and protect a large number of innocent players from toxic interactions.
Limitations and Future Work
- Language barriers: The reliance on English word valence for chat analysis limits its direct applicability to non-English chat logs, though the authors found some success with EUW data despite linguistic diversity.
- Adaptability of toxic players: Toxic players might adapt their behavior to evade detection, for example, by using sarcasm.
- Meta-game shifts: Changes in the game's "meta" (strategies and tactics) can influence in-game performance metrics, potentially affecting model accuracy over time.
Future work suggested includes:
- Real-time detection: Developing systems that can identify and flag toxic behavior as it happens, rather than after the fact.
- Adaptive interfaces: Creating interfaces that can warn players exhibiting toxic behavior in real-time.
- Investigating in-game penalties: Exploring the effectiveness of immediate, in-game penalties (e.g., degrading champion status) as a deterrent.
- Qualitative analysis: Combining quantitative findings with user interviews to gain deeper insights into reviewer decision-making processes.
Summary updated