Predicting Trolling Behaviors Using In-Game Timeline Data: A League of Legends Case Study on Divergent Mechanisms Behind AFK and Item Tier-Dropping

This study challenges the prevailing view that trolling in online games is driven primarily by malicious intent, proposing instead that different trolling behaviors arise from distinct performance trajectories. Using approximately 2.9 million League of Legends match timelines obtained through the official Riot Games Developer API, the research models two behaviors: Away From Keyboard (AFK) inactivity and intentional item tier-dropping. Timeline events were segmented into 0–10, 10–20, and 20+ minute windows, and relative performance was measured through KDA and gold differences at the lane and team levels. The analysis is explicitly scoped as a single-title case study rather than a general account of MOBA games.

Decision-tree classifiers were trained to identify interpretable thresholds, and SHAP values were used to quantify which performance disparities most strongly predict each behavior. The findings indicate that AFK is most closely associated with early and sustained disadvantage relative to a lane opponent, consistent with withdrawal following diminished perceived influence. Item tier-dropping, by contrast, is most strongly predicted by a sustained gold advantage around the 30-minute mark, with subsequent splits shaped by KDA advantage differences across time windows.

The authors interpret the advantage-based pattern as less consistent with hostility-oriented trolling frameworks, since self-weakening primarily affects one's own team. They suggest it may instead reflect strategic disengagement, protest against teammates, or time-efficiency considerations. The study acknowledges that it cannot directly establish player intent and does not control for rank bracket, queue type, or competitive stakes, which are identified as important directions for future work. Proposed extensions include combining behavioral logs with surrender-vote data, interviews, or chat analysis.