Justice League: Time-series Game Player Pattern Detection to Discover Rank-Skill Mismatch

Abstract

When rank and skill do not coincide in competitive games, this might be a sign of issues such as boosting, smurfing, and trolling occur. The fair gaming culture of online gaming is disrupted and offended by cheating like boosting, smurfing, and trolling. The player's play style must be used to determine the rank that appears in account information. In this study, we classified League of Legends' low and high tiers using a sequence-based CNN-LSTM model. Using input perturbation, the model can explain its own importance for certain features. The experimental progress: First, we selected features that show a difference between tiers by an extracted score estimating a cumulative sum graph. Second, we construct the dataset format variously with variable or fixed sequence length, compare performance, and analyze the pros and cons. Finally, we consider the possibility of early detection by measuring performance over game elapsed time. Along with the experiment, a rank classification performance of the model achieved AUC 0.9036 and found that we can distinguish from the 24 minutes after the start of the game. In addition, We derived that ccReduction and MinionsKilied were the information that had the most influence on skills among various features.