Categorical Learning-based Line-up Prediction in the Drafting Process of MOBA Games

Abstract

In recent years, the e-sports industry has witnessed a rapid development, setting a wave of social and economic influence all over the world. Among these, multiplayer online battle arena (MOBA) games, represented by Dota2 account for the largest share in the number of games, audience, and prizes. As the beginning of the MOBA games, the drafting phase is of great importance due to its impact on the outcome of the game. Therefore, line-up prediction in the drafting process is an important research field in e-sports. Nonetheless, different MOBA games have different hero pools, different rules for drafting, and different non-player factors, while the available game data are sparse, hindering the prediction of line-up in the drafting phase. To solve this problem, we i) identify the key steps in lineup prediction through statistical analysis, ii) based on which we propose an orderless prediction method that achieves 7% higher accuracy than existing models, and iii) further improve the accuracy by 1.6% with an attention layer in line-up predictions. Our finding of orderless key steps in MOBA game lineup drafting may also be used to expand the lineup dataset, which is crutial to learning-based automatic game analysis.