Win Prediction from the Snowball Effect Perspectives

Abstract

The global E-sports market has been growing steadily. In particular, “League of Legends” holds large international competitions every year, and professional leagues are held in each region. This paper conducted a study to predict advantageous teams in real-time using the time series data of League of Legends. A dataset was built by collecting game data with the API provided by Riot Games. Existing win-loss prediction studies using time series data have a limitation in that they learn as the final win-loss team without considering the flow of the game. To compensate for this, we propose a method of classifying advantageous real-time teams based on global gold indicators and learning with time series models. We trained LSTM, GRU, and RNN models using 76 features that subdivided the collected in-game data by position. As a result, our experiments show that all three models achieve an accuracy of more than 91 %.