The Multiplayer Online Battle Arenas genre features matches characterized by strategic real-time confrontations. The ability to predict victory in real-time in these contexts is of considerable practical importance for the development of support tools for professional players and teams, as well as for the creation of analytical systems that aid in understanding game dynamics. This work conducted a literature review on victory and event prediction in MOBA games. This investigation identified a lack of studies analyzing the Transformer model for in-game victory prediction.
To investigate the potential of the Transformer model in predicting victories in MOBA games, two distinct architectures were conceived. The first treats each match attribute as a token, analogous to words in natural language tasks, allowing for the evaluation of the impact of attribute ordering on the victory prediction task. The second is the FT-Transformer, which optimizes the handling of tabular data by preserving the inherent structure of the table and employs column-specific embeddings, facilitating the identification of attribute interactions without relying on a predefined sequence.
Experiments in this work used two public datasets from the game League of Legends, one from professional tournament matches and another from ranked matches. For both versions of the Transformer model, a study of appropriate hyperparameters for the datasets was performed. The FT-Transformer presented superior performance on both datasets, outperforming other evaluated methods in terms of accuracy, recall, and F1-score. Meanwhile, the other Transformer version, while demonstrating the influence of attribute structure, yielded the worst results for the same metrics.