MOBARec-GCNFP: Champion Recommendation for Multi-Player Online Battle Arena Games Using Graph Convolution Network with Fewer Parameters

Abstract–Multi-player online battle arena (MOBA)games are one of the most successful strategy videogames. Ten players are divided into two teams in a single match, each of which needs to pick five characters, known as champions, to play against the other team. Champions are designed with unique skills, strengths, and weaknesses, allowing some champions pairings to have a higher winning rate. For example, a defensive champion often complements a powerful but fragile champion. In this paper, we recommend the last champion to pick, i.e., the 10th champion, known as the counterpick, which is the most important pick to best cooperate with the already selected champions and counter the opponents’ selections. Previous studies recommending champions suffer from the accuracy and the cold start recommendation problem. To tackle the above problems, we propose MOBARecGCNFP, which adopts lightweightgraph neural networks. We solve the cold start recommendation problem with a newly designed GCN by initializing the match embeddings using champions’ embeddings. Our experimental evaluation confirmed that our method outperformed the previous methods, including fully connected NN and SR-GNN, improving Recall@l from 0.1208 to 0.1272 (5.3%) on the League of Legends dataset and from 0.2551 to 0.2586 (1.4%) on the Honor of Kings dataset.