BanPickMaker: Meta-Aware Cold-Start Champion Recommendation for Evolving League of Legends Drafts

Abstract

Multiplayer online battle arena games require a strategic ban-pick phase, where teams must consider champion strengths, compositions, interactions, and the current meta. Existing approaches often rely on champion-ID embeddings or historical interactions, limiting their ability to generalize to fully draft-unseen champions and changing metagames. This thesis proposes BanPickMaker, a meta-aware champion recommendation model for cold-start ban-pick scenarios. BanPickMaker represents each champion using LLM-derived semantic profiles, patch-level champion-lane statistics, and causal rolling professional-meta features, while a draft-state encoder models the current draft context and legal candidate set. Experiments on temporal and champion-level cold-start benchmarks show that BanPickMaker outperforms frequency-based baselines and achieves stronger top-ranked recommendation performance. The results demonstrate that semantic champion representations and match-derived meta signals provide an effective and practical approach for cold-start and meta-adaptive draft recommendation in evolving esports environments.