Hierarchical Transformers for Group-Aware Sequential Recommendation
Date:
This talk presents HT4Rec, a hierarchical transformer-based model designed for group-aware sequential item recommendation in MOBA video games, specifically addressing the challenge of generating contextual recommendations as game events unfold. Unlike previous approaches that suggest static item sets based only on character attributes and game participants, HT4Rec leverages both a bi-directional contextual encoder (capturing interactions among all players) and a unidirectional sequential encoder (modeling temporal dependencies), thereby learning user and group patterns for more accurate next-item suggestions. Extensive experiments on the Dota 2 dataset show HT4Rec outperforming existing sequential and non-sequential baselines; ablation studies highlight the critical role of group context and multi-head attention, while attention weight analysis offers insights into the model’s interpretability in capturing game dynamics. The model’s versatility is further validated in movie recommendation tasks, demonstrating its flexibility for other group-based sequential environments per findings in the published paper.