Knowledge-aware modeling of group commonality for group recommendation

Group Recommendation (GR) aims to offer recommendations that satisfy the entire group. Due to the inherent sparsity of group-item interactions in GR, relying solely on individual-level preference aggregation is often insufficient for producing high-quality recommendations. In contrast, mining group-level commonality that reflects shared behavioral patterns can help mitigate this challenge. Some existing methods attempt to model group commonality based on the number of overlapping users across groups. However, this approach often fails in sparse settings where shared users between groups are absent, leaving the data sparsity issue unresolved. To tackle these issues, we propose a novel model based on Knowledge-Aware Modeling of Group Commonality for Group Recommendation (ComRec). ComRec eliminates the reliance on overlapping users by modeling fine-grained commonality from the item side. Specifically, we construct a Group Collaborative Knowledge Graph (G-CKG) by integrating group members’ interactions, membership relations, and item knowledge, enabling the capture of multi-hop relational paths for each member. We then extract fine-grained commonality by fusing multiple relational representations with an orthogonal constraint to ensure signal independence. A novel commonality attention mechanism further aggregates member entity representations to derive the overall group-level commonality representation. Beyond modeling group commonality, we further consider the specific group composition by introducing a user-based fine-tuning module that refines the group representation through member-level differences. The results show that our model significantly outperforms existing methods in terms of classification accuracy and interpretability on Yelp and MovieLens-20M datasets, while effectively addressing the data sparsity issue in GR.