Personalized Group Recommendation Model Based on Hybrid Graph Neural Network

To address the issues of data sparsity and insufficient utilization of group interaction information in group recommendation systems, a personalized group recommendation model based on hybrid graph neural network is proposed. Firstly, the hypergraph convolutional neural network is designed based on the group interaction hypergraph to capture potential information that may exist between members and projects, the graph convolutional network is employed based on the group-item bipartite graph to capture high-order interaction information between groups and projects, and the graph attention network is utilized based on the projected hypergraph to dynamically learn the influence of different groups. Secondly, a double-stage gated adaptive fusion strategy is proposed to adaptively aggregate multi-view information to obtain the final group representation. Finally, the prediction score is calculated using the final group representation and the item representation from the hypergraph view. The experimental results on Mafengwo and CAMRa2011, demonstrate that the GRHG model outperforms mainstream models in terms of recommendation performance. Compared to the suboptimal model CubeRec, GRHG achieves improvements of 3.78%, 1.91%, 5.82%, and 4.88% in HR@5, HR@10, NDCG@5, and NDCG@10 on the Mafengwo dataset, respectively. Similarly, on the CAMRa2011 dataset, GRHG improves by 0.55%, 1.11%, 0.94%, and 0.34% in HR@5, HR@10, NDCG@5, and NDCG@10, respectively, validating the effectiveness of the GRHG model.