Information, Год журнала: 2025, Номер 16(5), С. 422 - 422
Опубликована: Май 21, 2025
Most current graph neural network (GNN)-based social recommendation systems mainly extract negative samples from explicit feedback, and are unable to accurately learn the boundaries of similar positive samples, which leads misjudgment user preferences. For this reason, we propose introduce hop-mixing technique synthesize hard for users fully explore their Firstly, sample information is injected into original in each layer generate augmented that very samples. Then super-enhanced with highest inner product score identified layer, finally, aggregated pooled obtain final Subsequently, a fusion mechanism used aggregate representations user–item bipartite graph. Comparative experiments on two real datasets ten baseline models conducted, results show proposed method has certain performance advantages over other state-of-the-art models.
Язык: Английский