Multiplex Memory Network for Collaborative Filtering

Hyperbolic Neural Collaborative Recommender

“…In addition, most deep recommendation models have less focus on explicitly modeling user-user or item-item high-order semantic correlations, while such relations could provide valuable information to inference user or item features. Although some existing works [5,15] utilize the co-occurrence relation (co-engage between users or co-engaged between items) to define the neighbors for users and items, we argue that such co-occurrence relation is macrolevel and coarse-grained. For instance, Figure 1(a) shows a simple user-item interaction in the movie domain.…”

“…Note that it only focuses on the user's neighbors without accounting for the information about similar items. • MMCF is a memory-based model, which models user-user and item-item co-occurrence contexts by memory networks [15]. Different from our methods, it only focuses on co-occurrence relations and ignores high-order semantic transitive relations.…”

“…Early models like Matrix Factorization (MF) embed users and items in a shared latent space and model the user preference to an item as the inner product between user and item embeddings [17]. However, due to the complex interaction between users and items, the shallow representations in MF-based methods lack expressiveness to model features [11,15]. As deep learning developed, some recommendation approaches utilize neural networks to capture complex interaction behaviors, which enhance the performance of previous shallow models [4,5,9,11,15,42,46].…”

“…However, due to the complex interaction between users and items, the shallow representations in MF-based methods lack expressiveness to model features [11,15]. As deep learning developed, some recommendation approaches utilize neural networks to capture complex interaction behaviors, which enhance the performance of previous shallow models [4,5,9,11,15,42,46].…”

Heterogeneous Side Information-based Iterative Guidance Model for Recommendation

“…Besides, to make full use of the heterogeneous side information, we use the attention network to learn the importance of different types of heterogeneous side information. Some studies have shown that different types of heterogeneous side information make different contributions to the learning of user and item representations [11], [24]. Discriminatively treating the heterogeneous side information helps ILIG comprehend the user and item accurately.…”