HGCN-GRU: temporal memory learning for hypergraph sequential recommendation

Sequential recommendation systems aim to predict users’ future interactions by modeling the temporal evolution of their preferences. Hypergraph-based representations are particularly well suited to this setting because they can capture high-order, non-pairwise dependencies among multiple users and items within a unified relational structure, enabling the modeling of complex collaborative patterns beyond conventional pairwise interactions. Despite this advantage, existing hypergraph-based sequential recommenders often rely on shallow temporal propagation, such as residual gating or single-step hypergraph convolution across adjacent snapshots, which limits their ability to preserve long-range semantic dependencies across distant time periods. Moreover, when trained solely with supervised signals, these models suffer from data sparsity, as the limited interaction density in real-world scenarios provides insufficient supervisory evidence for learning discriminative representations. In this work, we propose HyperGraph Convolutional Network-Gated Recurrent Unit (HGCN-GRU), a sequential recommendation framework that addresses both limitations through two complementary contributions. First, we replace the residual gating mechanism of existing hypergraph-based models with a Gated Recurrent Unit coupled with a temporal attention module. Unlike single-step propagation, the GRU models the sequence of per-period hypergraph embeddings as a temporal process, enabling the accumulation of semantic information across distant time steps, while the attention module adaptively emphasizes the most informative historical periods. Second, we introduce a self-supervised contrastive learning objective that generates auxiliary training signals directly from the hypergraph structure via stochastic hyperedge dropout augmentation. By aligning user representations derived from two independently augmented views, the contrastive module compensates for the scarcity of supervised signals in sparse interaction settings without requiring additional annotations. Extensive experiments on three benchmark datasets of varying density, namely Amazon Review, MovieLens-1 M, and Yelp, demonstrate that HGCN-GRU consistently outperforms seven competitive baselines across all evaluation metrics. The proposed model consistently outperforms the strongest baseline across datasets with varying interaction densities. The improvements remain substantial across datasets, including the sparsest setting, demonstrating the effectiveness of the proposed framework under different levels of interaction sparsity. Ablation studies further confirm that the GRU-based temporal memory and the contrastive objective address complementary weaknesses of existing hypergraph recommendation frameworks.