Abstract:

With the increasing demand for user personalization, diversity objectives in recommender systems have received growing attention. Graph Neural Networks (GNNs) provide a new paradigm for alleviating recommendation homogeneity and improving result diversity by modeling high-order associations between users and items. However, existing GNN-based methods mainly rely on local neighbor aggregation, which tends to introduce redundant information and lacks the ability to distinguish heterogeneous interest structures. In contrast, Transformers have strong global modeling capabilities and can capture long-range dependencies, but their computational complexity is relatively high, and they have difficulty to effectively capturing users’ latent diverse interests. To address these problems, a diversity-aware recommendation model named SGMDRec (Synergistic Graph-Mamba for Diversified Recommendation) was proposed to jointly optimize the accuracy and diversity of recommendation results. The model was composed of two core modules: dynamic programming neighbor selection based on high-dimensional prefix sums and a GNN-Mamba alternating network. The former constructed an efficient diverse subset selection strategy based on high-dimensional prefix sums. Through low-dimensional mapping, symbolic encoding, and category sparsity weighting, diverse neighbor subsets with category coverage potential were selected from candidate neighbors, and a category-aware representation enhancement mechanism was incorporated to characterize fine-grained differences among items within the same category. The latter first employed GNN to perform message passing between nodes and aggregate local structural information in the user-item interaction graph. Then, to overcome the computational efficiency bottleneck of traditional Transformer, a Mamba module was introduced for lightweight global sequence modeling. In particular, to enhance the graph-structure awareness of Mamba, a graph-aware state transition mechanism was further designed. By integrating node degree information, the parameter configuration of the State-Space Model (SSM) was dynamically adjusted, enabling the model to capture both local collaborative relations and global interest evolution characteristics. Experimental results showed that SGMDRec could effectively improve the category coverage and diversity of recommendation results while maintaining recommendation accuracy, achieving a better balance between accuracy and diversity. The results verified the effectiveness and scalability of the proposed method in diversified recommendation tasks and provide an efficient graph-aware modeling scheme for balancing accuracy and diversity in recommender systems.

Key words: recommender systems, diverse recommendation, graph neural networks, graph-aware modeling, dynamic programming neighbor selection