Fuzzy-enhanced variable weight graph convolutional networks for recommender systems

Recommender systems play an essential role in alleviating information overload by delivering personalized suggestions to users across domains such as e-commerce, restaurant services, and digital media. In recent years, graph-based approaches, particularly those leveraging graph convolutional networks (GCNs), have shown strong performance by modeling high-order connectivity. However, their effectiveness remains constrained by three critical challenges: the sparsity of user–item interactions, the presence of noisy or transient behaviors that distort preference modeling, and the underutilization of contextual information contained in reviews and product descriptions. To address these limitations, we propose a novel framework, termed fuzzy and variable weight graph convolutional network (FVW-GCN). The framework incorporates a fuzzy relation modeling module that enriches the adjacency structure by applying fuzzy C-means clustering to semantic embeddings extracted from pre-trained language models, thereby improving connectivity for sparse and long-tail items. In addition, a variable-weight GCN module is introduced, where a tuning GCN learns localized weight matrices from sampled subgraphs, which are then used by a tuned GCN to adaptively refine embeddings and suppress noisy signals. Through this combination, FVW-GCN effectively strengthens meaningful relations while reducing the influence of unreliable interactions. Extensive experiments conducted on benchmark datasets demonstrate that FVW-GCN consistently outperforms state-of-the-art baselines across several standard evaluation metrics, including recall, normalized discounted cumulative gain, and hit ratio. These results confirm the robustness and effectiveness of the proposed framework, highlighting its potential to support more accurate, diverse, and user-centric recommendation services in real-world applications.