Trust-based core social graph convolution: An innovative framework for location recommendation

With the increasing popularity of location-based social networks (LBSNs), recommendation based on LBSNs has attracted wide attention in academic and industrial domains. Traditional recommendation systems face critical challenges in handling data sparsity and underutilized social trust. Existing traditional models particularly struggle to incorporate both trust-level differences and implicit relationships effectively within large-scale personalized recommendations. This limitation directly leads to aggravated data sparsity, further exacerbating cold-start scenarios and significantly reducing recommendation accuracy for long-tail items. Existing widely-adopted deep learning models, such as graph convolutional networks (GCNs), apply equal propagation weights to all neighbor nodes, which not only causes feature convergence between high-trust and ordinary users (over-smoothing) but also incurs substantial computational overhead. We propose the Trust-Based Core Graph Collaborative Filtering (TCGCF) framework, integrating trust-weighted core graph analysis with trust-constrained graph convolution. TCGCF captures both explicit and implicit trust relationships, enhances personalization, and mitigates over-smoothing. Our trust-weighted core graph analysis identifies influential users in information propagation, while the trust-constrained convolution scheme enables precise, differentiated information flow. Experiments on real-world datasets demonstrate that TCGCF improves recommendation accuracy and computational efficiency, outperforming existing models in precision, recall, and suitability for large-scale applications.