Abstract:To address the challenges of friend recommendation under social network sparsity, this study proposes a graph hybrid learning-based friend recommendation model (GHLFR). The model integrates graph contrastive learning with self-training, and an enhanced graph convolutional neural network is designed to strengthen the propagation of user features. Meanwhile, the features are effectively fused, thereby mitigating social network sparsity. Specifically, contrastive views of different types are generated via a contrastive view generator, which contains a graph generator and a graph denoiser. These views are adopted for graph contrastive learning to capture the consistency of users’ social and behavioral features. A graph convolutional layer incorporating feature cross-interaction is designed to optimize the neighborhood aggregation process, thus enabling in-depth mining of user features. Then, users’ social and interest features are effectively fused by adopting an effective strategy. Finally, high-quality pseudo-links are selected via self-training-based prediction to iteratively optimize the model. By conducting experimental comparison and analysis with baseline models on three real-world datasets, including Ciao, Epinions, and Last.FM, the results demonstrate that the proposed model outperforms other models in terms of multiple evaluation metrics.