Random Virtual Embeddings Bootstrap High-Degree Item Diffusion for Recommendation

The data sparsity and cold start issues reduce the recommendation accuracy of recommendation models based on Graph Neural Networks (GNN). Existing research mostly attributes this to insufficient exploration of user-item interaction information, resulting in methods that primarily focus on how to effectively extract user-interacted items and their domain information. However, this raises a problem: sparse interaction data leads to incomplete capture of user preferences, meaning learned user features have locality. Overly focusing on interacted user embedding learning can result in a narrow range of node embeddings involved in node embedding learning, making it difficult to recommend uninteracted hidden preference items. To address this, this paper combines graph neural networks and designs a novel self-supervised contrastive learning recommendation model, REBD. REBD preserves the personalized recommendation advantage of traditional models focusing on obtaining user interests from existing interactions by designing item graph convolution and contrastive learning data enhancement methods. Meanwhile, the model introduces virtual embeddings to diffuse high-degree item information, prompting the model to learn more uniformly unified node embeddings to mitigate the narrowness of node embeddings, forming a new recommendation pattern: high-degree item recommendations beyond interactively associated items. Finally, testing on three commonly used real datasets across different domains shows that REBD’s recommendation performance surpasses various existing recommendation models. Comparing with one of the recent state-of-the-art models MMSSL on the dataset with the most severe data sparsity, the recommendation accuracy has significantly improved by over 30%, confirming the advantage of our model.