A data augmentation model integrating supervised and unsupervised learning for recommendation

Recommendation models based on Graph Neural Networks (GNNs) are typically employed within a supervised learning paradigm. However, the label data is extremely sparse across the entire interaction space, hindering the model's ability to learn high-quality embedding representations. Data augmentation techniques can alleviate the overfitting problem caused by insufficient label data by generating additional training samples. Therefore, we fused supervised learning tasks with unsupervised learning tasks, and applied different data augmentation techniques to learn the generation process, proposing a new recommendation model (DARec). In supervised learning tasks, we leverage the powerful generative capability of diffusion models for data augmentation. In unsupervised learning tasks, we enhance the user-item interaction graph and the knowledge graph (KG) by employing edge dropout. Unlike existing data augmentation methods, DARec does not rely on traditional labeled data; instead, it generates supervisory signals from the input data itself to train the model. This approach enables the model to learn feature representations of the data without explicit labels, thereby leveraging a large amount of unlabeled data to enhance learning efficiency. Moreover, it endeavors to minimize damage to the original interaction matrix and graph structure as much as possible. Validation on three representative public datasets shows that our DARec model outperforms several state-of-the-art recommendation models.