Knowledge Transfer-Enhanced Contrastive Learning for Long-Tailed User Recommendation

In recommendation systems, Graph Convolutional Networks (GCNs) have shown superior performance by exploiting high-order relationships. However, generating accurate recommendations for long-tailed users with sparse interactions in large-scale scenarios remains a challenge. We propose a knowledge transfer-enhanced contrastive learning framework for long-tailed user recommendation. Our framework incorporates self-supervised signals through contrastive learning to differentiate head and tail nodes processing, thereby mitigating model bias and improving the utilization of tail nodes. Specifically, we design an automated dropout module that creates pseudo-tail nodes from head nodes with the aid of knowledge transfer. Furthermore, we construct a GCN encoder and a contrastive learning-based graph autoencoder to guide the model in learning better node representations. To ensure that the generated tail nodes are consistent with the distribution of tail nodes in the original data, a Generative Adversarial Network (GAN) is adopted. The experimental results on three benchmark datasets demonstrate that our model substantially outperforms existing state-of-the-art baselines.