A Hybrid Embedding and Graph Neural Network Framework for Cross-domain Recommendation

Cross-domain recommendation seeks to mitigate data sparsity by facilitating the transfer of information between interconnected areas.This work puts forward a hybrid deep learning framework that combines embeddingbased representation learning with a graph neural network (GNN)-based architecture to improve the accuracy of recommendations on Amazon and Yelp datasets.The first module uses shared and domain-specific embeddings with global bias terms.It was trained in two steps: first, with MSE-based supervised warm-up, and then with Bayesian Personalised Ranking (BPR) fine-tuning.The second module builds on a BiTGCF-inspired GNN by using a heterogeneous user-item bipartite graph to show higher-order dependencies and domain alignment.The results of the experiments reveal that Recall@K and NDCG@K both get better over time.On Amazon, the model achieved Recall@10 = 0.0989 ± 0.003 and NDCG@10 = 0.0451 ± 0.002, while on Yelp it reached Recall@10 = 0.0141 ± 0.001 and NDCG@10 = 0.0064 ± 0.0005.These metrics were computed from re-audited experiment logs using the same top-K evaluation pipeline, averaged across five random seeds.The proposed hybrid design successfully integrates domain representation learning and graph propagation, providing a scalable framework for resilient cross-domain recommender systems.