Deep Learning for Cross-Domain Recommendation with Spatial-Channel Attention

Abstract

With the advancement of information technology, recommendation systems have become essential for enhancing user experience and improving information retrieval efficiency. Traditional recommendation methods perform well in single-domain scenarios but face challenges in cross-domain tasks, including data sparsity, user interest drift, and feature heterogeneity. To address these issues, this study proposes a cross-domain recommendation algorithm based on an improved Spatial-Channel Attention Mechanism (SCAM). By integrating spatial and channel attention, the method effectively captures user interest patterns across domains and optimizes cross-domain feature fusion. Additionally, this study examines the impact of different item feature inputs and model parameters (e.g., embedding size and attention heads) on recommendation performance. Experimental results demonstrate that the proposed method outperforms existing approaches across multiple evaluation metrics (HR@10, NDCG@10, Precision@10, and Recall@10) and exhibits strong generalization ability. Future research can further optimize the model architecture by incorporating graph neural networks and reinforcement learning to enhance the intelligence of recommendation systems.