Neural Feature Crossing for Cross-Domain Recommendation in Knowledge Management Systems

Recently, the rapid growth of digital knowledge resources has created an urgent need for advanced recommendation systems that capable to operate across various domains. This paper introduces KnowledgeXRec, a domain-aware hybrid recommendation framework that models both shared and domain-specific user-item interactions within a unified end-to-end architecture. Domain-specific feature crosses capture patterns unique to educational and cultural domains, while shared layers enable cross-domain knowledge transfer. Learnable embedding is represented categorical features such as learning style, age group, and content attributes, and the model is jointly optimized using binary cross-entropy loss with Adam optimizer. Findings provide evidence that KnowledgeXRec consistently outperforms seven state-of-the-art baselines on the OULAD (educational) and Goodreads (cultural) datasets. The model achieves enhancement of up to 15.6% in accuracy, 18.3% in diversity, and 54.8% in novelty metrics, and statistical analysis indicates the significance of these gains (p < 0.01). Overall, KnowledgeXRec offers a scalable and adaptable architecture for intelligent knowledge management systems, which enable effective cross-domain recommendations while maintain a strong domain-specific performance. This study illustrates that combining neural feature interactions with domain adaptation strategies introduce challenges in knowledge-intensive recommendation scenarios, such as cold-start problems, data sparsity, and heterogeneous domains. A new direction for future work is to extend the framework to multi-modal data, temporal interactions, and privacy-preserving federated learning, can further improve its applicability in diverse real-world knowledge management environments.