Practical Challenges and Methodologies in Next Basket Recommendation (NBR)

Recommendation systems are evolving rapidly, with increasing emphasis on deep and sophisticated model architectures to handle complex user-item interactions, resulting in larger models for better performance. However, the deployment of these models faces challenges due to high computing requirements and latency issues. To address this, a General Bidirectional framework (BiDiGen) is proposed, integrating user preferences and item similarities for personalized and context-aware recommendations. BiDiGen employs knowledge distillation to transfer pre-trained knowledge from large teacher models to smaller student models, striking a balance between efficiency and effectiveness. The framework utilizes a bidirectional approach where teachers and students mutually learn from each other, along with hypergraph construction to capture higher-order correlations efficiently. Additionally, a denoising generator is proposed for identifying relevant items, and a sampling scheme is suggested to determine knowledge transfer between teachers and students, ultimately training them with distillation and collaborative filtering losses.