Fast and Scalable Recommendation Retrieval Model with Mixed Attention, Knowledge Distillation and Approximated Vector Retrieval

Users' purchasing behaviors are influenced by both long-term preferences and recent interactions within digital platforms. Sequential recommendation systems are critical for predicting next best actions based on user history. Although deep neural networks have advanced this field, they often suffer from limited memory scopes and an overemphasis on popular items, reducing recommendation diversity. To overcome these challenges, we propose integrating external relational information through knowledge graphs synthesized from historical interactions, embedding them into the multi-head attention mechanism. Following training, knowledge distillation transfers the rich information from a complex teacher model into a lightweight student model, ensuring both accuracy and fast inference. Additionally, to accelerate retrieval during recommendation, we incorporate Approximate Nearest Neighbor approaches. These algorithms enable rapid vector similarity search, dramatically reducing query response times without compromising retrieval quality, which is essential for real-time serving at scale. Our approach enhances personalization, improves diversity, and ensures efficient large-scale deployment.