CAST: Modeling Semantic-Level Transitions for Complementary-Aware Sequential Recommendation

A research team led by Qian Zhang has introduced CAST (Complementary-Aware Semantic Transition), a novel AI framework designed to revolutionize sequential recommendation systems. Traditional models often fail by mistaking spurious correlations from co-purchase data for true complementary relationships, leading to poor suggestions. CAST addresses this by shifting the modeling paradigm from the item level to the semantic level. It uses a semantic-level transition module to track dynamic changes directly within discrete semantic codes (like product specifications), preserving fine-grained details that are typically lost when aggregated into coarse item representations.

The framework's second key innovation is a complementary prior injection module that integrates LLM-verified knowledge of complementary patterns directly into the model's attention mechanism. This guides the system to prioritize genuine 'goes-with' relationships over simple popularity bias. In experiments across multiple e-commerce datasets, CAST demonstrated a significant performance leap, achieving up to a 17.6% improvement in Recall and a 16.0% gain in NDCG compared to state-of-the-art models. Remarkably, it also delivered a 65x acceleration in training speed, making it both more effective and computationally efficient for real-world deployment.