Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling
Researchers have developed a new framework for modeling extremely long user behavior sequences in short-form video recommendation systems. The system uses content-native Semantic IDs instead of traditional item IDs to reduce embedding table size and improve generalization to new content. Additionally, a Global-Aware Compression Transformer condenses user sequences, significantly lowering memory and computational requirements. AI
IMPACT Enables more effective personalization in short-form video platforms by handling longer user histories.