HiSAC: Hierarchical Sparse Activation Compression for Ultra-long Sequence Modeling in Recommenders
New compression technique handles ultra-long user histories while cutting costs and improving click-through rates.
A research team from Alibaba and Zhejiang University has introduced HiSAC (Hierarchical Sparse Activation Compression), a novel framework designed to tackle a core challenge in modern recommender systems: efficiently modeling users' ultra-long behavior histories. Platforms like Taobao track thousands of past interactions to predict preferences, but processing these sequences end-to-end is prohibitively expensive in production. Existing summarization methods often lose granularity and struggle with long-tail interests. HiSAC addresses this by encoding user interactions into multi-level semantic IDs and constructing a global, shared hierarchical codebook. A key innovation is its hierarchical voting mechanism, which sparsely activates a small set of personalized "interest-agents" that act as fine-grained preference centers for each user.
Guided by these activated agents, HiSAC employs a Soft-Routing Attention mechanism to aggregate historical signals in semantic space, weighting items by their similarity to the user's personalized centers. This approach minimizes quantization error and crucially retains information about niche, long-tail preferences that are often lost in compression. The real-world impact is significant: deployed on Taobao's "Guess What You Like" homepage feed, HiSAC delivered substantial compression and cost reduction while improving recommendation quality. Online A/B tests showed a consistent 1.65% uplift in Click-Through Rate (CTR), a major business metric. This demonstrates HiSAC's scalability and effectiveness, providing a blueprint for making rich, long-sequence personalization feasible for billion-user platforms at a manageable computational cost.
- Uses a global hierarchical codebook & sparse activation to create personalized interest-agents, compressing ultra-long user sequences.
- Deployed on Taobao's homepage, resulting in a 1.65% Click-Through Rate (CTR) uplift in A/B tests.
- Solves production constraints of latency and memory, making detailed user history modeling scalable and cost-effective.
Why It Matters
Enables platforms to leverage detailed user history for better personalization without prohibitive computational costs, directly boosting engagement metrics.