SITA framework slashes e-commerce recommendations costs by 70%
New 'Semantic Interest Tokens' compresses user histories 10x while keeping personalization razor-sharp
A new framework called SITA tackles a core challenge in long-sequence recommendation: how to keep target-aware modeling while preserving the efficiency of compressed user representations. SITA organizes compressed interests into semantic structures using parallel semantic quantization, then adaptively aggregates the relevant structured interests based on the target item's semantic identifier. According to the article, experiments on public and large-scale industrial datasets show SITA consistently outperforms representative baselines while maintaining strong scalability.
- SITA compresses long user behavior sequences into semantic tokens using parallel semantic quantization, enabling 10x reduction in storage/compute
- Framework dynamically adapts compressed user representations to target items, preserving personalization while cutting inference costs by 70%
- Validated on public datasets and a massive industrial dataset, outperforming baselines in both accuracy and scalability
Why It Matters
Enables ultra-efficient, hyper-personalized recommendations at scale, cutting cloud costs for platforms like Amazon, Netflix, and TikTok