Research map / Label-efficient and robust learning / Recognition and long-tail

UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UxSID, which uses semantic IDs as queries over compressed item-agnostic user interest anchors so that ultra-long user sequences can be pre-computed offline per user-SID pair and served with O(1) lookup, reporting gains on XLong, KuaiRec-Big and a Kuaishou ads A/B test.

Problem

Ultra-long user behavior modeling in recommender systems faces a trade-off: item-specific search is costly online and biased by its key space, while item-agnostic compression loses target-specific interest signals.

Approach

An intermediate path: compress the user history into item-agnostic interest anchors, then query them with semantic IDs of the target item (with a gated two-stage attention), so semantically similar items share compressed interest memories and embeddings can be precomputed per user-SID key.

Key results

• On XLong, UxSID reaches 0.8408 AUC versus 0.8154 for TWIN and 0.8135 for C-Former.
• On the Kuaishou industrial CTCVR task, UxSID reaches 0.8626 AUC, above TWIN's 0.8609.
• A one-week online A/B test on Kuaishou advertising shows +0.337% revenue, +0.231% cost and +0.111% exposure.

Surprising findings

• Adding first-layer SID as a sparse feature to all baselines improves them only slightly, and UxSID still leads, which the authors use to argue gains are architectural.
• Gains on the industrial CTCVR AUC are small (about 0.17-0.18 points over TWIN and SIM-Soft).

Limitations

• Offline storage grows with users times unique SIDs (about 2.56 TB estimated), and embeddings must be refreshed weekly.
• Public-benchmark gains are reported with a single model variant per dataset, and the industrial gains over baselines are small in AUC.

Experimental setup

Datasets: XLong KuaiRec-Big Kuaishou industrial advertising logs

Benchmarks: XLong KuaiRec-Big Kuaishou industrial dataset

Metrics: AUC UAUC WUAUC Int.R@K exposure cost revenue

Baselines: DIN SIM-Hard SIM-Soft ETA SDIM MIRRN TWIN C-Former

Builds on: DIN SIM TWIN C-Former RQ-VAE LETTER

Closest papers by meaning

• Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.88)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.88)
• LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling (NeurIPS 2026 · similarity 0.88)
• Fine Tuning Out-of-Vocabulary Item Recommendation with User Sequence Imagination (NeurIPS 2024 · similarity 0.87)
• MARS: Multi-resolution Adaptive Routing for Sequential Recommendation (NeurIPS 2026 · similarity 0.87)
• SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.86)