Baidu's MuSeR compresses 10^5-length user histories with multi-interest modeling to boost long-term recommendation
_reachsumit · x · 2026-09-22
Baidu published MuSeR, a retrieval framework for scalable long-sequence recommendation with multi-interest modeling.
- Pain point: industrial recommenders truncate user histories to a few hundred actions under latency/memory budgets, wasting long-term interest signals; sparse ID embeddings struggle to represent multiple heterogeneous intents.
- Three components: hierarchical temporal compression (full resolution for recent actions, pooled older segments, fitting 10^4–10^5 interactions per user), disentangled multi-query interest extraction with orthogonality regularization, and multimodal semantic alignment augmenting item IDs with LLM-distilled textual summaries.
- Engineering & results: async user-representation refresh with adaptive caching and hierarchical beam-search retrieval; consistent Recall@K gains over strong long-sequence baselines on three public benchmarks and a large industrial dataset.
More from Research
- Tencent ARC Lab releases GameHorizon Suite to benchmark VLMs and agents on AAA gameplay — yshan2u · 2026-09-22
- Stanford turns papers into AI agents; two unrelated studies surface unreported ADHD genetic link — VraserX · 2026-09-22
- Xiaomi MiMo 2.6 Pro formalizes Li–Yorke chaos theorem in 6,000+ lines of verified Lean — bookwormengr · 2026-09-22
- Apollo, the first advanced LLM for Ancient Greek, aims to restore tattered papyri — nordicinst · 2026-09-22
- Two Minute Papers: Jev claims 200x AI speedup via System One models — but there's a catch — Two Minute Papers · 2026-09-22
- New LLM fills gaps in tattered ancient Greek papyrus records — Wired AI · 2026-09-22