Mutable Sketches for Fresh Embeddings

Mutable sketches store user preferences in KP-trees, fit one low-rank projection, and refresh user embeddings as ratings arrive.

TL;DR — Mutable sketches target stale user embeddings in two-stage recommendation. The abstract says they store each user's preferences in a KP-tree, fit one low-rank projection, and recompute embeddings as ratings arrive. Reported KuaiRec results are 0.810 RMSE at 1.8% data read versus ALS at 0.822 RMSE and 100%, plus 8x faster per-batch updates and <1 ms personalization after a first rating. Citation metadata is not supplied in the abstract.

Core idea

The abstract frames the problem directly: 'A common bottleneck in two-stage recommendation is embedding staleness.' The specific failure mode is also explicit: 'when a user rates a new item, their embedding remains fixed until the next retrain cycle.'

Mutable sketches are proposed as a user-update mechanism. The abstract says they 'store each user's preferences in a KP-tree,' 'fit a low-rank projection once,' and 'recompute embeddings on-the-fly as ratings arrive.'

The safest reading is not that the whole recommender never trains. The abstract says the projection is fitted once, while user embeddings can be refreshed after new ratings without waiting for another full retrain cycle.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.