PERSONALIZATION · 2026-11-11
Personalization — Paper Daily, 2026-11-11
Today's papers in this track all answer one question: at which step of the system, and in what form, should information about the user be used. SepLRM pulls long-sequence encoding out of individual ranking models so they can share it, and Wayfair searches long-term history in sponsored search by both the candidate product and the query. Airbnb's cover photo experiments show that narrowing candidates by search intent first brings more bookings than simply pushing up clicks. UAPM keeps unresolved readings of behavior outside the user profile until later evidence settles them. Meta's Your Algorithm lets people say in their own words what they want to see, and connects these preferences to the whole path from retrieval to the device.
4.45
On recommendation platforms, large models for different stages and business lines each encode the same long user history, and this repeated cost gets harder to justify as models and sequences multiply. Existing sharing schemes mostly pass downstream scores or compressed user representations, so the detail is already gone.
4.40
Implicit feedback mixes in accidental clicks and impulse purchases, and existing denoising methods almost all use training loss to find noise, but loss only grows after the model has been misled and is also disturbed by training stage and negative sampling. NDL from NetEase adds a signal that ignores predictions: the cosine similarity between an interaction's representation and the mean representation of that user's past interactions, where larger deviation means more suspicion, and this signal is available before the model makes a prediction error.
4.25
Airbnb wants to choose cover photos for hosts, and there are two difficulties. Over 90% of listing photos have never been a cover and so have no click data to learn from, and the most clickable photo may show a nearby attraction that does not represent the listing. The authors first screen candidates with indirect cues, use a VLM to check photos against the listing description, and then pick by predicted CTR, which raises HomePage CTR by 1.5%.
4.20
When sponsored search ranking uses long-term behavior, the usual approach searches history for actions similar to the candidate product, which can miss history highly relevant to the current query but only loosely tied to the candidate. Beyond recent behavior, Wayfair retrieves one batch of long-term actions by candidate and another by query, then feeds them to a Transformer with two context tokens for the query and the candidate.
4.10
Putting an LLM in the online CTR path is too expensive, so existing methods mostly compute LLM outputs offline and store them, which only moves the compute offline and must be redone whenever the item or feature catalog changes. SlimFM tries to run the LLM directly online, first applying structured pruning to LLaMA3, removing the least important groups of dependent parameters, 44% in total. LoRA fine-tuning then makes the model produce one embedding per feature field, such as MovieID, and similarity between fields reweights the feature interactions of the CTR model. On FM, the pruned model keeps 99.77% of the full model's mean AUC, and inference is up to 3.4 times faster than with existing LLM-enhanced methods. It holds no clear accuracy lead, reaching the highest AUC among the LLM-enhanced methods in fewer than half of the model and dataset combinations, so its main value lies in latency. Teams wanting semantic features online gain a route that avoids offline recomputation, with evidence so far from four public datasets.
3.90
ChatGPT Ads lets advertisers write natural-language hints about what a product is, who it suits, and when it helps, and these hints feed ad ranking separately from the ad card shown to users, yet how to write them has remained unclear. The authors split them into seven styles from marketing frameworks, add one that mixes all seven, and run over 1.19 million paired comparisons with and without hints. Adding hints moves an ad up by 0.56 to 1.16 positions on average, but the seven styles differ little and the mix has no consistent advantage. Measured by rank gain, the difference shows up in the ranker: embedding rankers favor hints about usage situations, while LLM rankers favor hints about target audiences, and which style wins depends on the ranker and the metric. Adding more hints generally improves rank further, with diminishing marginal gains. In these experiments only the evaluated ad receives hints and the rankers are offline, so the effect when competitors also write hints remains a separate question.
3.75
When an LLM maintains a user profile, new behavior often has several readings: a fantasy fan rating a fantasy drawing guide 3 out of 5 may be tiring of fantasy, or may simply dislike drawing guides. Existing methods must either write it into the profile at once or discard it, and a wrong entry keeps misleading later predictions. UAPM adds two stores outside the profile that the downstream model cannot read, one holding unresolved readings with their evidence and one recording, per user, how much evidence is needed before a conclusion. In training, a Teacher that sees the user's future behavior guides each decision of a Student, and at deployment the Student uses only current information. Under the same profile length, it beats the best baseline by up to 5%, and removing either store worsens almost every metric. Held entries rarely move straight into the profile and mainly serve as evidence for later profile edits. Storing undecided readings separately is an idea worth borrowing, at the cost of more tokens spent on maintaining the profile.
3.25
User tastes change over time, and RNN-based sequence models tend to focus on the latest behaviors and lose long-term interests. This paper, published in TKDE in 2023, handles this with two attention modules. The first uses user embeddings to weight the items in a recent time window and summarizes them into the current context. The second projects user embeddings into several spaces, each representing one aspect of taste, and the current context sets the weight of each aspect. The paper reports a 3.25% gain in HR@10 over the strongest baseline, but the gain is uneven, and on Pinterest it beats SASRec by only 0.32%. Ablations show the structure must be tuned per dataset: stacking more layers has little effect, and adding positional encoding is slightly worse on Pinterest. Evaluation ranks each test item only against randomly sampled negatives, which is worth remembering when comparing these numbers with full-catalog ranking results.
3.15
When multi-objective recommendation weights metrics such as relevance and member disagreement into one reward, variance can differ widely across candidates, so whether variance-independent exploration like UCB1 or variance-aware exploration like UCB-V wins depends on the data, and the variance regime is not known in advance. UCB-min is simple: in every round it computes both exploration bonuses for each candidate and keeps the smaller one. Each candidate chooses for itself, with low-variance candidates mostly using the variance-aware bonus and high-variance candidates mostly falling back to the variance-independent one. On MovieLens group recommendation, variance is large, UCB-V2 over-explores and ends with cumulative regret of 406, while UCB-min follows UCB1 and stops at 233. On low-variance Steam, it instead keeps up with UCB-V2. It needs no prior knowledge of the variance regime and no reward normalization, only one extra bonus. Note that UCB-minV2, the variant used in most experiments, currently has no theoretical guarantee, and UCB1-Tuned reaches lower regret on Steam. For online exploration, this is a default option that takes very little change.