PERSONALIZATION · 2026-11-06
Personalization — Paper Daily, 2026-11-06
An unseen item at the bottom of a page is not necessarily disliked, and a once-preferred answer may not stay preferred. ICE incorporates browsing depth into page evaluation, while PEARL restores exploration using new feedback. FocalSwap holds targets and candidates fixed while changing historical search behaviour: conclusions about a signal’s value also depend on what replaces it.
4.80
LEAD STORY ·
Meta
MetaExpanding a recommender changes interacting embedding dimensions, interaction depth and width; sweeping them separately quickly exhausts training budgets. Meta’s Axia ties the hidden widths of multiple interaction modules together, controlling embedding dimension and interaction depth separately. It profiles inference FLOPs, selects configurations using probabilistic quality and cost predictions, and updates predictions after training. Both adaptive and random search receive 20 full training runs in the main comparison of observed quality–cost frontiers. Across two deployment cycles and 34 model iterations, normalised cross-entropy improved by an average 0.66%, an observational comparison with prior deployments. A small set of experiments can provide evidence for scaling decisions.
4.80
Ranking already scores many candidates that never reach the screen. This work reuses the high-scoring part of those results to help users encounter additional categories. It accumulates these candidates per user over a month, keeps categories absent from recent exposure, and adds the results to recall.
4.80
AlibabaA displayed page does not guarantee that users reach its tail; adding item CTRs can overestimate the page’s value. ICE, a whole-page re-ranking evaluator, predicts browsing depth, multiplies each position’s impression probability by its conditional click probability and scores candidate pages. A two-week Taobao experiment kept the sequence generator unchanged; the complete framework increased clicks by 0.70% and commercial revenue by 3.67%. The intervention changes how a page is evaluated, rather than generating more possible orders.
4.45
Worse recommendations after removing search history do not establish that its content provided the gain. FocalSwap fixes pre-search recommendation history, candidates and the model, then compares masking all search inputs with replacing them using real searches matched by category and behavioural profile. On 66139 shared UniSAR examples, masking supported a positive contribution from the original search state; matched replacements did not. Primary tests changing only the latest search also established no positive contribution. Attribution depends on the reference input, making observed-search replacement a useful check alongside masking ablations in search-augmented recommendation.
4.40
PinterestTraining longer on the same recommendation data can hurt because of the embedding table rather than the entire model. Extending Pinterest Homefeed training from one epoch to two reduced its offline save metric by nearly 6%; freezing ID embeddings eliminated the regression. The team then trained the surrounding network first and enabled embedding adaptation later. Combined with systems changes such as overlapping lookup communication and computation, production deployment improved online saves by 1.09%. Embedding tables and other network components converge at different rates, so they need not share a single update schedule.
4.15
The same user–item interaction table can yield different recommendations when solved through item or user relationships. This paper replaces EASE’s item-to-item linear weights with user-to-user weights and provides a hybrid closed-form solution: mix interaction records using user-to-user weights, then score candidates using item-to-item weights. On MovieLens, the hybrid raises NDCG@100 from EASE’s 40.95 to 43.51; on Gowalla, both new variants perform worse. Gains depend on relationship structure, while a user-to-user matrix also makes memory and inversion costs scale with the user population.
4.15
A frozen tokenizer can assign two users’ profiles the same token sequence, leaving the recommender unable to distinguish them through those tokens alone. At each step, reference vectors represent groups: the nearest vector to the remaining profile information supplies its token, and the unrepresented remainder is encoded next. OptST freezes the original encoder and reference vectors, learning a low-rank correction matrix for each step. Multiplying each reference vector by that matrix and adding the result to the original vector yields the effective assignment position; recommendation losses train the matrix. On Games offline, the complete system raised AUC from the frozen-tokenizer variant’s 0.7469 to 0.7532.
3.75
After preferences change, updating a scoring model with new feedback alone may be slow to find better responses. PEARL uses a small selector for LLM responses: besides predicted ratings, candidates receive an uncertainty-based exploration bonus, weighted more heavily when a user’s recent observed ratings differ more from predictions. After PRISM tests artificially turned half the users’ high response ratings into low ratings and their low ratings into high ratings, mean reward at epoch 10 exceeded an equally online-updated model choosing only the highest predicted rating by 0.040. LLM weights stayed fixed. This supports active exploration under controlled drift, not online evidence of naturally changing preferences.
3.45
Asset recommendation must consider more than what users will buy: popular assets can move together. DSER first identifies investors with strong historical returns, ranking assets by their purchase counts weighted by returns. It preserves the ranking’s front and replaces trailing entries sequentially with the asset having the lowest return correlation to the current basket. The provider specifies the replacement count; no model training is required. In the main FAR-Trans experiment, benefits primarily came from higher-return replacements rather than lower basket volatility. With four replacements over six non-overlapping half-year windows, Sharpe improved only four times, while relevance to users’ actual purchases fell every time. Benefits depended on market phase; relevance costs were more stable.
3.10
What users usually like and what they want in the current session need not coincide. Tasteply lets users organise liked posts by purpose, representing each group as a point on a map. Moving closer gives that group more influence on recommendations, while a separate slider adjusts the remaining stable preference signal. In an exploratory study with 5 researchers, this mixing suited casual exploration better; some preferred fixing one group when seeking specific information. The interface demonstrates how to expose preference controls, not evidence that recommendation quality has improved.