PERSONALIZATION · 2026-11-10

Personalization — Paper Daily, 2026-11-10

Besides the already published Kuaishou full-screen ranking paper, today's personalization track adds two newly surfaced papers: one on signal form in persona simulation for cold-start ads, and one on personalizing only the order of explanation sentences.

4.55
LEAD STORY · Kuaishou

1. STCRank: Spatio-temporal Collaborative Ranking for Interactive Recommender System at Kuaishou E-shop

Kuaishou

In a full-screen e-commerce feed where each swipe down brings the next item, the two ranking objectives of whether the user buys and whether they keep swiping fight each other. Kuaishou E-shop's data show that after a purchase, users leave right away without swiping 61.3% of the time. With the usual labels, every one of those purchases becomes a negative sample for the swipe objective, so the learned swipe score ends up pushing down the items that sell. The authors drop purchase-then-exit samples from the swipe negatives. Ranking also stops picking the top-scoring item one screen at a time. Instead it scores different orderings of the next few items: each item's value is multiplied by the probability that the user swipes far enough to reach it, and the ordering with the highest total wins. Adding this look-ahead ranking to the full-screen feed alone raised online purchases by 2.10%. Teams running immersive feeds should check their objectives for conflicting labels first.

3.85

3. Not Your Diary, Just Your Topic: Grading the Wrong-User Control

A common personalization check replaces a user's own history with someone else's and watches accuracy drop. That swap also changes the topics in the history, so the drop may not mean the system recognized the person. On a frozen language model and LaMP-2 news classification, a user's own history beats using no history by 2.3 points. Keeping only each article's category label and dropping the text already reproduces 86% of the gap between own history and a random user's history. Synthetic histories are stitched from other users' articles that are closest in text to the target profile. When each slot must also match the category label at the same position in the original history, a user's own history beats this synthetic history by only 0.9 points. On unseen users, a user's own history beats similar-user histories by only 0.65 points, which is not yet strong evidence. Personalization tests should look at unseen users and compare against content-matched histories.

3.50

4. What Matters for LLM Persona Simulation: Signal Form in Cold-Start E-Commerce Advertising

Before an ad campaign launches there are no historical logs, so cold-start performance is hard to estimate. When large language models simulate persona-conditioned user responses, what matters is which signals go into the prompt, not how richly the persona is written. On real Google Ads e-commerce campaign data across five demographic groups, the method anchors on basic demographics without noisy extra attributes. The prompt notes that users often ignore banner ads even when the product seems relevant, instead of inserting numeric click-through rates. Mean absolute error between predicted and observed clicks across ad groups is 24.9 clicks. That is 51.3% lower than the best tested non-LLM baseline and 77.3% lower than an LLM setup that uses explicit click-through-rate anchors. Minimal demographic descriptions also produce lower mean error than richer persona configurations.

3.45

5. PAGSL: Path-Augmenting Graph Structure Learning for explainable scholar recommendation

Scholar recommendations often come with shallow, generic explanations, and highly connected nodes in the graph can pull those explanations off course. This work builds a subgraph for each scholar and that scholar's research attributes, adds type-specific edges between scholars who share an attribute, then extracts relational paths between a user and a candidate scholar for scoring. It masks lower-scoring paths so the remaining explanation is shorter and more informative. On the Scholat, AMiner, and ResearchGate academic datasets, recommendation F1 rises 6.31% and accuracy rises 6.29% over the strongest baseline. Ablations show that both heterogeneous convolution and path representation learning contribute to the gains.

3.05

6. Learning from How Users Interpret: Personalizing Recommendation Explanation Order

Personalized explanations usually change their content or presentation. This study asks whether regularities in how users previously interpreted text can be used to reorder explanation sentences while leaving the wording fixed. In a within-participant study with 122 people, prior interpretations were summarized by what information they selected, what they treated as more salient, and how they organized relations. Sentences that matched the more salient, more often selected information in that profile were placed earlier. After reading sentences ordered by their own profile, participants answered 86.5% of objective comprehension questions correctly. The same sentences ordered by another person's profile scored 66.0%, and a language model that ordered them from the reader's own profile scored 73.0%. There was no reliable overall effect on whether people accepted the recommendation.