PERSONALIZATION · 2026-10-25
Personalization — Paper Daily, 2026-10-25
Today's personalization theme is making use of what users actually saw. LinkedIn puts people who were shown but not opened into its training objective, a carousel bandit updates only on positions that were really browsed, and Kuaishou has several agents write ad scripts per user. Other papers handle recommendation under hard constraints, including a shared household fridge, medication sets and ad auctions inside assistants.
4.08
LEAD STORY ·
LinkedInLinkedIn On LinkedIn, a software engineer who starts looking for a job usually wants to open the profiles of recruiters, hiring managers and HR, not other engineers. In professional networking, who should be recommended is often complementary, not similar. Yet most recommender systems assume that the more alike two profiles are, the more one should be recommended, so they keep showing him his peers. PAIR changes the training goal: instead of learning who resembles whom, it learns from real logs which of the people shown the user clicked. In a live comparison experiment, clicks on recommendations rose 7.5%, paid-member renewals rose 2.1%, and pages did not get slower.
4.08
A recommender places a film in a row of items, and the user does not click it. Should the system record that the user doesn't like it? If the user never swiped far enough to see it, that record is wrong. Most online learning-to-rank algorithms cannot tell these two cases apart and have to guess: first estimate roughly how many people see each position, then scale click counts by that proportion. In a carousel interface that scrolls sideways, no guessing is needed, because the client can already record the furthest position the user reached this time. The authors use this signal in ranking algorithms and learn only from positions the user actually saw. In a simulated environment built from real users' eye-tracking data, the new algorithm does as well as a rival that knows in advance how likely each position is to be seen, while it does not need those probabilities itself.
3.75
In cross-domain recommendation, users differ widely in how much their interests overlap between two domains, and deciding per user how much to transfer and in which direction works better than one uniform transfer. On the Movie, Book, Sport and Cloth domains of Amazon review data, it beats the strongest baseline on hit rate and NDCG by 3.91% to 16.62%. Each user has a semantic representation in each domain, and transfer moves the source-domain representation along a quadratic Bezier curve toward the target-domain one. The curve's control point is computed by a small network from the user profiles an LLM writes for both domains. How far it moves depends on interest overlap: items are clustered into interest prototypes, and the more alike the user's distributions over prototypes are in the two domains, the further it goes. The curve is split into equal steps. At each step, prototypes are averaged with weights from their similarity to the current position, and a correction network reads both to output a correction; the new position is the next curve point plus all corrections so far, and the next step starts there.
3.73
JJiangxi University of Software Professional Technology
The purchase signal of views, favourites and add-to-carts fades with time, and something added to the cart last week without buying may no longer be wanted today. Estimating how much signal each behavior still carries, from its type and age, and down-weighting stale ones in ranking clearly improves purchase prediction. BLMRec gives each behavior an initial strength and a decay rate, with the remaining signal decaying exponentially over time. Training labels come from several windows from 1 to 30 days after the behavior. A behavior still counts as valid if, within the window, the user moves a step toward ordering the same item, such as view then cart or cart then buy. Buying an item of the same category or one often bought with it also counts; otherwise it counts as stalled. Each window gives its own label, the model's estimate at each window end must match it, and the losses are summed. In a harder test packed with items that were browsed or carted but never bought, its Recall@20 is 8.8% higher than the baseline RMBRec.
3.63
KuaishouPersonalizing video ads usually just means picking the best-fitting one from a library of ready-made creatives; the creatives themselves don't differ by person. Kuaishou's MAPS instead writes an ad script and generates a video for each user on the spot: it first infers preferences from viewing history, then writes the script using the product information. The script-writing model is trained with reinforcement learning that rewards scripts that both fit the user and describe the product accurately; after the ad is shown, the user profile is corrected based on the user's feedback. In an online A/B test, compared with the method that retrieves creatives, average revenue per user rose 9.85%.
3.55
To evaluate shopping recommendation policies offline, the policy differences a user simulator measures must match online A/B results. Most simulators fail: checked against a real online ranking-policy comparison, other dedicated simulators measure at most 0.297 of the real click-through lift. UserArena builds such a test bed from real e-commerce logs and also offers a way to train the simulator. An LLM first reads a user's full shopping record and summarises their shopping intent, such as stable preferences and what they compare products on, and only a frozen teacher model sees this summary. The student model sees only the current page and earlier actions and generates its own next action; the teacher, using the intent summary, gives a target distribution for every token the student generates, and the student learns to match it. Rerunning the two policies with this simulator gives a lift of 0.713 of the real online lift, the closest of all simulators.
3.55
A family shares one fridge but each member has their own dietary restrictions, so recommending for each person separately easily has them competing for the same ingredients, or suggests a dish someone can't eat. The authors treat it as an allocation problem. A learned model first scores every food for every member from taste, pantry stock and dietary rules. Foods are then handed out round by round: each round one member takes their highest-scoring remaining food, and the member chosen is the one who raises the product of the family's satisfaction the most, so no one eats especially badly. Finally, hard rules filter the result so that no one's restrictions are violated. On 100 four-person households, the satisfaction gap between members was 42% smaller than with resolving conflicts only when they arise.
3.45
Inserting ads into an LLM answer is different from search ad slots: once users click the first ad they stop reading, so earlier ads steal clicks from later ones, and advertisers may therefore not bid truthfully. Instead of reordering the ads, the authors randomly omit some ads, omitting earlier ones more often. Each ad's keep probability is one over one plus the sum of the standalone click rates of that ad and the selected ads after it. For two ads that each have a standalone click rate of one half, keeping the first half the time and the second two thirds of the time gives both a one-in-four chance of being clicked. On that basis they pick the combination with the largest total click value and charge by a VCG-style rule, so bidding truthfully pays best. The paper proves that this gets at least half of the best total value achievable when the order can be freely rearranged.
3.45
Medication recommendation has a hard constraint: a set of drugs can't contain too many conflicting pairs, and conflicts depend on the whole set, not on any single drug. The usual approach picks drugs by score, then deletes conflicting drugs one by one and adds substitutes. The authors keep the same model's scores fixed and replace this step with a single solve: within the conflict limit, directly find the whole set with the highest total score. Under every level of constraint from loose to strict, solving for the whole set overlaps more with what doctors actually prescribed, and the stricter the constraint, the bigger the advantage. How scores are turned into the final drug list itself affects the result.
3.20
Fairness adjustments in recommendation usually fix one objective at training time, or apply one fixed reranking rule after ranking, yet the candidates and exposure differ on every request. FAIRGate has a language model pick one policy from a fixed menu on each request, such as guaranteeing a floor for long-tail movies or balancing exposure across genres. The list is actually changed by a deterministic program, which then checks whether the policy is met and whether the relevance loss is within budget. On MovieLens, applying policies in a fixed order satisfies the chosen policy for only 9.10% of requests, versus 60.70% for FAIRGate.