PERSONALIZATION · 2026-10-27
Personalization — Paper Daily, 2026-10-27
The personalization track has 8 papers today, 4 with deep dives. Meituan's HGE estimates a store ad's expected sales by splitting it into whether the user orders, which voucher, and how many, and is fully deployed. JD.com's iFusion uses a diffusion model to write recent sessions into long-term interest step by step. MARS finds that long-history caches forget recent behavior and replaces one memory with several that forget at different speeds. Another paper synthesizes negatives with a diffusion model to avoid false negatives. Teams working on ad ranking or long sequences can start with the first three.
4.70
LEAD STORY ·
MeituanLocal-services and live-commerce ads are ranked and billed per store or livestream room, but users actually buy one voucher or one product inside it, so directly predicting how much a click will sell gives a noisy label that is hard to learn. This Meituan paper splits a store's expected sales into parts: first whether the user orders at all, then each voucher's purchase probability and quantity, multiplied by its price and summed. A store-level regression tower covers samples whose voucher-level signal is too sparse. Each sub-task gets a far more stable label. In a 13-day online A/B test, ad revenue per search rose 1.42%, and the method is fully deployed.
4.55
CTR models usually encode long-term and short-term behavior separately, then concatenate or gate them into one interest vector, so one-off noise in recent behavior, such as stocking up during a sale or buying for someone else, gets written straight into long-term preference. This JD.com ICLR 2026 paper treats long-term interest as a draft: it adds noise, then denoises step by step conditioned on the latest sessions to produce the fused interest vector. Each session splits in two: averaging keeps repeated preferences, attention catches sudden interests, and the stable path gets about twice the weight. In a 7-day online A/B test, CTR rose 2.44% while serving latency grew only about 0.3%.
4.25
To save compute, long-history recommenders compress a user's thousands of actions into a small set of cached vectors. This Duke and Meta paper finds that such caches barely remember a user's latest actions while very old ones survive best: recent interests get buried under old ones. MARS writes the history into four memory tracks at once. With each new action, every track fades old content at its own rate: one halves it within a few actions, another only after hundreds. When reading, a small router scores the tracks and each cached vector reads only the top two. On KuaiRec, AUC rises from 0.793 to 0.804, with about 2% more scoring latency.
4.00
Recommenders train on negatives drawn from items a user never clicked. Random picks teach little, while picking the items the model scores highest often turns things the user would like but never saw into false negatives. This AAAI 2026 paper from Harbin Institute of Technology and others synthesizes negatives with a diffusion model: it noises a liked item's vector step by step, learns to restore it, and takes a half-restored vector as the negative. Which step is used depends on the model's score for the item: a higher score means it takes more noise to become unrecognizable, so a noisier step is chosen. A synthetic negative matches no real item, so it never penalizes what the user wants. On weaker models such as NGCF, Recall@10 rises by up to 69%; on MixGCF and others that already mine hard negatives, the gain is only 1% to 5%.
4.00
Multi-behavior recommenders use auxiliary behaviors such as views, favorites and add-to-cart to make up for sparse target behavior like purchases, but many auxiliary interactions have nothing to do with the target and pull user and item embeddings off course. GCIB works on two levels. On the graph, it learns to drop edges from the auxiliary graph so that what remains reflects the target-behavior graph while carrying as little else from the original as possible; the dropped edges are mostly irrelevant interactions. On features, it contrasts the denoised auxiliary features with the target features as two views. Auxiliary behavior fills gaps without being taken wholesale. The paper reports beating existing methods on several real-world datasets.
3.75
In federated recommendation each device sees only a fragment of a user's behavior, yet most methods still represent a user with one fixed embedding, as if preferences were settled and precisely knowable. With little data, that estimate is brittle. This paper represents a user's preference as a distribution instead: a diffusion model generates several preference vectors that all fit the user's observed behavior, and the user is represented by that set rather than a single point. Evidence is not squeezed into a single point too early. On federated recommendation benchmarks, the paper reports steadier rankings, more robustness under ambiguous feedback, and consistent gains over baselines.
3.45
Recommendation and reinforcement learning often use an ensemble to estimate uncertainty: each decision samples one member, and disagreement among members signals uncertainty. The sampling is usually fixed and uniform, regardless of which member has been doing well. This paper proposes Bayesian Ensemble (BE), which treats choosing a member as a bandit problem of its own: it updates each member's selection probability by Bayesian inference on the rewards it observes. BE is a thin layer on top of existing ensembles. The authors attach it to bandit learning and to DQN, and report better results and good efficiency in both synthetic and real-world environments.
3.05
The same user may want a bold collaborator when brainstorming with an AI assistant and a strict critic when editing, so one fixed persona rarely fits. This paper lets the assistant switch personas from light feedback: the user simply prefers one of two answers, and the assistant picks the best of several hundred pre-built personas. Each round, the current top persona defends its spot against a challenger. A neural model estimates how far each candidate trails the champion, a light statistical model estimates how much one comparison would cut uncertainty about the ranking, and the candidate with the best ratio of the two is chosen. A new query starts from the win-loss records of similar earlier queries. Tested on seven personalized generation benchmarks with ten comparisons per query and two simulated judges per benchmark, it ranks first in 13 of those 14 results.
RecSys 2026
ICLR 2026
AAAI 2026
ICML 2026