PERSONALIZATION · 2026-11-03
Personalization — Paper Daily, 2026-11-03
Today's recommendation track has two engineering papers from production systems. Google's Light Heads let YouTube's ranking models take on new targets without retraining, cutting joint experiment time by more than half; Alibaba's RecIS arranges sparse feature processing and dense computation into one pipeline so GPUs wait less. BrowseFlow, from Taobao and Tsinghua, takes a different angle on user simulation: it learns the range of browsing outcomes a user might produce on the same feed instead of copying one log. Several other papers look at side effects of recommendation, including how the feedback loop in location recommendation leaves restaurants empty and makes people of different incomes meet less, and how to remove one user interaction's influence when the model is hosted by someone else.
4.20
LEAD STORY ·
GoogleGoogle The bigger a multi-task ranking model gets, the slower it becomes to add a new prediction target. YouTube's answer keeps the new target out of the main model and makes it a small prediction head on top of the shared layers (the paper calls it a Light Head). It only reads the shared layers' output, sends no gradient down, and is reset at the start of every training run, and every ranking model reads it from the same config. With the framework, a full experiment cycle covering the ranker and a downstream model shrank from 24 days to 11. The cost is an AUC about 0.01 lower than a fully trained head on data-rich tasks such as clicks.
4.10
Alibaba
Tsinghua UniversityShown the same feed, the same user might scroll to item 20 and click two things this time, and leave at item 8 the next; the log keeps only one of these. BrowseFlow argues that a user simulator should learn the distribution of such outcomes rather than copy that one log. It splits a browsing session into two things, how deep the user scrolls and at which positions they click or like, and generates those key events directly instead of predicting a long run of scroll actions step by step. This matters because in Taobao browsing data only 5.48% of actions are positive feedback such as clicks. On three datasets, its generated sessions come closest to real ones both in overall distribution and in click-count distribution.
3.90
In many apps the recommendation model is hosted by an outside provider, and the operator can only call an API for scores without touching the model's parameters. When a user then asks to remove a mis-click or an outdated interaction, existing recommendation unlearning methods do not apply, because they retrain or fine-tune the model. CEU leaves the model alone and changes only its output. The model provider precomputes how much each past interaction contributes to each recommendation, that is, how much the item's score would drop without that interaction, and serves it through the API. On a removal request, the operator subtracts that contribution from the affected items' scores and re-ranks. On three public datasets it forgets about as well as full retraining. One removal request takes at most 6.37 seconds, while full retraining takes from minutes to over an hour.
3.65
If a recommender ignores which group a user belongs to and serves the average taste of all users, minority groups keep losing out. The authors prove that whenever the overall favourite set of items is not the best one for some group, this policy's cumulative gap to each group's best recommendations grows linearly with time. The hard part is that groups overlap, so one person can belong to several at once. The authors offer two routes. One learns preferences separately for every combination of groups, which has a guarantee but leaves each combination with little data once there are many. The other learns per broad group first, and a selector uses past feedback to decide which of its groups' recommendations each combination follows at the moment. On two datasets, the second route wins when combinations share their broad groups' taste, and the first wins when they differ.
3.45
Location recommendations change where people go to eat, and the platform retrains on the visits it steered, closing a loop. Offline accuracy cannot measure this loop's long-term effect, so the authors ran long-term simulations on real mobility data from four US metro areas, accounting for travel distance and venue capacity. Under every algorithm and in every city, at least a fifth of food venues end up with no visitors, and more than half of that loss comes from repeated retraining. The paper measures this as income segregation: the more the people someone meets at venues lean toward one income band, the higher it is. More accurate algorithms give higher segregation, and the most accurate is 15% to 21% above residents' real visits. Re-ranking on top of the existing recommender to favour less popular venues eases both problems.
3.45
In recommender training, the first half handles thousands of feature fields and trillion-scale embedding tables with small, random operations, while the second half is an ever larger Transformer doing dense, regular computation. With both on the same GPUs, the GPUs often sit waiting for data. Alibaba's RecIS optimizes the whole path from raw features to dense computation as one unit. It merges scattered feature transforms into larger operators and spreads embedding lookups evenly across GPUs by hashing, removing duplicate transfers. It also prepares the inputs for upcoming steps while the current step runs its dense computation. On industrial recommender models, throughput reaches up to 1.15 times that of TorchRec and 2.62 times that of PICASSO.
3.15
A newly launched question platform has question texts but no answer logs to train on yet. Can a model trained on other platforms predict whether a student will get the next question right? Carrying question-text semantics over as features works poorly, because question styles differ a lot across platforms. TIZSKT uses text only as an index for finding similar questions. The model itself learns only how right and wrong answers evolve, without question IDs or text. At prediction time, the questions this student just answered on the new platform, and whether they got them right, are known. The model uses the text to pick those resembling the current question, summarizes how the student did on them, and combines that with its learned answering state to compute the chance of a correct answer. The new platform needs no training, and on three public datasets it has the highest mean AUC.
3.05
A recommender has to serve users, content creators and the platform, plus developers, regulators and society at large, and each party has several goals. Multi-stakeholder, multi-objective recommendation has been studied a fair amount, but few ask whether such systems can be trusted. This position paper argues that trust is not only about fewer biases and privacy problems; it also means actively respecting human values and handling conflicts between parties' goals. Its example: a creator promoting healthy-diet videos is harmless in itself, but pushing them to a user with an eating disorder can do harm. The paper has no new method or experiments and lists four open problems: whom to treat fairly under multiple objectives, how to turn human values into measurable goals, the computing cost of more objectives, and EU rules requiring platforms to explain what drives their recommendations.
RecSys 2026 Workshop
UCAS