PERSONALIZATION · 2026-11-02
Personalization — Paper Daily, 2026-11-02
Today's personalization track is about better inputs for ranking models and more trustworthy offline evaluation. Two long-sequence papers from Meta stand out: one gives history embeddings the user's own perspective, the other lets reading more history and stacking more layers be tuned separately. A third asks how to pick the best model offline in ad systems where impression probabilities are never logged. Also in the issue are a traffic simulator for supporting new creators, an online experiment on allocating advertiser subsidies, and a reminder that classical PageRank can match newer models on temporal link prediction.
4.05
The same cooking video is a pastime for a casual viewer and source material for a food creator. Yet a recommender gives it the same starting embedding in both users' history sequences. PRISM replaces each history event with a new vector. That vector expresses how this user related to this item at that moment. It reuses the intermediate states of every layer that the upstream ranking model already computed while scoring, and an autoencoder compresses them into one embedding per scored user–item pair. These are stored, and later sequence models read them in time order. On a billion-user recommendation surface at Meta, downstream NE drops by 0.15% on average. At a similar history length, VISTA, another approach that caches history summaries, drops by only 0.07%. The gain also grows with longer histories. The input to a history sequence can itself be personalized.
4.05
Long-sequence recommenders can improve in two ways: read more history, or stack more layers. Current models tie the two together. A sequence encoder pays for every extra event in every layer, while models that first select relevant events select only once, so later layers never see what was left out. LWR keeps the history in storage that every layer can query. Each candidate item carries a small fixed set of vectors, a kind of scratchpad: some are produced from the candidate's own features, and the rest are learned in training. Only this scratchpad moves up through the layers. At each layer, the scratchpad's current content produces a few queries that run cross-attention over the full history, and the results interact with ranking features to update the scratchpad. History length and depth can then be tuned separately. With 10K history events, LWR beats HSTU by 0.19% in NE while using 69.6% less forward compute. Reading more and stacking deeper can be scaled separately.
4.05
Before an ad system launches a new ranking model, it must pick the best candidate offline from logs. In multi-stage delivery, impressions pass through auctions, filters and caching, and their selection probabilities are never logged, so IPS cannot be used. The usual substitute scores new policies with a click-predicting reward model, but new policies push impressions toward ads that are rare in the logs, where that model is least accurate. The authors replay the candidates on past requests and train a classifier to tell whether an impression comes from the logs or from a given candidate. For each logged impression, the classifier gives the ratio of its probability under each candidate to its probability under the logs. The average of these ratios over all candidates becomes its weight when training one shared reward model, so impressions that look like what the candidates would show count more. Five candidates in a production CTR ranking system all ran in A/B tests. Picking a winner on each of 50 random log subsamples, the rate of picking the truly best one rises from 0.36 without weighting to 0.40.
3.65
Recommenders allocate traffic by engagement, so big creators keep growing and newcomers struggle to be seen. Testing traffic policies that support newcomers directly in production is costly, so the authors build a simulator of about 42K creators whose makeup and overall metrics are calibrated to the real platform, and compare traffic allocation methods inside it. The best one gives creators a weekly guaranteed view target, and the allocation algorithm uses how far a creator still falls short of it as its reward, so the small creators furthest behind get supported traffic first. With it, 90% of the guaranteed targets are met each week, and small creators' organic reach grows by about 33% per week. The authors also discuss when fairness conclusions drawn from a simulator can be trusted.
3.65
Temporal link prediction guesses who will interact with whom next. Recent methods mostly learn ever more complex temporal representations, yet a classical method can match or beat them. The authors run Personalized PageRank directly on the historical interaction graph, adding only a sliding window or a time decay that shrinks the weight of old edges, with no training at all. Across 14 datasets, the time-decay version ranks first on 9 and second on 4, even on datasets with few repeated edges. The reason is that paths in the history graph leading to new destinations already carry a strong signal. Current benchmarks may overstate the value of complex temporal models, and new methods should be compared against classical baselines like this one.
3.50
Ad platforms give advertisers billing discounts, usually set by operators through tier-based rules. Those rules cannot capture how each account responds to different discount levels. CASH needs to predict how much revenue each account brings at each discount level, and it does so in two ways at once: an LLM that carries business knowledge, and a numerical model learned from historical data. It measures each one's errors on accounts whose outcomes are already known, for each kind of account and discount level, and only a side whose error is small enough is licensed to predict for that level. Each account then uses one side's predictions and only that side's licensed levels. When both sides have licensed levels, a scorer trained on which side erred less in the past picks one; when neither does, the account gets no discount. The system then picks every account's discount level together under the total budget to maximize total revenue. In a one-month online experiment covering 60% of the platform's traffic, with the subsidy rate essentially unchanged, the platform's revenue per account rises by 15.35%, and longer-term metrics such as retention improve as well.
3.40
Link prediction on text-attributed graphs, such as predicting whether two papers will cite each other, is usually done by combining a language model with a GNN, but how best to fuse the two is still unsettled. Bridge-LP takes a simple route. It lets the language model's text embeddings bypass the GNN layers and join the final node embeddings used for scoring, so the loss gradient flows straight back to the language model and training is more stable. And during training it re-encodes text only for the nodes in the current batch of edges, using cached embeddings for all other nodes, which saves a lot of compute. Compared with 40 existing models on 9 datasets, it beats the strongest baseline on Cora by 43.06% in MRR. Neither design is expensive, so teams working on graph-based recommendation can try each one separately.
3.40
E-commerce recommendation has to handle three things at once: long-term collaborative preferences, short-term purchase intent, and new users and items with almost no interactions. DeepCommerceRec learns them with two encoders. One learns structural user-item relations on a heterogeneous graph, and the other learns how interests change over time from behavior sequences. A small MLP reads both vectors plus signals such as interaction count and cold-start flags, and outputs a weight between 0 and 1 for each dimension to mix the two. For a new user, the sequence representation is replaced by a weighted combination of their few most recent items, with more recent ones weighted higher. On three public datasets, Amazon, YooChoose and RetailRocket, the authors report better HR@10, NDCG@10 and MRR than a range of baselines. The method mostly combines existing parts, and the most useful idea for real systems is adjusting the fusion weights by how cold a user or item is.
3.25
A common fairness check for rankings asks whether different user groups get the same average utility, but this can hide problems. The authors simulate a job platform where the historical data miss women's interest in managerial jobs and men's interest in administrative jobs. Women are then under-served on managerial jobs and men on administrative ones, yet the two gaps cancel out and the overall gap stays near zero. The authors therefore require equal utility across groups within each job category. They prove this is equivalent to the average utility in each category staying the same when only the share of each group changes and each group's own makeup stays fixed. They then search mixtures of existing ranking policies for the one with the highest utility that meets the constraint, validating it on MovieLens-1M. Looking only at overall fairness metrics can miss bias inside categories.
3.05
Variables visible in training logs are not always usable at deployment. A clinical decision tool, for example, may be trained with patients' socioeconomic information but barred from using it as an input when deployed, to prevent discrimination. The authors call this case, complete logs but restricted deployment, runtime confounding. In reinforcement learning it is harder, because earlier decisions shape later states and actions, so the bias accumulates over time. Their approach estimates the logging policy's action probabilities and a return model with all variables, while the new policy sees only deployment-time variables. The new policy's probability of the logged action divided by the logging policy's reweights the logs, and the return model corrects the result. Both policy learning and off-policy evaluation build on this estimate, and the evaluation gives valid confidence intervals. They also give a test that measures how much return is lost by not being allowed to use those variables. Features that recommendation and ad systems log but cannot use online for privacy or compliance reasons pose the same kind of problem.