PERSONALIZATION · 2026-10-29
Personalization — Paper Daily, 2026-10-29
The personalization track has 10 papers today, 3 with deep dives. The common thread is that user interest is neither single nor fixed. COMPOTE uses the training objective to make several prediction heads each own a region of items, so multi-interest vectors do not collapse into one. TRIP treats how long before this impression a behavior happened as its real position. BehaveRec shows that long-term history helps only users with enough of it. On cold start and engineering, Roblox allocates traffic to new creatives with hierarchical Bayes plus random sampling, and Meta Andromeda brings near-ranking-quality models into ads retrieval. Also included are handling cross-day drift of daily embeddings, compression choices that keep rare-item recall under a storage budget, and one paper each on multimodal complementary recommendation and multi-scenario recommendation. Retrieval teams can start with COMPOTE and Andromeda, and ranking teams with TRIP.
4.85
AI-generated content makes it easy for advertisers to produce many creatives, but each gets very few impressions. This Roblox paper adds a creative pre-ranking layer between recall and ad ranking. The historical play rate μ of a new creative's campaign, game and target audience serves as the prior mean, and the prior strength κ is computed in closed form from the variance of true differences among sibling creatives: the smaller the differences, the larger κ. Adding the creative's own plays and non-plays gives a Beta posterior, and one random score drawn from it is used for ranking. Borrowed experience dominates while data is scarce and gives way to the creative's own record as data grows. On replayed live traffic, it lifts the ranker's predicted play-through rate by 8.14%, against 2.80% for a random shortlist of the same size, the pipeline gets only 0.2% slower, and it now serves production.
4.50
In unified Transformer CTR models like OneTrans, the authors tried six positional encodings on Kuaishou's KuaiRand data, and none beat using no positional encoding at all. Building on this, the Wuhan University and Tencent paper proposes TRIP (the authors' name for the whole method). Each behavior encodes both its rank in the sequence and how long before this prediction it happened. Behaviors are then grouped by that time distance: the seconds before this prediction are put on a base-2 log scale and rounded, so each group spans twice the time of the previous one. Behaviors in the same group see each other, while across groups a behavior sees only earlier groups. For the same history, change the impression time and every behavior's distance from now changes, something language-model positional encodings never had to handle. On four public datasets AUC improves over the strongest baseline by 1.47% relative on average, and by up to 4.45% on the Tencent ad dataset TAAC (a public Tencent ad-click sequence dataset).
4.20
A common problem in multi-interest retrieval is that the interest vectors end up pointing the same way. Criteo and Sorbonne's COMPOTE (the authors' name, short for collapse-resistant multiple-hypothesis learning) puts several prediction heads on a shared causal Transformer. Each sample splits its gradient across heads in proportion to the probability each gives the true item, raised to the power 1/τ: at high temperature τ all heads learn together, and as τ falls toward 0 only the head that guessed best is updated. At inference each item is ranked by its highest cosine similarity across heads. Multi-interest retrieval often lacks not more vectors but a training objective that makes the vectors divide the work. At the same parameter count, heads trained by averaging or with a routing gate point in exactly the same direction, while COMPOTE's heads are clearly apart. It wins 11 of 12 metrics on three Amazon datasets.
4.20
Are long-term history and community signals always useful? BehaveRec's answer is that it depends on how much history a user has. It combines three signals: a causal Transformer over the current session, a bidirectional encoder over 90 days of history, and a user graph built from behavioral similarity, with a gate deciding per user how much of each to use. On industrial data with over ten million users, Recall@5 is 3.2% higher than SASRec, and a larger SASRec still trails, so the gain is not about model size. The more useful result is the stratification: users with little history see almost no gain, and the gain grows steadily with history, with essentially none below about 10 interactions, from which the authors derive a deployment threshold of about 20 interactions. These conclusions hold for category-level intent prediction; the item-level check was a single preliminary run with mixed direction.
4.20
In sequential recommendation, the part that represents user preference and the history used for scoring often look at different evidence. RELAY has one learned query attend over the history to form a preference representation, then reuses the same attention weights for two memories. Direct history boosts items the user already engaged with. Transition memory holds training-set counts of item b following item a, on a log scale, and each past item boosts its frequent successors in proportion to its attention weight. Preference representation and scoring use the same selected history, so the two sides stay in step. Each candidate has gates deciding which channel to trust more, and ranking takes one forward pass. On six datasets, NDCG@5 is 17.88% higher on average than the previous best model, with 65% fewer predictor parameters and about 1.75 times faster inference.
4.20
Multimodal recommenders often build item graphs from content similarity, so similar substitutes connect tightly while complements that are bought together but look different (a camera and a memory card) stay isolated. MAGE works in two steps. Offline, it splits each item's co-purchased items by text similarity into substitutes and complements. The text vectors of the item and its substitutes form a matrix for SVD; directions with larger singular values represent substitutes more and are removed more strongly. Each complement has its substitute-direction parts removed by these weights, and the averaged residual is added to the item's text vector as a complement direction, with a capped magnitude. Online, image features propagate only over the similar-item graph, augmented text features over the co-purchase graph, and a gate fuses them. Pulling complements closer geometrically first keeps propagation from circling only among similar items. On four Amazon datasets, Recall@20 is about 2.8% to 6.5% higher than the strongest baseline.
4.10
Many teams retrain item co-occurrence embeddings daily and feed them to CTR models, but the same item's vector changes direction and length every day, so multi-day training data actually mixes several versions. This industrial empirical study compares many ways to handle it and reaches three conclusions. First, orthogonally aligning each new day's vectors to the previous day makes vectors more consistent across days, and GAUC rises 2.6%. Second, warm-starting from the previous day's vectors works well, but vector norms keep growing; multiplying the whole embedding table by one factor each day, so the mean item vector length returns to a reference value, holds them down and gives the highest GAUC gain, 4.0%. Third, the dimension-reducing encoder inside the downstream CTR model itself absorbs the remaining drift, so upstream stabilization and downstream learning complement each other. All numbers are percentages relative to a baseline; raw values are not published.
4.05
When multi-scenario recommenders add text and image information, they usually feed the vectors as ordinary features into the original parameter-sharing structure. DuTri's authors show with controlled experiments that this is not optimal, so they split the model in two. The collaborative branch applies CP decomposition to the three-way tensor of user, scenario and item interactions, yielding user, scenario and item factors. User and item factors are shared across scenarios and carry what they have in common; scenario factors modulate their interaction and carry how each scenario differs. Each set of factors is then quantized with its own residual codebooks and reconstructed as a structured preference representation. The multimodal branch uses the user and scenario preferences from the collaborative branch to decide how to fuse text and images, and contrastive learning aligns the two branches' item representations. Collaborative and multimodal signals need different modeling, so learning them separately and letting preferences guide fusion fits better than mixing them into one structure. On two real datasets, Ant's AntM2C and Amazon, it beats existing multimodal and multi-scenario baselines overall.
4.05
MetaMeta's next-generation ads retrieval engine Andromeda brings near-ranking-quality models to retrieval, scoring every eligible ad. The new version replaces the single user tower with several specialized user towers as experts, and encodes the ad tower with real-time request context. An attention-based gate in the interaction layer activates only the most relevant experts per ad, cutting inference compute by about 300 times versus the 2024 version while approaching ranking-model quality. The bottleneck shifted from compute to memory bandwidth and network transfer, so they fused embedding lookup, the interaction model and scoring for each user-ad pair into one GPU kernel, keeping intermediate results in registers and shared memory instead of writing them to GPU memory. They also put targeting-rule matching, scoring and top-K selection on the same GPU machine, returning only the final results. This lets the system score far more ads within the same latency; since 2025 it has lifted conversions on Facebook by 2.4%.
3.95
With a fixed embedding storage budget, should you use a small full-precision model, a distilled student, or a higher-dimensional low-precision model? This paper compares several options with exactly the same byte count on six datasets, focusing on recall of rare items. The result is consistent. Factorizing a large teacher's user-by-item score matrix into fairly high-dimensional user and item vectors, stored at only 4 bits per number, uses 13% of the teacher's storage and keeps rare-item recall roughly on par with the teacher. The low-dimensional full-precision option at the same byte count has the lowest rare-item recall in every setting. Spending bytes on dimensions rather than precision better preserves rare items. The authors also find that judging only by overlap with the teacher's hits picks the wrong best option in three of seven settings, because compressed models can find rare items the teacher misses, so rare-item recall should be measured directly.
NeurIPS 2026