PERSONALIZATION · 2026-10-28
Personalization — Paper Daily, 2026-10-28
The personalization track has 10 papers today, 4 with deep dives. Alibaba Taobao's HiRank carries the behavior sequence through every layer of the ranking model and lifts online orders and GMV. Pinterest does not let reinforcement learning set bids; it only lets it pick one parameter step for the PID controller each day. Another paper warns that recalibrating CTR on just a few hours of fresh data often makes things worse. One more learns sample, feature and embedding-dimension pruning together in one model. Teams working on ranking or ad bidding can start with HiRank and the Pinterest paper.
4.55
LEAD STORY ·
Alibaba /
Renmin University of China /
CASAlibabaRenmin University of ChinaCAS Most production ranking models squeeze the user's behavior sequence into one fixed vector before it reaches the feature-interaction module, so later interaction layers can no longer see which clicks or add-to-carts actually happened. Alibaba Taobao's HiRank keeps the behavior sequence alive through every layer of the backbone, and builds the retrieval query over long-term history from the user's recent behavior rather than from static profile fields such as age and gender. On Taobao homepage CVR prediction it lifts GAUC from 0.7615 to 0.7801 over the strongest baseline at the same parameter count; in an online A/B test, orders per user rose 2.27% and GMV per user rose 2.86%.
4.45
There are three common ways to trim the input side of a CTR model: drop low-quality training samples, switch off useless features, and prune useless embedding dimensions. Earlier methods do one at a time. This paper from Shenzhen University and collaborators argues that all three are the same operation, multiplying one embedding tensor by a mask at different granularities. They also interact: once a feature is switched off entirely, pruning its dimensions no longer matters. On Avazu with a plain DNN, any single selection raises AUC by only about 0.001. Learning all three together in one model, with penalties that make the levels' decisions fit each other, takes AUC from 0.7893 to 0.7930. Between samples and features, for example, the penalty sums each feature embedding's cosine distance to the sample's average embedding, weighted by the feature gates, and multiplies it by the sample weight. A heavily weighted sample that opens wide gates for outlying features is penalized hard.
4.35
In ad systems, recalibrating a CTR model on the last few hours of fresh data is usually treated as a routine step that can only help. Testing it day by day on ten days of Avazu ad logs, this paper finds that a calibrator fitted on the first 6 hours of a day makes calibration error worse about a third of the time, and in the worst case doubles it. Stretching the fitting window to the first 12 hours makes this harm essentially disappear. A uniform sample from that span, cut to as few rows as one hour of data holds, does almost as well as the full span and clearly better than using only the latest hour. What decides calibration quality is how much time the data spans, not how many rows it has.
4.15
Two items being consumed by the same people does not necessarily mean they appeal to the same taste. This University of Antwerp paper proves that even when every user picks items completely independently and at random, large differences in user activity are enough to make lift, cosine and Pearson show positive correlation between item pairs, which the authors call phantom correlation. On MovieLens-1M, 81% of item pairs have co-occurrence counts within a factor of two of what an activity-only model predicts, versus 35% for an independence model that treats all users as equally active. So before computing item similarity, the activity baseline should be removed first; the real taste signal lies in the deviation from that baseline.
4.15
Session-based recommendation has a counterintuitive problem. The longer the session, and the more information it carries, the worse the next click is predicted, whether the encoder is a GRU, a graph network or a Transformer. This paper from the University of Technology Sydney and collaborators traces the cause to the decoder, because common decoders lean on the last few clicks, drowning out earlier but relevant ones. MIDA replaces the decoder. It clusters the session's raw item embeddings by similarity into a few interest groups, Stable interest is a weighted sum of the session's clicks, with clicks in larger clusters weighted more. The average cluster size then feeds a gate whose share mixes the last click with stable interest. Larger average clusters mean a more focused session and give stable interest a bigger share. It can sit behind any encoder. Gains grow with session length: with a GCE-GNN encoder on Yoochoose, the Recall@5 improvement rises from 0.85% on short sessions to 3.43% on sessions of 16 or more clicks. The cost is noticeably slower training.
4.15
The item-by-item weight matrix of a linear recommender such as EASE already contains topics. This Chiang Mai University paper proves that if user behavior follows a latent-factor model, the top K eigenvectors of that matrix span the space of K topics. The authors build TEA on this, so one model both makes recommendations and yields topics directly. On six datasets its topic quality is significantly better than LDA's, while recommendation accuracy ties the best method.
4.05
Item similarities such as cosine and conditional probability are mostly rules of thumb: nobody can say exactly why they work or how to improve them. This University of Antwerp paper derives neighborhood-based collaborative filtering as Naive Bayes classification, obtains two new similarity measures, and finds that the key is to treat the zeros in the interaction matrix as unseen rather than disliked. The multinomial version reaches Recall@10 of 0.221 on Amazon Books, about 30% above cosine kNN at 0.170, with only 4% of EASE's fitting time; on the MIND news data it nearly matches EASE, but on the ML25M movie data it falls behind cosine kNN.
4.05
KDD 2026UUniversity of GeorgiaEEducation University of Hong KongUUniversity of Science and Technology of ChinaHHong Kong Polytechnic UniversityIn sequential recommendation, simply concatenating item-text embeddings with ID embeddings, or feeding both into one encoder, mixes signals that play different roles. This paper splits the information used to predict the next item into three parts: a part shared by text and collaborative behavior, a part only the text has, and a part only collaborative behavior has. RIDRec learns one representation per part. Contrastive learning aligns the text-side and behavior-side representations, and an orthogonality constraint keeps the two unique representations from overlapping with the shared one or with each other. Each representation predicts the next item, and a gate weights the three according to the current sequence. On four public datasets, Amazon Toys, Beauty, Sports and Yelp, the authors report significant gains over a range of existing sequential recommenders, and the code is open source.
3.90
Most cross-domain sequential recommendation papers assume the raw behavior logs of two businesses can be pooled for training and that users appear in both. In practice the data is often isolated by business rules, and many users exist in only one domain. FLADRec, from Skoltech, Sber AI Lab and collaborators, passes only a fixed-length user vector between domains. The source-domain model compresses the user into a vector, and a lightweight adapter maps it into the target-domain embedding space. That vector is added to the embedding of every item in the target sequence, attaching the same user context to each step, and the fused sequence goes into SASRec. Users with no source history skip the fusion and are recommended as usual. On large-scale data with half the users overlapping (QK articles to QK videos), NDCG@10 is 24% higher than single-domain SASRec, and training is 3.7 to 11.6 times faster than ABXI, which needs pooled data. When pooling is allowed, though, it still trails ABXI on some datasets.
3.87
PinterestIn ad auto-bidding, a PID (proportional-integral-derivative) controller keeps adjusting bids according to spend pacing, yet one set of control parameters has to serve hundreds of thousands of campaigns with very different budgets, durations and traffic. Pinterest does not let reinforcement learning set bids directly. Instead, once a day it picks a multiplier for each campaign's proportional and integral terms, either keeping the default or moving it a fifth up or down, and the PID runs as usual. In a seven-day online A/B test, effective cost per click (eCPC) fell a significant 0.25% with revenue essentially flat; two months after launch, against a holdout, campaigns with more than a week left had nearly 1% lower eCPC.
CIKM 2026