PERSONALIZATION · 2026-11-12
Personalization — Paper Daily, 2026-11-12
Today's main thread in this track is whether evaluation and reuse in recommender systems can really be trusted. Meta uses scaling curves drawn from a few small experiments to predict the outcome of a full-scale A/B test. INTENT evaluates new users, new periods and new items separately, and finds that much of the drop after the training cutoff comes from user behavior itself changing. A second thread is how to hand upstream work to downstream models: SeqFold compresses the states of a frozen long-sequence model into short codes fetched by timestamp, and LUMIR trains multimodal business representations apart from the downstream model. An audit of review distillation and a short paper on tourism data quality remind us that a student agreeing with its teacher, or rich data sources, do not by themselves make results trustworthy.
4.45
LEAD STORY ·
Meta /
NUS
Add a feature or change the architecture of a recommendation model, and a small-scale experiment may say it does nothing, while full-scale training shows it helps. The Meta authors hit exactly this on an LLM recommender: adding movie genres to the prompt lowered HR@10 at small scale and clearly raised it at full scale.
4.35
Downstream search and recommendation models all want to reuse the output of a frozen long-sequence user model, but the full multi-layer hidden state for one user at one point in time averages several million bytes, too much to store or ship. Coupang's SeqFold compresses each layer's hidden states into a fixed number of interest slots: a learned gate decides how much each past action puts into each slot, older actions are decayed more, and each slot is finally quantized into a short binary code.
4.15
Once deployed, sequence models in banking and recommendation face a period after the training cutoff, users never seen in training, and newly appearing items, yet the usual evaluation just holds out each user's last few events. INTENT splits the test into seven cells by whether each of these three was seen in training, and picks the cutoff date on five public datasets with one rule fixed in advance.
3.85
When feature interactions are taken from a black-box model such as XGBoost and written into an interpretable additive model like a motor insurance tariff, the usual practice is to keep the top-scoring pairs, but a high score may just reflect fitted noise. LENS trains one model on each of two disjoint data splits, computes for every feature pair the interaction pattern left after removing each feature's own effect in the first model, and checks whether the second model reproduces it. On French motor insurance data, vehicle age with brand is reproduced at 78%, while vehicle age with region also ranks high but is reproduced at only 3%, so it is dropped. Replication alone is not enough, because both models can share a very weak pattern, so the reproduced strength must also clear a floor. On synthetic data it recovers 91% of the material interactions without selecting a single false pair. Only an interaction that two independent fits both learn deserves a place in the model.
3.80
In next-POI recommendation, a business's metadata, reviews and photos are usually trained jointly with the downstream sequence model, so changing the downstream model means retraining the fusion. LUMIR separates this step: an LLM first turns all three modalities into text that is then encoded. Parts of the modalities are randomly masked, two incomplete versions of the same business are pulled together in embedding space, and businesses that follow each other in user trajectories are pulled together too. The learned business embeddings are then frozen and handed directly to BERT4Rec, SASRec and GRU4Rec. On a multimodal Yelp dataset, with BERT4Rec the HR@10 is 24.2% above the strongest baseline, and recommendations cover more businesses. With GRU4Rec the gain is much smaller, and aligning adjacent businesses mainly helps BERT4Rec, which sees context on both sides.
3.80
For e-commerce items with few interactions, collaborative filtering cannot learn reliable representations, so which helps most: item images, text descriptions or user reviews? MPIUR puts all three into one graph model: an LLM extracts aspect, opinion and sentiment triplets from reviews, and reviews serve as features on user-item edges. Images and text are fused with per-item weights, and a popularity-weighted alignment loss pulls items with fewer interactions harder toward their content representation. On the Amazon Musical Instruments and Baby categories, it has the best global nDCG@10 and Recall@10, with the advantage concentrated on items with fewer than 10 interactions. The ablation shows that removing reviews hurts more than removing images and text, dropping sparse-item nDCG@10 from 0.2347 to 0.1869. On popular items, BM3, which uses no reviews, is better on several metrics.
3.60
If an LLM labels millions of reviews and a small model is distilled from those labels for production, does agreeing with the teacher mean agreeing with editorial judgment? On more than five million reviews from a Turkish e-commerce marketplace, the authors had Gemini sort reviews into three grades by whether they give concrete product evidence, then trained BERTurk and other compact encoders on those labels. Adding the product category before the review slightly raises the student's agreement with the teacher on products that first appeared after the training cutoff. But on 16,000 later reviews checked by one company annotator who never saw the teacher's labels, the category prefix lowers agreement with both the teacher and the annotator. Agreeing with the teacher only shows the student learned the teacher, not that it matches human standards, and the conclusion also depends on which data it is measured on. On a single GPU the student processes reviews 50.2 times faster than the teacher pipeline.
3.15
Graphs whose nodes carry both text and images are everywhere in e-commerce, social and content platforms. This survey sorts multimodal graph learning methods into five levels by where modalities and graph structure are combined, and offers one criterion for comparing them: whether a node's text representation can be influenced by its neighbors' images. A product with a poor title, for example, can be recognized from the images of products bought with it only under that condition, and modality-decoupled models such as MMGCN cannot do this before fusion. The authors also define model-independent dataset descriptors, such as the share of edges linking nodes of the same class and the average number of neighbors. They urge evaluations to include strong baselines without the graph, including an MLP on the same features and a multimodal LLM that ignores graph structure. A gain from graph structure counts only after it beats strong baselines that use no graph.
3.10
Multimodal recommendation increasingly relies on reviews, photos and LLM-generated summaries from social platforms, yet the quality of this data is rarely checked. This short paper uses two tourism datasets, from Yelp and from Instagram in the Lake Constance region, to list the problems encountered in practice. Fake accounts generate reviews and check-ins, and LLM summaries contain made-up content. Modalities also contradict each other: a summary built from most reviews may call a restaurant clean and quiet while user photos show it cluttered and noisy, and contradictions between modalities directly reduce the reliability of the fused representation. Most posts come from young people and locals, while the people receiving recommendations are often older tourists. The authors suggest downweighting low-quality modalities during fusion, replacing free generation with structured summaries that have fixed fields, and modeling tourists and locals separately.