PERSONALIZATION · 2026-11-18

Personalization — Paper Daily, 2026-11-18

Today's personalization track asks who should be able to see and change the signals in a recommender, and where limited budget should be spent. One line keeps a long behavior history at full resolution for unified ranking and sees higher merchandise value in advertising traffic; another turns natural-language interests into editable serving addresses, so people can direct retrieval and exposure. A multimodal paper first lets a large model write coarse and fine features, then fuses them in the middle of a graph network, while an uplift model for promotional pricing warns that product text may leak the treatment decision. Others treat a newly opened location as an intervention on a district's return, use a simple rule to find patients more likely to benefit from home blood-pressure telemonitoring, and first decide whether an edit suggestion should be shown in a conversation. Product entity resolution, feature algebra, and live EEG music recommendation each make the same practical point: match the structure to data noise, efficiency goals, and the mood of the moment, rather than defaulting to the heaviest model.

4.85
LEAD STORY · Tencent

1. TokenFormer V2: Breaking the Resolution-Complexity Dilemma in Unified Recommendation

Industrial recommenders need fine-grained long behavior histories, yet they must stay inside a strict serving budget: keeping every event is expensive, and compact summaries throw evidence away. This work decouples representation resolution from interaction density. It encodes behaviors in local groups, expands range through hierarchical merging and global synchronization, and still lets the final predictor read every encoded event directly.

4.20

3. Take a Step Back: A Dual-Granularity Raw Feature Empowered Intermediate Fusion Graph Convolutional Network for Multimodal Recommendation

WSDM 2027UBCHK PolyUHKUHKUSTUniversity of Luxembourg📖 Deep dive →

Multimodal recommenders try to learn multi-level representations from sparse interactions, and often use contrastive learning to align modalities, which can squeeze out information that is unique to one modality. This work suggests taking a step back: large models generate coarse and fine raw features before training, as a complement to the original item features, rather than replacing the original text and images.

3.93

4. Regularizing Treatment Leakage for LLM-based Uplift Estimation

JD.com

When a large model reads product text to estimate the uplift of promotional pricing, titles may already contain words such as “clearance” that reveal the treatment decision. The model can take that shortcut, then fail when a real decision is still pending and the text does not yet carry those words. LULR freezes the encoder and adds a mutual-information penalty on the causal head's shared representation, to reduce dependence between that representation and the treatment. On a million-record industrial pricing randomized trial, the offline metric was better in four of five random seeds, but the difference was not statistically significant and policy values were nearly tied. In a separate one-week online test, the complete deployed pricing policy had 5.45% higher sales volume and 1.91% higher GMV than the control group. That is short-term business evidence for the full policy, and it does not isolate the regularizer. For teams feeding product text into an uplift model, blocking treatment leakage matters more than simply adding more text.

3.31

5. Real-World End-to-End Product Entity Resolution

International Hellenic UniversityNational and Kapodistrian University of Athens

Matching the same product across e-commerce catalogs is hard: titles mix promotional wording and messy formatting, and many near-lookalikes actually come from different brands. This evaluation compares four zero-shot cascades built from lexical, semantic, and hybrid retrieval plus generative verification, and releases RETAILERS, a real-world dataset with cross-lingual noise and dense brand collisions. No single architecture wins on every dataset: lexical retrieval is faster and lighter on short titles, while semantic retrieval handles wording changes better but also pulls in same-brand near-lookalikes. On the noisier RETAILERS-D split, hybrid retrieval plus remote GPT verification reaches an F1 of 0.666 and is about 34 times faster than local Gemma. A cross-encoder is a more practical default matcher, while a large model fits accuracy-critical or smaller-scale settings better. For product entity resolution, choose the cascade from the data noise and the latency budget, rather than defaulting to the heaviest generative verifier.

3.31

6. UrbanROI: Causal Modeling for Regional ROI Forecasting under POI Entry

KDD 2027NTU SingaporeChina Agricultural UniversityMeituan

When a new point of interest opens in a commercial area, it can change the region's advertising or business return, but common forecasts watch a single store or one time series. They do not model competition among nearby locations, and they cannot separate the path if nothing had opened from the gain that comes with entry. UrbanROI builds a regional graph from user requests, spatial proximity, and request-level commercial signals, and constructs treated and control graphs for the entered and unentered states. Removing the regional graph or the causal decomposition both hurts accuracy. On real data from Beijing, Shanghai, Shenzhen, Hangzhou, and Guangzhou, it beats the strongest baseline on every error metric. In a semi-synthetic effect-recovery test, sign accuracy rises from 71.2% with temporal information only to 85.0%. For regional ads and merchant onboarding, treat a new location as an intervention, rather than fitting one more accurate curve.

3.31

7. Edit Recommendation: Can LLMs Anticipate Edit Requests from Users, and When?

NTT

After a large model writes an email or a passage, users often send edit requests. Anticipating those requests before they arrive would cut a round of review, but some real edits introduce information that is absent from the prior context, such as a new name, constraint, or goal, and are inherently hard to predict. This work builds a benchmark from WildChat conversation logs and adds decision-model gating: it filters cases that are unlikely to be correct before suggestions are shown. At one gating threshold, the trained model raises the share of shown five-suggestion lists that contain a hit by 12 to 25 percentage points, and the gain also holds for predictors that were not used in training. Edit suggestions should not be pushed after every reply. For teams recommending edits in a chat, deciding when to show a suggestion is more practical than only chasing a higher hit rate.

3.19

8. Algebraic Interaction: A Recipe for Neural Architecture Generation

Tsinghua University

Convolution, attention, and state-space models are usually rewritten as one undivided mechanism, so a change in efficiency or symmetry often means rebuilding both the information pathway and the feature interaction. This work expresses them as bilinear products over a position algebra and a feature algebra, so the original information pathway can stay fixed while only the feature interaction changes. Replacing the feature algebra in Vision Mamba with a structure-constant algebra avoids a large matrix-valued hidden state, speeds training and inference by about 3 times, and keeps CIFAR-10 accuracy at 86.350%, close to the original 86.818%. Linear attention also gains accuracy at a comparable parameter count, and an algebra of three-dimensional rotation representations supplies rotational equivariance by construction. For people designing network architectures, the feature algebra is itself a design variable; positional routing does not have to be rebuilt first.

3.12

9. EE-Motion: An EEG-based Web Architecture for Live Affective Music Recommendation

Politecnico di BariUniversidad Autónoma de MadridWideverse

Mood drifts continuously in a single listening session, but recommenders still lean on static profiles and offline training, and they struggle to serve anonymous visitors who have no history and cannot be retrained on the spot. This web system pairs a consumer EEG headband with the browser over Bluetooth, maps frontal alpha asymmetry and an arousal index onto a valence-arousal plane, and uses an affective adapter to score diverse backbones, including graph networks and sequential Transformers, from the live coordinates. An anonymous visitor only picks a few favorite tracks, with no user table. Without a headband, a cursor on the same plane still shows how the ranking moves with the coordinates. In a local test, ACE and SASRec with the adapter both finish a cycle well below the 125 millisecond deadline of an 8 hertz EEG stream, so requests do not queue up. Signals stay in session memory and are deleted afterwards, and they are not used as a clinical mood diagnosis.

3.06

10. Who Benefits from Telemonitoring? Uplift Modelling of a Nationwide Home Blood Pressure Monitoring Programme

Home blood-pressure telemonitoring is now a national programme, but devices and follow-up staff are limited, and the benefit is unlikely to be the same for every patient. A hypertension telemonitoring programme in Singapore used propensity-score matching on more than sixteen thousand enrolled patients: the six-month blood-pressure control rate rose 5 percentage points overall, and 12 percentage points among those whose blood pressure was not controlled at baseline. Individual uplift estimates then suggested larger gains for people with a higher baseline blood-pressure grade, a more recent diagnosis, and a lower medication and diagnosis burden, and realized gains were also larger at the top of the predicted-uplift ranking. A simple rule written from those features even outperformed every learned model on the uncontrolled-baseline subgroup. That is a usable clue for writing enrolment criteria when care capacity is scarce, though the estimates are still imprecise and not fully cleared of confounding.