PERSONALIZATION · 2026-11-13

Personalization — Paper Daily, 2026-11-13

Today's personalization track runs along two lines. One is accounting for cost and signal in recommendation. Meta's ROO merges the impressions of one request into a single sample and spends the saved compute on a bigger user model, and ShopMind gives simulated users a mental state that changes as they browse, then carries learned mental states into real CTR prediction. The other is making personalized predictions explainable and guaranteed. One paper has an LLM summarize long counterfactual removal lists into a single preference, and another proves that feature interactions make provable explanations hard. Two medication recommendation papers control drug interactions and the severity of missed drugs, and a rating-model paper checks whether missing answers could overturn the current prediction.

4.60
LEAD STORY · Meta

1. Request-Only Optimization for Recommendation Systems

Recommendation training samples are usually cut per impression. One request sends the user several items, each impression becomes a sample, and the same user's features and long behavior sequence are copied into every one of them. Meta counted three of its products and found 4 to 7 impressions per request on average. ROO packs all impressions from one request into a single sample, so user-side features are stored once and computed once. The saved compute can go straight into a bigger user-side model. For retrieval, the offline gain that impression-level HSTU bought with 6.8 times the compute is reached with almost no extra compute per sample once training is per request. Long-sequence models are expensive largely because they recompute the same user.

3.75

2. From Counterfactual Edits to Human Narratives: LLM-Driven Interpretations for Recommender System Explanations

A counterfactual explanation tells the user which past interactions, once removed, would make a recommendation disappear. The trouble is that on MovieLens-1M, pushing the top movie out of the top 5 often means removing from a dozen to several dozen watched movies, and nobody can read a reason off such a long list. The authors put an LLM in the middle, sampling from the movies to remove and from the rest of the history, giving only year and genres, and asking for one sentence on what the removed movies share. The scoring wants that sentence to match the removed movies and not the rest of the history, otherwise it is a generic line that fits anyone. One round of DPO on these scores improves correctness in nearly every setting. How good the explanation is depends heavily on the upstream counterfactual method, and ACCENT and SPINRec produce the shortest removal lists and the most accurate summaries.

3.75

3. ShopMind: A Mental World Model for Personalized and Realistic E-commerce User Simulation

LLM simulators of e-commerce users usually predict the next action straight from the history and the current page. Yet the same discounted product looks like a bargain to a price-sensitive user and may signal lower quality to a quality-minded one, and behavior logs alone cannot tell these two apart. ShopMind keeps an evolving mental state for each simulated user, recording views on products, unresolved doubts and the current intention, rolls candidate actions forward before choosing, and lets a scorer trained on real trajectories pick the most human-like one. Against the strongest baseline, personalization improves by 18.0% and the realism of whole shopping sessions by 19.6%. The by-product is more useful. A model trained on the synthetic pairs to infer mental states from behavior raises AUC when added to CTR prediction on real Taobao logs. Having an LLM guess mental states zero-shot makes CTR prediction worse instead.

3.25

4. Clinical Evidence-Guided Residual Learning for Medication Recommendation with Dynamic DDI Regularization

Medication recommendation turns a health record into a set of drugs, and a common approach copies the prescriptions of similar patients. But if one similar patient took B and C and another took B and A, copying a whole prescription brings in drugs with weak support. EviDrug makes the call drug by drug, first building an initial prescription from the patient's own history and retrieved similar cases, then applying a bounded correction to each drug's score based on that drug's own evidence. A second design targets the drug-drug interaction (DDI) penalty. During training it thresholds predictions into prescription sets, measures the real DDI rate, and adjusts the penalty strength every epoch. On MIMIC-IV, AP beats the strongest baseline by 2.26 percentage points, and compared with prediction-only training, test DDI drops by 9.0% while AP and F1 barely move. A smooth proxy loss can change while the thresholded set of drugs does not change at all.

3.15

5. Severity-Controlled Prediction Sets for Medication Recommendation

Medication recommendation is usually evaluated only on whether a drug was missed. Missing a drug counts as the same error whether or not the predicted set holds a close substitute. CSC uses the drug classification tree (ATC) to give each missed drug a severity, light when the set has a drug from the same small class and maximal when it only has drugs far away in the tree. It then picks one global threshold on a calibration set over a trained model's scores, so that the average severity for test patients stays under a tolerance chosen in advance, with no retraining. On MIMIC-IV with the tolerance set to 0.25, the measured average severity is 0.2439. The authors also state that taking the top-k by score at a matched set size gives an even lower average severity. Its use is that the threshold follows directly from a risk tolerance, and it does not pick better at the same budget. Counting near misses and far misses on a category tree separately is an idea recommendation evaluation can borrow.

3.15

6. Do Feature Interactions Break Explainability? Provable Explanations for Any-Order Neural Additive Models

One way to explain a prediction with a guarantee is to find the smallest set of features that, once fixed, keeps the prediction unchanged however the other features vary. For general neural networks this cannot be computed at scale, but for additive models with no feature interactions at all (NAMs) it is easy. The authors prove that allowing interactions of even a fixed order makes the worst case as hard as for general networks, requiring exponentially many verifier calls. There are two ways out, a sparse interaction structure with limited overlap, or a stricter definition that verifies interaction terms separately, and then a linear number of calls suffices. On second-order models over nine tabular datasets, the sparse version always gives smaller explanations than a greedy feature-removal baseline, and on HELOC, where the gap is largest, it uses 21.1% of the features against 55.2% for the baseline. Feature interactions are where the hardness of explanation comes from, so provable explanations need sparse interactions by design.

3.15

7. Can the Remaining Answers Change the Prediction? Worst-Case Stability Checks for Personalized Rating Models

Personalized rating models often have to predict before the user has answered everything, for example predicting a clinician's remaining symptom ratings from the half already given. The question is whether the unanswered ones could flip the current prediction. The authors borrow the logic of calling an election, where the leader can be declared once they win under every way the uncounted votes could fall. As long as the completed model adds up one contribution per answer, taking the worst case for each missing answer separately gives an exact bound, without enumerating combinations. On ratings from 108 clinicians with half the history available, only 5.4% of predictions can be confirmed this way. Revealing first the answers that narrow the worst case most raises this to 13.7% at the same budget. What is confirmed is agreement with the completed model, which is not the same as being correct. Certificates like this are suited to deciding when to stop asking questions.