PERSONALIZATION · 2026-11-16

Personalization — Paper Daily, 2026-11-16

Today's personalization track asks what a recommender should actually optimize. One paper turns user curiosity into a computable score, adds it to ranking, and deploys it on a video platform. Another finds that users click harmful content more, so models trained only on interactions push more of it, and adds harmfulness signals to user and item representations. Work from Reddit makes the auxiliary click-through tower of a conversion model match the production model it is actually paired with. Kuaishou's KuaiRecV2 scores models on a daily-updated candidate pool with several metrics at once and exposes forgetting and instability that offline evaluation hides. A last paper swaps the scorer on top of LightGCN, finds small gains, and is candid about the limits of its experiments.

4.30

2. Serving-Aligned Entire-Space Conversion Modeling via Click-Through Rate Knowledge Distillation

Arizona State UniversityReddit📖 Deep dive →

Ranking often multiplies a click-through prediction by a conversion prediction to get the probability of a click followed by a purchase. Many conversion models are trained with ESMM, which carries its own auxiliary click-through tower trained together with the conversion tower, but only the conversion tower is deployed and multiplied by a separately trained production click-through model.

4.05

3. Measuring, Revealing and Mitigating Harmful Content Propagation in Recommender Systems

Recommenders chase user interest, and few ask whether what they push is harmful. The authors sampled 100 recommendation lists produced by BERT4Rec and found that nearly half of the items contained gender discrimination, misinformation, or offensive language. They had an LLM label the items in three news and advertising datasets as harmful or healthy with a reason, then had humans review the labels.

3.85

4. Sheaf-Inspired Structured Scoring Extensions of LightGCN for Collaborative Filtering

In graph collaborative filtering, after embeddings are propagated over the user-item graph, LightGCN's final step scores a pair with a plain inner product. Borrowing the idea of a sheaf from mathematics, the authors keep the propagation and change only the scorer, so that user and item embeddings pass through a linear transformation that depends on the pair before being compared. They try diagonal, rank-one, and full-matrix transformations, with a parameter-matched MLP scorer as the control. On Amazon-Book, Gowalla, MovieLens-1M, and Yelp2018, the selected structured scorers edge out both the inner product and the MLP on NDCG@20, for example from 0.031028 to 0.031302 on Amazon-Book, a very small gap. The MLP often covers more of the catalog. The authors state the caveats plainly: the candidates were picked after the fact from test results, Amazon-Book has only one seed, and no actual sheaf is built, so the gains cannot be credited to sheaf geometry. The paper is best read as an honest negative reference.

3.80

5. KuaiRecV2: Benchmarking Large-Scale Continual Learning for Diversified and Multi-task Recommendation

NeurIPS 2026USTCKuaishou

One reason recommenders look good offline and disappoint online is that offline data is too static: the candidate pool never changes, and diversity and multiple behaviors go unscored. Kuaishou's KuaiRecV2 puts these factors into one benchmark built from real data. It holds 30 days of behavior from 124 thousand users, the candidate video pool is updated first-in, first-out every day, and models must keep training day by day as they would in production. Evaluation looks at retrieval accuracy, diversity, and several behaviors such as likes and long views together, and a method counts as an advance only if every metric improves. Running several generative recommenders on it, the authors find that models forget old knowledge and that the SIDs generated under the same prefix are nearly identical. More counterintuitively, an RL-only model does improve on day 20 but drops sharply on day 21 once the pool changes, and only alternating with SFT keeps it stable. Aligning multiple behaviors with RL alone is fragile when the candidate pool shifts.