Weekly Top 10 · W47

2026-11-16 ~ 11-22 · Personalization

Generative · Personalization · AI Search · Infra · Embodied AI

  1. 1. When Does a Recommendation Spark Curiosity? Theory, Measurement, and Field Evidence · 📖 Deep dive

    Recommenders are often asked to break users out of filter bubbles, yet a user's curiosity has never had a definition a system can compute. Existing proxies such as unexpectedness and variety-seeking only look at how…

    Appeared in: 11-16

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

    Appeared in: 11-16

  3. 3. RACE: Ranking-Uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation · 📖 Deep dive

    On-device sequential recommenders often compress every request with the same ratio, so simple interests waste compute and complex ones do not get enough capacity. Pruning and decomposition usually freeze the model and…

    Appeared in: 11-17

  4. 4. Measuring, Revealing and Mitigating Harmful Content Propagation in Recommender Systems · 📖 Deep dive

    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…

    Appeared in: 11-16

  5. 5. 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…

    Appeared in: 11-16

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

    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…

    Appeared in: 11-16

  7. 7. Learning nonlinear interactions and cross-feature couplings with product-unit residual networks

    When a tabular model must predict an outcome that depends on crossings among several factors, an ordinary multilayer perceptron often buries those crossings in hidden layers…

    Appeared in: 11-17

‹ Previous week