PERSONALIZATION · 2026-10-22
Personalization — Paper Daily, 2026-10-22
Community search over short user-generated posts, questions, answers and comments. Key metrics are search relevance (nDCG) and CTR. Focus: semantic matching, query understanding, short-text retrieval, ranking models and LLM applications in search.
4.75
LEAD STORY ·
KuaishouKuaishou ECHO makes 100K-event histories affordable by retaining reusable original records before candidate-aware retrieval. A full-history teacher supervises selection, while retrieval and aggregation share relevance parameters. Kuaishou reports a 0.613% increase in app usage time in an online test.
Key insight: Compression can preserve individually addressable events rather than only latent interest vectors. Reusable user-level selection moves work out of the request path and leaves candidate-specific matching to a smaller evidence set.
4.55
TencentKuaFu avoids repeatedly encoding the same content across users and profiling tasks. It caches compact item representations, assembles them chronologically and uses staged training to preserve evidence. Tencent reports a 1.37% overall GMV increase against a long-term holdout.
Key insight: Make items the unit of reuse. Encode shared content once, while preserving actions and sequence order on the user side. The cache then scales with item inventory, supporting more frequent multi-task profile refreshes.
4.35
EASformer addresses a mismatch between historical tokens containing observed actions and candidates with no observed action. It separates environment, action and state before matching them. The larger of two reported online eCPM lifts is 3.94%.
Key insight: Attention should compare semantically aligned representations. Separating environmental matching from action prediction also lets masked historical actions reuse the candidate-prediction pathway for self-supervision.
4.30
YoukuSPAR reduces early branch errors in semantic-ID recommendation by retrieving user evidence at every decoding step and balancing traffic across sibling branches. A Youku online experiment reports a 3.95% increase in average watch time per user.
Key insight: Personalization should change with the generated prefix. Inspect user-evidence retrieval and branch-load imbalance together to diagnose early decoding errors that prevent the right item from remaining reachable.
4.10
This study jointly evaluates recommendation relevance and fairness using distance to a data-derived Pareto frontier. Experiments on six real-world datasets test robustness and show that existing joint metrics can disagree with the frontier-based evaluation.
Key insight: A multi-objective score needs an explicit reference boundary. Distance to an achievable trade-off makes the evaluation easier to interpret, while the choice of which frontier region matters remains a business decision.
4.10
SpotifySpotify compares pairwise transition models and Transformer recommenders under shared splits and evaluation protocols. On four of five benchmarks, at least one pairwise model outperforms both Transformer reproductions.
Key insight: Useful history does not prove that higher-order interactions are necessary. Smoothed transitions and recency weighting deserve strong baseline status when attributing gains to a new sequence architecture.
4.00
MOSS continuously samples, scores and diagnoses semantic failures in production discovery systems. Strategy-aware prompts connect relevance judgments to actionable alerts. An offline evaluation of a backfill-blending fix reports a 76.4% reduction in alert rate.
Key insight: Monitoring must understand the intended recommendation relationship. A recurring loop of targeted sampling, strategy-specific prompts and diagnostic follow-up can expose serving defects. The reported alert reduction is an offline repair estimate, not an online CTR lift.
3.95
Unlearning interactions in a semantic-ID recommender can damage neighboring items that share tokens. Neighborhood-Aware Unlearning controls local probability redistribution while removing the target behavior. Across evaluated settings, it retains more than 95% of retrained-reference utility.
Key insight: Global recommendation metrics can hide local damage. Audit both forgotten interactions and their shared-code neighborhoods when removing spam or withdrawn behavior, and constrain the redistribution rather than repairing only global utility.
3.80
This work ranks multimodal candidates by reading option probabilities from a single choice question instead of asking separate yes/no questions. On image retrieval, the choice format gains 5.4–7.3 Hit@1 points over per-candidate yes/no scoring.
Key insight: Joint comparison provides a useful relative signal. For multimodal UGC reranking, one candidate-set decision can avoid inconsistent independent scores, but its value depends on the candidate set, image budget and latency.
3.45
OpenAIOpenAI describes a stateful ad session that follows intent across conversational turns. A separate delivery service invokes retrieval, ranking, relevance and auction tools. The article explains three central architecture choices but gives no controlled CTR comparison.
Key insight: Maintaining evolving user state changes the quality of ranking inputs. Separating ad selection from answer generation supports independent delivery constraints and reduces redundant work; the evidence here is architectural rather than an attributed ranking lift.
RecSys 2026 Workshop