PERSONALIZATION · 2026-10-19
Personalization — Paper Daily, 2026-10-19
Community search over short user-generated posts, answers and comments. Primary metrics are search relevance (nDCG) and CTR.
4.15
LEAD STORY ·
LinkedInLinkedIn TransX caches reusable behavior memory and separates sequence encoding from candidate-conditioned scoring. Offline prefill and nearline refresh keep history processing off the request path. Its base model raised CTR by 6.02% over the production baseline in an online experiment.
Key insight: Compute history when behavior changes, then reuse the state across candidates. This shifts serving work toward candidate discrimination, while introducing explicit obligations around freshness, version compatibility and recovery.
4.05
RecSoup makes accuracy–diversity trade-offs adjustable by reusing objective-specific checkpoints. It prunes task-vector updates and learns preference-dependent merge coefficients. Parameter construction was about 165–228 times faster than a hypernetwork across three benchmarks.
Key insight: Controllability can come from combining existing specialists. Pruning and coefficient learning solve different problems: preserving useful updates and choosing their mixture. The latency claim concerns parameter construction, not the full recommendation pipeline.
4.05
Agentic-Priors asks a language model for interpretable trade-offs, maps them to priors over choice-model coefficients and updates them with observed choices. Across seven datasets, cold-start accuracy exceeded an uninformed prior by 5–40 percentage points.
Key insight: Use language-model knowledge as an uncertain starting point rather than a fixed label. Bayesian updates allow individual observations to take over. The approach is most useful before enough historical choice data exist.
3.70
GoogleGoogle studies training divergence as a gradient–weight–loss cascade and maps interventions to each stage. The combined treatment improved AUC by 0.38% relative to baseline on the second binary classification task.
Key insight: Inspect gradients, weights or activations, and task losses on one timeline. A loss spike may be the end of a cascade. Intervention effects differ across classification and regression tasks.
3.70
IIIT (ISM) DhanbadSSt. Ann’s College
LLMTransRec combines item identity, genre metadata and BERT-derived text representations in a Transformer for next-item prediction. It reports offline tests on four benchmarks; the federated component is a feasibility study. The accessible page does not quantify the main gains.
Key insight: Semantic content and interactions can share a compact sequence backbone. The appendix specifies BERT representations, rather than an online generative model. Treat the federated result as a proof of concept.
3.60
Korea University
Chung-Ang UniversityCalendarRec adds separate month, day and weekday representations to a Mamba-based sequential recommender. On health-functional-food click logs, NDCG@10 improved by 14.2% relative to Mamba4Rec.
Key insight: Calendar position captures recurring demand that interaction order alone misses. Separate temporal embeddings offer a concrete feature-design option; the headline result is from health-food e-commerce.
3.40
CCTU
Language embeddings can expose item identities through similarity matching against public descriptions. Experiments on two datasets also show that item leakage can enable user re-identification from sparse ratings. Obfuscation reduces leakage but harms recommendation utility.
Key insight: Unreadable vectors are not anonymous records. Public catalogs can provide the auxiliary information needed for identification. Evaluate both item recovery and user linkage when assessing embedding-sharing risks.
3.30
VVIT
This study separates ranking quality from serving correctness in open-network commerce. In controlled trace replays, routing and identifier fixes raised route resolution success from 60.0% to 99.8%. These are reproducible system experiments, not production A/B results.
Key insight: Ranking metrics cannot detect broken destinations or inconsistent item identities. Track accessibility alongside relevance. The study tests those failure surfaces independently with model-only and reliability-only configurations.
3.05
MARS combines a question knowledge graph with momentum-sensitive proficiency updates and time-budgeted routing. In its negative-marking experiment, overtime violations fell from 50% to 5.6%. The evaluation concerns adaptive assessment rather than media recommendation.
Key insight: The rate at which a user state changes can inform the next decision. MARS links rating stability to reward and remaining time, offering a transferable idea for budget-constrained personalization.
CIKM 2026