PERSONALIZATION · 2026-10-25
Personalization — Paper Daily, 2026-10-25
Premium people retrieval conditions on who you already saw, carousel bandits update only observed browse depth, and Kuaishou agents write ad scripts per user—today’s personalization thread is moving exposure, depth, and creative production out of post-hoc bias fixes and into the objective and the system.
4.08
LEAD STORY ·
LinkedInLinkedIn LinkedIn finds that for Premium people-browsemap retrieval, the production member embedding Holistic-171 only reaches MAR@100 0.288, and an off-the-shelf 0.6B LLM embedding scores 0.192—neither was trained under the exposure condition of who the viewer already saw. PAIR treats exposure as the conditioning for preference learning; a same-scale 0.6B model lifts MAR@100 to 0.450 (56% relative over Holistic, 2.3× over the base encoder). Online A/B on Premium Browsemaps: CTR +7.5%, Premium retention +2.1%, no measurable latency regression. Who you already saw is preference signal, not noise: condition the objective on exposure and a small model can beat a broad member vector.
4.08
In carousels and feeds users stop at a browsing depth; positions below that depth are neither clicked nor exposed, yet classical position-based models (PBM) still treat them as “seen but not clicked.” This paper proposes OD-UCB / OD-TS / OD-DMED under observable depth: update item stats only from truly exposed positions, with an instance-dependent log upper bound for OD-UCB. On synthetic shallow/deep environments and RecGaze-parameterized runs, OD-TS matches PBM-TS on final mean regret while all three beat PBM-UCB. When browse depth is observed, do not treat unseen slots as negative feedback.
3.75
Cross-domain sequential recommenders often move preference via global alignment and ignore how consistent a user is across domains, so irrelevant semantics leak in. BGAFlow builds a shared semantic space, evolves it with state-adaptive semantics, and gates transfer per user so the amount of transfer scales with that user’s cross-domain consistency instead of one global objective. Transfer granularity belongs at “how consistent is this user across these two domains,” not at one global alignment loss.
3.73
JJiangxi University of Software Professional Technology
In multi-behavior recommendation, how much browse / favorite / add-to-cart still predict purchase can go stale before the prediction moment. BLMRec encodes each behavior by type and temporal distance from prediction time, uses multi-window lifetime learning for short-term activation and long-term persistence, then feeds validity and stagnation signals into candidate purchase ranking so expired actions are not treated as strong intent. Behaviors are not permanent labels—model a lifetime per behavior type or ranking gets pulled by stale signals.
3.63
KuaishouMost personalized video ads still retrieve among pre-produced creatives, so personalization stays coarse and gaps across user groups persist. Kuaishou’s MAPS runs four collaborative agents (profile → script → creative → closed loop) to generate script-driven video ads per user; versus a retrieval baseline, CTR +2.87%, CVR +2.74%, ARPU +9.85%. When personalization is capped by pre-produced creatives, put creative production inside a multi-agent loop instead of chasing similarity in feature space.
3.55
Agentic shopping recommenders need world models that hold over long interactions, yet few resources offer fine-grained real-user trajectories or offline checks against production A/B. UserArena releases 64,729 session trajectories from 11,010 users (566,102 items, 13 action types), paired with an interactive environment and trajectory-level evaluation for offline policies vs production A/B. Without long-horizon real trajectories and an A/B anchor, agent recommendation “world models” only look good in toy sims.
3.55
Households share a pantry and members follow different dietary rules, so per-person recommenders collide on resources and break constraints. FairFRESH casts household food recommendation as constrained fair division: learn per-member utilities, allocate under welfare objectives, enforce diet rules with a deterministic filter, and explain with Shapley/Banzhaf. Learned utilities hit Recall@20 0.293; fair allocation cuts Gini disparity 42% vs conflict resolution, lifts Recall@20 from 0.180 to 0.197 (within ~2% of a smoothed Nash upper bound), and every recommendation satisfies the rules. The hard constraints are shared stock and diet rules—after utility learning you still need an enforceable filter.
3.45
Generative assistants now host ads inside answers: an ad’s CTR depends on the hosting passage, and because users often stop at the first click, earlier ads steal clicks from later ones—raising a bid can even push a rival ahead of you. Existing mechanisms either approximate welfare but lack known truthfulness, or are truthful while welfare degrades with more ads. This paper redesigns the mechanism for that click competition, folding passage dependence and early-stop externalities into allocation and payment. In-assistant ads are not a drop-in position auction—passage sets CTR and early stop creates externalities, so truthfulness and welfare must be rebuilt together.
3.45
Generative medication recommenders ultimately emit drug sets, and DDI constraints act on the whole set; the usual fix is predict-then locally delete interacting drugs. Holding FLAME’s frozen MIMIC-III scores, output size, and DDI constraint fixed, this paper compares stronger local repair strategies against an exact global MILP decoder that jointly picks the highest-scoring feasible set: local repair itself adds avoidable error; global decoding is stabler. Do not rely on post-hoc local repair for set-level constraints—with scores fixed, a global feasible decoder removes the extra error local repair introduces.
3.20
Fairness-aware recommenders often bake one objective into training or apply a fixed re-rank after scoring; in production the right action changes with the candidate pool, exposure debt, preference history, and utility budget. FAIRGate inserts a model-agnostic policy-selection agent between the base ranker and the final top-K, choosing from a predefined library (long-tail exposure, group balance, concentration control, calibration protection, …) per request instead of inventing fairness on the fly. Deployed fairness is “which validated policy to pick,” not redefining fairness on every request.