Cross‑Domain Multi‑Objective Modeling and Long‑Term Value Exploration in NetEase Yanxuan Recommendation System

The recommendation system at NetEase Yanxuan consists of four stages—recall, coarse ranking, fine ranking, and re‑ranking—where fine ranking handles multiple business metrics that often suffer from sparse conversion data, cold‑start issues, and a mismatch between business goals and algorithmic objectives.

Initially, the fine‑ranking model was a single‑objective CTR deep‑learning model; it later incorporated user behavior sequences, evolved to multi‑objective modeling, and finally to cross‑domain multi‑objective modeling to address data sparsity across various scenes such as new users, new items, and newly launched modules.

Sample construction uses clicks and conversions as positive samples and treats exposed but unclicked items as negatives, with additional optimizations like top‑item down‑sampling, exposure aggregation, fake exposure filtering, and handling of false‑positive samples.

Feature engineering categorizes features into numeric, categorical, sequential, and embedding types, applying normalization, RankGauss, hashing, and embedding techniques accordingly.

Model architecture iterations moved from single‑task networks to multi‑task frameworks such as MMOE and PLE, where MMOE shares expert networks across tasks with task‑specific gates, and PLE adds task‑specific experts to mitigate task dominance.

Position bias—users preferring items shown earlier—was mitigated by adding a debias module that learns bias features (e.g., slot position, device type) and combines them with the main MMOE output, yielding a 4.95% increase in clicks per user and a 1.70% rise in click‑through rate.

Loss functions were enhanced by replacing cross‑entropy with Focal Loss to focus on hard samples and by applying GradNorm to balance gradient magnitudes across tasks, resulting in CTR improvements of 6.92% and CTCVR gains of 5.87% in online A/B tests.

Cross‑domain multi‑objective modeling introduces a domain field feature, feeding it into a domain tower that interacts with a STAR network (shared tower + domain‑specific towers) to capture both commonalities and differences among scenes, achieving up to 10.8% uplift in conversion rate for low‑traffic scenes.

Long‑term value exploration includes a mixed‑ranking strategy based on Thompson Sampling, which models user preferences for different card types via a Beta distribution, and a session‑length objective added to the multi‑objective loss to encourage longer user sessions, both showing positive effects on retention metrics.

Multi‑scene modeling addresses the challenge of heterogeneous traffic by moving from a single shared model (which dilutes performance) and per‑scene models (which are data‑starved) to a shared MMOE framework augmented with scene‑specific towers and a bias network, achieving 5‑10% AUC improvements and notable gains in pCTR for small scenes.

The presentation concluded with a Q&A covering cold‑start handling for new users, model maintenance across many scenes, labeling strategies for session length, and the relationship between multi‑scene and multi‑objective modeling.