Multi-Scenario Modeling for NetEase Cloud Music Recommendation: Architecture, Challenges, and Results

Background – Multi‑scenario modeling is an algorithmic optimization tightly coupled with business, aiming to handle differences and commonalities across various recommendation scenes such as daily recommendation, streaming recommendation, playlist recommendation, and more.

Overall Architecture – The system unifies data, scene, and top‑level tasks into a single framework, separating public (shared) and private (scene‑specific) feature spaces. Public features capture common user interests, while private towers (SEN) handle scene‑specific knowledge, combined with an MMOE structure for multi‑task learning.

Modeling Goals – (1) Serve all recommendation scenes with a single model to improve effectiveness and capture unified user interest representations; (2) Reduce machine and human costs, increase development efficiency, and promote technology co‑building.

Key Difficulties – The "double seesaw" problem (multi‑task and multi‑scene balancing) and the challenge of replacing many scene‑specific small models with one universal large model.

Key Modules

Unified modeling with public and private networks.

Public network design focuses on extracting common user features.

Private network (SEN) provides scene‑specific towers, allowing flexible addition of new scenes.

Cross‑domain multi‑task model with task‑master logic to isolate gradients for tasks that do not belong to a scene.

Model lightweight design using hierarchical attention instead of heavy LSTM‑based long‑sequence encoders.

Application Effects – After deployment, core recommendation scenes saw >10% lift, many minor scenes >15% lift, and overall next‑day retention improved by 1%; the model also benefited other NetEase businesses, increasing new‑user and retention metrics by 0.2% each, while simplifying the tech stack and saving resources.

Outlook – Future work aims to extend the unified model to more NetEase services such as podcasts and live streaming, further increasing the model’s impact across business lines.

Q&A Highlights

New scenes can be added by inserting a private tower in the SEN network.

Iterative training cycles refer to offline full‑model training, not weekly recommendation pushes.

Gradient isolation and task‑masking prevent interference between scenes and tasks.

Hierarchical attention processes long and short sequences together, with short sequences acting as targets.

Public features benefit low‑activity users the most, as they provide a shared interest base.

Overall, the presentation demonstrates how a unified, multi‑scenario AI model can improve recommendation quality, operational efficiency, and scalability across a large music platform.