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

Background – Multi‑scenario modeling is a business‑driven algorithm optimization task that aims to handle differences and commonalities across various recommendation scenes (daily recommendation, streaming recommendation, playlist recommendation, etc.) in NetEase Cloud Music.

Modeling Goals – Use a single model to serve all recommendation scenes, improve effectiveness by jointly modeling user behavior across scenes, and reduce both machine and labor costs to boost development efficiency.

Overall Framework – The system consists of a unified public‑domain network that captures shared user interests and multiple private‑domain networks (SEN) that preserve scene‑specific features. The architecture combines a common MMOE layer for multi‑task learning with scene‑specific towers, and adopts a hierarchical attention mechanism to replace heavy LSTM‑based long‑term interest modeling.

Key Modules – Unified modeling design (public domain), private‑domain network design (scene‑specific towers), task‑master logic for multi‑task gradient isolation, and model lightweighting via hierarchical attention.

Modeling Challenges – The "double seesaw" problem (multi‑task and multi‑scene trade‑offs), balancing shared and private features, handling feature drift, and ensuring efficient offline training with massive heterogeneous samples.

Application Effects – After deployment, core recommendation scenes saw >10% lift in click‑through rate, many minor scenes >15% improvement, and overall user retention increased by ~1%; the model also benefited other NetEase businesses.

Outlook – Plans to extend the unified model to additional business lines such as podcasts and live streaming, further increasing its impact.

Q&A Highlights – The model supports adding new domains and tasks, uses full‑sample training for new scenes, employs task masking to decouple gradients, and adopts hierarchical attention to jointly process long and short sequences.