How NetEase Cloud Music Solved Cold‑Start with Large‑Scale Graph Neural Networks

Recommendation systems are now a standard feature of major internet companies, yet many data sources such as social and information networks exist in non‑Euclidean spaces, making traditional collaborative‑filtering methods based on historical behavior insufficient.

NetEase Cloud Music, a national music app, has shifted from a simple music tool to a content community and faces severe cold‑start problems in its live‑streaming recommendation scenario, where a large proportion of users have never interacted with live streams.

To overcome these challenges, the R&D team selected Baidu's open‑source distributed graph learning framework PGL (built on PaddlePaddle) as the foundation for large‑scale graph neural network training. The graph models heterogeneous entities—songs, DJs, queries, radio IDs—and their relationships (user‑host, user‑song, query‑host) to construct a unified graph.

Using algorithms such as DeepWalk, Metapath2Vec, and GraphSage, the team learned strong graph embeddings for entity IDs. These embeddings enable vector‑based recall, transferring users' historical actions in music, playlists, and Mlog domains to the live‑streaming domain, thereby improving the relevance of recommended hosts.

Cloud Music’s data graph contains billions of edges even after pruning, which exceeds the capacity of many open‑source GNN frameworks. PGL’s native distributed graph storage, distributed sampling, and integration with PaddlePaddle’s Fleet API allow efficient training on “thin” compute nodes without requiring prohibitively expensive hardware.

Experimental results demonstrate that graph‑based recommendation substantially boosts view counts, especially for new users and new hosts, confirming the effectiveness of cross‑domain knowledge transfer and large‑scale distributed training.