NetEase Cloud Music Recommendation System: Architecture, Challenges, and AI‑Driven Solutions

NetEase Cloud Music has become a market leader in mobile music apps, surpassing 200 million users in 2016, largely due to its high‑quality recommendation system that applies AI algorithms to provide personalized music suggestions.

The platform supports multiple recommendation scenarios, including Daily Recommendation, Private FM, and Playlist Recommendation, each tailored to different user intents and consumption patterns.

Compared with e‑commerce recommendation, music recommendation differs in that the items (songs) can be repeatedly consumed, have longer interaction times, and evoke highly subjective emotional responses, requiring richer user behavior analysis.

Recall System Exploration

1. Real‑time interest vector modeling : Users and songs are embedded into a shared low‑dimensional space; similarity search retrieves top‑k candidates using a high‑performance vector engine (n‑search). Optimizations include replacing average‑pooling with self‑attention and using click sequences instead of search queries to reduce over‑fitting.

2. Dynamic multi‑interest modeling : An information‑capsule network extracts multiple interest vectors from user listening histories, adapting the number of interests to sequence length and using attention to incorporate target information during training.

3. Music knowledge graph : Entities such as songs, albums, artists, videos, and MLOGs are linked via multi‑hop relationships, enabling cross‑entity recall based on user history.

4. Long‑short‑term interest mining : Separate long‑term and short‑term interest spaces are trained with dedicated losses, enriched with wide features and song‑profile information to balance persistent and recent preferences.

5. Multi‑path fusion recall : Multiple recall channels (vector similarity, knowledge graph, session‑aware models) are combined to capture the emotional and contextual aspects of music.

Ranking Model Evolution

The ranking pipeline models the user behavior chain (exposure → click → play → feedback) and has progressed from simple linear models to FTRL, deep neural networks, and sophisticated sequence models.

Key advancements include:

Deep Interest Network with attention to preserve important sequence embeddings.

User interest evolution model that integrates attention with ARGRU to capture temporal changes in taste.

Session‑aware multi‑behavior modeling that separates user actions (play, like, skip) and treats sessions as coherent sub‑sequences.

These models incorporate rich song embeddings (genre, language, artist, album) and side information to improve prediction accuracy.

AI Thinking for Music Recommendation

The authors emphasize that music is an emotional art form while AI provides rational tools; combining both yields a healthier recommendation ecosystem that discovers long‑tail, niche tracks and enhances user enjoyment.

The presentation concludes with gratitude to the audience and invites further community engagement.