Recommendation System Architecture and Coarse Ranking Design for Tencent Music's Quanmin K‑Song

The article introduces the four major mobile music products of Tencent Music, highlighting that Quanmin K‑Song accounts for over 1.5 billion monthly active users and plays a pivotal role in content distribution and the relationship between producers and consumers.

Business Background : Quanmin K‑Song offers diverse recommendation scenarios, including content, live, karaoke rooms, and user‑generated feeds, with specific modules for high‑quality UGC, follow‑stream, and local social recommendations.

Recommendation System Architecture and Challenges : The pipeline consists of four stages—recall, coarse ranking, fine ranking, and re‑ranking. The recall layer filters massive item pools using index‑based, social‑graph, and model‑based methods, followed by content‑ratio filtering to manage the daily 5 million new uploads.

Coarse Ranking Layer : Acts as a bridge between recall (millions of candidates) and fine ranking (thousands). It balances scoring performance and ranking precision, incorporates ecosystem controls (timeliness, tone, diversity), and employs model distillation or feature distillation to improve expressiveness while keeping inference cost low.

Coarse‑Ranking Algorithm Design : Two main routes are explored—extending recall with lightweight serving or compressing fine‑ranking models using a dual‑tower architecture. The dual‑tower model decouples user and item features, generates user embeddings in real‑time, and computes inner‑product scores for efficient large‑scale sorting.

Model Distillation : A teacher‑student framework transfers knowledge from a large, complex fine‑ranking model (teacher) to a lightweight coarse‑ranking model (student) using soft logits and temperature‑scaled softmax, improving both accuracy and efficiency.

Diversity‑Control Design : Discusses the importance of diversity from system and user perspectives, outlines three technical solutions—rule‑based shuffling, embedding‑based dispersion, and Determinantal Point Process (DPP). DPP jointly models relevance and diversity by maximizing the determinant of a similarity‑relevance matrix, with greedy algorithms reducing computational complexity.

Online Gains : Deployment of the coarse‑ranking module and DPP‑based diversity control yields significant improvements in interaction and playback metrics, as well as broader creator coverage and content freshness.

Further Considerations : Future work includes enhancing teacher‑student training to surpass teacher performance, integrating prior knowledge into similarity matrices, and adjusting relevance scores to balance platform‑level objectives with user preferences.

The article concludes with acknowledgments of the speakers, a call for community engagement, and promotional information about the DataFunTalk platform.