Optimizing QQ Music Ranking Models: From Pairwise Methods to Multi‑Objective Learning and Causal Inference

Background – QQ Music’s recommendation pipeline has progressed through three stages: initial focus on brand products (personalized stations, daily playlists), supporting long‑tail creator content via a billion‑yuan incentive plan, and finally expanding to multiple content categories such as video, long‑form audio, and live streams.

User Perception Model – The ranking model aims to capture diverse user actions (play, collect, share, skip, etc.) and rank songs accordingly. Pairwise training constructs behavior‑based pairs (e.g., collect > full play > skip) to address sample‑size imbalance across actions.

Model Structure Optimizations – Three modules were introduced: CIN for high‑order feature crossing, Behavior Attention to focus on recent relevant actions, and ID‑Cross to model co‑action relationships between historical song sequences and candidate items. These enhancements improve both shared and individual user preferences.

Learning Objective Upgrade – A multi‑objective framework (CGC) predicts separate targets for each user behavior, allowing shared and independent information to be learned jointly. This mitigates the bias of under‑trained rare actions and yields higher click‑through and collection rates.

Causal Inference for Long‑Tail Support – To counter the Matthew effect, user‑only and item‑only sub‑models estimate curiosity and item popularity biases. By subtracting these factors during inference, the model captures true user preference, boosting long‑tail content distribution and overall click performance.

Cross‑Domain Recommendation – An MV‑CoNet architecture extends MDN with dual‑mapping matrices between content domains, enabling knowledge transfer from dense domains (e.g., personalized stations) to sparse ones (e.g., long‑form audio). This increases user exposure to diverse categories and overall platform stickiness.

Module Personalization – Top‑N user‑preferred categories are predicted, then combined with a monetization score to generate multiple ranking strategies. An exploration‑exploitation (EE) layer selects the strategy that maximizes both user experience and revenue, leading to higher DAU diversity and category value.

Results & Outlook – The combined optimizations deliver noticeable gains in clicks, collections, and a 10% increase in song distribution while reducing “painful” (扎心) recommendations. Future work includes deeper causal bias removal, reinforcement‑learning‑based cross‑domain methods, and further refinement of module structures.