Optimizing QQ Music Ranking Model: From User Perception to Multi‑Category Traffic Exploration

Speaker: Glad, Senior Algorithm Engineer at Tencent Music, edited by Han Xiaoting (Peking University), presented on the DataFunTalk platform.

Background Introduction: QQ Music's recommendation service has progressed through three stages: initial focus on brand products (personal radio, daily 30 songs, playlists), support for music creators via a billion‑yuan incentive plan, and expansion to multiple content categories (video, long‑form audio, live streams), each requiring tailored ranking strategies.

User Perception Model: The team shifted from traditional pointwise models to a pair‑wise ranking approach that constructs behavior pairs (e.g., favorite > share > full play > skip). To address sample imbalance across behaviors, they introduced a multi‑behavior attention mechanism and redesigned the model architecture with three modules: CIN for high‑order feature interactions, behavior‑attention for focusing on recent relevant actions, and ID‑Cross for memorizing co‑action information between user history and candidate items.

Optimizing Learning Objectives: Recognizing limitations of single‑objective pair‑wise models—namely, ignoring user‑specific differences and insufficient learning for sparse behaviors—the team adopted a multi‑goal framework (CGC) that predicts each behavior separately while sharing common knowledge. Causal inference components (user‑only and item‑only models) estimate and remove user curiosity and item popularity biases, thereby mitigating the Matthew effect and improving true user preference estimation.

Supporting the Music Ecosystem: To promote long‑tail native content, the ranking system incorporates causal inference to counteract popularity bias, and the recommendation pipeline adds recall‑layer and re‑ranking techniques (e.g., balanced recall, relationship‑graph recall, re‑ranking tilt) that elevate under‑represented tracks without compromising user experience.

Multi‑Category Traffic Exploration: As QQ Music adds new categories, the team tackles fragmentation and fixed homepage layout by (1) cross‑domain recommendation using an MV‑CoNet model that shares user embeddings across categories and learns dual‑mapping matrices between them, and (2) module personalization that predicts top‑N preferred categories per user, assigns monetary value to each category, and dynamically generates multiple ranking strategies for exploration‑exploitation optimization.

Summary & Outlook: The presented optimizations have enhanced user discovery, increased playback and collection metrics, reduced "painful" (扎心) recommendations, and boosted overall traffic by ~10%. Future work includes further bias reduction in causal models, reinforcement‑learning‑based traffic optimization, and continued refinement of cross‑domain and personalization architectures.