How QQ Music Recommendation System Understands Your Preferences
The QQ Music recommendation system tackles cold‑start by first mixing Chinese and English tracks, then builds a six‑dimensional user profile (content, social, scenario, crowd, time, blacklist) and tags songs with six attributes, using content‑based, collaborative, matrix‑factorization and neural‑network models plus implicit co‑listening links, while acknowledging that final wisdom still comes from human listeners.
In the article “S‑Tech: Main Creator Talk”, the QQ Music recommendation team explains how their system interprets listeners’ preferences, likening it to a “God’s hand” that silently delivers the right music.
The narrative begins with a fictional story of Robert, an Alaskan pipe‑fitter who discovers the power of music recommendation during a winter night, illustrating the emotional impact of personalized music.
The system, referred to as RS (QQ Music Recommendation System), faces the classic “cold‑start” problem: when it knows nothing about a new user, it first presents a song and observes the user’s reaction (like, skip, collect) to infer preferences.
RS prioritizes recommendation factors in the order: language > artist > genre. Because 90 % of QQ Music users listen to Chinese and English songs, the system initially mixes both languages and records actions such as “collect” or “skip” to build a user profile.
User profiling consists of six dimensions: content preference (artist, language, era, genre), social attributes (age, gender, region), scenario preference (charts, playlists), crowd attributes (activity level), listening time slots, and blacklist.
Music profiling tags each track with six categories: artist information, audio features (MFCC, pitch), popularity statistics, genre, emotional tags, and instrumentation.
Based on these profiles, RS employs several recommendation strategies:
Content‑based recommendation (e.g., “if you like country music, recommend Country Road”).
Neighborhood (collaborative‑filtering) recommendation, where users with similar listening histories exchange song suggestions.
Matrix‑factorization and neural‑network (RBM) hidden‑factor models to discover indirect similarities between songs across languages.
The article also discusses how “implicit” connections between songs are derived from co‑listening data, and how “first‑order” and “second‑order” links help expand the recommendation space.
Overall, the piece illustrates that while algorithms can map preferences, the ultimate wisdom of recommendation still originates from human listeners.
Tencent Music Tech Team
Public account of Tencent Music's development team, focusing on technology sharing and communication.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.