China’s music and video streaming unicorn NetEase Music has a database of 10 million songs, 400 million users, and a net worth of over US$1.14 billion.
Chinese netizens are all ears for the company’s “hearty” AI-powered music recommendations. In an interview with Synced, NetEase Data Scientist Jia Xu and Product Manager Bowen Shen explained the NetEase system, which learns how to predict what songs will resonate with a user’s particular taste in music.
AI recommendation systems are based on the same tech used by news websites, online shopping websites, dating apps, and social media feeds. Consumers generally appreciate a service that learns their preferences and makes accurate predictions, and such systems have become a powerful tool in targeted marketing.
Amazon engineers first put the recommendation systems to work — if a customer purchased A, they might also want to put B into their basket — using a common method called collaborative filtering. This works well when supplied with large amounts of user preference data, which NetEase has and continues to accumulate.
NetEase’s collaborative filtering follows two approaches: the first recommends tracks by linking between users who have similar tastes, the second takes a single track as a vector for similarity calculation, and recommends it based on the user’s play history.
In his blogpost Recommending Music on Spotify with Deep Learning, the creator of Spotify’s music recommendation system and current DeepMind researcher Sander Dieleman explained the shortfalls of collaborative filtering, pointing out that the models are “content-agnostic”, have problems dealing with albums with multiple tracks of different styles, and that “new and unpopular songs cannot be recommended” due to a lack of user data, posing the “cold-start problem.”
To solve these challenges NetEase turned to deep learning, using input data with features such as song genre, artist, album information, lyrics, beats, user comments, VIP download preferences, and price. The information is projected into a low-dimensional latent space to train deep learning models.
The system establishes a vector position for each song, encoding all relevant information including user preferences, so the recommendation system can make suggestions even when the song or user have little or no play history. The vectors are “normalized” to overlook popularity, which ensures that users can also discover and recommend new songs. NetEase also uses a machine learning ranking model to prioritize recommendations on a daily basis.
NetEase is the radio station of the future, a space where even those with the most esoteric tastes can discover new music. Close to half a billion users can’t be wrong.
Journalist: Meghan Han| Editor: Michael Sarazen
Pingback: Synced | Chinese New Year 2018 Synced 30 Best AI Use Case Award