TL;DR: Across a broad spectrum of product features, Spotify’s array of recommender systems play a pivotal role in tailoring personalized user experiences and helping creators grow.
We developed a Named Entity Disambiguation (NED) method to assist human curators in finding and correcting infrequent errors in a music catalog.
We study the problem of constructing personalized playlists for users of music streaming services, in our case Spotify.
Spotify’s search system plays a vital role in helping users explore the catalog and discover new content.
Every day, music is enjoyed, created, and discovered by billions of people around the globe – and yet, existing AI systems largely struggle to model the nuances that make music different from other forms of audio.
7am I usually wake up around 7am and have breakfast with my husband before he goes to work. We chat... View Article
Existing spoken Language Identification (SLI) solutions focus on detecting languages from short audio clips. Podcast audio, on the other hand, poses peculiar challenges...
TL;DR: On online platforms such as Spotify, recommender systems are increasingly tasked with improving users’ long-term satisfaction...
Reinforcement learning (RL) is an established tool for sequential decision making. In this work, we apply RL to solve an automatic music playlist generation problem...
Launching a radical change to a habitual feature, used by millions every day, presents an interesting challenge.
7am I usually wake up around 7am – I’m not a morning person by nature, but I have two young... View Article
Allowing users to discover new entities such as books, music, and movies is an important goal for online platforms. This can be achieved for example by recommending entities that the user has not yet interacted with. Another way users can find new entities is by exploring the catalog with the search system.
Answering causal questions with machine learning algorithms is a challenging yet critical task.
Quantifying cause and effect relationships is of fundamental importance in many fields, from medicine to economics. The gold standard solution to this problem is to conduct randomised controlled trials, or A/B tests.
Understanding cause and effect relationships in Spotify data to inform decision-making is crucial for best serving Spotify’s users and the company.
Recommender systems typically look to users' past consumption to predict what they may want next. In practice, this approach tends to work best when what the user wants is similar to what they have consumed recently, and when it is relatively easy for that person to evaluate new items.
A new approach to calibrating recommendations to user interests. Users’ interests are multi-faceted and representing different aspects of users’ interest in their recommendations is an important factor for recommender systems....
TL;DR: Survival analysis provides a framework to reason about time-to-event data; at Spotify, for example, we use it to understand and predict the way users might engage with Spotify in the future. In this work, we bring temporal-difference learning, a central idea in reinforcement learning, to survival analysis. We develop a new algorithm that trains a survival model from sequential data by leveraging a temporal consistency condition, and show that it outperforms direct regression on observed outcomes......
Cutting-edge research in Machine Learning, Language Technologies, User Modeling, Audio Intelligence, Search and Recommender Systems are some of the key areas we feel incredibly enthusiastic about at Spotify...
There are many A/B tests we might like to run, but which are too technically challenging, risky in terms of user impact or even impossible to perform. For instance, in the classic example of whether smoking causes lung cancer, forcing a randomly selected group of people to smoke is unethical if we believe it might damage their health. In the context of technology companies, if we want to understand if app crashes cause users to churn, we would have to randomly select a subgroup of users and crash their apps on purpose – not something we would want to consider given we do not want to break their trust....
Dynamic topic modeling is a well established tool for capturing the temporal dynamics of the topics of a corpus....
What is Speaker Diarization? Speaker diarization is the process of logging the timestamps of when various speakers take turns to talk...
A large number of new podcasts are launched every month on Spotify and other online media platforms. In this work,... View Article
Here at Spotify, we are highly dedicated to cutting-edge research in various areas in Machine Learning, User Modeling, Personalization, and... View Article