What Is The Secret of Spotify's Recommendation System Success?
Online services such as Amazon, Netflix and YouTube propose new content to their users based on their browsing history and personal preferences. Out of all these systems, there is one that stands out - Spotify's recommendation engine. Every week a special "Discover weekly" playlist is prepared for each of over 100 million application users, consisting of 30 songs that they have never heard before, but which they are most likely to enjoy. It is a popular option thanks to which users broaden their musical horizons.
What is Spotify?
Spotify was founded in 2008 in Stockholm. It is a cloud-based digital music platform that provides access to over 50 million songs across various devices and a rapidly growing number of podcasts and movies. This streaming platform pays the performers of the streamed songs proportionally to the number of streams their songs have accumulated. Users pay a monthly subscription of PLN 19.99 for using the mobile application or use the free option with ads.
How does Spotify collect data about users?
Spotify doesn't use just one revolutionary recommendation model that would be responsible for generating all of personalized playlists. Spotify's recommendation system combines some of the best strategies used by different kinds of models to create one powerful and comprehensive solution. To compose Discover Weekly playlists, the platform uses 3 very important elements of the recommendation engines:
- Collaborative Filtering Model that analyzes the behavior and preferences of a given user and compares them to others to predict new songs that they might like. Machine learning (ML) and artificial intelligence (AI) algorithms look for connections between different users so that they can better understand their personal preferences.
- Natural Language Processing Model (NLP) , which is responsible for analyzing the lyrics of both songs, their descriptions and other information obtained from the web. This part of Spotify's system constantly searches the Internet for blog posts, forums, and tabloid news to find out what people are saying about specific artists and songs. Thanks to this, the system has a good understanding of which songs are currently trending and what style they represent.
- Audio model that can analyze audio tracks looking for similarities in beat, tempo, and used instruments. It is essentially the same technology that is used in all kinds of facial recognition programs. In case of Spotify, it has been modified to parse audio files instead of pixels. Additionally, unlike the first two types, the audio models also take into account new songs that have just appeared on the platform.
All these elements combine into one comprehensive and extensive recommendation system, generating personalized playlists every week for each user of the service. All this effort brings impressive results. Within five years since the recommendation system was launched, listeners have listened to over 2.3 billion hours of music from automatically generated Discover Weekly playlists. Moreover, the system is able to introduce users to performers they have never listened to before, and whose songs perfectly match their musical taste. These users are also likely to check out the rest of the work of the artists they discovered through the “Discover weekly” playlist.
Importance for the eCommerce industry
Can practices used by Spotify in its recommendation system find their way into the eCommerce industry? Of course! The algorithms behind generating "Discover Weekly" playlists clearly show that the multitude of data sources and the use of several recommendation models is the key to success and maximizing profits. This is the reason why end-to-end recommendation engines like Recostream offer a wide range of recommendation models based on both product characteristics and descriptions as well as user behavior.
Models such as "other customers also bought" or "popular" can be implemented in different places in the store, providing customers with a constant selection of products that would most likely be of interest to them. We wrote more about the models and their location on the website here.
The example of Spotify also shows the importance of personalizing customer experience and recommendation systems today. Presenting users with content tailored to their needs and tastes is extremely important and can certainly bring tangible results in the form of an increased average session length or higher engagement, and the comprehensiveness of the recommending engine is important both in the entertainment and eCommerce industries.