How does Spotify suggest music to me?

A lot of my friends have recently debated whether to use Youtube music or Spotify (Apple music users are not friends, sorry).

Now, to run this debate on factual grounds, I wanted to get hands on more than just “I like the UI of Spotify better”. So, as a software developer trying to build an understanding of real world algorithms, I decided to dig deep.

Understanding the VOYAGER Algorithm

The core of Spotify’s recommendation system is the VOYAGER algorithm. The VOYAGER algorithm transforms the way we interact with music by creating a dynamic, hierarchical clustering system. Here’s how it works:

1. Entity Creation:

Each song is treated as an entity, much like a file, containing essential details such as the song’s name, genre, and release date. This data forms the foundation of the algorithm.

Creating an entity file for each song

2. Constructing the Profile Space:

Imagine a vast digital space where each song entity resides. This space is populated by such profiles according to their characteristics.

A 3D space where each file is placed according to the values of artist, genre and release_date

3. Randomized Division:

To organize this space, the algorithm begins by randomly dividing it into two halves. This division isn’t arbitrary; it aims to separate the songs into groups with shared characteristics.

This splitting process continues, creating smaller and more refined groups until clusters of similar songs emerge.

This 3D space with all the songs is then split according to similarity

4. Cluster Formation:

These groups of songs, now termed clusters, are collections of music with similar traits. Clustering helps in managing the vast amount of music data efficiently.

5. Tree of Pools and Sub-Pools:

The clusters are structured hierarchically, forming a tree-like architecture. This tree consists of pools (larger groups) and sub-pools (smaller, more specific groups).

This structure allows for intuitive navigation and retrieval of music based on various criteria matching your likings.

recommendation tree according to the various songs shortlisted in one group

Personalised Search and Suggestions:

The algorithm leverages this tree structure to offer personalized music suggestions. By analyzing user-specific data such as search history, preferred genres, and listening habits, it can tailor recommendations.

The system also prioritizes but is not limited to genres based on listening time and identifies the user’s willingness to explore new but related genres.

The algorithm takes full advantage of the extensive data collected from users. Similar to how Spotify elegantly presents user data in their yearly Wrapped feature, VOYAGER utilizes this data to enhance user experience.

The algorithm not only organizes music efficiently but also ensures a highly personalized and engaging listening experience for users.