Spotify Recommendation System using Content-Based Filtering

Recommendation systems can work in very different ways depending on what data they rely on. This video moves from a collaborative filtering approach, covered in an earlier video, to content-based filtering, and builds a working Spotify song recommendation system using the audio characteristics of the tracks themselves rather than how other users have behaved.

Content-based filtering versus collaborative filtering

Collaborative filtering relies on the preferences and similarities between users, recommending content that similar users have already watched or interacted with. Content-based filtering works differently: it relies on the content, features, or characteristics of the items a user has already interacted with or liked. The example given is straightforward: if a user has watched a movie and liked its genre or its actors, other movies sharing those same characteristics become good candidates for recommendation, regardless of what other users have done.

Getting data from the Spotify API

To build the system, the project pulls data directly from the Spotify API. Anyone with a Spotify account can create an app on the Spotify developer dashboard and obtain a client ID and client secret. Those credentials are used to generate an access token, which in turn is used with a specific playlist ID to retrieve all the tracks in that playlist, along with the audio features of each individual track. The core data used for the recommendation system includes the track name, the artist, album details, and the audio features associated with each song.

Choosing the features that define similarity

After pulling the playlist data and reviewing what came back, the project settles on a set of audio features to build recommendations around, including danceability, energy, and key, among other characteristics of a song. These values are normalized so they can be compared consistently across tracks.

How the recommendation function works

The content-based recommendation function starts by finding the index of the input song, the one the user wants recommendations for. Using that index, it calculates similarity scores between that song and every other song in the dataset, then retrieves the top N+1 most similar songs. The "plus one" matters because the most similar song to any given track is itself, so the function needs one extra slot to return N genuinely different recommendations. The output includes each recommended song's track name, artist, and album, and the approach can be extended to return additional features as needed.

Where content-based filtering falls short

The video is candid about the tradeoffs of this approach. Recommendations tend to stay similar to items the user has already interacted with, which limits diversity and can hold back the discovery of new interests. There is also a cold start problem: content-based filtering struggles with new users or new items because it depends on historical interaction data or established item attributes, and without enough of that data, its recommendations become unreliable. Because it only analyzes item attributes, it may not fully capture the complexity of user preferences, making it less accurate in some cases than collaborative filtering. Finally, there is a risk of over-specialization, where the system overly focuses on specific features and misses broader patterns or associations that could otherwise improve recommendation quality.

Key takeaways

• Content-based filtering recommends items based on the item's own features rather than on similarities between users.
• Spotify's API provides the track, artist, album, and audio feature data needed to power the recommender, accessed via a client ID, client secret, and access token.
• The recommendation function scores similarity against an input song and returns the top N most similar tracks, excluding the song itself.
• Weaknesses include limited diversity, a cold start problem for new users or items, and a risk of over-specialization on narrow features.
• This approach directly complements the collaborative filtering method covered in an earlier video in the series.

Who this is for

This video is useful for anyone building a first recommendation system and wanting to understand content-based filtering as an alternative or complement to collaborative filtering. It is part of a Humanitarians AI series exploring different recommendation approaches using real Spotify data.