This blog post is the meat behind the Spotify article. Wanted to make it a top-level submission so more people could see the algorithms.
This blog post is the meat behind the Spotify article. Wanted to make it a top-level submission so more people could see the algorithms.
Didn’t the Spotify post say they didn’t use this algorithm because it’s slow and memory hungry?
They used this algorithm, but modified it somehow. (fta: “However, his algorithm is complicated and could be very slow in some cases, so we modified it to better suit our needs”). The Spotify post doesn’t explain the difference though, as the high-level description is the same as the one from Martin Fiedler’s post.
Based on the reference to the dithering algorithm, maybe they push the probability of an artist reoccurring to the next slot instead of stretching each artists' tracks into the full length of the playlist (and ending up with many sparse arrays).