Problem
Spotify's Shuffle was uniformly random for five years, using a standard randomization method. Listeners still complained that it felt repetitive: some songs and artists surfaced again and again, others seemed buried. A truly random order does not guarantee an even spread, but people expect one, so statistical randomness and perceived randomness diverge. The goal was not to make Shuffle less random, but to make it feel fairer and fresher.
Example from the article: a listener who just played "Bohemian Rhapsody" or "Sweet Caroline" should not meet those songs again at the start of the next shuffled order.
Solution in brief
Generate several uniformly random orders of the playlist, score each for freshness using the listener's recent plays, and serve the best-scoring one. The randomness itself is untouched; only the choice among random candidates is informed by history. Spotify calls this mode "Fewer Repeats" and made it the default for Premium users.
Questions this case answers
- How to make a random ordering of a list feel fair to users when a uniformly random order often puts recently seen items first
- How to bias a random permutation toward fresher items using recent-consumption history, without a learned model and without added latency
- How to make a shuffled playlist feel random to listeners without giving up randomness?
- How to reduce repeats of recently played tracks at the top of a shuffled order?
Tools and methods
- Random permutation sampling — produces the candidate orders; the baseline "Standard Shuffle" uses the Mersenne Twister generator.
- Freshness scoring — penalizes recently played tracks that appear early in a candidate order.
- Best-of-N selection — serves the candidate with the highest freshness score.
Google or a one-time email code. No password.