Skip to content

Building Daily Mixes without audio features

How Melodius groups a listener's artists into Daily Mixes using only listening history: trusting listening time over skip flags, decaying old plays, session co-occurrence, catalog similarity and sticky groups.

3 min read

Spotify builds Daily Mixes partly from audio features: how songs sound. Melodius has no such data. What it has is listening history. This is how that history alone becomes up to six themed mixes that feel like they belong to you.

No audio features

Without tempo, energy or timbre, the only question that can be answered is behavioural: which artists does this listener play together, and what does the catalog say about them? The grouping is pure and deterministic, so the same history always produces the same mixes, which makes it testable.

What counts as a play

  • Listening time, not skip flags. The recorded skip flag turned out to be unreliable: in the data, "skipped" plays averaged longer listening than "completed" ones. So a play counts once at least 30 seconds were heard, and hearing half a track earns full credit.
  • Decay. A play loses half its weight every three weeks, so taste can move on.
  • Likes count like one full, fresh play.
  • Avoidance. A track or artist abandoned within 30 seconds three times, and never really played, is something to keep out.
typescript
const HALF_LIFE_DAYS = 21;
const weight = Math.min(1, heardShare * 2) * 0.5 ** (ageDays / HALF_LIFE_DAYS);

When are two artists similar?

Similarity between two of a listener's artists is a weighted blend of four signals:

  • Sessions, 0.4: played in the same listening sessions more than chance would explain. Plays more than 30 minutes apart are different sessions, and long sessions are capped so one marathon does not link everything.
  • Catalog, 0.3: the catalog's similar-artist lists.
  • Era, 0.1: release years within twenty years of each other.
  • Genre, 0.1: shared genre tags, where both have them.

Grouping, and keeping groups stable

Artists merge into groups while their average similarity stays above the threshold. A mix that changes theme every day feels random, so yesterday's grouping pulls its artists together with a stickiness bonus, and a new group continues yesterday's mix when they share at least 30% of their artists' weight. Your "Mix 2" stays recognisably your Mix 2.

Lessons

  1. Check your signals against the data before trusting them. The skip flag looked authoritative and was not.
  2. Weak signals should support strong ones, not create links alone.
  3. Stability is a feature. Recommendations that reshuffle daily feel worse even when each day is "better".

The plays feeding this include those heard in Discord, see counting plays from a voice channel.

Illustrated avatar of Kumar Manas

Written by Kumar Manas

Full-stack developer in India. I build Melodius and Krypton.

Related projectMelodius: music streaming platform