Problem
A consumption product wants to tell each user something meaningful about how they use it, not only which items they like. Top-artist lists do not do this: two people with the same favorite artists can behave very differently. Clustering on raw logs gives segments that are hard to name and explain to users. The team needs a small, stable set of segments that every user can be assigned to from ordinary play logs.
Example from the article: Spotify Wrapped 2022 gave every listener one of 16 “Listening Personalities”, each a combination of four binary attributes.
Solution in brief
Define four independent behavioral axes, each with a score computed directly from play history. Normalize each score across the whole listening population and cut it at the median to get a letter. Four letters give 16 interpretable segments that are not forced to be equal in size.
Questions this case answers
- How to describe users' consumption style with a few binary, interpretable segments computed directly from event logs, independent of what they consume
- How to normalize per-user behavioral scores against the whole user population so that each axis splits users roughly in half
- How to describe each user's listening style with a few interpretable axes, separate from which content they like?
- How to turn raw play logs into user segments that people understand without a clustering model?
Tools and methods
- Four behavioral axes — Familiarity/Exploration, Timelessness/Newness, Loyalty/Variety, Commonality/Uniqueness, each scored from play logs.
- Population normalization — compare each score with the distribution over all listeners to assign the letter.
- Behavioral segmentation — 2^4 = 16 segments from four binary splits, not from a clustering algorithm.
Google or a one-time email code. No password.