How group recommendations work

Ordinary recommendation services work on one person. This one takes several: you enter a list of Last.fm usernames and mark each one either “likes” (their library feeds the shared pool) or “might like” (they receive the recommendations). Any count of each is allowed, so one mechanism covers three different questions.

Mark everyone “likes” and the service looks for shared favourites: artists the whole group already plays. Mark everyone “might like” and it becomes the group version of ordinary recommendations. Mix the marks and the donors introduce the recipients to the donors' music.

For every member the service loads their Last.fm top artists over all time, the last year and the last month, picks the search starting points from them and fetches a similar-artist list for each. Every candidate is then scored separately for every person: how much they already play it relative to their own top artist, and how well it is predicted to fit them.

The finished table can be re-ranked in the browser without a second search: a slider from “liked by everyone” to “liked by anyone”, a separate weight per member, a filter on how many people an artist counts for, genre chips, and a toggle that hides what the recipients already play. The result link can be shared with the group.