I have seen lots of evidence from this community that Spotify users are fed up with the suggestion algorithms. It seems that all previous posts on this topic have been closed, "because the algorithm team is always working to improve the radio/suggestion algorithm." In the interest of reiterating the message, I am here going to discuss some core failures of the current radio algorithmic functions that I haven't seen brought up before, and I will try to relate them back to previous user feedback. 1. Instead of playing songs that are similar in musical content or feeling to the seed song, the algorithm seems to simply choose "songs that people who like the seed song will likely like." There is a crucial distinction between these two ideas, and it relates back to much of the user base's algorithm dissatisfaction. Because everyone pretty much already knows "the songs that people who like their type of music like." These are the songs that usually come up on a song radio, and the root of much discontent is that users already know these songs or that they hear from the same few artists over and over. By giving more sway to the actual musicality of the song, Spotify can play us songs that we don't know but that are musically similar to the seed, and in that way it can pleasantly surprise us and widen our musical palette. As it stands the algorithm frustratingly manages to consistently play songs that i. we already know and ii. that are not musically consistent with the seed. 2. It seems that the "response to listener action" part of the radio algorithm is too slow to be effective within a single listening session. I hear that Spotify learns from skipped songs and songs played in entirety, but it's not clear whether this input helps shape song radios or just account-scale recommendations and daily mixes. Many users listen to under ten songs from a song radio before starting a new one, so they either don't skip enough songs to tune the algorithm to just what they want before they change stations, or, if Spotify remembers listener inputs from previous iterations of the same song's radio queue, the mood may have changed and the listener's desires for the station today may not match those from yesterday. The solution to both of these problems is, again, a stronger radio queue basis on musical elements, so that users can get exactly the style they want and they don't need to do much tuning. There's a HUGE metric that I like to use to asses the song radio algorithm, and it happens to be one of its biggest failures right now: Start two different song radios from two songs on the same album that have very different musical feel. If you get mostly the same songs on the two stations, you have good reason to feel bad. From the Community forum history, Spotify teams have been working on the algorithm for years. Evidently something big is missing from their efforts. Please, please make it so that two radio stations started from dissimilar songs on the same album actually play different songs, songs that better reflect the musicality of the seed songs.
... View more