Return to site

Recommendation systems: scope for differentiation

Recommendation is about extending listeners music universe beyond what they know and like. It empowers listeners once they have exhausted all their songs/artists search capabilities with further navigation celerity.

Music services, even before the digital revolution, have been relying on several points of entry in the music catalogue: filter by genres, decades, selections of hits, of new releases/what’s trending, by curators/influencers, playlists by context (moods/activities), and provided means for sharing content and playlists. A song is a 3 minute experience, and the question of what to listen next keeps coming back, contrary to other creative contents (movies, books,...). Hence the historic format of the album, which provides a minimum acceptable duration, along its artistic intention.

People in their day to day life encounter many situations where music can be listened to while doing something else: in transport (cars, traveling,...), while eating, doing sport...or with other people (party,...). In those situations their sight and hands may be busy doing another activity, their hearing is available to listen to music. Music can also be more functional and directly stimulate the activity (dance, yoga, cheering up,...).

The digital revolution provides listeners with devices, apps and algorithms that allow to better capture those listening situation opportunities, and to adapt to each context: rich UI on PC, simplified UI on mobile phones and in cars, voice control in the car and smart speakers. Even during situations where interaction with screens is limited, listeners can enjoy rich navigation, to the point of inviting designers to create a zero interface where no interaction is possible. Digital brings higher granularity (the possibility to provide multiple types of playlists, similar artists and songs), higher frequency of updates of selections, a much deeper dive in the catalogue, and personalization (personalized recommendations/playlists/UI to each listener).

"Discover weekly : when a streaming service can pick songs from a catalogue rich of tens of millions, why keep only 30 ?"

Music platform have tried many UI possible interactions to provide those entry points (genres, new releases, tops, mood/activities,...). In the mid 2000’s Pandora had set a landmark with smart personalized radios from seed artists and songs. Since and after years of experimentation to enrich music listening experience (personalization, sharing, gamification,...), it seems that on top of Pandora smart radios there are 2 new winning recommendation models: "Rapcaviar" and "Discover weekly".

broken image

"Rapcaviar", with 10 million followers, is a playlist consisting of 50 songs, renewed at different pace (a song can stay from one day to months), the same for all listeners (not personalized), that listener can subscribe to. It is a format very close to FM playlisting, between continuous feeds of social networks and fixed playlists.

broken image

"Discover weekly" provides each listener a personalized playlist of 30 songs, renewed every week. It is counter-intuitive because when a streaming service can pick songs from a catalogue rich of tens of millions, why keep only 30 ? Why not proposing an infinite scroll when algorithm allows it ? By offering a fixed selection of songs available for one week, it actually brings key decisive benefits: it urges to listen to the selection before the end of the week (there is no history of former selections), you can repeat play-back as many times as you want and see the lineup in the same order (which is crucial to a proper memorization of the new songs), it meets you the same day every week. The power of this feature is to create an habit and a ritual.

Spotify keeps innovating by pioneering new recommendation features: "daily mix" (daily personalized playlists, one by genre listened to by the user), "radar" (personalized selection of new releases), and lots of exotic types of selections: “your hidden face”, “your time capsule”,...but those innovations do not seem to be widely used.

"By and large the offer of recommendation systems is not personalized"

By and large the offer of recommendation systems is not personalized, which can come as a surprise at the age of IA and hyper-personalized experiences provided by the GAFAs and their likes. Most playlists and the UI are almost the same for all listeners. They are numerous points of entry, but there are the same for all.

Update frequency of playlists seems to revolve around one week. But this seems long for many, and too short for others. Even if saturation is calculated for each listener, a saturated song is removed at the same time for all listeners.

Onboarding UX on most platforms takes time before there is some personalization taking place.

All streaming platforms are providing the usual entry points (genres, new releases, moods, social sharing,...), with more or less emphasis, hence a standard of playlist is emerging: 15-50 track playlists, renewed around once a week, not personalized. It offers around 15 main genres (hip-hop, rock/pop, electro, soul/r&b,...), around 15 main moods/activities (at work, party, at the gym, romantic diner, rainy day,...), and numerous curators whether internal (editorial team) or external (journalists, labels, artists, stars, venues managers, amateurs...).

"By adding more and more entry points the UIs of most of streaming platforms have become richer but quite complicated and less readable"

By adding more and more entry points the UIs of most of streaming platforms have become richer but quite complicated and less readable. It brought more granularity and diversity for recommendations, but a lot of curated playlists provide little of the value expected by such feature: good selection of tracks, a well defined genre or context, a clearly recognised influencer. On the same platform lots of (popular) songs belong to a wide range of playlists, and some playlists are called with the same name or are expected to define the same universe. It does not make the UI very effective. Here is for instance an interesting post on how to redesign Spotify UX.

Algorithms through personalization helps simplifying UIs, but provide limited context which is key to expose listeners to new content. There is room for a better trade-off to strike between algo/UI and for providing more explicability and affordance.

Listening to audio tracks is not the only music experience that can be subjected to recommendations: concert, videos, podcast, lyrics, scores...Some platforms have integrated those complementary contents, but as a separate category and not so much via recommendations, though this can constitute a valuable user experience.

Songkick, Bandsintown, created stand alone apps that make geolocalized and personalized recommendations of concerts. Pandora and Spotify have also added some features, but hardly the others platforms.

There is still lots to offer.

"simplified and rationalized UI designed for specific audiences"

Going forward there are several angles for music experiences providers to play: more personalization, simplified and rationalized UI designed for specific audiences, including music related content other than audio recordings in the recommendation,...

To better assess this potential, looking at the listeners is going to help to identify which angles to play, depending on targeted audience and strategy. That will be the subject of the next post.