Music recommendation: have platforms tried everything ?

It was taken for granted that the advent of music streaming platforms through algorithmic recommendations would increase personalization, discoverability and diversity of the content listened. But there has been many allegations over the last couple of years that algorithmic recommendations create filter bubbles and barely increase diversity.

Since their emergence streaming services have been offering various recommendation features, and listeners have been more or less using them, with various impact on the extent of discoverability and diversity of the content. Here is an interesting and recent article by Valerio Velardo providing some insight on this matter.

It is indeed a challenge to make listeners wander outside their comfort zone, the music universe they know and like. There are natural limits, like the time that a listener can dedicate to listening to music and discover new content.

In a series of articles, we are going to look at the offer of recommendation features, the impact of this offer on listeners, how recommendation algorithm are built, and the recommendation features within music platforms more global strategy.

We will argue that recommendation is still in its infancy and that beyond algorithm, the way recommendations are proposed to listeners (UX and algorithms) leaves space for new initiatives.

We will especially focus on Spotify because it has introduced key innovative recommendation features, but we will look also at the other major services (Youtube, Pandora, Napster, Apple, Google, Amazon, Deezer,...).

We will address the following topics:

• The offer of recommendation systems:
  • Recommendation is about extending a listener music universe beyond what he knows and likes.
  • Recommendation systems rely mainly on several points of entry: hits, new releases, genres, moods/activities, friends/relatives, curators/influencers, trending, similar artists and songs.
  • The digital revolution (devices, UI, algorithms) allows to better cater for those music UX needs and offers considerable room for innovation.
  • recommendations sytems have converged around 3 models: Pandora smart radios, Spotify Rapcaviar and Discover weekly
  • A surprising lack of personalization overall, and overloaded UI
  • Other music related content available for recommendation: concerts, podcasts, lyrics, scores…
  • There is scope for rationalising, differentiating UX and targeting more precisely various types of listeners.

• The listeners

  • Listening to music is on the rise: 2 hours/day
  • The challenge of being interested by a new artist
  • Need for listeners to contextualize discoveries: what is the background of the artist, who are listening to him, who does recommend it, is it new and trending ?
  • The appropriation cycle: discovery, repetition, pleasure, saturation
  • Various types of behaviors towards music: extent of engagement towards music, personal distinction versus being part of the community, being genre focused or diverse, genres preferences, socio-demographics
  • The limits of the listener: time limit, curiosity limit, need to contextualize the discovery, immediate pleasure versus risk taking. 
  • building its identity: frozen at 32 ?
  • Listeners use of recommendation systems

• Making recommendations:

  • How are the algorithms produced ?

  • Human versus algorithmic recommendations, and hybrid types

  • the 3 types of similarity: social, content, context

  • Similarity: a relevant model ?

  • Benchmark of algorithms provided by the major streaming platforms: proportion of user based, content based, context

  • Satisfaction metrics: dictatorship of the skip rate ?

  • The impact of algos on diversity, memorisation,...

• The recommendation within platforms strategy

  • KPIs, listeners satisfaction, discovery, diversity

  • UX: where are the features increasing diversity, risk taking, serendipity,...?

  • Have platforms tried everything?

• Filling the recommendation gap