Are Algorithms Enough for Music Discovery?

Growing up in rural Devon my teenage experience with the world of music was limited to rifling through the discount cassette tapes at Woolworth’s or deliberately hunting out obscure bands by listening to John Peel’s radio show at night.  It’s so different for today’s teenagers who have a huge variety of ways to ‘discover’ music from YouTube through to any number of music streaming services with their own recommendation services. And yet, when we ask people where they find new music, the vast majority continue to say the radio.

In a world of multiple channels and algorithmic music recommendations, it seems strange that a broadcast medium would still be the place most people listen to new music. Why are more people not finding new music from sites where they can curate their preferences and music is streamed to them that fits their preferences?

I would argue that it’s partly because we still have a strong affection for music based radio – a low effort source of entertainment that is tapped into as part of our daily routine. The fact that all our cars are equipped with FM radio helps of course. And it’s also a communal experience – we typically enjoy sharing music.

But it does beg the question whether recommendation systems will ever be able to properly reflect our musical tastes. Because if they do, why are we not using them more?

Studies have found that music discovery via music streaming service recommendations is still much lower than the radio.  But as Nate Silver, a man you may expect to be a strong advocate of computer-based prediction said: “It is questionable whether any computer will be able to capture the subtlety and personalisation that real human beings demonstrate across social contexts.”

So is music preference too slippery to really be able to pin down using technology? Are we simply too unpredictable, too dependent on the context we find ourselves in, to be able to accurately predict our preferences?

Duncan Watts, a scientist at Microsoft explored this very issue and specifically wanted to understand the role of social influence in music preferences. To this end, he created an artificial music site with over 14,000 consumers from a teen related website. All participants in the study were asked to rate a list of previously unheard songs from unknown bands. They were required to assign a rating to the song and were then given an option to download it. This corresponds to an ‘individual’ model of decision making, where we are making decisions without reference to others.  As might be expected, there was a normal distribution of preferences for the different songs, with the most popular songs being around three times as popular as the least popular.

The second group of individuals did exactly the same task but with a crucial difference: they could see the number of times that the songs had been downloaded by others. When the consumers could see the preferences of others (in the form of downloads) there was a significant shift in consumers’ preferences, with just a few songs being hugely popular and the majority of songs getting much lower ratings. In this scenario, the ratio between the most popular and the least popular was at least thirty to one.

And the tracks that were popular when selected individually (i.e. not seeing what had been downloaded) bore little relationship to the tracks that were selected when consumers could see what had been downloaded by others. So we can see that interactions between individuals ended up drastically enhancing small fluctuations to produce outcomes that would have been very difficult to have predicted.

So, on this basis, music preferences may often have less to do with attributes of the music itself and more to do with a shared experience with others. Hence whilst sites that rely on algorithms based on musical features alone will help ‘sift’ the huge choice available, it at best looks debatable how effective they are in identifying music choice which really hits the spot, as this is primarily a social, rather than individual, choice.

We would, therefore, expect social-recommendation to become an increasingly popular source of music discovery and indeed the likes of Spotify, Rdio and Songbird now include ‘frictionless’ integrations with Facebook. We are also seeing the rise of purpose-built sharing sites such as ‘This is My Jamm’.

Whilst these services theoretically help to deliver more powerful recommendations, at the moment many still create more noise than value due to the user simply having too many social connections. You may not take your music cues from your family, for example. However, in principle, a discovery that integrates the social cues of your friends or those you may consider ‘experts’ would seem to be a sensible bet for the future.

If we accept the premise that music discovery is primarily a social phenomenon then this will explain the continued popularity of radio, but also the rise in ‘human-curated’ sites such as Pitchfork. Targeted at music enthusiasts, feisty Pitchfork music reviews provide guidance around independent music. There are also a number of bespoke expert services which will fulfil a similar role; record-store Rough Trade has started a subscription service where they send subscribers their six recommended tracks of the week.

It’s this kind of personalised, trusted experience (created by humans rather than algorithms) that may well end driving the market. And this has got more sophisticated as the likes of Shuffler.fm offer ‘audio magazines’ pulling together tracks from a wider selection of influential blogs from a virtual panel of music ‘tastemakers’.  And surely these very same principles explain the continued success of radio – after all, it is a human-curated music channel which has a long history of social engagement.

So whilst huge investments are currently made to distil preference from analysis of tracks played, there is plenty of evidence that social effects are equally – if not more – critical in determining choice.  Increasingly, new analysis techniques will explore the type of social influence at play to assess whether it is, for example, mediated by ‘indiscriminate copying’ or ‘influencer effect’. By identifying the ‘species’ of social influence in operation for particular genres of music it should be possible for recommendation sites to more accurately shape their offering.

A better understanding of these effects and the associated marketing implications is an area which increasingly needs to be the focus of activity not only for music sites but for content discovery more generally

This article first appeared on MediaTel’s Newsline

By Colin Strong