Music Recommendations By The Numbers

Nowadays, there is nothing preventing the independent musician from entering the music market – everyone can sell their music online, quickly and easily. The problem for music consumers is that they are being flooded with choices, so they need to find a reliable music recommendation system. Most algorithms of this type use collaborative filtering, but there are alternatives…

Recommending More Listeners

Collaborative filtering relies on the ‘wisdom of crowds‘ – basically, if you analyse the listening behavior of a large group of people who like a particular song, you can use this data to suggest new songs to someone else who also likes the seed song.

Back when I was at school, this dynamic was enacted in the real world – if you were a fan of The Cure, chances are that you also listened to The Smiths, and if you didn’t, then one of your fellow Cure-fans would probably introduce you.

Now that the process has moved online, the scale has vastly increased, and collaborative filtering is quite effective at providing decent new music recommendations when dealing with popular artists or tracks.

Recommending The Unheard

The danger of CF is that listeners can sometimes be sucked into a ‘similarity vortex’ – a set of artists may be so closely linked that the recommendations keep coming back on themselves, with very little ‘discovery’ taking place. It may be hard for new artists to break into this circle, and the recommender may not help listeners to break out.

It is also true that CF requires large numbers of listeners for it to work properly, so the more esoteric your tastes, the less likely it is to satisfy your needs.

An alternative approach is to set listener behaviour aside and focus on the properties of the music itself. This is akin to what Pandora is attempting with its Music Genome Project – categorising songs by their sonic, musical and structural qualities, rather than by who is actually listening to them.

Have You Heard The Mufin, Man?

The idea of algorithmically categorising music has been around for quite a while – I previously wrote about hit prediction software that makes record company executives’ jobs easier by calculating the odds of a particular track becoming a blockbuster.

Paul Lamere is a researcher at Sun Labs who deals with music recommendations at a very in-depth level. He recently posted his impressions of a new content-based music suggestion site, Mufin.

Lamere’s overview of Mufin is worth reading, as is his strategy for exposing a CF filter masquerading as a content-based filter. In some cases, it seems as if the system actually resorts to random-number-based recommendations.

However, the art and science involved in musical content filtering is extremely complex, and will require vast investments of research and ingenuity to get right. For the sake of all those who seek new and obscure music (without having to wade through hours of mud), let’s hope someone works it out…