Music Recommendation & HerdIt

This week I had the chance to attend a tutorial at the ACM Conference on Multimedia on Music Recommender Systems presented by Oscar Celma. It was a very informative talk, touching on some of the foundational issues in music recommenders: relevancy, serendipity, transparency, and context. There was also some discussion of the tradeoffs between content based recommendations versus those made based on human added metadata. For instance, content based recommendations have the cross-genre problem of potentially recommending songs from a different genre which share some similar musical features. The assumption in the presentation is that this is bad, though in some sense, serendipity may call for some cross genre pollination.

I wanted to pick at a point that bothered me a bit: the tension between relevancy and serendipity. Relevancy on the one hand calls for a user centric model which takes into account how interesting a particular recommendation is for a particular user. Relevancy means that the recommendations made are in fact meaningful and perhaps “useful” or at least appreciated by the user. On the other hand the virtue of serendipity is espoused as something to strive for. The value judgement is that people shouldn’t be constrained to things they already know or are familiar with, but should also be exposed to things outside of their comfort zone. And music aside (especially in an information domain like politics) I think serendipity IS something major to strive for. But doesn’t this compete with relevancy for attention? A personalized / recommended news page that includes “serendipitous” results risks presenting results to the user that are in fact not relevant at all. I would have appreciated a more earest discussion of the tradeoffs between these factors.

There are various methods that commercial systems are using to make music recommendations. The two big ones discussed in this tutorial were and pandora. Pandora relies on an “army” of paid specialists who listen to each indexed some and rate it based on 400 attributes on a 10 point scale. This clearly cannot scale as there is simply too much labor involved in the process, but the result is impressive. Another tack on the problem is to produce these annotations using non-experts, something I’ve thought about in my design of games like PhotoPlay and AudioPuzzler. Some folks at UCSD have designed a social game for Facebook called HerdIt to try to collect affective music data in the process of people playing the game. The hope is that this data can inform machine algorithms and eventually produce better recommendation systems.