DBTune blog

To content | To menu | To search

Tag - notube

Entries feed

Tuesday 27 October 2009

Music recommendation and Linked Data

We just presented yesterday at ISMIR a tutorial about Linked Data for music-related information. More information on the tutorial is available on the tutorial website, and the slides are also available.

In particular, we had two sets of slides dealing with the relationship between music recommendation and linked data. As this is something we're investigating within the NoTube project, I thought I would write up a bit more about it.

Let's focus on artist to artist recommendation for now. If we look at last.fm for recommendations for New Order, here is what we get.

Artists similar to New Order, from last.fm

Similarly, using the Echonest API for similar artists, we get back an ordered list of artists similar to New Order, including Orchestral Manoeuvres in the Dark, Depeche Mode, etc.

Now, let's play word associations for a few bands and musical genres. My colleague Michael Smethurst took the Sex Pistols, Acid House and Public Enemy, and draw the following associations:

Sex Pistols associated words

Acid House word associations

Pubic Enemy word associations

We can see that among the different terms in these diagrams, some refer to people, to TV programmes, to fashion styles, to drugs, to music hardware, to places, to laws, to political groups, to record labels, etc. Just a couple of these terms are actually other bands or tracks. If you were to describe these artists just in musical terms, you'd probably be missing the point. And all these things are also linked to each other: you could play word associations for any of them and see what are the connections between Public Enemy and the Sex Pistols. So how does that relate to recommendations? When recommending an artist from another artist, the context is key. You need to provide an explanation of why they actually relate to each other, whether it's through common members, drugs, belonging to the same independent record label, acoustically similar (if so, how exactly), etc. The main hypothesis here being that users are much more likely to be accepting a recommendation that is explicitly backed by some contextual information.

On the BBC website, we cover quite a few domains, and we try to create as much links as possible between these domains, by following the Linked Data principles. From our BBC Music site, we can explore much more information, from other BBC content (programmes, news etc.) to other Linked Data sources, e.g. DBpedia, Freebase and Musicbrainz. This provides us with a wealth of structured information that we would ultimately want to use for driving and backing up our recommendations.

The MusicBore I've described earlier on this blog kind of uses the same approach. Playlists are generated by following paths in Linked Data. Introduction of each artists is done by generating a sentence from the path leading from the seed artist to the target artist. The prototype described in that paper from the SDOW workshop last year also illustrates that approach.

So we developed a small prototype of these kind of ideas, rqommend (and when I say small, it is very small :) ). Basically, we define "relatedness rules" in the form of SPARQL queries, like "Two artists born in Detroit in the 60s are related". We could go for very general rules, e.g. "Any paths between two artists make them related", but it would be very hard to generate an accurate textual explanation for it, and might give some, hem, not very interesting connections. Then, we just go through these rules on an aggregation of Linked Data, and generate recommendations from them. Here is a greasemonkey script injecting such recommendations with BBC Music (see for example the Fugazi page). It injects Linked Data based recommendations, along with the associated explanation, within BBC artist pages. For example, for New Order:

BBC Music recs for New Order

To conclude, I think there is a really strong influence of traditional information retrieval systems on the music information retrieval community. But what makes Google, for example, particularly successful is to exploit links, not the documents themselves. We definitely need to go towards the same sort of model. Exploiting links surrounding music, and all the cross-domain information that makes it so rich, to create better music recommendation systems which combine the what is recommended with the why it is recommended.

Tuesday 12 May 2009

Yahoo Hackday 2009

C

We went to the Yahoo Hackday this week end, with a couple of people from the C4DM and the BBC. Apart from a flaky wireless connection on the Saturday, it was a really great event, with lots of interesting talks and interesting hacks.

On the Saturday, we learned about Searchmonkey. I tried to create a small searchmonkey application during the talk, but eventually got frustrated. Apparently, Searchmonkey indexes RDFa and eRDF , but doesn't follow <link rel="alternate"/> links towards RDF representations (neither does it try to do content negotiation). So in order to create a searchmonkey application for BBC Programmes, I needed to either include RDFa in all the pages (which, hem, was difficult to do in an hour :-) ) or write an XSLT against our RDF/XML representations, which would just be Wrong, as there are lots of different ways to serialise the same RDF in an RDF/XML document.

We also learned about the Guardian Open Platform and Data Store, which holds a huge amount of interesting information. The license terms are also really permissive, even allowing commercial uses of this data. I can't even imagine how useful this data would be if it were linked to other open datasets, e.g. DBpedia, Geonames or Eurostat.

I got also a bit confused by YQL, which seems to be really similar to SPARQL, at least in the underlying concept ("a query language for the web"). However, it seems to be backed by lots of interesting data: almost all of Yahoo services, and a few third-party wrappers, e.g. for Last.fm. I wonder how hard it would be to write a SPARQL end-point that would wrap YQL queries?

Finally, on Saturday evening and Sunday morning, we got some time to actually hack :-) Kurt made a nice MySpace hack, which does an artist lookup on MySpace using BOSS and exposes relevant information extracted using the DBTune RDF wrapper, without having to look at an overloaded MySpace page. It uses the Yahoo Media Player to play the audio files this page links to.

At the same time, we got around to try out some of the things that can be built using the linked data we publish at the BBC, especially the segment RDF I announced on the linked data mailing list a couple of weeks ago. We built a small application which, from a place, gives you BBC programmes that feature an artist that is related in some way to that place. For example, Cardiff, Bristol, London or Lancashire. It might be bit slow (and the number of results are limited) as I didn't have time to implement any sort of caching. The application is crawling from DBpedia to BBC Music to BBC Programmes at each request. I just put the (really hacky) code online.

And we actually won the Backstage price with these hacks! :-)

This last hack illustrates to some extent the things we are investigating as part of the BBC use-cases of the NoTube project. Using these rich connections between things (programmes, artists, events, locations, etc.), it begins to be possible to provide data-rich recommendations backed by real stories (and not only "if you like this, you may like that"). I mentioned these issues in the last chapter of my thesis, and will try to follow up on that here!