dbrec - Intelligent music recommendations for and from the Web of Data
Mon, 2010-02-01 15:02 — alex
In addition to the Social Semantic Web, you probably know that one of my main research interest concerns Linked Data, not only in publishing but also in consuming it. And well, I also enjoy music and the possibilities that LOD offers in that context, as we've wrote with Yves mid-2008.
So, I recently worked deeper on the use of Linked Data for music recommendations and I'm happy to announce dbrec, a service providing recommendations for the 39,000+ artists available in the DBpedia dataset (i.e. identified as instances of dbpedia-owl:MusicalArtist or dbpedia-owl:Band). The recommendations are computed using an algorithm for Linked Data Semantic Distance and take into account the various links that connect two resources, either directly (e.g. artists having played together) or indirectly (e.g. being on the same label or having covered the same song). Moreover, dbrec, explains the recommendations to the user, by keeping in mind the various links that have been used to compute the recommendations. For instance, the following screenshot shows why Big Brother and the Holding Company is suggested for a search on Janis Joplin.
dbrec is fully based on Semantic Web and Linked Data technologies and, in addition, exposes all the recommendations publicly (under a Creative Commons license) in RDFa using the dedicated LDSD ontology. For more details, you can check the homepage of the service, and start exploring the recommendations. Hey ! Ho ! Let's Go !
Comments
Great work, and yet...
This is cool work, and the inclusion of explanations is great. As a semantic-technology demonstration, it's intriguing. And yet, as a music-recommendation system, ignoring the technology, it's pretty bad. Poking around at various examples in various genres, I find that the only decent recommendations are when personnel overlap happens to lead to stylistically similar bands. Everything else leads to randomness, and even the personnel connections produce garbage plenty of the time. Compare dbrec's list for any band to any other recommendation system you like. dbrec loses badly in every case I tried.
This is not a condemnation of your technical work, but it's a clear statement about the weakness of your data for the purpose to which you're trying to put it. The current musical-artist connections in dbpedia are not meaningful enough to be the basis of a musically-plausible recommendation system. Bad data + good processing = bad analysis.
And in an academic/technical context, this would be an interesting finding. But if you're going to promote this to a music audience, as opposed to a semantic-web audience, you should think very carefully about whether "intelligent music recommendations" is what you really want to say at the top. That is a statement about how you wish people to evaluate your work, and one that I think will lead to people who care about music forming a low opinion of your idea of "intelligent", and by association a low opinion of this "semantic technology" you say underlies it.
glenn mcdonald
Hi glenn, thanks for that
Hi glenn, thanks for that inspiring comment.
I (obviously) don't think the recommendations are but, while you may come up with weird things, as geolocation-relatedness that may not make sense at all from a musical-taste point of view, or for bands / artists where the DBpedia linkage is weak - and so are the recommendations. However, you know why these artists are recommended and you can make your mind regarding the recommendation. I'm also planning to improve the UI to provide more control on the explanation features, by filtering some properties that one do not consider as relevant for the explanation. In general, from the current evaluation I'm running, it compares well with other systems (I hope to get the full evaluation published in the future).
In general, I do not think the DBpedia data and the links are too bad. They may not be well-suited for music-similarity based on the music signal itself, indeed, but as the algorithm is domain-agnostic, it can be applied as well to a more "music-oriented" dataset (i.e. modeling rythm, beat, etc.) with the MO as suggested in another comment, to give more signal-oriented comments.
Oh and finally, I still think that "intelligent" is a key point here, by explaining these recommendations and providing these explanations openly for further reuse.
LDSD?
Are you going to describe the algorithm, or is it secret?
For comparison, here's a comparison metric I invented ("invented" being a rather grandiose word for pretty simple math) and use:
http://www.furia.com/page.cgi?type=log&id=282
As for "intelligent", I think your explanations are a great idea, but transparency is not the same as intelligence!
Congrats
Hi, I'm happy to explore this pretty good use case of LOD. As far as I can see there is no usage of the Music Ontology. That would be the next great step, but hey - one step after another ;) Well done!
Cheers,
zazi
Thanks. Indeed, so far I'm
Thanks. Indeed, so far I'm only relying on DBpedia (and I'm not sure there are any MO links used in the dataset) but applying the same algorithm to a dataset based on MO would definitely also be interesting.