kurtisrandom

LinkedBrainz Project

2010-08-10T08:01:00.000-07:00

I've recently started working on a new JISC funded grant to publish MusicBrainz Next Generation Schema and Advanced Relationships as Linked Data. The project is called LinkedBrainz and I will be neglecting this blog even more in favor of posting at the LinkedBrainz site.

No, I have not finished my thesis but hopefully soon :-)

The Return of CatfishSmooth

2010-05-15T17:34:00.000-07:00

I've really let this blog languish lately - my excuses include writing thesis, preparing for birth of daughter (literally any day now), and moving back across the Atlantic.

To make up for it, I've done something that is pretty cool (I hope you'll agree). I've given it the name CatfishSmooth for no particular reason other than the fact it is a funny domain name I happen to own (some of my more astute readers might recall the previous award-winning CatfishSmooth incarnation).

CatfishSmooth is a music artist navigation tool built on the web of linked data. We fuse data from DBpedia, DBTune, Echonest, BBC, Last.fm and MusicBrainz to find arbitrary connections between music artists. For example, if we examine the page for James Brown we see a dizzying array of artist similarities. We see a list of artists who's hometown is also Augusta Georgia, a list of artists how are also American soul singers, a list of artists who are also incarcerated celebrities, and several other lists.

We also get some audio and youtube videos courtesy of the amazing Echonest. And finally, we also include along side the random linked-data connections some more traditional "similar artist" recommendation from the Echonest and Last.fm.

So how does it all work you ask? Most of the random connections come from DBpedia - the linked data translation of Wikipedia. We specifically leverage the YAGO types to create really simple SPARQL queries that yield some interesting lists of artists (i.e. incarcerated celebrities). If you don't speak semantic web, you probably understood very little of that last sentence. So let's look at the big picture.

There's tons of reasonably well-structured data about music artists all over the web. Unfortunately, these various data resources are not always easy to link up. How do we know some webpage is talking about James Brown "the Godfather of Soul" and not James Brown the drummer for the post-rock group Veil Maker??? Disambiguating all this data is really the crux of what CatfishSmooth does and by no means does CatfishSmooth do it alone.

MusicBrainz is the ultimate resource in music artist/album/track disambiguation. A unique MusicBrainz

id is created for each artist, album, and track. A wide variety of metadata is maintained by an active community of users ensuring a high level of accuracy and a wide breadth of coverage. In a sense we treat MusicBrainz as the center of the musical universe. With a MusicBrainz id it is a simple matter find information from last.fm (which used mbids directly) or the Echonest (which supports mbids through the Rosetta Stone project).

MusicBrainz also serves as our entry point to the world of linked data. The BBC uses mbids for its music artist pages (which are awesomely published as linked data). We use the amazing sameAs.org service to go

from BBC mbid URIs to DBpedia URIs. With DBpedia URIs we're plugged into all the knowledge of Wikipedia. Most of our lists come from here although some come from the DBTune.org/musicbrainz resource.

With this matching done, we fire off a slew of API calls and SPARQL queries to generate the different elements of the page. CatfishSmooth currently has absolutely no local data store. Everything is "in the cloud" (oh I hate myself for writing that but it's true). We've sort-of created the views and the controllers of an MVC web application. The model is RDF (and some web APIs) so we can just plug in to linked data. Future incarnations will likely be backed by a triple store that aggregates and collects new data from users.

CatfishSmooth is far from a finished work, but I was keen to release it in its current state because (1) I thought the crowd at SF Music Hack Day might take a liking to it, (2) I think it's pretty fun to use as-is, (3) I'm not sure how much time I will have to continue working on it in the very near future. Planned future improvements include:

better media support - more youtube videos, play.me, playdar, and/or spotify support
semantic playlist building - another idea that probably warrants an entirely new post
more links! - from BBC, MySpace, Discogs, and others

One final note, CatfishSmooth is a bit of an experiment. It is prone to crash and you will probably see the python stack trace when it does. I need to setup a bug reporting system also. So take it for what it's worth. Also you'll find it is not so useful for "long tail" artists - the connections are really only found for artists that have detailed MusicBrainz and Wikipedia entries. But hopefully this thing is still fun and useful. I would love to get any ideas or feedback.

Enter HeVeA

2010-01-22T15:32:00.000-08:00

In my last post, we discussed problems with converting LaTeX documents into HTML. I briefly described DocBook, the most compelling alternative I found to LaTeX. I got a lot of response re: DocBook from people telling me I am crazy and what horrible experiences they'd had with DocBook. But I was undaunted, and I forged ahead. I have converted my ISMIR 2009 paper into DocBook and then to HTML. I did this in about half a day, mostly converting LaTeX commands to DocBook XML by hand using the Emacs nXML mode.

You can view the results at http://docs.kurtisrandom.com/ismir2009-docbook/

Or you can download a tarball with my source as well. I applied the most basic docbook-xsl transform. On Ubuntu, you can apt-get docbook-xsl and Macports seems to include this stuff as well. I used xsltproc to apply the transform with some arguements:


xsltproc --output index.html \
--stringparam bibliography.numbered 1 \
--stringparam bibliography.collection \
     ./bib.xml \
/usr/share/xml/docbook/stylesheet/nwalsh/html/docbook.xsl \
db-ismir09.xml

We are using --stringparam to pass some optional parameters to the docbook.xsl to get a numbered bibliography and to specify the external bibliography file. These a complete reference list of these parameters is provided on the docbook sourceforge page.

You may notice, not everything went according to plan. I had to spend a lot of time converting the bibliography by hand - the tools I found to do this didn't really work. So in the end, I left the bibliography a bit of a mess with some left-over latex in there.

Furthermore, the PDF version was a bit of a disaster. I used the dblatex utility to generate the PDF as follows:


dblatex --output=db-ismir09.pdf -T simple -L bib.xml db-ismir09.xml

While I was able to get the Docbook tool chain up and working rather quickly, I did have lots of questions. My posts to the mailing list and to the IRC channel were virtually ignored (I had one pleasant reply from a jsmith on IRC pointing me to a masters thesis written in docbook which was helpful but includes no citations or bibliography - the thesis is really a technical software documentation). This was really off-putting for me and (sigh) has led me back to LaTeX.

Luckily, thanks to my long-time friend and colleague Ben Fields, I discovered HeVeA. This is more modern and complete LaTeX to HTML package that seems very promising. I was able to convert the same ISMIR 2009 paper to HTML in a manner of minutes.

http://docs.kurtisrandom.com/ismir2009-hevea/

This actually looks much better than the Docbook version IMHO. The only major problem I see is that the URLs in my LaTeX don't seem to automagically become href's in the HTML as I had hoped. There was also some funkiness with the images I had to fix by hand. But this package seems to be well documented and I am optimistic.

In addition to HeVeA, Ben suggested TtH which I had tried before and had problems getting it to work (probably user error to be fair). But I hope to give it a real try again this week.

So in conclusion, in my brief experience Docbook is not as horrible as everybody said. But, the Docbook user community seemed to give me the cold shoulder (sorry I'm a lamer noob who has questions). It seems using Docbook instead of LaTeX would, in the end, create a lot more problems than it would solve. I still like the idea of Docbook - it seems so much more "future proof" than LaTeX - but I'm going to have to stick to the beaten path on this one. Looks like LaTeX wins.

UPDATE:

I've managed to use TTH to create an HTML version of the same paper.

http://docs.kurtisrandom.com/ismir2009-tth/

Looks pretty good, maybe the best yet, but not sure what it's doing with the figures - seems to create a link to the image instead of displaying it with an appropriate <img> tag. Not bad at all, but it seems HeVeA allows more control and is better documented.

Beyond TeX? Discovering DocBooks

2010-01-21T06:01:00.000-08:00

As you may or may not know, it is about time for me to start writing my PhD thesis. In preparation for writing what will be the longest and perhaps most important document I've ever written, I've been exploring what document authoring tools are available.

Now most PhD students would simply use LaTeX and get on with it. LaTeX is a document markup language front end for the 30 year-old typesetting solution know as TeX. The intent of these tools has always been to allow anyone to create good-looking printed documents. And these TeX-based tools perform this task rather well despite their ancient syntax and tedious compilation work flow. But while LaTeX, ConTeXt, and other tools based on TeX are definitely compelling in their handling of reference citations and typesetting for print, I find they fall short in one very important respect - the Web. Compiling LaTeX documents into HTML is a cumbersome and error-prone process. The best tool I have found for rendering LaTeX documents for the Web is by far TeX4HT. But this tool is not without it's quirks and, out-of-the-box, it does not handle some very common LaTeX tags including \footnote and \mbox. I my brief experiments I've been unable to render all but the most basic of LaTeX documents as HTML using TeX4HT. Perhaps there is a helpful guide on this topic I am missing...

My thesis is about music and the Web. Therefore, I want to write my thesis to be published on the Web. For me, Web publishing is not an after thought - it is a priority. My experience with TeX4HT has left me searching for a solution not based on LaTeX despite a lengthy round of humorous taunting from my peers.

I have gotten some interesting suggestions regarding LaTeX alternatives. Fellow Linked-data-enthusiast Tom Morris suggested LyX (with some reservations) - however, it is basically a GUI on top of LaTeX/TeX and does not really allow me to write for the Web. There were also some murmurrings of LyX being a bit buggy and not well-suited to creating a very large document. Other suggestions were generally some TeX derivative (i.e. ConTeXt) or some software that makes LaTeX authoring less painful (i.e. RecSys heavy-weight Xavier Amatriain suggested Kile which is quite good, but LaTeX pain was never really my problem anyway) or some alternative document conversion solution (i.e. Pandoc).

Finally, Chris Lowis from the BBC's Web team suggested DocBook. DocBook was one of those things I had heard about vaguely long ago but I didn't really grok it. So I did a bit of research, starting with the Wikipedia page on DocBook. Put simply, DocBook and XML-based (or SGML-based) semantic markup language for writing documentation. It is "presentation neutral" meaning that the same source can easily be converted to HTML, XHTML, or PDF. Using an XSLT style sheet, you can convert DocBook source into just about anything. In fact, it seems the preferred method for creating PDFs actually involves a DocBook->LaTeX->PDF conversion. I've had some experience writing and applying XSLT to create RDF/XML from some proprietary XML schema. I found XSLT to be relatively easy - much less confounding than TeX style files anyway.

So the great experiment begins. I will be trying out DocBook over the next few days (weeks?). So far I've found writing DocBook source with Emacs (or Aquamacs on OS X) in nXML mode to be a joy - I've got autocomplete and validation right there as I type. Tonight I will attempt to do some rendering to PDF. Depending on how painful this process proves... well, we'll see.

Will Kurt fall in love with DocBooks? Will he run back to LaTeX in tears begging for forgiveness? Will he complete his thesis on time? Stay tuned for the exciting conclusion!

UPDATE:

Had some early, not-so-encouraging but hilarious responses re: DocBook.

From IRC:

" [...]: I chuckled at your thing about latex and docbook just now. I wrote the ------ user manual entirely in docbook and it was the most painful experience imaginable. I would never, ever, ever, ever attempt that again unless I had a nice user interface to make all the xml invisible to me" -(some badass developer)

"[...]: docbook is terrible. we used docbook for iteration 1 of the ----- manual, too. it was almost impossible to assemble a toolchain to compile it." -(some guy with long hair)

From twitter:

"@kurtjx Dude, if you're considering DocBook to get around LaTeX authoring woes... UR DOING IT RONG." -(another guy with long hair)

Echonest Artist Graph

2009-11-26T04:10:00.000-08:00

I've just recently had a chance to mess around with the Echonest data set that Paul Lamere and Justin Donaldson presented in their tutorial on Music Visualization at ISMIR 2009. Unfortunately I wasn't able to attend the tutorial because we were in the other room talking about Linked Data and music. But I'm a big fan music viz. I wanted to do a "remote hack" using the data set for Boston Music Hackday ala Adam Lindsay's Remix Shell but alas the weekend got away from me.

But finally I made some pretty pictures.

I suppose first we had better discuss the data set. This is a "small" dump from the Echonest API that includes about 70k artists and their connections to "similar artists" as decided by Echonest (which in my experience does a pretty good job). The dataset also includes Echonest's notions for hotttness and familiarity values for each artist.

Now about the picture. We take the largest connected component (turns out that includes about 54k artists) and create an undirected graph where each artist is a node and each similarity relation is an edge. We use the awesome igraph library to handle the graph structure. Luckily igraph includes some nice graph layout options including the super fast Dynamic Recursive graph Layout algorithm (DrL) which we use here. This is similar to what Tamas did for the Last.fm graph. Also note the color corresponds to the "hotness" - hotter artists are more red - and the size corresponds to "familiarity" - the more familiar artists are larger.

Of course it would be nice to see some artist names and some more interactivity, but hopefully that's coming soon along with some interesting details about the structure of the Echonest artist network. But now it's turkey time :-)

update: We've got names on the artist nodes now. You can zoom in using the incredible embedded seadragon widget above!

Note you can embed the above interface with this code:


<script src="http://seadragon.com/embed/gwo.js?width=auto&height=400px"></script>

More on the Similarity Ontology

2009-10-28T17:15:00.000-07:00

I recently gave a talk at ISMIR 2009 about MuSim the similarity ontology. The talk was well attended and well received, but I wanted to review the core concepts and promote the demo implementation as well.

In a nut shell, similarity is very complex and when we say two things are similar we really mean that they are similar in some specific sense. So our ontology focusing on reifying similarity - not saying just what things are similar, but also how was it determined they are similar. As such, we treat similarity as a class rather than a property.

We define a class sim:Similarity for describing similarity statements. We then define the class sim:AssociationMethod for describing a method for determining similarity. By associating a similarity statement of type sim:Similarity with a method of type sim:AssociationMethod we are describing in what sense the elements involved in our statement are similar. We can further reify our method by providing provenance (who made the method) and even fully disclose our method by pointing to a graph describing our workflow. The diagram below illustrates a basic example involving two music tracks.

Here is the same example in Turtle:


@prefix sim: <http://purl.org/ontology/similarity/> .
@prefix mo: <http://purl.org/ontology/mo/> .
@prefix foaf: <http://xmlns.com/foaf/0.1/> .
@prefix : <#> .

:TrackA a mo:Track .
:TrackB a mo:Track .

_:SimilarityStatment a sim:Similarity ;
sim:element :TrackA ;
sim:element :TrackB ;
sim:method :TimbreSim .

:TimbreSim a sim:AssociationMethod ;
foaf:maker :Me ;
sim:description <http://some.graph.uri> .

:Me a foaf:Person .

Hopefully this example is somewhat illustrative, although if you're not familar with RDF or the Turtle syntax for serializing RDF it's probably not terribly useful. The important thing to understand is that :TimbreSim is our method for deriving our similarity statement. We specify the person who made this methdo (:Me) and we point to a Named Graph that describes our workflow. We could also attach a textual description and other information.

Ofcourse this example is only illustrative and not very realistic. Enter our demo implementation classical.catfishsmooth.net . Here we've taken some public domain classical music found on Musopen as well as some information about classical composers and their network of influence as specified by the Classical Music Navigator - this gives us a source of composer-to-composer similarity. We also use the Sonic-Annotator and some of the QMUL VAMP Plugins to analyze the audio files and get a couple methods of determining track-to-track similarity.

In this implementation, we have 3 similarity methods:

Note that MuSim would not be very interesting without the SPARQL query language. We can query a collection of similarity statements and combine similarity methods in interesting ways. In our catfishsmooth implementation we store our similarity triples and related data in 4Store - a scalable open-source RDF store from Garlik. This allows us to include a SPARQL endpoint and, even better, a Snorql SPARQL explorer. In the Snorql explorer, you will find a series of example queries on the right-hand side of the interface. The most advanced query allows us to combine all three similarity methods. We use a recording of Wagner's piece Der Meistersinger von Nürnberg as the seed. We find other tracks that are in the same key, similar by timbre, and composed by composers who had influenced Wagner.

I've had some exciting (but brief) discussions with Mert Bay and Stephen Downie about integrating MuSim with the MIREX framework such that particpants in the audio similarity task might optionally apply their algorithm to some creative commons dataset (probably Jamendo) and publish MuSim data for re-use and additional evaluation.

We haven't really covered here the workflow description syntax or concepts. This will be the focus of much future work. Our design leaves this specification open-ended, but we have used the N3-Tr framework developed by Yves Raimond in his PhD thesis to describe workflows. You can see some examples in our implementation.

For more details on MuSim you can read our ISMIR paper (pdf) or view the ontology specification.

Snorql and SNARPL

2009-09-29T05:45:00.000-07:00

I'm back sliding in just before the end of September to maintain my once-a-month posting pace. This will be a short post to informally anounce the birth (or re-birth) of the Snorql/SNARPL project. Snorql is an AJAXy front end create by Richard Cyganaik to allow people to test SPARQL queries against a triple store. Snorql allows you to 'dive right in' to SPARQL. You have used Snorql if you've ever used the SPARQL Web interface for a D2R Server instance such as the DBtune.org/musicbrainz interface. Snorql has a lot of nice features such as built-in namespaces and the ability to view the query and the results simultaneously. It turns out, it is rather easy to drop Snorql on top of any triple store that supports JSON result formats - for example Openlink Virtuoso (see DBPedia.org/snorql/) or Garlik's 4Store.

I've created a hacky mash up of Snorql and Soundmanager 2 I call SNARPL (Snorql Network Audio Results PLayer). Basically, any result that returns links to audiofiles on the Web will be 'playable' - click the link and hear the audio.

You can check it all out on Github: http://github.com/kurtjx/SNORQL

Warning: as of this writing the SNARPL code is not very pretty or well documented.

“Total Eclipse Of The Heart” Flow Chart

2009-08-18T10:19:00.001-07:00

This just goes to show that music is all about connections on all levels ;-)

Listen to Bonnie Taylor's "Total Eclipse of the Heart" and follow the flow chart:

ISMIR tutorial site is up

2009-07-10T01:19:00.000-07:00

In case you missed it, we've launched a site to serve as the resource for our tutorial at ISMIR 2009 in Kobe Japan. The tutorial is called Share and Share alike - you can say anything about music with Linked Data. The tutorial will focus on applying Semantic Web technologies to music informatics. If you can't make it to Kobe, this site will hopefully still be very useful and interesting. If you would like to suggest any web resources please tag them on delicious as linked_music_data. Enjoy!

sparql myspace

2009-06-22T05:46:00.000-07:00

As part of the dbtune.org myspace service, I've started caching data in a sparql endpoint. Anytime anyone queries the service, the rdf is dumped into a triple store. When the same query is repeated, the data just comes from the triple store. If the data about a particular artist is more than 30 days old, the page is re-translated from myspace.

Here is an example query that lists all artist users from Hungary by name in order of total friends.

PREFIX mysp:<http://purl.org/ontology/myspace#>
PREFIX foaf:<http://xmlns.com/foaf/0.1/>
PREFIX mo:<http://purl.org/ontology/mo/>

SELECT ?name ?friends from <http://dbtune.org/myspace/>
WHERE {?artist mysp:country ?country ;
a mo:MusicArtist ;
foaf:name ?name ;
mysp:totalFriends ?friends .
filter ( regex(str(?country), 'Hungary') )
}

UPDATE: we now use foaf:based_near and Geonames URIs for countries related to music artists. This means the SPARQL Query is a bit different and much faster as we no longer need the REGEX:
PREFIX mysp:<http://purl.org/ontology/myspace#>

PREFIX foaf:<http://xmlns.com/foaf/0.1/>
PREFIX mo:<http://purl.org/ontology/mo/>

SELECT ?name ?friends from <http://dbtune.org/myspace/>
WHERE {?artist foaf:based_near <http://sws.geonames.org/719819/> ;
a mo:MusicArtist ;
foaf:name ?name ;
mysp:totalFriends ?friends ;
mysp:genreTag <http://purl.org/ontology/myspace#Hip%20Hop> .
}
ORDER BY DESC (?friends)

You can see the results here or you can try your own queries using the endpoint here.

Due to human error on my part, the first 6.5 million triples do not include the genre tags for music artists. Terribly sorry about that. Also, we do not include the 'total plays' information as we are consistently getting the wrong value for this field for some mysterious reason. In the end, my PhD is not _only_ about reverse engineering the myspace website ;-)

I'll keep this thing running until we're low on disk space or the server catches fire, which ever comes first.

All the code for the service is in the motools project. I'm using CherryPy to handle requests, a strange menagerie of BeautifulSoup, Regex, and string matching for the screen scraping (yes I've heard of XSLT), and the Chris Sutton classic MoPy for the RDF serialization. The backend triple store is Virtuoso which we connect to using the ODBC interface.

We will talk about uses for this resource and others during our upcoming tutorial at ISMIR 2009 in Kobe Japan. If you can't make it, don't worry, the website for the tutorial is coming soon and we plan to _everything_ there.

Open Hackday London wrap up

2009-05-14T03:48:00.000-07:00

UPDATE: catfishsmooth has been offline for a while now because Myspace finally changed the way they host their music. Sorry guys. Hopefully we'll return in another form sometime in the near future.

Last weekend I participated in the Yahoo! sponsored Open Hackday London. A good time was had by all. Your typical hackday setup - loads of free cola and food, presentations about various APIs and such, and then a friendly hack competition.

Our hack was entitled Boss of Myspace. We actually walked away with the BBC Backstage prize and even got some positive reviews as well. I created what was basically a human-readable version of our dbtune.org/myspace service. Yves contributed some SPARQL for converting mo:available_as properties into XSPF playlists. However, in the end we used the Yahoo! Media Player which was happier just swallowing embedded html links to audio files. We also used the super cool Yahoo! BOSS search API on the front end to allow for a more fuzzy search - you can enter something close to the artist name and still get the results you want (hopefully).

Ben Heitman from Deri contributed some minimalist css just in time for the deadline. Finally, Ben Fields contributed a really nice feature (post deadline actually working while the other projects at the Hackday were being presented). Ben used the Echonest API to get similar artist recommendations. Just today, I managed to integrate these in the service with some of my first jQuery coding ;-) They are clickable and seem to work really well for head, medium, and hi-tail artists. Planning to present recommendations from the Last.fm API side-by-side in the future as well as recs from other sources.

Please note that I _think_ the streaming audio will not work in the States - strange thing I can't explain but have encountered before when working with Myspace audio streams. Let me know if anybody State-side can test this and confirm.

Here are some of my personal favorite entry points:

Enjoy!!!

And oh yeah! Yves made a nice little RDF-based hack as well that he describes on his blog.

WebSci09 report

2009-04-01T04:22:00.000-07:00

I recently attended WebSci09 in Athens Greece. This was the first conference of it's kind focusing on web science as a new research discipline. See WSRI's nifty cross-discipline intersection diagram below:

The conference was quite engaging and interesting and rather cross disciplinary. I met a lot of really cool people and had some good discussions about my new pet project MuSim, but more about that in another post. All (or what seems to be all) of the papers are available through WSRI's cool new experimental on-line journal. A few papers I found particularly interesting include Patricia Victor et al's Trust- and Distrust-based recommendations, Denny Vrandecic et al's analysis of new features and user responses in Wikipedia, and Yeung and Noll's approach to measuring expertise in collaborative tagging systems.

And let us not forget one of the few music-centric entries in this inagural WebSci conference, Jeff Pan and Stuart Taylor's very cool but very alpha semantic-web-powered MusicMash. Using their own TrOWL backend, the site aggregates data about a particular music artist - throws it in a triple store as Music Ontology RDF - and serves up a nice human-readable webpage as well. Although I must say, I've had mixed success playing with the website.

Actually my poster presentation was not strictly music-related. I presented the k-pie graph layout algorithm which allows for the visualization of a network where the nodes have a set of semantic labels associated with them. However my main application and demo where very music-centric. The visualization for a sample of Myspace artists and genres looks something like this:

I have created an open source implementation that uses the Jung framework if you're interested you can get it from this public svn. I also provided an interactive demo at the conference that allowed one to click through the graph and listen to music. The code for that is _really_ hacky and I'm not releasing it yet, but if you're really keen just contact me. Also, I'm not convinced it was truly useful as a visual music discovery interface - really wish I could attend Paul and Justin's upcoming tutorial at ISMIR 09 but I'm afraid I'll be busy presenting another tutorial ;-)

Just to wet your appetite a bit, here's a screen shot of what the interactive demo looks like - fun to play with, but really needs a lot of work.

plz click to enlarge.

In summary, WebSci09 was a success, met lots of cool people and I believe there are many more successful WebSci's on the horizon.

Classical Music in the Web of Data

2009-02-05T07:33:00.000-08:00

I have finally re-published the classical music composer influence data set I've been working with in a proper Linked Data fashion. You might remember this data set from this relatively recent post or this much older post where I describe some complex networks statistics and the original collection of the data set. This rather modestly sized data set contains just under 8,000 triples. Of course you can make SPARQL queries against the data here or you can browse around w/ your favorite data browser or even your plain-old html web browser. I've implemented content negotiation using the data-resource-page paradigm you might recognize from DBpedia.org.

I've done most of this with Openlink Virtuoso - the all singing all dancing web of data server solution. Although some things in Virtuoso didn't work exactly right at first and it's a bit rough around the edges, it is an awesome piece of software. It's so feature-rich it can be a bit daunting, but the Openlink guys are quick to help on the mailing list.

In the near future we will be publishing our MySpace data set from ISMIR 2008 in a similar fashion...

Last post of the year

2008-12-30T13:42:00.000-08:00

I have been slacking on the blog posts and I really have no excuse. I've been working on a GUI for browsing structured data about classical music composers. There is a very alpha version of the software available here.

I've also been busy with my new favorite CMS drupal working on the omras2 website. You can browse a fancy list of omras-related publications here.

Finally, really big news in my personal life - I got married to my lovely wife Larisa!!!

Sparqling a funk legend

2008-10-28T05:24:00.000-07:00

A while ago I had the privilege of witnessing funk pioneer, JB Horns member, and life-long saxophone badass Maceo Parker perform. It was fantastic. Maceo led the band with swagger and style - using many of the same hand gestures and moves Soul Brother #1 used to lead the Horns back in the day. He finished his encore with the funk classic "Pass the Peas".

Afterwords I was curious - did Maceo write "Pass the Peas" or was it Fred Wesley or maybe Pee Wee Ellis??? So I decided to ask the Semantic Web. We'll use the SPARQL (pronounced "sparkle") to make queries. There is a nice gentle tutorial from IBM here which also describes using the Java-based Jena package. If you're a python type you might try sparql-wrapper. However, to start out and to see what data is where you'll probably want to use a web-based interface (most SPARQL endpoints will have some web-based interface where you can enter a query). We'll start with the Virtuoso SPARQL endpoint for DBpedia to see what wikipedia knows about Maceo Parker...

PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
SELECT DISTINCT ?Concept
WHERE 
{<http://dbpedia.org/resource/Maceo_Parker> 
        rdf:type ?Concept}

If you're familiar with SQL you might be able to guess what this means. We're asking DBpedia to tell us all the concepts related to Maceo Parker. The "PREFIX" keyword is setting up a namespace. Then we are going to "SELECT" any values of the variable "?Concept" that fit the graph pattern we establish in our "WHERE" clause. That pattern is Maceo_Parker rdf:type ?Concept - or any concept that describes what type of entity Maceo Parker is. You can copy and paste the query into the DBpedia interface and see that Maceo Parker is a Person, a AfricanAmericanMusician, a AmericanJazzSaxophonist, and a number of other things.

Well that's a good amount of information right there. We can infer we've got the right Maceo Parker but unfortunately DBpedia doesn't have a lot of discography data, so we can't really answer our original question. Luckily my good friend, former colleague, and Semantic Web mentor Yves Raimond created a SPARQL endpoint that queries the Musicbrainz database. The service is hosted on DBtune.org/musicbrainz and a web interface can be found here.

So let's construct a new query.

PREFIX owl: <http://www.w3.org/2002/07/owl#>
PREFIX foaf: <http://xmlns.com/foaf/0.1/>
PREFIX dc: <http://purl.org/dc/elements/1.1/>

SELECT ?maceo ?songs ?title WHERE {
?maceo owl:sameAs 
  <http://dbpedia.org/resource/Maceo_Parker> .
?songs foaf:maker ?maceo .
?songs dc:title ?title .
}

Again, we start with a few namespace prefixes. This time we're going to select three different variables - "?maceo", "?songs", and "?title". We're using the same resource URI for Maceo Parker as we did when querying DBpedia. However, the DBtune interface refers to Maceo with a different URI. But the DBtune/musicbrainz is inter-linked with DBtune using the "owl:sameAs" property. So the first line of our WHERE clause will find DBtune/musicbrainz URIs that are the same as the DBpedia URI for Maceo Parker. The second line will fill the variable ?songs with URIs for entities for which Maceo Parker is the "foaf:maker". In DBtune/musicbrainz the foaf:maker property is use to associate a song with the song writer. Finally in the third line of our WHERE clause we retrieve all the titles of the songs made by Maceo Parker. Again, you can try to query yourself here - you will find that Maceo Parker is indeed credited with writing "Pass the Peas" as well as a quite lenghty list of other funk gems.

ICMC 2008 Belfast

2008-09-01T01:29:00.000-07:00

I spent last week in Belfast attending the 2008 ICMC. Here's the view from the flat in which we stayed...

There was a very full concert schedule as well as a really diverse and well-paced academic program. There are of course a few papers that seemed to really standout. Maybe I'm falling for a bit of a gimmic, but I was really impressed with Nick Collin's generative music system Infno. The system basically generates electronic dance music automatically. The few examples Nick generated live during his talk were actually really impressive - strong in melody and rhythm but very much lacking in sound design. Nick suggested that his generative electronic dance music algorithms could be used to create "ground truth" data for MIR classification tasks - a very interesting idea but there be dragons.

George Tzanetakis gave a nice demo of Marsyas-0.2. I briefly looked at Marsyas a long time back but never really used the stuff. Now that George has created some really nice Python bindings, I think Marsyas is going to become a big part of my work. George demonstrated the bindings by doing some live coding in Python - haha, take that you mother Chuckers!!!

Bob Sturm from UC Santa Barbra quite deservingly won the best paper award (sorry can't find the PDF online). Unfortunately, I missed the talk (@9am) but there seems to be a lot of buzz around the "atomic decomposition" method he uses for analysis and synthesis of audio. I reckon this will become a hot topic in the MIR community.

Also, the SoundSpots installations by Rob van Rijswijk and Jeroen Strijbos was my favorite artistic piece of the conference. A series of parabolic speakers are hanging from the ceiling such that you can only really hear the spearkers' output when standing directly below them. Additional loud speakers around the room are playing one of several brilliant compositions artfully crafted specifically for the installation - truly a top-notch sound installation.

And finally a big-up to Henry Penttinen, Nick Bryan, and the rest of the CCRMA kids for the Mobile Phone Orchestra. Brilliant.

2008-08-20T06:37:00.001-07:00

Piano Lab - andrew robertson

Mr. Robertson jams on piano guts
Piano Lab 01

the piano lab
Posted via Pixelpipe.

A strange and wonderful flower

2008-08-07T09:39:00.000-07:00

Can anybody help me identify this lovely flower that just bloomed a few weeks ago in my garden? I've recently moved to a new flat and the previous tenant was obviously an avid gardener. I feel obligated to do my part to maintain where I can...

I feel like maybe this is a plant that is not really at home in the London climate. Will it die in the winter? I can't bring it inside as it is definitely planted directly in the garden...

Periodic spectral re-ordering of genre tags based on network structure

2008-07-25T07:17:00.000-07:00

Phew!!! that title is a mouth-full. But basically all we're doing here is re-ordering the list of genre tags found on myspace based on ''top friends'' connections. If people who tag themselves as ''Death Metal'' tend to have friends who tag themselves as "Grindcore" then these genres will appear close to each other in the list. Follow the link for an interactive scroll through the genre list (sorry couldn't sort out embedding the processing applet in blogger :/)

- scrolling list from 'friendship' network structure
- scrolling list from co-occurrence network

We use the same methods that are sometimes used to organize nodes in a graph for visualization. Here are nodes are musical genres and our weighted edges are defined by the number of myspace artist friendships span any pair of genres. Alternatively we can use ge In the first step, we build an adjacency matrix of genres A. If artist i and artist j are ''friends'' in our myspace sample, we add one count to A(i,j) and repeat for every artist in our sample space. Once we've completed our adjacency matrix we perform a spectral re-ordering. Conceptually, we want to move the non-zero values in our matrix to the diagonal. Viewing this directly as an optimization problem is intractable for very large matrices, but we can use an eigenvector approach to make things faster.

Turns out we can be even a bit more clever and try to move non-zeros to either the diagonal or the off-diagonal corners. This is periodic spectral re-ordering and requires us to calculate a normalized Laplacian as well as a subdominant eigenvector pair. The 'friendship-based' genre list above is generated using this method. Below you can see the original and periodically re-ordered adjacency matrix:

After the adjacency matrix is re-ordered, we actually have a re-ordered list of genre tags. The tags that have more 'friendship' connections in the network get moved closer together in the list. You may notice the results aren't perfect. Keep in mind we have a small sparse sample and the distribution of genre labels follows nearly a power law - ''Hip Hop'' and ''Rap'' are much much more common than ''Chinese traditional'' or ''Showtunes'' for example. With a larger dataset (we're using about 19k artists here) the genre listing might improve - moving conceptually similar genres closer and dissimilar ones further apart. As it stands we have some interesting combinations - ''Southern Rock'' and ''Chinese Pop'' can be found directly adjacent to each other in the list. Note, however, that these genre tags appear 113 and 39 times respectively while the ''Hip Hop'' tag occurs 25,536 times.

Note we also construct the genre tag adjacency matrix based on co-occurrence - if an artist is tagged as ''Tango'' and ''Salsa'' put a count in the corresponding row/column of the adjacency matrix. Then we have a genre list based on genre tag co-occurrence. Although it is difficult to quantify, this seems to provide a slightly better listing. Perhaps this is a more coherent list than the one created using the friendship network - not as many unusual pairings in the order. What do you think?

Special thanks to Gabriella Kalna from the University of Strathclyde for the periodic spectral re-ordering Matlab code.

Happy Flu

2008-07-22T09:27:00.000-07:00

I'm breaking my non-blogging streak with the Happy Flu. Add this to your own blog and become part of the epidemic ;)

NetSci 2008

2008-06-27T01:45:00.000-07:00

Oh boy have I been busy. Last weekend I was at the Mashed hackfest with Yves and Ben. We did a nice little hack to link personal audio collections to BBC radio shows (see Yves' blog post and Ben's) . Not so sexy compared to some of the other hacks but we did get some blog nods.

Then on Tuesday, I took the train to the lovely and charming East Anglican city of Norwich for the NetSci 2008 conference. It was a really great cross-discipline meeting of minds - people working on problems in molecular biology, physics, sociology, politics, (ehm) music, and probably others I'm forgetting. And guess what - everybody is using network science to tackle all these diverse problems.

I gave a talk about the topology of the Myspace artist network (pdf slides). In the same session there was a nice talk from Anmol Madan (MIT Media Lab) about the Mob Media system which uses hacked iphones to track face-to-face interactions and music sharing habits in a real-world social network. Really interesting work. Also in the same session a nice talk from Annelies Kamran on Networks in Global Politics and a talk from Mason Porter on Communities in the US Congress.

I was also really excited by the periodic spectral re-ordering (pdf) methods presented by the lovely and brilliant Gabriella Kalna from the University of Strathclyde. I have some nice plans to apply these methods to social tags and genre labels (stay tuned for the exciting conclusion;)

The Giant Instrument

2008-05-16T06:42:00.000-07:00

This week C4DM Presents completed work on our sound installation for Shunt Lounge called the Giant Instrument. The concept is you play the instrument by moving in and out of these spotlights. I wanted to give the users a really fun interactive experience and sort-of recreate the feeling you get when you first learn to play a musical instrument. The work was a collaboration between Andrew Robertson, Adam Stark, Jean-Baptise Theibault, Steve Welbourn, and myself. The team really came together and pulled off something great in a short time - exceeding all expectations. Chris Frauenberger and Stuart Battersby took some really great pics of the installation.

Free your Music, the Cash will follow

2008-03-15T06:27:00.000-07:00

The main-stream media is still calling it an experiment, but it seems the donation-download model for music distribution is really taking off. While forward-looking music entities like Magnatune and Jamendo have been using this approach for some time, it has only recently gained the attention of mass media with the donation-download release of Radiohead's In Rainbows. And now NIN's four volume opus Ghosts is making waves as well as making Trent Reznor loads of cash. Of course both these artists have a well-established fan base that borders on fanatical. What about more obscure artists? Or what about those just starting out? Well prior to the release of Ghosts, poet-rapper Saul Williams released Niggy Tardust (which was actually a collaboration with Trent Reznor) using the donation-download model. This seems to have worked out quite well for him with 154,449 downloads and 28,322 donations of $5 or more. And nearly zero marketing costs and, more importantly, no major label interference. Although Saul is an established artist in his own right, he has a considerably more 'underground' appeal than Radiohead or NIN. So while the RIAA has been busy sueing grandmothers and fetuses, these pioneers have shown the world we don't need them and we won't miss them...

Myspace to RDF service on DBTune

2008-03-15T05:35:00.000-07:00

Working with YvesR and BenF, we've created a Myspace to RDF service on dbtune. This service provides URIs and corresponding FOAF and Music Ontology RDF representations. The representation includes top friends, depictions, names, and URIs for available tracks.

See Yves' blog post for more details...

60x60

2008-03-02T04:12:00.000-08:00

From time to time, I actually work on _music_ and not just music-related technologies. Last fall I created a short piece for this collaborative project called 60x60 - where 60 artists create their own 60 second track and the tracks are merged into an hour long musical collage. I think the result is rather nice - very eclectic and diverse yet somehow coherent. Mine is minute 35...

Soul Brother #1

2008-02-24T05:29:00.000-08:00

(1933-2006)

Can we let mob rule run music?

2008-02-15T04:39:00.001-08:00

A lot of attention has been paid to listener preferences with respect to music recommendation, and with good reason. After all, music recommendation systems are meant to support the listener in her effort to find new, enjoyable music. And of course there is loads of data on listening habits. But we must be careful not to forget the experts - the artists, the musicians, the critics, and the musicologists. And we must remember, music is not a popularity contest - it is a complicated facet of our identity as individuals and as a society as well as one of the most important aspects of culture.

I think anybody reading this blog would agree that the artists at the top of the pop charts are often of dubious merit. Is this what happens when leave music to mob rule? Or can we blame the evil record labels of yesteryear for this phenomenon? I think we are all excited about the new digital age where the 'long tail' is accessible and increasingly navigable. But what happens to this long tail when we allow the same mob who put Brittney Spears on the top of the charts to dictate how we populate and search this new digital music space?

In case you think I'm just pretentious and that there is no evidence of problems in the mob rule of music, let's look at a recent study by Salganik, Dodds, and Watts. A group of nearly 15k listeners were presented a collection of about 50 unheard songs. One group was instructed to rate each song with no extra information. The other group was given the same instructions but also the average ratings for each song given by other members of the group. It was found that, "increasing the strength of social influence increased both inequality and unpredictability of success." For the group with social influence, the success of a song was only loosely correlated with success in the control group. The conclusion - the success of a song in a market with social influence depends little on the quality of that song.

So if we rely on techniques based solely on social influence (i.e. collaborative filtering) to make music recommendations what happens in the future? Will we loose our most talented jazz musicians, our historic folk artists, and our great composers in the vast obscurity of the long tail? Will classical music still matter?

Maybe the diversity of interests and the strength of new technologies will prevail - preserving the greatness of our musical heritage and expanding the horizons of our musical future. I'm sure there will always be groups of dedicated listeners who truly care about good music. And by definition, there is always room in the long tail. And maybe I'm going a bit Andrew Keen (who has some really nice points but I don't entirely agree with).

But as potential gate keepers for this new digital music space, why should we risk it? Why should we submit to mob music rule? I don't deny that collaborative filtering and related technologies are tremendously useful and exciting. I am just arguing that because they are inherently social, these methods cannot be our only gateway into the new digital music space.

Do I have a better idea? Well, maybe. This argument is behind my interest in networks of artists. By studying networks of artist collaboration or musical influence we are in a sense studying a map of the cultural context of music. Perhaps navigating these networks in an intelligent way could help power music recommendation systems that are less social and more cultural...

Wagner is the center of the Universe?

2008-02-01T02:09:00.000-08:00

I've just finished a little project mapping the network of classical composer influences published at Classical Music Navigator. This awesome site was published over ten years ago by Charles H. Smith, an accomplished library scientist. The site contains a listing of 444 classical composers. For each composer there are links for "musical influences" and "has influenced" enabling the construction of a directed network of composer influence. Some initial analysis indicates a few interesting things - the graph is not fully connected, the degree distribution follows an exponential decay (more on this to come), and it shows a high level of community structure (C=0.21). When we look at a k-cores visualization of the network, we notice one node conspicuously at the center. This bright center of the classical universe is non other than Wilhem Richard Wagner with a total degree of 93. With the original author's permission, I hope to re-publish this data on the Semantic Web...

Acid Brass Graph

2008-01-29T03:11:00.001-08:00

Oh man I've been busy - writing papers, keeping up with my busy social schedule, organizing my lint collection - no time for blogging. Just so I have a post for January, here's a link to a nice post by Tony Effik - fellow Londoner and social graph enthusiast. He points to the work of Jeremy Deller who in 2004 created a music piece called Acid Brass. Tony points out the accompanying artwork resembles a graph and goes on to do some complex network analysis. Turns out the graph has a relatively low clustering coefficient and Tony points out there is extensive use of "artistic license" in creating nodes and links. Still, I'm glad to learn about this important artistic endeavor - it seems to support my belief that, at least from a cultural perspective, music is best modeled in terms of connections.

David Byrne on music distribution

2007-12-30T13:29:00.000-08:00

I found David Byrne's thoughts on the new paradigms in music distribution to be very interesting and insightful. He identifies 6 distinct approaches an artist can take to music distribution.

Finding artist communities by network topology

2007-12-22T12:42:00.000-08:00

In examining artist relationship networks, we can use community finding algorithms from the complex network literature to identify groups of artists. Various community grouping methods can be employed including betweeness link removal, fast and greedy modularity maximization, eigenvector community finding, or an algormerative approach - grouping artists using audio-based similarity.

Measuring Assortativity w/ respect to Genre

2007-11-26T07:51:00.000-08:00

Assortativity is a measure of the mixing patterns in a complex network. For a wide variety of networks, nodes tend to connect to other nodes that are similar in some way. A 1992 study regarding heterosexual couples in the San Francisco area is often cited as an example. The study shows that individuals typically partner with other individuals of the same race. In social networks this is called homophily - people tend to connect with people similar to themselves. These patterns can be quantified by calculating the assortativity coefficient, r, for a given network [1]. This will be a value between zero and one, where one corresponds to perfect homophily and zero corresponds to no assortative mixing.

In our investigation of artist-centric social networks (Myspace, Last.fm, etc) we are exploring the assortativity with respect to genre. In the myspace friendship network do artists tend to link with other artists of the same genre? If so, we would expect a high r value for our sampled network.

Focusing on the myspace data set - we see that each artist (node) belongs to between 0 and 3 distinct genre groups. The myspace infrastructure allows an artist to specify up to 3 genre labels with 119 distinct genres. Some artists specify no genre, some artists specify a genre that is clearly incorrect (perhaps expressing disdain for the concept of genre in general), but most artists seem to give at least some thought to their self-ascribed genre label. In theory, this is of benefit to the artist in the context of visibility to potential listeners.

Because an artist may belong to between 0 and 3 groups, our assortativity coefficient calculation is slightly different from the one described in the literature. Each edge will contribute between 0 and 9 connections between genre labels. So an artist labeled as 'electro' linked to an artist labeled 'electro/jazz/breaks' will contribute an r value of 1/3 instead of 1. This will lead to a slightly deflated assortativity coefficient. Using the standard calculation methods, a 6100 node sampling of the myspace network resulted in r=0.14 - which suggests little assortative mixing. However, a more clever approach might be warranted.

Perhaps the current approach can be modified to accommodate multiple group membership. The simplest approach would be to only count one genre match (or mismatch) per edge. This method results in r = 0.64 - which implies a high level of assortativity. Although it is not entirely a kludge, it's not very eloquent either. Perhaps a new measure of assortativity can be applied. For example a clustering algorithm could be applied to the sample set, and then the co-occurrences of genre labels within a cluster could be counted. Stay tuned for the exciting conclusion...

Songbird is rad

2007-11-20T09:16:00.000-08:00

I've been vaguely following the development of Songbird for a while now - with version 0.3, it seems to be really maturing now. Songbird is a next generation media player that emphasizes connectivity and discovery by having rich web integration. Furthermore, this is an open source project. In the near future we hope to develop plugins based on our research for Songbird. Check out this neat demo video:

Myspace Pirate Radio and Local Modularity

2007-11-15T07:07:00.000-08:00

To re-cap, our research is focused on artist-centric social networks and leveraging these networks for recommendation. We have a simple proof-of-concept web application up and running. It is a sort of Myspace Pirate Radio that works by performing a simple "snowball" search around a target artist. Basically, if an artist is in the 'top friends' of the target artist, her music is added to the radio station. This is admittedly rather simple and not terribly academically interesting but the results seem to quite nice and initial user feedback suggests the system is highly addictive. So I like to think, to at least some extent, this validates the artist-network approach to recommendation.

The next step is to build a "community" around a target artist. This is not an entirely trivial task - especially when it is impossible to have knowledge of the entire network a priori. However, there are some nice solutions to this problem in the literature. The local modularity community identification algorithm was first purposed by Clauset 2005 and brought to my attention by Hinne 2007 in the context of the WWW graph. Basically, the algorithm defines a set of nodes as a 'boundary' around the community. A ratio (local modularity) between the number of links from the boundary set to the rest of the community and to nodes outside the community is calculated. The boundary is extended in a greedy fashion to maximize this ratio (more connectivity within the community). This process is repeated until the ratio reaches a threshold or a pre-determined number of nodes is added to the community. In the next week or so, I will be experimenting with this approach and hopefully implementing it in the Myspace Pirate Radio.

Open Social and MySpace

2007-11-05T05:17:00.000-08:00

Well, I'm an asshole. At about the same exact time as I was writing about Myspace having no API:

...there is no API for accessing the myspace social graph (and I'm sure News Corp won't be changing this anytime soon)...

Myspace joined Google's OpenSocial endeavor. While there is some chatter claiming Google's API falls short of providing true social network portability, it seems undeniable that this is a step in the right direction. And although this is probably a ploy by Myspace to try to undermine the momentum of Facebook, I still must say, thanks Rupert Murdoch.

What implications will this have for our work? Stay tuned to find out.

Sampling massive online social networks

2007-11-02T05:59:00.000-07:00

Myspace boasts well over 130 million users and the network is still growing at an explosive rate. The task of sampling this graph is by no means trivial. For starters, there is no API for accessing the myspace social graph (and I'm sure News Corp won't be changing this anytime soon), so we must resort to web scrapping, which is slow and painful. But even if there were an API, many of the standard complex network analysis metrics would be impractical to calculate on such a scale. So we must take a statistically-significant sub-sample of the graph.

There are a few common graph sampling methods - node sampling, link sampling, snowball sampling... And actually, the most intuitive approach appears to be the best. We pick one seed node (a myspace artist) and perform a breadth-first search, getting all of the seed's friends, then all there friends and so on. This is called snowball sampling because the sample grows in exponentially larger layers like a snowball rolling down a hill. Some previous work indicates this is (currently) the preferred method for sampling online social network services [1, 2].

However, we have a slightly different problem because we are examining the myspace graph from a different perspective. Our approach is unique in that we apply two filtering methods:

We are (for now) only interested in 'artist' myspace users
We only include the 'top friends' on a profile page as new nodes

We only include artist accounts because we are interested in the artist graph. Artist-based social graphs have been analyzed before, but the source of this graph information has always been critical reviews such as AllMusicGuide. The myspace artist graph will only contain relationship explicitly specified by the artists themselves.

The second filtering method is more troublesome with respect to graph sampling. Myspace allows users to specify 'top friends'. These are friend accounts that are displayed on the users profile page. This means our artist graph becomes a directed graph (if I am your top friend, you are not necessarily my top friend). This is a blessing and a curse for reasons I won't get into here. Using only the top friends in our graph reduces the noise inherent in the network - many friends may just be fans or spammers.

However, restricting our sampling to the top friends has the effect of artificially limiting the out-degree of every node in our graph. Myspace allows users to specify anywhere between 4 and 40 top friends but most users seem to stick with the default of 16 top friends. Of course, even for an artist account, not all top friends will be artists. So the out-degree distribution will vary from node to node. Of course, the in-degree of each node will be a more interesting measure for our graph. An artist with a high in-degree can be assumed to be rather influential as he is specified as a top friend by many other artists. There are a few important questions that arise related to the restricted out-degree:

When the out-degree of each node is restricted, is snowball sampling still a valid approach?
Are the standard complex network metrics still as meaningful when the out-degree is restricted.

I don't know the answer. I'm asking...

MyPySpace SF project

2007-11-02T04:05:00.000-07:00

I've been working with Ben Fields on a projected related to social networks and music recommendation. The basic idea is to crawl the Myspace network focusing on *artists* and the relations they specify. The assumption is that relationships specified by the artists themselves are in some sense musically meaningful and could form the basis for good music recommendations. We hope to combine this approach with audio-based similarity techniques.

The project is available on sourceforge, but everything is very alpha.

Hello

2007-11-01T04:27:00.000-07:00

So I'm starting a blog. I am research student at Queen Mary's Centre for Digital Music and I thought this would be a good way to log my progress. My research relates to music information retrieval - basically finding the music you want amongst today's overwhelming volume of digital material. Most of my experience (up till now) has been in audio signal processing. But now I'm exploring a combination social networks and audio-based analysis as a means to create music recommendation systems...