The 2009 Music Industry Panel has just kicked off, and since Paul Lamere is on the panel, he won’t be able to blog or twitter about the events like he has so far… so I figured maybe I should do a brief write up instead.

The panel is comprised of:

1. Paul Lamere – The Echo Nest
2. Tom Butcher – Microsoft
3. Peter (?) – Gracenote
4. Oscar Celma – Barcelona Music and Audio Technologies
5. Norman Casagrande – last.fm
6. Keiichiro Hoashi – KDDI R&D Laboratories
7. Kunio Kashino – NTT Laboratories
8. Malcolm Slaney – Yahoo

Paul is acting as an ad hoc moderator, and started off asking:  What are the tough problems that people in the industry are facing right now?

Tom :  We’re actively trying to grow our music market, and come up with recommendation engines that operate at large scales.  Overcoming cold start problems, and evaluating user assessments in a general way is difficult.

Peter : We  have special challenges as a b2b company, since we have a lot of external constraints.  We have to cater to the needs of the customers, and accommodate what they want to pay for.

Oscar : In our case, we have a lot of practical problems:  metadata cleaning, etc.  I’m interested in semantic music descriptions, and music discovery in the long tail (less popular music).  We have clients & music from smaller markets like Turkey.

Norman : Metadata is one of our biggest problems as well. We try to leverage our users to correct some of the metadata, but you cannot rely on this.  We also face scalability data, in the face of real-time data needs for our users.  We have to be able to allow users to build a profile, and then provide recommendations immediately.  We also focus on understanding user “types” and general listening behaviors.  We also face the limitations that the record labels impose on us.  We are provided a pool of tracks (several thousand), and then we apply some user listening behavior metrics on this data.  Then, once we have a collection of candidate tracks (from the label), and the label restricts which songs can be played where.  We end up filtering a lot of the music out because of the limitations of how we can match against their catalog. Evaluation is also really important, but AB test can be tricky… it is easy to become too biased, and it’s important to design user tests correctly.

Peter : There is a lot of weird grey areas here, you can’t put cover art in embedded devices for instance.  The music publishers have the right to lyrics, which we have to explicitly clear with the publisher, and sometimes they are too slow, or not willing to clear rights in certain cases.

Keiichiro : We do our own music distribution service, and we are trying to compete with the “giants” such as iTunes, to increase sales and preserve the rights of the artists.  On the research side, we are trying to convince the music people to know that MIR is a “good thing”.  Our service has a lot of MIR functionality, but most of the services are based on conventional search.  My research is on more discovery search, but many business people are not convinced that you can present unknown songs to the users as recommendations.

Kunio : In Japanese language, there is many different ways of writing the titles of different songs.  So, we can’t always identify music based on simple labels.  We also have difficulties with dealing with all the different variations in recordings, and deciding if the recordings are actually the same.  We are actively researching this.

Malcolm : We don’t have a big problem scaling, but we have a problem deciding on how to provide the right content to the right user.  On average we get 2.2 words per query… so we have only 2 words to figure out what the user wants… Madonna the music star?  Madonna the painting?  etc.

Paul : If you look at mirex as where people are spending their time, and we see classification of genre, artist, etc… and then we look at the companies actually doing something with music (apple & itunes, guitar hero, google music, pandora), they don’t tend to use these techniques.  For the panel, the question is… why do you think that is?  Has the industry not caught up, or is research in the wrong area.

Malcolm : None of my people are begging for these technologies

Tom : We are not doing content analysis yet, it’s hard to do currently with our dataset.  When we launched the collaborative filtering algorithm, we had a lot of good data for that.  The product people don’t come to us and ask about content analysis technology, they want to know how to launch a service in Germany.  I disagree with the fact that people aren’t using MIR techniques in the industry (SongSmith).  It takes someone that knows about the research to able to see the solutions and opportunity for MIR techniques.

Keiichiro : There were two concerns for using MIR techniques:  The cost of performing content analysis, and the fact that we didn’t have the rights for songs.  We had to ask record labels and the artists if they agreed to allow us to analyze/use their content.

Oscar : I don’t think the mirex tasks need to be useful for the industry, maybe in 10 years it will prove useful or will be mature.

Peter : In all of these situations, industry comes down to a question of profitability.  The customers will only buy if they’re convinced of profitability.  Apple made decisions on recommendation technology, and the decision was cost based on the incremental sales it would provide.  Also, number one, everyone is working on new things.  The first is that different combined techniques are being considered over individual approaches.  The second is to ensure that the solutions work very well globally.  It can’t just work for certain types of music.  You can’t exclude yourself from certain international markets.

Tom : You mentioned Harmonix (Guitar Hero).  The technology in these games has changed how we consume music.  Not only are they fun, but the music in these games is monetized and consumed in new ways.  It’s more interactive, so you can charge more for it.  That’s a good way to make money.

Paul :  A nice thing about the industry is that we have more direct contact with users.  There’s not as much research around taking advantage of  user data.  What kinds of data do you have that you would want to share with researchers.

Malcolm :  I think AOL proved that giving out user data is a firable offense.  We give out CF data that has been thoroughly scrubbed.  We have a rating database available of 700 million rating data.

Norman : We have an API for a lot of data.  But it’s also true that I’m not a user of this API.  Quite often, we’re not offering something that’s not useful to researchers.  For instance, we only return the last 200 tracks for a user.  It used to be a practical limitation, but not necessarily anymore.  We are willing to negotiate directly with researchers if they require more information.

Tom : What Netflix did with their rating data really opened the door.  The contest really drove innovation in the field.  The other thing I’d like to mention is that Microsoft has a great internship program with the research labs.  In that setting you have access to all the data you want.

Paul : We’ll open up from questions from the audience

Question : What kinds of skills are you looking for?

Malcolm : I interview a lot of people.  From the research jobs, I’m looking for someone who can teach me something new.  For the engineering jobs, I want to see that you can build something.  show your passion.  Show something nobody has done before.

Oscar : We are looking for music lovers.  We use typical tools like C++ , but we mainly need people with open minds who are willing to try new things

Keiichiro : We need people that can think of things that corporate people don’t.  However, we only hire people who speak Japanese… so that might disinterest most here (it’s a small problem).

Norman : When we are hiring at last.fm, we are looking for people that have previous experience with small toy datasets.  Something that shows passion, etc.  Our team is very small, so the person we want is someone that can do a lot of coding, research, and development.  Also ping-pong skills are important.

Peter : The people that work in Markus’s team have to be Jacks of all trades.  We need people that can build complete systems.  People with motivation and drive.

Paul : When I was at Sun Microsystems, we were in a feud with Microsoft.  It never ended to surprise me how many resumes I got in Microsoft Word.  Paying attention to the company you are applying to makes a lot of difference   The people that we hire we tend to invite first.  What that translates to is that we are looking for people already doing stuff in their spare time, and making it public somehow.  That is a good indicator that they have a lot of passion.

Kunio : We are looking for people with a vision of technology and services.  The successful students are expected to have their own core technologies, and an ability to express their visions.

Question : I’d like to ask you about query by humming technologies, as well as illegal downloading issues.

Tom : I haven’t seen any commercial applications of this (query by humming).  It’s not in the works for our mobile services.

Keiichiro : We actually have a query by humming service.  If anyone can beat our service, we would be interested.

Paul : In the west, there doesn’t seem to be much interest in query by humming.  Is there more interest in Japan?

Keiichiro : I’m not sure how much it is actually being used, but we like to try a number of approaches.

Peter : We don’t have a working example, but it is something we’re interested in.  As the technology improves, we will start considering it more strongly.

Question : What are the differences between the use of symbolic data and content data?

Malcolm : I think symbolic data means text to us.

Paul : We use all of the data we can.

Norman : One of the biggest challenge we face is how to combine those different sources.  It’s hard to figure out what is best.  Which one do you trust?  What are the priorities.

Question : Are you scared of Google’s music service?

Malcolm : We already compete with them.

Oscar : We are looking forward to crawling their data.

Norman : We are interested in stealing their engineers.

Tom : I think Apple is really the big player in music.  Google music is based in iLike, which was started in Seattle.  If you’re feeling entrepenurial, that’s a great example of how to make some money.

Paul : I like what Google is doing.  If you search for a banner there, there’s music links in there, which lets you listen once or twice to a given track.  Some of the PR said that people went to Google first for “discovery”.  I don’t think that anyone will actually find anything new given how Google works.

Question : What are you all excited about these days?

Tom : From my experience, what we are looking for is ways to augment the way they consume music.  We don’t own the content, we don’t generate the content.  Any revenue we generate is around the services we build around consumption.

Keiichiro : I agree with Tom.  When the music service started at my company, nobody knew how people were really going to use music (as an alarm clock, or as an e-mail notification, etc.).  We are looking for new experiences and ideas (like our radio example in the demo session).

Paul : One thing I’m really excited about is services like Spotify’s new API.  As Spotify spreads its domains, other companies might do so as well.  I think labels might realize that music can flow more freely, and they might relax their copyright control.  That will lead to a lot of new ways of interacting and consuming music.

Question : I’m interested in music creation/composition.  What can you say about supporting tools for music creation?  What do you think is the next generation of MIR technologies useful for composition?

Tom : We don’t represent the manufacturers of music creation software.  We had this application called SongSmith.  You sing unaccompanied into a microphone, and the application generates the accompaniment.  That idea was generated by one of the developers.

Oscar : The Yamaha Vocaloid technology came out of MTG at UPF, which we spun off of.

Question : How good does a technology need to be to make it as a commercial product?  What is your criteria for quality?

Tom : That’s a good question.  However, it doesn’t matter.  The technology could be good or bad, but if it satisfies a consumer need, then it’s “good enough”.

Malcolm : At the high level I agree, but at the low level, we just look at traffic on our website.

Question : A lot of you to one extent or another have been following this conference for some time.  What do you think is a highlight from this conference that has helped at your company?

Oscar : I’m interested in merging content with context, which has been presented here.

Malcolm : We’ve come a long way since using simple MFCC features… but it’s hard to point out a single example.

Norman : The MIR has been building up datasets and tools over time, so bits and pieces come from many places.

Paul : All of the work on tags has been very useful.

Question : Do you work with Non-western music?  Do you have specific problems with that?

Norman : We are in that situation.  We have problems with label changes due to languages.  Those are issues we have to deal with.

Keiichiro : Most of the music we work with is Non-western.  We had some cases when we worked with ringtones.  Some are not music at all.  That messes up the music features we were expecting.  If metadata was present, we could solve that problem.

Norman : There is also an issue with bias.  The majority of our users are in a few countries.  When you compute similarities, you are biasing towards the cultures there.

Peter : We put a lot of effort in getting to a baseline performance of good coverage, and good support of local cultural sensibilities.  We have over 2000 genres, and we try to use local experts for editorializing in certain contexts so that everything is consistent to the region.  This also applies for mood data and classifier.  We try to do our best to make sure the training data has information from all those places in the world.  There is a long ways to go still.  We spend a lot of time in how to map classes and categories to labels and terminologies appropriate for different markets.

Malcolm : Yahoo’s goal is to organize all the world’s information.  Most of the revenue and users come from bigger countries, so it’s doubly hard to work elsewhere.

Paul : We talked about long term, grand challenges earlier.  What are the big problems being faced in the next 5 years?

Malcolm : Maybe give us 10 years.  What our product people are asking for is understanding content.  We want to understand user content.  We want to be able to identify music embedded in other content (music in videos).  How do we match music in other media?

Norman : One of the big problems is finding the “little gems” in the long tail.  Some of the popular music deserves to be less so.  From a research perspective, coming up with a better qualitative measurement would be really cool.  I don’t think there’s a generic value for quality, so it should be based on user interest.  I don’t think the quality of the content based recommendation is there yet.  I also like analysis of sub-song quality… identify smaller parts of songs that people really like.

Oscar : We are interested in understanding the user better.  We’d really like to improve that.

Peter : I also have user profiling and personalization at the top of my list.  There’s been some great advances, but it would be nice to have a solid personality profile for people, that would enable a better understanding of what people like, etc.  Navigation and visualization is also needed, and there are a lot of unsolved problems in this domain, especially in mobile devices, cars, etc… not only to look cool, but to actually be usable.

Tom : I’d like to add to that.  I’d like a way to bring things to customers, without them having to ask for it.  The more the user has to ask for, the more chance there is for frustration.  We think about recommendation experience as a single modality.  It may not fit in all situations, but closing the gap on user expectations and needs is important.

Malcolm : In the real world, you don’t get much information from the user.  We need more technologies that need less data to drive recommendations.

Question : How do you feel about interoperability of the different services?

Norman : I think it’s awesome.  As long as our users are using it, and are aware of last.fm and provide new data, then it’s great.

Paul : There are some need for standards.  Some standards are proposed, but businesses have been slow to adopt.  I think there may be some interesting things come down the line there.

Tom : I’m very disappointed with my company’s stance on this issue.  We are a closed model like Apple.  I’d like our group to take a leadership role in this area.

Question : Do you know a technology that was good enough on its own without any marketing?

Paul : Shazam was an example like that

Malcolm : Most of Google’s stuff is like that

Paul and I recently finished up our tutorial on Using Visualizations for Music Discovery. The fast paced, 3 hour long presentation covered nearly every instance of music corpus visualization that we had found in the literature, or in casual “fan” based renditions.  This included visualizations of artist connections generated from millions of playlists by high powered supercomputers, all the way down to personal hand-drawn sketches on the back of a notebook.

I’ve embedded the slides from the tutorial above.  However, the format of the online slides do not include any of the video segments that we included, nor does it include the demo visualizations that we generated live on our laptops.  While our demo code (as well as the rest of the tutorial materials) is available online, I also prepared some back up video versions of my demos, just in case something went horribly wrong during the presentation.  I thought it would be a good idea to post these videos online for posterity.

The demo focused on exploring the “acoustic space” of music involving the lute and the folk guitar. Both instruments can be similar acoustically, but generally take part in culturally distinct genres. In the realm of audio-based genre classification, renaissance music that features the lute can get confused for folk music songs that feature the guitar. I wanted to focus on representative lute and folk music for the demo, and see if visualization helps to show the confusion that a classification might run into, and perhaps lead to ways of resolving it.

Rather than show all of the acoustic features at once, I chose to look at them one at a time. The first feature that I wanted to focus on was pitch information, or a general representation of what pitches, keys, etc. were present in the music.  For this and all demo videos, the blue nodes corresponds to lute music, while the red nodes correspond to folk music.  However, there are some “mislabeled” songs, as we will discover later:

The visualization shows how adjacent songs do in fact have similar pitched melodies, but can have very different styles or genre. This makes sense, since most popular genres use the same western arrangement of major and minor chords, etc. So, pitch information is essentially useless to separate the lute and folk music in our data.

The next acoustic data I looked at was timbre, which is a representation of the song’s tone or color.  This type of data is typically the “best” information to use for genre classification, and I was curious to see how the timbre separated these genres:

Interestingly enough, the timbre space “almost” separates the music clearly.  The softer tones of certain lute music are distinct from the rest of the folk music, which has harsher or brassier coloring.  However, the lute has a variety of timbres that it is able to create, and once the lutist uses this harsher plucking action, it sounds more like conventional folk guitar music.  At this point, it may be possible to make a “pretty good” separation using a hyperplane, but as indicated by the visualization, there would most likely be some significant error using timbre alone.

Luckily, our database of music included non-acoustic features, such as tags. The tags were simple terms or phrases that were applied to songs by human listeners.  Using appropriate text retrieval techniques allows us to see how the music was separated according to how it was tagged:

Here we see the two styles of music are clearly separated.  However, there is a problem… I had originally used the tag data to identify/label lute and folk music in the first place.  I quickly found out that some of this music was mislabeled… i.e. classical lute music was labeled as folk, etc.  So, even though the separation was clean, we now have exposed a new form of error.

The final approach that I tried was to mix the term and the timbre data in a “hybrid” representation:

This had the best result of any of the approaches that I had tried. It is no longer possible to separate the music by color (some of the red folk music is included among the blue lute cluster), I show that these red dots were actually mislabeled. In this fashion, the combined term + timbre features were able to “correct” each other, and a more valid representation of the genres were presented.

The code for this is all available on the website.  The code is written in R, and requires the packages rgl, RGtk2, as well as a GUI interface package I wrote called sculpt3d.

Also, if you’ve made it this far, you should check out  (the co-presenter) Paul’s blog.  He’ll be live-blogging the rest of the conference as it continues through the week.

Paul and I have just launched a new blog:
http://visualizingmusic.com/

from the about page:

As the world of online music grows, tools for helping people find new and interesting music in these extremely large collections become increasingly important. In this blog we survey the state-of-the-art in visualization for music discovery in commercial and research systems.

Paul and I wanted to create an ongoing, up-to-date resource of all the different instances of musical “discovery” visualizations that we have come across.  We’re defining this as broadly as we can, while avoiding musical signal visualizations, such as all the iTunes/winamp plugins.  Essentially, music discovery visualizations should help you understand relationships between songs, and hopefully help someone identify new and relevant content that they had not heard or considered before.

There is an incredible variety of techniques used to express musical relationships with images, rather than words.  There’s many people who believe the best way to communicate about music is not through words.  As the famous saying goes: “Writing about music is like dancing about architecture”.

Be sure to check it out, and if you have any suggestions for new visualizations, or comments in general, don’t hesitate to contact either Paul or myself.

Paul Lamere and I recently got our tutorial accepted to ISMIR 2009.  It is titled Using Music Visualizations for Discovery and will be a survey of all the different ways that researchers have approached visualizing music collections, and how we can use these methods to find new and interesting music.

At first glance, this perhaps appears to be a bit of a ‘fringe’ topic.  Visualizations are often seen as more of an art form than a straightforward research tool.  However, I believe that music is a special form of information.  On one hand it behaves according to a set of physical rules, following certain spectral behaviors as an acoustic signal.  On the other hand, it behaves as a cultural icon, following another set of rules in a network of contextual associations and systems of meaning.  On yet another hand, it is understood in a musicological system of compositional symbols and structures.  To study and understand only one aspect of music is to have an incomplete grasp of the role it plays in the other systems.  You could spend an incredible amount of time understanding how musical information is related in each context.  Perhaps this is why music information retrieval projects spanning different music metadata types are somewhat rare.

Paul and I are believers that visualizations can be effective “big picture” overviews of music corpus information in many different contexts, and can be a great first step towards understanding how each form of music metadata reflects an important way of understanding music.  We hope to use our combined expertise to produce a useful overview of the current state of the art in music corpus visualization, as well as provide some useful open source tools for researchers to investigate their own musical information visually.

We’ve been gathering a lot of existing relevant projects already, and have tagged them all on del.icio.us.  There’s a lot of great projects on there, and it’s worth a look.  If you think we’ve missed something, tag the respective website with “MusicVizIsmir2009″ and we’ll check it out and add it to our collection.

Paul has some information available on his blog…. but just don’t believe everything he says about me, I’m the understudy in this group

I already talked about why I liked the approach, and now I’d like to distinguish it a bit semantically.  The ISMIR paper builds on a 2005 Nature paper, which talks about how the low level auditory system is “tuned” to pick up speech related sounds with a very high efficiency using gammatone like wavelet filters.  However, labeling the technique  “sparse” in general is a bit misleading.  This is because the relative “sparsity” of the signal representation depends on the gammatone basis functions, the encoding routine, as well as the signal characteristics themselves. So, respectively, these techniques use biologically observed gammatone-like wavelets, using a matching pursuit algorithm to minimally (sparsely) encode signals, on “natural” occurring sounds.

To characterize their technique more plainly, it’s appropriate to call it a form of neural audio analysis.  While the standard matching pursuit encoding method doesn’t completely emulate the mechanism for a neural response  (partly because it is usually performed as an offline process), it does come far closer to mimicking what our brain is doing over other techniques like short time Fourier transforms.

Furthermore, neural audio analysis differs from other more general wavelet analyses (using general Gabor atoms, etc.) in that it is using a predefined, limited, neurologically-based dictionary of wavelet functions.  For better or for worse, these chirps and blips are the building blocks of how we perceive sound.  They may perform very efficiently for speech, but they are probably lacking (non-sparse) for many other types of signals.

How did we get these wavelet filters?  The hard way!  De Boer et. al literally had to wire a cat brain into a signal processing loop in order to extract the individual neural responses to audio stimulus.  Cats were chosen because they have auditory faculties that are very similar to humans.  Luckily, the Boston Ear Lab provides a database of these filters so that other researchers don’t have to repeat this same experiment on the poor cats any more.

The point is, using these neurological wavelet filters is important, even if they’re non-efficient ways of representing certain signals.  They come closest to representing how our brains are responding to an audio stimuli, and therefore provide better building blocks for forming higher level representations of sound.  I think that even though the underlying techniques are not new, the potential for “neural audio analysis” is still fairly untapped.

As a side note, I’m glad that Manzagol et. al wrote that paper, it’s finally putting my “Mind” category to use.

All plots taken from either De Boer & De Jongh 1978 or Smith & Lewicki 2004,2005.

1. As symbolic notation (A composer’s score, or digital MIDI file)
2. As waveform data (MP3′s, etc.)
3. As a social/cultural phenomenon (Holiday music, pop music, playlists/streams)

The first form of analysis is perhaps the easiest; If we have the symbolic representations of notes, events, and other temporal data, then we can do an enormous amount of music related retrieval.  However, not all forms of music are available in this form… others are impractical or impossible to put into this form.

The second form of analysis forms the bulk of the research in MIR.  However, most of the techniques use a process involving Short Time Fourier Transforms (STFT). These techniques are some of the most important and useful techniques for general computation that digital signal processsing has yet to produce.

People… very bright people… have dedicated a good part of their lives towards optimizing and improving these techniques.  The method and manner of the transforms are highly tuned, and well understood.  However, the plot above shows how STFTs will never yield time relative coding.  STFTs work by cutting a signal into a set of segments, and then transforming them into a set of energy over a range of frequencies.

The problem is that all the frequency content must be extracted at once from a segment.  Short segments will limit the ability to extract low frequency informaiton, while long segments “smear” the activity in high frequency bands during the segment duration.

On top of the timing issue, the STFT output is not really a good representation of what we “hear”.  We are much more sensitive to changes in a limited frequency range.  This range has been modeled with the Mel Scale, and is commonly used in “Cepstral Coefficients” (MFCCs).  However, there is still a large divide between these representations and the nuances of musical perception.

The (main) point of MIR is to characterize music the way that we hear it. Unfortunately, our brains don’t use STFT, so it’s a struggle to start from STFT and work towards a model of human hearing and cognition.

However, promising work has been done in the realm of speech and hearing.  By analyzing the coherent cochlear activity of the inner ear, we’ve found that certain neurons “encode” brief patterns of sound.  In essence, our minds have a consistent “dictionary” of these little patterns of sound (wavelets), and they use them to represent every sound that passes into the ears.

Recent work by Smith and Lewicki show fascinating evidence that human speech is optimally/efficiently encoded by these neurons.  This couldn’t have happened by accident, and it means that we (as a culture at large) have fashioned our various spoken languages to take advantage of this neural activity.

The question is then, have we as a culture done something similar with music?  I had been interested in digging into this topic for a long time, so I’m glad that Pierre-Antoine Manzagol, Thierry Bertin-Mahieux and Douglas Eck have already done some initial analysis, and received a best paper award for their efforts.

So, as a newcomer to the conference, it’s interesting to see “three waves” converging in Philadelphia:  The past, (which includes conventional acoustic analysis and has now become state of the art in industry), the present (which includes the wide array of social analyses that are finding their feet), and the future (the sparse/neural encoding methods that appear to have tremendous potential).

In fact there’s so many that it’s hard to sort them out since they aren’t easy to parse in his index.  Here’s an overview:

1. Tutorials
   1. Chuck Tutorial
   2. Tutorial Overview
2. Day One
   1. Day One Overview
   2. Geometry of Musical Chords
   3. Plenary Session 1
   4. Plenary Session 2
   5. Day One Posters
3. Day Two
   1. Music Recommendation Industry Panel
   2. Plenary Session 3
   3. Plenary Session 4 (Couldn’t find this one)
   4. Day Two Posters
4. Day Three
   1. Keynote
   2. Plenary Session 5
   3. Mirex Panel
   4. Mining the MySpace Social Graph
   5. Plenary Session 6
   6. Day Three Posters (Mirex)
5. Day Four
   1. Plenary Session 7,8
   2. Plenary Session 9 (he had to prepare his Explaura demo, and didn’t cover them)

Thanks again Paul!

I’m at the 2008 ISMIR conference on music information retrieval, presenting a poster paper with my co-author, Claudio Baccigalupo.  I talked a bit about the paper in a previous post.

I was surprised at how prominently “social” sources for musical understanding were featured in the projects and demos, so our work on using playlists/streams to understand relationships between genres was right at home.  There were also some fantastic demos, including the “Aura Music Explaura” which allows for “steerable recommendations” by using tags.

The way that Francois and Paul (the project leads) implemented the tag interface was particularly brilliant.  By leveraging the common visual motif of a tag cloud, the interface communicates the frequencies of the tag labels applied to the music.  However, the tags themselves are interactive.  By dragging the tag term boundaries, one can change the “weight” of the tag.  This also increases or decreases its relative visual size.  Doing so will alter the importance of the tag in selecting similar artists.  So, hypothetically someone could alter the amount of “emo” they want in their music results.  Paul gave me a quick demo, but I really hope this ends up publicly accessible soon, it’s killer.

Claudio Baccigalupo and I have a paper at ISMIR entitled Uncovering affinity of artists to multiple genres from social behaviour data.  The paper details a project we worked on for the past year or so involving popular music listening activity from a pool of MusicStrands (MyStrands) users.

We provide not only the paper, but also the dataset and the code used in our analysis.  All of this is available at the website we have set up for the project.

The main contribution of the project is an analysis and illustration of genres as “fuzzy sets” rather than boolean labels.  Through a co-occurence analysis of hundreds of thousands of user playlists, a frequency based “affinity” metric is formed between artists and genres.  This affinity metric is a more detailed expression of the style of a given artist’s music.  The idea and awareness of predominant genres are a trivial part of any person’s understanding of the vast corpus of popular music.  However, genres typically are used as boolean categorical labels.  I.e. an artist is understood to be associated with only one given genre.

By expressing a connection to multiple genres through our affinity metric, a more detailed picture of the artist emerges.  We give a lot more examples in the website, so be sure to check it out.