A classic problem in traditional collaborative filtering recommendation is the 'cold start' problem. It is hard to generate recommendations for new items because there isn't enough taste data about the new items to make reliable correlations with other items. That's where content analysis comes in. The cold start problem can be alleviated by basing recommendations on similarity of content as well as the wisdom of the crowds. New items can be analyzed and enrolled into a recommender, making these items available and recommendable.

However, using content-based techniques doesn't guarantee the elimination of the cold-start problem. Pandora, everyone's favorite Internet Radio, uses content-analysis to drive their customized radio. However, since Pandora performs all of their analysis by hand, there may be some lag before your favorite artist makes it into the Pandora catalog.

There's another content-based recommender - BookLamp.org. The BookLamp F.A.Q describe BookLamp as a

"book recommendation system that uses the full text of a book to match it to other books based on scene-by-scene measurements of elements such as pacing, density, action, dialog, description, perspective, and genre, among others. In other words, BookLamp.org is a Pandora.com for books, based on an author's writing style. If you match against multiple books, the self-learning system adjusts your formulas to make the match specific to your tastes. As the system moves out of beta, it will also incorporate human feedback into the recommendation systems, blending the strengths of social networks with the strengths of computer analysis. Ultimately, we want users to be able to create and share their own formulas, creating a community of book lovers that have tools to discover and share books in a way never before possible. Because the system matches books through objective data from the text itself instead of relying solely on social networks to generate recommendations, the recommendations are impervious to outside influences such as advertising or author marketing. It also allows you to match to a far greater detail than alternative systems. With BookLamp, you can request a book similar to Stephen King's The Stand, but half the length, first person, literary mainstream fiction, with slightly more dialog, less description, and a rising action level across the first 10 scenes. If that's what you're looking for".

It is a neat idea, and sounds very similar to the types of things we are doing with Search Inside the Music and Project Aura. Using content analysis gives you better ways to help people discover new items. However, BookLamp has its own cold start problem. Again, from the BookLamp FAQ:

Does BookLamp Work? Can I use it right now to find a book to read?

The simple answer to this question is that while BookLamp works, it doesn't have enough books in the database to work well. While the technology behind the system is capable of finding you books to read right now, BookLamp will remain a technology demonstration until we have a large enough database of books to give the system enough data to make realistic recommendations. Without more books, not only will most users have a hard time finding a book to match against, but the system will have a limited number of books that are capable of being matches. In other words, if we don't have a book in the database that matches, we won't be able to recommend a book for you. Additionally, with so few books in the database, we're not able to match against all the metrics that we would like. In order to be the most effective, BookLamp needs to match against 7 to 8 metrics; with less than 300 books in the database, we're having to make recommendations after matching against only 3 or 4 metrics. To get any matches at all, we've had to turn down the sensitivity of the measures (see the next question) a bit already.We estimate that it will take a database of at least 10,000 books to make BookLamp a usable system. The more, the better.

So BookLamp has a bit of a problem, with only 300 books in its database, it is not going to be the best book recommender. And unlike music, it is not so easy to enroll a new book - scanners and page turners are involved. So BookLamp is trying to figure out its next step. If I were them, I'd build a recommender for the Gutenberg project with its over 25,000 titles. Of course there are no NY Times best sellers in the bunch, but it would be a great way to fine tune the content-analysis while providing a service to a worthy project.


Sure, Pandora is content-analysis.. but not really. Pandora is more of a tag or label approach. The tags come from musical properties of the content of the song, yes that is true. But they are very coarse, blunt hammers. Reducing the content analysis of a song to a 1 through 5 label of the amount of vibrato robs the recommender of the ability to actually map out a complex vibrato acoustic space, and compare the vibrato of two different songs directly, within that space. By making that final label decision/classification, it robs the system of a lot of the subtlety that could really make the difference in content-based systems. Same thing with lots of the other Pandora labels.. when things like rhythm or voice gender or etc. get tagged. A decision gets made too early, which limits the search space.

A better analogy might be to the color space. Suppose I am trying to compare / assess the similarity of two images, via a color histogram. The best way to do this is to start with a color histogram of every RGB pixel value. In this manner, you can learn functions over the entire color space (all 16 million values).

The Pandora approach, on the other hand, starts by reducing these 16 million values to a very small set, say {red, orange, yellow, green, blue, indigo, violet}. They use human experts to label the content of an image with these seven color labels, and then they do their similarity calculations in this impoverished space.

What I would really like to see, what no one has still yet managed to create, is a content-based system that uses hundreds of features (rhythm, vibrato, vocal timbre, harmony, etc.), each feature in its *full* musicological function space, to do recommendation.

I wonder how many of these problems we could overcome, if we were able to take things to that level.

Posted by jeremy on April 25, 2008 at 01:43 PM EDT #

I don't like how you present content-based as a solution to collaborative cold-start problem. Sounds as if it's only a fallback solution. I think it deserves to be considered as a whole different approach that doesn't need to exist solely to alleviate the problems of collaborative engines.

I know, it's not what you really think of content-based system; I just have a gripe with the introduction.


About the dataset coldstart problem, it's something we've faced in our research, though digital music is easier to come by than digital books. Our official solution is to deal with the content owners themselves, and without having to get a direct access to the digital files (copyrights maze), you can get the "features" you want with a simple standalone analyzer program. Makes testing the system difficult, but

Gutenberg project is obviously a very interesting dataset to bootstrap BookLamp.org, and once they show their recommendation's quality on this dataset, perhaps they'll convince some editing companies to run the "feature extractor" on their books, so NYTimes best sellers finally make their way into the BookLamp database.



WhileI tend to share your interest for full-space feature scales, I think the Pandora approach has some redeeming qualities. Human involvement, I think, is adding a layer of editoriality, authority to the real features, something an automated system could hardly ever get. Also, while you can get hundreds of automatic features for a song, a human listener can select which are the important features and which are not. This can be both good and bad, but I'm not sure Pandora's approach is so much on the bad side...

Posted by Marc-O on April 28, 2008 at 12:48 PM EDT #

Marc-O: I apologize if I implied that the Pandora approach is "so much" on the bad side. There is a lot of merit to what Pandora does, most of all for the fact that they actually have a working system out there, in mass use. That by itself is 90% of the solution, and they should be commended for it, most definitely.

My comment was targeted mostly at the notion that the feature space that Pandora uses is impoverished. I mean, they represent vibrato on a 5-point scale (or at least they did, back in 2005, when I attended a talk by Tom Conrad).

I just don't think you can accurately represent music on such an impoverished level. And that fact is independent of the issue of "human involvement and editoriality" versus automated machine methods.

First of all, it ignores the fact that even automated methods can have "authority to the real features", as you say. Even automated methods are designed by humans, and the intelligence that goes into algorithms are imbued with human authority and editoriality. Juan Bello and I have a good example of this, our ISMIR 2005 paper (Robust Mid-Level Representation for Harmonic Content in Music Signals) in which we use automated, machine learning methods (HMMs). But rather than just use the machine learning methods as pure number crunchers, as had been done for years previously, we imbue the machine learning methods with musicological, human-centered intelligence. Ie., we use human-based musical knowledge to constrain and guide the algorithms.

But again, that is not the main issue. The main issue is that, in order to allow humans to label hundreds of thousands of songs in a reasonable amount of time, Pandora uses an impoverished feature space. Representing vibrato on a 5-point scale is like representing color with the 7 ROYGBIV colors of the rainbow. Sometimes there are colors, such as blue-green (aka cyan aka teal) that don't really fit anywhere into this label set. Cyan is not really blue, and it is not really green. But a human, doing a Pandora-style labeling of colors in an image, is going to have to pick one or the other. And that is just not as good as having a richer representational space for a feature.

Posted by jeremy on April 28, 2008 at 02:17 PM EDT #

Marc-O: FWIW, even automated methods rely on humans to judge which features are important and which are not. Humans build into the intelligence of the algorithms a certain preference for some features and not for others.

Conversely, even human methods rely on machine learning to judge which features are important and which are not. At that same 2005 presentation I attended, where Tom Conrad explained the inner workings of Pandora, he talked about how Pandora uses your thumbs up/down feedback to automatically adjust and re-weight certain feature values, so as to steer Pandora's recommender engine toward song with the types of features that you like, and away from ones that you don't.

Posted by jeremy on April 28, 2008 at 02:28 PM EDT #

The involvement of machine learning you speak of is beside the point of tagging, which it what I was referring to. Yes, Pandora (I guess) does some machine learning to model users against the tags. I was referring to machine learning for the tagging process, which they don't seem to do (yet).

As for the human involvement in machine learning, yes, it's often present. The "human authority" I spoke of still differentiate the tagging process though. I think the fact that a tag was applied by a human rather than by a machine is important information for the user. That is of course, debatable, but for the general public, I think this information (human tagging) is worth something. A human decision (and error) is easier to explain, and to forgive, than the result of a neural network...

Or perhaps I'm wrong, and people judge primarilly from the quality of the results, whatever created them...

I agree the 5-point scaling system is weak. It's most possibly a trade-off to accelerate things in the labelling. However, is the absolute accuracy of the tagging really that relevant to the use Pandora makes of it? So far I don't see Pandora (which I actually don't know very much I admit) trying to really classify things precisely as much as making good content-based recommendations. Perhaps even if the precision of the tagging ain't great, it's possible to make good recommendations. Perhaps a gain in precision isn't that important to the recommendations anymore.

Also, how precise can the scale be so that human taggers still do the job consistently (among themselves)? A 5-point scale is simpler to use, and probably hides just fine the fact that one person's 7.1 is another's 8.4.

It's still beside the point. Pandora's pretty much stuck with the system they've created in the first place; they're mostly prisoners of the state of their database so far, because re-tagging would cost too much. I think in the end, that might be the biggest flaw in their method: the inherent inertia of the dataset.

Posted by Marc-O on April 28, 2008 at 03:55 PM EDT #

>I think the fact that a tag was applied by a human rather than by >a machine is important information for the user. That is of >course, debatable, but for the general public, I think this >information (human tagging) is worth something. A human decision >(and error) is easier to explain, and to forgive, than the result >of a neural network...

Yes, I certainly agree with you that, in terms of explanatory analysis, having a human tag is better than having a machine-generated tag.

The main point of my original comment, however, was to suggest moving beyond tags for content-based analysis. My point was to suggest that (human-driven machine-learning) algorithms should (1) map raw acoustic data into a large musicological representational space, and then (2) do the matching directly, in that representational space.

Why use tags as an intermediary? Why use tags at all, either human OR machine-driven?

For example, if I am trying to match the rhythmic constructs of two songs, one approach would be to label the rhythmic genre of each song (e.g. "cha cha" or "bossa nova"), and then match two songs if their labels match. I do not distinguish between a human-labeled rhythm, or a machine-labeled rhythm. At this level of abstraction, it doesn't matter how the label was arrived at. What matters is that you've collapsed your representation of the entire song to a single label.

I contrast that approach with one that NEVER tags or labels the music. It only looks at the music itself.. the beats, the timbres of the beats (was this a high hat? Was that a snare? was that a bass kick?), and it looks at the metric levels of the beats. And then to determine the rhythmic similarity of two songs, it looks at whether various kicks and hits and syncopations occur in similar periodicities.

In other words, it never bothers assigning a semantic label to the song at all. It just looks at the actual content of two songs, and does a direct, musically-meaningful match.

That's the sort of recommendation engine that I would like to see more of.

>Or perhaps I'm wrong, and people judge primarilly from the >quality of the results, whatever created them...

I agree with you here. At the end of the day, it shouldn't matter to the end user whether or not human labels, machine labels, or no labels are used. What matters is the final quality, yes.

I'm just saying that, overall, I think it is a more productive research path, I think that we are going to get more from content-based methods, by relying on direct similarity measures rather than going through lots of labeling.

>I agree the 5-point scaling system is weak. It's most possibly a >trade-off to accelerate things in the labelling.

Yes, that is my guess as to why it is used, too.

>However, is the absolute accuracy of the tagging really that >relevant to the use Pandora makes of it?

I don't think it is. Like you say, what matters is the final results.

But this is the whole reason for my original comment. Why bother tagging at all, if the accuracy of the tags don't matter, or only matter a little bit? Why not just match up all the beats and timbres, directly? Why go through this tagging step, with its inherent inaccuracies, at all?

Posted by jeremy on April 28, 2008 at 05:10 PM EDT #

To draw this back to my other example, in the colorspace domain, check out this entry on the blue-green label for color:


This is, and has always been, the main difficulty of using labels or tags. In some cultures, blue == green ("grue"). In others, blue, green and cyan are all different colors.

So instead of relying on labels at all, why not just describe colors in terms of their absolute color-space values, regardless of what we as humans label the colors?

"Blue" becomes RGB(0, 0, 255)
"Green" becomes RGB(0, 128, 0),
"Cyan" becomes RGB(0, 255, 255), and
"Teal" becomes RGB(0, 128, 128)

So now, worrying about whether the label "blue" matches with the label "green", or with "teal" or whatever, we just look at the distance in color space between RGB values. Blue and green are a lot closer to each other than to, say, deep pink: RGB(255, 20, 147).

So again, why use labels at all? Why not just measure similarity, directly?

Posted by jeremy on April 28, 2008 at 05:48 PM EDT #

Sorry, I mean "Green" is RGB(0, 255, 0).

Silly me.

Posted by jeremy on April 28, 2008 at 08:46 PM EDT #

alright, let me state first that, on a content-based level, I want the same thing as you. I'm all for matching real measures instead of tag-approximations. Actually, that's what we try to do/want to do (at my job).

I think in the end, it comes down to Pandora's decision to choose humans over computers to put information on the songs. They're locked in the pattern they chose. While I think there is great power attainable via the machines, it's a choice that asks for more work, that is more uncertain as to what is possible and what isn't (a priori AND a posteriori - sometimes, what the machine says is difficult to understand, and to explain).

Still, and I don't plan here to draw this argument much longer because I think we agree on all that matters, I think that sometimes, the human tags, the discretization of a real-value scale, can gather some precious value, mostly invisible to the machine.

Because it's, in the end, a human that applies the tag, that person is passing the information through it's own ears/brain, and what exits might be sometimes more than what a machine can hear/compute. The real-value scale of musical features might not always matter in a simple linear or logarithmic scale (hell, it could even be non-monotonic), and the discretization used via tags might map the real-value scale to a discrete tag-scale that is more likely to make sense to whatever algorithm we use for the final recommendations.

If I take your color example, yes, in the end I'd prefer to use RGB myself. But passing through color names (tags) instead of the real color might have some gain. We could simply assume that the tags are simple approximation, discretization of the 3D color space, but what if they are not. What if some color tags represent large zones while others are very small? In that case maybe a large difference in numbers on the RGB scale might mean almost nothing to the end users, the human, and that assumed meaningless difference is correctly squiched by the human tagging process!

Take infrared for instance. A machine could see it just fine, and care about it a lot (if you let it), while it's totally invisible to the humans. Same for sounds. There are things physically present that people do not hear. The cultural differences for tagging might actually mean that the physical-to-label tagging is different from one culture to another, and that different cultures do not care the same about different features.

Ok, here I said "might" a lot, and I lack the knowledge to really make valid hypotheses out of that. I also only assumed the upsides of human tagging, without mentionning all the problems it can cause, especially the one that caused this post (hence this discussion) in the first place, the "cold-start" time it takes for one element to be added.

Posted by Marc-O on April 29, 2008 at 12:58 PM EDT #

I think I get where you're coming from. Just like you agree that working in raw/pure ("RGB") content space is more interesting, I also agree that there is a strong need for "explanatory" information retrieval and recommendation. I want search engines and recommenders to not only give me information, but tell me why they gave me that information. And perhaps having human tags on attributes is a good way to do it.

Maybe this is a discussion best suited for in-person, at a cafe or in between sessions at a conference. Will you be at ISMIR this year? Can we continue the discussion there?

Posted by jeremy on April 29, 2008 at 09:16 PM EDT #

ISMIR yes. I'm really interested in going there, so yes, we could probably continue the discussion there.

Posted by Marc-O on April 30, 2008 at 09:55 AM EDT #

Sounds good. Who are ya/where do you work? (So I know who to approach :-)

Posted by jeremy on April 30, 2008 at 01:42 PM EDT #

Post a Comment:
Comments are closed for this entry.

This blog copyright 2010 by plamere