OK, so I want you to watch this cool video from Pandora. But before you do, a bit about why.
I talk a lot about how how people organize things is unique. How i would put ten books on the shelf, or songs into my play lists, or CD’s in my car, are not necessarily the same way you would. That’s because the way things are related is not prescribed by the things being organized, but by my way of thinking, which is influenced by how I know things. It’s why organizational schema like DDC and LC just aren’t universal. You have to learn them, and even then, it can be VERY confusing.
Furthermore it is the relationship between items that is as if not more important than the way we describe things themselves. So the fact that Mt. Everest is 29,029 feet tall only takes on real meaning when related to the fact that the plane I am using to fly over it only goes up to 29,000 feet. This is why keywords and tags are so problematic…they lack connective tissue.
These ideas are behind things like Scapes and Reference Extract and play a big role in the whole New Librarianship thing. It is also why I say we should scrap catalogs and start fresh not with inventory systems, but with knowledge discovery and building systems.
Anyway, the video. I’m always looking for good examples of this sort of organziationa nd discovery by relation. I think Pandora has done a brilliant job:
By the way, this is also an excellent example of why cataloging is not the ony way to organize information. These are the kind of tools and connections that librarians should be making…or at the very least aware of. Imagine how your music collection might look using these kind of tools. Take all your music, plug it into Pandora and see what kind of recommendations it pops out for your next acquisition.
Hi Dave! Music is near and dear to my heart so I can’t help but weigh in. Wouldn’t Pandora’s recommendation system be *linear*, not *non-linear*? (And I don’t mean to be pedantic here; I think the distinction is important to understand.) Through this example we see how it traces Kanye through Blondie (and beyond), one artist at a time. One can thus trace a line from Kanye to Blondie. A non-linear recommendation would be discontinuous, like recommending some doom metal band for a Kanye fan, no? Which begs the question: why recommend a doom metal band for a Kanye fan? What’s the reason behind it? If there’s a reason (similar rhythmic tempo, another user likes both, e.g.), then it’s linear… if there’s not a reason, then it’s non-linear. Is my line of thinking unsound?
If no, then one non-linear recommendation system, off the top of my head, would involve a random ‘surprise me!’ button. But my guess is that few people would use this, perhaps beyond an initial curiosity stage. I’m getting a bit off topic now, but I’ve encountered very few people who really go beyond a comfort zone when it comes to music. (And the older they get the more entrenched they become, if I may generalize further.)
Any thoughts on this issue?
Good point, and thank you for keeping me honest on my terminology. I should have used the word hierarchical or deductive. It is discontinuous in the sense that it is not the same facet of the song (title, beat, lyrics) that are being used to link from one item to the next.
Perhaps a better way of thinking about it is to think of each artist as a node (with an array of properties) in an n-dimensional web. Kanye relates to Can along dimension A (he samples them), and Can relates to Kraftwerk on dimension B (same scene in time and space), and Kraftwerk relates to Eno on dimension C (Eno was inspired by Kraftwerk), and so on. So you can trace Kanye to Eno through certain properties of a set of nodes (Can…Eno) – let’s say 4 degrees of separation.
But what’s interesting, to me, is *who* makes these connections between nodes. Because the person who made the video, e.g., doesn’t seem to know that much about the artists he’s talking about. (Or s/he’s dumbing it down for marketing purposes by including only recognizable names?) If someone said to me, “I like Kraftwerk… what do you suggest I listen to?” I might recommend one or two of the 25 artists that sites like Pandora recommend. To me, as a hard-core fan of music, Pandora rarely makes interesting connections. It’s a bit like Music Appreciation for Dummies in this way.
Perhaps the point is that I’m in the minority, the long tail. Or perhaps this is an empirical question. How can we actually measure effectiveness with sites like Pandora? Bob is interviewed about his satisfaction with Pandora, and he says that yes, he liked 4 of the 5 artists that were suggested to him. But what if he had been suggested these other 5 artists?
I’m starting to move away from a cogent point so I’ll quit now. I’ve probably played my hand though: I can count on one hand the number of times these massive-scale recommendation systems have pointed me toward something interesting. New finds tend to be unexpected ‘happy accidents’. There might be a point to be made here about physical stacks in the library but this post has gone WAY too long now…
Actually as you put it is a great example of what I am talking about. Current catalog systems pre-define what the nodes are and the possible set of inter-relations. For example you chose artists as the central node type, most library systems will choose album (very rarely song), and you even chose the facts of linkages. The point is not that you agree with Pandora’s way of recommending, it is that they made connections on a shifting set of facets and nodes that created a very unique connection and organizational schema. Your system is equally unique. In neither of these is an attempt to come up with some universal set of nodes and facets. Instead the system should allow for both – interesting algorithms and curated collections based on personal schema of understanding.
What we need to avoid is a system of deductive classification that focus on items (nodes) with no regard to relations. Equally we need to avoid inventory systems that preclude multiple member generated systems of organization. We must also avoid the equating of information organization to traditional descriptive cataloging. My guess is that the Pandora example used is not a series of human choices, but rather a human attempt to make sense of a series of neural net connections (actually my guess is it is an artificial example meant to highlight the non-deterministic system of relations).
Thanks for the thoughtful response, Dave. This is my last comment on this post, I promise!
I think you’re right about the Pandora example – artificial and meant to highlight… (And yes, I meant to say that artist was only one type of node!) And I think you’re right about the need for systems that allow “interesting algorithms and curated collections” – the key being *both* (incl the latter). B/c as a user – to reiterate the point – these ‘collective intelligence’ systems (maybe not the best term) like Pandora have yielded very little fruit for me. Much more fruit has been yielded through interactions with individuals with idiosyncratic tastes that are similar to mine in some kind of inexpressible spirit – not in terms of genres, artists or even similar sounds. For the CI system it’s a numbers game that produces a ‘tyranny of the majority’ effect.
I can’t be alone in this; there must be others like me. I just don’t want our story to get lost as web 2.0 is seen increasingly as a panacea. These systems are useful in many ways but not without their limitations. (And I know you aren’t saying this; I’m just using your blog post as a convenient venue for my rally cry :). Thanks again for listening and responding. Cool blog.