Amazon’s Recommendations are Probabilistic

Amazon.com is a fascinating website. It’s one of the first eCommerce websites, but it started with a somewhat unique strategy. The initial launch of the site included such a comprehensive implementation of functionality that there are sites today that are still struggling to catch up. Why? Because much of the functionality that Amazon implemented early and continued to improve didn’t directly attempt to solve the problems most retailers face: What products do I offer? How often do we change our offerings? And so on. Instead, Amazon attempted to set up a self-organizing system based on past usage and user preferences.

For the first several years of Amazon’s existence, they operated at a net loss due to the high initial cost in setup. Competitors who didn’t have such expenses seemed to be doing better. Indeed, Amazon’s infamous recommendations were often criticized, and anyone who has used Amazon regularly has certainly had the experience of wondering how in the world they managed to recommend something so horrible. But over time, Amazon’s recommendations engine has gained steam and produced better and better recommendations. This is due, in part, to improvements in the system (in terms of the information collected, the analysis of that information, and the technology used to do both of those things). Other factors include the growth of both Amazon’s customer base and their product offerings, both of which improved their recommendation technology.

As I’ve written about before, the important thing about Amazon’s system is that it doesn’t directly solve retailing problems, it sets up a system that allows for efficient collaboration. By studying purchase habits, product ratings, common wishlist items, etc… Amazon is essentially allowing it’s customers to pick recommendations for one another. As their customer base and product offerings grow, so does the quality of their recommendations. It’s a self-organizing system, and recommendations are the emergent result. Many times, Amazon makes connections that I would have never made. For instance, a recent recommendation for me was the DVD set of the Firefly TV series. When I checked to see why (this openness is an excellent feature), it told me that it was recommended because I had also purchased Neal Stephenson’s Baroque Cycle books. This is a connection I probably never would have made on my own, but once I saw it, it made sense.

Of course, the system isn’t perfect. Truth be told, it probably won’t ever be perfect, but overall, I’d bet that its still better than any manual process.

Chris Anderson (a writer for Wired who has been exploring the Long Tail concept) has an excellent post on his blog concerning these systems, which he refers to as “probabalistic systems:”

When professionals–editors, academics, journalists–are running the show, we at least know that it’s someone’s job to look out for such things as accuracy. But now we’re depending more and more on systems where nobody’s in charge; the intelligence is simply emergent. These probabilistic systems aren’t perfect, but they are statistically optimized to excel over time and large numbers. They’re designed to scale, and to improve with size. And a little slop at the microscale is the price of such efficiency at the macroscale.

Anderson’s post is essentially a response to critics of probabilistic systems like Wikipedia, Google, and blogs, all of which have come under fire because of their less-than-perfect emergent results. He does an excellent job summarizing the advantages and disadvantages of these systems and it is highly recommended reading. I reference it for several reasons. It seems that Amazon’s website qualifies as a probabilistic system, and so the same advantages and disadvantages Anderson observes apply. It also seems that Anderson’s post touches on a few themes that often appear on this blog.

First is that human beings rarely solve problems outright. Instead, we typically seek to exchange one set of disadvantages for another in the hopes that the new set is more desirable than the old. Solving problems is all about tradeoffs. As Anderson mentions, a probabilistic system “sacrifices perfection at the microscale for optimization at the macroscale.” Is this tradeoff worth it?

Another common theme on this blog is the need for better information analysis capabilities. Last week, I examined a study on “visual working memory,” and it became apparent that one thing that is extremely important when facing a large amount of information is the ability to figure out what to ignore. In information theory, this is referred to as the signal-to-noise ratio (technically, this is a more informal usage of the terms). One of the biggest challenges facing us is an increase in the quantity of information we are presented with. In the modern world, we’re literally saturated in information, so the ability to separate useful information from false or irrelevant information has become much more important.

Naturally, these two themes interact. As I concluded last week’s post: ” Like any other technological advance, systems that help us better analyze information will involve tradeoffs.” While Amazon, Wikipedia, Google or blogs may not be perfect, they do provide a much deeper look into a wider variety of subjects than their predecessors.

Is Wikipedia “authoritative”? Well, no. But what really is? Britannica is reviewed by a smaller group of reviewers with higher academic degrees on average. There are, to be sure, fewer (if any) total clunkers or fabrications than in Wikipedia. But it’s not infallible either; indeed, it’s a lot more flawed that we usually give it credit for.

Britannica’s biggest errors are of omission, not commission. It’s shallow in some categories and out of date in many others. And then there are the millions of entries that it simply doesn’t–and can’t, given its editorial process–have. But Wikipedia can scale to include those and many more. Today Wikipedia offers 860,000 articles in English – compared with Britannica’s 80,000 and Encarta’s 4,500. Tomorrow the gap will be far larger.

The good thing about probabilistic systems is that they benefit from the wisdom of the crowd and as a result can scale nicely both in breadth and depth.

[Emphasis Mine] The bad thing about probabilistic systems is that they sacrifice perfection on the microscale. Any individual entry at Wikipedia may be less reliable than its Britannica counterpart (though not necessarily), and so we need to take any single entry with a grain of salt.

The same is true for blogs, no single one of which is authoritative. As I put it in this post, “blogs are a Long Tail, and it is always a mistake to generalize about the quality or nature of content in the Long Tail–it is, by definition, variable and diverse.” But collectively they are proving more than an equal to mainstream media. You just need to read more than one of them before making up your own mind.

I once wrote a series of posts concerning this subject, starting with how the insights of reflexive documentary filmmaking are being used on blogs. Put simply, Reflexive Documentaries achieve a higher degree of objectivity by embracing and acknowledging their own biases and agenda. Ironically, by acknowledging their own subjectivity, these films are more objective and reliable. Probabilistic systems would also benefit from such acknowledgements. Blogs seem to excell at this, though it seems that many of the problems facing Wikipedia and other such systems is that people aren’t aware of their subjective nature and thus assume a greater degree of objectivity than is really warranted.

It’s obvious that probabilistic systems are not perfect, but that is precisely why they work. Is it worth the tradeoffs? Personally, I think they are, provided that such systems properly disclose their limitations. I also think it’s worth noting that such systems will not fully replace non-probabilistic systems. One commonly referenced observation about Wikipedia, for instance, is that it “should be the first source of information, not the last. It should be a site for information exploration, not the definitive source of facts.”