Comments on: A Blooper from “The World’s Best Retailer”

By: Daniel Tunkelang

Daniel Tunkelang — Thu, 02 Apr 2009 17:43:16 +0000

A friend of mine pointed out that Amazon’s public relations team may have bigger fish to fry when it comes to damage control. I hadn’t been aware of the “rape simulator” game incident:

http://www.belfasttelegraph.co.uk/news/local-national/amazon-selling-3d-rape-simulator-game-14183546.html

From a public relations perspective, that’s probably a lot more of an issue than the girl scout cookies query–and the game is truly disturbing. But I’m admittedly more concerned with understanding the quirks of their search functionality.

By: Daniel Tunkelang

Daniel Tunkelang — Tue, 31 Mar 2009 15:30:00 +0000

Well, I can see how a coincidence of independent and non-malicious user-supplied tags might conspire. But I'd love to know whether the results were accidental or intentional. Now it seems possible that someone from the outside could have manipulated the rankings through strategically placed user tags, e.g., making good use of infrequently assigned tags. Think evil!

By: Ken

Ken — Tue, 31 Mar 2009 15:18:24 +0000

Wisdom of the crowds, my friend. Can’t go wrong there…

By: Daniel Tunkelang

Daniel Tunkelang — Tue, 31 Mar 2009 15:15:39 +0000

Ah, so user-supplied tags may be the culprit? The plot thickens…

By: Ken

Ken — Tue, 31 Mar 2009 15:11:18 +0000

Maybe it was based on your purchasing history David. The speculums and wolf urine are gone, but under the first girl scout costume listed, I do see them under “Customers Who Viewed This Item Also Viewed”. They might have left open a positive feedback mechanism, and its amplifying random associations, although the appearance in that section may have been effect rather than cause. If they’re smart, they would not include those who viewed the item by using the “Also Viewed…” links in the stats to generate the “Also Viewed…” links, or use those to fill out sparse search results. It seems clear that the speculums and wolf urine were included not because of a lexical match to the name or description, but because of an association with something that did match, and the “Also Viewed” seems like a good candidate. It may also have been some common tags that were applied. The GS outfit was tagged with “Morgan Freeman”, which could explain why some saw products related to him. The Wolf Urine is also tagged with “valentines day gift”, and the GS outfit may have been so tagged at some point. Pity I couldn’t eyeball it before it was fixed. I wonder if we can get this to happen to an obscure search phrase just by having a few dozen people look at a related then an unrelated product in sequence, and for another pair have users tag them similarly with unique (obscure) values.

By: Daniel Tunkelang

Daniel Tunkelang — Tue, 31 Mar 2009 04:38:54 +0000

Actually, machines are perfectly able to follow instructions, and most sites that cater to heterogeneous audiences (at least in the prudish United Sates) go to great lengths to avoid presenting “adult” content to users who don’t clearly want to find it. Google turns on moderate safe search by default; other sites require you to click through an age verification page; etc.

Of course, Amazon’s content here isn’t quite in that category. Still, it’s not that hard to manually label the areas of the sites that might cause offense and make those products harder to find by accident. We actually had to do something like that in our first ecommerce deployment at Endeca, for a site that sold a wide variety of movies.

Still, no algorithmic or heuristc approach is perfect, which is why transparency is the best policy.

By: Josh Young

Josh Young — Tue, 31 Mar 2009 04:26:13 +0000

For sure, algorithms should work to avoid this kind of really unfortunate–if hilarious–blunder. And, I agree, it would help if Amazon were more transparent about it.

I can’t help but suspect, however, that part of what shocks us as humans is based on our mistakenly anthropomorphizing code. I’m certainly no expert in AI, but my sense is that machines cannot themselves make moral decisions. “Magical theories of reference” don’t work.

By: Daniel Tunkelang

Daniel Tunkelang — Tue, 31 Mar 2009 04:02:18 +0000

To err is human. But to return results like these requires an opaque relevance ranking algorithm. Seriously, they would look a lot better if they could explain what happened. Given the amount of discussion on the web, I imagine they’ll do so. I certainly hope so: I’d love to understand how things could go so wrong. It’s a bit early for an April Fool’s joke, and this is far too over the top to have been sanctioned by management.

By: Daniel Lemire

Daniel Lemire — Tue, 31 Mar 2009 03:51:47 +0000

Fascinating. And unavoidable.

By: Daniel Tunkelang

Daniel Tunkelang — Tue, 31 Mar 2009 03:50:29 +0000

I can live with that. It’s the wolf urine and, um, instruments, that are harder to explain to my family and friends.

By: Jason Adams

Jason Adams — Tue, 31 Mar 2009 03:49:08 +0000

The problem with searching for girl scout cookies on amazon is that you get recommended “sexy” female costumes and girls clothing for a while.

By: Daniel Tunkelang

Daniel Tunkelang — Mon, 30 Mar 2009 19:52:35 +0000

You can follow the discussion in Amazon's forums here. Also, it looks like someone tweeted about this a week ago.