What do they call the feature on amazon.com when they say, "People who bought this book also bought...these other books"?
September 18, 2006 12:56 PM Subscribe
What do they call the feature on amazon.com when they say, "People who bought this book also bought...these other books"?
What's the technical term for the "people who bought this also bought..." on amazon.com and various other sales websites?
What's the technical term for the "people who bought this also bought..." on amazon.com and various other sales websites?
An older term with some relevance is: "collaborative filtering"
posted by Good Brain at 1:07 PM on September 18, 2006 [1 favorite]
posted by Good Brain at 1:07 PM on September 18, 2006 [1 favorite]
Suggestive selling is the term I've always heard.
posted by jbickers at 4:13 PM on September 18, 2006
posted by jbickers at 4:13 PM on September 18, 2006
People also call it a "recommender" system, if you're looking in the literature.
posted by maschnitz at 5:35 PM on September 18, 2006
posted by maschnitz at 5:35 PM on September 18, 2006
Up-selling.
posted by JamesMessick at 6:56 PM on September 18, 2006
posted by JamesMessick at 6:56 PM on September 18, 2006
Thanks, vacapinta.
The Firefly recommendation system was primitive in comparison. Of the versions of recommendation systems that Amazon has filed a patent for, this patent application sounds most what is being asked about:
Use of product viewing histories of users to identify related products
It has a section which details "Use of Purchase Histories to Identify Related Items" which seems to address the question that ostranenie poses.
In the summary section, the inventors describe some of the limitations of collaborative filtering that they (not me) see:
1. "...users of online stores frequently do not take the time to explicitly rate the products, or create lists of their favorite products."
2. Collaborative filtering would be based upon ratings or a user profile of a searcher. Not helpful if you are shopping for something for someone else. This process provides recommendations based upon other searchers' purchases and product views.
3. In a collaborative filtering system, new items without ratings wouldn't get recommended, and may not build enough of a critical mass to get ranked until a good number of ratings have been received.
4. A realtime comparison amongst users profiles when there are thousands, or tens of thousands of users would take to long or be too computationally expensive.
5. A collaborative filtering system wouldn't take into account "current preferences" of users of the system, or that a user is searching for a particular type or category of item.
posted by bragadocchio at 7:11 PM on September 18, 2006
The Firefly recommendation system was primitive in comparison. Of the versions of recommendation systems that Amazon has filed a patent for, this patent application sounds most what is being asked about:
Use of product viewing histories of users to identify related products
It has a section which details "Use of Purchase Histories to Identify Related Items" which seems to address the question that ostranenie poses.
In the summary section, the inventors describe some of the limitations of collaborative filtering that they (not me) see:
1. "...users of online stores frequently do not take the time to explicitly rate the products, or create lists of their favorite products."
2. Collaborative filtering would be based upon ratings or a user profile of a searcher. Not helpful if you are shopping for something for someone else. This process provides recommendations based upon other searchers' purchases and product views.
3. In a collaborative filtering system, new items without ratings wouldn't get recommended, and may not build enough of a critical mass to get ranked until a good number of ratings have been received.
4. A realtime comparison amongst users profiles when there are thousands, or tens of thousands of users would take to long or be too computationally expensive.
5. A collaborative filtering system wouldn't take into account "current preferences" of users of the system, or that a user is searching for a particular type or category of item.
posted by bragadocchio at 7:11 PM on September 18, 2006
I remember collab-filtering even before Firefly, when it was called RINGO (and Pattie Maes was something of an internet rockstar).
Perhaps it was the novelty, but the demo version always felt more accurate than Amazon's various recommendations do now. That's probably because it was based upon active expression of preference rather than the additional features in Amazon's patent. For something that was bright, shiny and new, it was easy to procure volunteers... which makes me wonder if the Firefly team got paid for all of that data, and not just the model.
posted by holgate at 12:18 AM on September 19, 2006
Perhaps it was the novelty, but the demo version always felt more accurate than Amazon's various recommendations do now. That's probably because it was based upon active expression of preference rather than the additional features in Amazon's patent. For something that was bright, shiny and new, it was easy to procure volunteers... which makes me wonder if the Firefly team got paid for all of that data, and not just the model.
posted by holgate at 12:18 AM on September 19, 2006
As someone who has done a bit of work on collaborative filtering (hence my collab-filtering newsbot): the difference between RINGO, Firefly, MyNews, and Amazon was a pretty significant breakthrough: previous recommendation systems emphasized similarities between people (person-to-person collab filtering) whereas Greg Linden at Amazon turned the process on its head and based the recommendations on similar items (item-to-item collab filtering). For various technical reasons, i2i works much better for e-commerce than p2p, so Amazon's innovation was far from trivial.
If you care about this topic Greg's blog (linked above) is a very good read. BTW, Greg is now running findory, an i2i collab-filtering newsbot (whereas say memigo, my own newsbot is p2p). Personally I think p2p makes a bit more sense in news filtering, but diversity is a good thing!
posted by costas at 12:49 AM on September 19, 2006
If you care about this topic Greg's blog (linked above) is a very good read. BTW, Greg is now running findory, an i2i collab-filtering newsbot (whereas say memigo, my own newsbot is p2p). Personally I think p2p makes a bit more sense in news filtering, but diversity is a good thing!
posted by costas at 12:49 AM on September 19, 2006
Patti Maes still is a rock star. Love her InterestMap (pdf) paper.
I'll second the recommendation on Greg Linden's blog. If you dig through his archives, he has a series of posts about what it was like to work for Amazon that are very good. In one of them, he talks about some of the challenges he faced while doing some of the programming involving recommendations.
posted by bragadocchio at 8:45 AM on September 19, 2006
I'll second the recommendation on Greg Linden's blog. If you dig through his archives, he has a series of posts about what it was like to work for Amazon that are very good. In one of them, he talks about some of the challenges he faced while doing some of the programming involving recommendations.
posted by bragadocchio at 8:45 AM on September 19, 2006
Witchcraft.
posted by oxford blue at 11:30 PM on September 19, 2006
posted by oxford blue at 11:30 PM on September 19, 2006
This thread is closed to new comments.
posted by chrisamiller at 12:58 PM on September 18, 2006