Earlier this summer, I wrote about personalized coupons. After thinking about the topic some more, I realized that personalized Internet coupons are not so simple, since customers can easily compare and share deals on various internet coupons, leading to perceived–if not actual–inequity. One way to be fair while utilizing data mining is to offer generic coupons with personalized recommendations, thus making the coupon seem more appealing.
I can’t help but notice personalized recommendations everywhere. Many online retailers take advantage of this by mining my purchases and even items that I have viewed. Online retailers have a wealth of data to mine. Even catalog companies seem to do a decent job of sending personalized recommendations with very little data. Retailers such as LL Bean or Lands End will send a semi-personalized mini-catalog to customers who receive gifts, whom they presumably know nothing about except for the gift that was sent–such as a catalog filled with only children’s clothing when a gift for a child is sent.
Before I proceed, I have a question for you: What was your first purchase with Amazon?
The first purchase I had from Amazon more than a decade ago was a Rage Against the Machines album. Over the years, Amazon has made many Rage Against the Machine recommendations to me, despite other purchases and frequent browsing. Amazon has a complex set of rules for recommending products that are clearly useful: I have added many things to my wish list based on recommendations. Amazon’s approach doesn’t seem to work that well for me, seeing as I rarely make purchases there, but I suspect that my infrequent shopping is more related to the lack of Amazon coupons (rather than their recommendations) as well as my extreme frugality. As a result, I have a love-hate relationship with Amazon’s recommendations.
I was impressed, however, with Amazon’s MP3 recommendations for me, when I recently visited their MP3 store for the first time. This really shocked me, since I was expecting a bunch of Rage Against the Machine recommendations. Most of the songs that Amazon recommended were not from bands/artists whose music I purchased or even browsed. I listened to many of the songs, and I really liked them. It got me thinking about how valuable personalized recommendations are, particularly for things like music, movies, books, and other items that we consume frequently. The perfect vacuum recommendation algorithm is not as valuable, since I plan to purchase few vacuums over the course of my lifetime. But I’m always on the lookout for better music, movies, and books (and so is everyone else, as evident by Netflix’s domination and the interest in their contests).
This week, I read two articles that touch on the science behind these recommendations (at least what could be reported in a news article).
- Wired article about Hunch, collects information from users via survey questions to make better recommendations.
- WSJ article about [x+1], Inc., secretly collects a small amount of information from users to make better recommendations.
Both companies essentially try to figure out which stereotypes might apply to consumers in order to make more tailored offers or recommendations. For example, the output of the [x+1] algorithm could be used to determine which credit card offers to make to a first-time web visitor on a bank site. No matter how little is known about someone, personalized recommendations could be made. That surprised me, since it seems like I often provide a lot of information (through browsing at Amazon or rating movies at Netflix) to get personalized recommendations. It was also a little Big-Brotheresque–I couldn’t help but wish that the [x+1] algorithm was much, much worse.
Personalized recommendations used to be so time-consuming back in the day. I’d get book, music, and movie recommendations from friends when growing up. And, let’s be honest, they were often terrible! I was a voracious reader as a teen, and I would outpace my friend’s recommendations. And none of my friends really read fantasy or SF novels, so I was totally lacking recommendations in those genres. In the end, I would pick a bunch of books from the library until I found a series that was addictive. But I had to read a lot of bad books along the way. It’s so easy to find good books now.
How did you get recommendations before the Internet?
Where do you get your best personalized recommendations now? I like Pandora, Netflix, and Amazon MP3. And google, if you consider search engine results to be the ultimate personalized recommendation.
Related posts:
Punk Rock Operations Research 
August 11th, 2010 at 10:12 am
To answer your questions: my first Amazon purchase was books (but don’t ask me which); before the Internet I didn’t rely on recommendations; and after the Internet I still don’t (although I occasionally look at customer ratings on e-tail sites like Amazon).
As far as personalized coupons go, I think Amazon tried some sort of user-tailored pricing strategy a while back and got screamed at when customers compared notes and found they’d paid different prices for the same book. IIRC their response was it that it was a (discontinued) experiment.
But Hal Varian was writing about information-based price discrimination years ago, and one method that sounds feasible to me is bundling. (See, for instance, section 3 of http://people.ischool.berkeley.edu/~hal/Papers/price-info-goods.pdf.) If Amazon sees you are shopping for X and their data mining suggests you might also like Y (complementary or unrelated product), they can quote you a price for an X-Y bundle. It’s fairly unlikely very many customers will compare notes and have been offered the same bundle (let alone on the same date — consumers are accustomed to sale prices appearing and disappearing), so differential pricing likely won’t create hard feelings.
August 11th, 2010 at 3:10 pm
I’m not sure about my first amazon.com order, but I looked through my order history and the earliest two they have on record are Sheldon Ross’ “Introduction to Stochastic Dynamic Programming” and then The Shins “Oh, Inverted World” from Fall 2002.
More generally about recommendation systems: I find that generally they do not work very well for me, although I do sometimes get reasonable recommendations from Pandora or iTunes “Genius”; it does seem helpful that iTunes pretty much has access to all of the music I own (unlike amazon.com, which sees only a small fraction of my shopping history).
If you have diverse tastes (especially in music), obviously you might be difficult to figure out. Often, if I really like some music, I don’t necessarily want to buy an album by a new artist that makes music with lots of similar “features”.
My personal opinion is that these systems work best if they attempt for each shopper to build a custom group of “similar people”, and then recommend to you things that your group also likes and/or buys.
August 12th, 2010 at 7:45 am
I don’t recall my first ever purchase via Amazon. I think it was a book. But I know my last order was a Liftgate Strut for our minivan. I had to install new ones after the old ones weren’t holding up the liftgate anymore. It’s amazing what you can find on Amazon.
I am very interested in recommendation systems. I found the Netflix Prize forum (http://www.netflixprize.com/community/) a fascinating insight into the algorithms that drive recommendation systems. In fact I was meaning to do a blog post on that subject.