The Recommender System at Amazon.com
Get instant personalized recommendations based on your prior purchases the moment you log on. ~Amazon.com
Project Description AND Objectives:
The idea that recommender systems may give relevant and useful information about what is being recommended, what is similar and potentially useful, was the basis for our first assignment. We looked at how Amazon.com recommended books purchased by those who had bought a particular book. However, Amazon.com seemed to have two methods of recommending books: one for fiction and one for non-fiction. Fiction books recommended other fiction books by the same author, and eventually recommended similar subjects written by authors of similar background and of the same literary genre, typically books I already was aware of because of the book jacket. Non-fiction books recommended other non-fiction books of similar subject, with a wide variety of authors.
I edlook at how this system quantifies books without qualifying them, except in the review process, which doesn’t seem that helpful. I say this because my own experience with the recommender system, where I purchased four books recommended because others bought them as they had also purchased the book I originally purchased, was disappointing.
For example, I purchased Plein Air Painters of the North, a non-fiction book with some excellent art history essays and a lot of color plates of Northern Californian artist’s works, where the subject matter was typically landscape and painted during the period from 1885 to 1940. Amazon recommended California Impressionism. The reviews were all very positive. However, the essays in the recommended book, California Impressionism, were not as academically focused, nor as historically relevant, and the works tended more towards flowery still-life or portraiture and overall, while it is a handsome book, I found it half as informative and helpful in the research I was doing.
Amazon.com’s reviewers, anyone who wishes to publish a review on the site, tend often to be written by those who either loved the book, or hated it, but often the people who seem willing to go to the trouble of reviewing are the ones who are enthusiastic about a work. Not being professional reviewers or writers usually, they do not always articulate very well why the book is good or not good, but rather they often pass on adulatory comments.
I compared the “recommendation” systems in real space at book stores, to Amazon.com’s system, as well as look at a couple of other recommended systems, to see if there might be a way to improve the recommendation system at Amazon.com. Areas looked at include how employees in real space stores recommend works, what other methods of recommending stores use, and how useful those recommendations are. Then, comparing those real space recommendations, to online recommendations, I looked at where the differences are or patterns of differences that might help to see where this system is not as relevant as real space.
I also will use my personal experiences with Amazon.com since I go to the site five or six times a week, and purchase about twice a week. I’ve been purchasing there for four years.
“ONLINE READERS ARE IN NEED OF TOOLS TO HELP THEM COPE with the mass of content available on the World-Wide Web. In traditional media, readers are provided assistance in making selections. This includes both implicit assistance in the form of editorial oversight and explicit assistance in the form of recommendation services such as movie reviews and restaurant guides.” - Marko Balabanovi´c and Yoav Shoham, “Content Based Collaborative Recommendation”, Communications of the ACM, Volume 40 , No. 3 (Mar. 1997)
FINAL PROJECT REALIZATION:
Data Collection from Real Space:
I went to Cody’s Books, and talked with two employees. I asked each one to recommend some fiction, based upon the fact that I had recently read Blindness, by Jose Saramago. Neither had read this book, but they had read others by Saramago, and were familiar with his style, the fact that he is Portuguese, and his style of magical realism.
The first employee recommended the following books, in no particular order (comments are the employees):
Ralph Elliosn, The Invisible Man Really recommended. Said it starts wide, and is open, and then as the subject closes and goes back in time, the style matches this by becoming more traditional. It also follows the path of slowly becoming more and more aware of the blackness of the main character.
The second employee recommended the following, in no particular order (comments are the employees):
She said that she had not read Blindness, but this book was watery, it was always raining, and it was a dreamscape. And she really recommended it.
She said this was dense, dark, about a man who can only see his landlady as a big blue skirt, he’s a misogynist and hates everyone, hard to read but funny and great.
It’s a first novel.
All time classic.
Hot Teen Sex.
Much better than the movie, maybe a forgotten book as it’s very dated 80’s stuff, but really good. Harsh.
Customers tell her they stay up until 5am reading this book.
Bruno Schultz was a writer, who while in a concentration camp, got into the good graces of a guard, and another prisoner was envious and shot him. His work was published after the war. Two novels only. But really good.
Other Recommender Systems:
I also looked at the Reel.com system, just to compare and get ideas about how other systems might work. Reel.com works much the way Amazon.com does. It recommends movies by the director or with actors from the movie currently being looked at, or very similar genres. For example, looking at The African Queen, recommended “close movie matches” included An Affair to Remember, Casablanca, The Rainmaker, Summertime, and To Have and Have Not. All but one have Katherine Hepburn or Humphry Bogart. But Reel.com also has “creative movie matches”, and these included Romancing the Stone, Rooster Cogburn, Six Days, Seven Nights and Two Mules for Sister Sara. While these creative matches might not be the greatest choices, for one reason or another, they are matches. For example, Romancing the Stone is about two people who fall in love while going through the jungle to find Kathleen Turner’s sister and to get a large jewel. But it’s set in 1984 verses The African Queen, of 1951, which is about two people who go through the jungle because there is no other way out, and fall in love. But they are different in sensibility in a lot of ways, although I’m sure there are plenty of people who like both movies. However, according to Efron and Geisler in Is it all About Connections? Factors Affecting the Performance of a Link Based Recommender System , Reel.com uses both a weighted system of descriptions manually placed by Reel.com reviewers, as well as director, actor and genre information to recommend movies. However, Efron and Geisler have devised a new recommender system, RecEx, to try to solve this problem, and conclude that the recommendation idea still needs evaluation.
I noticed that the books that were recommended by employees, when looked up on Amazon.com, all listed other books by the same author under the recommended section. This confirmed my assignment results, and those discussed in class, where the conclusion was the Amazon.com will, for a fiction book, first recommend books by the same author, and then list books that are by people of similar background and writing genre. The owner of a work of fiction can typically look at the book jacket to find other books by the same author. And I was aware of the other Amazon.com recommended books by the same author, and so the recommendation system results just become something to move past to get to something else more useful.
The Cody’s employee recommendations were subjective, but they made the dynamic leap that a reader of a serious work of fiction, who might like some magical realism but also might like some other premises as well, towards fiction, or even poetry or short stories, that was more interesting and desirable than the Amazon.com recommendations. It was interesting to note that several of their recommendations overlapped either by book, or by author, and yet, they also recommended works by women, people of other races and distinctly different cultural experiences, and very different subject matters. And yet, as far as the books I was familiar with, these multidimensional recommendations made sense even though they could not be related back (other than one time) to Saramago, Portuguese authors, or even magical realism in exactly the same vein as Blindness.
In fact, every time I go to a bookstore in Santa Cruz on the main street, I walk in, and ask an employee to recommend a couple of books. I’ve never been disappointed; this is how I came upon Jose Saramago’s Blindness. It may be that the presence of a person, emphatically communicating their enjoyment of a book, or why they liked it or the fact that it’s a first novel, makes more of an impression than reading an enthusiastic review online. However, these same details are available on Amazon.com and I believe can be exploited to make the recommender system richer.
The human recommenders seemed to have more insight into a broader array of multidimensional literature, and what readers might like based on a particular choice. I found myself getting really interested in the choices they were recommending that I was not familiar with. Could the Amazon.com system reflect a more sophisticated recommendation system beyond just the similarity model? It might be too expensive for Amazon.com to pay professional reviewers, or even to pay for reviews published elsewhere, other than the editorial reviews they put above the Reviews section. However, it seems as though Amazon.com might dramatically improve their recommender system by using a filtering system to blend recommendations for similar books “others have purchased who purchased this book” with a number of other pieces of information, to achieve broader and hopefully more qualitative recommendations.
However, Amazon.com would have to go beyond their seeming current practice of using classification of author and book, and matching like books, to something more sophisticated and useful. This may involve asking returning users questions about past purchases, or blending different kinds of information with the current classification data.
The crucial thing missing from the traditional geographies is the failure to appreciate how environments are conceived by people as opposed to simply perceived by people. ~Munro, Hook & Benyon
Ideas for making the AMAZON.com system better:
- Allow returning users to recommend books recently purchased, without necessarily reviewing them, with a one click system. Then blend these rankings with purchase recommendations. This would hopefully achieve a greater sampling than the reviewers section, which seems to attract readers of a work who really liked it, giving it a high number of stars. The idea would be to achieve a larger and more reliable pool of purchasing recommenders verses the smaller pool of reviewers.
- Allow users to chose words and or categories to describe the work, to help Amazon.com see how users see the works they have purchased. This could be done after a purchase when a returning user logs on weeks later. These words and categories could be matched to works under review by others and details like, “first novel”, “magical realism”, “historical” or “dreamscape” could be weighted and blended with recommendations.
- User profiles could be blended with those lists of similar profiles for books purchased by others, to gain a more refined recommendation.
- After a recommended book listing, a rating average (a la Sfgate.com’s list of movies with critics’ ratings from 1 to 10), could be listed, averaging Amazon reviews and critical reviews. The New York Times Bestseller List might be used, although it has in the past been a reflection of Barnes & Noble sales. Maybe a book critics rating could be started in order to average numerical critical reviews.
- Top Amazon.com reviewers with high helpfulness ratings and similar tastes could have their other reviewed books blended into recommendations, and their purchases weighted more heavily.
People are active participants in reshaping the space. There is a dynamic relationship between people, their activities in space, and the space itself. All three are subject to change. ~Munro, Hook & Benyon
Instead of this for Blindness:
Customers who bought this book also bought:
We might get these recommendations:
Customers who bought this book also bought:
CACM special section on Recommender Systems edited by Paul
Resnick and Hal Varian, Communications of the ACM, Volume 40 , No. 3 (Mar. 1997).
Is it all About Connections? Factors Affecting the Performance of a Link Based Recommender System by Miles Efron and Gary Geisler, August 23, 2001. Presented at the ACM SIGIR 2001 conference.
“Footprints in the Snow” by Alan
Munro, Kristin Hook and David Benyon, from Social Navigation of Information Space,