What have we discovered about the Amazon database behind this individualized magic? We already know that for every unique book ID, Amazon must store a tree of subject area categories and subtopics, subcategories, that the book is located in, and a record of how each book ranks in terms of Amazon's own recent sales within each of the subject area categories and subcategories in which it is cross-listed. We are about to see what other dynamic revenue metrics Amazon tracks for each book, and how it stores and uses that information. The top book on the list of 12, remember out of 1,200 possible that Amazon has retrieved and is showing me, is Information theory, Inference, and Learning Algorithms by David MacKay. Which happens to be one of my favorite books which I first read when my brother gave it to me many years ago, and which we now use as a text book in my Duke Master of Engineering Management Data Mining course. I can click on the book to get more information about it. Look carefully at the new page layout. First, we get Frequently Bought Together with a suggestion that we buy the Cover and Thomas Elements of Information Theory. Note the subtle selling suggestion, that learning information theory in relation to machine learning requires buying more than one book. Because people frequently, quote frequently unquote, bought the Cover and Thomas Elements of Information Theory textbook, which is also excellent together with the MacKay book. To my surprise, Amazon is not offering a discount for purchasing both together. Obviously because they have found that doing so does not increase sales enough to justify it. We don't actually know what percentage of buyers of MacKay also bought Cover and Thomas or vice versa. But Amazon is clearly tracking every purchase of more than one book at a time, and maintaining an up-to-date database for every book it sells of the other books sold during the same user session to the same customer ID. Amazon offers viewers the top 100 books in terms of frequency that people bought at the same time that they bought the MacKay textbook. This kind of data is called co-occurrence data. It can be represented in a co-occurrence matrix, much more on this later. Co-occurring sales is obviously a very important dynamic metric to Amazon. A question now arises to which I do not know the answer. If you do know, please send me a note. Does Amazon list the top 100 co-occurring sales books in simple order from the highest co-occurrence to the lowest? If yes, then Cover and Thomas is not actually the book bought most frequently with MacKay. That would be Pattern Recognition and Machine Learning by Christopher Bishop. Of course, Amazon did not say that it was bought together with MacKay most frequently, only frequently. It is likely that the frequently bought together book offered is chosen through A/B testing experiments by Amazon that suggests that showing Cover together with MacKay increased sales or profits by more than showing Bishop or maybe any of the other 98 co-occurring books on the list. In this case, Amazon has created best upsell revenue potential in the frequently bought together marketing slot. On the other hand, if they are presenting the 100 books out of ranked order, maybe Cover and Thomas really is the book most frequently bought with MacKay. But, Amazon knows we've already been given information about it and we didn't bite on their offer to buy both together. So Amazon makes a small change in the true ordering in the list of 100 to put the second-most popular book to buy together with MacKay first. Because it is a better use of the top slot to display a new title to us. It is a mystery to me. But in either case, the combined placement in the Frequently Bought Together display and the top 100 frequently bought together display, is certainly optimized through A/B testing to maximize sales. In addition to tracking what books people buy together at the same time, Amazon maintains a co-occurrence database of every book that a web user clicked on while browsing. Wow, that is a lot of detail. Clickstream data about everything I ever looked at but did not buy. How do I know that Amazon does this? Because, at the bottom of the same page is a separate co-occurrence list of the books people most commonly bought during the same session when they viewed the MacKay textbook without buying it. So we know that Amazon is maintaining a co-occurrence matrix that lists for every visit that results in the sale of at least one book to other possibilities. First, when book one is bought, it has a list of all other books bought. And second, when book one is bought, it has a list of all other books viewed. By using this second set of records, Amazon can identify what books were ultimately purchased by people who viewed but did not buy any given book. So, to summarize, Amazon's first line of promotion is the profit maximizing frequently bought together metric. Their second metric is a kind of recommendation engine. What's likely to be also be bought based on the assumption that we do ultimately purchase the first book. The third metric is a recommendation based on the purchases people made who looked at but did not buy the first book.
