Chapter 8. Regression – Recommendations Improved
At the end of the last chapter, we used a very simple method to build a recommendation engine: we used regression to guess a ratings value. In the first part of this chapter, we will continue this work and build a more advanced (and better) rating estimator. We start with a few ideas that are helpful and then combine all of them. When combining, we use regression again to learn the best way to combine them.
In the second part of this chapter, we will look at a different way of learning called
basket analysis, where we will learn how to make recommendations. Unlike the case in which we had numeric ratings, in the basket analysis setting, all we have is information about shopping baskets, that is, what items were bought together. The goal is to learn recommendations. You have probably already seen features of the form "people who bought X also bought Y" in online shopping. We will develop a similar feature of our own.
Remember where we stopped in the previous chapter: with a very basic, but not very good, recommendation system that gave better than random predictions. We are now going to start improving it. First, we will go through a couple of ideas that will capture some part of the problem. Then, what we will do is combine multiple approaches rather than using a single approach in order to be able to achieve a better final performance.
We will be using the same movie recommendation dataset that we started off with in the last chapter; it consists of a matrix with users on one axis and movies on the other. It is a sparse matrix, as each user has only reviewed a small fraction of the movies.
Using the binary matrix of recommendations
One of the interesting conclusions from the Netflix Challenge was one of those obvious-in-hindsight ideas: we can learn a lot about you just from knowing which movies you rated, even without looking at which rating was given! Even with a binary matrix...
The methods we have discussed so far work well when you have numeric ratings of how much a user liked a product. This type of information is not always available.
Basket analysis is an alternative mode of learning recommendations. In this mode, our data consists only of what items were bought together; it does not contain any information on whether individual items were enjoyed or not. It is often easier to get this data rather than ratings data as many users will not provide ratings, while the basket data is generated as a side effect of shopping. The following screenshot shows you a snippet of Amazon.com's web page for the book War and Peace, Leo Tolstoy, which is a classic way to use these results:
This mode of learning is not only applicable to actual shopping baskets, naturally. It is applicable in any setting where you have groups of objects together and need to recommend another. For example, recommending additional recipients to a user writing an e-mail is done by...
In this chapter, we started by improving our rating predictions from the previous chapter. We saw a couple of different ways in which to do so and then combined them all in a single prediction by learning how to use a set of weights. These techniques, ensemble or stacked learning, are general techniques that can be used in many situations and not just for regression. They allow you to combine different ideas even if their internal mechanics are completely different; you can combine their final outputs.
In the second half of the chapter, we switched gears and looked at another method of recommendation: shopping basket analysis or association rule mining. In this mode, we try to discover (probabilistic) association rules of the customers who bought X are likely to be interested in Y form. This takes advantage of the data that is generated from sales alone without requiring users to numerically rate items. This is not available in scikit-learn (yet), so we wrote our own code (for a change...
