Programming Collective Intelligence

Chapter 8. Building Price Models

So far we have examined several classifiers, most of which are well suited for predicting to which category a new piece of data belongs. However, Bayesian classifiers, decision trees, and support-vector machines (which you’ll see in the next chapter) are not the best algorithms for making predictions about numerical data based on many different attributes, such as prices. This chapter will look at algorithms that can be trained to make numerical predictions based on examples they have seen before, and even display probability distributions for the predictions to help the user interpret how the prediction is being made.

We’ll be looking at how to use these algorithms for building models that predict prices. Economists consider prices, particularly auction prices, to be a good method of using collective intelligence to determine the real value of something; in a large market with many buyers and sellers, the price will usually reach the optimal value for both sides of the transaction. Price prediction is also a good test for algorithms of this kind, since there are usually many different factors to consider when determining a price. When considering bidding on a laptop, for example, you have to take into account processor speed, installed RAM, hard drive size, screen resolution, and other factors.

An important part of making numerical predictions is determining which variables are important and in what combinations. In the laptop example, there are likely ...