Chapter 1

The Learning Problem

If you show a picture to a three-year-old and ask if there is a tree in it, you will likely get the correct answer. If you ask a thirty-year-old what the denition of a tree is, you will likely get an inconclusive answer. We didnt learn what a tree is by studying the mathematical denition of trees. We learned it by looking at trees. In other words, we learned from data. Learning from data is used in situations where we dont have an analytic solution, but we have data that we can use to construct an empirical solution. The premise covers a lot of territory, and indeed learning from data is one of the most widely utilized techniques in science, engineering, and economics, among other elds. In this chapter, we present examples of learning from data and formalize the learning problem. We also discuss the main concepts associated with learning, and the dierent paradigms of learning that have been developed.

1. The Learning Problem

1.1. Example

bu y? ? ed tion lock b om c s c es a fers e ik e lik l pr

r ste

s? om sT C

is ru

e?

e lik

viewer
add contributions from each factor

Match movie and viewer factors movie

predicted rating

m To

Figure 1.1: A model for how a viewer rates a movie

co

r? ste bu nt ck nte blo co tent n tio on ac d y c me

Cr uis n ei it?

1.1

Example

What is common between nancial forecasting, medical diagnosis, computer vision, and search engines? They all have successfully utilized learning from data. The repertoire of applications of learning from data is quite impressive. Let us start here with a specic application in order to see how learning from data works. Consider the problem of predicting how a movie viewer would rate the dierent movies out there. This is an important problem if you are a company that rents out movies, since you want to recommend to dierent viewers the movies they will like. Good recommendation systems are so important to business that the movie rental company Netix oered a prize of one million dollars to anyone
c
AM L

Abu-Mostafa, Magdon-Ismail, Lin. DRAFT, August 25, 2011.

12

1. The Learning Problem

1.1. Example

who could improve their recommendations by a mere 10%. The problem captures the essence of learning from data. There is no question that how people would rate a given movie is related to how they rated other movies, and to how others rated that movie. One can even assert a few qualitative rules, such as similar viewers are likely to rate a future movie similarly, or viewers who have preferred romantic movies to action movies are likely to rate a chick ick higher than a car chase movie. However, these rules do not provide a specic prediction of how a viewer will rate a particular movie. Learning from data does. Let us see how that would work. The main diculty in this problem is that the criteria that viewers use to rate movies are quite complex. Trying to model those explicitly is no easy task, so it may not be possible to come up with an analytic solution. However, we know that the historical rating data reveal a lot about how people rate movies, so we may be able to construct a good empirical solution. There are tons of data available to movie rental companies, since they often ask their viewers to rate the movies that they have already seen. Figure 1.1 illustrates a specic approach that was widely used in the milliondollar competition. Here is how it works. You describe a movie as a long array of dierent factors, e.g., how much comedy is in it, how complicated is the plot, how handsome is the lead actor, etc. Now, you describe each viewer with corresponding factors; how much do they like comedy, do they prefer simple or complicated plots, how important are the looks of the lead actor, and so on. How this viewer will rate that movie is now estimated based on the match/mismatch of these factors. For example, if the movie is pure comedy and the viewer hates comedies, the chances are he wont like it. If you take dozens of these factors describing many facets of the movie content and viewer taste, the conclusion based on matching all the factors will be a good predictor of how the viewer will rate the movie. The power of learning from data is that this entire process can be automated, without any need for analyzing movie content or viewer taste. The learning algorithm simply starts with random factors. As it gets data about how viewers rated movies in the past, it ne-tunes the factors until they conform to how viewers rate movies. This produces factors that are specically related to movie rating, and also determines the correct level that dierent viewers and dierent movies have for each factor. Since the algorithm generates factors that are derived from ratings data, these factors may not have an intuitive interpretation like comedy content for example. In fact, they can be quite subtle, even incomprehensible, because the algorithm is only trying to nd the best way to predict how a viewer would rate a movie, not necessarily explain to us how it is done. This type of algorithm performed quite well in the million-dollar competition,
c
AM L

Abu-Mostafa, Magdon-Ismail, Lin. DRAFT, August 25, 2011.

13

1. The Learning Problem

1.1. Example

and was an important part of the winning solution.

AM L

Abu-Mostafa, Magdon-Ismail, Lin. DRAFT, August 25, 2011.

14