Prediction Revisited

3Co-occurrence

So far, we have only considered attributes separately. When we observe more than one attribute, how do we know if their joint occurrence is typical? The answer lies in learning how they tend to co-occur. Whereas previously we looked at pairs of observations, we now investigate pairs of attributes.

Co-occurrence Conceptually

Imagine you are a sales analyst for an online retailer, and you want to forecast how much a customer will spend based on a collection of attributes about this person. For example, suppose you know this person's age, annual income, time spent on your website, and education level. Before you take the leap of predicting, which we will come to later, you must first understand the typical relationship between these attributes. Only then can you gauge whether this customer is an archetype or an anomaly.

First, consider the relationship between age and income. Here, your observation is a single customer; age and income form a pair of attributes. You select at random a person who is 25 years old and earns $75,000 per year. What does the alignment of these attributes say about this shopper? Is this shopper:

• Younger with lower income, implying a positive relationship?
• Younger with higher income, implying a negative relationship?
• Older with higher income, implying a positive relationship?
• Older with lower income, implying a negative relationship?

Start with what you know. By applying the techniques in Chapter 2, you know that:

• The average age of ...