Classification Trees
- By Vikram Kole
The goal of a classification tree is to sequentially partition the data to maximize the differences in the dependent variable. It is often referred to as a decision tree. The true purpose of a classification tree is to classify the data into distinct groups or branches that create the strongest separation in the values of the dependent variable.
Classification trees are very good at identifying segments with a desired behavior such as response or activation. This identification can be quite useful when a company is trying to understand what is driving market behavior. It also has an advantage over regression in its ability to detect nonlinear relationships. Classification trees are “grown” through a series of steps and rules that offer great flexibility. In Figure, the tree differentiates between responders and non responders. The top node represents the performance of the overall campaign. Sales pieces were mailed to 10,000 names and yielded a response rate of 2.6%. The first split is on gender. This implies that the greatest difference between responders and non responders is gender. We see that males are much more responsive (3.2%) than females (2.1%).
If we stop after one split, we would consider males the better target group. Our goal, though, is to find groups within both genders that discriminate between responders and non responders. In the next split, these two groups or nodes are considered separately. The second-level split from the male node is on income. This implies that income level varies the most between responders and non responders among the males. For females, the greatest difference is among age groups. It is very easy to identify the groups with the highest response rates. Let’s say that management decides to mail only to groups where the response rate is more than 3.5%. The offers would be directed to males who make more than $30,000 a year and females over age 40. A definite advantage of classification trees over other techniques is their ability to explain the results. We can often develop complex logistic models for scoring and build a tree to explain the results to the marketers. Although the outcome is never identical, the tree does a good job of uncovering key drivers in the market. Due to their broad applicability, classification trees will continue to be a valuable tool for all types of target modeling.
Posted on April 30th, 2008 by Vikram Kole
Filed under: Business Intelligence
Leave a Reply