Decision Tree

Predicting Response at BookBinders: Decision Trees

Recursive partitioning algorithms (or decision trees) are a versatile tool for uncovering patterns or relationships in data. They are especially useful when there is a large set of potential predictors and when you are not sure which are most important or what the relationships between the predictors and the target (dependent) variable are. In the case of a binary target variable, decision tree algorithms iteratively search through the data to find which predictors best separate the two categories of the target variable.

So far, we have used RFM segmentation and logistic regression to predict the response to the mailing offer for “The Art History of Florence.” Now we will see how decision trees compare as an alternative.

Tree using Exhaustive CHAID

We’ll start with a tree using the exhaustive CHAID algorithm. Our target variable is BUYER (whether or not they bought The Art History of Florence) and all other variables will be potential predictors. To see how the tree grows, let’s take it one step at a time, beginning with the ‘root node’. Because decision trees are prone to ‘overfitting’, we will split the dataset in two: two-thirds of the observations will be used to develop the model (the training sample) and the remaining one-third will be used to test the model (the validation or test sample) to see how well it does on ‘new’ observations.

Next, we’ll grow the tree one level. After specifying this, the SPSS software searches through the potential predictor variables to see which one ‘best’ separates the buyers from the non-buyers, and we see that gender is selected:

To see how well this one-level tree classifies buyers and non-buyers we can look at the classification table and ‘risk estimate’. In SPSS, the risk estimate is the percent of customers incorrectly classified:

Misclassification Matrix | | |

| Actual Category | |

| No | Yes | Total |

Predicted...