A classifier predicts which category a data point belongs to using one or more predictors. Its outcome is discrete—such as “purchased” or “not purchased”—rather than a continuous value such as a price. Choose classification when the decision calls for a category; use regression when it calls for a continuous prediction. section 9.1.1 section 9.4
Table 9.1 Classes of models, their purposes, and examples
|
Class
|
Purpose
|
Example
|
|---|---|---|
|
Regression
|
Predict outcomes represented with continuous data
|
Predicting student test scores
|
|
Classification
|
Predict outcomes represented with categorical data
|
Predict which students will graduate on time
|
|
Clustering
|
Identify underlying patterns and natural “groups” in the data
|
Find clusters of student performance across classes
|
|
Dimension reduction
|
Identify shared variance among predictors to simplify model input
|
Find 2 underlying components to represent scores in 12 classes
|
|
Survival analysis
|
Predict the time until an event occurs
|
Predict the time to withdraw from an intensive graduate curriculum
|
The model-class comparison places classification beside regression and clustering, making the choice of method depend on the outcome needed. section 9.1.1
Categorizing support tickets
A support team needs to assign each ticket a priority—Low, Medium, or High—and a type, such as Billing or Delivery, before routing work. A supervised classifier is one candidate; a keyword-based rules system is another. An estimate that 20 keywords cover 60% of tickets is not a tested accuracy result. Compare the approaches on correct categorization and manual work saved, and check whether the system has unintended negative effects on groups of customers. section 8.2.2
Predicting trial conversion
A marketing team can use predicted conversion likelihood to tailor engagement emails. One workflow prepares training and test sets, tries logistic regression and tree-based models such as random forest, and evaluates out-of-sample predictions with accuracy, precision, recall, and AUC. In the reported iteration, those measures were 84%, 74%, 65%, and 0.76, respectively. Hourly predictions were then made available through a data-warehouse table linked to customer IDs. section 9.3.2
