Ch. 3 From Correlation to Supervised Segmentation (Intro (The topic of…
Ch. 3 From Correlation to Supervised Segmentation
The topic of this chapter is predictive modeling with supervised segmentation. Supervised segmentations means that we are segmenting based on a target we want to estimate or predict. An example of supervised segmentation is the churn problem where we try to predict which customers are likely to leave. We will also learn about how to find and select informative attributes. Information is a quantity that reduces uncertainty. Lastly we will learn about tree induction which is incorporates the idea of supervised segmentation and selection of attributes.
A predictive model focuses on estimating the target of interest. Data mining is used when we have many attributes and are not sure what the segments should be.
Selecting Informative Attributes
Churn example. Round, square , black white. Attributes: shapes and colors . Target variable write off YES/NO. Purity measure called entropy. Entropy = p1(log(p1))-p2(log(p2)). P is the relative percentage. low=pure.
Example: Attribute Selection with Information Gain
Poisonous and edible mushrooms. Unbalanced data due to unbalance in target.
Supervised Segmentation with Three-Structured Models
Upside down tree with the most significant values first. Ending in leafs that would be seen as decision points . the tree is a supervised segmentation , because each leaf contains a value for the target. Be careful to not overfit the model.
Scatterplot, with x and y axis as the most significant attributes .
Trees as Sets of Rules
These decisions are based on IF, THEN statements . For example IF (Balance <50K) AND (age<50) THEN class= Write-off.
Equation for binary class probability estimation: p(c) =(n+1)/(n+m+2) . n is the number of examples in the leaf belonging to class c, m is the number of examples not belonging to class c.
Example: Addressing the Churn Problem with Three Induction
house, overage, long calls. Test, redesign
Supervised segmentation. attributes and target. Decision tree, good visualization
Models, Induction, and Prediction
Models are simplified representations of reality (ex. map).
Prediction is forecasting of an unknown event or estimate an unknown value. Predictive modeling is used to predict a target value, descriptive modeling is trying to find the attributes associated with the target. Induction is the opposite of deduction. Deduction starts with general rules and facts and create other facts from them.