CISC483 Lecture Notes - Lecture 25: Absolute Difference, Data Mining

• School: University of Delaware
• Department: Computer and Information Sciences
• Course: CISC483
• Professor: Carberry Sandra

Document Summary

As the tree is built, construct a linear model for the instances at each internal node as well as for leaf nodes. Compute the value predicted by the leaf node. Each node contains a regression formula of the form w0+w1a1+ +wkak, where the ai are attributes and the wi are weights. At each interior node, consider pruning the regression formula by removing terms: Estimate the error rate at the node n, using the formula below: where n = # of training instances that reach node n v = # of attributes in the linear regression model for the node. En is the average of the absolute difference between predicted value and actual value for the training instances that retch node n. Compare this with a revised error estimate enr that we get if we remove a term from the regression formula at node n. Greedily remove terms as long as the resultant error rate is lower.