数据挖掘讲座 Classification：Model Overfittingand Classifier Evaluation.pdfVIP

0 1 0 2 , g n i r Lecture 12: e e n i Classification: Model Overfitting and g n E a Classifier Evaluation t a D t n e g i l l e t n I A Review of Decision Tree  Generally, it is a SEARCH problem. - Select a hypothesis: Tree structure. - Set the objective.  Global: Minimizing the misclassification error.  Local: Maximizing the impurity gain. - Search from the hypothesis space to find the very hypothesis that optimize the objective.  Top down, greedy, iterative.  Stopping criteria: Same label in a leaf; same attribute values in a node; early stopping. 2 Strongnesses and Issues  Advantages: - Inexpensive to construct - Extremely fast at classifying unknown records - Easy to interpret for small-sized trees - Accuracy is comparable to other classification techniques for many simple data sets  Issues: - Overfitting - Missing attribute values - Data fragmentation, decision boundary, tree replication 3 Classification Errors  Generalization errors - Expected error of a model over random selection of records from same distribution  Training errors (apparent errors) - Errors committed on the training set  Test errors - Errors committed on the test set 4 Example Data Set Two class problem: +, o 3000 data points (30% for