1. MACHINE LEARNING 10 Decision Trees

2. Motivation  Parametric Estimation  Assume model for class probability or regression  Estimate parameters from all data  Non-Parametric  Find “similar”/”close” data points  Fit local model using these points  Costly computation of distance from all training data Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 2

3. Motivation  Pre-split training data into region using small number of simple rules organized in hierarchical manner  Decision Trees  Internal decision nodes have splitting rule  Terminal leaves have class labels for classification problem or values for regression problem Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 3

4. Tree Uses Nodes, and Leaves Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 4

5. Decision Trees  Start from univariate decision trees  Each node looks only at single input feature  Want smaller decision trees  Less memory for representation  Less computation for a new instance  Want smaller generalization error Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 5

6. Decision and Leaf Node  Implement simple test function f m (x)  Output: labels of branches  f m (x) discriminant in d-dimensional space  Complex discriminant is broken down into hierarchy of simple decisions  Leaf node describes a region in d-dimensional space with same value  Classification label  Regression value Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 6

7. Classification Trees  What is the good split function?  Use Impurity measure  Assume N m training samples reach node m    Node m is pure if for all classes either 0 or 1  Need values in between Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 7

8. Entropy  Measure amount of uncertainty on a scale from 0 to 1  Example: 2 events  If p1=p2=0.5, entropy is 1 which is maximum uncertainty  If p1=1=1-p0, entropy is 0, which is no uncertainty Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 8

9. Entropy Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 9

10. Best Split Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 10  Node is impure, need to split more  Have several split criteria (coordinates), have to choose optimal  Minimize impurity (uncertainty) after split  Stop when impurity is small enough  Zero stop impurity=>complex tree with large variance  Larger stop impurity=>small tress but large bias

11. Best Split  Impurity after split: N mj of N m take branch j.  N i mj belong to C i  Find the variable and split that min impurity  among all variables  split positions for numeric variables Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 11

12. Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 12 ID3 algorithm for Classification and Regression Trees(CART)

13. Regression Trees  Value not a label in a leaf nodes  Need other impurity measure  Use Average Error Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 13

14. Regression Trees Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 14  After splitting:

15. Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 15 Example

16. Pruning Trees Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 16  Number of data instances reach a node is small  Less then 5% of training data  Don’t want to split further regardless of impurity  Remove subtrees for better generalization  Prepruning: Early stopping  Postpruning: Grow the whole tree then prune subtrees Set aside pruning set Make sure pruning does not significantly increase error

17. Decision Trees and Feature Extraction  Univariate Tree uses only certain variable  Some variables might not get used  Features closer to the root have greater importance Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 17

18. Interpretability  Conditions that are simple to understand  Path from the root =>one conjunction of test  All paths can be defined using set of IF_THEN rules  Form a rule base  Percentage of training data covered by the rule  Rule support  Tool for a Knowledge Extraction  Can be verified by experts Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 18

19. Rule Extraction from Trees Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 19 C4.5Rules (Quinlan, 1993)

20. Rule induction  Learn rules directly from data  Decision-tree is a breadth-first rule construction  Rule induction: depth-first construction  Start  Learn rules one by one  Rule is a conjunction of conditions  Add condition one by one by certain criteria  Entropy  Remove samples covered by rule from training data Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 20

21. Ripper algorithm  Assume two classes K=2, positive and negative examples  Add rules to explain positive examples, all other examples are classified as negative  Foil algorithm: add condition to a rule to maximize information gain Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 21

22. Multivariate Trees Based on for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1) 22