1. Decision Trees

3. DEFINE: Set X of Instances (of n-tuples x = ) –E.g., days decribed by attributes (or features): Sky, Temp, Humidity, Wind, Water, Forecast Target function y, e.g.: –EnjoySport X  Y = {0,1} (example of concept learning) –WhichSport X  Y = {Tennis, Soccer, Volleyball} –InchesOfRain X  Y = [0, 10] GIVEN: Training examples D –positive and negative examples of the target function: FIND: A hypothesis h such that h(x) approximates y(x). General Learning Task

4. Hypothesis Spaces Hypothesis space H is a subset of all y: X  Y e.g.: –MC2, conjunction of literals: –Decision trees, any function –2-level decision trees (any function of two attributes, some of three) Candidate-Elimination Algorithm: –Search H for a hypothesis that matches the training data Exploits general-to-specific ordering of hypotheses Decision Trees –Incrementally grow tree by splitting training examples on attribute values –Can be thought of as looping for i = 1,...,n: Search H i = i-level trees for hypothesis h that matches data

17. Decision Trees represent disjunctions of conjunctions = (Sunny ^ Normal) v Overcast v (Rain ^ Weak)

18. Decision Trees vs. MC2 Yes Yes No MC2 can’t represent (Sunny v Cloudy) MC2 hypotheses must constrain to a single attribute value if at all Vs. Decision Trees:

20. Learning Parity with D-Trees  How to solve 2-bit parity:  Two step look-ahead  Split on pairs of attributes at once  For k attributes, why not just do k-step look ahead? Or split on k attribute values? =>Parity functions are the “victims” of the decision tree’s inductive bias.

22. = I(Y; xi)

26. Overfitting is due to “noise” Sources of noise: –Erroneous training data concept variable incorrect (annotator error) Attributes mis-measured –Much more significant: Irrelevant attributes Target function not deterministic in attributes

27. Irrelevant attributes If many attributes are noisy, information gains can be spurious, e.g.: 20 noisy attributes 10 training examples =>Expected # of depth-3 trees that split the training data perfectly using only noisy attributes: 13.4 Potential solution: statistical significance tests (e.g., chi-square)

28. Non-determinism In general: –We can’t measure all the variables we need to do perfect prediction. –=> Target function is not uniquely determined by attribute values

29. Non-determinism: Example HumidityEnjoySport 0.900 0.871 0.800 0.750 0.701 0.691 0.651 0.631 Decent hypothesis: Humidity > 0.70  No Otherwise  Yes Overfit hypothesis: Humidity > 0.89  No Humidity > 0.80 ^ Humidity <= 0.89  Yes Humidity > 0.70 ^ Humidity <= 0.80  No Humidity <= 0.70  Yes

30. Rule #2 of Machine Learning The best hypothesis almost never achieves 100% accuracy on the training data. (Rule #1 was: you can’t learn anything without inductive bias)

39. Hypothesis Space comparisons Hypothesis Space H# of Semantically distinct h Rote Learning MC2 1 + 3 k 1-level decision tree k n-level decision tree Task: concept learning with k binary attributes

40. Decision Trees – Strengths Very Popular Technique Fast Useful when –Instances are attribute-value pairs –Target Function is discrete –Concepts are likely to be disjunctions –Attributes may be noisy

41. Decision Trees – Weaknesses Less useful for continuous outputs Can have difficulty with continuous input features as well… –E.g., what if your target concept is a circle in the x1, x2 plane? Hard to represent with decision trees… Very simple with instance-based methods we’ll discuss later…

43. decision tree learning algorithm; along the lines of ID3