1. Supervised Learning I, Cont’d Reading: Bishop, Ch 14.4, 1.6, 1.5

2. Administrivia I Machine learning reading group Not part of/related to this class We read advanced (current research) papers in the ML field Might be of interest. All are welcome Meets Fri, 2:00-3:30, FEC325 conf room More info: http://www.cs.unm.edu/~terran/research/reading_group/ Lecture notes online

3. Administrivia II Microsoft on campus for talk/recruiting Feb 5 Mock interviews Hiring both undergrad, grad, interns Office hours Wed 9:00-10:00 and 11:00-noon Final exam time/day Incorrect in syllabus (last year’s. Oops.) Should be: Tues, May 8, 10:00-noon

4. Yesterday & today Last time: Basic ML problem Definitions and such Statement of the supervised learning problem Today: HW 1 assigned Hypothesis spaces Intro to decision trees

5. Homework 1 Due: Tues, Jan 30 Bishop, problems 14.10, 14.11, 1.36, 1.38 Plus: A) Show that entropy gain is concave (anti- convex) B) Show that a binary, categorical decision tree, using information gain as a splitting criterion, always increases purity. That is, information gain is non-negative for all possible splits, and is 0 only when the split leaves the data distribution unchanged in both leaves.

6. Feature (attribute): Instance (example): Label (class): Feature space: Training data: Review of notation

7. Finally, goals Now that we have and, we have a (mostly) well defined job: Find the function that most closely approximates the “true” function The supervised learning problem:

8. Goals? Key Questions: What candidate functions do we consider? What does “most closely approximates” mean? How do you find the one you’re looking for? How do you know you’ve found the “right” one?

9. Hypothesis spaces The “true” we want is usually called the target concept (also true model, target function, etc.) The set of all possible we’ll consider is called the hypothesis space, NOTE! Target concept is not necessarily part of the hypothesis space!!! Example hypothesis spaces: All linear functions Quadratic & higher-order fns.

10. Space of all functions on Visually... Might be here Or it might be here...

11. More hypothesis spaces Rules if (x.skin==”fur”) { if (x.liveBirth==”true”) { return “mammal”; } else { return “marsupial”; } } else if (x.skin==”scales”) { switch (x.color) { case (”yellow”) { return “coral snake”; } case (”black”) { return “mamba snake”; } case (”green”) { return “grass snake”; } } } else {... }

12. More hypothesis spaces Decision Trees

13. More hypothesis spaces Decision Trees

14. Finding a good hypothesis Our job is now: given an in some and an, find the best we can by searching Space of all functions on

15. Measuring goodness What does it mean for a hypothesis to be “as close as possible”? Could be a lot of things For the moment, we’ll think about accuracy (Or, with a higher sigma-shock factor...)

16. Aside: Risk & Loss funcs. The quantity is called a risk function A.k.a., empirical loss function Approximation to true (expected) loss: (Sort of) measure of distance between “true” concept and approximation to it All functions on

17. Constructing DT’s, intro Hypothesis space: Set of all trees, w/ all possible node labelings and all possible leaf labelings How many are there? Proposed search procedure: 1. Propose a candidate tree, t i 2. Evaluate accuracy of t i w.r.t. X and y 3. Keep max accuracy t i 4. Go to 1 Will this work?

18. A more practical alg. Can’t really search all possible trees Instead, construct single tree Greedily Recursively At each step, pick decision that most improves the current tree

19. A more practical alg. DecisionTree buildDecisionTree(X,Y) { // Input: instance set X, label set Y if (Y.isPure()) { return new LeafNode(Y); } else { Feature a=getBestSplitFeature(X,Y); DecisionNode n=new DecisionNode(a); [X0,...,Xk,Y0,...,Yk]=a.splitData(X,Y); for (i=0;i<=k;++i) { n.addChild(buildDecisionTree(Xi,Yi)); } return n; } }

20. A bit of geometric intuition x 1 : petal length x 2 : sepal width

21. The geometry of DTs Decision tree splits space w/ a series of axis orthagonal decision surfaces A.k.a. axis parallel Equivalent to a series of half-spaces Intersection of all half-spaces yields a set of hyper-rectangles (rectangles in D>3 dimensional space) In each hyper-rectangle, DT assigns a constant label So a DT is a piecewise-constant approximator over a sequence of hyper-rectangular regions