A Tutorial on Support Vector Machine

Transcription

1 A Tutorial on School of Computing National University of Singapore

2 Contents Theory on Using with Other s

3 Contents Transforming Theory on Using with Other s

4 What is a classifier? A function that maps instances to classes. Instance usually expressed as a vector of n attributes. Example: Fit-And-Trim club : Gender, Weight, Height Class (Member): Yes, No Is Garvin a member of the Fit-And-Trim club? Person Gender Weight Height Member Garvin Male ?

5 Training a classifier: Given: Training data (a set of instances whose classes are known). Output: A function, selected from a predefined set of functions. Example: Fit-And-Trim club Training instances: Person Gender Weight Height Member Alex Male Yes Betty Female Yes Charlie Male Yes Daisy Female No Eric Male No Fiona Female No Possible classification rule: Weight more than 74.5?

6 Training a classifier: Aim to minimize both training error (errors on seen instances) and generalization error (errors on unseen instances). A classifier that has low training error but high generalization error is said to overfit the training data. Example: Fit-And-Trim club Suppose we use person name as an attribute, and the trained classifier uses the following classification rules: Alex Yes Betty Yes Charlie Yes Daisy No Eric No Fiona No This classifier severely overfits: it achieves 100% accuracy on the training data, but is unable to classify any unseen test instance.

7 are real numbers, so each instance x = (x 1,x 2,...,x n ) T R n is a n-dimensional vector. Classes are +1 (positive class) and 1 (negative class). Classification rule: y = sign[f(x)] = where the decision function f( ) is { +1 if f(x) 0 1 if f(x) < 0 f(x) = w 1 x 1 +w 2 x w n x n +b = w T x+b for some weight vector w = (w 1,w 2,...,w n ) T R n and bias b R. Training a linear classifier means tuning w and b.

8 Example: A course is graded based on two written tests. Weightage: Test 1 30%, Test 2 70%. Students pass if the total weighted score is at least 50%. : x 1 = Test 1 score, x 2 = Test 2 score. To pass, we need: 0.3x x 2 50 Decision function of linear classifier: f(x) = 0.3x x 2 50 Positive class = pass, negative class = fail.

9 f(x) = w T x+b = w 1 x 1 +w 2 x w n x n +b f(x) = 0 is a hyperplane in the n-dimensional real space R n. Training: Find a hyperplane that separates positive and negative instances. Exercise 1 Show that w is orthogonal to the hyperplane.?

10 Given a linear classifier with: f(x) = w 1 x 1 +w 2 x w n x n +b The following are equivalent classifiers: f(x) = (aw 1 )x 1 +(aw 2 )x (aw n )x n +(ab) for any constant a > 0. f(x) = w 1 x w i (ax i)+...+w n x n +b for any constant a > 0 and where w i = w i a. f(x) = w 1 x w i (x i +a)+...+w n x n +b for any constant a and where b = b aw i. In other words, possible to scale whole problem, and scale and shift individual attributes.

11 Using training data, we can normalize each attribute x i : 0 x i 1. x i has mean 0 and standard deviation 1. Normalization makes training easier by avoiding numerical difficulties. For linear classifier with normalized attributes: f(x) = w 1 x 1 +w 2 x w n x n +b Larger w i or w 2 i means x i more important or relevant. Can be used for attribute ranking or selection. If x i is missing, can be used to decide whether to acquire it.

12 Transforming If attribute values are ordered, use the ordering. Consider attribute values {Small, Medium, Large}. Map Small to 1. Map Medium to 2. Map Large to 3. If attribute values have no ordering information, create one numeric attribute whose values are {0, 1} for each discrete attribute value. Consider attribute values {Red, Green, Blue}. Create three attributes x Red, x Green, x Blue. Map Red to x Red = 1, x Green = 0, x Blue = 0. Map Green to x Red = 0, x Green = 1, x Blue = 0. Map Blue to x Red = 0, x Green = 0, x Blue = 1. If attribute values are finite sets, use the above method.

13 Contents Theory on Using with Other s

14 Narrow margin Wide margin ρ ρ Wide margin gives better generalization performance.

15 Support vector machine (SVM) is a maximum margin linear classifier. Training data: D = {(x 1,y 1 ),(x 2,y 2 ),...,(x N,y N )}. Assume D is linearly separable, i.e., can separate positive and negative instances exactly using a hyperplane. SVM requires: Training instance xi with class y i = +1: f(x i ) = w T x i +b +1. Training instance xi with class y i = 1: f(x i ) = w T x i +b 1. Combined, we have y i f(x i ) = y i (w T x i +b) 1. Support vector: A training instance x i with f(x i ) = w T x i +b = ±1.

16 ρ = 2/ w support vectors w f(x) = +1 f(x) = 0 f(x) = 1 Test instance x: Greater f(x) Greater classification confidence. Maximize ρ Minimize w 2 Minimize w T w. ( w 2 = w 2 1 +w w2 n is the length or 2-norm of the vector w) Exercise 2 Show that ρ = 2 w 2.

17 Primal problem: Minimize 1 2 wt w Subject to y i (w T x i +b) 1 Primal problem is convex optimization problem, i.e., any local minimum is also global minimum. But more convenient to solve the dual problem instead.

18 Let f(u 1,u 2,...,u k ) be function of variables u 1, u 2,..., u k. Partial derivatives: Partial derivative f u i means differentiate w.r.t. u i while holding all other u j s as constants. Example: f(x,y) = sinx+xy 2, f x = cosx+y2, f y = 2xy. Stationary point when f u i = 0 for all i. Partial derivatives w.r.t. vectors: Let u = [u 1,u 2,...,u n ] T. [ ] f = f u = f T. u 1, f u 2,..., f u n Stationary point when f = f u = 0. Exercise 3 Verify that u at u = u ut a = a and u ut u = 2u.

19 Primal problem: Minimize f(u) Subject to g 1 (u) 0,g 2 (u) 0,...,g m (u) 0 Lagrangian function: L(u,α,β) = f(u)+ h 1 (u) = 0,h 2 (u) = 0,...,h n (u) = 0 m α i g i (u)+ The α i s and β i s are Lagrange multipliers. n β i h i (u) Optimal solution must satisfy Karush-Kuhn-Tucker (KKT) conditions: L u = 0 g i(u) 0 h i (u) = 0 α i g i (u) = 0 α i 0

20 KKT conditions: Dual problem: L u = 0 g i(u) 0 h i (u) = 0 Maximize Subject to α i g i (u) = 0 α i 0 f(α,β) = inf L(u,α,β) u KKT conditions (inf means infimum or greatest lower bound) Relation between primal and dual objective functions: f(u) f(α,β) Under certain conditions (satisfied by SVM), equality occurs at optimal solution.

21 Primal problem: Lagrangian: Minimize 1 2 wt w Subject to 1 y i (w T x i +b) 0 L(w,b,α) = 1 2 wt w+ KKT conditions: = 1 2 wt w+ α i (1 y i (w T x i +b)) α i α i y i w T x i b α i y i L N w = w α i y i x i = 0 L N b = α i y i = 0 1 y i (w T x i +b) 0 α i (1 y i (w T x i +b)) = 0 α i 0

22 Since w = N α iy i x i, we have: w T w = α i y i w T x i = Note also N α iy i = 0. L(w,b,α) = 1 2 wt w+ = α i 1 2 α i j=1 j=1 α i α j y i y j x T i x j α i y i w T x i b α i y i α i α j y i y j x T i x j

23 Dual problem: Maximize α i 1 2 Subject to α i 0 α i y i = 0 j=1 α i α j y i y j x T i x j Dual problem is also convex optimization problem. Many software packages available and tailored to solve this form of optimization problem efficiently.

24 Solving w and b: w = N α iy i x i. Choose support vector x i, then b = y i w T x i. Decision function: f(x) = N α iy i x T i x+b. Support vectors: KKT condition: α i (1 y i (w T x i +b)) = 0. α i > 0 y i (w T x i +b) = 1 x i is support vector. Usually relatively few support vectors, many α i s will vanish.

25 Narrow margin Not linearly separable Solution: Allow training instances inside the margin or on the other side of the separating hyperplane (misclassified).

26 Soft margin SVM ε i ε i ε i Allow training instance (x i,y i ) to have y i (w T x i +b) < 1, or y i (w T x i +b) = 1 ε i for ε i > 0, but penalize it with a constant factor C > 0. ε i

27 Primal problem: Dual problem: Minimize Subject to Maximize Subject to 1 2 wt w+c ε i y i (w T x i +b) 1 ε i ε i 0 α i α i C α i y i = 0 j=1 α i α j y i y j x T i x j Exercise 4 Derive the dual problem for the soft margin SVM.

28 Recall: 0 α i C. x i is support vector if α i > 0. Two types of support vectors: x i is free support vector if α i < C. x i is bounded support vector if α i = C. Exercise 5 What are the characteristics of free support vectors and bounded support vectors? (Hint: Consider the KKT conditions.)

29 Possible to weigh each training instance x i with λ i > 0 to indicate its importance. Primal problem: Minimize Subject to Cost-sensitive SVM: 1 2 wt w+c λ i ε i y i (w T x i +b) 1 ε i ε i 0 Different costs for misclassifying positive and negative instances. Set λ i proportional to misclassification cost of x i. Exercise 6 Derive the dual problem for this SVM.

30 Nonlinear feature map Φ : R n R m from attribute space to feature space. Training instances linearly separable in feature space. Φ(x) = (x,x 2 ) T Cover s theorem: Instances more likely to be linearly separable in high dimension space.

31 Dual problem: Maximize Subject to Decision function: Observations: α i α i C α i y i = 0 f(x) = j=1 α i α j y i y j Φ(x i ) T Φ(x j ) α i y i Φ(x i ) T Φ(x)+b Computationally expensive or difficult to compute Φ(x). But often Φ(x) T Φ(x ) has simple expression.

32 Dual problem: Maximize Subject to Decision function: α i α i C α i y i = 0 f(x) = j=1 α i y i K(x i,x)+b α i α j y i y j K(x i,x j ) The kernel is the function K(x,x ) = Φ(x) T Φ(x ). Exercise 7 Express w T w and b in terms of the kernel. What is the margin of the kernelized SVM?

33 K(, ) can take any expression that satisfies Mercer s condition to ensure K(x,x ) = Φ(x) T Φ(x ) for some feature map Φ( ): Mercer s condition: matrix K formed by K ij = K(x i,x j ) is positive semidefinite, i.e., eigenvalues of K are all nonnegative, or v T Kv 0 for all vectors v. No need to compute or even know explicit form of Φ( ). Examples: Linear kernel: K(x,x ) = x T x. Polynomial kernel: K(x,x ) = (x T x +k) d. Radial basis function (RBF) kernel: K(x,x ) = exp( x x 2 2σ 2 ). Suggestion: Try linear kernel first, then RBF kernel.

34 Contents Theory on Using Estimating Handling with Other s

35 Dual problem: Maximize Subject to RBF kernel: α i α i C α i y i = 0 j=1 ( x x K(x,x ) 2 ) = exp 2σ 2 α i α j y i y j K(x i,x j ) Hyperparameters include C of SVM formulation and kernel parameters, e.g., σ 2 of RBF kernel. Important to select good hyperparameters, has large impact on SVM performance.

36 Traditional way of parameter tuning: Use a training set and a validation set. Do a grid search on the parameters, e.g., when using RBF kernel for each C {2 10,2 8,...,2 0,...,2 8,2 10 } and σ 2 {2 10,2 8,...,2 0,...,2 8,2 10 }: Train on training set. Test on validation set. Choose C and σ 2 that gives best performance on validation set. Extensions of this idea: k-fold cross-validation and leave-one-out cross-validation. Some SVM implementations have integrated (and more optimized) parameter tuning, others require user to perform own tuning before training.

37 Estimating Probability of test instance x belonging to a class can be estimated by: P(y = 1 x) = 1 1+exp(Af(x)+B) A and B are parameters to be determined from training data.

38 Handling For k > 2 classes, decompose into multiple two-class SVMs. Pairwise SVMs: Train ( k 2) SVMs, one for each pair of classes. For a test instance, each SVM casts a vote on a class. Classify test instance as the class that gets the most votes. One-against-all SVMs: Train k SVMs, the ith SVM considers class i as positive and all other classes as negative. Classify test instance to class i if the ith SVM gives the greatest f( ) value.

39 : Text processing: Each text document is an instance. Each word is an attribute. Attribute values are word counts. Image processing: Each image is an instance. Each pixel is an attribute. Attribute values are pixel values. Cover s theorem: Instances more likely to be linearly separable in high dimension space. These applications already have many attributes, and each instance is a sparse vector, so linear SVM often performs well.

40 SVM decision function: f(x) = w 1 x 1 +w 2 x 2 +w 3 x 3 +w 4 x 4 +b 2-norm of weight vector is w 2 = w 2 1 +w2 2 +w2 3 +w2 4. Test instance with missing attribute values: x = (x 1,?,x 3,?) Substitute with constants attributewise: x = (x 1,x 2,x 3,x 4 ) Decision function becomes: f(x) = w 1 x 1 +w 2 x 2+w 3 x 3 +w 4 x 4+b = w 1 x 1 +w 3 x 3 +b New weight vector is w (x) = (w 1,w 3 ) T and its 2-norm decreased to w (x) 2 = w 2 1 +w2 3. Can use f(x), w 2 and w (x) 2 to estimate probability of misclassification, and hence whether to acquire missing attribute values.

41 Contents Theory on Using with Other s of s

42 Advantages of SVM: SVM can handle linear relationships between attributes, but typical decision tree split only on one attribute at a time. Typically better on datasets with many attributes. Handles continuous values well. Advantages of decision tree: Decision tree is better at handling nested if-then-else type of rules, which SVM is not good at. Typically better on datasets with fewer attributes. Handles discrete values well.

43 Advantages of SVM: Faster classification compared to neighbours. Smooth separating hyperplane. Less suspectible to noise. Advantages of neighbours: Training is instantaneous nothing to do. Local variations are considered.

44 often has the dubious honour of being placed among the last in evaluations. is fast and easy to implement, hence often treated as a baseline. One key handicap of naïve Bayes is its independence assumption.

45 A linear SVM can be seen as a neural network with no hidden layers. But training algorithm is different. SVM design is driven by sound theory, neural network design is driven by applications. Backpropagation algorithm for training multi-layer neural network does not find the maximal margin separating hyperplane. Neural network tend to overfit more than SVM. Neural network have many local minima, SVM has only one global minima. Multi-layer neural networks require specifying number of hidden layers and number of nodes at each layer, but SVM does not need them. What do the learned weights in a trained multi-layer neural network mean?

46 of s Possible to combine classifiers: SVM with decision tree: Instead of splitting based only on one attribute, split using a linear classifier trained by SVM. SVM with neighbours: Use neighbours to classify test instances inside the margin. SVM with naïve Bayes: Train classifier, produces conditional probabilities P(x i c). Modify every (training and testing) instance x by multiplying each attribute x i by P(x i c). Train and test on the modified instances.

47 Contents Theory on Using with Other s

48 SVM is a maximum margin linear classifier. Often good for classifying test instances in high dimensional space. Lagrangian theory primal and dual problems. When training instances are not linearly separable: Use soft margin. Use kernel. Usage tips: Transforming non-numeric attributes. Normalize attributes. Tune parameters when training. with other classifiers.

49 Introduction to SVM, with applications: [Burges, 1998], [Osuna et al., 1997]. Practical guide on using SVM effectively: [Hsu et al., 2003]. LIBSVM manual, including weighted instances: [Chang and Lin, 2001]. Tutorial on Lagrangian theory: [Burges, 2003], [Klien, 2004]. SVM with posterior probabilities: [Platt, 2000]. Attribute selection using SVM: [Guyon et al., 2002]. Handling multiple classes: [Hsu and Lin, 2002], [Tibshirani and Hastie, 2007]. SVM combined with... Decision tree: [Bennett and Blue, 1998], [Tibshirani and Hastie, 2007]. neighbours: [Chiu and Huang, 2007]. : [Li et al., 2007].

50 References I Bennett, K. P. and Blue, J. A. (1998). A support vector machine approach to decision trees. In IEEE World Congress on Computational Intelligence, pages Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery, 2(2): Burges, C. J. C. (2003). Some notes on applied mathematics for machine learning. In Advanced Lectures on Learning, pages Chang, C.-C. and Lin, C.-J. (2001). LIBSVM: a library for support vector machines. Software available at Chiu, C.-Y. and Huang, Y.-T. (2007). Integration of support vector machine with naïve bayesian classifier for spam classification. In International Conference on Fuzzy Systems and Knowledge Discovery (FSKD), pages Guyon, I., Weston, J., Barnhill, S., and Vapnik, V. (2002). Gene selection for cancer classification using support vector machines. Learning, 46(1-3): Hsu, C.-W., Chang, C.-C., and Lin, C.-J. (2003). A practical guide to support vector classification. Technical report, Department of Computer Science, National Taiwan University. Hsu, C.-W. and Lin, C.-J. (2002). A comparison of methods for multiclass support vector machines. IEEE Transactions on s, 13(2):

51 References II Klien, D. (2004). Lagrange multipliers without permanent scarring. Available at Li, R., Wang, H.-N., He, H., Cui, Y.-M., and Du, Z.-L. (2007). Support vector machine combined with k-nearest neighbors for solar flare forecasting. Chinese Journal of Astronomy and Astrophysics, 7(3): Osuna, E. E., Freund, R., and Girosi, F. (1997). Support vector machines: Training and applications. Technical Report AIM-1602, Artificial Intelligence Laboratory, Massachusetts Institute of Technology. Platt, J. (2000). Probabilistic outputs for support vector machines and comparison to regularized likelihood methods. In Advances in Large Margin s, pages Tibshirani, R. and Hastie, T. (2007). Margin trees for high-dimensional classification. Journal of Learning Research (JMLR), 8: