Support Vector Machines for Classification and …

1 ISIS Technical Report Support Vector Machines for Classification and regression Steve Gunn 14 May 1998. Contents 1 Introduction 5. Statistical Learning Theory ..6. VC .. 7. Structural Risk Minimisation .. 7. 2 Support Vector Classification 9. The Optimal Separating Hyperplane ..9. Linearly Separable Example .. 13. The Generalised Optimal Separating Hyperplane ..14. Linearly Non-Separable Example .. 16. Generalisation in High Dimensional Feature Space ..17. Polynomial Mapping 19. 3 Feature Space 21. Kernel Polynomial .. 21. Gaussian Radial Basis 21.

2 Exponential Radial Basis Function .. 22. Multi-Layer 22. Fourier Series .. 22. 22. B splines .. 23. Additive Kernels .. 23. Tensor Product 23. Implicit vs. Explicit Bias ..23. Data Normalisation ..24. Kernel Selection ..24. 4 Classification Example: IRIS data 25. Applications ..28. 5 Support Vector regression 31. Linear -Insensitive Loss Function .. 32. Quadratic Loss Function .. 33. Huber Loss Function .. 33. Example .. 34. Non Linear .. 35. Comments .. 38. 6 regression Example: Titanium Data 39. Applications ..42. Image Speech and Intelligent Systems Group 7 Conclusions 43.

3 References 45. Appendix - Implementation Issues 47. Support Vector Support Vector MATLAB SVM Toolbox ..52. Image Speech and Intelligent Systems Group 1 Introduction The problem of empirical data modelling is germane to many engineering applications. In empirical data modelling a process of induction is used to build up a model of the system, from which it is hoped to deduce responses of the system that have yet to be observed. Ultimately the quantity and quality of the observations govern the performance of this empirical model.

4 By its observational nature data obtained is finite and sampled; typically this sampling is non-uniform and due to the high dimensional nature of the problem the data will form only a sparse distribution in the input space. Consequently the problem is nearly always ill posed [16] in the sense of Hadamard [9]. Traditional neural network approaches have suffered difficulties with generalisation, producing models that can overfit the data. This is a consequence of the optimisation algorithms used for parameter selection and the statistical measures used to select the best' model.

5 The foundations of Support Vector Machines (SVM) have been developed by Vapnik [21] and are gaining popularity due to many attractive features, and promising empirical performance. The formulation embodies the Structural Risk Minimisation (SRM) principle, which has been shown to be superior, [8], to traditional Empirical Risk Minimisation (ERM) principle, employed by conventional neural networks. SRM minimises an upper bound on the VC dimension ('generalisation error'), as opposed to ERM that minimises the error on the training data.

6 It is this difference which equips SVM with a greater ability to generalise, which is the goal in statistical learning. SVM were developed to solve the Classification problem, but recently they have been extended to the domain of regression problems [20]. In the literature the terminology for SVM can be slightly confusing. The term SVM is typically used to describe Classification with Support Vector methods and Support Vector regression is used to describe regression with Support Vector methods. In this report the term SVM will refer to both Classification and regression methods, and the terms Support Vector Classification (SVC) and Support Vector regression (SVR).

7 Will be used for specification. This section continues with an introduction to the structural risk minimisation principle. In section 2 the SVM is introduced in the setting of Classification , being both historical and more accessible. This leads onto mapping the input into a higher dimensional feature space by a suitable choice of kernel function. The report then considers the problem of regression . Illustrative examples are given to show the properties of the techniques. Image Speech and Intelligent Systems Group Statistical Learning Theory Generalisation error = Estimation error + Approximation error Ultimately we would like to find the function, f, which minimises the risk, R[ f ] = ( y f (x)) P(x, y )dxdy.

8 2. (1). X Y. However, P(x,y) is unknown. It is possible to find an approximation according to the empirical risk minimisation principle, 1 l (. Remp [ f ] = y f (x i ) ). 2. (2). l i =1. which minimises the empirical risk, f$n ,l (x) = arg min Remp [ f ] (3). f Hn Empirical risk minimisation makes sense only if, lim Remp [ f ] = R[ f ]. l . which is true from the law of large numbers. However, it must also satisfy, lim min Remp [ f ] = min R[ f ]. l f Hn f Hn which is only valid when Hn is 'small' enough. This condition is less intuitive and requires that the minima also converge.

9 The following bound holds with probability 1- , Image Speech and Intelligent Systems Group h(ln 2hl + 1) ln( 4 ). R[ f ] Remp [ f ] +. l VC Dimension The VC dimension is a scalar value that measures the capacity of a set of functions. Figure 1 VC Dimension Illustration The set of linear indicator functions, Equation has a VC dimension equal to n+1. The VC dimension is defined as, There exists a set of points xn such that these points can be separated in all 2^n possible configurations, and that no set xm exists where m>n satisfying this property.

10 Figure 1illustrates how three points in the plane can be shattered by the set of linear indicator functions whereas four points cannot. In this case the VC dimension is equal to the number of free parameters, but in general that is not the case. the function Asin(bx) has an infinite VC dimension [ref]. SRM principle creates a structure Structural Risk Minimisation Create a structure such that S h is a hypothesis space of VC dimension h, S1 S2 K S1 K. SRM consists in solving the following problem h(ln 2hl + 1) ln(4 ) . min Remp [ f ] +.