Short Course Robust Optimization and Machine Learning. Lecture 6: Robust Optimization in Machine Learning

Transcription

1 Short Course Robust Optimization and Machine Machine Lecture 6: Robust Optimization in Machine Laurent El Ghaoui EECS and IEOR Departments UC Berkeley Spring seminar TRANSP-OR, Zinal, Jan , 2012

2 Outline Machine

3 Outline Machine

4 Supervised learning problems Many supervised learning problems (e.g., classification, regression) can be written as min w L(X T w) where L is convex, and X contains the data. Machine

5 Penalty approach Often, optimal value and solutions of optimization problems are sensitive to data. A common approach to deal with sensitivity is via penalization, e.g.: min x L(X T w) + Wx 2 2 (W = weighting matrix). How do we choose the penalty? Can we choose it in a way that reflects knowledge about problem structure, or how uncertainty affects data? Does it lead to better solutions from machine learning viewpoint? Machine

6 Support Vector Machine Support Vector Machine (SVM) classification problem: min w,b m (1 y i (zi T w + b)) + i=1 Z := [z 1,..., z m] R n m contains the data points. y { 1, 1} m contain the labels. x := (w, b) contains the classifier parameters, allowing to classify a new point z via the rule y = sgn(z T w + b). Machine

7 Robustness to data uncertainty Assume the data matrix is only partially known, and address the robust optimization problem: min w,b max U U m (1 y i ((z i + u i ) T w + b)) +, i=1 where U = [u 1,..., u m] and U R n m is a set that describes additive uncertainty in the data matrix. Machine

8 Measurement-wise, spherical uncertainty Assume Machine where ρ > 0 is given. U = {U = [u 1,..., u m] R n m : u i 2 ρ}, Robust SVM reduces to min w,b m (1 y i (zi T w + b) + ρ w 2 ) +. i=1

9 Link with classical SVM Classical SVM contains l 2 -norm regularization term: min w,b m (1 y i (zi T w + b)) + + λ w 2 2. i=1 where λ > 0 is a penalty parameter. With spherical uncertainty, robust SVM is similar to classical SVM. Machine When data is separable, the two models are equivalent...

10 Separable data Machine Maximally robust classifier for separable data, with spherical uncertainties around each data point. In this case, the robust counterpart reduces to the classical maximum-margin classifier problem.

11 Interval uncertainty Assume U = {U R n m where ρ > 0 is given. : (i, j), U ij ρ}, Machine Robust SVM reduces to min w,b m (1 y i (zi T w + b) + ρ w 1 ) +. i=1 The l 1 -norm term encourages sparsity, and may not regularize the solution.

12 Separable data Machine Maximally robust classifier for separable data, with box uncertainties around each data point. This uncertainty model encourages sparsity of the solution.

13 Other uncertainty models We may generalize the approach to other uncertainty models, retaining tractability: Measurement-wise uncertainty models: perturbations affect each data point independent of each other. Other models couple the way uncertainties affect each measurement; for example we may control the number of errors across all the measurements. Norm-bound models allow for uncertainty of data matrix that is bounded in matrix norm. A whole theory is presented in [1]. Machine

14 Consider standard l 1 -penalized SVM: φ λ (X) := min w,b Constrained counterpart: ψ c(x) := min w,b 1 m m (1 y i (w T x i + b)) + + λ w 1 i=1 m (1 y i (xi T w + b)) + : w 1 c i=1 Machine Basic goal: solve these problems in the large-scale case. Approach: use to sparsify the data matrix in a controlled way.

15 Thresholding data We threshold the data using an absolute level t: { 0 if xi,j t (x i (t)) j := 1 otherwise This will make the data sparser, resulting in memory and time savings. Machine

16 Handling thresholding errors Handle thresholding errors via robust counterpart: (w(t), b(t)) := arg min w,b Above problem is tractable. max Z X t m (1 y i (w T z i + b)) + + λ w 1. i=1 Machine The solution w(t) at threshold level t satisfies 0 1 m m (1 y i (x T i w(t) + b(t))) + + λ w(t) 1 φ λ (X) 2t λ. i=1

17 Results 20 news groups data set Machine Dataset size: 20, , 000. Thresholding of data matrix of TF-IDF scores.

18 Results UCI NYTimes Dataset 1 stock 11 bond 2 nasdaq 12 forecast 3 portfolio 13 thomson financial 4 brokerage 14 index 5 exchanges 15 royal bank 6 shareholder 16 fund 7 fund 17 marketing 8 investor 18 companies 9 alan greenspan 19 bank 10 fed 20 merrill Top 20 keywords for topic stock. Dataset size: 100, , 660, 30,000,000 non-zeros. Thresholded dataset (by TF-IDF scores) with level ,000 non-zeroes (2.8 %). Total run time: 4317s. Machine

19 Robust SVM with Data: boolean Z {0, 1} n m (eg, co-occurence matrix) Nominal problem: SVM min w,b m (1 y i (zi T w + b)) +, i=1 Uncertainty model: assume each data value can be flipped, total budget of flips is constrained: { } U = U = [u 1,..., u m] R l m : u i { 1, 0, 1} l, u 1 k. Machine

20 Robust counterpart where min w,b m (1 y i (zi T w + b) + φ(w)) +, i=1 φ(w) := min s k w s + s 1 Penalty is a combination of l 1, l norms. Problem is tractable (doubles number of variables over nominal). Still needs regularization. Machine

21 Results UCI Internet advertisement data set Machine Dataset size: k = 0 corresponds to nominal SVM problem. Best performance at k = 3.

22 Refined model We can impose u i {0, 1 2x i }. This leads to a new penalty: with min w,b m (1 y i (xi T w + b) + φ i (w)) +, i=1 n φ i (w) := min kµ + (y i w j (2x ij 1) µ) + µ 0 Problem can still be solved via LP. j=1 Machine

23 Results UCI Heart data set Machine Dataset size: k = 0 corresponds to nominal SVM problem. Best performance at k = 1.

24 Outline Machine

25 Nominal problem where min L(Z T θ), θ Θ Z := [z 1,..., z m] R n m is the data matrix L : R m R is a convex loss function Θ imposes structure (eg, sign) constraints on parameter vector θ Machine

26 Loss function: assumptions We assume that L(r) = π(abs(p(r))), where abs( ) acts componentwise, π : R m + R is a convex, monotone function on the non-negative orthant, and { r ( symmetric case ) P(r) = ( asymmetric case ) r + with r + the vector with components max(r i, 0), i = 1,..., m. Machine

27 Loss function: examples l p-norm regression hinge loss Huber, Berhu loss Machine

28 Robust counterpart min max θ Θ Z Z L(ZT θ). where Z R n m is a set of the form Z = {Z + : ρd, }, with ρ 0 a measure of the size of the uncertainty, and D R l m is given. Machine

29 Generic analysis For a given vector θ, we have max Z Z L(ZT θ) = max u where L is the conjugate of L, and is the support function of D. u T Z T θ L (u) + ρφ D(uv T ), φ D(X) := max X, D Machine

30 Assumptions on uncertainty set D Separability condition: there exist two semi-norms φ, ψ such that φ D(uv T ) := max D ut v = φ(u)ψ(v). Does not completely characterize (the support function of) the set D Given φ, ψ, we can construct a set D out that obeys condition The robust counterpart only depends on φ, ψ. WLOG, we can replace D by its convex hull. Machine

31 Largest singular value model: D = { : ρ}, with φ, ψ Euclidean norms. Any norm-bound model involving an induced norm (φ, ψ are then the norms dual to the norms involved). Measurement-wise uncertainty models, where each column of the perturbation matrix is bounded in norm, independently of the others, correspond to the case with ψ(v) = v 1. Machine

32 Other examples Bounded-error model: there are (at most K ) errors affecting data Machine D = = [λ 1 δ 1,..., λ mδ m] R l m : δ i 1, i = 1,..., m,. m λ i K, λ {0, 1} m for which φ( ) =, ψ(v) = sum of the K largest magnitudes of the components of v. i=1

33 (follow d) The set { D = = [λ 1 δ 1,..., λ mδ m] R l m } : δ i { 1, 0, 1} l, δ 1 k models measurement-wise uncertainty affecting (we can impose δ i {x i 1, x i } to be more realistic) In this case, we have ψ( ) = 1 and φ(u) = u 1,k := min w k u w + w 1. Machine

34 Main result For a given vector θ, we have where min θ max Z Z L(ZT θ) = min θ,κ Lwc(Z T θ, κ) : κ φ(u T θ) L(r, κ) := max v v T r L (v) + κψ(v) is the worst-case loss function of the robust problem. Machine

35 Worst-case loss function The tractability of the robust counterpart is directly linked to our ability to compute optimal solutions v for L(r, κ) = max v v T r L (v) + κψ(v) Dual representation (assume ψ( ) = is a norm): L(r, κ) = max ξ L(r + κξ) : ξ 1 When ψ is the Euclidean norm, robust regularization of L (Lewis, 2001). Machine

36 When ψ( ) = p, p = 1,, problem reduces to simple, tractable convex problem (assuming nominal problem is). For p = 2, problem can be reduced to such a simple form, for the hinge, l q-norm and Huber loss functions. Machine

37 Lasso In particular, the least-squares problem with lasso penalty min θ X T θ y 2 + ρ θ 1 is the robust counterpart to a least-squares problem with uncertainty on X, with additive perturbation bounded in the norm n 1,2 := max 2 ij. 1 i l j=1 Machine

38 Globalized robust counterpart The robust counterpart is based on the worst-case value of the loss function assuming a bound on the data uncertainty (Z Z): min max θ Θ Z Z L(ZT θ). The approach does not control the degradation of the loss outside the set Z. Machine

39 Globalized robust counterpart: formulation In globalized robust counterpart, we fix a rate of degradation of the loss, which controls the amount of degradation of the loss as the data matrix Z goes away from the set Z. We seek to minimize τ, such that : L((Z + ) T θ) τ + α, where α > 0 controls the rate of degradation, and is a matrix norm. Machine

40 Globalized robust counterpart For the SVM case, the globalized robust counterpart can be expressed as: min w,b m (1 y i (zi T w + b)) + : m θ 2 α, i=1 which is a classical form of SVM. For l p-norm regression with m data points, the globalized robust counterpart takes the form min X T θ y p : κ(m, p) θ 2 α θ Machine where κ(m, 1) = m, κ(m, 2) = κ(m, ) = 1.

41 can address problems with chance constraints min θ max EpL(Z p P (δ)t θ) where δ follows distribution p, and P is a class of distributions Results are more limited, focused on upper bounds. Convex relaxations are available, but more expensive. Approach uses Bernstein approximations (Nemirovski & Ben-tal, 2006). Machine

42 Robust regression with chance constraints: an example φ p := min θ max x (ˆx,X) Regression variable is θ R n E x A(x)θ b(x) p x R q is an uncertainty vector that enters affinely in the problem matrices: [A(x), b(x)] = [A 0, b 0 ] + i x i[a i, b i ]. The distribution of uncertainty vector x is unknown, except for its mean ˆx and covariance X. Objective is worst-case (over distributions) expected value of l p-norm residual (p = 1, 2). Machine

43 Main result (Assume ˆx = 0, X = I WLOG) For p = 2, the problem reduces to least-squares: Machine φ 2 2 = min θ q A i θ b i 2 2 i=0 For p = 1, we have (2/π)ψ 1 φ 1 ψ 1, with ψ 1 = min θ q A i θ b i 2 i=0

44 Example: robust median As a special case, consider the median problem: min θ q θ x i i=1 Now assume that vector x is random, with mean ˆx and covariance X, and consider the robust version: φ 1 := min θ max E x (ˆx,X) x q θ x i i=1 Machine

45 Approximate solution We have (2/π)ψ 1 φ 1 ψ 1, with ψ 1 := n i=1 (θ ˆx i ) 2 + X ii Amounts to find the minimum distance sum (a very simple SOCP). Machine

46 Geometry of robust median problem std Nominal vs. robust median nominal points std dev. median robust median Machine x

47 Outline Machine

48 Machine A. Bental, L. El Ghaoui, and A. Nemirovski. Robust Optimization. Princeton Series in Applied Mathematics. Princeton University Press, October C. Caramanis, S. Mannor, and H. Xu. Robust optimization in machine learning. In S. Sra, S. Nowozin, and S. Wright, editors, Optimization for Machine, chapter 14. MIT Press, 2011.