MIT/ORC Spring seminar

Convex Optimization in Classification Problems

Laurent El Ghaoui

Department of EECS, UC Berkeley

[email protected]

1
 goal

• connection between classification and LP, convex QP has a long

history (Vapnik, Mangasarian, Bennett, etc)
• recent progresses in convex optimization: conic and semidefinite
programming; geometric programming; robust optimization
• we’ll outline some connections between convex optimization and
classification problems

joint work with: M. Jordan, N. Cristianini, G. Lanckriet,

C. Bhattacharrya

2
 outline

 convex optimization
• SVMs and robust linear programming
• minimax probability machine
• learning the kernel matrix

3
 convex optimization

standard form:

min f0 (x) : fi (x) ≤ 0, i = 1, . . . , m

x

• arises in many applications

• convexity not always recognized in practice
• can solve large classes of convex problems in polynomial-time
(Nesterov, Nemirovski, 1990)

4
 conic optimization

special class of convex problems:

min cT x : Ax = b, x ∈ K
x

where K is a cone, direct product of the following ”building blocks”:

K = Rn+ linear programming

K = {(y, t) ∈ Rn+1 : t ≥ kyk2 } second-order cone programming,
quadratic programming
K = {x ∈ Rn×n : x = xT  0} semidefinite programming

fact: can solve conic problems in polynomial-time

(Nesterov, Nemirovski, 1990)

5
 conic duality

dual of conic problem

min cT x : Ax = b, x ∈ K
x

is
max bT y : c − AT y ∈ K ∗
y

where
K ∗ = {z : hz, xi ≥ 0 ∀ x ∈ K}
is the cone dual to K

for the cones mentioned before, and direct products of them, K = K ∗

6
 robust optimization

conic problem in dual form: maxy bT y : c − AT y ∈ K

→ what if A is unknown-but-bounded, say A ∈ A, where A is given?

robust counterpart: maxy bT y : ∀A ∈ A, c − AT y ∈ K

• still convex, but tractability depends on A

• systematic ways to approximate (get lower bounds)
• for large classes of A, approximation is exact

7
 example: robust LP

linear program: minx cT x : aTi x ≤ b, i = 1, . . . , m

assume ai ’s are unknown-but-bounded in ellipsoids

T −1

Ei := a : (a − âi ) Γi (a − âi ) ≤ 1

where âi : center, Γi  0: ”shape matrix”

robust LP: minx cT x : ∀ ai ∈ Ei , aTi x ≤ b, i = 1, . . . , m

8
 robust LP: SOCP representation
robust LP equivalent to
1/2
min cT x : âTi x + kΓi xk2 ≤ b, i = 1, . . . , m
x

→ a second-order cone program!

interpretation: smoothes boundary of feasible set

9
 LP with Gaussian coefficients

assume a ∼ N (â, Γ), then for given x,

Prob{aT x ≤ b} ≥ 1 − 

is equivalent to:
âT x + κkΓ1/2 xk2 ≤ b
where κ = Φ−1 (1 − ) and Φ is the c.d.f. of N (0, 1)

hence,

• can solve LP with Gaussian coefficients using second-order

cone programming
• resulting SOCP is similar to one obtained with ellipsoidal
uncertainty

10
 LP with random coefficients

assume a ∼ (â, Γ), i.e. distribution of a has mean â and covariance

matrix Γ, but is otherwise unknown

Chebychev inequality:

Prob{aT x ≤ b} ≥ 1 − 

is equivalent to:
âT x + κkΓ1/2 xk2 ≤ b
where r
1−
κ=

leads to SOCP similar to ones obtained previously

11
 outline

• convex optimization
 SVMs and robust linear programming
• minimax probability machine
• kernel optimization

12
 SVMs: setup

given data points xi with labels yi = ±1, i = 1, . . . , N

two-class linear classification with support vector:

min kak2 : yi (aT xi − b) ≥ 1, i = 1, . . . , N

• problem is feasible iff there exists a separating hyperplane between

the two classes
• if so, amounts to select one separating hyperplane among the
many possible

13
 SVMs: robust optimization interpretation

interpretation: SVMs are a way to handle noise in data points

• assume each data point is unknown-but-bounded in a sphere of
radius ρ and center xi
• find the largest ρ such that separation is still possible between the
two classes of perturbed points

14
 variations

can use other data noise models:

• hypercube uncertainty (gives rise to LP)

• ellipsoidal uncertainty (→ QP)
• probabilistic uncertainty, Gaussian or Chebychev (→ QP)

15
 separation with hypercube uncertainty
assume each data point is unknown-but-bounded in an hypercube Ci :

xi ∈ Ci := {x̂i + ρP u : kuk∞ ≤ 1}

where centers x̂i and ”shape matrix” P are given

robust separation:

leads to linear program

min kP ak1 : yi (aT x̂i − b) ≥ 1, i = 1, . . . , N

16
 separation with ellipsoidal uncertainty

assume each data point is unknown-but-bounded in an ellipsoid Ei :

xi ∈ Ei := {x̂i + ρP u : kuk2 ≤ 1}

where center x̂i and ”shape matrix” P are given

robust separation leads to QP

min kP ak2 : yi (aT x̂i − b) ≥ 1, i = 1, . . . , N

17
 outline

• convex optimization
• SVMs and robust linear programming
 minimax probability machine
• kernel optimization

18
 minimax probability machine

goal:
• make assumptions about the data generating process
• do not assume Gaussian distributions
• use second-moment analysis of the two classes

let x̂± , Γ± be the mean and covariance matrix of class y = ±1

MPM: maximize  such that there exists (a, b) such that

inf Prob{aT x ≤ b} ≥ 1−

x∼(x̂+ ,Γ+ )

inf Prob{aT x ≥ b} ≥ 1−

x∼(x̂− ,Γ− )

19
 MPMs: optimization problem

→ two-sided, multivariable Chebychev inequality:

T 2
(b − a x̂) +
inf Prob{aT x ≤ b} =
x∼(x̂,Γ) (b − aT x̂)2+ + aT Γa

MPM problem leads to second-order cone program:

1/2 1/2
min kΓ+ ak2 + kΓ− ak2 : aT (x̂+ − x̂− ) = 1
a

complexity is the same as standard SVMs

20
 dual problem

express problem as unconstrained min-max problem:

1/2 1/2
min max uT Γ+ a − v T Γ− a + λ(1 − aT (x+ − x− ))
a kuk2 ≤1, kvk2 ≤1

exchange min and max, and set ρ := 1/λ:

1/2 1/2
min ρ : x+ + Γ+ u = x− + Γ− v, kuk2 ≤ ρ, kvk2 ≤ ρ
ρ,u,v

geometric interpretation: define the two ellipsoids

n o
1/2
E± (ρ) := x̂± + Γ± u : kuk2 ≤ ρ

and find largest ρ for which ellipsoids intersect

21
 robust optimization interpretation

assume data with label + generated arbitrarily in ellipsoid

n o
1/2
x+ ∈ E+ (ρ) := x̂+ + Γ+ u : kuk2 ≤ ρ

and similarly for data with label −

MPM finds largest ρ for which robust separation is possible

aT x − b = 0
x̂+
PSfrag replacements
x̂−

22
 experimental results

Linear kernel Gaussian kernel BPB

α TSA α TSA
Twonorm 80.2 % 95.2 % 86.2 % 95.8 % 96.3 %
Breast cancer 84.4 % 97.2 % 92.7 % 97.3 % 96.8 %
Ionosphere 63.3 % 84.7 % 94.9 % 93.4 % 93.7 %
Pima diabetes 31.2 % 73.8 % 33.0 % 74.6 % 76.1 %
Heart 42.8 % 81.4 % 50.5 % 83.6 % unknown

23
 variations

• minimize weighted sum of misclassification probabilities

• quadratic separation: find a quadratic set such that

inf Prob{x ∈ Q} ≥ 1 − 
x∼(x̂+ ,Γ+ )

inf Prob{x 6∈ Q} ≥ 1 − 
x∼(x̂− ,Γ− )

→ leads to a semidefinite programming problem

• nonlinear classification via kernels
(using plug-in estimates of mean and covariance matrix)

24
 outline

• convex optimization
• SVMs and robust linear programming
• minimax probability machine
 learning the kernel matrix

25
 transduction

the data contains both

• labeled points (training set)
• unlabeled points (test set)

transduction: given labeled training set and unlabeled test set,

predict the labels on the test set

26
 kernel methods

main goal: separate using a nonlinear classifier

aT φ(x) = b

where φ is a nonlinear operator

define the kernel matrix

Kij = φ(xi )T φ(xj )

(involves both labeled and unlabeled data)

fact: for transduction, all we need to know to predicts the labels is

the kernel matrix (and not φ(·) itself!)

27
 kernel methods: idea of proof
at the optimum, a is in the range of the labeled data:
X
a= λ i xi
i

=⇒ solution of classification problem depends only on the values of

kernel matrix Kij for labeled points xi , xj

in a transductive setting, the prediction of labels also involves Kij

only, since for an unlabeled data point xj ,
X
T
a φ(xj ) = λi φ(xi )T φ(xj )
i

involves only Kij ’s

fact: all previous algorithms can be ”kernelized”

28
 partition training/test

in a transductive setting, we can partition the kernel matrix as follows:

 
Ktr,tr Ktr,t
K= 
T
Ktr,t Kt,t

where subscripts tr and t stand for ”training” and ”test”, respectively

29
 kernel optimization

what is a ”good” kernel?

• margin: kernel ”behaves well” on the training data,

→ condition on the matrix Ktr,tr
• test error: kernel yields low predicted error
→ condition on the full matrix K
• also, to prevent overfitting, the blocks in K should be ”entangled”
→ will restrict the search space with affine constraint

30
kernel optimization and semidefinite programming

main idea: kernel can be described via the Gram matrix of data points,
hence is a positive semidefinite matrix

→ semidefinite programming plays a role in kernel optimization

31
 margin of SVM classifier

kernel-based SVM problem for labeled data points:

min kak2 subject to yi (aT φ(xi ) + b) ≥ 1, i = 1, . . . , N

a,b

(classifier depends only on training set block of kernel matrix Ktr )

margin of optimal classifier is γ = 1/ka∗ k2

geometrically:

γ −1 = distance between the convex hulls of the two classes

(can work with ”soft” margin when data is not linearly separable)

32
 generalization error

how well the SVM classifier will work on the test set?
from learning theory (Bartlett, Rademacher), generalization error is
bounded above by
√
Tr K
γ(Ktr )
where γ(Ktr ) is the margin of the SVM classifier with training set
block kernel matrix Ktr
hence, the constraints

Tr K = c, γ(Ktr )−1 ≤ w

ensure an upper bound on the generalization error

33
 margin constraint

using a dual expression for the SVM problem, margin constraint

γ(Ktr ) ≥ γ

writes as LMI (linear matrix inequality) in K

 
G(Ktr ) e+ν+λ·y
 0
(e + ν + λ · y)T γ −1

where
• G(Ktr ) is linear in Ktr
• e = vector of ones
• λ, ν are new variables

34
 avoiding overfitting

the trace constraint Tr K = c is not enough to ”entangle” the matrix

K
we impose an affine constraint on K of the form
X
K= µi Ki
i

where Ki ’s correspond to given different, known kernels and µi ’s will

be our new variables

35
 optimizing kernels: example problem

goal: find a kernel matrix that

• is positive semidefinite and has a given trace

K  0, Tr K = c

• belongs to an affine space (here Ki ’s are known)

X
K= µi Ki
i

• satisfies a lower bound γ on the margin on the training set, γ(Ktr )

the problem reduces to a semidefinite programming feasibility problem

36
 experimental results

K1 K2 K3 K∗
Breast cancer d=2 σ = 0.5
margin 0.010 0.136 - 0.300
TSE 19.7 28.8 11.4
Sonar d=2 σ = 0.1
margin 0.035 0.198 0.006 0.352
TSE 15.5 19.4 21.9 13.8
Heart d=2 σ = 0.5
margin - 0.159 - 0.285
TSE 49.2 36.6

37
 wrap-up

• convex optimization has much to offer and gain from interaction

with classification
• described variations on linear classification
• many robust optimization interpretations
• all these methods can be kernelized
• kernel optimization has high potential

38
 see also

• Learning the kernel matrix with semidefinite programming

(Lanckiert, Cristianini, Bartlett, El Ghaoui, Jordan, submitted to
ICML 2002)
• Minimax probability machine
(Lanckiert, Bhattacharrya, El Ghaoui, Jordan) (NIPS 2001)

39