Efficient Regression in Metric Spaces via Approximate Lipschitz Extension

Efficient Regression in Metric Spaces

via Approximate Lipschitz ExtensionLee-Ad Gottlieb Ariel University

Aryeh Kontorovich Ben-Gurion UniversityRobert Krauthgamer Weizmann Institute

2

Regression A fundamental problem in Machine Learning:

Metric space (X,d) Probability distribution P on X [-1,1] Sample S of n points (Xi,Yi) drawn iid ~P

1-1

1

0

0

-1

1

3

Regression A fundamental problem in Machine Learning:

Metric space (X,d) Probability distribution P on X [-1,1] Sample S of n points (Xi,Yi) drawn iid ~P

Produce: Hypothesis h: X → [-1,1] empirical risk:

expected risk: q={1,2}

Goal: uniformly over h in probability, And have small Rn(h) h can be evaluated efficiently on new points

1-1

1

0

?

4

A popular solution For Euclidean space:

Kernel regression (Nadaraya-Watson) For vector v, let Kn(v) = e-(||v||/)2

Hypothesis evaluation on new x

1-1

1

0

?

5

Kernel regression Kernel Regression

Pros Achieves minimax rate (for Euclidean with Gaussian noise) Other algorithms: SVR, Spline regression

Cons: Evaluation for new point: linear in sample size Assumes Euclidean space: What about metric space?

6

Metric space (X,d) is a metric space if

X = set of points d = distance function

Nonnegative: d(x,y) ≥ 0 Symmetric: d(x,y) = d(y,x) Triangle inequality: d(x,y) ≤ d(x,z) + d(z,y)

Inner product ⇒ norm Norm ⇒ metric d(x,y) := ||x-y|| Other direction does not hold

Regression for metric data? Advantage: often much more natural

much weaker assumption Strings - edit distance (DNA) Images - earthmover distance

Problem: no vector representation No notion of dot-product (and no kernel) Invent kernel? Possible √logn distortion

7

AACGTA

AGTT

8

Metric regression Goal: Give class of hypotheses which generalize well

Perform well on new points

Generalization: Want h with Rn(h): empirical error R(h): expected error

What types of hypotheses generalize well? Complexity: VC, Fat-shattering dimensions

VC dimension Generalization: Want

Rn(h): empirical error R(h): expected error

How do we upper bound the expected error? Use a generalization bound. Roughly speaking (and whp)

expected error ≤ empirical error + (complexity of h)/n

More complex classifier ↔ “easier” to fit to arbitrary {-1,1} data

Example 1: VC dimension complexity of the hypothesis class VC-dimension: largest point

set that can be shattered by h -1

+1

+1

9

-1

Fat-shattering dimension Generalization: Want

Rn(h): empirical error R(h): expected error

How do we upper bound the expected error? Use a generalization bound. Roughly speaking (and whp)

expected error ≤ empirical error + (complexity of h)/n

More complex classifier ↔ “easier” to fit to arbitrary {-1,1} data

Example 2: Fat-shattering dimension of the hypothesis class Largest point set that can be

shattered with min distance from h+1

10

-1

Efficient classification for metric data 11

Generalization Conclustion: Simple hypotheses generalize well

In particular, those with low Fat-Shattering dimension

Can we find a hypothesis class For metric space Low Fat-shattering dimension?

Preliminaries: Lipschitz constant, extension Doubling dimension

+1

-1

12

Preliminaries: Lipschitz constant The Lipschitz constant of function f: X →

the smallest value L satisfying xi,xj in X

Denoted by (small smooth)

+1

-1≥ 2/L

13

Preliminaries: Lipschitz extension Lipschitz extension:

Given a function f: S → for S⊂ X with constant L Extend f to all of X without increasing the Lipschitz constant Classic problem in Analysis

Possible solution

Example: Points on the real line f(1) = 1 f(-1) = -1

picture credit: A. Oberman

14

Doubling Dimension Definition: Ball B(x,r) = all points within distance r>0 from x.

The doubling constant (of X) is the minimum value >0 such that every ball can be covered by balls of half the radius First used by [Ass-83], algorithmically by [Cla-97]. The doubling dimension is ddim(X)=log2(X) [GKL-03] Euclidean: ddim(Rn) = O(n)

Packing property of doubling spaces A set with diameter D>0 and min. inter-point

distance a>0, contains at most (D/a)O(ddim) points

Here ≥7.

Applications of doubling dimension Major application

approximate nearest neighbor search in time 2O(ddim) log n

Database/network structures and tasks analyzed via the doubling dimension Nearest neighbor search structure [KL ‘04, HM ’06, BKL ’06, CG ‘06] Spanner construction [GGN ‘06, CG ’06, DPP ‘06, GR ‘08a, GR ‘08b] Distance oracles [Tal ’04, Sli ’05, HM ’06, BGRKL ‘11] Clustering [Tal ‘04, ABS ‘08, FM ‘10] Routing [KSW ‘04, Sli ‘05, AGGM ‘06, KRXY ‘07, KRX ‘08]

Further applications Travelling Salesperson [Tal ’04, BGK ‘12] Embeddings [Ass ‘84, ABN ‘08, BRS ‘07, GK ‘11] Machine learning [BLL ‘09, GKK ‘10 ‘13a ‘13b]

Message: This is an active line of research… Note: Above algorithms can be extended to nearly-doubling spaces [GK ‘10]

15

q

G

2

11

H

2

11

1

16

Generalization bounds We provide generalization bounds for

Lipschitz (smooth) functions on spaces with low doubling dimension [vLB ‘04] provided similar bounds using covering numbers and

Rademacher averages

Fat-shattering analysis: L-Lipschitz functions shatter a set → inter-point distance is at

least 2/L Packing property → set has (diam L)O(ddim) points

Done! This is the Fat-shattering dimension of the smooth classifier on doubling spaces

17

Generalization bounds Plugging in Fat-Shattering dimension into known bounds, we

derive key result:

Theorem: Fix ε>0 and q = {1,2}. Let h be a L-Lipschitz hypothesis

[P(R(h)) > Rn(h) + ε] ≤ 24n (288n/ε2)d log(24en/ε) e-ε2n/36

Where d ≈ (1+1/(ε/24)(q+1)/2) (L/(ε/24)(q+1)/2)ddim

Upshot: Smooth classifier provably good for doubling spaces

18

Generalization bounds Alternate formulation:

d With probability at least 1- where

Trade-off Bias-term Rn decreasing in L Variance-term (n,L,) increasing in L

Goal: Find L which minimizes RHS

19

Generalization bounds Previous discussion motivates following hypothesis on sample

linear (q=1) or quadratic (q=2) program computes Rn(h)

Optimize L for best bias-variance tradeoff Binary search gives log(n/) “guesses” for L

For new points Want f* to stay smooth: Lipschitz extension

20

Generalization bounds To calculate hypothesis, can solve convex (or linear) program

Final problem: how to solve this program quickly

21

Generalization bounds To calculate hypothesis, can solve convex (or linear) program

Problem: O(n2) constraints! Exact solution is costly

Solution: (1+)-stretch spanner Replace full graph by sparse graph Degree -O(ddim)

solution f* perturbed by additive error Size: number of constraints reduced to -O(ddim)n Sparsity: variable appears in -O(ddim) constraints

G

2

11

H

2

11

1

22

Generalization bounds To calculate hypothesis, can solve convex (or linear) program

Efficient approximate LP solution Young [FOCS’ 01] approximately solves LP with sparse constraints our total runtime: O(-O(ddim) n log3 n)

Reduce QP to LP solution suffers additional 2 perturbation O(1/) new constraints

23

Thank you! Questions?