Shattering (machine learning)

The concept of shattering of a set of points plays an important role in Vapnik–Chervonenkis theory, also known as VC-theory. Shattering and VC-theory are used in the study of empirical processes as well as in statistical computational learning theory.

Contents

Definition

Suppose we have a class C of sets and a given set A. C is said to shatter A if, for each subset T of A, there is some element U of C such that

U \cap A = T.\,

Equivalently, C shatters A when the power set P(A) is the set { U ∩ A | U ∈ C }.

For example, the class C of all discs in the plane (two-dimensional space) cannot shatter every set A of four points, yet the class of all convex sets in the plane shatters every finite set on the (unit) circle. (For the collection of all convex sets, connect the dots!)

We employ the letter C to refer to a "class" or "collection" of sets, as in a Vapnik–Chervonenkis class (VC-class). The set A is often assumed to be finite because, in empirical processes, we are interested in the shattering of finite sets of data points.

Example

Suppose we have a set A of 4 points on the unit circle, and we wish to know if it is shattered by the class C of all discs.

To test this, we attempt to draw a disc around every subset of points in A. First, we draw a disc around the subsets of each isolated point, which turns out to be easy. Next, we try to draw a disc around every subset of point pairs. This turns out to be easy for adjacent points, but impossible for points on opposite sides of the circle. As visualized below:

Because there is some subset which can not be isolated by any disc in C, we conclude then that A is not shattered by C. And, with a bit of thought, we can prove that NO set of 4 points is shattered by this C.

However, if we redefine C to be the class of all elliptical discs, we find that we can still isolate all the subsets from above, as well as the points which were problems. Thus, this specific set of 4 points is shattered by the class of ellipitical discs. Visualized below:

With a bit of thought, we could generalize that any set of finite points on a unit circle could be shattered by the class of all convex sets (visualize connecting the dots).

Shatter coefficient

To quantify the richness of a collection C of sets, we use the concept of shattering coefficients (also known as shatter coefficients or the growth function). For a collection C of sets s⊂Ω, Ω being any space, often a probability space, we define the nth shattering coefficient of C as

S _C(n) = \max_{x_1,x_2,\dots,x_n \in \Omega } \operatorname{card} \{\,\{\,x_1,x_2,\dots,x_n\}\cap s, s\in C \}

where \operatorname{card} denotes the cardinality of the set.

S _C(n) is equal to the largest number of subsets of any set A of n points that can be formed by intersecting A with the sets in collection C.

Here are some facts about S_C(n):

1.S_C(n)\leq 2^n for all n because \{s\cap A|s\in C\}\subseteq P(A) for any A\subseteq \Omega.
2. If S_C(n)=2^n, that means there is a set of cardinality n, which can be shattered by C.
3. If S_C(N)<2^N for some N>1 then S_C(n)<2^n for all n\geq N.

The third property means that if C cannot shatter any set of cardinality N then it can not shatter sets of larger cardinalities.

Vapnik–Chervonenkis class

The VC dimension of a class C is defined as

VC(C)=\underset{n}{\min}\{n:S_C(n)<2^n\}\,

or, alternatively, as

VC_0(C)=\underset{n}{\max}\{n:S_C(n)=2^n\}.\,

Note that VC(C)=VC_0(C)%2B1.

If for any n there is a set of cardinality n which can be shattered by C, then S_C(n)=2^n for all n and the VC dimension of this class C is infinite.

A class with finite VC dimension is called a Vapnik–Chervonenkis class or VC class. A class C is uniformly Glivenko–Cantelli if and only if it is a VC class.

References