Title: Three Algorithms for Nonlinear Dimensionality Reduction
1Three Algorithms for Nonlinear Dimensionality
Reduction
- Haixuan YangGroup Meeting
- Jan. 011, 2005
2Outline
- Problem
- PCA (Principal Component Analysis)
- MDS (Multidimentional Scaling)
- Isomap (isometric mapping)
- A Global Geometric Framework for Nonlinear
Dimensionality Reduction. Science, 292(22),
2319-2323, 2000. - LLE (locally linear embedding)
- Nonlinear Dimensionality Reduction by Locally
Linear Embedding. Science, 292(22), 2323-2326,
2000. - Eigenmap
- Laplacian Eigenmaps and Spectral Techniques for
Embedding and Clustering. NIPS01.
3Problem
- Given a set x1, , xk of k points in Rl, find a
set of - points y1, , yk in Rm (m ltlt l) such
that yi represents xi as accurately as
possible. - If the data xi is placed in a super plane in
high dimensional space, the traditional
algorithms, such as PCA and MDS, work well. - However, when the data xi is placed in a
nonlinear manifold in high dimensional space,
then the linear algebra technique can not work
any more. - A nonlinear manifold can be roughly understood
as a distorted super plane, which may be twisted,
folded, or curved.
4PCA (Principal Component Analysis)
- Reduce dimensionality of data by transforming
correlated variables (bands) into a smaller
number of uncorrelated components - Reveals meaningful latent information
- Best preserves the variance as measured in the
high-dimensional input space. - Nonlinear structure is invisible to PCA
5First, a graphical look at the problem
Band 2
Two (correlated) Bands of data
Band 1
6Regression LineSummarizes the Two Bands
Band 2
Band 1
7Rotate axes to create two orthogonal
(uncorrelated) components
PC1
Band 2
PC2
Reflected X- and y-axes
Band 1
8Partitioning of Variance
PC1
Var(PC1)
Band 2
Var(PC2)
PC2
Band 1
9PCA algorithm description
- Step 1 Calculate the average x of xi .
- Step 2 Estimate the Covariance Matrix by
- Step 3 Let ?p be the p-th eigenvalue (in
decreasing order) of the matrix M, and vpi be the
i-th component of the p-th eignvector. Then set
the p-th componet of the d-dimentional coordinate
vector yi equal to
10MDS
- Step 1 Given the distance d(i, j) between i and
j. - Step 2 From d(i, j), get the covariance matrix
M by
11An example of embedding of a two dimentional
manifold into a three dimentional space
Not the true distance
The true distance
12Isomap basic idea
- Learn the global distance by the local distance.
- The local distance calculated by the Euclidean
distance is relatively accurate because a patch
in the nonlinear manifold looks like a plane when
it is small, and therefore the direct Euclidean
distance approximates the true distance in this
small patch. - The global distance calculated by the Euclidean
distance is not accurate because the manifold is
curved. - Best preserve the estimated distance in the
embedded space in the same way as MDS.
13Isomap algorithm description
- Step 1 Construct neighborhood graph
- Define the graph over all data points by
connecting points i and j if they are closer than
e (e-Isomap), or if i is one of the n nearest
neighbors of j (k-Isomap). Set edge lengths equal
to dX(i,j). - Step 2 Compute shortest paths
- Initialize dG(i,j) dX(i,j) if i and j are
linked by an edge dG(i,j) 8 - otherwise. Then compute the shortest path
distances dG(i,j) between all - pairs of points in weighted graph G. Let
DG( dG(i,j) ). - Step 3 Construct d-dimensional embedding
- Let ?p be the p-th eigenvalue (in decreasing
order) of the matrix t(DG), and vpi be the i-th
component of the p-th eignvector. Then set the
p-th componet of the d-dimentional coordinate
vector yi equal to .
14An example each picture, a 4096
(6464)-dimensional point, can be mapped into
2-dinesional plane
15Another example the 3-dimentional points are
maped into 2-dimentional plane
16LLE basic idea
- Learn the local linear relation by the local data
- The local data is relatively linear because a
patch in the nonlinear manifold looks like a
plane when it is small. - Globally the data is not linear because the
manifold is curved. - Best preserve the local linear relation in the
embedded space in the similar way as PCA.
17LLE algorithm description
- Step 1 Discovering the Adjacency Information
- For each xi find its n nearest neighbors,
. - Step 2 Constrcting the Approximation Matrix
- Choose Wij by minimizing
- Under the condition that
- Step 3 Compute the Embedding
- The embedding vectors yi can be found by
minimizing
18An example 4096-dimentional face pictures are
embedded into a 2-dimentional plane
19Eigenmap Basic Idea
- Use the local information to decide the embedded
data. - Motivated by the way that heat transmits from one
point to another point.
20Eigenmap
- Step 1 Construct neighborhood graph
- The same as Isomap.
- Step 2 Compute the weights of the graph
- If node i and node j are connected, put
- Step 3 Construct d-dimensional embedding
- Compute the eigenvalues and eigenvectors for
the generalized eigenvector problem
, where D is a diagonal matrix, and
21Cont.
- Let f0,,fk-1 be the solutions of the above
equation, - ordered increasingly according to their
eignvalues, - Lf0?0Df0
- Lf1?1Df1
-
- Lfk-1?k-1Dfk-1
- Then yi is determined by the ith component of the
d - eigenvectors f1,,fd .
-
22An example 256-dimentional speech data is
represented in a 2-dimentional plane
23Conclusion
- Isomap, LLE and Eigenmap can find the
meaningful low-dimensional structure hidden in
the high-dimensional observation. - These three algorithms work well especially in
the nonlinear manifold. In such a case, the
linear methods such as PCA and MDS can not work.