Dimensionality Reduction

1
Dimensionality Reduction

• CS 685 Special Topics in Data Mining
• Spring 2009
• Jinze Liu

2
Overview

• What is Dimensionality Reduction?
• Simplifying complex data
• Using dimensionality reduction as a Data Mining
  tool
• Useful for both data modeling and data
  analysis
• Tool for clustering and regression
• Linear Dimensionality Reduction Methods
• Principle Component Analysis (PCA)
• Multi-Dimensional Scaling (MDS)
• Non-Linear Dimensionality Reduction

3
What is Dimensionality Reduction?

• Given N objects, each with M measurements, find
  the best D-dimensional parameterization
• Goal Find a compact parameterization or
  Latent Variable representation
• Given N examples of find
  where
• Underlying assumptions to DimRedux
• Measurements over-specify data, M gt D
• The number of measurements exceed the number of
  true degrees of freedom in the system
• The measurements capture all of the significant
  variability

4
Uses for DimRedux

• Build a compact model of the data
• Compression for storage, transmission,
  retrieval
• Parameters for indexing, exploring, and
  organizing
• Generate plausible new data
• Answer fundamental questions about data
• What is its underlying dimensionality?How many
  degrees of freedom are exhibited?How many
  latent variables?
• How independent are my measurements?
• Is there a projection of my data set where
  important relationships stand out?

5
DimRedux in Data Modeling

• Data Clustering - Continuous to Discrete
• The curse of dimensionality the sampling density
  is proportional to N1/p.
• Need a mapping to a lower-dimensional space that
  preserves important relations
• Regression Modeling Continuous to Continuous
• A functional model that generates input data
• Useful for interpolation
• Embedding Space

6
Todays Focus

• Linear DimRedux methods
• PCA Pearson (1901) Hotelling (1935)
• MDS Torgerson (1952), Shepard (1962)
• Linear Assumption
• Data is a linear function of the parameters
  (latent variables)
• Data lies on a linear (Affine) subspace

where the matrix M is m x d
7
PCA What problem does it solve?

• Minimizes least-squares (Euclidean) error
• The D-dimensional model provided by PCA has the
  smallest Euclidean error of any D-parameter
  linear model.
• where is the model predicted by the
  D-dimensional PCA.
• Projects data s.t. the variance is maximized
• Find an optimal orthogonal basis set for
  describing the given data

8
Principle Component Analysis

• Also known to engineers as the Karhunen-Loéve
  Transform (KLT)
• Rotate data points to align successive axes with
  directions of greatest variance
• Subtract mean from data
• Normalize variance along each direction, and
  reorder according to the variance magnitude from
  high to low
• Normalized variance direction principle
  component
• Eigenvectors of systems Covariance
  Matrixpermute to order eigenvectors in
  descending order

9
Simple PCA Example

• Simple 3D example
• gtgt x rand(2, 500)
• gtgt z 1,0 0,1 -1,-1 x 001 ones(1,
  500)
• gtgt m (100 rand(3,3)) z rand(3, 500)
• gtgt scatter3(m(1,), m(2,), m(3,), 'filled')

10
Simple PCA Example (cont)

• gtgt mm (m- mean(m')' ones(1, 500))
• gtgt E,L eig(cov(mm ))
• gtgt E
• E
• 0.8029 -0.5958 0.0212
• 0.1629 0.2535 0.9535
• 0.5735 0.7621 -0.3006
• gtgt L
• L
• 172.2525 0 0
• 0 116.2234 0
• 0 0 0.0837
• gtgt newm E (m - mean(m)' ones(1, 500))
• gtgt scatter3(newm(1,), newm(2,), newm(3,),
  'filled')
• axis(-50,50, -50,50, -50,50)

11
Simple PCA Example (cont)
12
PCA Applied to Reillumination

• Illumination can be modeled as an additive
  linear system.

13
Simulating New Lighting

• We can simulate the appearance of a model under
  new illumination by combining images taken from a
  set of basis lights
• We can then capture real-world lighting and use
  it to modulate our basis lighting functions

14
Problems

• There are too many basis lighting functions
• These have to be stored in order to use them
• The resulting lighting model can be huge, in
  particular when representing high frequency
  lighting
• Lighting differences can be very subtle
• The cost of modulation is excessive
• Every basis image must be scaled and added
  together
• Each image requires a high-dynamic range
• Is there a more compact representation?
• Yes, use PCA.

15
PCA Applied to Illumination

• More than 90 variance is captured in the first
  five principle components
• Generate new illumination by combining only 5
  basis images

V0
for n lights
16
Results Video
17
Results Video
18
Results Video
19
MDS What problem does it solve?

• Takes as input a dissimilarity matrix M,
  containing pairwise dissimilarities between
  N-dimensional data points
• Finds the best D-dimensional linear
  parameterization compatible with M
• (in other words, outputs a projection of data in
  D-dimensional space where the pairwise distances
  match the original dissimilarities as faithfully
  as possible)

20
Multidimensional Scaling (MDS)

• Dissimilarities can be metric or non-metric
• Useful when absolute measurements are
  unavailable uses relative measurements
• Computation is invariant to dimensionality of
  data

21
An example map of the US

• Given only the distance between a bunch of cities

22
An example map of the US

• MDS finds suitable coordinates for the points of
  the specified dimension.

23
MDS Properties

• Parameterization is not unique Axes are
  meaningless
• Not surprising since Euclidean transformations
  and reflections preserve distances between points
• Useful for visualizing relationships in high
  dimensional data.
• Define a dissimilarity measure
• Map to a lower-dimensional space using MDS
• Common preprocess before cluster analysis
• Aids in understanding patterns and relationships
  in data
• Widely used in marketing and psychometrics

24
Dissimilarities

• Dissimilarities are distance-like quantities that
  satisfy the following conditions
• A dissimilarity is metric if, in addition, it
  satisfies
• The triangle inequality

25
Relating MDS to PCA

• Special case when distances are Euclidean
• PCA eigendecomposition of covariance matrix MTM
• Convert the pair-wise distance matrix to the
  covariance matrix

26
How to get MTM from Euclidean Pair-wise Distances
i
Law of cosines
j
Definition of a dot product

• Eigendecomposition on b to get VSVT
• VS1/2 matrix of new coordinates

27
Algebraically
So we centered the matrix
28
MDS Mechanics

• Given a Dissimilarity matrix, D, the MDS model is
  computed as follows
• Where, H, the so called centering matrix, is a
  scaled identity matrix computed as follows
• MDS coordinates given by (in order of decreasing

29
MDS Stress

• The residual variance of B (i.e. the sum of the
  remaining eigenvalues) indicate the goodness of
  fit for the selected d-dimensional model
• This term is often called MDS stress
• Examining the residual variance gives an
  indication of the inherent dimensionality

30
Reflectance Modeling Example
The top row of white, grey, and black balls have
the same physical reflectance parameters,
however, the bottom row is perceptually more
consistent.

• From Pellacini, et. al. Toward a
  Psychophysically-Based Light Reflection Model for
  Image Synthesis, SIGGRAPH 2000
• Objective Find a perceptually meaningful
  parameterization for reflectance modeling

31
Reflectance Modeling Example

• User Task Subjects were presented with 378
  pairs of rendered spheres an asked to rate their
  difference in glossiness on a scale of 0 (no
  difference) to 100.
• A dissimilarity 27 x 27 dissimilarity matrix was
  constructed and MDS applied

32
Reflectance Modeling Example

• Parameters of a 2D embedding space were
  determined
• Two axes of gloss were established

33
Limitations of Linear methods

• What if the data does not lie within a linear
  subspace?
• Do all convex combinations of the measurements
  generate plausible data?
• Low-dimensional non-linear Manifold embedded in a
  higher dimensional space
• Next time Nonlinear Dimensionality Reduction

34
Summary

• Linear dimensionality reduction tools are widely
  used for
• Data analysis
• Data preprocessing
• Data compression
• PCA transforms the measurement data s. t.
  successive directions of greatest variance are
  mapped to orthogonal axis directions (bases)
• An D-dimensional embedding space
  (parameterization) can be established by modeling
  the data using only the first d of these basis
  vectors
• Residual modeling error is the sum of the
  remaining eigenvalues

35
Summary (cont)

• MDS finds a d-dimensional parameterization that
  best preserves a given dissimilarity matrix
• Resulting model can be Euclidean transformed to
  align data with a more intuitive parameterization
• An D-dimensional embedding spaces
  (parameterization) are established by modeling
  the data using only the first d coordinates of
  the scaled eigenvectors
• Residual modeling error (MDS stress) is the sum
  of the remaining eigenvalues
• If Euclidean metric dissimilarity matrix is used
  for MDS the resulting d-dimensional model will
  match the PCA weights for the same dimensional
  model