Principal Component Analysis

Group (Subgroup)

DREAM3D Review (Dimensionality Reduction)

Description

This Filter performs principal component analysis on a data space defined by the user-selected scalar Attribute Arrays. Specifically, the dimensionality of the data space analyzed is equal to the number of selected arrays, where each array represents one dimension. The arrays may be of any primitive type, and are not required to share the same primitive type. This Filter uses a matrix approach to compute the principal components; the user may opt to use either the correlation or covariance matrix. The algorithm proceeds as follows:

  1. Form a \f$ m \f$ x \f$ n \f$ matrix, \f$ \mathbf{X} \f$, where the columns are the input Attribute Arrays (\f$ n \f$ is the number of selected arrays (columns) and \f$ m \f$ is the number of tuples for the arrays (rows))
  2. When forming \f$ \mathbf{X} \f$, the values for each of the arrays are cast to double precision
  3. If the correlation approach is being used, standardize each column of \f$ \mathbf{X} \f$ such that it has mean of 0 and unit variance; this approach may be useful if each Attribute Array has different scale
  4. Center \f$ \mathbf{X} \f$ by subtracting the column-wise mean from each row value; call this centered matrix \f$ \mathbf{B} \f$
  5. Compute the covariance/correlation matrix, \f$ \mathbf{C} \f$, as follows:

\f[ \mathbf{C} = { 1 \over {m-1} } \mathbf{B}^{*} \otimes \mathbf{B} \f]

where \f$ * \f$ is the conjugate transpose; if \f$ m = 1 \f$, then the prefactor denominator is set to 1

  1. Perform the eigen decomposition of \f$ \mathbf{C} \f$ to find the eigenvalues and eigenvectors; the eigenvalues are stored in ascending order, and the eigenvectors correspond to the same order as the eigenvalues.

The computed eigenvalues and eigenvectors are stored in a new Attribute Matrix. The number of eigenvalues/eigenvectors computed is \f$ n \f$; the dimensionality of the eigenvectors is also \f$ n \f$.

The user may opt to project the data space to a lower dimensionality using the computed eigenvectors. A lower dimensionality, \f$ d \f$, must be specified; the \f$ d \f$ eigenvectors that have the highest eigenvalues are used to project the data space. The \f$ d \f$ eigenvectors form a \f$ d \f$ x \f$ n \f$ matrix that is used to post-multiply each row of the centered matrix \f$ \mathbf{B} \f$. The result is an Attribute Array that is \f$ m \f$ tuples long with \f$ d \f$ dimensions. It may be useful to visualize this space as a point cloud; this can be accomplished by creating a Vertex Geometry using the projected array as coordinates. Note that a Vertex Geometry requires three coordinates for point positions, so if \f$ d < 3 \f$, additional components must be added to the projected data space. It is possible to create an array of all zeros and then append it to the projected array to get the correct dimensionality. Also note that Vertex Geometry coordinates must be 32-bit floating point values, but the projected space created by this Filter will be 64-bit floating point; the different precision can be created by converting the primitive type of the projected array. Finally, the data values from the original arrays may be visualized within this projected space by moving the original Attribute Arrays into the Vertex Attribute Matrix of the new Vertex Geometry.

Parameters

Name Type Description
Matrix Approach Enumeration Whether to use the correlation or covariance matrix approach
Project Data Space bool Whether to project the input data space to a lower dimensionality based on the principal component eigenvectors
Number of Dimensions for Projection int32_t The number of dimensions on which to project the data space, if Project Data Space is checked

Required Geometry

None

Required Objects

Kind Default Name Type Component Dimensions Description
Any Attribute Array None Any (1) Attribute Arrays that form the data space for the principal component analysis

Created Objects

Kind Default Name Type Component Dimensions Description
Generic Attribute Matrix PrincipalComponentAnalysis Generic N/A Attribute Matrix in which to store the results of the eigen analysis
Attribute Array PrincipalComponentEigenvalues double (1) Eigenvalues from the principal component analysis
Attribute Array PrincipalComponentEigenvectors double (n) Eigenvectors from the principal component analysis
Attribute Array ProjectedDataSpace double (Number of Dimensions for Projection) Projected data space, if Project Data Space is checked

Example Pipelines

Please see the description file distributed with this plugin.

DREAM3D Mailing Lists

If you need more help with a filter, please consider asking your question on the DREAM3D Users mailing list: https://groups.google.com/forum/?hl=en#!forum/dream3d-users