Method and system for analyzing multi-variate data using canonical decomposition

• US 7,409,321 B2
• Filed: 08/10/2006
• Issued: 08/05/2008
• Est. Priority Date: 07/12/2000
• Status: Expired due to Fees

• Current Assignee: Cornell Research Foundation Incorporated (Cornell University)
• Original Assignee: Cornell Research Foundation Incorporated (Cornell University)
• Inventors: Schiff, Nicholas, Victor, Jonathan, Repucci, Michael
• Primary Examiner(s): Nghiem; Michael P
• Application Number:
  US11/463,728
• Publication Number:
  US 20070005391A1
• Time in Patent Office: 726 Days
• Field of Search: 702/179, 702/182, 702/189, 702/190, 702/196, 707/100, 707/101, 707/104.1
• US Class Current: 702/189
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06F 2218/08 Feature extraction
• CPC Class Codes: G06Q 90/00 Systems or methods specially adapted for administrative, commercial, financial, managerial or supervisory purposes, not involving significant data processing

Abstract

A canonical decomposition (CD) method that includes building a multi-variate linear autoregressive (“MLAR”) model from an original data set or from a reduced set derived by data reduction methods from the original data set. The MLAR analysis is followed by seeking a coordinate transformation of the MLAR model to obtain the best possible match with one or more canonical forms representing relationships among components. For multi-variate data with a truly hierarchical structure, CD accurately extracts the underlying sources of the system.

Claims

1. A computer-assisted method of determining dynamic characteristics of data, the data being causal and having translation-invariant statistics with respect to a variable, which method comprises the steps of:

receiving signals representing the data, the signals originating from a set of components, the set of components having dynamic relationships with each other;

calculating a set of multi-linear autoregressive coefficients of the data, the multi-linear autoregressive coefficients yielding an array of square matrices, each square matrix reflecting the dynamic relationships among the components for a certain value of the variable;

storing the array of square matrices in a memory of a computer;

transforming each square matrix to a substantially canonical form corresponding to a selected canonical form suitable for analyzing the dynamic relationships, yielding a set of canonical form matrices representing a transformation of the components;

analyzing the selected canonical form and transformed components to determine the dynamic characteristics of the data with respect to the variable; and

means for outputting the transformation of the components from the computer.

2. A method according to claim 1, wherein the dynamic relationships are hierarchical.

3. A method according to claim 1, wherein the dynamic relationships are dependent.

4. A method according to claim 1, wherein the dynamic relationships are independent.

5. A method according to claim 1, wherein the dynamic relationships are partially dependent and partially independent.

6. A method according to claim 1, wherein the receiving step further comprises applying a data reduction method to the data.

7. A method according to claim 6, wherein the data reduction method comprises Principal Component Analysis.

8. A method according to claim 1, wherein the variable is time.

9. A computer-assisted method for determining an orthogonal rotation for analyzing data, the data being causal and having translation-invariant statistics with respect to a variable, the data originating from a set of components, the set of components having dynamic relationships to each other, which method comprises the steps of:

creating a multi-variate linear autoregressive model of the data, the model describing influences of one component on another component with respect to the variable;

storing the model in a memory of a computer;

decorrelating random terms in the model so that the decorrelated random terms are orthogonalized to produce a new model with autoregression coefficients, the random terms driving channels of the components;

identifying an orthogonal rotation that preserves the influences from the creating step and the orthogonality of the random terms from the decorrelating step, and that transforms the new model with autoregression coefficients into a canonical form; and

a means for outputting the transformation of the components from the computer.

10. A method according to claim 9, wherein the creating step further comprises performing data reduction on the data.

11. A method according to claim 10, wherein the data reduction performed is Principal Component Analysis decomposition.

12. A method according to claim 9, wherein the components are temporal, T_p,n, and are modeled according to:

$T_{p,n}=R_{p,n}+{\mathrm{\mu <br><br>}}_p+\underset{q=1}{\overset{P}{\sum <br><br>}}<br><br><br><br>\underset{l=1}{\overset{L}{\sum <br><br>}}<br><br><br><br>A_{q,p,l}<br><br>T_{q,n-l}$

whereby, P represents principal components;

L represents a predetermined number of lags;

N represents a total number of samples of the data observed at evenly spaced intervals;

R_p,n is a matrix of residual values of dimension P×

N;

u_p is a mean value of a pth component;

A is a three dimensional array of model coefficients of dimension P×

P×

L, and n is a particular sample or interval of time.

13. A method according to claim 9, wherein a set of residuals R_MLAR is calculated according to:

$R_{\mathrm{MLAR}}=\underset{p=1}{\overset{P}{\sum <br><br>}}<br><br><br><br>\underset{n=1}{\overset{N}{\sum <br><br>}}<br><br><br><br>R_{p,n}^2$

whereby, N represents a total number of samples of the data observed at evenly spaced intervals; and

P represents principal components.

14. A method according to claim 9, wherein the canonical form has zero and nonzero elements, the nonzero elements being located in positions only along the main diagonal and anywhere at locations right of the main diagonal.

15. A method according to claim 9, further comprising predicting future dynamic characteristics of the data using the canonical form.