Method for visualizing feature ranking of a subset of features for classifying data using a learning machine

• US 8,126,825 B2
• Filed: 04/04/2011
• Issued: 02/28/2012
• Est. Priority Date: 05/01/1998
• Status: Expired due to Term

• Current Assignee: Curtis Anderson, Health Discovery Corporation, James Roberts, Joe Mckenzie, Jules B. Paderewski, Julian N. Stern, Memorial Health Systems Incorporated, Timothy P. O'Hayer
• Original Assignee: Health Discovery Corporation
• Inventors: Guyon, Isabelle
• Primary Examiner(s): Holmes, Michael B
• Application Number:
  US13/079,198
• Publication Number:
  US 20110184896A1
• Time in Patent Office: 330 Days
• Field of Search: 706/12
• US Class Current: 706/12
• CPC Class Codes: G06F 18/2115 by evaluating different subsets according to an optimisation criterion, e.g. class separability, forward selection or backward elimination
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G16B 25/00 ICT specially adapted for hybridisation; ICT specially adapted for gene or protein expression
• CPC Class Codes: G16B 25/10 Gene or protein expression profiling; Expression-ratio estimation or normalisation
• CPC Class Codes: G16B 40/00 ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding
• CPC Class Codes: G16B 40/20 Supervised data analysis
• CPC Class Codes: G16B 40/30 Unsupervised data analysis
• CPC Class Codes: G16B 50/00 ICT programming tools or database systems specially adapted for bioinformatics
• CPC Class Codes: G16B 50/30 Data warehousing; Computing architectures

Abstract

A method for enhancing knowledge discovery from a dataset uses visualization of a subset features within a dataset that provide the best separation of the dataset into classes. One or more classifiers are trained using each subset of features and the success rate of the classifiers in accurately classifying the dataset is calculated. The success rate is converted into a ranking that is represented as a visually distinguishable characteristic. One or more tree structures may be displayed with a node representing each feature, and the visually distinguishable characteristic is used to indicate the scores for each feature subset. Connectors between the nodes may be used to indicate unconstrained and constrained feature sets. Nodes within a constrained path may be substituted for a feature within the preferred, unconstrained path if that feature is impractical to measure.

Claims

1. A method for enhancing knowledge obtained from a dataset by visualizing subsets of features selected from a plurality of features that describe the dataset, the method comprising:

downloading the dataset into a processor programmed for executing one or more learning machine classifiers;

training the one or more classifiers with each subset of features;

calculating a success rate of the one or more classifiers trained on each subset of features;

assigning a rank to each subset of features according to the success rate of the trained classifier in accurately classifying the dataset;

assigning a visually distinguishable characteristic to each rank; and

displaying a graph at a user interface display, the graph comprising a plurality of representations of subsets of features, wherein each representation of the subset of features comprises the visually distinguishable characteristic corresponding to the rank of the subset of features.

2. The method of claim 1, wherein the subsets of features are nested subsets and the graph comprises a tree comprising a plurality of nodes representing the features and having one or more root nodes corresponding to the smallest number of features in a subset of features.

3. The method of claim 2, wherein a terminal node at a depth D of a subset of features has a visually distinguishable characteristic corresponding to a combined rank of the subset of features.

4. The method of claim 2, wherein the tree comprises a plurality of connectors between a plurality of nodes, wherein the connectors corresponding to an unconstrained connector path are distinguishable from a constrained path.

5. The method of claim 4, wherein a constrained path is selected when one or more features in a subset of features within an unconstrained path is impractical.

6. The method of claim 5, wherein an alternative feature is selected from a constrained path when a feature in a subset of features within an unconstrained path is impractical.

7. The method of claim 1, wherein the features comprises genes or proteins that are differentially expressed in a disease relative to normal.

8. The method of claim 7, wherein the subsets of genes or proteins are nested subsets and the graph comprises a tree comprising a plurality of nodes representing the genes or proteins and having one or more root nodes corresponding to the smallest number of genes or proteins in a subset of genes or proteins.

9. The method of claim 8, wherein a terminal node at a depth D of a subset of genes or proteins has a visually distinguishable characteristic corresponding to a combined rank of the subset of genes or proteins.

10. The method of claim 8, wherein the tree comprises a plurality of connectors between a plurality of nodes, wherein the connectors corresponding to an unconstrained connector path are distinguishable from a constrained path.

11. The method of claim 10, wherein a subset of features comprises a group of proteins and a constrained path is selected when one or more proteins in the group of proteins within an unconstrained path is impractical to dose or measure.

12. The method of claim 10, wherein an alternative protein is selected from a constrained path when a selected protein in a group of proteins within an unconstrained path is impractical impractical to dose or measure.

13. The method of claim 1, wherein the visually distinguishable characteristic is color or shading.

14. The method of claim 1, wherein the success rate is calculated using a leave-one-out method.

15. The method of claim 1, wherein the one or more learning machine classifiers is one or more support vector machines.

16. The method of claim 15, wherein the subset of features is selected by recursive feature elimination.

17. A computer program product embodied on a computer readable medium for enhancing knowledge discovered from a dataset by visualizing nested subsets of features selected from a plurality of features that describe a dataset, the computer program product comprising instructions for executing learning machine classifiers and further for causing a computer processor to:

receive the dataset;

train the one or more classifiers with each subset of features;

calculate a success rate of the one or more classifiers trained on each subset of features;

assign a rank to each subset of features according to the success rate of the trained classifier in accurately classifying the dataset;

assign a visually distinguishable characteristic to each rank; and

display one or more trees at a user interface, each tree comprising a plurality of nodes, each node representing a feature, wherein a representation of the subset of features comprises the visually distinguishable characteristic corresponding to the rank of the subset of features.

18. The computer program product of claim 17, wherein the tree comprises one or more root nodes corresponding to the smallest number of features in a subset of features.

19. The computer program product of claim 17, wherein a terminal node at a depth D of a subset of features has a visually distinguishable characteristic corresponding to a combined rank of the subset of features.

20. The computer program product of claim 17, wherein the tree comprises a plurality of connectors between the plurality of nodes, wherein the connectors corresponding to an unconstrained connector path are distinguishable from a constrained path.

21. The computer program product of claim 20, wherein a subset of features comprises a group of proteins and a constrained path is selected when one or more proteins in the group of proteins within an unconstrained path is impractical to dose or measure.

22. The computer program product of claim 21, wherein an alternative protein is selected from a constrained path when a selected protein in a group of proteins within an unconstrained path is impractical to dose or measure.

23. The computer program product of claim 17, wherein the success rate is calculated using a leave-one-out method.

24. The computer program product of claim 17, wherein the one or more learning machine classifiers is one or more support vector machines.

25. The computer program product of claim 24, wherein the subset of features is selected using recursive feature elimination.