Classification of rare events with high reliability

• US 20030204507A1
• Filed: 04/25/2002
• Published: 10/30/2003
• Est. Priority Date: 04/25/2002
• Status: Abandoned Application

• Current Assignee: Agilent Technologies, Inc.
• Original Assignee: Agilent Technologies, Inc.
• Inventors: Li, Jonathan Qiang, Barford, Lee A., Smith, David R., Heumann, John M.
• Application Number:
  US10/132,626
• Publication Number:
  US 20030204507A1
• Time in Patent Office: Days
• US Class Current: 707/100
• CPC Class Codes: G06V 30/2504 Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches

Abstract

Hierarchical classification of samples. First stage classification identifies most members of the majority class and removes them from further consideration. Second stage classification then focuses on discriminating between the minority class and the greatly reduced number of majority class samples lying near the decision boundary.

Claims

1. A hierarchical classifier for classifying data samples into a first majority result class or a second minority result class, the hierarchical classifier comprising a first stage classifier which classifies input samples into the first result class, or passes the samples on to a second stage classifier which classifies samples from the first stage classifier into the first result class or the second result class.

2. A hierarchical classifier according to claim 1 where the first classifier removes most data samples which are members of the first majority input class by classifying those data samples as members of the first result class.

3. A hierarchical classifier according to claim 1 where the second classifier maps an input sample on to a value which is compared to a threshold value.

4. A hierarchical classifier according to claim 3 where the threshold value is adjustable.

5. A hierarchical classifier according to claim 1 where the first stage classifier comprises a single substage which classifies samples into the first result class or the second result class.

6. A hierarchical classifier according to claim 1 where the first stage classifier comprises a plurality of substages in series in which each substage classifies samples from the first stage classifier into the first result class or passes samples on to the next substage.

7. A hierarchical classifier according to claim 1 where the second stage classifier comprises a single substage which classifies samples from the first stage classifier into the first result class or the second result class.

8. A hierarchical classifier according to claim 1 where the second stage classifier comprises a plurality of substages which classify samples from the first stage classifier into the first result class or the second result class.

9. A hierarchical classifier according to claim 8 where the second plurality of substages are applied in series.

10. A hierarchical classifier according to claim 8 where the second plurality of tests are applied in parallel, each of the tests providing a weight which is summed to classify samples from the first stage classifier into the first result class or the second result class.

11. The method of training a hierarchical classifier for classifying data samples which are members of a first majority input class or a second minority input class into a first result class or a second result class comprising:

selecting a first classification model,

training the first model,

selecting a second classification model, and

training the second classification model.

12. The method of claim 11 where the step of training the second classification model includes the step of minimizing overall cost.

13. The method of claim 12 where cost parameters used in minimizing overall cost are specified by the user.

14. The method of claim 11 where the second classification model uses a threshold value.

15. The method of claim 14 where the threshold value used by the second classification model may be altered without retraining either the first or second stages.