Method, system, and computer program product for visualizing an evidence classifier

• US 5,930,803 A
• Filed: 04/30/1997
• Issued: 07/27/1999
• Est. Priority Date: 04/30/1997
• Status: Expired due to Term

• Current Assignee: RPX Corporation
• Original Assignee: Silicon Graphics Incorporated (Graphics Properties Holdings, Inc.)
• Inventors: Becker, Barry G., Kohavi, Ron, Sommerfield, Daniel A., Tesler, Joel D.
• Primary Examiner(s): Black, Thomas G.
• Assistant Examiner(s): Coby, Frantz
• Application Number:
  US08/841,341
• Time in Patent Office: 818 Days
• Field of Search: 707/100, 707/104, 345/342, 345/346, 345/121
• US Class Current: 1/1
• CPC Class Codes: G06F 16/34 Browsing; Visualisation therefor
• CPC Class Codes: Y10S 707/99945 Object-oriented database structure processing
• CPC Class Codes: Y10S 707/99948 Application of database or data structure, e.g. distributed, multimedia, or image

Abstract

A method, system, and computer program product visualizes the structure of an evidence classifier. An evidence inducer generates an evidence classifier based on a training set of labeled records. A mapping module generates visualization data files. An evidence visualization tool uses the visualization data files to display an evidence pane and/or a label probability pane. A first evidence pane display view shows a normalized conditional probability of each label value, for each attribute value. The first evidence pane display view can be a plurality of rows of pie charts. Each pie slice in a pie chart has a size which is a function of the normalized conditional probability of each label value for the respective attribute value. A second evidence pane display view shows relative conditional probabilities of a selected label value, for each attribute value. The second evidence pane display view can be a plurality of rows of bars. Bar height is a function of a conditional probability of a respective attribute value conditioned on the selected label value. A first label probability pane display view shows a pie chart of prior probabilities of each label value based on the training set. A second label probability pane display view shows a pie chart of posterior probabilities of each label value based on at least one selected attribute value. An importance slider controls filtering of attributes based on the importance of the attributes to a classification of unlabeled records. A count slider filters out attribute values having relatively low record counts. The evidence classifier visualization tool further provides sorting of attributes and/or attribute values. A subtracting minimum evidence capability is provided.

Claims

1. A computer-implemented method for visualizing the structure of an evidence classifier, the evidence classifier being generated from a training set of labeled records, each record having at least one attribute value and a corresponding class label, and the evidence classifier being capable of assigning class labels to unlabeled records based on attribute values found in the unlabeled records, comprising the steps of:

displaying an evidence pane including at least one of a first evidence pane display view and a second evidence pane display view, wherein said first evidence pane display view shows a normalized conditional probability of each label value for each attribute value, and said second evidence pane display view shows relative conditional probabilities of a selected label value, for each attribute value; and

displaying a label probability pane including at least one of a first label probability pane display view and a second label probability pane display view, wherein said first label probability pane display view shows prior probabilities of each label value, and said second label probability pane display view shows posterior probabilities of each label value based on at least one selected attribute value.

2. The method of claim 1, wherein said evidence pane displaying step comprises the step of displaying a first evidence pane display view having a plurality of rows of charts, each row corresponding to a respective attribute, each chart in a row corresponding to a respective attribute value, wherein each chart shows a normalized conditional probability of each label value for said respective attribute value.

3. The method of claim 2, wherein said first evidence pane display view displaying step comprises the step of displaying charts comprising pie charts, wherein pie slices in a respective pie chart have sizes which are a function of said normalized conditional probabilities of each label value for said respective attribute value.

4. The method of claim 3, wherein said first evidence pane display view displaying step further comprises the step of displaying a pie slice graphical attribute representative of a label value for each pie slice.

5. The method of claim 4, wherein said pie slice graphical attribute comprises color, and further comprising the step of assigning each pie slice in said first evidence pane display view a color representative of a label.

6. The method of claim 3, wherein said first evidence pane display view displaying step further comprises the step of displaying a pie chart graphical attribute that is a function of the number of records in the training set associated with the evidence classifier.

7. The method of claim 6, wherein said pie chart graphical attribute comprises height, and further comprising the step of assigning each pie chart a height that is a function of the number of records in the training set associated with the evidence classifier;

whereby a user can view heights of pie charts to determine the reliability of a classification.

8. The method of claim 3, further comprising the step of:

sorting pie charts for a nominal value attribute based on the size of slices for a particular label.

9. The method of claim 3, further comprising the steps of:

calculating said normalized conditional probability of each label value for said respective attribute value based on a conditional probability P(A|L), where P is the conditional probability that a random record chosen only from records with label L takes the attribute value A, said conditional probability P being determined based on record counts made with respect to the training set; and

mapping each calculated normalized conditional probability to a respective pie slice.

10. The method of claim 9, wherein said calculating step includes the step of determining said conditional probability P based on record counts made with respect to the training set and a Laplace probabilistic correction.

11. The method of claim 1, wherein said label probability pane displaying step comprises the step of displaying a first label probability pane display view having a chart that shows a prior probability for each label value.

12. The method of claim 11, wherein said chart comprises a pie chart, and further comprising the step of assigning pie slices in said pie chart sizes which are a function of prior probabilities for respective label values.

13. The method of claim 12, further comprising the steps of:

counting the number of records having a class label;

counting the total number of records;

dividing the number of records with a class label by the total number of records to calculate said prior probabilities for label values; and

mapping said calculated prior probabilities for each class label to said sizes of respective pie slices.

14. The method of claim 1, wherein said evidence pane displaying step comprises the step of displaying a second evidence pane display view having a plurality of rows of bars, each row corresponding to a respective attribute, each bar in a row corresponding to a respective attribute value, including the step of assigning each bar a height that is a function of a relative conditional probability of said selected label value conditioned on a respective attribute value.

15. The method of claim 14, wherein said bar height assigning step comprises assigning each bar a height representing at least one of evidence for or evidence against said selected label value.

16. The method of claim 15, wherein said bar height assigning step comprises the steps of:

calculating an evidence for value, z, for each bar height based on a negative logarithm of a difference between one and the size of a pie slice matching said selected label in the corresponding first evidence pane display view; and

mapping said calculated evidence for values, z, to respective bar heights.

17. The method of claim 15, further comprising the steps of:

calculating an evidence against value, z'"'"', for each bar height based on a negative logarithm of the size of a pie slice matching said selected label in the corresponding first evidence pane display view; and

mapping said calculated evidence against values, z'"'"', to respective bar heights.

18. The method of claim 14, further comprising the step of subtracting minimum evidence, said subtracting step comprising, for each bar, the steps of determining a minimum height value representing an approximate minimum height over all label values, and subtracting said minimum height value across all label values, whereby small differences are magnified in bar heights in said second evidence pane display views for different selected label values.

19. The method of claim 14, further comprising the step of assigning each bar a color that is a function of at least one of a corresponding confidence interval and number of records.

20. The method of claim 1, wherein said label probability pane displaying step comprises the step of displaying a second label probability pane display view having a chart that shows posterior probabilities of each label value based on at least one selected attribute value.

21. The method of claim 20, wherein said chart comprises a pie chart, and further comprising the step of assigning pie slices in said pie chart sizes which are a function of the respective posterior probabilities of each label value based on said at least one selected attribute value.

22. The method of claim 21, further comprising the steps of:

multiplying conditional probabilities of all said attribute values with the prior probabilities of each label value based on Bayes Rule to calculate said posterior probabilities of each label value based on at least one selected attribute value; and

mapping said posterior probabilities to said sizes of pie slices in said pie chart in said second label probability pane display view.

23. The method of claim 1, further comprising the step of:

displaying an importance controller that permits a user to control the filtering of attributes based on the importance of the attributes to a classification of unlabeled records.

24. The method of claim 7, further comprising the step of maintaining the sum of the heights of said pie charts in each row constant such that the distribution of pie chart heights in each row represents a histogram showing the way records are distributed over the attribute values for each attribute.

25. The method of claim 1, further comprising the step of binning a continuous attribute into discrete bins having binning intervals such that class distributions in each bin are different.

26. The method of claim 1, further comprising the step of sorting at least one of said attributes and said attribute values.

27. The method of claim 1, further comprising the step of:

displaying a count controller that permits a user to filter attribute values corresponding to relatively low counts.

28. The method of claim 1, further comprising the steps of:

permitting a user to select an attribute value in said first evidence pane; and

displaying said second label probability pane display view that shows posterior probabilities of each label value based on said at least one selected attribute value, thereby imitating behavior of the evidence classifier.

29. A system for visualizing the structure of an evidence classifier, the evidence classifier being generated from a training set of labeled records, each record having at least one attribute value and a corresponding class label, and the evidence classifier being capable of assigning class labels to unlabeled records based on attribute values found in the unlabeled records, comprising:

means for displaying an evidence pane including at least one of a first evidence pane display view and a second evidence pane display view, wherein said first evidence pane display view shows a normalized conditional probability of each label value for each attribute value, and said second evidence pane display view shows relative conditional probabilities of a selected label value, for each attribute value; and

means for displaying a label probability pane including at least one of a first label probability pane display view and a second label probability pane display view, wherein said first label probability pane display view shows prior probabilities of each label value, and said second label probability pane display view shows posterior probabilities of each label value based on at least one selected attribute value.

30. A computer program product comprising a computer useable medium having computer program logic recorded thereon for enabling a processor in a computer system to visually represent the structure of an evidence classifier, the evidence classifier being generated from a training set of labeled records, each record having at least one attribute value and a corresponding class label, and the evidence classifier being capable of assigning class labels to unlabeled records based on attribute values found in the unlabeled records, said computer program logic comprising:

means for enabling the processor to display an evidence pane including at least one of a first evidence pane display view and a second evidence pane display view, wherein said first evidence pane display view shows a normalized conditional probability of each label value for each attribute value, and said second evidence pane display view shows relative conditional probabilities of a selected label value, for each attribute value; and

means for enabling the processor to display a label probability pane including at least one of a first label probability pane display view and a second label probability pane display view, wherein said first label probability pane display view shows prior probabilities of each label value, and said second label probability pane display view shows posterior probabilities of each label value based on at least one selected attribute value.

31. A system for visualizing the structure of an evidence classifier, the evidence classifier being generated from a training set of labeled records, each record having at least one attribute value and a corresponding class label, and the evidence classifier being capable of assigning class labels to unlabeled records based on attribute values found in the unlabeled records, comprising:

means for generating a visualization data file defining an evidence pane; and

means for displaying said evidence pane based on said visualization file;

wherein said evidence pane includes at least one of a first evidence pane display view and a second evidence pane display view,

said first evidence pane display view shows a normalized conditional probability of each label value, for each attribute value, and

said second evidence pane display view shows relative conditional probabilities of a selected label value, for each attribute value.

32. The system of claim 31, wherein:

said visualization data file generating means further includes means for defining a label probability pane; and

said displaying means displays said label probability pane; and

wherein, said label probability pane includes at least one of a first label probability pane display view, and a second label probability pane display view;

said first label probability pane display view shows prior probabilities of each label value, and

said second label probability pane display view shows posterior probabilities of each label value based on at least one selected attribute value.

33. An evidence visualizer for visualizing the structure of an evidence classifier comprising:

an evidence inducer mapping module that defines at least one of an evidence pane and a label probability pane; and

a display that displays each evidence pane and label probability pane defined by said evidence inducer mapping module.

34. An integrated data mining system comprising:

an evidence inducer;

means for configuring said evidence inducer to generate a first data file representing structure of an evidence classifier and a second data file representing structure of a decision-tree classifier;

means for visualizing said evidence classifier structure based on said first data file; and

means for visualizing said decision-tree classifier structure based on said second data file.