Method and apparatus for explaining categorization decisions
AssigneeXerox Corporation
InventorsEric Gaussier, Cyril Goutte
Life of this patent
Estimating the term…
Abstract
Feature selection is used to determine feature influence for a given categorization decision to identify those features in a categorized document that were important in classifying the document into one or more classes. In one embodiment, model parameters of a categorization model are used to determine the features that contributed to the categorization decision of a document. In another embodiment, the model parameters of the categorization model and the features of the categorized document are used to determine the features that contributed to the categorization decision of a document.
Claims20
A method for explaining a categorization decision, comprising:
- computing a score for each of a plurality of classes using model parameter values of a selected categorization model and feature values of a document to be categorized;
- computing a contribution for each of a plurality of features in the document, where the contribution is computed using document-dependent feature selection;
- specifying a categorization decision using at least the class with the highest score;
- assigning an importance to each of a plurality of the features using the computed contributions to identify for different ones of the plurality of features, their influence on the categorization decision.
A method for explaining a categorization decision, comprising:
- computing a score for each of a plurality of classes using model parameter values of a selected categorization model and feature values of a document to be categorized;
- computing a contribution for each of a plurality of features in the document, where the contribution is computed using model-dependent feature selection;
- specifying a categorization decision using at least the class with the highest score;
- assigning an importance to each of a plurality of the features using the computed contributions to identify for different ones of the plurality of features, their influence on the categorization decision;
- wherein the contribution is computed by computing the contribution εf of feature f to a categorization decision using:
- ɛ
- i
- =
- w
- i
- =
-
- 1
- 2
-
- 2
-
- (
- f
-
- (
- 1
- i
- )
- -
- f
-
- (
- -
- 1
- i
- )
- )
- ,
- ɛ
- ii
- =
- w
- ii
- =
-
- 1
- 2
-
- (
- f
-
- (
- 1
- i
- )
- +
- f
-
- (
- -
- 1
- i
- )
- )
- -
- f
-
- (
- 0
- )
- ,
- ɛ
- ij
- =
- w
- ij
- =
-
- 1
- 2
-
- (
- f
-
- (
- 1
- i
- +
- 1
- j
- )
- -
- f
-
- (
- 1
- i
- )
- -
- f
-
- (
- 1
- j
- )
- =
- f
-
- (
- 0
- )
- )
- ,
- where
- εi is the frequency contribution from feature fi,
- εii is the squared contribution from feature fi, and
- εij is the contribution from the combination of features fi and fj,
- i and j are any two features in an original feature space.
Dependencies are inferred from the claim text.