METHOD AND APPARATUS FOR MODEL-SHARED SUBSPACE BOOSTING FOR MULTI-LABEL CLASSIFICATION

• US 20090157571A1
• Filed: 12/12/2007
• Published: 06/18/2009
• Est. Priority Date: 12/12/2007
• Status: Active Grant

• Current Assignee: SINOEAST CONCEPT LIMITED (Tencent Holdings Limited)
• Original Assignee: International Business Machines Corporation
• Inventors: Smith, John R., Yan, Rong, Tesic, Jelena
• Granted Patent: US 7,958,068 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data

Abstract

A computer program product includes machine readable instructions for managing data items, the instructions stored on machine readable media, the product including instructions for: initializing a plurality of base models; minimizing a joint loss function to select models from the plurality for a plurality of labels associated with the data items; and at least one of sharing and combining the selected base models to formulate a composite classifier for each label. A computer system and additional computer program product are provided.

Claims

1. A computer program product comprising machine readable instructions for managing data items, the instructions stored on machine readable media, the product comprising instructions for:

initializing a plurality of base models;

minimizing a joint loss function to select models from the plurality for a plurality of labels associated with the data items; and

at least one of sharing and combining the selected base models to formulate a composite classifier for each label.

2. The computer program product as in claim 1, wherein the plurality of base models comprises at least one of:

a learning algorithm, a decision tree model, a N model (k-Nearest Neighbors), a support vector machines (SVMs) model and a Gaussian mixture model.

3. The computer program product as in claim 1, wherein at least one of the base models is learned from multi-modal information.

4. The computer program product as in claim 3, wherein the multi-modal information comprises at least one of visual information, audio information and text information.

5. The computer program product as in claim 3, wherein the multi-modal information comprises feature information.

6. The computer program product as in claim 5, wherein the feature information comprises at least one of color, texture, shape, appearance, sound, duration, arrangement, inter-relationship and order.

7. The computer program product as in claim 1, wherein at least one of the plurality of base models is adapted for modeling one of a visual topic, an event topic, a genre, and a type.

8. The computer program product as in claim 1, wherein at least one of the plurality of base models is learned from labeled data.

9. The computer program product as in claim 8, wherein where the labeled data is derived from at least one of metadata, annotation, tagging, associated text or closed caption transcripts and crowd-sourcing.

10. The computer program product as in claim 1, wherein at least one of the plurality of base models is learned from selected data samples.

11. The computer program product as in claim 1, wherein at least one of the plurality of base models is learned from a feature subspace.

12. The computer program product as in claim 1, wherein the joint loss function comprises at least one of a weighting for a label, a sum of the learning losses for each label.

13. The computer program product as in claim 1, wherein certain ones of the plurality of the base models can be shared with a classifier.

14. The computer program product as in claim 1, wherein the composite classifier is a combination of base models for generating prediction outputs.

15. The computer program product as in claim 1, wherein a model-shared subspace boosting is customized for a classification objective.

16. A computing system for managing multi-mode data items, the system comprising:

at least one processor for processing a computer program product, the product comprising machine readable instructions for managing data items, the instructions stored on machine readable media coupled to the computing system, the product comprising instructions for;

initializing a plurality of base models;

minimizing a joint loss function to select models from the plurality for a plurality of labels associated with the data items; and

at least one of sharing and combining the selected base models to formulate a composite classifier for each label.

17. A computer program product comprising machine readable instructions for managing data items, the instructions stored on machine readable media, the product comprising instructions for:

initializing a plurality of base models comprising at least one of;

a learning algorithm, a decision tree model, a kNN model (k-Nearest Neighbors), a support vector machines (SVMs) model, a Gaussian mixture model and a model learned from multi-modal information;

minimizing a joint loss function to select models from the plurality for a plurality of labels associated with the data items; and

at least one of sharing and combining the selected base models to formulate a composite classifier for each label;

wherein the multi-modal information comprises at least one of visual information, audio information, text information and feature information, the feature information comprising at least one of color, texture, shape, appearance, sound, duration, arrangement, inter-relationship and order;

wherein at least one of the plurality of base models is learned from at least one of a feature subspace, selected data samples and labeled data, the labeled data being derived from at least one of metadata, annotation, tagging, associated text or closed caption transcripts and crowd-sourcing; and

wherein the joint loss function comprises at least one of a weighting for a label, a sum of the learning losses for each label.

18. The computer program product as in claim 17, wherein certain ones of the plurality of the base models can be shared with a classifier.

19. The computer program product as in claim 17, wherein the composite classifier is a combination of base models for generating prediction outputs.

20. The computer program product as in claim 17, wherein a model-shared subspace boosting is customized for a classification objective.