System and method for large-scale multi-label learning using incomplete label assignments

Patent metadata

Field	Value
Publication number	US-2016140451-A1
Application number	US-201414543133-A
Country	US
Kind code	A1
Filing date	Nov 17, 2014
Priority date	Nov 17, 2014
Publication date	May 19, 2016
Grant date	—

Abstract

At least one label prediction model is trained, or learned, using training data that may comprise training instances that may be missing one or more labels. The at least one label prediction model may be used in identifying a content item's ground-truth label set comprising an indicator for each label in the label set indicating whether or not the label is applicable to the content item.

First claim

1 . A method comprising:
training, using a computing device, an initial level of a stacked model for use in making a labeling prediction, the initial level being trained using feature information for each training instance of a plurality of training instances, at least one training instance of the plurality is missing at least one label of a plurality of labels, the feature information corresponding to a plurality of features associated with the training instance of the plurality; generating, using the computing device, a labeling prediction for each training instance of the plurality using the initial level of the stacked model, the labeling prediction comprising a label applicability prediction for at least one label of the plurality of labels missing from the training instance's set of labels; training, using the computing device, one or more additional levels of the stacked model, each additional level being trained using information for each training instance of the plurality, each training instance's information comprising the labeling prediction from a previous level of the stacked model, the feature information corresponding to the plurality to features, and information indicating the training instance's set of labels; and identifying, using the computing device, a labeling prediction for a content item using the stacked model, the labeling prediction identifying for each label of the plurality whether the label is applicable to the content item.

2 . The method of claim 1 , the identifying further comprising:
using the content item's feature information to identify the content item's labeling prediction for the initial level of the stacked model; and using information comprising the labeling prediction from the previous level of the stacked model, the content item's feature information, and information indicating the content item's set of labels to identify the content item's labeling prediction for each additional level of the stacked model.

3 . The method of claim 1 , the initial level and each additional level of the stacked model comprising a plurality of weights, each weight in the plurality corresponding to a label of the plurality of labels.

4 . The method of claim 1 , the generating further comprising:
generating the labeling prediction for each training instance of the plurality using the initial level of the stacked model, the feature information corresponding to a plurality of features associated with the training instance and the training instance's set of labels.

5 . The method of claim 1 , further comprising:
generating, using the computing device, a labeling prediction for a training instance of the plurality using each additional level of the one or more additional levels of the stacked model.

6 . The method of claim 5 , the generating a labeling prediction for each training instance of the plurality using each additional level of the one or more additional levels of the stacked model further comprising:
generating the labeling prediction for a training instance of the plurality using each additional level of the stacked model and correlations between labels of the plurality.

7 . The method of claim 5 , the generating a labeling prediction for each training instance of the plurality using each additional level of the one or more additional levels of the stacked model further comprising:
generating the labeling prediction for a training instance of the plurality using each additional level of the stacked model, the previous level's labeling prediction for the instance, feature information corresponding to a plurality of features associated with the training instance and the training instance's set of labels.

8 . The method of claim 1 , generating a labeling prediction for each training instance of the plurality using the initial level of the stacked model further comprising:
generating the labeling prediction for the training instance using cross validation, such that the training instance is excluded from generating a model that is used to generate the labeling prediction for the training instance.

9 . The method of claim 1 , one of the one or more additional levels of the stacked model is a final level of the stacked model, and the final level of the stack model is used to generate the labeling prediction for a content item.

10 . A system comprising:
a processor; a storage medium for tangibly storing thereon program logic for execution by the processor, the stored logic comprising:
training logic executed by the processor for training an initial level of a stacked model for use in making a labeling prediction, the initial level being trained using feature information for each training instance of a plurality of training instances, at least one training instance of the plurality is missing at least one label of a plurality of labels, the feature information corresponding to a plurality of features associated with the training instance of the plurality;
generating logic executed by the processor for generating a labeling prediction for each training instance of the plurality using the initial level of the stacked model, the labeling prediction comprising a label applicability prediction for at least one label of the plurality of labels missing from the training instance's set of labels;
training logic executed by the processor for training one or more additional levels of the stacked model, each additional level being trained using information for each training instance of the plurality, each training instance's information comprising the labeling prediction from a previous level of the stacked model, the feature information corresponding to the plurality to features, and information indicating the training instance's set of labels; and
identifying logic executed by the processor for identifying a labeling prediction for a content item using the stacked model, the labeling prediction identifying for each label of the plurality whether the label is applicable to the content item.

11 . The system of claim 10 , the identifying logic executed by the processor further comprising:
using logic executed by the processor for using the content item's feature information to identify the content item's labeling prediction for the initial level of the stacked model; and using logic executed by the processor for using information comprising the labeling prediction from the previous level of the stacked model, the content item's feature information, and information indicating the content item's set of labels to identify the content item's labeling prediction for each additional level of the stacked model.

12 . The system of claim 10 , the initial level and each additional level of the stacked model comprising a plurality of weights, each weight in the plurality corresponding to a label of the plurality of labels.

13 . The system of claim 10 , the generating logic executed by the processor further comprising:
generating logic executed by the processor for generating the labeling prediction for each training instance of the plurality using the initial level of the stacked model, the feature information corresponding to a plurality of features associated with the training instance and the training instance's set of labels.

14 . The system of claim 10 , program logic for execution by the processor further comprising:
generating logic executed by the processor for generating a labeling prediction for a training instance of the plurality using each additional level of the one or more additional levels of the stacked model.

15 . The system of claim 14 , the generating logic execu

Assignees

• Yahoo Inc

Inventors

• Li Jia
• Chang Yi
• Kong Xiangnan

Classifications

• G06N99/005 Primary

  Physics · mapped topic

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G06N20/00 Primary

  Machine learning · CPC title