LEARNING-BASED IMAGE PAGE INDEX SELECTION

• US 20140105488A1
• Filed: 10/17/2012
• Published: 04/17/2014
• Est. Priority Date: 10/17/2012
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Geng, Bo, Zhou, Dengyong, Hua, Xian-Sheng, Wu, Zhong
• Granted Patent: US 9,070,046 B2
• Time in Patent Office: Days
• US Class Current: 382/161
• CPC Class Codes: G06F 16/951 Indexing; Web crawling techniques
• CPC Class Codes: G06F 18/2193 based on specific statistical tests

Abstract

Architecture that performs image page index selection. A learning-based framework learns a statistical model based on the hyperlink (URL-uniform resource locator) previous click information obtained from the image search users. The learned model can combine the features of a newly discovered URL to predict the possibility of the newly-discovered URL being clicked in the future image search. In addition to existing web index selection features, image clicks are added as features, and the image clicks are aggregated over different URL segments, as well as the site modeling pattern trees to reduce the sparse problem of the image click information.

Claims

1. A system, comprising:

a data sampling and labeling component that samples training webpages having images to create positive webpage samples and negative webpage samples;

a feature generation component that extracts features from the training webpages;

a model creation component that creates a selection model based on the sampled training webpages and the features;

a validation component that evaluates and validates the selection model using benchmark data to obtain an optimal selection model;

a prediction component that predicts an image webpage static rank based on the optimal selection model; and

a microprocessor that executes computer-executable instructions associated with at least one of the data sampling and labeling component, feature generation component, model creation component, validation component, or prediction component.

2. The system of claim 1, further comprising an index component that selects top ranked webpages as index targets based on the image webpage rank.

3. The system of claim 1, wherein the features are extracted in association with a webpage URL (uniform resource locator), the features include at least one of link-based features, click features, URL pattern features, or webpage property features.

4. The system of claim 1, wherein the data sampling and labeling component samples the training webpages within a specified time window and past clicks are used as features to train the selection model to predict a selection score for future discovered URLs.

5. The system of claim 1, wherein the data sampling and labeling component employs image clicks as labeling criteria to create the positive webpage samples and negative webpage samples.

6. The system of claim 1, wherein the feature generation component employs image clicks from at least one of a browser log, toolbar log, or a search engine log as features for image page index selection.

7. The system of claim 1, wherein the feature generation component utilizes image click aggregation over different URL levels as features.

8. The system of claim 1, wherein the feature generation component utilizes image click aggregation over a website host to model URL patterns for the host, the URL patterns used as features.

9. A method, performed by a computer system executing machine-readable instructions, the method comprising acts of:

sampling and labeling webpages having images as training webpages to create positive webpages and negative webpages;

extracting features from the training webpages;

creating a selection model based on the training webpages and the features;

evaluating and validating the selection model using benchmark data to obtain an optimal selection model; and

predicting an image webpage rank based on the optimal selection model.

10. The method of claim 9, further comprising creating an image index based on the predicted image page rank.

11. The method of claim 9, further comprising accumulating and merging the features according to a specific time, and representing each training webpage by the features at that specific time.

12. The method of claim 9, further comprising learning the selection model based on URL click information of image search users.

13. The method of claim 9, further comprising creating the positive webpage and negative webpage sampling data based on a user click of an image of a vertical image search.

14. The method of claim 9, further comprising applying weight decay to each image click to accumulate the image clicks.

15. The method of claim 9, further comprising employing an image click as a labeling criterion and an image click as a URL feature.

16. The method of claim 9, further comprising employing image click aggregation over URL segments, and image click aggregation over website hosts to model URL patterns.

17. A computer-readable medium comprising instructions that when executed by a processor cause the processor to perform a method comprising acts of:

sampling and labeling training webpages having images to create positive webpages and negative webpages derived from user clicks as labeling criteria;

extracting historical features from the training webpages, the historical features of each webpage accumulated and merged to a specific time;

creating a statistical selection model based on the sampled training webpages and the features;

evaluating and validating the selection model using benchmark data to obtain an optimal selection model;

applying the optimal selection model to newly-discovered URLs to compute corresponding selection scores; and

selecting a set of top-ranked URLs as candidates for indexing based on the scores.

18. The computer-readable medium of claim 17, the method further comprising an act of modeling a web host site as a tree structure that represents URL patterns of the host site.

19. The computer-readable medium of claim 17, the method further comprising an act of aggregating image clicks over URL segments that include domain, host, L1-path and suffix, the image clicks of the different segments are utilized as different features.

20. The computer-readable medium of claim 17, the method further comprising an act of utilizing image clicks from at least one of a browser log, toolbar log, or a search engine log as features for image page index selection, in addition to webpage index selection features.