FEATURE NORMALIZATION AND ADAPTATION TO BUILD A UNIVERSAL RANKING FUNCTION

• US 20100293175A1
• Filed: 05/12/2009
• Published: 11/18/2010
• Est. Priority Date: 05/12/2009
• Status: Abandoned Application

• Current Assignee: Oath Inc. (Verizon Communications Inc.)
• Original Assignee: Oath Inc. (Verizon Communications Inc.)
• Inventors: Vadrevu, Srinivas, Tseng, Belle L.
• Application Number:
  US12/464,660
• Publication Number:
  US 20100293175A1
• Time in Patent Office: Days
• US Class Current: 707/759
• CPC Class Codes: G06F 16/951 Indexing; Web crawling techniques
• CPC Class Codes: G06F 16/953 Querying, e.g. by the use of web search engines

Abstract

To increase the amount of training data available to train a machine learning ranking function, data from multiple markets are normalized in such a manner as to optimize a measurement of quality of the ranking function trained on the various sets of normalized training data. Furthermore, the feature scores of training data from individual markets are adapted to conform to the distributions of feature scores from a base market. Such adapted training data from the various markets may be used to train a single, robust ranking function. Adaptation of feature scores in a particular training data set involves mapping feature scores of the particular training data set to feature scores of a base training data set to conform the distributions of the feature scores in the particular training data set to the distributions of the feature scores in the base training data set.

Claims

1. A computer-executed method comprising:

determining a first data item from a first set of data, wherein the first data item includes a first original feature score for a particular feature;

calculating a first calculated feature score for the particular feature of the first data item based at least in part on a first set of values and the first original feature score;

determining a first evaluation score based at least in part on the first calculated feature score;

wherein the first set of values are selected based at least in part on optimizing the first evaluation score; and

wherein the method is performed by one or more computing devices.

2. The computer-executed method of claim 1, further comprising:

determining a set of rules based at least in part on the first calculated feature score;

creating a ranking function based at least in part on the set of rules;

ranking, based at least in part on the ranking function, one or more data items; and

wherein determining the first evaluation score further comprises basing the first evaluation score at least in part on the set of rules.

3. The computer-executed method of claim 2, further comprising:

determining a second data item from a second set of data, wherein the second data item includes a second original feature score for the particular feature;

calculating a second calculated feature score for the particular feature of the second data item based at least in part on a second set of values and the second original feature score; and

determining a second evaluation score based at least in part on the second calculated feature score;

wherein the second set of values are selected based at least in part on optimizing the second evaluation score.

4. The computer-executed method of claim 3,

wherein the first original feature score is measured according to a first scale;

wherein the second original feature score is measured according to a second scale; and

wherein the first scale is different than the second scale.

5. The computer-executed method of claim 2,

wherein the one or more data items correspond to data items in the first set of data;

wherein each data item of the one or more data items comprises a feature score for the particular feature; and

wherein the step of ranking the one or more data items further comprises;

calculating a calculated feature score for the particular feature of each data item of the one or more data items based, at least in part, on the first set of values to produce a set of one or more calculated feature scores,

providing the set of one or more calculated feature scores to the ranking function, and

ranking, by the ranking function, the one or more data items, based at least in part on the set of one or more calculated feature scores.

6. The computer-executed method of claim 3,

wherein the first set of data originates from a first market and the second set of data originates from a second market;

wherein a first data item of the one or more data items originates from the first market;

wherein a second data item of the one or more data items originates from the second market; and

wherein a third data item of the one or more data items originates from a third market.

13. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 1.

14. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 2.

15. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 3.

16. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 5.

7. A computer-executed method comprising:

determining a first feature score associated with a particular feature of data in a first data set;

determining a first subset of data, of the first data set, having the first feature score for the particular feature;

associating the first feature score with a first distribution of relevance scores associated with the first subset of data;

determining a second feature score associated with the particular feature of data in a second data set;

determining a second subset of data, of the second data set, having the second feature score for the particular feature;

associating the second feature score with a second distribution of relevance scores associated with the second subset of data;

determining whether a difference between the first distribution and the second distribution is below a specified threshold; and

in response to determining that the difference between the first distribution and the second distribution is below the specified threshold, changing the second feature score to be the first feature score;

wherein the method is performed by one or more computing devices.

8. The computer-executed method of claim 7,

wherein the step of changing the second feature score to be the first feature score further comprises changing the second feature score in each data item of the second subset of data to be the first feature score;

and

the method further comprising;

determining a set of rules based at least in part on the first subset of data and the second subset of data,

creating a ranking function based at least in part on the set of rules, and

ranking one or more data items based at least in part on the ranking function.

9. The computer-executed method of claim 8,

wherein the one or more data items correspond to data items in the second data set;

wherein each data item of the one or more data items comprises a feature score for the particular feature; and

wherein the step of ranking the one or more data items further comprises;

determining whether a particular data item of the one or more data items includes the second feature score, and

in response to determining that the particular data item includes the second feature score, ranking, by the ranking function, the particular data item based at least in part on the first feature score.

10. The computer-executed method of claim 7,

wherein the data in the first data set comprises a query/document pair comprising a corresponding query and a corresponding document;

wherein the particular feature is one of;

(a) a feature of the corresponding query, (b) a feature of the corresponding document, or (c) a feature of the query/document pair;

wherein the query/document pair is associated with a particular graded relevance score; and

wherein the particular graded relevance score of the query/document pair is determined by a human.

11. The computer-executed method of claim 7, further comprising:

determining a third feature score associated with the particular feature of data in the first data set;

determining a third subset of data, of the first data set, having the third feature score for the particular feature;

associating the third feature score with a third distribution of relevance scores associated with the third subset of data;

determining a fourth feature score associated with the particular feature of data in the second data set;

determining a fourth subset of data, of the second data set, having the fourth feature score for the particular feature;

associating the fourth feature score with a fourth distribution of relevance scores associated with the fourth subset of data;

determining whether a difference between the third distribution and the fourth distribution is below the specified threshold;

in response to determining that the difference between the third distribution and the fourth distribution is below the specified threshold, changing the fourth feature score to be the third feature score in the fourth subset of data;

determining a fifth feature score associated with the particular feature of data in the second data set, wherein the fifth feature score is between the second feature score and the fourth feature score;

determining a particular value based at least in part on the first feature score and the third feature score; and

changing the fifth feature score to be the particular value.

12. The computer-executed method of claim 8,

wherein the first data set originates from a first market and the second data set originates from a second market;

wherein a first data item of the one or more data items originates from the first market;

wherein a second data item of the one or more data items originates from the second market; and

wherein a third data item of the one or more data items originates from a third market.

17. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 7.

18. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 8.

19. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 9.

20. One or more storage media storing instructions which, when executed by one or more computing devices, cause performance of the method recited in claim 11.