Personalized information retrieval search with backoff

• US 20080082485A1
• Filed: 09/28/2006
• Published: 04/03/2008
• Est. Priority Date: 09/28/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Mei, Qiaozhu, Church, Kenneth Ward
• Granted Patent: US 7,783,636 B2
• Time in Patent Office: Days
• US Class Current: 707/3
• CPC Class Codes: G06F 16/3331 Query processing

Abstract

Query logs are accessed to obtain queries, user information that specifies a user from which the query was received, a long with a selected result that was selected by the specified user who authored the query. This query log information is used to identify classes of users that looked for a similar result given a similar query. Those classes can then be used by a search engine in order to rank or provide search results to a user in response to a query input by the user.

Claims

1. A method of retrieving information, comprising:

receiving a query and user information indicative of a user;

receiving search results based on the query;

applying to the search results a user class model that models a plurality of classes of users based on user information and the query, to determine likely results desired by the user given the user information and the query; and

returning the search results to the based on application of the user class model.

2. The method of claim 1 wherein applying a user class model comprises:

applying a user class model that models a first set of user classes; and

applying a backing off model that models a second set of user classes.

3. The method of claim 2 wherein applying a backoff model comprises:

applying a plurality of backoff models that model successively larger user classes.

4. The method of claim 2 wherein applying a user model comprises:

applying a user class model that models grouping of users based on a user identifier.

5. The method of claim 4 wherein applying a user class model comprises:

applying a user class model that models grouping of users into classes based on a first portion of an IP address of the user and wherein applying the backoff model comprises applying a backoff model that models grouping of users based on a second, smaller portion of the IP address.

6. The method of claim 4 wherein applying a user class model comprises:

applying a user class model that models grouping of users into classes based on a first portion of a global unique user identifier (GUID) and wherein applying the backoff model comprises applying a backoff model that models grouping of users based on a second, smaller portion of the GUID.

7. The method of claim 2 wherein applying the user class model comprises:

applying a user class model that models grouping of users into classes based on a first geographic location of the user and wherein applying the backoff model comprises applying a backoff model that models grouping of users based on a second, larger geographic location of the user.

8. The method of claim 2 wherein applying the user class model comprises:

applying a user class model that models grouping of users into classes based on a first time window within which the user submitted the query and wherein applying the backoff model comprises applying a backoff model that models grouping of users based on a second, larger time window.

9. The method of claim 8 wherein the first time window comprises a day of the week and wherein the second time window comprises multiple days of the week.

10. The method of claim 1 and further comprising:

generating advertising based on application of the user class model.

11. The method of claim 1 and further comprising:

suggesting to the user a user action based on application of the user class model.

12. A system for conducting information retrieval searches for a user, comprising:

a search engine including a user class model that models a first plurality of different groups of users, the search engine configured to receive a query and execute the query over a network and to receive user information identifying a characteristic of the user and use the user class model to rank search results, provided for the user, based on the user information.

13. The system of claim 12 wherein the user class model includes a backoff model that models a second plurality of groups of users, larger than the first groups.

14. The system of claim 13 wherein the user class model comprises a model for each group in the first plurality of groups and the second plurality of groups.

15. The system of claim 812 wherein the search engine is configured to rank the search results by scoring each of the search results with the user class model.

16. A method of generating a search engine, comprising:

accessing query logs including queries previously executed, user information specifying a characteristic of a user that input each query, and a search result selected by the user;

training a user class model that models groups of users based on the user information, the query and the search result;

training a backoff model that models larger groups of users based on the user information, of the query and search result; and

configuring the search engine to receive and launch an input query, to receive user information for the user that input the input query and to access the user class model and backoff model to score search results based on user information for the user that inputs the input query.

17. The method of claim 16 and further comprising:

weighting the user class model and the backoff model with weights; and

combining the user class model and the backoff model using the weights.

18. The method of claim 17 wherein weighting comprises:

weighting the backoff model higher than the user class model.

19. The method of claim 18 wherein training a backoff model comprises:

training a plurality of backoff models each modeling a larger group of users; and

weighting each backoff model.

20. The method of claim 19 wherein combining comprises:

linearly combining the user class models and each of the backoff models.