Using Aggregate Location Metadata to Provide a Personalized Service

• US 20120317104A1
• Filed: 06/13/2011
• Published: 12/13/2012
• Est. Priority Date: 06/13/2011
• Status: Abandoned Application

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Radlinski, Filip, Bennett, Paul N., White, Ryen W., Yilmaz, Emine
• Application Number:
  US13/158,483
• Publication Number:
  US 20120317104A1
• Time in Patent Office: Days
• US Class Current: 707/724
• CPC Class Codes: G06F 16/9535 Search customisation based on user profiles and personalisation
• CPC Class Codes: G06F 16/9537 Spatial or temporal dependent retrieval, e.g. spatiotemporal queries

Abstract

Functionality is described herein which generates a plurality of item models based on the aggregate behavior of users, such as the aggregate behavior of the users in selecting network-accessible sites and/or issuing particular queries. In one implementation, each item model estimates a probabilistic distribution of locations for an individual, given that the individual selects a particular item (e.g., a particular site or query). The functionality can use the item models to provide a personalized service to an end user. For example, in one scenario, the functionality can generate a plurality of location-based features based on the item models. The functionality can then learn a ranking model based on the location-based features. In a real-time phase of operation, a query processing system uses the ranking model to personalize search results for an end user.

Claims

1. A training system, implemented using computing functionality, for generating item models for use in providing a personalized service, comprising:

a data collection module for providing location-tagged data, the location-tagged data identifying one or more of;

sites that have been selected by a group of data-providing users with respect to respective locations of the data-providing users; and

queries that have been issued by the group of data-providing users with respect to respective locations of the data-providing users;

a data store for storing the location-tagged data;

an item model generation module for generating a plurality of item models based on the location-tagged data, the plurality of items models including one or more of;

at least one site model that estimates a probabilistic distribution of locations for an individual, given that the individual selects a particular site; and

at least one query model that estimates a probabilistic distribution of locations for the individual, given that the individual issues a particular query; and

a data store for storing one or more of said at least one site model and said at least one query model.

2. The training system of claim 1, wherein each of said at least one site model and at least one query model comprises a Gaussian mixture model that comprises a weighted combination of Gaussian components.

3. The training system of claim 1, further comprising a feature generation module for generating a group of location-based features based on:

user online activity data that describes online activity performed by the data-providing users; and

one or more of said at least one site model and said at least one query model.

4. The training system of claim 3,

wherein the group of location-based features comprises one or more non-contextual features, selected from among:

a feature based on an aggregate popularity of a particular site;

a feature based on an aggregate popularity of a particular query;

a feature based on an entropy of said at least one site model;

a feature based on an entropy of said at least one query model;

a feature based on a divergence of said at least one site model from a background site model, the background site model describing a distribution of locations for plural selections of sites;

a feature based on a divergence of said at least one query model from a background query model, the background query model describing a distribution of locations for plural selection of queries;

a feature based on a mean width of said at least one site model;

a feature based on a mean width of said at least one query model; and

a feature based on a divergence between said at least one site model and said at least one query model,

and wherein the group of location-based features also comprises one or more contextual features, selected from among;

a feature based on an assessed location associated with the individual;

a feature based on a probability of the assessed location of the individual, given that the individual selects a particular site;

a feature based on a probability of the assessed location of the individual, given that the individual issues a particular query;

a feature based on uncertainty associated with the assessed location of the individual, given that the individual selects a particular site;

a feature based on uncertainty associated with the assessed location of the individual, given that the individual issues a particular query;

a feature based on a probability that the individual has selected a particular site, given the assessed location of the individual;

a feature based on a probability that the individual has selected a particular query, given the assessed location of the individual;

a feature based on a percent of a probability mass associated with said at least one site model that is within a particular distance of the assessed location of the individual;

a feature based on a percent of a probability mass associated with said at least one query model that is within a particular distance of the assessed location of the individual;

a feature based on a distance between the assessed location of the individual and a mean of said at least one site model;

a feature based on a distance between the assessed location of the individual and a mean of said at least one query model;

a feature based on a distance between the assessed location of the individual and a nearest mixture component of said at least one site model; and

a feature based on a distance between the assessed location of the individual and a nearest mixture component of said at least one query model.

5. The training system of claim 3, wherein the group of location-based features includes a feature based on a probability that the individual has selected a particular site or issued a particular query, given the assessed location of the individual.

6. The training system of claim 3, wherein the group of location-based features includes a feature based on a probability of the assessed location of the individual, given that the individual selects a particular site or issues a particular query.

7. The training system of claim 3, wherein the group of location-based features includes at least one of:

a feature based on a divergence of said at least one site model from a background site model, the background site model describing a distribution of locations for plural selections of sites; and

a feature based on a divergence of said at least one query model from a background query model, the background query model describing a distribution of locations for plural selection of queries.

8. The training system of claim 3, wherein the group of location-based features includes a feature based on uncertainty associated with the assessed location of the individual, given that the individual selects a particular site or issues a particular query.

9. The training system of claim 3, further comprising a ranking model generation module for generating at least one ranking model based, in part, on at least the location-based features, together with labels applied to the user online activity data.

10. The training system of claim 1, further comprising functionality for:

using said at least one query model to identify at least one of;

one or more queries that are sensitive to location; and

one or more queries that not sensitive to location; and

producing a dataset based on said using said at least one query model.

11. A computer readable storage medium for storing computer readable instructions, the computer readable instructions providing a query processing system when executed by one or more processing devices, the computer readable instructions comprising:

logic configured to receive a query from an end user;

logic configured to associate the query with an assessed location;

logic configured to generate a group of query-time features based, in part, on at least one item model, said at least one item model estimating a probabilistic distribution of locations for an individual, given that the individual selects a particular item; and

logic configured to use at least one ranking model, together with the query-time features, to provide at least one recommended item that is assessed as being suitable for the end user, given the assessed location that is associated with the end user.

12. The computer-readable storage medium of claim 11, wherein the query-time features include a first group of general-purpose features and a second group of location-based features, said logic configured to use said at least one ranking model comprising:

logic configured to use a general-purpose ranking model, together with the first group of general-purpose features, to generate a candidate list of one or more recommended items; and

logic configured to use a location-based ranking model, together with the second group of location-based features, to re-rank said one or more recommended items in the candidate list.

13. A method, implemented using computing functionality, for providing a personalized service, comprising:

receiving user selection data which defines selections of items by a group of data-providing users;

associating each instance of the user selection data with a metadata observation, to provide metadata-tagged data;

storing the metadata-tagged data in a data store;

generating at least one item model based on the metadata-tagged data, said at least one item model estimating a probabilistic distribution of metadata observations associated with an individual, given that the individual selects a particular item;

storing said at least one item model in a data store; and

applying said at least one item model to provide the personalized service to an end user,

said receiving, associating, storing the metadata-tagged data, generating, storing said at least one item model, and applying being performed by the computing functionality.

14. The method of claim 13, further comprising associating each instance of the metadata-tagged data with a weight that indicates a reliability of the instance of metadata-tagged data.

15. The method of claim 13, wherein:

the metadata observation that is associated with each instance of the user selection data comprises a location, and

said at least one item model estimates a probabilistic distribution of locations associated with the individual, given that the individual selects the particular item.

16. The method of claim 15, wherein:

the items selected by the group of data-providing users comprise sites, and

said at least one item model comprises a site model that estimates a probabilistic distribution of locations associated with the individual, given that the individual selects a particular site.

17. The method of claim 15, wherein:

the items selected by the group of data-providing users comprise queries, and

said at least one item model comprises a query model that estimates a probabilistic distribution of locations associated with the individual, given that the individual issues a particular query.

18. The method of claim 15, wherein said at least one item model comprises a compact model that is represented by a set of model parameters.

19. The method of claim 15, wherein said at least one item model conveys the probabilistic distribution of locations on a discrete region-by-region basis.

20. The method of claim 19, further comprising associating a level of uncertainty associated with each region.