# US 8,543,518 B2 · Deducing shadow user profiles for ad campaigns

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 8,543,518 B2
- Document ID: US8543518B2
- Status: Granted
- Kind: B2
- Application number: 12/843,762
- Filed: 2010-07-26
- Published: 2012-01-26
- Granted: 2013-09-24
- Earliest claimed: 2010-07-26
- Assignees: Yahoo! Inc.
- Applicants: Amit Kumar Verma; Viswanathan Ramaiyer; Hastagiri Prakash
- Inventors: Amit Kumar Verma; Viswanathan Ramaiyer; Hastagiri Prakash
- Web page: https://patentgrep.com/patent/US8543518B2

## CPC

- G06Q 30/0241: Advertisements
- G06Q 30/02: Marketing; Price estimation or determination; Fundraising
- G06Q 30/0201: Market modelling; Market analysis; Collecting market data
- G06Q 30/0251: Targeted advertisements
- G06Q 30/0269: {based on user profile or attribute}
- G06Q 30/0277: {Online advertisement}
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR

## Abstract

A method and a system are provided for deducing shadow user profile attributes for ad campaigns aimed at target users. In one example, the system extracts tagged data from source data. The tagged data includes label information associated with an actual profile for a user. The tagged data is associated with the user. The system prepares the tagged data by splitting the tagged data into datasets, including at least training data and test data. The system generates one or more individual models based on the tagged data, wherein the one or more individual models provide the ability to deduce attributes of a profile for the user. The system then generates a composite model based on the individual models. The composite model includes a combination of the individual models that are associated with the user. The system may charge a premium for ad campaigns that are aimed at target users who are each assigned one or more shadow profile attribute values. The system may determine the premium based on the confidence level with which the one or more attribute values fits to the one or more users. The system is applicable to both display advertising and sponsored search advertising.

## Claims

### Claim 1

A computer-implemented method for deducing one or more user profile attributes using a shadow profile, the method comprising:
extracting, at a computer, tagged data from source data, wherein the tagged data comprises label information associated with an actual profile for a user, and wherein the tagged data is associated with the user;
preparing, at a computer, the tagged data by splitting the tagged data into datasets, including at least training data and test data;
generating, at a computer, one or more individual models based on the tagged data, wherein the one or more individual models provide the ability to deduce attributes of a profile for the user;
testing, at a computer, the one or more individual models using the test data to determine a confidence level representing the predicted accuracy of the individual model;
generating, at a computer, a composite model based on the individual models, wherein the composite model comprises a combination of the individual models that are associated with the user and achieve a confidence level which is greater than a predefined threshold value;
generating, at a computer, a plurality of values associated with a user profile attribute using the composite model;
storing, at a computer, the plurality of values associated with a user profile attribute in a shadow profile associated with the user, wherein the shadow profile comprises a plurality of attributes, each attribute comprising at least one value and at least one confidence level;
updating, at a computer, the at least one confidence level of the shadow profile based on new or modified label information associated with the actual profile; and
exposing, at a computer, at least one attribute associated with a shadow profile to facilitate the selection of an advertisement to present to the user associated with the shadow profile.

### Claim 2 (depends on 1)

The method of claim 1, further comprising at least one of:
receiving, at a computer, additional tagged data, including tuning data for model tuning;
validating, at a computer, the one or more individual models;
validating, at a computer, the composite model;
tuning, at a computer, the one or more individual models; and
tuning, at a computer, the composite model.

### Claim 3 (depends on 1)

The method of claim 1, wherein the source data includes at least one of:
search history;
browsing pattern;
item-click history;
group memberships;
user-expressed interests;
mail records; and
instant message records.

### Claim 4 (depends on 1)

The method of claim 1, wherein the tagged data includes information associated with at least one of:
gender;
age group; and
income level.

### Claim 5 (depends on 2)

The method of claim 2, wherein the tagged data includes total tagged data, and wherein the training data includes about 80% of the total tagged data, and wherein the test data includes about 10% of the total tagged data, and wherein the tuning data includes about 10% of the total tagged data.

### Claim 6 (depends on 1)

The method of claim 1, wherein generating one or more individual models includes at least one of:
selecting, at a computer, categories that contribute to a predefined threshold percentage of total data;
defining, at a computer, data points of a classifier based on the categories;
generating, at a computer, an output of the classifier, wherein the output is a deduced profile for the user; and
generating, at a computer, one or more individual models by using machine learning techniques.

### Claim 7 (depends on 1)

The method of claim 1, wherein generating one or more individual models includes using, at a computer, a machine learning technique that includes at least one of:
one or more decision trees;
AdaBoost; and
support vector machines.

### Claim 8 (depends on 1)

The method of claim 1, wherein generating one or more individual models includes training, at a computer, the one or more individual models by using training data, wherein training includes training the one or more individual models to form associations between source data and profile attributes.

### Claim 9 (depends on 1)

The method of claim 1, wherein generating the composite model includes at least one of:
assigning, at a computer, a weight to each output of each individual model, wherein each weight is based on a training error for each individual model;
assigning, at a computer, a lesser weight to individual models with higher training error; and
assigning, at a computer, a greater weight to individual models with lower training error.

### Claim 10 (depends on 1)

The method of claim 1, further comprising updating, at a computer, the composite model by using an incremental learning technique.

### Claim 11

A system for deducing one or more user profile attributes, the system comprising:
a computer system configured for:
extracting tagged data from source data using a shadow profile, wherein the tagged data comprises label information associated with an actual profile for a user, and wherein the tagged data is associated with the user;
preparing the tagged data by splitting the tagged data into datasets, including at least training data and test data;
generating one or more individual models based on the tagged data, wherein the one or more individual models provide the ability to deduce attributes of a profile for the user;
testing the one or more individual models using the test data to determine a confidence level representing the predicted accuracy of the individual model;
generating a composite model based on the individual models, wherein the composite model comprises a combination of the individual models that are associated with the user and achieve a confidence level which is greater than a predefined threshold value;
generating a plurality of values associated with a user profile attribute using the composite model;
storing the plurality of values associated with a user profile attribute in a shadow profile associated with the user, wherein the shadow profile comprises a plurality of attributes, each attribute comprising at least one value and at least one confidence level;
updating the at least one confidence level of the shadow profile based on new or modified label information associated with the actual profile; and
exposing at least one attribute associated with a user profile to facilitate the selection of an advertisement to present to the user.

### Claim 12 (depends on 11)

The system of claim 11, wherein the computer system is further configured for:
receiving additional tagged data, including tuning data for model tuning;
validating the one or more individual models;
validating the composite model;
tuning the one or more individual models; and
tuning the composite model.

### Claim 13 (depends on 11)

The system of claim 11, wherein the source data includes at least one of:
search history;
browsing pattern;
item-click history;
group memberships;
user-expressed interests;
mail records; and
instant message records.

### Claim 14 (depends on 11)

The system of claim 11, wherein the tagged data includes information associated with at least one of:
gender;
age group; and
income level.

### Claim 15 (depends on 12)

The system of claim 12, wherein the tagged data includes total tagged data, and wherein the training data includes about 80% of the total tagged data, and wherein the test data includes about 10% of the total tagged data, and wherein the tuning data includes about 10% of the total tagged data.

### Claim 16 (depends on 11)

The system of claim 11, wherein generating one or more individual models further configures the computer system for at least one of:
selecting categories that contribute to a predefined threshold percentage of total data;
defining data points of a classifier based on the categories;
generating an output of the classifier, wherein the output is a deduced profile for the user; and
generating one or more individual models by using machine learning techniques.

### Claim 17 (depends on 11)

The system of claim 11, wherein generating one or more individual models further configures the computer system for using a machine learning technique that includes at least one of:
one or more decision trees;
AdaBoost; and
support vector machines.

### Claim 18 (depends on 11)

The system of claim 11, wherein generating one or more individual models further configures the computer system for training the one or more individual models by using training data, wherein training includes training the one or more individual models to form associations between source data and profile attributes.

### Claim 19 (depends on 11)

The system of claim 11, wherein generating the composite model further configures the computer system for at least one of:
assigning a weight to each output of each individual model, wherein each weight is based on a training error for each individual model;
assigning a lesser weight to individual models with higher training error; and
assigning a greater weight to individual models with lower training error.

### Claim 20 (depends on 11)

The system of claim 11, wherein the computer system is further configured for updating the composite model by using an incremental learning technique.

### Claim 21

A non-transitory computer readable storage medium comprising one or more instructions for deducing one or more user profile attributes, wherein the one or more instructions are configured for causing one or more processors to perform the steps of:
extracting tagged data from source data, wherein the tagged data comprises label information associated with an actual profile for a user, and wherein the tagged data is associated with the user;
preparing the tagged data by splitting the tagged data into datasets, including at least training data and test data;
generating one or more individual models based on the tagged data, wherein the one or more individual models provide the ability to deduce attributes of a profile for the user;
testing the one or more individual models using the test data to determine a confidence level representing the predicted accuracy of the individual model;
generating a composite model based on the individual models, wherein the composite model comprises a combination of the individual models that are associated with the user and achieve a confidence level which is greater than a predefined threshold value;
generating a plurality of values associated with a user profile attribute using the composite model;
storing the plurality of values associated with a user profile attribute in a shadow profile associated with the user, wherein the shadow profile comprises a plurality of attributes, each attribute comprising at least one value and at least one confidence level;
updating the at least one confidence level of the shadow profile based on new or modified label information associated with the actual profile; and
exposing at least one attribute associated with a user profile to facilitate the selection of an advertisement to present to the user.