US 10,832,167 B2Granted

Interest prediction for unresolved users in an online system

AssigneeFacebook, Inc.

InventorsGoran Predovic, Chaochao Cai

Life of this patent

Estimating the term…

Loading patent…

Abstract

Disclosed is an online system that infers interests of unresolved users for whom the interests are not known. The online system determines certain features about the unresolved users, but does not have certain information about the users themselves (e.g., their interests), so instead infers these attributes based on the features of the user. The online system provides the features as input to a classifier trained to predict a particular interest, and the classifier outputs a prediction of whether the user has the corresponding interest. In one embodiment, the online system trains a classifier for various interest values by forming training sets for the interests using the features for users who are logged into the online system and hence have known interests.

Claims15

Dependent claims

A computer-implemented method performed by an online system, the method comprising:

receiving content items from content providers for display to users of the online system, each content item associated with an audience defining interests of users to whom the content item is to be provided;
for each interest of one or more interests:
forming a training set of users having the interest in their user profile in the online system,
deriving features of the users of the training set, the features indicating which applications the users have installed on client devices of the users, each feature represented by a feature value, and
training a classifier to predict the interest for a user by providing the derived features as input to a machine learning algorithm;
responsive to detecting an opportunity to provide one of the received content items to an unresolved user who the online system is unable to match to a user profile of the online system, applying one or more of the trained classifiers to predict one or more interests for the unresolved user by:
deriving features of the unresolved user, the features indicating which applications the unresolved user has installed on a client device of the unresolved user, each feature represented by a feature value,
providing the derived features as input to one of the trained classifiers, and
obtaining, as an output from the trained classifier, a prediction of whether the unresolved user has an interest corresponding to the trained classifier, wherein obtaining the prediction comprises:
computing averaged feature values by averaging the feature values of the users of the training set,
computing a degree of similarity of the averaged feature values of the users of the training set and the feature values of the unresolved user by computing a dot product of the averaged feature values of the users of the training set and the feature values of the unresolved user, and
responsive to determining that the degree of similarity is above a threshold value, predicting that the unresolved user has the interest;
and
based on the predicted interests of the unresolved user, providing for display to the unresolved user a content item having an audience that includes at least one of the predicted interests for the unresolved user.

11 more elements

A computer-implemented method performed by an online system, the computer-implemented method comprising:

training a classifier using a training set of users having an interest known by the online system, the training set comprising feature values for a first set of features for the users of the training set and the classifier configured to predict an interest of an unresolved user of the online system for whom the online system does not have information about;
for the unresolved user, determining feature values for the first set of features for the unresolved user;
obtaining, as an output of the trained classifier applied to the feature values for the unresolved user, a prediction of whether the unresolved user has the first interest, wherein obtaining the prediction comprises:
computing averaged feature values by averaging feature values for the users of the training set,
computing a degree of similarity of the averaged feature values for the users of the training set and the feature values for the unresolved user by computing a dot product of the averaged feature values of the users of the training set and the feature values of the unresolved user, and
responsive to determining that the degree of similarity is above a threshold value, predicting that the unresolved user has the interest;
and
selecting content to provide to the unresolved user based on the prediction.

5 more elements

A non-transitory computer-readable storage medium having instructions that when executed by a processor of an online system perform actions comprising:

training a classifier using a training set of users having an interest known by the online system, the training set comprising feature values for a first set of features for the users of the training set and the classifier configured to predict an interest of an unresolved user of the online system for whom the online system does not have information about;
for the unresolved user, determining feature values for the first set of features for the unresolved user;
obtaining, as an output of the trained classifier applied to the feature values for the unresolved user, a prediction of whether the unresolved user has the first interest, wherein obtaining the prediction comprises:
computing averaged feature values by averaging feature values for the users of the training set,
computing a degree of similarity of the averaged feature values for the users of the training set and the feature values for the unresolved user by computing a dot product of the averaged feature values of the users of the training set and the feature values of the unresolved user, and
responsive to determining that the degree of similarity is above a threshold value, predicting that the unresolved user has the interest;
and
selecting content to provide to the unresolved user based on the prediction.

5 more elements

Dependencies are inferred from the claim text.