US 10,380,501 B2Granted

Lookalike evaluation

AssigneeFacebook, Inc.

InventorsHaibin Cheng, Xian Xu, Yang Pei

Life of this patent

Estimating the term…

Loading patent…

Abstract

Lookalike models can select users that are predicted to share characteristics with a specified set of seed users. The processing requirements for lookalike models can be decreased by identifying features that have low impact on model accuracy, and therefore can be excluded from creating models. Also, by identifying preferred seed sources and training parameters, accurate lookalike models can be created with less overhead and in less time. The features and training parameters can be identified by obtaining a sample seed set, extracting seeds with a defined set of features, and using the remaining training seeds to train a model. Performance of this model can be compared to a standard model to see if the model performs well. If so, features excluded from the features used to create the model, a seed source, or training parameters used to create the model can be selected.

Claims20

Dependent claims

A method for selecting preferred seed sources, comprising:

obtaining multiple seed groups, each seed group including one or more users of an online system, from multiple seed sources, each seed source including users of the online system having at least one common feature;
selecting subsets of the seed groups as training seed groups;
generating unions from the training seed groups by selecting training seed groups that have common evaluation data, each union including two or more training seed groups including users having at least one common feature;
using two or more of the unions that include training seed groups having at least one different feature to train two or more corresponding lookalike models, wherein different lookalike models are trained using training seed groups having at least one different feature;
comparing performances of two or more lookalike models each trained using different unions that each include training seed groups having at least one different feature;
selecting, based on the comparisons, training seed groups included one or more unions corresponding to lookalike models having a performance above a threshold level as preferred seed sources;
and
training one or more additional lookalike models using the selected preferred seed sources.

5 more elements

A system for selecting preferred training parameters, comprising:

one or more processors;
a memory;
a model parameter tester, executed by the one or more processors, that performs operations comprising:
obtaining multiple seed groups from multiple seed sources, each seed group including one or more users of an online system and each seed source including users of the online system having at least one common feature;
creating subsets of the seed groups as training seed groups;
training multiple lookalike models using different training seed groups each of the multiple lookalike models is trained with a training seed group having at least one different feature from training seed groups used to train other lookalike models of the multiple lookalike models;
comparing performances of each of the lookalike models trained using training seed groups having at least one different feature from other training seed groups;
selecting one or more lookalike models having a performance above a threshold level;
for each selected lookalike model of one or more of the multiple trained lookalike models:
scoring the at least one feature of the training seed group used to train the selected lookalike model that is different from features of training seed groups used to train the other lookalike models of the multiple lookalike models;
identifying the preferred training parameters based on a comparison of results of the scoring;
and
a model builder that builds one or more additional lookalike models using the preferred training parameters.

10 more elements

A computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to perform operations for selecting preferred seed features, the operations comprising:

obtaining multiple seed groups from multiple seed sources, each seed group including one or more users of an online system and each seed source including users of the online system having at least one common feature;
creating subsets of the seed groups as training seed groups;
training a plurality of lookalike models using different training seed groups, each of the plurality of lookalike models trained with
a training seed group having at least one different feature from training seed groups used to train other lookalike models;
comparing performances of each of the lookalike models trained using training seed groups having at least one different feature from other training seed groups;
selecting one or more lookalike models having a performance above a threshold level;
for each selected lookalike model of the plurality of lookalike models:
scoring the at least one feature of the training seed group used to train the selected lookalike model that is different from features of training seed groups used to train other lookalike models of the plurality of lookalike models;
identifying the preferred seed features based on a comparison of scores for features of different training seed groups used to train at least one selected lookalike model;
and
training one or more additional lookalike models using the identified preferred seed features.

8 more elements

Dependencies are inferred from the claim text.