Content targeting using content context and user propensity
Disclosed herein are techniques for machine-learning systems and methods for generating content objects using AI models. A method described herein includes predicting a propensity metric using a machine-learning propensity model describing a propensity of a user to interact with a tag. The method includes generating, using a content-tagging machine-learning model, a set of features characterizing the content object. The method includes determining, for each user in a set of users, a score that predicts a propensity of the user interacting with a particular content object. The method includes selecting a subset of users of the set of users based on the scores determined for the set of users. The method also includes facilitating output of the particular content object to each of the subset of users.
1 . A computer-implemented method comprising:
accessing, for each user of a set of users, a user metadata that corresponds to characteristics of the user;
predicting, for each user of the set of users and for each content tag of one or more content tags, a propensity metric, wherein:
the propensity metric for each user of the set of users are generated by processing the user metadata using a machine-learning propensity model that is configured as a non-binary classifier trained on a training set,
the training set comprises a plurality of content objects labeled with tags, a plurality of user metadata paired to the tags and defining a ground truth for training the machine-learning propensity model that is configured to map a user metadata of the plurality of user metadata to the one or more content tags, and
the propensity metric indicates a probability of each of the user of the set of users to interact with the plurality of content objects associated with one or more content tags;
accessing a particular content object of the plurality of content objects that comprises image data or text data or both;
generating, using a content-tagging machine-learning model, a first set of qualitative features characterizing elements of the particular content object, wherein the content-tagging machine-learning model is a classifier model;
determining, for each user in the set of users, a score that predicts a propensity of the user interacting with the particular content object, the score being based at least on the propensity metric generated by the machine-learning propensity model associated with the user and the first set of qualitative features generated by the content-tagging machine learning model;
selecting a first subset of users of the set of users based on the scores determined for the set of users;
generating, when a predicted level of interaction of the first subset of users with the particular content object is below a threshold level, a new content object having a second set of qualitative features by processing at least part of the content object by adding, editing or removing visual elements and/or text corresponding to the first set of qualitative features using a generator neural network, the new content object including image data, wherein:
the generator neural network is trained to reference the machine-learning propensity model in accordance with mapping generated by the machine-learning propensity model, and
the generator neural network is configured to generate content objects comprising one or more features of the set of qualitative features with which users are predicted to interact based on one or more results generated by the machine-learning propensity model;
determining, by a discriminator neural network, authenticity of the second set of qualitative features associated with the image data, wherein the authenticity indicates whether the second set of qualitative features corresponds to a recognizable or realistic content object in the image data;
determining, for each user in the set of users, a new score that predicts a propensity of the user to interact with the new content object, the new score being based at least on the propensity metric associated with the user and the second set of qualitative features;
selecting a second subset of the set of users based on the new score; and
facilitating output of the new content object to each of the second subset of users.
2 . The method of claim 1 , further comprising:
accessing seed data comprising seed tuples, each seed tuple of the seed data comprising a respective content object and a respective feature corresponding to the respective content object; and
training the generator neural network and a discriminator neural network in a generative adversarial network (GAN), wherein, in the GAN, the generator neural network learns to generate content objects comprising one or more features of a feature set and the discriminator neural network learns to recognize generated content objects that are authentic.
3 . The method of claim 2 , wherein the generator neural network is a sequence-to-sequence neural network, and the discriminator neural network is a binary classifier.
4 . The method of claim 3 , wherein the method further comprises: training the machine-learning propensity model, wherein the training comprises:
accessing the training set;
inputting a user metadata of the plurality of user metadata to the machine-learning propensity model;
predicting, as an output of the machine-learning propensity model, a subset of tags of the tags;
determining a value of an objective function, wherein the objective function comprises an estimate of an error between the output of the machine-learning propensity model and the ground truth; and
determining a set of coefficients corresponding to the tags and metadata, at least in part by minimizing the objective function.
5 . The method of claim 1 , wherein processing the user metadata comprises:
identifying a subset of user characteristics using the user metadata for each user of the set of users, based on a relevance to the particular content object.
6 . The method of claim 1 , wherein the machine-learning propensity model comprises a multinomial logistic regression model.
7 . A content system comprising:
one or more data processors; and
a non-transitory computer readable storage medium containing instructions which, when executed on the one or more data processors, cause the one or more data processors to perform a set of actions including:
accessing, for each user of a set of users, a user metadata that corresponds to characteristics of the user;
predicting, for each user of the set of users and for each content tag of one or more content tags, a propensity metric, wherein:
the propensity metric for each user of the set of users are generated by processing the user metadata using a machine-learning propensity model that is configured as a non-binary classifier trained on a training set,
the training set comprises a plurality of content objects labeled with tags, a plurality of user metadata paired to the tags and defining a ground truth for training the machine-learning propensity model that is configured to map a user metadata of the plurality of user metadata to the one or more content tags, and
the propensity metric indicates a probability of each of the user of the set of users to interact with the plurality of content objects associated with one or more content tags;
accessing a particular content object of the plurality of content objects that comprises image data or text data or both;
generating, using a content-tagging machine-learning model, a first set of qualitative features characterizing elements of the particular content object, wherein the content-tagging machine-learning model is a classifier model;
determining, for each user in the set of users, a score that predicts a propensity of the user interacting with the particular content object, the score being based at least on the propensity metric generated by the machine-learning propensity model associated with the user and the first set of qualitative features generated by the content-tagging machine learning model;
selecting a first subset of users of the set of users based on the scores determined for the set of users;
generating, when a predicted level of interaction of the first subset of users with the particular content object is below a threshold level, a new content object having a second set of qualitative features by processing at least part of the content object by adding, editing or removing visual elements and/or text corresponding to the first set of qualitative features using a generator neural network, the new content object including image data, wherein:
the generator neural network is trained to reference the machine-learning propensity model in accordance with mapping generated by the machine-learning propensity model, and
the generator neural network is configured to generate content objects comprising one or more features of the set of qualitative features with which users are predicted to interact based on one or more results generated by the machine-learning propensity model;
determining, by a discriminator neural network, authenticity of the second set of qualitative features associated with the image data, wherein the authenticity indicates whether the second set of qualitative features corresponds to a recognizable or realistic content object in the image data;
determining, for each user in the set of users, a new score that predicts a propensity of the user to interact with the new content object, the new score being based at least on the propensity metric associated with the user and the second set of qualitative features;
selecting a second subset of the set of users based on the new score; and
facilitating output of the new content object to each of the second subset of users.
8 . The content system of claim 7 , wherein the instructions, when executed on the one or more data processors, further cause the one or more data processors to perform actions comprising:
accessing seed data comprising seed tuples, each seed tuple comprising a respective content object and a respective feature corresponding to the respective object; and
training the generator neural network and a discriminator neural network in a generative adversarial network (GAN), wherein, in the GAN, the generator neural network learns to generate content objects comprising one or more features of a feature set and the discriminator neural network learns to recognize generated content objects that are authentic.
9 . The content system of claim 8 , wherein the generator neural network is a sequence-to-sequence neural network, and the discriminator neural network is a binary classifier.
10 . The content system of claim 7 , wherein processing the user metadata comprises:
identifying a subset of user characteristics using the user metadata for each user of the set of users, based on a relevance to the content object.
11 . The content system of claim 7 , wherein the machine-learning propensity model comprises a multinomial logistic regression model.
12 . A non-transitory computer-readable memory storing instructions that, when executed by one or more data processors of a computer system, cause the one or more data processors to perform a set of actions including:
accessing, for each user of a set of users, a user metadata that corresponds to characteristics of the user;
predicting, for each user of the set of users and for each content tag of one or more content tags, a propensity metric, wherein:
the propensity metric for each user of the set of users are generated by processing the user metadata using a machine-learning propensity model that is configured as a non-binary classifier trained on a training set,
the training set comprises a plurality of content objects labeled with tags, a plurality of user metadata paired to the tags and defining a ground truth for training the machine-learning propensity model that is configured to map a user metadata of the plurality of user metadata to the one or more content tags, and
the propensity metric indicates a probability of each of the user of the set of users to interact with the plurality of content objects associated with one or more content tags;
accessing a particular content object of the plurality of content objects that comprises image data or text data or both;
generating, using a content-tagging machine-learning model, a first set of qualitative features characterizing elements of the particular content object, wherein the content-tagging machine-learning model is a classifier model;
determining, for each user in the set of users, a score that predicts a propensity of the user interacting with the particular content object, the score being based at least on the propensity metric generated by the machine-learning propensity model associated with the user and the first set of qualitative features generated by the content-tagging machine learning model;
selecting a first subset of users of the set of users based on the scores determined for the set of users;
generating, when a predicted level of interaction of the first subset of users with the particular content object is below a threshold level, a new content object having a second set of qualitative features by processing at least part of the content object by adding, editing or removing visual elements and/or text corresponding to the first set of qualitative features using a generator neural network, the new content object including image data wherein:
the generator neural network is trained to reference the machine-learning propensity model in accordance with mapping generated by the machine-learning propensity model, and
the generator neural network is configured to generate content objects comprising one or more features of the set of qualitative features with which users are predicted to interact based on one or more results generated by the machine-learning propensity model;
determining, by a discriminator neural network, authenticity of the second set of qualitative features associated with the image data, wherein the authenticity indicates whether the second set of qualitative features corresponds to a recognizable or realistic content object in the image data;
determining, for each user in the set of users, a new score that predicts a propensity of the user to interact with the new content object, the new score being based at least on the propensity metric associated with the user and the second set of qualitative features;
selecting a second subset of the set of users based on the new score; and
facilitating output of the new content object to each of the second subset of users.
13 . The computer-readable memory of claim 12 , further comprising:
accessing seed data comprising seed tuples, each seed tuple comprising a respective content object and a respective feature corresponding to the respective object; and
training the generator neural network and a discriminator neural network in a generative adversarial network (GAN), wherein, in the GAN, the generator neural network learns to generate content objects comprising one or more features of a feature set and the discriminator neural network learns to recognize generated content objects that are authentic.
14 . The computer-readable memory of claim 13 , wherein the generator neural network is a sequence-to-sequence neural network, and the discriminator neural network is a binary classifier.
15 . The computer-readable memory of claim 12 , wherein processing the user metadata comprises:
identifying a subset of user characteristics using the user metadata for each user of the set of users, based on a relevance to the particular content object.