Content for targeted transmission

A computer-implemented method, comprising:

• receiving a set of user metadata for a group of users, the user metadata comprising data relating to user browsing history, purchase history, term usage history, social media posts and actions, location information, or a combination thereof;
• identifying a subset of users of the group of users, wherein the subset of users share a common attribute found in the set of user metadata;
• selecting at least one user from the subset of users;
• selecting a term from a category of terms related to a piece of digital content;
• determining a first value by comparing a first total of appearances of the term and a second total of appearances of all terms from the category of terms within the set of user metadata corresponding to the at least one user;
• determining a second value for the subset of users by determining a portion of users of the subset of users for which the term appears in the set of user metadata corresponding to the subset of users;
• determining an affinity user score based at least in part on the first value and the second value;
• randomly splitting a training dataset into a number of chunks, wherein the training dataset comprises historical user metadata associated with a first historical time range;
• iteratively training a field-aware factorization machine (FFM) model by sequentially feeding the chunks of the training dataset into the FFM model;
• applying updated coefficients to a testing dataset after each training iteration, wherein the testing dataset comprises historical user metadata associated with a second historical time range different than the first historical time range and the updated coefficients represent correlations between behaviors and different feature combinations;
• validating the FFM model after each training iteration, wherein validating the FFM model comprises inputting the testing dataset into the FFM model and determining validation metrics to evaluate the updated coefficients based on resulting outputs of the FFM model;
• terminating training of the FFM model in response to the validation metrics converging of declining across training iterations, wherein the updated coefficients associated with the highest validation metrics are retained and utilized by the trained and validated FFM model;
• determining, via the trained and validated FFM model, correlations between a desired behavior and different feature combinations;
• converting the set of user metadata into a set of features;
• determining, via the trained and validated FFM model, a machine-learning user score based at least in part on a comparison between the different feature combinations and the set of features, wherein each feature of the set of features is assigned a dynamic weight based on one or more attributes of the user metadata forming the respective feature, and the dynamic weight corresponds to an amount of influence the respective feature has on the machine-learning user score;
• and
• providing the affinity user score and the machine-learning user score to presentation logic to cause the presentation logic to provision targeted content associated with the piece of digital content to one or more electronic devices associated with the at least one user based on the affinity user score and the machine-learning user score, wherein the presentation logic is configured to provide the targeted content to the one or more electronic devices when the affinity user score, the machine-learning user score, or both are below respective upper threshold values and above respective lower threshold values, and exclude the one or more electronic devices from receiving the targeted content when the affinity user score and the machine-learning user score are outside respective predetermined ranges defined by the respective upper threshold values and the respective lower threshold values.

A non-transitory, computer-readable medium, comprising computer-readable instructions that, when executed by one or more processors of one or more computers, cause the one or more computers to:

• receive a set of user metadata for a group of users, the user metadata comprising data relating to user browsing history, purchase history, term usage history, social media posts and actions, location information, or a combination thereof;
• identify a subset of users of the group of users, wherein the subset of users share a common attribute found in the set of user metadata;
• select at least one user from the subset of users;
• select a term from a category of terms related to a piece of digital content;
• determine a first value by comparing a first total of appearances of the term and a second total of appearances of all terms from the category of terms within the set of user metadata corresponding to the at least one user;
• determine a second value for the subset of users by determining a portion of users of the subset of users for which the term appears in the set of user metadata corresponding to the subset of users;
• determine an affinity user score based at least in part on the first value and the second value;
• randomly split a training dataset into a number of chunks, wherein the training dataset comprises historical user metadata associated with a first historical time range;
• iteratively train a field-aware factorization machine (FFM) model by sequentially feeding the chunks of the training dataset into the FFM model;
• apply updated coefficients to a testing dataset after each training iteration, wherein the testing dataset comprises historical user metadata associated with a second historical time range different than the first historical time range and the updated coefficients represent correlations between behaviors and different feature combinations;
• validate the FFM model after each training iteration, wherein validating the FFM model comprises inputting the testing dataset into the FFM model and determining validation metrics to evaluate the updated coefficients based on resulting outputs of the FFM model;
• terminate training of the FFM model in response to the validation metrics converging or declining across training iterations, wherein the updated coefficients associated with the highest validation metrics are retained and utilized by the trained and validated FFM model;
• determine, via the trained and validated FFM model, correlations between a desired behavior and different feature combinations;
• convert the set of user metadata into a set of features;
• determine, via the trained and validated FFM model, a machine-learning user score based at least in part on a comparison between the different feature combinations and the set of features, wherein each feature of the set of features is assigned a dynamic weight based on one or more attributes of the user metadata forming the respective feature, and the dynamic weight corresponds to an amount of influence the respective feature has on the machine-learning user score;
• and
• provide the affinity user score and the machine-learning user score to presentation logic to cause the presentation logic to provision targeted content associated with the piece of digital content to one or more electronic devices associated with the at least one user based on the affinity user score and the machine-learning user score, wherein the presentation logic is configured to provide the targeted content to the one or more electronic devices when the affinity user score, the machine-learning score, or both are below respective upper threshold values and above respective lower threshold values, and exclude the one or more electronic devices from receiving the targeted content when the affinity user score and the machine-learning user score are outside respective predetermined ranges defined by the respective upper threshold values and the respective lower threshold values.