| CPC G06N 3/084 (2013.01) [G06N 3/02 (2013.01); G06N 3/042 (2023.01); G06N 3/0455 (2023.01); G06N 3/0464 (2023.01); G06N 3/0475 (2023.01); G06N 3/094 (2023.01); G06N 20/00 (2019.01); G06T 7/20 (2013.01); G06T 2207/20084 (2013.01)] | 15 Claims |
|
1. A method for automated prediction of user data, comprising:
a) obtaining, by at least one processing unit, user action features, represented as an array of first vectors;
b) obtaining, by at least one processing unit, user features, represented as an array of second user feature vectors;
c) creating a predictive model of actions, the predictive model of actions comprising:
at least one user action features encoder made in the form of at least one neural network model;
at least one user feature processing model, wherein said at least one user feature processing model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
at least one user prediction model, wherein said at least one user prediction model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
d) training the predictive model of actions using an error backpropagation method, including the following steps;
feeding, by at least one processing unit, the array of first vectors to the user action features encoder to generate at least one first latent state action feature vector of at least one user;
feeding, by at least one processing unit, the array of second user feature vectors to the feature processing model to generate at least one second latent state user feature vector of at least one user;
concatenating, by at least one processing unit, said first latent state action feature vector and second latent state user feature vector;
feeding, by at least one processing unit, the concatenation of first latent state action feature vector and second latent state user feature vector to the at least one user prediction model to generate a predicted value of the user data of at least one user;
optimizing, by at least one processing unit, a difference between the predicted value of the user data of at least one user and a respective actual value of the user data of at least one user using an error backpropagation method for all components of the predictive model of actions, wherein the respective actual value of the user data is obtained from the user action features;
e) performing automated prediction of the user data by means of the obtained predictive model of actions of at least one user.
|
|
6. An automated user data prediction system comprising one or more processors and one or more storage devices storing instructions that when executed by the one or more processors cause the one or more processors to perform operations comprising:
a) obtaining, by at least one processing unit, user action features, represented as an array of first vectors;
b) obtaining, by at least one processing unit, user features, represented as an array of second user feature vectors;
c) creating a predictive model of actions, the predictive model of actions comprising:
at least one user action features encoder made in the form of at least one neural network model;
at least one user feature processing model, wherein said at least one user feature processing model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
at least one user prediction model, wherein said at least one user prediction model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
d) training the predictive model of actions using an error backpropagation method, including the following steps;
feeding, by at least one processing unit, the array of first vectors to the user action features encoder to generate at least one first latent state action feature vector of at least one user;
feeding, by at least one processing unit, the array of second user feature vectors to the feature processing model to generate at least one second latent state user feature vector of at least one user;
concatenating, by at least one processing unit, said first latent state action feature vector and second latent state user feature vector;
feeding, by at least one processing unit, the concatenation of first latent state action feature vector and second latent state user feature vector to the at least one prediction model to generate a predicted value for the user data of at least one user;
optimizing, by at least one processing unit, a difference between the predicted value for the user data of at least one user and a respective actual value of user data of at least one user using an error backpropagation method for all components of the predictive model of actions, wherein the respective actual value of user data is obtained from the user action features;
e) performing automated prediction of the user data by means of the obtained predictive model of actions of at least one user.
|
|
11. A non-transitory computer-readable storage medium containing instructions to be executed by at least one processing unit, which, when executed by said at least one processing unit, cause automated prediction of user data through at least the following steps:
a) obtaining, by at least one processing unit, user action features, represented as an array of first vectors;
b) obtaining, by at least one processing unit, user features, represented as an array of second user feature vectors;
c) creating a predictive model of actions, the predictive model of actions comprising:
at least one user action features encoder made in the form of at least one neural network model;
at least one user feature processing model, wherein said at least one user feature processing model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
at least one user prediction model, wherein said at least one user prediction model is made in the form of at least one neural network model and configured to dynamically select an architecture depending on one or more of: dimensions of the input data, distribution of features contained in the input data;
d) training the predictive model of actions using an error backpropagation method, including the following steps;
feeding, by at least one processing unit, the array of first vectors to the user action features encoder to generate at least one first latent state action feature vector of at least one user;
feeding, by at least one processing unit, the array of second user feature vectors to the feature processing model to generate at least one second latent state user feature vector of at least one user;
concatenating, by at least one processing unit, said first latent state action feature vector and second latent state user feature vector;
feeding, by at least one processing unit, the concatenation of first latent state action feature vector and second latent state user feature vector to the at least one user prediction model to generate a predicted value of the user data of at least one user;
optimizing, by at least one processing unit, a difference between the predicted value of the user data of at least one user and a respective actual value of user data of at least one user using an error backpropagation method for all components of the predictive model of actions, wherein the respective actual value of user data is obtained from the user action features;
e) performing automated prediction of the user data by means of the obtained predictive model of actions of at least one user.
|