Deep learning models processing time series data

A method of Artificial Intelligence (AI) data processing, comprising:

• accessing historical data including a plurality of time series data sets collected at different time steps for different entities of an entity hierarchy, wherein each of the plurality of time series data sets is associated with a corresponding variable;
• generating a plurality of transformed time series data sets for each of the plurality of time series data sets,
• wherein the corresponding variable is associated with a time series data set of the plurality of time series data sets can be expressed in the transformed time series data set in terms of other corresponding variables associated with other time series data sets of the plurality of time series data sets;
• training deep learning (DL) models including a primary model and an extended model forming a set of sequential models to generate forecasts for the corresponding variables,
• wherein the primary model is trained using first-party data from local data sources that include the plurality of time series data sets, and the primary model provides an output using the plurality of transformed time series data sets, and
• wherein the extended model is a long short term memory (LSTM) model which is trained via supervised methods on third-party data that is different than the plurality of time series data sets, and the extended model receives the output of the primary model in addition to the third-party data for training and to generate the forecasts;
• receiving a user request for the forecasts for the corresponding variables,
• wherein the user request refers to one or more of the different entities, and one or more predetermined future time steps;
• providing values related to the one or more different entities and the one or more predetermined future time steps from the user request to the DL models;
• outputting the forecasts for the corresponding variables for the predetermined future time steps via a graphical user interface (GUI) based on the user request;
• and
• automatically updating the DL models on new data associated with one or more of the corresponding variables,
• wherein the DL models are configured for automatic updates based on each time one of the plurality of time series data sets or the first-party data is updated for the primary model, and
• wherein the DL models are configured for automatic updates based on any portion of the third-party data is updated for the extended model,
• wherein the method further comprising:
• determine whether processing by the extended model is necessary;
• in response to a determination that processing by the extended model is not necessary, output the output of the primary model via the GUI without accessing the third-party data, wherein the forecasts are generated based the plurality of transformed time series data sets;
• and
• in response to a determination that processing by the extended model is necessary, access the third-party data and provide the output from the primary model along with the third-party data to the extended model to generate the forecasts, wherein the forecasts enable visualization of what-if scenarios via the GUI in one or more of textual, graphical, or image modalities.

A non-transitory processor-readable storage medium comprising machine-readable instructions that cause a processor to:

• access historical data including a plurality of time series data sets generated at different time steps for different entities of an entity hierarchy, wherein each of the plurality of time series data sets is associated with a corresponding variable;
• generate a plurality of transformed time series data sets for each of the plurality of time series data sets, wherein the corresponding variable associated with one time series data set of the plurality of time series data sets can be expressed in the transformed time series data set in terms of other corresponding variables associated with other time series data sets of the plurality of time series data sets;
• train deep learning (DL) models including a primary model and an extended model forming a set of sequential models to generate forecasts for the corresponding variables,
• wherein the primary model is trained using first-party data from local data sources that include the plurality of time series data sets, and the primary model provides an output using the plurality of transformed time series data sets, and
• wherein the extended model is a long short term memory (LSTM) model which is trained via supervised methods on third-party data that is different than the plurality of time series data sets, and the extended model receives the output of the primary model in addition to the third-party data for training and to generate the forecasts;
• receive a user request for the forecasts for the corresponding variables, wherein the user request refers to one or more of the different entities, and a predetermined future time period including one or more predetermined future time steps;
• provide values related to the one or more different entities and the one or more future time steps from the user request to the DL models;
• output the forecasts for the corresponding variables for the predetermined future time steps via a graphical user interface (GUI) based on the user request;
• and
• automatically update the DL models on new data associated with one or more of the corresponding variables,
• wherein the DL models are configured for automatic updates based on each time one of the plurality of time series data sets or the first-party data is updated for the primary model, and
• wherein the DL models are configured for automatic updates based on any portion of the third-party data is updated for the extended model,
• wherein the machine-readable instructions further cause the processor to:
• determine whether processing by the extended model is necessary;
• in response to a determination that processing by the extended model is not necessary, output the output of the primary model via the GUI without accessing the third-party data, wherein the forecasts are generated based the plurality of transformed time series data sets;
• and
• in response to a determination that processing by the extended model is necessary, access the third-party data and provide the output from the primary model along with the third-party data to the extended model to generate the forecasts, wherein the forecasts enable visualization of what-if scenarios via the GUI in one or more of textual, graphical, or image modalities.