Customer journey management engine

1. A tangible, non-transitory, machine-readable medium storing instructions that when executed by one or more processors effectuate operations comprising:

obtaining, with one or more processors, a first training dataset, wherein:

the first training dataset comprises a plurality of subject-entity records, the subject-entity records each describe a different subject entity;

each subject entity is a different member of a first population of entities that have interacted over time with an actor entity;

each subject-entity record describes attributes of a respective subject entity among the first population;

each subject-entity record describes a time-series of events involving a respective subject entity among the first population;

the events are distinct from the attributes and the events have respective event types stored within an ontology of event types that describe both interrelatedness and similarity of events in the time series of events;

at least some of the events are question events that are caused by the actor entity, wherein the question events are asked until a level of confidence resulting in a reduced set of decisions is reached; and

at least some of the events are subject responses that are caused by a respective subject entity among the first population;

training, with one or more processors, a first machine-learning model on the first training dataset by adjusting parameters of the first machine-learning model to optimize a first objective function that indicates an accuracy of the first machine-learning model in predicting subsequent events in the time-series given prior events in the time-series and given attributes of subject entities among the first population;

obtaining, with one or more processors, a set of candidate question sequences including candidate question events, a subset of which are to be asked by the actor entity, the set including a plurality of different candidate question sequences, wherein at least some subject entities respond to candidate question events among the subset via interactive user interface elements;

forming, with one or more processors, virtual subject-entity records by appending the set of candidate question sequences to time-series of at least some of the subject-entity records, wherein:

a given subset of the virtual subject-entity records includes a plurality of virtual-subject-entity records that each include at least part of a time-series from a respective subject-entity record in the first training dataset; and

at least some of the plurality of virtual-subject-entity records in the given subset each have a different member of the set of candidate question sequences appended to the at least part of the time-series from the respective subject-entity record in the first training dataset;

forming, with one or more processors, a second training dataset by:

predicting responses of the subject entities to at least some of the set of candidate question sequences with the first machine-learning model based on the virtual subject-entity records; and

associating subject entities or attributes thereof with corresponding predicted responses in the second training dataset;

training, with one or more processors, a second machine-learning model on the second training dataset by adjusting parameters of the second machine-learning model to optimize a second objective function that indicates an accuracy of the second machine-learning model in predicting the predicted responses in the second training set given attributes of subject entities corresponding to the predicted responses;

using, with one or more processors, the adjusted parameters of the second machine-learning model to select questions for a given subject entity to reduce a knowledge gap of the actor entity regarding the given subject entity's future behavior, thereby increasing probability of occurrence of a future event involving the given subject entity; and

storing, with one or more processors, the adjusted parameters of the second machine-learning model in memory.

2. The medium of claim 1 , wherein:

at least some of subject-entity records of the plurality of subject-entity records indicate corresponding subject entities have interacted within a time range with the actor entity.

3. The medium of claim 2 , wherein the time range is a trailing time range.

4. The medium of claim 1 , wherein:

at least some of subject-entity records of the plurality of subject-entity records indicate corresponding subject entities have interacted within a geolocation range with the actor entity.

5. The medium of claim 1 , wherein:

the first machine-learning model is configured to predict responses of the plurality of subject-entity records given previous time-series of events and attributes of the plurality of subject-entity records.

6. The medium of claim 1 , wherein the operation comprise:

iterating training of a plurality of models to increase the accuracy of the model across each iteration responsive to next questions and responses.

7. The medium of claim 1 , wherein:

the first machine learning model comprises a Hidden Markov model.

8. The medium of claim 1 , wherein:

the first machine learning model comprises a long short-term memory model.

9. The medium of claim 1 , wherein:

the first machine learning model comprises a dynamic Bayesian network.

10. The medium of claim 1 , wherein:

the first machine learning model comprises a neural network classifier.

11. The medium of claim 1 , wherein:

the second machine learning model is an unsupervised model configured to translate inputs into a vector representation that maps to a candidate action.

12. The medium of claim 1 , wherein:

the second machine learning model is a random decision forest model that includes a plurality of weighted trained decision trees.

13. The medium of claim 1 , wherein:

the second machine learning model is a gradient-boosted trees model that includes a plurality of weighted trained decision trees.

14. The medium of claim 1 , wherein:

training the first machine-learning model comprises steps for training a supervised time-series forecasting model; and

training the second machine-learning model comprises steps for training a supervised classification model.

15. The medium of claim 1 , wherein:

the operations comprise steps for causing subject entities to respond in a targeted manner based on a trained model.

16. The medium of claim 1 , wherein:

the events are stored in an ontology of event types that describes interrelatedness or similarity between the events.

17. The medium of claim 1 , wherein the question events comprise:

an interactive user interface element for which a response within the interactive user interface element is collected, which may be a selection or other user input;

an advertisement, for which a response may be the interaction with a particular portion of the advertisement; and

an article about a product, for which a response may be the interaction with a particular portion of the article.

18. The medium of claim 1 , wherein at least some of the events are subject responses that are caused by the respective subject entity among the first population.

19. The medium of claim 18 , wherein the subject responses comprise:

a canceled response;

a delayed response;

a direct response; and

an indirect response.

20. The medium of claim 1 , wherein at least some of the plurality of virtual-subject-entity records in the given subset each have a different member of the set of candidate question sequences appended to the at least part of the time-series from the respective subject-entity record in the first training dataset.

21. A method comprising:

obtaining, with one or more processors, a first training dataset, wherein:

the first training dataset comprises a plurality of subject-entity records, the subject-entity records each describe a different subject entity;

each subject entity is a different member of a first population of entities that have interacted over time with an actor entity;

each subject-entity record describes attributes of a respective subject entity among the first population;

each subject-entity record describes a time-series of events involving a respective subject entity among the first population;

the events are distinct from the attributes and the events have respective event types stored within an ontology of event types that describe both interrelatedness and similarity of events in the time series of events;

at least some of the events are question events that are caused by the actor entity wherein the question events are asked until a level of confidence resulting in a reduced set of decisions is reached; and

at least some of the events are subject responses that are caused by a respective subject entity among the first population;

training, with one or more processors, a first machine-learning model on the first training dataset by adjusting parameters of the first machine-learning model to optimize a first objective function that indicates an accuracy of the first machine-learning model in predicting subsequent events in the time-series given prior events in the time-series and given attributes of subject entities among the first population;

obtaining, with one or more processors, a set of candidate question sequences including candidate question events, a subset of which are to be asked by the actor entity, the set including a plurality of different candidate question sequences, wherein at least some subject entities respond to candidate question events among the subset via interactive user interface elements;

forming, with one or more processors, virtual subject-entity records by appending the set of candidate question sequences to time-series of at least some of the subject-entity records, wherein:

a given subset of the virtual subject-entity records includes a plurality of virtual-subject-entity records that each include at least part of a time-series from a respective subject-entity record in the first training dataset; and

at least some of the plurality of virtual-subject-entity records in the given subset each have a different member of the set of candidate question sequences appended to the at least part of the time-series from the respective subject-entity record in the first training dataset;

forming, with one or more processors, a second training dataset by:

predicting responses of the subject entities to at least some of the set of candidate question sequences with the first machine-learning model based on the virtual subject-entity records; and

associating subject entities or attributes thereof with corresponding predicted responses in the second training dataset;

training, with one or more processors, a second machine-learning model on the second training dataset by adjusting parameters of the second machine-learning model to optimize a second objective function that indicates an accuracy of the second machine-learning model in predicting the predicted responses in the second training set given attributes of subject entities corresponding to the predicted responses;

using, with one or more processors, the adjusted parameters of the second machine-learning model to select questions for a given subject entity to reduce a knowledge gap of the actor entity regarding the given subject entity's future behavior, thereby increasing probability of occurrence of a future event involving the given subject entity; and

storing, with one or more processors, the adjusted parameters of the second machine-learning model in memory.