Predictive, machine-learning, time-series computer models suitable for sparse training sets

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, for a plurality of entities, entity logs, wherein:

the entity logs comprise events involving the entities,

a first subset of the events are actions by the entities,

at least some of the actions by the entities are targeted actions, and

the events are labeled according to an ontology of events having a plurality of event types;

transforming, with one or more processors, each entity log into a collection of features to which the predictive machine learning model is capable of responding;

training, with one or more processors, based on the transformed collection of features from the entity logs, a predictive machine learning model to predict whether an entity characterized by a set of inputs to the model will engage in a targeted action in a given duration of time in the future; and

storing, with one or more processors, the trained predictive machine learning model in memory.

2. The medium of claim 1 , comprising predicting whether a given entity will engage in a targeted action with the trained predictive machine learning at least in part by:

obtaining a given entity log of the given entity;

determining a plurality of features from the given entity log, the plurality of features having fewer dimensions than the given entity log; and

inputting the determined plurality of features into the trained predictive machine learning model to cause the model to output a value indicative of the whether the given entity will engage in a targeted action within the given duration of time in the future, wherein:

the trained model is configured to make predictions based on features indicative of recency of events prior to targeted actions.

3. The medium of claim 1 , wherein:

the predictive machine learning model is based on a plurality of decision trees combined with an ensemble procedure; and

at least some of the decision trees are trained with classification and regression tree (CART) learning by recursively splitting a feature space of inputs to the predictive machine learning model along different dimensions of the feature space at values of respective dimensions that locally optimize the respective split to minimize entropy of Gini impurity of targeted actions and non-targeted actions on each side of respective splits.

4. The medium of claim 3 , wherein the ensemble procedure comprises boosting.

5. The medium of claim 3 , wherein the ensemble procedure comprises random forest or rotation forest.

6. The medium of claim 1 , wherein at least some of the features are determined by classifying whether a given type of non-targeted-action event occurred with a range of time prior to a targeted action.

7. The medium of claim 1 , wherein:

the trained model is configured to make predictions based on features indicative of frequency of events prior to targeted actions; and

at least some of the features are determined by classifying an amount of times a given type of non-targeted-action event occurred with a range of time prior to a targeted action.

8. The medium of claim 1 , wherein at least some of the features are determined based on an amount of time a first type of non-targeted-action event occurred prior to a targeted action.

9. The medium of claim 1 , wherein a given one of the features is determined based on a first amount of time a first type of non-targeted-action event occurred prior to a targeted action and based on a second amount of time a second type of non-targeted-action event occurred prior to a targeted action.

10. The medium of claim 1 , wherein a given one of the features is determined based on first amounts of time and based on a first number of times a first type of non-targeted-action event occurred prior to a designed targeted action.

11. The medium of claim 10 , wherein the given one of the features is determined based on second amounts of time and based on a second number of times a second type of non-targeted-action event occurred prior to the designed targeted action.

12. The medium of claim 1 , wherein:

a second subset of events are actions taken on the entities by others;

at least some features upon which the model is trained are based on events in the second subset; and

at least some features upon which the model is trained are based on events in the first subset.

13. The medium of claim 1 , wherein:

the entities include consumers;

the events include communications to consumers by an enterprise;

the events include purchases by consumers from the enterprise;

the events include non-purchase interactions by consumers with the enterprise; and

the entity logs are obtained from a customer relationship management system of the enterprise.

14. The medium of claim 13 , wherein:

the enterprise is a credit card issuer and the trained predictive machine learning model is configured to predict whether a consumer will default;

the enterprise is a lender and the trained predictive machine learning model is configured to predict whether a consumer will borrow;

the enterprise is an insurance company and the trained predictive machine learning model is configured to predict whether a consumer will file a claim;

the enterprise is an insurance company and the trained predictive machine learning model is configured to predict whether a consumer will sign-up for insurance;

the enterprise is a vehicle seller and the trained predictive machine learning model is configured to predict whether a consumer will purchase a vehicle; or

the enterprise is a seller of goods and the trained predictive machine learning model is configured to predict whether a consumer will file a warranty claim.

15. The medium of claim 1 , wherein:

at least some of the features are not temporality or sequentially adjacent a targeted action.

16. The medium of claim 1 , wherein:

at least some of the features overlap in respective ranges of time or sequence to which they pertain.

17. The medium of claim 1 , wherein:

the trained predictive machine learning model is configured to output a plurality of scores each indicative of likelihood of an entity engaging in a targeted behavior in a different respective duration of time in the future.

18. The medium of claim 1 , wherein:

the trained predictive machine learning model is configured to output a plurality of scores each indicative of likelihood of an entity engaging in a different respective targeted behavior in the given duration of time in the future; and

the operations comprise:

inputting more than 1,000 entity logs corresponding to more than 1,000 entities into the trained machine learning model; and

selecting a subset of the more than 1,000 entities predicted to engage in a targeted action by the trained predictive machine learning model.

19. The medium of claim 18 , comprising:

causing one or more messages to be sent to members of the selected subset of entities responsive to the selection.

20. The medium of claim 1 , wherein training comprises means for training.

21. The medium of claim 1 , the operations comprising:

steps for predicting whether an entity will engage in a targeted behavior in the future.

22. The medium of claim 1 , wherein the ontology of events comprise:

a primary level comprising a plurality of exogenous actions; and

a secondary level comprising a plurality of endogenous actions, wherein:

a numerical score is assigned to each exogenous action and endogenous act;

the numerical score assigned to each exogenous action is higher than the numerical score assigned to each endogenous act; and

an aggregate score is calculated based on a weighted combination of numerical scores.

23. The medium of claim 1 , wherein the trained model is configured to filter some of the features of entity logs, wherein the filtration comprises applying:

a dependency level among the features of entity logs determined by a Bayesian Network;

a logistic regression calculated by a Lasso, Boruta, or ElasticNet penalty function; or

a product moment correlation coefficient calculated by a Pearson correlation.

24. The medium of claim 23 , wherein the filtration comprises applying:

a dependency level among the features of entity logs determined by a Bayesian Network;

a logistic regression calculated by a Lasso, Boruta, or ElasticNet penalty function; and

a product moment correlation coefficient calculated by a Pearson correlation.

25. The medium of claim 1 , wherein the predictive machine learning model is a real-time model, configured to adjust model parameters as new entity log entries are obtained.

26. The medium of claim 1 , wherein the predictive machine learning model is configured to accommodate sparse entity logs by applying discriminative classification models.

27. The medium of claim 1 , wherein the events are further labeled according to a hierarchical taxonomy of events having a plurality of event types.

28. The medium of claim 1 , wherein the targeted actions are outcomes of an industrial process controlled over time.

29. A method, comprising:

obtaining, with one or more processors, for a plurality of entities, entity logs, wherein:

the entity logs comprise events involving the entities,

a first subset of the events are actions by the entities,

at least some of the actions by the entities are targeted actions, and

the events are labeled according to an ontology of events having a plurality of event types;

transforming, with one or more processors, each entity log into a collection of features to which the predictive machine learning model is capable of responding;

training, with one or more processors, based on the transformed collection of features from the entity logs, a predictive machine learning model to predict whether an entity characterized by a set of inputs to the model will engage in a targeted action in a given duration of time in the future; and

storing, with one or more processors, the trained predictive machine learning model in memory.