Ensemble model policy generation for prediction systems

1. A method, by a processor, for ensemble policy generation for prediction systems, comprising:

deriving, by the processor, a plurality of policies for a set of ensemble models to predict a plurality of target variables for streaming data such that the plurality of policies enables dynamic adjustment of the prediction system, wherein the plurality of policies indicate which types of models are to be included in the set of ensemble models, which specific instances of a model class of the types of models for a particular category are to be included in the set of ensemble models, a weight applied to each model in the set of ensemble models, and which models of the set of ensemble models are to be re-executed, retrained, and rebuilt such that the plurality of policies are derived in real-time as the streaming data is received;

generating, by the processor, the plurality of policies according to contextual information of the streaming data, model similarities, ensemble model definitions, prediction quality metrics, and ground truth data of the set of ensemble models;

identifying model similarity for the set of ensemble models according to prediction accuracy of a plurality of conditions based on processing of the streaming data;

wherein one or more error states of those of the set of ensemble models determined to have model similarity parameters over a predetermined threshold are correlated to one or more error states produced by constituent ones of those of the set of ensemble models determined to have the model similarity parameters over the predetermined threshold;

dynamically updating by the processor, automatically and with no user input, one or more of the plurality of policies according to the one or more error states of the set of ensemble models and a performance of those of the set of ensemble models having the model similarity parameters;

wherein the error states generated by a first model of the set of ensemble models of a first type is used to the plurality of policies for a second model of the set of ensemble models of a second type different than the first type;

initially training, by the processor, a classifier of the second model of the set of ensemble models of the second type using the updated one or more of the plurality of policies according to the error states generated by the first model of the set of ensemble models of the first type; and

predicting, by the processor, the plurality of target variables utilizing the updated one or more of the plurality of policies implemented within the set of ensemble models, wherein the prediction determined by the set of ensemble models is synthesized into a single score and presented via a display interface in communication with the processor.

2. The method of claim 1 , further including mapping sampled predictive qualities between those of the set of ensemble models having the model similarity features.

3. A system for ensemble policy generation for prediction systems, comprising:

one or more computers with executable instructions that when executed cause the system to:

derive, by a processor associated with the one or more computers and executing the executable instructions, a plurality of policies for a set of ensemble models to predict a plurality of target variables for streaming data such that the plurality of policies enables dynamic adjustment of the prediction system, wherein the plurality of policies indicate which types of models are to be included in the set of ensemble models, which specific instances of a model class of the types of models for a particular category are to be included in the set of ensemble models, a weight applied to each model in the set of ensemble models, and which models of the set of ensemble models are to be re-executed, retrained, and rebuilt such that the plurality of policies are derived in real-time as the streaming data is received;

generate, by the processor, the plurality of policies according to contextual information of the streaming data, model similarities, ensemble model definitions, prediction quality metrics, and ground truth data of the set of ensemble models;

identify model similarity for the set of ensemble models according to prediction accuracy of a plurality of conditions based on processing of the streaming data;

wherein one or more error states of those of the set of ensemble models determined to have model similarity parameters over a predetermined threshold are correlated to one or more error states produced by constituent ones of those of the set of ensemble models determined to have the model similarity parameters over the predetermined threshold;

dynamically update by the processor, automatically and with no user input, one or more of the plurality of policies according to one or more error states of the set of ensemble models and a performance of those of the set of ensemble models having the model similarity parameters;

wherein the error states generated by a first model of the set of ensemble models of a first type is used to dynamically adjust the plurality of policies for a second model of the set of ensemble models of a second type different than the first type;

initially train, by the processor, a classifier of the second model of the set of ensemble models of the second type using the updated one or more of the plurality of policies according to the error states generated by the first model of the set of ensemble models of the first type; and predict, by the processor, the plurality of target variables utilizing the updated one or more of the plurality of policies implemented within the set of ensemble models, wherein the prediction determined by the set of ensemble models is synthesized into a single score and presented via a display interface in communication with the processor.

4. The system of claim 3 , wherein the executable instructions sampled predictive qualities between those of the set of ensemble models having the model similarity features.

5. A computer program product for, by a processor, ensemble policy generation for prediction systems, the computer program product comprising a non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising:

an executable portion that derives, by the processor, a plurality of policies for a set of ensemble models to predict a plurality of target variables for streaming data such that the plurality of policies enables dynamic adjustment of the prediction system, wherein the plurality of policies indicate which types of models are to be included in the set of ensemble models, which specific instances of a model class of the types of models for a particular category are to be included in the set of ensemble models, a weight applied to each model in the set of ensemble models, and which models of the set of ensemble models are to be re-executed, retrained, and rebuilt such that the plurality of policies are derived in real-time as the streaming data is received;

an executable portion that generates, by the processor, the plurality of policies according to contextual information of the streaming data, model similarities, ensemble model definitions, prediction quality metrics, and ground truth data of the set of ensemble models;

an executable portion that identifies model similarity for the set of ensemble models according to prediction accuracy of a plurality of conditions based on processing of the streaming data;

wherein one or more error states of those of the set of ensemble models determined to have model similarity parameters over a predetermined threshold are correlated to one or more error states produced by constituent ones of those of the set of ensemble models determined to have the model similarity parameters over the predetermined threshold;

an executable portion that dynamically updates by the processor, automatically and with no user input, one or more of the plurality of policies according to one or more error states of the set of ensemble models and a performance of those of the set of ensemble models having the model similarity parameters;

wherein the error states generated by a first model of the set of ensemble models of a first type is used to dynamically adjust the plurality of policies for a second model of the set of ensemble models of a second type different than the first type;

an executable portion that initially trains, by the processor, a classifier of the second model of the set of ensemble models of the second type using the updated one or more of the plurality of policies according to the error states generated by the first model of the set of ensemble models of the first type; and

an executable portion that predicts, by the processor, the plurality of target variables utilizing the updated one or more of the plurality of policies implemented within the set of ensemble models, wherein the prediction determined by the set of ensemble models is synthesized into a single score and presented via a display interface in communication with the processor.

6. The computer program product of claim 5 , further including an executable portion that maps sampled predictive qualities between those of the set of ensemble models having the model similarity features.