Transforming predictive models

Patent metadata

Field	Value
Publication number	US-9483734-B2
Application number	US-201614988038-A
Country	US
Kind code	B2
Filing date	Jan 5, 2016
Priority date	Jan 12, 2015
Publication date	Nov 1, 2016
Grant date	Nov 1, 2016

Abstract

A system for translating a boosting algorithm includes an interface communicatively coupled to a processor. The interface is operable to receive a trained boosting model. The processor is operable to identify a plurality of split-node variables associated with the trained boosting model. Each of the plurality of split-node variables comprises a variable name, a cutoff point, and a weight. The processor may aggregate the split-node variables by variable name and cutoff point and then combine the weights of each of the plurality of split-node variables having the same variable name and cutoff point. The processor may then create a linear model based on the combined variables. The processor may further create a performance scorecard based on the combined weight, and create a model evaluation based on the error rate of the trained boosting model.

First claim

What is claimed is:

1. A method for translating a boosting algorithm, comprising:
receiving, at a hardware interface, a trained boosting model;
identifying, using a processor communicatively coupled to the interface, a plurality of one-level, binary split-node variables associated with the trained boosting model, wherein each of the plurality of one-level, binary split-node variables comprises a variable name, a cutoff point, and a weight;
identifying, using the processor, a group of one-level, binary split-node variables that have the same variable name and cutoff point within the plurality of one-level, binary split-node variables;
combining, using the processor, the weights of each of the one-level, binary split-node variables in the group of one-level, binary split-node variables to calculate a combined weight for the one-level, binary split-node variables in the group of one-level, binary split-node variables, wherein combining the weights comprises summing the weights of the one-level, binary split-node variables in the group of one-level, binary split-node variables;
creating, using the processor, a linear model based at least in part upon the variable name, the cutoff point, and the combined weight;
creating a performance scorecard based on the combined weight for the group of one-level, binary split-node variables; and
creating a model evaluation based on an error rate of the trained boosting model.

2. The method of claim 1 , wherein the trained boosting model is one selected from the group comprising: Discrete Adaboost, Real Adaboost, Gentle Adaboost, and Logitboost.

3. The method of claim 1 , wherein the linear model comprises conditional logic.

4. The method of claim 1 , wherein the linear model is in an if-then-else format.

5. The method of claim 1 , further comprising creating a model evaluation based on a Kolmogorov-Smirnov test of the trained boosting model.

6. A non-transitory computer readable storage medium comprising logic, the logic operable, when executed by a processor, to:
receive a trained boosting model;
identify a plurality of one-level, binary split-node variables associated with the trained boosting model, wherein each of the plurality of one-level, binary split-node variables comprise a variable name, a cutoff point, and a weight;
identify a group of one-level, binary split-node variables that have the same variable name and cutoff point within the plurality of one-level, binary split-node variables;
combine the weights of each of the one-level, binary split-node variables in the group of one-level, binary split-node variables to calculate a combined weight for the one-level, binary split node variables in the group of one-level, binary split node variables, wherein combining the weights comprises summing the weights of the one-level, binary split-node variables in the group of one-level, binary split-node variables;
create a linear model based on the variable name, the cutoff point, and the combined weight;
creating a performance scorecard based on the combined weight for the one-level, binary split-node variables; and
creating a model evaluation based on the error rate of the trained boosting model.

7. The non-transitory media of claim 6 , wherein the trained boosting model is one selected from the group comprising: Discrete Adaboost, Real Adaboost, Gentle Adaboost, and Logitboost.

8. The non-transitory media of claim 6 , wherein the linear model comprises conditional logic.

9. The non-transitory media of claim 6 , wherein the linear model is in an if-then-else format.

10. The non-transitory media of claim 6 , further comprising creating a model evaluation based on a Kolmogorov-Smirnov test of the trained boosting model.

11. A system for translating a boosting algorithm, comprising:
a hardware interface operable to:
receive a trained boosting model; and
a hardware processor operable to:
identify a plurality of one-level, binary split-node variables associated with the trained boosting model, wherein each of the plurality of one-level, binary split-node variables comprises a variable name, a cutoff point, and a weight;
identify a group of one-level, binary split-node variables that have the same variable name and cutoff point within the plurality of one-level, binary split-node variables;
combine the weights of each of the one-level, binary split-node variables in the group of one-level, binary split-node variables to calculate a combined weight for the one-level, binary split-node variables in the group of one-level, binary split-node variables, wherein combining the weights comprises summing the weights of the one-level, binary split-node variables in the group of one-level, binary split-node variables;
create a linear model based on the variable name, the cutoff point, and the combined weight;
creating a performance scorecard based on the combined weight for the one-level, binary split-node variables; and
create a model evaluation based on the error rate of the trained boosting model.

12. The system of claim 11 , wherein the trained boosting model is one selected from the group comprising: Discrete Adaboost, Real Adaboost, Gentle Adaboost, and Logitboost.

13. The system of claim 11 , wherein the processor is further operable to create a model evaluation based on a Kolmogorov-Smirnov test of the trained boosting model.

14. The system of claim 11 , wherein the linear model comprises conditional logic.

Assignees

• Bank Of America

Inventors

• Laxmanan Kasilingam Basker
• Chen Yudong
• Song Peng

Classifications

• G06N5/022 Primary

  Knowledge engineering; Knowledge acquisition · CPC title

• G06N5/04 Primary

  Inference or reasoning models · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F30/00

  Computer-aided design [CAD] · CPC title

• G06F17/50

  Physics · mapped topic