Machine learning of predictive models using partial regression trends

• US 10,229,368 B2
• Filed: 10/19/2015
• Issued: 03/12/2019
• Est. Priority Date: 10/19/2015
• Status: Active Grant

• Current Assignee: Kyndryl Incorporated
• Original Assignee: International Business Machines Corporation
• Inventors: Baughman, Aaron K., Malkin, Peter K., O'Connell, Brian M., Pan, Hubert E.
• Primary Examiner(s): Holmes, Michael B
• Application Number:
  US14/886,236
• Publication Number:
  US 20170109649A1
• Time in Patent Office: 1,240 Days
• Field of Search: 706 14
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

An input is selected from a set of inputs used by a prediction model to produce an initial predicted value of an outcome. A changed predicted value of the outcome is produced by removing the selected input from the inputs to the model. An actual value of the outcome is obtained. A label residual is computed using the actual value and the changed predicted value. A second prediction model is formed to predict a value of the selected input. A variable residual is computed using an actual value and the predicted value of the selected input. An expression is generated of a plot of the label residual and the variable residual. The selected input is transformed, to form a transformed selected input, where the model produces a second predicted value of the outcome by using the transformed selected input.

Claims

1. A method for improving a prediction model, the method comprising:

modifying a load prediction component for a server, the modifying enabling the load prediction component to improve with training, the modifying comprising;

selecting, at an application executing using a processor and a memory, an input from a set of inputs, the set of inputs being used by the prediction model to produce an initial predicted value of an outcome at a future time;

producing, using the processor and the memory, a changed predicted value of the outcome by removing the selected input from the set of inputs supplied to the model;

obtaining an actual value of the outcome at the future time;

computing a label residual using the actual value of the outcome and the changed predicted value;

forming a second prediction model to predict a value of the selected input at the future time;

computing a variable residual using an actual value of the selected input at the future time and the predicted value of the selected input;

generating an expression of a plot of the label residual and the variable residual; and

transforming the selected input, to form a transformed selected input, wherein the model produces a second predicted value of the outcome by using the transformed selected input.

2. The method of claim 1, further comprising:

determining, from the changed predicted value and the initial predicted value a relationship between a prediction made by the prediction model and the selected input that has been removed to produce the changed predicted output; and

using, responsive to the relationship exhibiting a greater than a threshold level of influence of the selected input on the prediction, the relationship to prevent the selected input from elimination from the set of inputs during a dimensionality reduction of the prediction model.

3. The method of claim 1, further comprising:

replacing, to form a revised set of input, the selected input in the set of inputs with the transformed selected input; and

providing the revised set of input to the model, wherein the second predicted value of the outcome is for a second future time, and wherein the second predicted value of the outcome is more accurate than a third predicted value that would be produced for the second future time when the model uses the set of inputs instead of the revised set of inputs.

4. The method of claim 1, further comprising:

computing several label residual values at different future times;

computing several variable residual values at the different future times; and

plotting the several label residual values with the several variable residual values that correspond at the different future times on the plot.

5. The method of claim 1, further comprising:

supplying the subset of inputs to the second model to predict the value of the selected input.

6. The method of claim 1, further comprising:

executing, as a part of the producing the changed predicted value, the prediction model using a subset of the inputs, the subset of the inputs excluding the selected input.

7. The method of claim 6, wherein the subset comprises all inputs in the set of inputs except the selected input.

8. The method of claim 1, wherein the outcome is a result of an event at the future time.

9. The method of claim 1, wherein the outcome is a dependent variable in an event, and wherein the predicted value is a value of the dependent variable at the future time.

10. The method of claim 1, further comprising:

producing the initial prediction for the future time by executing the prediction model using the set of inputs.

11. The method of claim 1, wherein the method is embodied in a computer program product comprising one or more computer-readable storage devices and computer-readable program instructions which are stored on the one or more computer-readable tangible storage devices and executed by one or more processors.

12. The method of claim 1, wherein the method is embodied in a computer system comprising one or more processors, one or more computer-readable memories, one or more computer-readable storage devices and program instructions which are stored on the one or more computer-readable storage devices for execution by the one or more processors via the one or more memories and executed by the one or more processors.

13. A computer program product for improving a prediction model, the computer program product comprising one or more computer-readable storage devices, and program instructions stored on at least one of the one or more storage devices, the stored program instructions comprising:

program instructions to modify a load prediction component for a server, the modifying enabling the load prediction component to improve with training, the modifying comprising

program instructions to select, at an application executing using a processor and a memory, an input from a set of inputs, the set of inputs being used by the prediction model to produce an initial predicted value of an outcome at a future time;

program instructions to produce, using the processor and the memory, a changed predicted value of the outcome by removing the selected input from the set of inputs supplied to the model;

program instructions to obtain an actual value of the outcome at the future time;

program instructions to compute a label residual using the actual value of the outcome and the changed predicted value;

program instructions to form a second prediction model to predict a value of the selected input at the future time;

program instructions to compute a variable residual using an actual value of the selected input at the future time and the predicted value of the selected input;

program instructions to generate an expression of a plot of the label residual and the variable residual; and

program instructions to transform the selected input, to form a transformed selected input, wherein the model produces a second predicted value of the outcome by using the transformed selected input.

14. The computer program product of claim 13, further comprising:

program instructions to determine, from the changed predicted value and the initial predicted value a relationship between a prediction made by the prediction model and the selected input that has been removed to produce the changed predicted output; and

program instructions to use, responsive to the relationship exhibiting a greater than a threshold level of influence of the selected input on the prediction, the relationship to prevent the selected input from elimination from the set of inputs during a dimensionality reduction of the prediction model.

15. The computer program product of claim 13, further comprising:

program instructions to replace, to form a revised set of input, the selected input in the set of inputs with the transformed selected input; and

program instructions to provide the revised set of input to the model, wherein the second predicted value of the outcome is for a second future time, and wherein the second predicted value of the outcome is more accurate than a third predicted value that would be produced for the second future time when the model uses the set of inputs instead of the revised set of inputs.

16. The computer program product of claim 13, further comprising:

program instructions to compute several label residual values at different future times;

computing several variable residual values at the different future times; and

program instructions to plot the several label residual values with the several variable residual values that correspond at the different future times on the plot.

17. The computer program product of claim 13, further comprising:

program instructions to supply the subset of inputs to the second model to predict the value of the selected input.

18. The computer program product of claim 13, further comprising:

program instructions to execute, as a part of the producing the changed predicted value, the prediction model using a subset of the inputs, the subset of the inputs excluding the selected input.

19. The computer program product of claim 18, wherein the subset comprises all inputs in the set of inputs except the selected input.

20. A computer system for improving a prediction model, the computer system comprising one or more processors, one or more computer-readable memories, and one or more computer-readable storage devices, and program instructions stored on at least one of the one or more storage devices for execution by at least one of the one or more processors via at least one of the one or more memories, the stored program instructions comprising:

program instructions to modify a load prediction component for a server, the modifying enabling the load prediction component to improve with training, the modifying comprising

program instructions to select, at an application executing using a processor and a memory, an input from a set of inputs, the set of inputs being used by the prediction model to produce an initial predicted value of an outcome at a future time;

program instructions to produce, using the processor and the memory, a changed predicted value of the outcome by removing the selected input from the set of inputs supplied to the model;

program instructions to obtain an actual value of the outcome at the future time;

program instructions to compute a label residual using the actual value of the outcome and the changed predicted value;

program instructions to form a second prediction model to predict a value of the selected input at the future time;

program instructions to compute a variable residual using an actual value of the selected input at the future time and the predicted value of the selected input;

program instructions to generate an expression of a plot of the label residual and the variable residual; and

program instructions to transform the selected input, to form a transformed selected input, wherein the model produces a second predicted value of the outcome by using the transformed selected input.