Predictive model validation
View Patent ↗Models are generated using a variety of tools and features of a model generation platform. For example, in connection with a project in which a user generates a predictive model based on historical data about a system being modeled, the user is provided through a graphical user interface a structured sequence of model generation activities to be followed, the sequence including dimension reduction, model generation, model process validation, and model re-generation. In connection with a project in which a user generates a predictive model based on historical data about a system being modeled, the user is enabled to validate the model development process with cross-validation between at least two subsets of the historical data; the validated model development process is enabled to be reapplied.
1. A machine-based method comprising:
in connection with a project, selecting a first subset of historical data about a system being modeled for use in generating a tentative predictive model for the system, the first subset of historical data comprising less than all of the historical data;
selecting a process for developing a predictive model for the system from among multiple possible model development processes;
applying the selected model development process to the first subset of historical data to generate the tentative predictive model;
selecting a second subset of the historical data, the second subset being less than all of the historical data and being at least a portion of a complementary dataset of the first subset or being randomly selected from the historical data and independent of the first subset;
applying the tentative predictive model to the selected second subset, determining whether results of applying the tentative predictive model to the selected second subset validate that the selected model development process will produce a final predictive model that is accurate for data that is not part of the historical data, if the selected model development process is so validated, then applying the validated model development process to a full set of historical data that includes the first and second subsets to generate a final predictive model, and using the final predictive model.
2. The method of claim 1 also including displaying on a user interface project goals to enable the user to assess performance of the project, wherein the project goals comprise at least one of: cumulative lift over an interval of interest, degree of monotonicity, or concordance scores.
3. The method of claim 2 also including
identifying that the tentative predictive model does not produce at least a predefined degree of lift for at least one of the subsets.
4. The method of claim 3 also including
enabling a user to choose interactively at least one model development criterion change or transformation or interaction of variables to improve a fit of the predictive model.
5. The method of claim 4 also including
graphically displaying and comparing measures of performance for a validation dataset and a training dataset.
6. A machine-based method comprising:
in connection with a project, selecting a model development process from multiple model development processes to apply on a first subset of less than all of a set of historical data to generate a first tentative predictive model, applying the selected model development process including (a) automatically transforming variables of the first subset of the historical data, (b) automatically generating the first tentative predictive model, and (c) automatically generating performance measures of the first tentative predictive model, determining a validity of the selected development process based on the performance measures of the first tentative predictive model;
applying the validated model development process to the full set of historical data to generate a second, final predictive model, and using the final predictive model.
7. The method of claim 6 also including
generating measures of the performance of the predicative model for the at least two datasets, the performance measures being generated separately percentile by percentile.
8. The method of claim 6 also including
graphically displaying and comparing measures of the performance for at least two datasets.
9. The method of claim 6 also including
persistently storing the validated model development process and a validated model for computing propensities for at least one target outcome variable, the propensities serving as indices of a score for non-historical data.
10. The method of claim 6 also including providing a user interface for assessing project goals against performance.
11. The method of claim 6 also including providing a user interface for selecting at least one subset of the historical data in addition to a training subset.
12. The method of claim 6 providing a user interface for displaying the performance of the first predictive model for at least two subsets of the historical data for an interval of interest.
13. The method of claim 6 enabling a user to choose interactively at least one transformation or interaction of variables to improve the model development process.
14. The method of claim 6 also includes cross-validating the final model using random portions of the historical data.
15. The method of claim 6 providing a user interface that enables the user to select at least one validation dataset and invoke a process for validating the model development process.
16. The method of claim 6 providing a user interface that enables the user to point and click to cause display of information about the validation of the model development process.
17. The method of claim 16 in which the information about the validation of the model development process includes at least one of: a statistical report card with a link to the statistical report chart, a cumulative lift chart with a link to the cumulative lift chart, and a non-cumulative lift chart with a link to the non-cumulative lift chart.
18. The method of claim 17 in which invocation of the link to the statistical report card causes display of the statistics of the validation of the model development process.
19. The method of claim 17 in which invocation of the link to the cumulative lift chart causes display of a cumulative lift chart.
20. The method of claim 17 in which invocation of the link to the cumulative lift chart causes display of a non-cumulative lift chart.
21. The method of claim 17 in which a user is enabled to choose interactively at least one performance criterion change or transformation or interaction of variables to improve the model development process.
22. The method of claim 6 also including providing a user interface that enables the user to select at least one machine automated model development process applied to the entire set of the historical data for the validated model development process.
23. The method of claim 6 also including providing a user interface that enables the user to point and click to cause display of information about the performance of the validated model development process applied to the entire set of historical data.
24. The method of claim 23 in which the information about the performance of the first model for two independent datasets includes at least one of the following:
a statistical report card with a link to the statistical report chart, a cumulative lift chart with a link to the cumulative lift chart, a non-cumulative lift chart with a link to the non-cumulative lift chart.
25. The method of claim 24 in which the invocation of the link to the statistical report card causes display of the statistics of the validation of the model development process.
26. The method of claim 24 in which the invocation of the link to the cumulative lift chart causes display of a cumulative lift chart.
27. The method of claim 24 in which the invocation of the link to the cumulative lift chart causes display of a non-cumulative lift chart.
28. The method of claim 6 also including storing the final model and validation results of the model development process persistently.