System and method for developing a propensity model
View Patent ↗A technique is provided for developing a propensity model for customer behavior. Multiple biased samples of customer characteristics and results from past activities are established. Initial propensity models are created for each biased sample. The propensity models established for each biased sample are processed separately from the propensity models established for the other biased samples. A genetic algorithm is used to evolve the propensity models. A select number of propensity models that best fit their respective biased samples are compared to a validation sample that is unbiased. A select number of these propensity models that best fit the validation sample are cross-bred into the propensity models established for each biased sample. The propensity models for each biased sample are then processed again using the genetic algorithms. However, a number of elite propensity models are maintained in their original form and not evolved using the genetic algorithm. This cycle continues until a stopping criterion is reached.
1. A computer-implemented method of developing a propensity model for an event, comprising:
establishing a plurality of biased samples of data representative of a previous event, wherein each of the plurality of biased samples of data representative of a previous event is biased differently;
establishing at least one propensity model for each biased sample of data;
processing each of the at least one propensity model with a genetic algorithm;
checking fitness of at least one propensity model for each biased sample against a validation sample of data, wherein the validation sample of data is unbiased; and
identifying a first propensity model for a first biased sample of data for cross-breeding with a second propensity model for a second biased sample based on fitness of the first propensity model to the validation sample of data.
2. The computer-implemented method of developing a propensity model as recited in claim 1 , comprising:
cross-breeding the first propensity model and the second propensity model with the genetic algorithm to establish a third propensity model.
3. The computer-implemented method of developing a propensity model as recited in claim 2 , wherein crossbreeding the first propensity model and the second propensity model with a genetic algorithm to establish a third propensity model comprises combining a portion of the first propensity model with a portion of the second propensity model to create a third propensity model.
4. The computer-implemented method of developing a propensity model as recited in claim 2 , wherein crossbreeding the first propensity model and the second propensity model with a genetic algorithm to establish a third propensity model comprises averaging a variable of the first propensity model with a corresponding variable in the second propensity model.
5. The computer-implemented method of developing a propensity model as recited in claim 2 , comprising processing the second propensity model with the genetic algorithm after cross-breeding the second propensity model with the first propensity model.
6. The computer-implemented method of developing a propensity model as recited in claim 2 , wherein a copy of the first propensity model is associated with the second biased sample of data and not evolved by the genetic algorithm.
7. The computer-implemented method of developing a propensity model as recited in claim 1 , wherein identifying a first propensity model based on fitness of the first propensity model to a validation sample of data comprises:
checking fitness of each propensity model to its biased sample of data;
identifying at least one propensity model for each biased sample of data to check for fitness to the validation sample of data based on fitness of each propensity model to its biased sample of data; and
checking each of the at least one propensity model for each biased sample for fitness to the validation sample of data.
8. The computer-implemented method of developing a propensity model as recited in claim 7 , wherein identifying a first propensity model comprises identifying as the first propensity model the propensity model that best fits the validation sample of data from among the at least one propensity model for each biased sample of data.
9. The computer-implemented method of developing a propensity model as recited in claim 7 , comprising:
storing a copy of each propensity model that best fits each of the biased samples of data and not evolving the copy with the genetic algorithm; and
processing each other propensity model of each biased sample of data with the genetic algorithm.
10. The computer-implemented method of developing a propensity model as recited in claim 1 , wherein processing each propensity model established for each biased sample of data with a genetic algorithm comprises processing a propensity model corresponding to a first biased sample with the genetic algorithm independently of processing a propensity model corresponding to a second biased sample with the genetic algorithm.
11. The computer-implemented method of developing a propensity model as recited in claim 1 , wherein the data comprises data representative of each entity of a plurality of entities and a previous result corresponding to each entity of the plurality of entities.
12. The computer-implemented method of developing a propensity model as recited in claim 11 , wherein each biased sample of data comprises the data representative of each entity of the plurality of entities having a specific previous result.
13. The computer-implemented method of developing a propensity model as recited in claim 12 , wherein each biased sample of data comprises the data representative of different limited sets of entities of the plurality of entities having a result different from the specific previous result.
14. A non-transitory machine-readable medium for developing a propensity model, comprising:
code operable to establish a plurality of biased samples of data representative of a previous event, wherein each of the plurality of biased samples of data representative of a previous event is biased differently;
code operable to establish a propensity model for each biased sample of data;
code operable to process each propensity model with a genetic algorithm;
code operable to check fitness of at least one propensity model for each biased sample against a validation sample of data, wherein the validation sample of data is unbiased; and
code operable to identify a first propensity model for a first biased sample of data for cross-breeding with a second propensity model for a second biased sample based on fitness of the first propensity model to the validation sample of data.
15. A computer-implemented method of developing a propensity model for an event, comprising:
establishing a plurality of biased samples of data;
establishing a plurality of propensity models that correspond to each of the plurality of biased samples of data;
processing each of the plurality of propensity models that correspond to each of the plurality of biased samples of data with a genetic algorithm;
identifying at least one propensity model that best fits its biased sample of data from each of the plurality of propensity models processed with the genetic algorithm;
identifying a propensity model that best fits a validation sample from the plurality of at least one propensity model that best fits its biased sample of data; and
adding the propensity model that best fits a validation sample to each of the plurality of propensity models that correspond to each of the plurality of biased samples of data.
16. The computer-implemented method of developing a propensity model as recited in claim 15 , wherein the validation sample of data is unbiased.
17. The computer-implemented method of developing a propensity model as recited in claim 15 , comprising processing each of the plurality of propensity models that correspond to each of the plurality of biased samples of data with a genetic algorithm after adding the propensity model that best fits a validation sample to each of the plurality of propensity models that correspond to each of the plurality of biased samples of data.
18. The computer-implemented method of developing a propensity model as recited in claim 17 , wherein processing each of the plurality of propensity models that correspond to each of the plurality of biased samples of data with a genetic algorithm after adding the propensity model that best fits a validation sample to each of the plurality of propensity models comprises cross-breeding the propensity model that best fits a validation sample with a propensity model in at least one of the plurality of propensity models that correspond to each biased sample of data.
19. The computer-implemented method of developing a propensity model as recited in claim 17 , wherein processing each of the plurality of propensity models that correspond to each of the plurality of biased samples of data with a genetic algorithm after adding the propensity model that best fits a validation sample to each of the plurality of propensity models comprises mutating the propensity model that best fits a validation sample.
20. The computer-implemented method of developing a propensity model as recited in claim 17 , comprising storing a copy of each of the plurality of propensity models that best fit its biased sample of data so that the copy of each of the plurality of propensity models that best fit its biased sample of data is not evolved by the genetic algorithm.
21. The computer-implemented method of developing a propensity model as recited in claim 15 , comprising:
identifying at least one propensity model that best fits its biased sample of data from each of the plurality of propensity models processed with the genetic algorithm after cross-breeding;
identifying a propensity model that best fits a validation sample from the plurality of at least one propensity model that best fits its biased sample of data after cross-breeding; and
cross-breeding the propensity model that best fits a validation sample from the plurality of at least one propensity model that best fits its biased sample of data after cross-breeding with a propensity model for each of the plurality of biased samples of data.