Method of building predictive models on transactional data

• US 6,873,979 B2
• Filed: 02/28/2001
• Issued: 03/29/2005
• Est. Priority Date: 02/29/2000
• Status: Active Grant

• Current Assignee: Experian Information Solutions Incorporated (Experian plc)
• Original Assignee: Marketswitch Corp. (Experian plc)
• Inventors: Reynberg, Anatoly, Fishman, Vladimir, Galperin, Yuri
• Primary Examiner(s): Knight, Anthony
• Assistant Examiner(s): Holmes, Michael B.
• Application Number:
  US09/795,224
• Publication Number:
  US 20010037321A1
• Time in Patent Office: 1,490 Days
• Field of Search: 706 15- 44, 705/35, 707/1, 707/3, 709/200
• US Class Current: 706/21
• CPC Class Codes: G06Q 10/04 Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" market predictions or forecasting for commercial activities G06Q30/0202
• CPC Class Codes: G06Q 99/00 Subject matter not provided for in other groups of this subclass

Abstract

A method of building predictive statistical models provides a dedicated aggregation module for each transactional record source. Each aggregation module aggregates the transactional records using a neural network function to produce a scalar output which can then be input to a traditional modeling function, which may employ either logistic regression, neural network, or radial basis function techniques. The output of the aggregation modules can be saved, and updated aggregation values can be updated by processing new transaction records and combining the new transaction values with the previous output values using a blending function. Parameters of the neural network in the aggregation module may be calculated simultaneously with the parameters of the traditional modeling module.

Claims

1. A method of building predictive models on transactional data, comprising:

providing an aggregation module for each transactional record source;

initializing output values of each aggregation module;

inputting a first transactional record from each transactional record source into said corresponding aggregation module;

calculating a first iteration of said output values for each aggregation module as;

ƒ

_kⁱ(1)=F(Ø

(Σ

t_pqw_i^m),0),

 

where;

φ

is a neural network element function;

F is a blending function that controls how fast a previous

transactional record become obsolete; and

W_i^m are weights of the neural network;

inputting a next transactional record from each transactional record source into said corresponding aggregation module;

updating said outputs values of each aggregation module as;

ƒ

_kⁱ(r+1)=F(Ø

(Σ

t_pqw_i^m),ƒ

_kⁱ(r));

repeating the two prior steps until all transactional records are processed; and

obtaining scalar values ƒ

_kⁱ as scalar inputs for traditional modeling.

2. The method of building predictive models on transactional data of claim 1, further comprising storing output values ƒ

_kⁱ.

3. The method of building predictive models on transactional data of claim 1, further comprising:

combining accumulated values of ƒ

_kⁱ with a traditional array of scalar inputs S as input attributes and;

using said input attributes in a traditional modeling module.

4. The method of building predictive models on transactional data of claim 3, further comprising selecting said traditional modeling module from the group consisting of logistic regression models, neural networks and radial basis function models.

5. The method of building predictive models on transactional data of claim 2, further comprising updating output values of each aggregation module in response to new transactional data by:

setting output values of each aggregation module equal to previously stored values ƒ

_kⁱ;

inputting a new transactional record from each transactional record source into said corresponding aggregation module;

updating said outputs values of each aggregation module as;

ƒ

_kⁱ=F(Ø

(Σ

t_pqw_i^m),ƒ

_kⁱ)

repeating the two prior steps until all new transactional records are processed; and

obtaining scalar values ƒ

_kⁱ as scalar inputs for traditional modeling.

6. The method of building predictive models on transactional data of claim 5, further comprising staring output values ƒ

_kⁱ.

7. The method of building predictive models on transactional data of claim 5, further comprising:

combining accumulated values of ƒ

_kⁱ with a traditional array of scalar inputs S as input attributes and;

using said input attributes in a traditional modeling module.

8. The method of building predictive models on transactional data of claim 7, further comprising selecting said traditional modeling module from the group consisting of logistic regression models, neural networks and radial basis function models.

9. A method for predicting a response to an offer comprising the steps of:

aggregating first transaction records from a first transaction source using a first aggregation module to produce a set of first output variables, the first transaction records corresponding to a plurality of customers, each of the first transaction records having a first number of attributes, a number of first transaction records associated with each customer varying from customer to customer, the first aggregation module comprising a neural network that weights the attributes from the first transaction records to produce the first output variables using a first set of aggregation weights;

aggregating second transaction records from a second transaction source using a second aggregation module to produce a set of second output variables, the second transaction records corresponding to a plurality of customers, each of the second transaction records having a second number of attributes, a number of second transaction records associated with each customer varying from customer to customer, the second aggregation module comprising a neural network that weights the attributes from the second transaction records to produce the second output variables using a second set of aggregation weights;

inputting the first output variables and the second output variables to a traditional modeling module;

inputting scalar variables into the traditional modeling module;

calculating, by the traditional modeling module, a predicted response to an offer; and

determining whether to make the offer based on predicted response.

10. The method of claim 9, wherein the traditional modeling module calculates the predicted response using a technique selected from the group consisting of logistic regression, neural network, and radial basis function.

11. The method of claim 9, further comprising the step of calculating the first aggregation weights and the second aggregation weights using a back propagation method.

12. The method of claim 11, wherein the first aggregation weights and the second aggregation weights and parameters of the traditional modeling module are calculated simultaneously.

13. The method of claim 9, further comprising the step of updating at least either the first output variables or the second output variable based on new transaction data according to the following formula:

ƒ

_kⁱ(r=1)=F(Ø

(Σ

_pqⁱw_i^m),ƒ

_kⁱ(r))

where

φ

is a neural network element function;

F is a blending function that controls how fast a previous transactional record become obsolete;

an

w_i^m are weights of the neural network of the aggregation module corresponding to the new transaction data.