Data manipulation and decision processing

• US 7,379,926 B1
• Filed: 02/11/2002
• Issued: 05/27/2008
• Est. Priority Date: 02/09/2001
• Status: Active Grant

• Current Assignee: Remington Partners
• Original Assignee: Remington Partners
• Inventors: Belniak, MacGregor, Moraleda, Jorge
• Primary Examiner(s): Hirl; Joseph P
• Application Number:
  US10/074,958
• Time in Patent Office: 2,297 Days
• Field of Search: 706/46, 706/12, 706/14, 717/102
• US Class Current: 706/46
• CPC Class Codes: G06N 5/00 Computing arrangements using knowledge-based models

Abstract

A device for making automated transaction decisions comprising a memory to store a series of historical records and results, where each record contains a set of data values for a corresponding set of variables, a transaction input unit to convey corresponding data values of a transaction for which a decision is to be made, a decision output unit to convey a recommended decision, a self-configuring modeler to configure itself based upon information stored in memory in order to predict expected results from any set of data values, and a controller coupled to the modeler and the transaction input unit to generate a decision output.

Claims

1. A method comprising:

dynamically processing a database to eliminate missing values in records, each record to contain a plurality of values;

preparing a statistical summary of processed data in the database;

constructing an alternating decision tree (“

ADTree”

) database structure to hold the statistical summary, wherein the ADTree database structure provides dynamic pruning of less frequently used branches when system memory is low;

searching through possible relationship models to find a high-scoring network; and

using the high-scoring network to compute a decision for a new case.

2. The method of claim 1, further comprising:

discretizing continuous variables in the database before preparing the plurality of statistical summaries.

3. The method of claim 1, further comprising:

modifying a relationship of the high-scoring network.

4. The method of claim 1 wherein the new case consists of values of a subset of fields of a record.

5. The method of claim 1 wherein records in the database represent bank customer transaction records, and the decision is to identify an at-risk customer of the bank.

6. The method of claim 1 wherein the statistical summary is a Bayesian model of correlations between data in records.

7. The method of claim 1 wherein the decision is one of a detection of an illegal financial transaction, a network fault diagnosis, or a prediction of a result of a pharmaceutical compound in an organism.

8. The method of claim 1, further comprising:

aggregating similar database records together; and

computing a frequency of occurrence for each of the aggregated records.

9. The method of claim 8 wherein aggregating comprises:

preparing a plurality of hashes to cluster the database records.

10. The method of claim 1 wherein the database structure comprises correlation statistics between nodes.

11. The method of claim 1, further comprising:

inferring a value of a non-observed variable based on a previous observation and the high-scoring network.

12. The method of claim 1, further comprising:

storing the high-scoring network in an eXtensible Markup Language (“

XML”

) format.