Unsupervised analytical review

• US 9,037,607 B2
• Filed: 03/14/2013
• Issued: 05/19/2015
• Est. Priority Date: 02/20/2012
• Status: Active Grant

• Current Assignee: Galisteo Consulting Group, Inc.
• Original Assignee: Galisteo Consulting Group, Inc.
• Inventors: Chew, Peter Alexander
• Primary Examiner(s): Wilder, Hunter
• Application Number:
  US13/829,053
• Publication Number:
  US 20140279301A1
• Time in Patent Office: 796 Days
• Field of Search: 705/30, 707/1, 707/6, 707/748, 707/706, 707/776, 707/780
• US Class Current: 707/776
• CPC Class Codes: G06F 16/273 Asynchronous replication or reconciliation
• CPC Class Codes: G06Q 40/10 Tax strategies

Abstract

Disclosed is a method generally applicable to any financial dataset for the purposes of: (1) determining the most important patterns in the given dataset, in order of importance; (2) determining any trends in those patterns; (3) determining relationships between patterns and trends; and (4) allowing quick visual identification of anomalies for closer audit investigation. These purposes generally fall within the scope of what in financial auditing is known as ‘analytical review’. The current method'"'"'s advantages over existing methods are that is fully independent of the financial data subject to analysis, requires no background knowledge of the target business or industry, and is both scalable (to large datasets) and fully scale-invariant, requiring no a priori notion of financial materiality. These advantages mean, for example, that the same method can be by an external auditor for many different clients with virtually no client-specific customization, directing his attention to the areas where more detailed audit investigation may be required. Compared with existing methods, the current method is extremely flexible, and because it requires no a priori knowledge, saves significant time in understanding the fundamentals of a business.

Claims

1. A computer implemented method of unsupervised financial analytical review, comprising:

extracting features, by a computer having stored a financial dataset, of each transaction in the financial dataset;

generating, by the computer, a transaction-by-attribute/value matrix from the extracted features of each transaction in the financial dataset;

calculating, by the computer, distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the distinctiveness weights signify how much more likely an attribute/value is to occur in conjunction with a particular transaction than may be expected on the basis of chance;

calculating, by the computer, materiality weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the materiality weights signify an importance or significance of an amount, transaction or discrepancy with a particular transaction;

calculating, by the computer, combined weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the combined weight is calculated by adding each of the materiality weights for each attribute/value in the transaction-by-attribute/value matrix to respective distinctiveness weights not equal to zero of the distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix;

applying, by the computer, the combined weights to the transaction-by-attribute/value matrix to generate a weighted transaction-by-attribute/value matrix;

factorizing, by the computer, the weighted transaction-by-attribute/value matrix into a transaction-by-concept matrix; and

generating, by the computer, output of the transaction-by-concept matrix enabling significant patterns and trends to be identified;

wherein the extracting, the generating of the transaction-by-attribute/value matrix, the calculating of the distinctiveness weights, the calculating of the materiality weights, the calculating of the combined weights, the applying of the combined weights, the factorizing, and the generating of the output are performed regardless of a number, type, or monetary amount of transactions in the financial dataset, regardless of a number or type of features associated with each transaction in the financial dataset, and regardless of a provenance of the financial dataset.

2. The computer implemented method of claim 1, wherein extracting features of each transaction in the dataset comprises unpivoting the dataset to create a list of transactions associated with attribute/values.

3. The computer implemented method of claim 1, wherein extracting features comprises:

tokenizing contents of fields in the dataset which contain unstructured text to create a list of transactions associated with text tokens; and

treating the text tokens as attribute/values.

4. The computer implemented method of claim 1, wherein generating a transaction-by-attribute/value matrix comprises:

listing unique transactions along one axis of the transaction-by-attribute/value matrix;

listing unique attribute/values along another axis of the transaction-by-attribute/value matrix; and

populating the transaction-by-attribute/value matrix with non-zero values recording a number of times particular attribute/values occur in conjunction with particular transactions.

5. The computer implemented method of claim 1, wherein calculating the distinctiveness weights comprises measuring pointwise mutual information between respective transactions and attribute/values which occur in conjunction with one another.

6. The computer implemented method of claim 1, wherein calculating the materiality weights comprises using a logarithm of each transaction'"'"'s absolute monetary amount, for transactions with a non-zero monetary amount.

7. The computer implemented method of claim 1, wherein factorizing the weighted transaction-by-attribute/value matrix comprises:

performing a truncated or full singular value decomposition on the weighted transaction-by-attribute/value matrix; and

generating the transaction-by-concept matrix as an output of the truncated or full singular value decomposition.

8. The computer implemented method of claim 1, wherein generating output comprises:

creating a series of scatterplots, one per concept;

creating one axis for each scatterplot in the series of scatterplots to correspond to a logical ordering of transactions;

creating another axis for each scatterplot in the series of scatterplots to correspond to importance in the concept; and

representing transactions as individual points on the series of scatterplots.

9. The computer implemented method of claim 1, wherein generating output comprises creating a table of transactions which lists importance of the transactions in each dimension of the transaction-by-concept matrix.

10. A computer implemented method of unsupervised financial anomaly detection, comprising:

extracting, by a computer having stored a financial dataset, features of each transaction in the financial dataset;

generating, by the computer, a transaction-by-attribute/value matrix from the extracted features of each transaction in the financial dataset;

calculating, by the computer, distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the distinctiveness weights signify how much more likely an attribute/value is to occur in conjunction with a particular transaction than may be expected on the basis of chance;

calculating, by the computer, materiality weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the materiality weights signify an importance or significance of an amount, transaction or discrepancy with a particular transaction;

calculating, by the computer, combined weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the combined weight is calculated by adding each of the materiality weights for each attribute/value in the transaction-by-attribute/value matrix to respective distinctiveness weights not equal to zero of the distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix;

applying, by the computer, the combined weights to the transaction-by-attribute/value matrix to generate a weighted transaction-by-attribute/value matrix;

factorizing, by the computer, the weighted transaction-by-attribute/value matrix into a transaction-by-concept matrix; and

generating, by the computer, output of the transaction-by-concept matrix enabling anomalies to be identified;

wherein the extracting, the generating of the transaction-by-attribute/value matrix, the calculating of the distinctiveness weights, the calculating of the materiality weights, the calculating of the combined weights, the applying of the combined weights, the factorizing, and the generating of the output are performed regardless of a number, type, or monetary amount of transactions in the financial dataset, regardless of a number or type of features associated with each transaction in the financial dataset, and regardless of a provenance of the financial dataset.

11. The computer implemented method of claim 10, wherein extracting features of each transaction in the financial dataset comprises unpivoting the financial dataset to create a list of transactions associated with attribute/values.

12. The computer implemented method of claim 10, wherein extracting features comprises:

tokenizing contents of fields in the financial dataset which contain unstructured text to create a list of transactions associated with text tokens; and

treating the text tokens as attribute/values.

13. The computer implemented method of claim 10, wherein generating a transaction-by-attribute/value matrix comprises:

listing unique transactions along one axis of the transaction-by-attribute/value matrix;

listing unique attribute/values along another axis of the transaction-by-attribute/value matrix; and

populating the transaction-by-attribute/value matrix with non-zero values recording a number of times particular attribute/values occur in conjunction with particular transactions.

14. The computer implemented method of claim 10, wherein calculating the distinctiveness weights comprises measuring pointwise mutual information between respective transactions and attribute/values which occur in conjunction with one another.

15. The computer implemented method of claim 10, wherein calculating the materiality weights comprises using a logarithm of each transaction'"'"'s absolute monetary amount, for transactions with a non-zero monetary amount.

16. The computer implemented method of claim 10, wherein factorizing the weighted transaction-by-attribute/value matrix comprises:

performing a truncated or full singular value decomposition on the weighted transaction-by-attribute/value matrix; and

generating the transaction-by-concept matrix as an output of the truncated or full singular value decomposition.

17. The computer implemented method of claim 10, wherein generating output comprises:

creating a series of scatterplots, one per concept;

creating one axis for each scatterplot in the series of scatterplots to correspond to a logical ordering of transactions;

creating another axis for each scatterplot in the series of scatterplots to correspond to a range of values in the output of factorization; and

representing transactions as individual points on the series of scatterplots.

18. The computer implemented method of claim 10, wherein generating output comprises tabulating transactions in order of their importance in each dimension of the transaction-by-concept matrix.

19. A method of unsupervised financial analytical review and anomaly detection, comprising the steps of:

extracting features, by a computer having stored a financial dataset, of each transaction in the financial dataset;

generating, by the computer, a transaction-by-attribute/value matrix from the extracted features of each transaction in the financial dataset;

calculating, by the computer, distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the distinctiveness weights signify how much more likely an attribute/value is to occur in conjunction with a particular transaction than may be expected on the basis of chance;

calculating, by the computer, materiality weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the materiality weights signify an importance or significance of an amount, transaction or discrepancy with a particular transaction;

calculating, by the computer, combined weights for each attribute/value in the transaction-by-attribute/value matrix, wherein the combined weight is calculated by adding each of the materiality weights for each attribute/value in the transaction-by-attribute/value matrix to respective distinctiveness weights not equal to zero of the distinctiveness weights for each attribute/value in the transaction-by-attribute/value matrix;

applying, by the computer, the combined weights to the transaction-by-attribute/value matrix to generate a weighted transaction-by-attribute/value matrix;

factorizing, by the computer, the weighted transaction-by-attribute/value matrix into a transaction-by-concept matrix; and

generating, by the computer, output of the transaction-by-concept matrix so as to transform the financial dataset into useful actionable information enabling identification of transactions or groups of transactions that merit closer scrutiny;

wherein the extracting, the generating of the transaction-by-attribute/value matrix, the calculating of the distinctiveness weights, the calculating of the materiality weights, the calculating of the combined weights, the applying of the combined weights, the factorizing, and the generating of the output are performed regardless of a number, type, or monetary amount of transactions in the financial dataset, regardless of a number or type of features associated with each transaction in the financial dataset, and regardless of a provenance of the financial dataset.