Computer-implemented semi-supervised learning systems and methods

• US 8,346,691 B1
• Filed: 09/06/2007
• Issued: 01/01/2013
• Est. Priority Date: 02/20/2007
• Status: Active Grant

• Current Assignee: SAS Institute Incorporated
• Original Assignee: SAS Institute Incorporated
• Inventors: Subramanian, Revathi, Desai, Vijay S., Gong, Hongrui
• Primary Examiner(s): Wong, Lut
• Application Number:
  US11/850,861
• Time in Patent Office: 1,944 Days
• Field of Search: 706/15, 706/12
• US Class Current: 706/15
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

Computer-implemented systems and methods for determining a subset of unknown targets to investigate. For example, a method can be configured to receive a target data set, wherein the target data set includes known targets and unknown targets. A supervised model such as a neural network model is generated using the known targets. The unknown targets are used with the neural network model to generate values for the unknown targets. Analysis with an unsupervised model is performed using the target data set in order to determine which of the unknown targets are outliers. A comparison of list of outlier unknown targets is performed with the values for the unknown targets that were generated by the neural network model. The subset of unknown targets to investigate is determined based upon the comparison.

Claims

1. A processor-implemented method for determining unknown targets to investigate, comprising:

receiving, using one or more processors, a target data set that includes known targets and unknown targets;

generating, using the one or more processors, a neural network model using the known targets, wherein generating includes training the neural network model using the known targets;

using the neural network model to score the unknown targets in the target data set to generate unknown target scores, wherein scoring is performed using the one or more processors;

determining, using the one or more processors, a neural network target set, wherein the neural network target set contains unknown targets having unknown target scores that meet a threshold;

performing, using the one or more processors, outlier detection analysis on the unknown targets in the target data set to determine a sorted list of outlier unknown targets, wherein the sorted list is sorted according to an outlying degree;

determining, using the one or more processors, an outlier detection target set, wherein the outlier detection target set includes a portion of the unknown targets in the sorted list, and wherein inclusion in the portion is based upon the outlying degree;

determining, using the one or more processors, a subset of unknown targets to investigate, wherein the subset of unknown targets to investigate contains unknown targets that appear in both the neural network target set and the outlier detection target set;

applying, using the one or more processors, labels to the unknown targets in the subset based upon an investigative analysis;

retraining, using the one or more processors, the neural network model using the labeled unknown targets; and

using the retrained neural network model and the outlier detection analysis to determine a new subset of unknown targets to investigate, wherein determining the new subset is performed using the one or more processors.

2. The method of claim 1, wherein the target data set includes one or more fraudulent entities or transactions to investigate.

3. The method of claim 2, wherein the unknown target scores are scores that indicate whether fraud has occurred.

4. The method of claim 1, wherein performing the outlier detection analysis includes:

compressing, using the one or more processors, one or more unknown targets into a lower dimensional representation;

uncompressing, using the one or more processors, the one or more unknown targets into a higher dimensional representation; and

determining, using the one or more processors, that one of the unknown targets is an outlier based upon a compression/uncompression error.

5. The method of claim 4, wherein compressing includes using a compression neural network.

6. The method of claim 1, wherein performing the outlier detection analysis includes using a nonlinear replicator neural network.

7. The method of claim 1, wherein when the subset of unknown targets are investigated they become newly known targets.

8. The method of claim 1, wherein the outlier unknown targets are identified as possibly anomalous.

9. The method of claim 8, wherein the outlier unknown targets are identified as possibly fraudulent activity.

10. The method of claim 1, wherein the target data set includes transaction data.

11. The method of claim 10, wherein one or more of the known targets in the target data set is labeled as a fraudulent transaction.

12. A system for determining unknown targets to investigate, comprising:

one or more processors;

a computer-readable storage medium containing instructions configured to cause the one or more processors to perform operations, including;

receiving a target data set that includes known targets and unknown targets;

generating a neural network model using the known targets, wherein generating includes training the neural network model using the known targets;

using the neural network model to score the unknown targets in the target data set to generate unknown target scores;

determining a neural network target set, wherein the neural network target set contains unknown targets having unknown target scores that meet a threshold;

performing outlier detection analysis on the unknown targets in the target data set to determine a sorted list of outlier unknown targets, wherein the sorted list is sorted according to an outlying degree;

determining an outlier detection target set, wherein the outlier detection target set includes a portion of the unknown targets in the sorted list, and wherein inclusion in the portion is based upon the outlying degree;

determining a subset of unknown targets to investigate, wherein the subset of unknown targets to investigate contains unknown targets that appear in both the neural network target set and the outlier detection target set;

applying labels to the unknown targets in the subset based upon an investigative analysis;

retraining the neural network model using the labeled unknown targets; and

using the retrained neural network model and the outlier detection analysis to determine a new subset of unknown targets to investigate.

13. A computer-program product, tangibly embodied in a non-transitory machine-readable storage medium, including instructions configured to cause a data processing apparatus to:

receive a target data set that includes known targets and unknown targets;

generate a neural network model using the known targets, wherein generating includes training the neural network model using the known targets;

use the neural network model to score the unknown targets in the target data set to generate unknown target scores;

determine a neural network target set, wherein the neural network target set contains unknown targets having unknown target scores that meet a threshold;

perform outlier detection analysis on the unknown targets in the target data set to determine a sorted list of outlier unknown targets, wherein the sorted list is sorted according to an outlying degree;

determine an outlier detection target set, wherein the outlier detection target set includes a portion of the unknown targets in the sorted list, and wherein inclusion in the portion is based upon the outlying degree;

determine a subset of unknown targets to investigate, wherein the subset of unknown targets to investigate contains unknown targets that appear in both the neural network target set and the outlier detection target set;

apply labels to the unknown targets in the subset based upon an investigative analysis;

retrain the neural network model using the labeled unknown targets; and

use the retrained neural network model and the outlier detection analysis to determine a new subset of unknown targets to investigate.