COMPUTER-IMPLEMENTED SEMI-SUPERVISED LEARNING SYSTEMS AND 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.

Description

TECHNICAL FIELD

This document relates generally to constructing and using computer predictive models and more particularly to using semi-supervised learning systems and methods for generating predictive models.

BACKGROUND

Computer predictive models have found applicability in many diverse areas. However, difficulty arises in using predictive models when the training targets are not fully known. A non-limiting example where predictive models encounter unknown targets is when predictive models are to assess whether fraud may have occurred with respect to monetary-related transactions. Current predictive model approaches have difficulty in discerning legitimate monetary-related transactions from fraudulent ones.

SUMMARY

In accordance with the teachings provided herein, systems and methods for operation upon data processing devices are provided for performing semi-supervised learning. For example, a method and system 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 (e.g., using an approach such as outlier detection analysis) is performed using the target data set in order to determine which of the unknown targets are outliers. A comparison of the 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.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram depicting a computer-implemented system for constructing predictive models.

FIG. 2 is a block diagram depicting neural network training combined with an outlier detection process.

FIG. 3 is a process flow diagram depicting predictive model construction and use.

FIG. 4 is a process flow diagram depicting application of a neural network to generate scores.

FIGS. 5 and 6 are process flow diagrams depicting techniques for performing outlier detection.

FIG. 7 is a flowchart depicting an operational scenario for using a non-linear replicator neural network (e.g., a compression neural network) in determining outliers.

FIG. 8 is a process flow diagram depicting generation of a rank order list from an outlier detection process.

FIG. 9 is a process flow diagram depicting an iterative training process for improving a neural network model.

FIG. 10 is a process flow diagram depicting an iterative training process for improving an outlier detection process.

FIG. 11 is a process flow diagram depicting a system for analyzing financial or monetary type activities with respect to whether fraud may have occurred.

FIG. 12 is a process flow diagram depicting the combination of results from a neural network model with results from an outlier detection process.

FIG. 13 is a block diagram depicting a networked environment wherein users can interact with a predictive model construction system.

FIG. 14 is a block diagram depicting a stand-alone computer environment wherein a user can interact with a predictive model construction system.

DETAILED DESCRIPTION

FIG. 1 depicts at 30 a computer networked environment wherein users 32 can interact with a predictive model construction system 34. System 34 allows the users 32 to build predictive models even when targets 42 of the modeling process are only partially known. This can be useful in many different situations, such as in semi-supervised learning situations.

Semi-supervised situations involve generation of predictive models typically (but not always) by means of a small amount of labeled data and a large amount of unlabeled data (e.g., collectively target data set 42). Semi-supervised situations can arise because the cost associated with the labeling process may render a fully labeled training set impractical, whereas acquisition of unlabeled data is relatively inexpensive. In such situations, semi-supervised learning can be of great practical value.

The users 32 can interact with the predictive model construction system 34 through a number of ways, such as over one or more networks 36. A server 38 accessible through the network(s) 36 can host the predictive model construction system 34. Data store(s) 40 can store the data to be analyzed (e.g., target data set 42) as well as any intermediate or final data calculations and data results.

The predictive model construction system 34 can be a web-based tool that provides users with flexibility and functionality for generating predictive models when the targets 42 are only partially known. Moreover, the predictive model construction system 34 can be used separately or in conjunction with other software programs, such as with other predictive model construction techniques.

With reference to FIG. 2, neural network training 50 combined with an outlier detection process 52 allow for the identification of which targets in the target data set 42 should be focused upon in order to ascertain the true (or at least a more accurate) value of the targets. The increased knowledge for some of the targets results in an evolution in the fidelity of the model's or models' predictive capability. More specifically, model training evolves from a state where the targets are partially known to a state where at least more of the targets have been identified.

FIG. 3 illustrates a process flow by which predictive model construction can be performed using the target data set 42. As discussed above, the target data set 42 contains only a partial set of known targets (e.g., known targets 100 and unknown targets 110). As an example of one of the uses, the process flow can help in determining which of the potentially many unknown targets should be focused upon in order to determine the true values for these unknown targets.

Process 50 constructs models for predicting target values for the entries contained in the target data set 42. Many different types of predictive models can be constructed, such as artificial neural network predictive models. An artificial neural network is constructed of interconnecting neurons designed to model the target(s) and which changes its structure based on how it is trained, such as through the training process 50. More specifically, neural networks are non-linear statistical data modeling tools. They can be used to model complex relationships between inputs and targets or to find patterns in data.

The training process 50 begins with a set of interconnected nodes and alters the strength (e.g., weights) of the connections in the network to produce outputs. The training process 50 is provided with the input data about the target data set 42 and a cost function to be minimized. The cost function can be any function of the target values and predicted target values from the model under construction, such as the norm of the difference between the predicted and original target values. With the input data and the cost function, process 50 generates a neural network model using the known targets 100 (or at least a portion of the known targets 100). The unknown targets 110 (or at least a portion of the unknown targets 110) are used with the generated neural network model to generate values (i.e., results 130) for the unknown targets 110.

An outlier detection process 52 is performed using the target data set 42 for determining which of the unknown targets 110 are outliers. An unknown target being identified as an outlier by process 52 is an indication of anomalous activity, such as the possibility that fraud may have occurred.

In training process 52, both the input data set and the output data set are the target data set 42. The learning process tries to reproduce the input data as the target. Let the vector x=(x₁, x₂, . . . x_p)^T represent an observation with p inputs to the unsupervised learning process with a mean μ and covariance Σ. Let the vector y=(y₁, y₂, . . . y_p)^T represent the same observation with p outputs from the unsupervised learning process. The cost function can be any function of the input data and target output of the model under construction—such as in this example, the Mahalanobis distance between the inputs and outputs which is defined as √{square root over ((x−y)^TΣ⁻¹(x−y))}{square root over ((x−y)^TΣ⁻¹(x−y))}. The difference between the inputs and outputs is also defined as the reconstruction error of the unsupervised learning process and can be represented as E=(x−y). The Mahalanobis distance based reconstruction error can thus be defined as √{square root over (E^TΣ⁻¹E)}. The covariance matrix can be expressed in terms of eigenvalue matrix Λ and eigenvector U as Σ=UΛU^T. Therefore, √{square root over (E^TΣ⁻¹E)} can be computed from √{square root over ((EU)^TΔ⁻¹(EU))}{square root over ((EU)^TΔ⁻¹(EU))}. Σ can be noisy. The computation of the reconstruction error can be done by using the first m eigenvalues Λ of the covariance matrix Σ. An alternate approach is to use the first m eigenvalues and a small value for the remaining eigenvalues. A more general approach could be to weight the different eigenvalues differently in computing the reconstruction error. The inputs with the highest reconstruction error are deemed to be anomalous inputs.

Process 140 performs a comparison between the list of unknown targets (that have been identified as outliers) and the values 130 for the unknown targets that were generated by the neural network model. The subset 150 of unknown targets to investigate is determined based upon the comparison by process 140.

FIGS. 4 and 5 provide additional details respectively for the neural network training process 50 and the outlier detection process 52. With reference to FIG. 4, results 130 are generated when the neural network is applied at process 200 to the unknown target data set 110. For example, if the neural network is to detect whether fraud has occurred with respect to a financial or monetary transaction, then process 200 can generate target value scores for the unknown target data set 110 as the results 130 of the neural network.

The scores of the neural network model can assume a number of different formats. As an illustration, the scores can be binary in nature (e.g., a value of 1 to indicate that a transaction constitutes fraud; and a value of 0 to indicate that fraud has not occurred). The scores could also encompass a range of values. For example, a continuous range from 0-100 can indicate the degree by which an item in the target data set 42 can be considered fraudulent, wherein a value of 0 can establish the lower end of the fraud spectrum (i.e., not a fraud event) and a value of 100 can establish the upper end of the fraud spectrum.

FIG. 5 illustrates that the outlier detection process 52 can detect whether an unknown target should be considered an outlier with respect to the other unknown targets by performing processes 300 and 310. Process 300 compresses the data related to the unknown targets 110 by reducing the multi-dimensional data set to a lower dimension. Process 310 then reverses the compression process by expanding the data set from the lower dimension back to a higher dimension. Outliers can then be detected based upon how the data set performed during the compression and uncompression processes 300 and 310.

Different techniques are available in order to perform the compression and uncompression processes 300 and 310 such as by using nonlinear replicator neural networks (e.g., autoregressive neural networks, etc.) as shown at 350 in FIG. 6. Non-linear replicator neural networks are generally known in the art and discussed in such references as: S. Hawkins, H. X. He, G. J. Williams, and R. A. Baxter “Outlier detection using replicator neural networks,” Proceedings of the Fifth International Conference and Data Warehousing and Knowledge Discovery 2002; G. J. Williams, R. A. Baxter, H. X. He, S. Hawkins, and L. Gu “A comparative study of RNN for outlier detection in data mining,” Proceedings of the 2002 IEEE International Conference on Data Mining; and O. Abdel-Wahhab and M. Fahmy, “Image compression using multi-layer neural networks,” Proceedings of the 2nd IEEE Symposium on Computers and Communications (ISCC 1997).

FIG. 7 depicts an operational scenario for using a nonlinear replicator neural network in determining outliers based upon an input set of unknown targets. With reference to FIG. 7, a nonlinear replicator neural network compresses at step 400 the input unknown target data into a lower dimensional representation. Step 410 reverses the process by uncompressing the intermediate details back into the original dimensional representation. Step 420 examines the error that may have occurred between the uncompressed form and input data. The error is greater for those target values that had difficulty in being compressed into the lower dimensional representation. Target outliers are determined based upon the amount of target error that was determined in step 420.

As shown in FIG. 8, the results 120 of the outlier detection process 52 can be a rank order list. The rank order list contains the unknown targets (that were determined to be outliers) in an order dictated by the error amount that they exhibited from the compression and uncompression processes. In other words, the unknown targets that exhibited the greatest amount of error would be ranked higher in the outlier list than the unknown targets that did not exhibit as much error. The results 130 of the neural network training process 50 can also be a rank order list format.

Process 140 compares the results 130 (e.g., the unknown target scores) from the neural network with the rank order list from the outlier detection process 52. The comparison process 140 can use output subset criteria 500 in determining which of the unknown targets should be included in the subset 150 of unknown targets to investigate. As an example, the output subset criteria 500 can specify that the subset 150 should include only those unknown targets that have a relatively high score as determined by the neural network model as well as those that were highly ranked as outliers by the outlier detection process. The combination of the analyses performed by the neural network and by the outlier detection process enhances the fidelity of the selection of which of the unknown targets constitutes anomalous behavior. It should be understood that many different types of criteria can be used in order to determine the subset of unknown targets to investigate and should be based upon the application at hand.

FIG. 9 depicts that an iterative training process can be employed for improving the neural network training. For example, after the subset 150 of unknown targets to investigate has been established, process 600 can investigate the true status or values of the unknown targets in the subset 150.

The investigation process 600 is resource efficient because the investigation (at this stage) only needs to focus on the subset 150 and not the entire (and potentially large) corpus of unknown targets. As an example of process 600, analyzing the subset 150 can include examination by human analysts of the unknown targets to determine which targets within the data subset 150 constitute anomalous behavior.

In addition to or as a supplement to manual investigation, analysis of the target subset 150 can include examination of the subset 150 by another software program that uses more resource intensive techniques to determine a more accurate set of values for the unknown targets in the subset 150.

In any event, process 600 results in more accurate target values being produced for the data subset 150. The target results are fed into process 50 so that the neural network model can be improved which results in more accurate results 130 for use by the comparison process 140 in subsequent iterations.

FIG. 10 illustrates then the retraining process can also be applied to the outlier detection process 52. After process 600 has investigated the subset 150 of unknown targets and ascertained the true values for the subset 150, these target values 610 can be made available to improve at 700 the outlier detection process. The nonlinear replicator neural network that was used in the outlier detection process 52 can be improved by training the model based upon the target values that are now known for the subset.

The predictive model construction approaches described herein can be utilized for many different purposes where target values are unknown. For example, predictive models can be constructed in order to analyze whether fraud may have occurred with respect to a financial transaction. FIG. 11 depicts a computer-implemented system for analyzing financial or monetary type activities with respect to whether fraud may have occurred (e.g., purchase card fraud or tax evasion fraud).

The target data set 800 includes both known targets 802 and unknown targets 808. The known targets 802 in this example include known fraud targets 804 or instances of fraud (e.g., an entity had filed a false tax return; an entity had used a stolen credit card to purchase an item; etc.) as well as known non-fraud targets (e.g., an entity had filed a correct tax return; an entity had made a legitimate credit card purchase; etc.). In this example, the number of legitimate type targets overwhelmingly outnumbers the known fraud targets. The unknown targets 808 constitute entities, such as individuals or organizations, whose transactions (e.g., filing a tax return; making a credit card purchase; etc.) do not contain an indication of whether fraud has occurred.

As discussed above, the target data set 800 contains only a partial set of known targets (e.g., known targets 802 and unknown targets 808). Process 810 trains a model (e.g., neural network) for predicting values or scores that are indicative of whether fraud has occurred with respect to the known targets 802.

With the input data and the cost function, process 810 generates a neural network model using the known targets 802. The unknown targets 808 are used with the generated neural network model to generate values (i.e., fraud-indicative scores 830) for the unknown targets 808.

An outlier detection process 820 is performed using the target data set 800 in order to determine which of the unknown targets 808 are outliers. An unknown target being identified in the results 840 as an outlier by process 820 is an indication of fraudulent activity. More specifically in this example, the results 840 of the outlier detection process 820 can be a rank order list. The rank order list contains the unknown targets that were determined to be outliers ordered by the error amount that they exhibited from the compression and uncompression processes. The unknown targets exhibiting a higher amount of error are considered more likely to involve fraudulent activity than the unknown targets that exhibit a smaller amount of error.

Process 850 performs a comparison between the list of unknown targets (that have been identified as outliers) and the scores 830 for the unknown targets that were generated by the neural network model. The subset 870 of unknown targets to investigate is determined based upon the comparison by process 850.

Process 850 compares the fraud-indicative scores 830 from the neural network with the rank order list from the outlier detection process 820. The comparison process 850 uses subset criteria 860 in determining which of the unknown targets should be included in the subset 870 of unknown targets to investigate. In this example, the subset criteria 870 specifies that the subset 860 should include only those unknown targets that have a relatively high score as determined by the neural network model as well as those that were highly ranked as outliers by the outlier detection process.

Process 880 performs an investigation of the unknown targets that passed the fraud-indicative criteria 860. As a result of the investigation 880, the true target values become known for the subset 870 and are used to retrain the neural network at 810. The retrained neural network is then applied to the remaining unknown targets in the target data set 800 in order to generate new fraud indicative scores.

Similarly, the target values 890 that are known for the subset 870 can also be used to refine the outlier detection process as indicated in 1000. The improved outlier detection process 820 performs outlier detection upon the remaining unknown targets 808 in order to produce a new fraud-indicative outlier list. Process 850 then uses the new fraud-indicative scores as well as the new fraud-indicative scores generated by process 810 when it is to perform its comparison operations. Process 850 results in a new subset of unknown targets to investigate. The investigation and retraining operations can continue until the models have reached a particular level of precision and/or until no more investigations are desired.

While examples have been used to disclose the invention, including the best mode, and also to enable any person skilled in the art to make and use the invention, the patentable scope of the invention is defined by claims, and may include other examples that occur to those skilled in the art. Accordingly the examples disclosed herein are to be considered non-limiting. As an illustration, FIG. 12 shows that the results 120 and 130 that are respectively generated via a processes 52 and 50 can be combined by process 1100. The combination of the results can be used to construct a more robust overall predictive model to determine whether an event or activity constitutes anomalous behavior. More specifically, the combination of the results can be achieved through optimizing the contributions of the results from processes 50 and 52 to the final overall model in terms of the model performance on known target data set.

As another illustration, the systems and methods disclosed herein could use different types of models as supervised models and unsupervised models. For example, linear regression models and logistic regression models can be used as supervised models in the disclosed operational scenarios; and principal component analysis type models can be used as unsupervised models in the disclosed operational scenarios.

As yet another illustration, the systems and methods disclosed herein may be implemented on various types of computer architectures, such as for example on a networked system, on a single general purpose computer, etc. For example, FIG. 13 depicts a networked environment wherein users 32 can interact with a predictive model construction system 34. The same server or different servers can contain various software instructions 1200 (e.g., software instructions for neural network training, software instructions for outlier detection, etc.) or modules of the predictive model construction system 34. Data store(s) 40 can store the data to be analyzed as well as any intermediate or final data calculations and data results of such software instructions or modules.

It should be understood that the analytical systems described herein (e.g., tax fraud analysis system, purchase card fraud analysis system, etc.) can be implemented in other ways, such as on a stand-alone computer for access by a user as shown at 1300 in FIG. 14.

It is further noted that the systems and methods may include data signals conveyed via networks (e.g., local area network, wide area network, internet, combinations thereof, etc.), fiber optic medium, carrier waves, wireless networks, etc. for communication with one or more data processing devices. The data signals can carry any or all of the data disclosed herein that is provided to or from a device.

Additionally, the methods and systems described herein may be implemented on many different types of processing devices by program code comprising program instructions that are executable by the device processing subsystem. The software program instructions may include source code, object code, machine code, or any other stored data that is operable to cause a processing system to perform methods described herein. Other implementations may also be used, however, such as firmware or even appropriately designed hardware configured to carry out the methods and systems described herein.

The systems' and methods' data (e.g., associations, mappings, etc.) may be stored and implemented in one or more different types of computer-implemented ways, such as different types of storage devices and programming constructs (e.g., data stores, RAM, ROM, Flash memory, flat files, databases, programming data structures, programming variables, IF-THEN (or similar type) statement constructs, etc.). It is noted that data structures describe formats for use in organizing and storing data in databases, programs, memory, or other computer-readable media for use by a computer program.

The systems and methods may be provided on many different types of computer-readable media including computer storage mechanisms (e.g., CD-ROM, diskette, RAM, flash memory, computer's hard drive, etc.) that contain instructions (e.g., software) for use in execution by a processor to perform the methods' operations and implement the systems described herein.

The computer components, software modules, functions, data stores and data structures described herein may be connected directly or indirectly to each other in order to allow the flow of data needed for their operations. It is also noted that a module or processor includes but is not limited to a unit of code that performs a software operation, and can be implemented for example as a subroutine unit of code, or as a software function unit of code, or as an object (as in an object-oriented paradigm), or as an applet, or in a computer script language, or as another type of computer code. The software components and/or functionality may be located on a single computer or distributed across multiple computers depending upon the situation at hand.

It should be understood that as used in the description herein and throughout the claims that follow, the meaning of “a,” “an,” and “the” includes plural reference unless the context clearly dictates otherwise. Also, as used in the description herein and throughout the claims that follow, the meaning of “in” includes “in” and “on” unless the context clearly dictates otherwise. Finally, as used in the description herein and throughout the claims that follow, the meanings of “and” and “or” include both the conjunctive and disjunctive and may be used interchangeably unless the context expressly dictates otherwise; the phrase “exclusive or” may be used to indicate situation where only the disjunctive meaning may apply.

Claims

1. A processor-implemented method for determining a subset of unknown targets to investigate, comprising: receiving a target data set;wherein the target data set includes known targets and unknown targets;generating a neural network model using the known targets;using the unknown targets with the neural network model to generate values for the unknown targets;performing outlier detection analysis using the target data set in order to determine which of the unknown targets are outliers;performing a comparison of list of outlier unknown targets with the values for the unknown targets that were generated by the neural network model;wherein the subset of unknown targets to investigate is determined based upon the comparison.

2. The method of claim 1, wherein the target data set is indicative of entities to investigate with respect to fraud.

3. The method of claim 2, wherein the fraud includes tax evasion fraud or purchase card fraud.

4. The method of claim 2, wherein the values for the unknown targets that were generated by the neural network model are scores indicative of whether fraud has occurred.

5. The method of claim 1, wherein the performing of the outlier detection analysis includes compressing at least a portion the unknown targets into a lower dimensional representation and uncompressing intermediate results of the compressing process into a higher dimensional representation; determining whether an unknown target is an outlier based error that exists between the uncompressed form and the unknown target.

6. The method of claim 5, wherein said compressing is performed by using a compression neural network approach.

7. The method of claim 1, wherein the performing of the outlier detection analysis includes using a nonlinear replicator neural network approach to determine which of the unknown targets are outliers.

8. The method of claim 1, wherein the list of outlier unknown targets is a rank order list of the unknown targets that was generated through use of a compression neural network approach.

9. The method of claim 1, wherein the performing of the comparison includes comparing based upon a pre-specified criterion or criteria the list of outlier unknown targets with the values for the unknown targets that were generated by the neural network model.

10. The method of claim 9, wherein the pre-specified criterion or criteria contain criterion or criteria for including in the subset of unknown targets to investigate those unknown targets that have a value that satisfies a value scoring threshold and that have been determined to be outliers by the outlier detection analysis.

11. The method of claim 1, wherein targets in the determined subset of unknown targets are investigated; wherein the targets in the determined subset become known because of the investigation.

12. The method of claim 11 further comprising: retraining the neural network using the known targets and the targets that have become known because of the investigation.

13. The method of claim 12 further comprising: using the retrained neural network upon the unknown targets in order to generate values for the unknown targets.

14. The method of claim 13 further comprising: performing a second outlier detection analysis in order to determine a second list of which of the unknown targets are outliers;performing a second comparison of the second list of outlier unknown targets with the values for the unknown targets that were generated by the retrained neural network model;wherein a second subset of unknown targets to investigate is determined based upon the second comparison.

15. The method of claim 1 further comprising: combining values for the unknown targets that were generated by the neural network model with outlier results generated by the outlier detection process in order to optimize the overall model performance.

16. The method of claim 1, wherein an unknown target being identified as an outlier by the outlier detection process is an indication of anomalous activity.

17. The method of claim 16, wherein an unknown target being identified as an outlier by the outlier detection process is an indication of fraudulent activity.

18. A data signal that is transmitted using a network, wherein the data signal includes the subset of unknown targets to investigate of claim 1; wherein the data signal comprises packetized data that is transmitted across the network.

19. A processor-implemented system for determining a subset of unknown targets to investigate, comprising: a data store to store a target data set;wherein the target data set includes known targets and unknown targets;model generation software instructions configured to generate a neural network model using the known targets;wherein the unknown targets are used with the neural network model to generate values for the unknown targets;software instructions to perform outlier detection analysis using the target data set in order to determine which of the unknown targets are outliers;comparison software instructions configured to perform a comparison of list of outlier unknown targets with the values for the unknown targets that were generated by the neural network model;wherein the subset of unknown targets to investigate is determined based upon the comparison.

20. Computer software stored on one or more computer readable mediums, the computer software comprising program code for determining a subset of unknown targets to investigate, said method comprising: receiving a target data set;wherein the target data set includes known targets and unknown targets;generating a neural network model using the known targets;using the unknown targets with the neural network model to generate values for the unknown targets;performing outlier detection analysis using the target data set in order to determine which of the unknown targets are outliers;performing a comparison of list of outlier unknown targets with the values for the unknown targets that were generated by the neural network model;wherein the subset of unknown targets to investigate is determined based upon the comparison.