MULTIVARIATE RISK ASSESSMENT VIA POISSON SHELVES

Abstract

Detecting fraudulent activity can be a complex, manual process. In this paper, we adapt statistical properties of count data in a novel algorithm to uncover records exhibiting high risk for fraud. Our method identifies shelves, partitioning data under the counts using a Student's t-distribution. We apply this methodology on a univariate dataset including cumulative results from phone calls to a customer service center. Additionally, we extend this technique to multivariate data, illustrating that the same method is applicable to both univariate and multivariate data.

Description

BACKGROUND

Field

Embodiments of the present invention relate to risk assessment using Poisson Shelves.

Background

In fraud prevention, it desirable to not only determine if certain transactions are at risk, but to also do so in an accurate and automated fashion. Suppose a customer service center records the number of times a particular phone number calls per day. If the same number calls multiple times over a series of days, is this activity deemed unusual and risky? What criteria should a company apply to automatically determine if calls from one number are more or less risky than other calls? How should a company treat future calls from a risky number? These are the kinds of questions that must be answered daily by security engineers.

Published methods on unsupervised fraud detection span many disciplines, and generally center around clustering and pattern identification [Phua et al., 2010]. More recent surveys of fraud detection methods such as [Lata et al., 2015] contain very few active unsupervised methods. SmartSifter is the most similar to our multivariate methodology, where a mixture model is used to combine different modes in the cumulative data. SmartSifter is primarily used for discovering financial fraud [Yamanishi et al., 2004].

BRIEF SUMMARY OF THE DISCLOSURE

Accordingly, the present invention is directed to risk assessment using Poisson Shelves that obviates one or more of the problems due to limitations and disadvantages of the related art.

Herein, we consider relevant works to our topic and why there is a need for our technique. We then follow with a methodology, considering both the univariate and multivariate cases, where we describe how to obtain partition values, to detect shelves, and what are considered extreme outliers given a shelf. We then apply our technique to IVR call datasets. We finally discuss the implications of our results.

In accordance with the purpose(s) of this invention, as embodied and broadly described herein, this invention, in one aspect, relates to a computer program product comprising computer executable code embodied in a non-transitory computer readable medium that, when executing on one more computing devices performs a method of performing risk assessment using count data. The method includes modeling the count data using a Poisson distribution; calculating a threshold value, w; partitioning the count data based on the threshold w; determining a number of surviving records from among the portioned count data based on the threshold w; identifying at least one shelf in the surviving records; and defining outliers for risk assessment according to the at least one the shelf.

In another aspect, the invention relates to a computer program product of identifying outliers for risk assessment using count data “Records” having y values. The method of this aspect includes computing a sample mean {circumflex over (λ)} of y values in Records; calculating threshold value ω={circumflex over (λ)}+n√{circumflex over (λ)}, from n=0 to n_max, where n is the number of standard deviation units in ω; for every ω, computing a number of surviving records according to |{r∈Records: r[y]>ω}|; identifying at least one shelf in surviving data by identifying an abrupt variation in generative parameters of a sequence of data; and defining outliers for further risk assessment according to {r∈Records: r[y]>{circumflex over (λ)}+n*√{circumflex over (λ)}}.

In yet another aspect, the invention relates to a computer program product of identifying outliers for risk assessment in multivariable count data having a first count variable x_i and a second count variable y_i. The method of this aspect includes calculating a mixture of independent Poissons: calculating a threshold value ω for every combination of mixture component c and dimension i; identifying at least one shelf in surviving data by identifying an abrupt variation in generative parameters of a sequence of data; defining outliers in component c; and providing a visual representation of the outliers.

In accordance with the purpose(s) of this invention, as embodied and broadly described herein, this invention, in one aspect, a method of performing risk assessment using count data including modeling the count data using a Poisson distribution; calculating a threshold value, ω; partitioning the count data based on the threshold ω; determining a number of surviving records from among the portioned count data based on the threshold ω; identifying at least one shelf in the surviving records; and defining outliers for risk assessment according to the at least one the shelf.

In yet another aspect, the invention relates a method of identifying outliers for risk assessment using count data “Records” having y values, the method including computing a sample mean {circumflex over (λ)} of y values in Records; calculating threshold value ω={circumflex over (λ)}+n√{circumflex over (λ)}, from n=0 to n_max, where n is the number of standard deviation units in ω; for every ω, computing a number of surviving records according to |{r∈Records: r[y]>ω}|; identifying at least one shelf in surviving data by identifying an abrupt variation in generative parameters of a sequence of data; and defining outliers for further risk assessment according to {r∈Records: r[y]>{circumflex over (λ)}n*√{circumflex over (λ)}}.

In yet another aspect, the invention relates to a method of identifying outliers for risk assessment in multivariable count data having a first count variable x_i and a second count variable y_i, the method including calculating a mixture of independent Poissons: calculating a threshold value ω for every combination of mixture component c and dimension i; identifying at least one shelf in surviving data by identifying an abrupt variation in generative parameters of a sequence of data; defining outliers in component c; and providing a visual representation of the outliers.

Additional advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

Further embodiments, features, and advantages of the risk assessment using Poisson Shelves, as well as the structure and operation of the various embodiments of the risk assessment using Poisson Shelves*, are described in detail below with reference to the accompanying drawings.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only, and are not restrictive of the invention as claimed.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying figures, which are incorporated herein and form part of the specification, illustrate risk assessment using Poisson Shelves. Together with the description, the figures further serve to explain the principles of the risk assessment using Poisson Shelves described herein and thereby enable a person skilled in the pertinent art to perform and use risk assessment using Poisson Shelves.

FIG. 1 is a graph showing cumulative call events for a series of approximately 17,000 calls to an IVR (Integrated Voice Recognition) system.

FIG. 2. is a plot showing n against the number of surviving records.

FIG. 3 is a correlation heatmap of the complete data in Table 2.

FIG. 4A shows surviving records for AcctCnt.

FIG. 4B shows the posterior probability of the current run r_t's length at each step.

FIG. 4C shows the probability of a changepoint occurring by looking N_ω=15 samples prior.

FIG. 5 illustrates determining the number of components for the mixture model.

FIG. 6 is a histogram of the distribution of points amongst the mixture model components.

FIG. 7 illustrates performing PCA to reduce the number of dimensions.

FIG. 8 are survival graphs for component 15 of a dataset of 330 datapoints.

FIG. 9 shows data like in FIG. 7 except points are grayscaled depending on if they are outliers or not.

FIG. 10 shows dataset univariate plots of the number of surviving records for every field given n where n is the number of standard deviation units in ω.

FIG. 11 shows dataset univariate plots of the number of surviving records for every field given n where n is the number of standard deviation units in ω.

FIG. 12 shows dataset multivariate plots for component 1, which has 4,595 points.

DETAILED DESCRIPTION

Reference will now be made in detail to embodiments of the risk assessment using Poisson Shelves with reference to the accompanying figures The same reference numbers in different drawings may identify the same or similar elements.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention cover the modifications and variations of this invention provided they come within the scope of the appended claims and their equivalents.

A method for determining the risk of a record using event frequency is described herein. An example applying the method to an example phone call problem is also described herein. The method is to help fraud and security analysts build systems using frequency in determining when future events may present risk. This method provides easy understanding and visualization for fraud and security analysts that may not have extensive background in statistics and outlier detection methodologies. The technique described herein may be categorized as a pattern recognition algorithm for detecting classes of data that appear to connote risk.

The intuition behind uncovering risk via frequency is best highlighted with a simple example. Consider every credit card transaction to be an event. Over a period of time, say an hour, we may expect to see several transactions for a single, given card. However, if 50 to 100 transactions occur for this card in that hour, the card may be at risk for fraud. Thus, counting the number of events is an intuitive way to determine risk. This intuition may be used to generate a rule for revealing risky records. As another example, consider FIG. 1, which shows cumulative call events for a series of approximately 17,000 calls to an IVR (Integrated Voice Recognition) system. In FIG. 1, the number of calls to the system is aggregated from each phone number over a period of time (one month). In the vast majority of cases, a phone number calls the IVR fewer than 10 times. However, there are some suspicious phone numbers that call more than 600 times in a one month period; this is approximately 20 calls made by one phone number per day!

The methodology described herein involves modeling these counts using a Poisson distribution and considering various thresholds for partitioning the data. The partitions are based on the mean and standard deviation of the data and also incorporate user choice, as what is considered fraudulent may differ from application to application. Records that are greater than a threshold value are considered surviving records, and plotting threshold parameters against the number of surviving records reveals distinct, shelf-like behavior. We exploit this phenomena to identify extreme outliers in detecting possible cases of fraud. We also modify and apply this methodology for multivariate datasets.

Truly risky calls are rare, and we want to minimize false positives. Thus, we must take special care in classifying the risk of individual calls. For example, suppose that 1 out of 1000 of the calls truly exhibit risky characteristics. Furthermore, suppose there exists a method that we can use to classify risk of calls from a given number with an accuracy of 98% with a false positive rate of 1%. Using this method to classify whether or not a given phone number is risky, what are the chances that the phone number is actually risky? Applying Bayes' Law:

$\begin{array}{c}P\left(\mathrm{risk}|\mathrm{classifier}\right)=\frac{P\left(\mathrm{classifier}|\mathrm{risk}\right)P\left(\mathrm{risk}\right)}{p\left(\mathrm{classifier}\right)}\\ =\frac{0.98-0.001}{0.01}=9.8\%\end{array}$

There is less than a ten percent likelihood that a given phone number identified as at risk for fraud is actually at risk under the assumptions given above.

Traditional classification methods (such as decision trees or SVMs) alone can thus be problematic, potentially causing a high likelihood of false positives. The technique described herein provides another level of assessment to pinpoint risky data before application of traditional classification methods of uncovering risk.

The technique described herein involves modeling count data using a Poisson distribution. Partition values are generated using the mean, standard deviation, and user choice. Because a Poisson distribution is used, the mean and variance are equivalent. By counting and plotting the number of data points greater than certain partition values (which we call surviving records), a distinct shelf-like pattern may emerge, which is the basis for our methodology. We begin by considering the univariate case and expand into multivariate data, as described below.

A random variable X has a Poisson distribution with parameter λ>0 if X can have integer values x=0, 1, 2, . . . with probability:

$P\left(X=x\right)=\frac{{\lambda }^x{e}^{-\lambda }}{x!}$

A feature of the Poisson distribution is that the mean (λ) is equivalent to the variance. That is:

E(X)=var(X)=λ

Counting events that occur independently over some period of time is typically modeled using the Poisson distribution. The phone call example discussed above can be modeled using a Poisson distribution.

A process satisfying the following three assumptions:

• • the probability of at least one occurrence of the event in a time interval is proportional to the interval length;
  • the probability of two or more occurrences of the event in a small time interval is negligible; and
  • the number of occurrences of the event in disjoint time intervals is mutually independentis called a Poisson process. In the phone number example, the random variable is the number of times a particular phone calls the service center. The longer the time period, the higher the chance that the same phone number will call multiple times. In a small time interval, the chances are very low. Finally, the number of calls in disjoint time intervals is also mutually independent. Among the objectives of the process described herein are to partition the counts so we can distinguish risky data from non-risky records and to develop a partition methodology that is independent of the units of counting; the same method should be applicable to different types of count data.

We will use the parameter λ from the Poisson distribution, which we obtain from the sample mean of the dataset. Let N represent the number of samples in the time period of interest and x_i be a datapoint. We calculate λ by:

$\tilde{\lambda }=\frac{{\sum }_{i=0}^N{x}_i}{N}$

We partition the data based on a threshold value, ω:

ω={circumflex over (λ)}+n×stdev(X)={circumflex over (λ)}+n√{square root over (λ)}

where n=0, 1, 2, . . . is a standard deviation (stdev) multiplier and is user chosen. By varying the value of n, we can determine the number of data points greater than co. See Table 1 for the number of surviving records given n for a sample dataset.

TABLE 1
	Surviving
n	Records	w
0	6831.0	1.74
1	690.0	4.32
2	279.0	6.90
3	133.0	9.48
4	91.0	12.06
5	68.0	14.64
6	47.0	17.22
7	38.0	19.80
8	30.0	22.37
9	28.0	24.95

Table 1 shows the number of surviving records given n for the sample dataset in FIG. 2. Also listed are the threshold value, ω, for every n as determined by λ+n√{square root over (λ)}. This table shows only the first 10 values of n.

If we plot n against the number of surviving records (see FIG. 2), a distinct shelf-like pattern emerges starting from approximately n=27 to n=45. FIG. 2 is a plot showing the number of surviving records given n where n indicates the number of standard deviation units to use to calculate the threshold value, ω. The y-axis is displayed in the log-scale. AcctCnt is the name of a field in our data. We use these shelves to identify extreme outliers in our data. In this example dataset, only 1 datapoint survives after the shelf.

We define an outlier as a data point that survives after the first shelf. Shelves can be detected by computing the probability distribution of the length of a current “run”. For example, one could consider n=0 to n=26 to be one run and n=27 to n=45 to be another run. At n=27, a changepoint has occurred, where a changepoint is an abrupt variation in the generative parameters of a sequence of data. Herein, we use a Student's t-distribution and the concepts in [Adams and MacKay, 2007] to determine changepoints in our data, however, any suitable method may be used to identify changepoints as used herein, including concept drift detection techniques.

Not every run may be equivalent to a shelf. A linear regression analysis on log counts could be performed to determine if the slope of a run is below a certain threshold. A shelf and its length can also be specified via trial and error or by the user per the needs of the application. For example, a user could argue that, although short, a shelf occurs from n=24 to n=26. One could experiment with different Student's t-distribution parameters or with different members of the exponential distribution family entirely to detect shelves setting the application. Other distributions besides a Student's t-distribution can and should be considered and may be more applicable to survival graphs.

The method described herein can be applied to any univariate count of events satisfying the requirements of a Poisson process. An example, as applied to the phone numbers is described below. We call a function that returns the final n* value of the first detected shelf given a list of n's and corresponding numbers of surviving records shelfDetect( ).

Let Records be defined as follows:

Records ={(x, y): phone # x calls y times}

To determine outliers:

• • 1. Compute {circumflex over (λ)}, the sample mean of the y values in Records;
  • 2. From n 0 to n_max, calculate ω={circumflex over (λ)}+n√{circumflex over (λ)};
  • 3. For every ω, compute the # of surviving records:

|{r∈Records: r[y]>ω}|

• • 4. Run shelfDetect( )to get n*
  • 5. Outliers are defined as:

{r∈Records : r[y]>{circumflex over (λ)}+n*√{circumflex over (λ)}}

An aspect of risk detection is the ability to evaluate the interdependency of multiple count variables. This is a complex problem because multivariate distributions for count data have behaviors that can be difficult to analyze and understand or have severe restrictions. For example, multivariate count models where the marginal distributions are Poisson can only apply to situations where there are positive dependencies between the variables. According to the technique described herein, regardless of the number of multivariate dimensions, a graph of n against the number of surviving records can be produced, as shown in FIG. 2. To create this graph, ω in the multivariate case is determined. Once one decides how to derive ω, the methods described above can be used to reveal outliers in the joint distribution of multiple columns of data as easily as it can with univariate data. A technique that can be used in calculating ω is described below. Other techniques can be substituted for calculating ω.

In [Inouye et al., 2017], several classes of multivariate distributions derived from the univariate Poisson are introduced. The first class where the marginal distributions are Poisson can apply to situations where there are positive dependencies between the variables. Negative relationships are always a possibility in real-world datasets, potentially restricting the use of this option. The second class involves a mixture of independent Poissons:

${\mathbb{P}}_{\mathrm{MixedPoi}}\left(x\right)={\int }_{{ℝ}_{++}^d}g\left(\lambda \right)\prod _{i=1}^d{\mathbb{P}}_{\mathrm{Poise}}\left(x_i|{\lambda }_i\right)d\lambda$

where x=[x₁, x₂, . . . x_d], d is the dimension of the data, λ=[λ₁, λ₂, . . . λ_d] is a vector where λ_i is the parameter of the Poisson distribution for x_i, and g(λ) is the mixing distribution. As we have counting data, we work in the domain of positive, real numbers, notated as R++.

Poisson mixture generalizations allow for positive and negative variable dependencies as opposed to Poisson marginal distributions. They can also model overdispersion where the variance of the data is larger than the mean, a common feature in real world data.

Typically, expectation-maximization is used to refine the parameters of the mixture model, iterating until convergence. For every combination of mixture component c and dimension i, ω can be calculated:

ω_ic={circumflex over (λ)}_ic+n√{square root over ({circumflex over (λ)}_ic)} to create the survival graphs like in FIG. 2.

Applying this procedure to the phone numbers example once more, let Records now be a set containing elements of the form (x, y₁, y₂, . . . y_d) where phone number x is associated with y_i count data. For example, yi could be how many times phone number x called in the time period, and y₂ could be how many social security numbers are referenced by phone number x. Let Records_c represent the records in component c. An outlier in a component c is then defined as:

{r∈Records_ic: ∃i s.t. r|y_i|>λ_ic+n*_ic√{square root over (λ_ic)}}

where y_i is the ith dimension of count data, and n*_ic is the final n value of the first detected shelf in the ith dimension for component c.

The methods described herein can be applied to historical call records to surface phone numbers that may have been (or are being) used for fraudulent access to a company's systems. As an example, let's look at the anonymized output of an IVR system servicing a financial company containing 16,819 phone numbers. Historical calls are accumulated and counted in the various columns of data. Some example values appear in Table 2.

TABLE 2
AcctCnt	CallCnt	Auth2Cnt	Auth3Cnt	Anth4Cnt	EPCnt1	EPCnt3
5	15	5	5	3	0	1
3	57	3	3	8	0	0
4	8	3	3	4	0	0
5	14	5	5	6	1	0
3	8	2	2	4	0	0

Negative relationships are present, deterring use of marginal Poisson distributions in [Inouye et al., 2017].

FIG. 3 is a correlation heatmap of the complete data in Table 2. Table 2 shows some sample data from a set of approximately 17,000 phone numbers from an IVR system over the course of a month. Every row represents counts for a single unique phone number. For example, whatever phone number is associated with the first row has called 15 times (CallCnt) and has referred to 5 unique accounts (AcctCnt) in its calls. Negative relationships are present, making use of marginal Poisson distributions less desirable.

We apply the process described herein to the data in Table 2 and plot the number of surviving records for every field given n where n is the number of standard deviation units in co. Survival plots for every field are available in FIGS. 10-12. FIG. 10 shows dataset univariate plots of the number of surviving records for every field given n where n is the number of standard deviation units in ω. FIG. 11 shows dataset univariate plots of the number of surviving records for every field given n where n is the number of standard deviation units in ω. FIG. 12 shows dataset multivariate plots for component 1, which has 4,595 points. ω is calculated for every dimension of i of the data (ω_i={circumflex over (λ)}_i+n√{circumflex over (λ)}_i). In FIGS. 10-12, the y-axis is displayed in log-scale.

We focus on the AcctCnt field for now and use [Adams and MacKay, 2007] to determine change points in our data. Using the Python implementation of [Adams and MacKay, 2007] on AcctCnt, the posterior probability (P(r_t|x_1:t)) of the current run r_t's length at each step (x_i for i=1 . . . t) can be displayed, using a logarithmic color scale (see FIG. 4B). There are several differences between [Adams and MacKay, 2007] and its implementation. [Adams and MacKay, 2007] suggests exponential-family distributions for the runs, and [Kulick, 2016] specifies this by using the Student's t-distribution as the predictive distribution associated with runs. In addition, [Kulick, 2016] uses a heuristic to generate the visual in FIG. 4C, which evaluates the probability of a changepoint occurring by looking N_ω samples prior. We use N_ω=15 in FIG. 4.

FIG. 4A shows surviving records for AcctCnt. FIG. 4B shows the posterior probability (P(r_t|x_1:t)) of the current run r_t's length at each step (x_i for i=1 . . . t) using a logarithmic grayscale. FIG. 4C shows the probability of a changepoint occurring by looking N_ω=15 samples prior. Note that because of this prior sample's requirement, the final visual is cut off early.

Two runs appear to dominate in the second visual of FIG. 4, one from n=0 to n=27 and another from n=27 to n 32 45. We can then analyze the summary statistics of each of those runs and see how they change over time. The mean and variance of the run from n=27 to n=45 do not change; this suggests that the second run is a potential shelf. All records surviving past n=45 would be considered outliers.

For AcctCnt, this gives rise to 1 outlier record; this particular phone number has called over 120 times in the span of 1 month! The surviving data past the first shelf exhibits extreme values that are in the long tail of the assumed Poisson distribution, and, thus, are candidates for risk of fraudulent or other aberrant usage.

Looking now at the multivariate case by considering all fields in dataset 1, we begin by fitting the data in Table 2 to a Poisson mixture model using the Python library Pomegranate [Schreiber, 2017]. As we wish to learn distribution parameters entirely from data, we use the from_samples class method which runs k-means to initialize the components. Returned clusters are used to initialize all parameters of the distributions. Thus, we are using a multivariate Gaussian for the mixing distribution, g(λ). [Steyn, 1976] also used multivariate Gaussians for the mixing distribution to provide more flexibility on the correlation structure.

To determine the number of components in the mixture model, the log-likelihood is graphed in FIG. 5, which illustrates determining the number of components for the mixture model, the log-likelihoods (averaged over 5 runs each) for varying numbers of components are considered.

At roughly 10 components, the log-likelihood begins to plateau. We choose 20 components and use posterior probabilities to predict the most likely component which generated each sample. With Pomegranate, this is efficient, taking 56.78 seconds on a laptop with an Intel Core i7-4710MQ CPU @ 2.50 GHz, 16 GB of RAM, running Ubuntu 16.04. The distribution of components is displayed in FIG. 6. FIG. 6 is a histogram of the distribution of points amongst the mixture model components.

As an additional experiment and sanity check on the mixture model, we performed principal component analysis (PCA) to reduce the number of dimensions to 3 and grayscale the points based on the component the point belongs to in FIG. 7. FIG. 7 illustrates performing PCA to reduce the number of dimensions to 3 and grayscale the points based on the component the point belongs to after determining what component every point belongs to using posterior probabilities.

We then use the parameters of the mixture model to determine co thresholds for every dimension and component. Recall that ω_ic={circumflex over (λ)}_ic+n√{circumflex over (λ)}_ic, where {circumflex over (λ)}_ic is the parameter of the Poisson distribution for dimension i and component c. Once survival graphs are created for every component and dimension, run shelfDetect( ). A point is an outlier if at least one of its dimensions i contains a shelf and has records surviving the first shelf (r[y_i]>{circumflex over (λ)}_ic+n*_ic√{circumflex over (λ)}_ic). Note that n*_ic is the final n value of the first detected shelf in the ith dimension and component c.

FIG. 8 shows survival graphs for a dataset having 330 datapoints. ω is calculated for every dimension of i of the data (ω_i={circumflex over (λ)}_i+n√{circumflex over (λ)}_i) to create these survival graphs. For example, looking at FIG. 8, in component 15, Auth4Cnt has a shelf starting from n=0 and ending at n=2 (supposing that we set a minimum shelf length of 2). All points in component 15 that have Auth4Cnt values greater than λ_{Auth4Cn{circumflex over ( )}t, comp 15}+2*(λ_{Auth4Cn{circumflex over ( )}t,? comp 15)}^1/2 are outliers. More outliers may exist if we consider other fields and their respective shelves. Depending on user-specified shelf length, any points past n=2 in Auth3Cnt might be considered an outlier.

Then all points in component 15 that have Auth3Cnt values greater than λ_{Auth3Cn{circumflex over ( )}t, comp 15}+2*(λ_{Auth3Cn{circumflex over ( )}t, comp 15})^1/2 are also outliers, and they may be different than the outliers discovered using Auth4Cnt.

FIG. 9 displays the data like in FIG. 7 except points are grayscaled depending on if they are outliers or not. Using a shelf size of 3, 13 outliers are present in the data.

We consider another dataset from an IVR system servicing a financial company containing 103,823 phone numbers over the course of a month. This dataset contains several fields not present in the dataset discussed above, and their definitions are provided in Table 3.

TABLE 3
Field    Description
AcctCnt  Number of unique accounts referenced by
         phone number x
CallCnt  Number of times the phone number x called
Auth2Cnt Number of social security numbers (SSNs)
         referenced by phone number x
Auth3Cnt Number of zip codes referenced by phone number x
Auth4Cnt Number of financial institutions referenced
         by phone number x
EPCnt1   Number of times the phone number x hit exit point
         1 which occurs when payment succeeds
EPCnt3   Number of times the phone number x hit exit point
         3 which occurs when caller exited prior to SSN entry

The differences between this dataset and the one discussed above highlight advantages of the present methodology. Although both datasets are generated by accumulating call data to an IVR from external phone numbers, they have different IVR call tree structures. A call tree determines the path callers traverse and the exit points available to them. Call trees can be very complex and vary widely in their functionalities. However, the shelf method is the same regardless of the call data collected by the IVR system and the structure of the call tree. This makes the present method useful for discovering risk in a wide variety of datasets and call tree structures. Another difference between the two datasets is that all correlations between the fields in this second dataset are positive, opening up the possibility of class 1 methods (marginal Poisson distributions) in the [Inouye et al., 2017] paper.

In the multivariate case, we once again analyze log-likelihood plots to determine the number of components (10) and use posterior probabilities to assign points to components. The survival graphs in the univariate and multivariate are provided in FIGS. 10-12.

We have demonstrated the existence of shelves and have shown that the records that survive the first shelf (phone numbers that exhibit relatively unusual activity) represent outliers compared to the expected regular distribution of calls according to the Poisson assumption.

In some cases, no shelf exists. In these cases, one of two things is true: (1) either the distribution of survivors is very “thin” (i.e., no survivors exhibit extreme counts) or (2) there is a relatively continuous slow drop in the number of survivors as we increase n.

In the first case, our method may not identify any extreme behavior and so, in turn, won't identify any risk in those calls. If the calls still contain risk, then it must be surfaced in a different way. In the second case, we may be missing risk and an addition to our algorithm to analyze the slope of the survivor counts as n changes may be provided to account for very small but continuous-looking shelves in the data.

The data may also exhibit multiple shelves. The algorithm and method presented herein picks the first shelf of a given length, but could be modified to look for the longest shelf, the last shelf of a given length, or provide a selection of such methods for an analyst. The point of picking the first matching shelf is to capture the initial significant deviation from the assumed distribution of survivors. This is a simple and consistent way of separating data that may belong to a different distribution that connotes risk, but the method is not so limited.

We used a mixture model to represent the dependencies between variables of the call data. There are other methods for generating multivariate distributions from count data that could be used in the present process. One interesting method is to use copulas to model the dependencies between the marginal parameters like in [Inouye et al., 2017]. Another possible methodology is to create a series of bivariate Poisson distributions over the call data and look for dependencies that match the marginal distributions.

Other possibilities include acquiring and applying this method to count data in a different domain. We have call data represented as examples herein, but other types of count data could be analyzed using the methodologies described herein. Also, different default distributions besides a Poisson may be used. If the data exhibits strong overdispersion, a negative binomial distribution may be more appropriate as the relationship between the mean and variance can be better controlled. In addition, other distributions instead of a Student's t-distribution may be used for determining shelves, but may be selected depending on whether the graphs, like FIG. 2, have a negative slope and shelves are zones of 0 variance. Pomegranate's from samples also assumes multivariate Gaussians for the mixing proportions, but may not be best suited because (1) the present data is discrete, and (2) λ must reside in the positive reals, but a normal distribution is defined on the reals [Inouye et al., 2017].

We have presented a method for determining potentially risky examples in count data. This method may be unsupervised and provides visual feedback on the results of the algorithm. In the examples provided herein, this method is demonstrated by application to data collected from IVR (Integrated Voice Recognition) systems in the financial industry and to call counts and other parameters related to the IVR call tree specification. In cases where shelves were found, we were able to isolate groups of outlier phone numbers that exhibited behavior indicative of fraud risk. In conjunction with other fraud and risk detection methods, we believe that this method is a powerful addition to a telephone fraud detection system.

The method disclosed herein is novel in the following ways:

• • The method is approximate, unsupervised, simple, and fast.
  • It is easy to visualize, and, thus, easy for nontechnical people to use and evaluate.
  • It relies solely on survival statistics from the dataset, so it is applicable to univariate and multivariate datasets, providing simple visualizations based on survival counts.

The present method differs from others by focusing on a simple algorithm for segregating survival data. The focus is not on calculating the mixture of distributions in the data like Smartsifter, but, instead, on providing tools for security engineers to identify and evaluate sources of potentially fraudulent activity.

Throughout this application, various publications may have been referenced. The disclosures of these publications in their entireties are hereby incorporated by reference into this application in order to more fully describe the state of the art to which this invention pertains:

[Adams and MacKay, 2007] Adams, R. P. and MacKay, D. J. (2007). Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742.

[Inouye et al., 2017] Inouye, D. I., Yang, E., Allen, G. I., and Ravikumar, P. (2017). A review of multivariate distributions for count data derived from the poisson distribution. Wiley Interdisciplinary Reviews: Computational Statistics, 9(3):e1398.

[Kulick, 2016] Kulick, J. (2016). Bayesian changepoint detection. https://github.com/hildensia/bayesian_changepoint_detection.

[Lata et al., 2015] Lata, L. N., Koushika, I. A., and Hasan, S. S. (2015). A comprehensive survey of fraud detection techniques. International Journal of Applied Information Systems, 10(2):26{132.

[Phua et al., 2010] Phua, C., Lee, V., Smith, K., and Gayler, R. (2010). A comprehensive survey of data mining-based fraud detection research. arXivpreprint arXiv: 1009.6119.

[Rodriguez, 2007] Rodriguez, G. (2007). Poisson models for count data. acesso, 10(1):2014.

[Schreiber, 2017] Schreiber, J. (2017). Pomegranate: fast and exible probabilistic modeling in python.

The Journal of Machine Learning Research, 18(1):5992{15997.

[Steyn, 1976] Steyn, H. (1976). On the multivariate poisson normal distribution. Journal of the American Statistical Association, 71(353):233{1236.

[Yamanishi et al., 2004] Yamanishi, K., Takeuchi, J.-I., Williams, G., and Milne, P. (2004). On-line unsupervised outlier detection using_nite mixtures with discounting learning algorithms. Data Mining and Knowledge Discovery, 8(3):275{300}.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example only, and not limitation. It will be apparent to persons skilled in the relevant art that various changes in form and detail can be made therein without departing from the spirit and scope of the present invention. Thus, the breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.

Claims

1. A system for performing risk assessment, comprising: a memory comprising executable instructions;a processor configured to execute the executable instructions and cause the system to: access a probabilistic model in a library of known probabilistic models; modeling said count data using a Poisson distribution using the probabilistic model running on a processor;calculate a threshold value, co, using the probabilistic model;partition the count data based on the threshold co, using the probabilistic model;determine a number of surviving records from among the partitionedcount data based on the threshold co, using the probabilistic model;identify at least one shelf in the surviving records; anddefine outliers for risk assessment according to the at least one shelf.

2. The system of claim 1, wherein the executable instructions, when executed, cause the system to identify the at least one shelf by identifying an abrupt variation in generative parameters of a sequence of data.

3. The system of claim 1, wherein the executable instructions, when executed, further cause the system to create a visual representation of the surviving records for identifying the at least oneshelf.

4. The system of claim 3, wherein the at least one shelf isuser-identified is based on the visual representation.

5. The system of claim 1, wherein the at least one shelf is identified using concept drift detection.

6. The system of 4, wherein the executable instructions, when executed, cause the system to identifying an abrupt variation in generative parameters of a sequence of data.

7. The system of claim 1, wherein the thresholdvalue co is calculated according to: ω=λ+n×stdev(X)=+n√λwhere n is the number of standard deviation units in ω, λ is a mean of a Poisson distribution and X is a random variable having integer values.

8. The system of claim 7, wherein the executable instructions, when executed, further cause the system to determine a number of datapoints greater than the threshold value co for a given numberof standard deviation units n.

9. The system of claim 1, wherein the library of known probabilistic models is Python Pomegranate.

10. A non-transitory computer readable medium comprising instructions that, when executed by a processor of a processing system, cause the processing system to perform a method, the method comprising: accessing a probabilistic model in a library of known probabilistic models; modeling said count data using a Poisson distribution using the probabilistic model running on a processor;calculating a threshold value, co, using the probabilistic model;partitioning the count data based on the threshold ω, using the probabilisticmodel;determining a number of surviving records from among the partitioned count data based on the threshold ω, using the probabilistic model;identifying at least one shelf in the surviving records; anddefining outliers for risk assessment according to the at least one shelf.

11. The non-transitory computer readable medium of claim 10, wherein identifying the at least one shelf comprises identifying an abrupt variation in generative parameters of a sequence of data.

12. The non-transitory computer readable medium of claim 10, the method further comprising creating a visualrepresentation of the surviving records for identifying the at least one shelf.

13. The non-transitory computer readable medium of claim 12, wherein the at least one shelf is user-identified basedon the visual representation.

14. The non-transitory computer readable medium of claim 13, wherein identifying the at least one shelf comprises identifying an abrupt variation in generative parameters of a sequence of data.

15. The non-transitory computer readable medium of claim 10, wherein the threshold value ω iscalculated according to: ω=λ+n×stdev(X)=λ+n√λwhere n is the number of standard deviation units in ω, λ is the mean of a Poisson distribution and X is a random variable having integer values.

16. The non-transitory computer readable medium of claim 15, the method further comprising determining a number of datapoints greater than the threshold value ω for a given number of standard deviationunits n.

17. The non-transitory computer readable medium of claim 10, wherein the library of known probabilistic models is Python Pomegranate.