Method, system, and computer program product for outlier detection

[0001] 1. Field of the Invention

[0002] This invention relates to the field of data mining. More specifically, the present invention relates to the detection of outliers within a large body of multi-dimensional data.

[0003] 2. Description of the Related Art

[0004] Organizations collect huge volumes of data from their daily operations. This wealth of data is often under-utilized. Data mining is a known technology used to discover patterns and relationships in data. It involves the process of analyzing large amounts of data and applying advanced statistical analysis and modeling techniques to the data to find useful patterns and relationships. These patterns and relationships are used to discover key facts that can drive decision making. This helps companies reap rewards from their data warehouse investments, by transforming data into actionable knowledge and by revealing relationships, trends, and answers to specific questions that cannot be easily answered using traditional query and reporting tools.

[0005] Data mining, also known generically as “knowledge discovery,” is a relatively young, interdisciplinary field that cross-fertilizes ideas from several research areas, including machine learning, statistics, databases, and data visualization. With its origins in academia about ten years ago, the field has recently captured the imagination of the business world and is making important strides by creating knowledge discovery applications in many business areas, driven by the rapid growth of on-line data volumes.

[0006] Until recently, it was highly impracticable to build large, detailed databases that could chronicle thousands, and from a statistical view, preferably millions, of data points related to, for example, consumer transactions or insurance claims. Deriving useful information from these databases (i.e., mining the databases) was even more impractical. With the advent of modern technology, however, building and processing large databases of information has become possible. For example, most organizations now enter all of their important business information into a computer system, and information that is not already entered into the system can be scanned into the system quickly and easily. In consumer transactions, a bar code reader can almost instantaneously read so-called basket data, i.e., when a particular item from a particular lot was purchased by a consumer, the items the consumer purchased, and so on, for automatic electronic storage of the basket data. Further, when the purchase is made with, for example, a credit card, the identity of the purchaser can be almost instantaneously known, recorded, and stored along with the basket data.

[0007] Likewise, processing power is now available at relatively inexpensive cost, making the mining of databases for useful information simple and efficient. Such database mining becomes increasingly problematic as the size of databases expand into the gigabyte and even the terabyte range, and much work in the data mining field is now directed to the task of finding patterns of measurable levels of consistency or predictability in the accumulated data.

[0008] Fayyad et al. (“From Data Mining to Knowledge Discovery: An Overview,” in Chapter 1, Advances in Knowledge Discovery and Data Mining, American Association for Artificial Intelligence (1996)) presents a good, though somewhat dated, overview of the field. Bigus (Data Mining with Neural Networks, McGraw-Hill (1996)) and Berry and Linoff (Data Mining Techniques: For Marketing, Sales, and Customer Support, John Wiley & Sons (1997)), among others, have written introductory books on data mining that include good descriptions of several business applications.

[0009] Predictive modeling is a well-known technique used for mapping relationships among actual data elements and then, based on the model derived from the mapping process, predicting the likelihood of certain behavior with respect to hypothetical future actions or occurrences. There are numerous techniques and method for conducting predictive modeling and the specific details of these know methods and techniques are not discussed further herein.

[0010] One of the first steps in the process of building predictive models is to review the characteristics (attributes) of data that is to be used. Typically, measures such as the mean, range, and the distribution of the data are considered for each attribute. A judgment is often made about what data values are considered to be “outliers”. Outliers are data points that fall outside of the norm.

[0011] Generally, there are two kinds of outliers. A value that is identifiably incorrect, e.g., the age of a person being 180 years, is an identifiably incorrect outlier. For modeling purposes, identifiably incorrect outliers are essentially removed from consideration altogether (e.g., they are considered as having a value of “null”), since they represent errant data.

[0012] A second type of outlier is a value that is considered extreme, but correct. This type of outlier represents correct data, but data that is so unusual that it inappropriately skews the analysis of the more normal data. For modeling purposes, it may be desired to represent this extreme, correct outlier by some value that is less extreme, to minimize the skewing impact of the outlier on the model. As an example, when analyzing annual incomes of Americans, Bill Gates'income, while accurate, would be considered an outlier relative to the majority of the population. For outliers of this type, a limit may be placed upon the data value (as opposed to removing it from the data set altogether) to “filter out” such correct but statistically misleading data points.

[0013] Identifying these types of outliers is relatively simple when considering a single variable at a time (e.g., the age of a person or the annual income of a person) and standard box plots can be used to help identify these outliers. However, it may be desirable to identify an analyze combinations of variables that are outliers, e.g., the age of an insurance customer (a first variable) and the amount of insurance coverage that that person carries (a second variable). There are many reasons why it might be desirable to identify and analyze such multi-variable outliers; they may, for example, represent unusual occurrences such as the existence of fraud.

[0014] The original analysis of one variable can be expanded to two variables by plotting their occurrence on a scatter plot. Likewise, the combination of three variables can be visualized in a three-dimensional plot. Beyond three variables, however, a problem exists because there is no dimension beyond the third dimension to use for plotting. In the complex environments that analysts currently strive to manage, there may be many hundreds of potential attributes to use in modeling and thus the “maximum of three variables” approach is inadequate.

[0015] Predictive models can be used to identify unusual combinations of inputs exceeding three. A weakness of the traditional modeling approach, however, is that it isn't focused on the goal of the model. As an example, a company may routinely compile a broad range of data related to individuals, e.g., demographic data, interests, hobbies, profession, likes and dislikes, etc., but may wish to create a model focused on characteristics relevant to a narrow subject area, e.g., work-related characteristics. Traditional predictive modeling approaches will use either all the data, including data unrelated to the narrow subject area, or someone must cull through the data and identify which characteristics to include. Thus, all characteristics for which data has been stored will be input to the model unless steps are taken to hand-select only the characteristics of interest. Time and effort considering attributes that will not be used in the model is time wasted and thus leads to inefficiency.

[0016] In accordance with the present invention, a random sampling of a subset of a data population is taken and the sampled data is used to build a predictive model using a cubic or multiquadric radial basis function, and then “scores” (i.e., predictions) are generated for each data point in the entire data population. This process is repeated on additional random sample subsets of the same data population. After a predetermined number of random sample subsets have been modeled and scores for all data points in the population are generated for each of the models, the average score and variation for each predicted data point is calculated. The data points are subjected to rank ordering by their variance, thereby allowing those data points having a high variance to be identified as outliers. By identifying them as outliers, these data points are “flagged” for investigation, thereby simplifying the ability to determine if the data points and the variables with which they are associated should be included in the model or removed from consideration for a particular model and/or making it easier to identify data points that identify a particular activity (e.g., fraud) but which might otherwise go unnoticed in a large body of multi-dimensional data.

[0020] The present invention involves predictive modeling. Along with a predicted value, it is highly desirous to identify the confidence that can be placed in that prediction. One method of identifying this confidence is called “bootstrapping” and involves repeatedly taking smaller samples from the complete data population and then using the sample to estimate the population's statistics overall. This allows a mean and a variance for the prediction to be estimated.

[0021] FIG. 1 illustrates an example of the steps of the present invention. In accordance with the present invention, bootstrapping is utilized to build multiple models from subsets of the population, and then the models are applied to all of the data points in the population. At step 102, a large population of data points related to a large number (e.g., >3) of variables is identified and then, at step 104, a random sample of a subset of the data points is selected. For each subset so selected, a predictive model is built using a cubic or multi-quadric radial basis function (step 106). All of the data points within the entire population are then plugged into the model to generate scores (predicted values) for each data point in the complete data population (step 108). At step 110, a determination is made as to whether or not additional subsets of the entire population are to be selected and modeled. If yes, the process proceeds back to step 104. This process continues until a predetermined number of subsets has been modeled. This process is repeated with more accurate results being achieved as the number of repetitions increases.

[0022] Once the desired number of repetitions have been completed, the average score (step 112) and variation (step 114) for each data point is calculated. The data points are then ranked according to their variance (step 116). By rank ordering the data points by their variance, those points that have high variances (e.g., exceed a predetermined “outlier threshold”) are identified as candidates for investigation as outliers (steps 118 and 120). Using this technique, a measure of confidence in the predictive value for each data point is also obtained.

[0023] The above-described method is able to highlight outliers because, in a sparsely populated region of the multi-dimensional (i.e., multi-variable) space, an instance where the outcome is unusual, as compared with the other instances in the region, causes the model fit to differ significantly, depending on whether the point is included in the sample or not. In a data set having a large number (e.g., >3) of variables, the impact of these outliers is difficult to “see”, and thus they cannot be easily investigated. By breaking down the modeling into subsets of the data, applying the models to the entire data set, and then ranking the data points according to their variance, the outliers are more visible, and thus they can be located and investigated to determine whether they are real, incorrect, irrelevant to the analysis, or indicative of something undesirable happening, e.g., fraud.

[0024] The method and usefulness of the present invention is demonstrated below using a small data set (six data points) in two dimensions, i.e., one independent variable (age) and one dependent variable (amount of coverage). It is stressed that this example is used for the purpose of explanation only, and that the concept explained in connection with this example is applicable for analyzing data sets containing any number of points in multiple dimensions, especially for dimensions exceeding three. It is noted that from a purely practical standpoint, one would never use this technique in a situation involving 3 or less variables; however, illustrating the techniques for variables exceeding 3 would be voluminous and is unnecessary for understanding the present invention. Further, the specific approach discussed in the example below utilizes radial basis function models. However, it is understood that any models with a non-linear, global-fitting characteristic can also be used.

[0025] The data points comprising the complete data set (i.e., the population) for the following example are set forth in Table 1 below. The data points in Table 1 represent the age of life insurance customers A through F and the coverage amount that these customers purchased.

[0026]

TABLE 1
(Age of insured/amount of coverage)
Customer Name	Age	Coverage
A	23	50 K
B	29	75 K
C	33	150 K
D	37	250 K
E	42	350 K
F	63	100 K

[0027] FIG. 2 is a curve of all of the data points of Table 1 applied to a standard quadratic function. By fitting a quadratic function to the data points, the equation representing the model illustrated by the FIG. 2 curve: −0.58x²+52.7x−907, is obtained. The curve of FIG. 2 provides a model to use to estimate the coverage that a particular individual at a particular age might purchase. Thus, from the data in Table 1 and the equation represented by the curve of FIG. 2, it can be estimated that a 58-year-old person would likely purchase approximately $200,000 worth of life insurance.

[0028] In addition, analysis of the curve and conjecture based on that analysis might allow a conclusion to be reached that the reason coverage purchased by an individual increases with age up to approximately their mid-forties, and then begins to drop, is because they may begin amassing significant savings at about the mid-forty mark and that life insurance becomes less necessary as the person goes beyond this point. Using conjecture, it might also be concluded that the expense of coverage for someone in their mid-forties and up increases so significantly that they reduce the amount of coverage as the price increases. These conclusions are examples of the types of uses an analyst might make of the model.

[0029] Table 2 and FIGS. 3 through 8 illustrate the method of the present invention in a much simplified form. In Table 2 and each of FIGS. 3–8, five of the six customers from Table 1 are “selected” to create the above-described subsets. Instead of using random samples, for the purpose of this example, the data points for a different one of the six customers is excluded from each subset. This is done because the small size of the complete set (six customers) prohibits a true and effective random sampling that would demonstrate the present invention; in a real situation, thousands or millions of data points pertaining to hundreds or thousands of characteristics might exist and thus a random sampling would function properly.

[0030]

TABLE 2
Prediction from model
built leaving out:	Customer A	Customer B	Customer C	Customer D	Customer E	Customer F
Customer A (Sample A)	−176.64	56.4899	174.3908	262.2766	329.9252	101.9176
Customer B (Sample B)	24.78218	153.0803	215.9017	260.5546	290.8212	107.9404
Customer C (Sample C)	5.808077	146.4996	215.7103	265.2543	299.5277	107.9104
Customer D (Sample D)	−1.64803	133.245	200.0304	248.3013	282.6037	110.7689
Customer E (Sample E)	22.70059	121.9395	171.0011	206.3847	231.3793	102.9742
Customer F (Sample F)	43.24799	92.0402	150.6626	230.1604	358.8888	1255.735
Average of Predictions	−13.6248	117.2157	187.9495	245.4886	298.8576	297.8744
Variance	6624.811	1348.899	711.8673	534.2966	1890.297	220210.2
Maximum - minimum	219.8876	96.59037	65.23907	58.8696	127.5095	1153.817

[0031] In Table 2, the quadratic function of each of the subsets is used to obtain predictions for each customer in the complete set, and the average, variance, and range of the predictions is also generated. By rank ordering the data points by their variance, those points that have high variances are identified as candidates for investigation as outliers. Referring to FIG. 3, the quadratic function of samples B, C, D, E, and F are plotted. Thus, as shown, the data points for Customer A, the insured who is age 23, having 50K of coverage, has been excluded. In FIG. 4, data points for Customer B have been excluded; in FIG. 5, data points for Customer C have been excluded; in FIG. 6, data points for customer D have been excluded; and in FIG. 7, data points for Customer E have been excluded. An analysis of each of these curves, while indicating slightly different estimated values for potential insured persons along the curve, do not differ significantly from the curve shown in FIG. 2 (which includes all six data points).

[0032] Referring now to FIG. 8, the data points for Customer F have been excluded, and is the curve for the remaining customer data points (A through E) differs significantly from the curve of FIG. 1 and the curves of FIGS. 3–7. One looking at the FIG. 8 curve would presume that as a person ages, the amount of insurance coverage that they obtain continues to increase.

[0033] From this simple example, it can be seen that by repeatedly sampling subsets from the data population and building new models, and then applying the models to the entire data population (thereby creating a probability distribution for the entire data population), an average prediction and the corresponding variance of that prediction can be derived. By using a non-linear model with enough degrees of freedom (e.g., quadratic, cubic, etc.) and a global fitting process (e.g., using all data points when applying each model), those points that are sensitive to inclusion/exclusion from the model can be readily identified. These sensitive data points are indicative of outliers. In the above example, the 63-year-old's insurance purchase is identified as an outlier and thus is identified as “unusual” and worthy of investigation. When sorting through thousands or millions of data records, identifying such outliers is extremely valuable.

[0034] A beneficial aspect of the present invention is self-evident by looking at the many dimensions (variables) that may be analyzed in the insurance industry. An insurance analyst, for example, might want to know the customer's:

[0035] Needs

[0036] Risk aversion

[0037] Cost sensitivity

[0038] Financial situation—ability to pay

[0039] Financial self confidence

[0040] Convenience versus cost trade off

[0041] Technology affinity—do they prefer to talk to a person or happy to go online?

[0042] Loyalty

[0043] Value

[0044] The following is a list of attributes that are used in the insurance industry to identify the above information to the analyst. It is not exhaustive and it is not intended to imply that attributes not on this list are not useful.

Demographic/Life Stage Variables

• [0045] # adults in the household
• [0046] # children in household
• [0047] Presence of college age (18–22) member in household
• [0048] # members of household
• [0049] Household income
• [0050] Highest education level in the household
• [0051] Presence of high school child
• [0052] Head of household gender
• [0053] Marital status
• [0054] Presence of middle school child
• [0055] Oldest principal driver
• [0056] Presence of pre school child
• [0057] Presence of young adult in household
• [0058] Credit rating
• [0059] Type of home
• [0060] Market value of home
• [0061] Home ownership status
• [0062] Metropolitan/rural location
• [0063] Length of residence
• [0064] Live in grand parents
• [0065] # of autos
• [0066] Zip code
• [0067] State
• [0068] Census tract/block group/block
  Product Variables
• [0069] # policies by product grouping
• [0070] Total premium by product grouping
• [0071] Average premium by product grouping
• [0072] Total household premium
• [0073] Date of most recent product purchase
• [0074] Date of first purchase
• [0075] Product grouping of first purchase
• [0076] # claims
• [0077] Date of most recent claim
• [0078] $ amount of claims in last X years
• [0079] # terminated policies
• [0080] # terminated policies by product grouping
• [0081] Date of most recent termination
• [0082] Competitiveness rating by product by state
• [0083] Savings due to multi line discount
  Coverage Choice Variables
• [0084] # of coverages by coverage type
• [0085] Limits chosen relative to default limit
• [0086] Deductibles chosen relative to default
  Payment Choices
• [0087] Payment option
• [0088] Payment frequency
  Relationship Variables
• [0089] Most recent mailing
• [0090] Most recent service contact
• [0091] Preferred service channel
• [0092] Date of last review with customer of their insurance needs
• [0093] Tenure with Company
• [0094] Most recent contact
• [0095] Household tenure with agent
• [0096] Solicitation history
  Life Events
• [0097] Birth of child
• [0098] Change of address
• [0099] Home purchase
• [0100] Marriage
• [0101] Age related e.g. go to college, retire

[0102] As can be seen, there are numerous variables that go into insurance analysis, and identifying outliers and their significance has heretofore been a monumental task. The present invention renders the analysis much more useful, efficient, and productive. It is understood that the present invention is not limited to insurance analysis; instead, the present invention can be used in any situation involving analysis of a body of multi-dimensional data.

[0103] The above-described steps can be implemented using standard well-known programming techniques. The novelty of the above-described embodiment lies not in the specific programming techniques but in the use of the steps described to achieve the described results. Software programming code which embodies the present invention is typically stored in permanent storage of some type, such as in the permanent storage of a workstation operated by a data analyst. In a client/server environment, such software programming code may be stored with storage associated with a server. The software no programming code may be embodied on any of a variety of known media for use with a data processing system, such as a diskette, or hard drive, or CD-ROM. The code may be distributed on such media, or may be distributed to users from the memory or storage of one computer system over a network of some type to other computer systems for use by users of such other systems. The techniques and methods for embodying software program code on physical media and/or distributing software code via networks are well known and will not be further discussed herein.

[0104] Although the present invention has been described with respect to a specific preferred embodiment thereof, various changes and modifications may be suggested to one skilled in the art and it is intended that the present invention encompass such changes and modifications as fall within the scope of the appended claims.