Data processing: measuring – calibrating – or testing – Measurement system – Statistical measurement
Reexamination Certificate
2002-06-28
2004-10-26
Barlow, John (Department: 2863)
Data processing: measuring, calibrating, or testing
Measurement system
Statistical measurement
Reexamination Certificate
active
06810357
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to systems and methods for evaluating and displaying the reliability of a data mining model which predicts the state of a variable with multiple possible states. More particularly, the present invention relates to systems and methods for displaying the accuracy of one or more predictive models in a data mining context.
BACKGROUND OF THE INVENTION
Data mining is the exploration and analysis of large quantities of data, in order to discover correlations, patterns, and trends in the data. Data mining may also be used to create models that can be used to predict future data or classify existing data.
For example, a business may amass a large collection of information about its customers. This information may include purchasing information and any other information available to the business about the customer. The predictions of a model associated with customer data may be used, for example, to control customer attrition, to perform credit-risk management, to detect fraud, or to make decisions on marketing.
To create and test a data mining model, available data may be divided into two parts. One part, the training data set, may be used to create models. The rest of the data, the testing data set, may be used to test the model, and thereby determine the accuracy of the model in making predictions.
Data within data sets is grouped into cases. For example, with customer data, each case corresponds to a different customer. Data in the case describes or is otherwise associated with that customer. One type of data that may be associated with a case (for example, with a given customer) is a categorical variable. A categorical variable categorizes the case into one of several pre-defined states. For example, one such variable may correspond to the educational level of the customer. There are various values for this variable. The possible values are known as states. For instance, the states of the educational level variable may be “high school degree,” “bachelor's degree,” or “graduate degree” and may correspond to the highest degree earned by the customer.
As mentioned, available data is partitioned into two groups—a training data set and a testing data set. Often 70% of the data is used for training and 30% for testing. A model may be trained on the training data set, which includes this information. Once a model is trained, it may be run on the testing data set for evaluation. During this testing, the model will be given all of the data except the educational level data, and asked to predict a probability that the educational level variable for that customer is “bachelor's degree”.
Running the model on the testing data set, these results are compared to the actual testing data to see whether the model correctly predicted a high probability of the “bachelor's degree” state for cases that actually have “bachelor's degree” as the state of the educational level variable. One method of displaying the success of a model graphically is by means of a lift chart, also known as a cumulative gains chart. To create a lift chart, the cases from the testing data set are sorted according to the probability assigned by the model that the variable (e.g. educational level) has the state (e.g. bachelor's degree) that was tested, from highest probability to lowest probability. Once this is done, a lift chart can be created from data points (X, Y) showing for each point what number Y of the total number of true positives (those cases where the variable does have the state being tested for) are included in the X% of the testing data set cases with the highest probability for that state, as assigned by the model.
As shown in
FIG. 1
, the conventional lift chart shows that there are 1000 total true positives in the testing set. This is not necessarily the number of cases in the testing data set. Some cases may have a different state for the variable than the one for which the test is being conducted. The number of true positives in the testing data set is the highest number shown on Y axis
10
. The X axis
20
correlates with the percentage of cases with the highest probabilities. Lift line
30
depicts the success of the model. For example, it can be seen that lift line
30
includes a point with (X, Y) coordinates are approximately (20, 500). This indicates that, in the 20% of the cases selected by the model as the most probable cases having the tested-for state of the variable, approximately 500 of the cases that are truly positive for the state of the variable are included. This is equivalent to getting 50% of the actual cases with the desired state in only 20% of the cases for which the test is conducted.
A model that randomly assigns probabilities would be likely to have a chart close to the random lift line
40
. In the top 10% of cases, such a model would find 10% of the true positives. Note that the X axis may also be expressed in the number of high probability cases, and the Y axis in percentages. A perfect model may also be considered. In a situation where there are N% true positives among the entire testing data set, the lift line would stretch straight from the origin to the point (N, Y
MAX
) (where Y
MAX
is the maximum Y value). This is because all of the true positives would be identified before any false positives are identified. The lift line for the perfect model would then continue horizontally from that point to the right. For example, if 20% of the cases had the tested for state, as shown in
FIG. 2
, a perfect model would have the perfect lift line
50
, extending from (0,0) to (20, 1000) and then from (20, 1000) to (100, 1000). Similarly, the worst case model would identify no true positives until the last N% of the testing population is included, and, as shown in
FIG. 3
for the case where there are 20% true positives, the worst case lift line
60
for such a model would extend from (0,0) to (80, 0) and then straight from (80,0) to (100, 1000).
As described above, in the prior art, a lift chart can be used to display and measure the prediction accuracy of a model for a given state of a variable. However, existing lift charts have several drawbacks. Because models often must pick, for a given case and variable, what state is most probable for that variable, it is useful to know with what accuracy they can perform this prediction for all states. Lift charts in the prior art do not allow the display of the general prediction accuracy of a model for a variable over all states. A model may perform well for one state but may perform badly when predicting another. Also, in the prior art, comparing lift charts for different models to evaluate the effectiveness over the entire population will not be useful, as all conventional lift chart lines have a Y value corresponding to all of the true positives at the X value corresponding to the entire population of the testing set. This is because when all cases are considered, all true positives are included.
Thus, there is a need for improved charts with which to understand the behavior of models for predicting the state of multiple-state variables.
SUMMARY OF THE INVENTION
In view of the foregoing, the present invention provides systems and methods for creating a chart displaying the accuracy of a model in predicting the state of a multi-state variable. The most likely state of the variable for each input case is predicted, as well as its associated probability. The input cases are sorted based on the predicted probabilities, regardless of which state of the variable was predicted. A chart is created which graphs the percentage of correct predictions for the most probable cases.
Other features and embodiments of the present invention are described below.
REFERENCES:
patent: 6473084 (2002-10-01), Phillips et al.
patent: 6535817 (2003-03-01), Krishnamurti
patent: 2002/0042786 (2002-04-01), Scarborough et al.
Jackson, J., “Data Mining: A Conceptual Overview”,Communications of the Association for Information Systems, 2002, 8, 267-296 (Pertinent pp. 287-289).
Lo, V.S.Y., “T
Kim Pyungchul
Tang Zhaohui
Barlow John
Lau Tung
Microsoft Corporation
LandOfFree
Systems and methods for mining model accuracy display for... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Systems and methods for mining model accuracy display for..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Systems and methods for mining model accuracy display for... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3304378