KR20200077315A - Method and apparatus for interpreting individual machine learning predictions - Google Patents

Abstract

According to an embodiment of the present invention, an apparatus for interpreting a machine learning model includes: a model training unit to generate a prediction model for providing a probability in which a result to be predicted occurs, by applying a predetermined machine learning algorithm to a training data set including a plurality of data representing predetermined transaction, and to calculate an influence degree on the prediction model for each parameter constituting data; a probability predicting unit to calculate prediction probability in which a result occurs, by applying a prediction model to the training data set; a relative frequency coefficient calculating unit to select an upper group having a top first ratio and a lower group having a bottom second ratio from the training data set, based on the prediction probability, and to calculate a relative frequency coefficient, which is a ratio of a proportion of data belonging to each category while belonging to the upper group to a proportion of data belonging to each category while belonging to the lower group, with respect to each category of each parameter; a correlation calculating unit to calculate, with respect to the plurality of data in the training data set, the sum of relative frequency coefficients of parameters, to calculate correlation coefficients of the sum of the relative frequency coefficients and the prediction probability, to calculate the sum of products of the influence degrees of the parameters to the relative frequency coefficients of the parameters, and to calculate the correlation coefficients of the sum and the prediction probability; and a control unit to allow the relative frequency coefficient calculating unit to calculate the relative frequency coefficient until the correlation coefficient converts to a minimum value, and to allow the correlation coefficient calculating unit to calculate the correlation coefficient.

Description

METHOD AND APPARATUS FOR INTERPRETING INDIVIDUAL MACHINE LEARNING PREDICTIONS}

The present invention relates to a machine learning model analysis apparatus and method for interpreting a prediction model generated as a result of machine learning and providing the analysis information to a user.

2. Description of the Related Art Recently, with the development of electronic technology, electronic devices have been used to solve problems of high difficulty. For example, the electronic device performs high-level problem solving using machine learning. Machine learning may refer to a technical field in which information processing capability is improved by itself based on data and data processing experience. The electronic device may generate a solution, that is, a prediction model for deriving a result to be predicted by acquiring information that has not been input through a self-learning process through machine learning.

When data is input to the prediction model, information related to a result to be predicted for the input data may be derived. At this time, in relation to the derived information, the user of the prediction model may have a question of how the corresponding result is derived, and accordingly, objectivity and results of the prediction model created using machine learning It may be required to provide a reliable analysis.

Korean Registered Patent No. 10-1388654 (registered on April 17, 2014)

The problem to be solved by the present invention is to generate a predictive model predicting the probability that a desired result is predicted using machine learning, and to provide an analysis in which objectivity and reliability are secured for judgment results based on the predictive model. It is to provide a method and apparatus for interpreting a learning model.

However, the problems to be solved by the present invention are not limited to those mentioned above, but are not mentioned, but include the purpose that can be clearly understood by those skilled in the art from the following description can do.

According to an embodiment of the present invention, a method for interpreting a machine learning model may apply a predetermined machine learning algorithm to a training data set composed of a plurality of data, each representing a predetermined transaction, and a probability that a result to be predicted may occur A model learning step of generating a prediction model that provides a probability prediction step of calculating a prediction probability that the result will occur by applying the prediction model to the training data set, based on the prediction probability, the training In the data set, the upper first ratio ratio upper group and the lower second ratio lower group are selected, and for each category of each variable constituting the data, the ratio of data belonging to each category among the data belonging to the upper group is selected. Calculating a relative frequency coefficient, which is a ratio of a ratio of data belonging to each category among data belonging to the subgroup, and calculating a sum of the relative frequency coefficients of each variable for each of the plurality of data in the training data set And calculating a correlation coefficient of the sum of the relative frequency coefficients and the predicted probability, and the calculating of the relative frequency coefficient and the calculating of the correlation coefficient are performed repeatedly until the correlation coefficient converges to a minimum value. .

Further, the first ratio and the second ratio are the same.

In addition, the first ratio and the second ratio are different from each other.

Further, calculating the influence degree of each variable on the prediction model, and multiplying the product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data. The method further includes providing visual information.

In addition, the time information is information indicating the position of the product of the relative frequency coefficient and the influence frequency of each variable of the new data among the range of the minimum and maximum values of the product of the relative frequency coefficient and the influence degree.

Further, calculating the influence degree of each variable on the prediction model, and multiplying the product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data. And providing a scoring value.

In addition, the relative frequency coefficient when the correlation coefficient converges to a minimum value is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.

According to an embodiment of the present invention, a machine learning model analysis method provides a probability of generating a result to be predicted by applying a predetermined machine learning algorithm to a training data set composed of a plurality of data each representing a predetermined transaction. A model learning step of generating a prediction model, a probability prediction step of applying a prediction model to the training data set to calculate a prediction probability that the result will occur, and calculating a degree of influence of each variable constituting the data on the prediction model Step, based on the predicted probability, select the upper first ratio upper group and lower second ratio lower group from the training data set, and for each category of each variable, among the data belonging to the upper group Calculating a relative frequency coefficient, which is a ratio of a ratio of data belonging to each category among data belonging to the subgroup to a ratio of data belonging to each category, and each variable for each of the plurality of data in the training data set And calculating a sum of the product of the relative frequency coefficient and the influence of each variable, and calculating a correlation coefficient between the sum and the predicted probability, wherein the calculating the relative frequency coefficient and calculating the correlation coefficient include It is repeatedly performed until the correlation coefficient converges to the minimum value.

Further, the first ratio and the second ratio are the same.

In addition, the first ratio and the second ratio are different from each other.

The method further includes providing, as visual information, a product of a relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.

In addition, the time information includes the relative frequency coefficient and the influence of the new data among the range of the minimum value and the maximum value of the product of the relative frequency coefficient and the influence of each variable when the correlation coefficient converges to the minimum value. Information indicating the location of the product.

The method further includes providing a product of the relative frequency coefficient of each variable and the influence of each variable when the correlation coefficient converges to a minimum value for new data as a scoring value.

In addition, the relative frequency coefficient when the correlation coefficient converges to a minimum value is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.

According to an embodiment of the present invention, a machine learning model analysis apparatus applies a predetermined machine learning algorithm to a training data set composed of a plurality of data, each representing a predetermined transaction, and provides a probability that a result to be predicted occurs A model learning unit that generates a predictive model and calculates the degree of influence of each variable constituting the data on the predictive model, and applies a predictive model to the training data set to calculate a predictive probability that the result will occur. Second, based on the predicted probability, the upper group of the first ratio and the lower group of the lower second ratio are selected from the training data set, and for each category of each variable, each category of data belonging to the upper group A relative frequency coefficient calculation unit for calculating a relative frequency coefficient that is a ratio of a ratio of data belonging to each category among data belonging to the sub-group to a ratio of data belonging to, for each of the plurality of data in the training data set, the Calculate the sum of the relative frequency coefficients of each variable, calculate the sum of the relative frequency coefficients and the correlation coefficient of the predicted probability, or calculate the sum of the product of the relative frequency coefficient of each variable and the influence of each variable Then, the correlation coefficient calculation unit for calculating the correlation coefficient of the sum and the prediction probability, and the relative frequency coefficient calculation unit until the correlation coefficient converges to the minimum value to calculate the relative frequency coefficient and the correlation coefficient calculation unit It includes a control unit for controlling to calculate the correlation coefficient.

In addition, an output unit that provides, as visual information, a product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.

In addition, a scoring unit that provides, as a scoring value, a product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.

In addition, the relative frequency coefficient when the correlation coefficient converges to a minimum value is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.

The machine learning model analysis method and system according to an embodiment of the present invention may enable individual machine learning prediction to be interpreted by allowing individual prediction results to be calculated for data representing each transaction.

The method and apparatus for interpreting a machine learning model according to an embodiment of the present invention may assist a decision of a person using the machine learning model by providing a rational basis for a factor from which the result is derived for a prediction result of machine learning .

The method and apparatus for interpreting a machine learning model according to an embodiment of the present invention can allow a user to easily make a decision by providing visual information with an effect of each variable of data on a result of a prediction model by machine learning.

The method and apparatus for interpreting a machine learning model according to an embodiment of the present invention can be applied to a prediction model based on a machine learning algorithm known to be difficult to interpret prediction results because it can be applied without being limited to the type of machine learning algorithm.

The method and apparatus for analyzing a machine learning model according to an embodiment of the present invention may have high utilization as a module when each step for analysis of the machine learning model is implemented separately. However, the effects that can be obtained in the present invention are not limited to the above-mentioned effects, and other effects that are not mentioned will be clearly understood by those skilled in the art from the following description. Will be able to.

1 shows an example of a functional block diagram of a machine learning model analysis apparatus according to an embodiment of the present invention.
Figure 2 shows the flow of each step of the machine learning model analysis method according to an embodiment of the present invention.
Figure 3 shows the flow of each step of the machine learning model analysis method according to another embodiment of the present invention.
4 shows an example of information for calculating a correlation coefficient provided by a method and apparatus for analyzing a machine learning model according to an embodiment of the present invention.
5 is a diagram illustrating an example of information visualizing an influence degree of a predictive model of each variable of data provided by a machine learning model analysis method and apparatus according to an embodiment of the present invention.
FIG. 6 illustrates another example of information visualizing an influence degree of each variable of data provided by a machine learning model analysis method and apparatus according to an embodiment of the present invention on a predictive model.
7 illustrates another example of information visualizing an influence degree of each variable of data provided by the machine learning model analysis method and apparatus according to an embodiment of the present invention on a predictive model.
8 illustrates another example of information visualizing an influence degree of each variable of data provided by a machine learning model analysis method and apparatus according to an embodiment of the present invention on a predictive model.

Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various forms, and only the embodiments allow the disclosure of the present invention to be complete, and those skilled in the art to which the present invention pertains. It is provided to fully inform the person of the scope of the invention, and the scope of the invention is only defined by the claims.

In describing the embodiments of the present invention, detailed descriptions of known functions or configurations will be omitted except when actually necessary in describing the embodiments of the present invention. In addition, terms to be described later are terms defined in consideration of functions in an embodiment of the present invention, which may vary according to a user's or operator's intention or practice. Therefore, the definition should be made based on the contents throughout this specification.

The present invention can be applied to various changes and may include various embodiments, and specific embodiments will be illustrated in the drawings and described in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood as including all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

Terms including ordinal numbers such as first and second may be used to describe various components, but the corresponding components are not limited by these terms. These terms are only used to distinguish one component from another.

When a component is said to be'connected' or'connected' to another component, it is understood that other components may be directly connected or connected to the other component, but other components may exist in the middle. It should be.

1 shows an example of a functional configuration of a machine learning model analysis apparatus according to an embodiment of the present invention. '… Wealth','… Terms such as'is meant a unit that processes at least one function or operation, which may be implemented by hardware or software, or a combination of hardware and software.

Referring to FIG. 1, the machine learning model analysis apparatus 100 includes a model learning unit 110, a probability prediction unit 120, a relative frequency coefficient calculation unit 130, a correlation coefficient calculation unit 140, and a control unit 150 , An output unit 160 and a scoring unit 170.

The model learning unit 110 may generate a prediction model that calculates a probability of generating a result to be predicted by applying a predetermined machine learning algorithm to a training data set. Each data constituting the training data set may represent respective transactions such as insurance claims and bank transactions. For example, in order to generate a model for predicting insurance fraud or overpayment of insurance claimants, data regarding pre-secured insurance claim details in the form of [Table 1] below may be used as a training data set. . The training data set in [Table 1] is composed of data on variables such as occupation, age, region of residence, and hospitals that have been treated. In addition, variables such as a path for applying for insurance payments and an amount for applying for insurance payments may be included. However, the training data set in [Table 1] and the prediction of insurance fraud or overpayment based on the training data are merely examples used for convenience of description, and application examples of the present invention are not limited thereto.

job age area hospital One Official 30s Jeonbuk Oriental Hospital 2 dealership 40s Gwangju general Hospital 3 housewife 60 cars Busan lawmaker

The model learning unit 110 of the machine learning model analysis apparatus 100 according to an embodiment of the present invention performs machine learning on a training data set of the form shown in [Table 1] to insure insurance fraud (overage insurance overcharge) Produce a prediction model that calculates the corresponding probability. As an algorithm used for machine learning, one suitable for a result to be predicted may be used among a plurality of machine learning algorithms, for example, a deep neural network (DNN). In this regard, the type of algorithm used for machine learning is not limited, and binary or multi-class machine learning algorithms can be used.

The probability predictor 120 may apply a prediction model to a training data set to calculate a predicted probability of a result to be predicted, for example, an insurance fraud result. For example, the probability prediction unit 120 applies the prediction model generated by the model learning unit 110 to the training data set of [Table 1] including data for a plurality of insurance payment application transactions, respectively The predicted probability of insurance fraud will be calculated for the insurance payment transaction. [Table 2] below is a result of calculating the predicted probability of insurance fraud by applying the (insurance fraud) prediction model generated by the model learning unit 110 to the training data set of [Table 1].

job age area hospital Prediction probability One Official 30s Jeonbuk Oriental Hospital 0.526828 2 dealership 40s Gwangju general Hospital 0.294994 3 housewife 60 cars Busan lawmaker 0.231529

The relative frequency coefficient calculating unit 130 sorts the training data set according to the predicted probability, selects the upper first ratio upper group and the lower second ratio lower group, and for each category of each variable constituting the data. A relative frequency coefficient, which is a ratio of a ratio of data belonging to each category among data belonging to a subgroup to a ratio of data belonging to each category among data belonging to a group, may be calculated.

For example, the relative frequency coefficient calculating unit 130 sorts the training data set of [Table 1] from the highest prediction probability to the lowest order based on the prediction probability of [Table 2], and the prediction probability of the top 5%. You can select the upper group and the lower group with predicted probability of lower 3%. Here, the first ratio and the second ratio may be the same as each other, or may be different from each other in some cases.

The relative frequency coefficient calculating unit 130 calculates a relative frequency coefficient that is a ratio of a ratio of data belonging to each category among data belonging to a subgroup to a ratio of data belonging to each category among data belonging to each category for each category of each variable. Can be calculated. For example, the variables of'age' in [Table 1] can be expressed in categories such as '10s', '20s', '30s', '40s', '50s' and '60s' The variables of'region' can be expressed in categories such as'Jeonbuk','Gwangju', and'Busan'. For example, with respect to the training data set in [Table 1], the relative frequency coefficient calculating unit 130 calculates the relative frequency coefficient of the “female” category of the “sex” variable, and the “women” of the variable “sex” in the upper group. The ratio of data to a category (a%) and the ratio of data to a'female' category (b%) in a subgroup are obtained, and the ratio of b to a is expressed by various mathematical expressions. It can be obtained from the frequency coefficient. This can be done for each category of categorized variables. Details of the relative frequency coefficient calculation method will be described later.

The relative frequency coefficient has a positive (+) or negative (-) directionality. If the relative frequency coefficient of a certain category is positive, the category decreases the probability of the result to be predicted, and if the relative frequency coefficient of a predetermined category is negative The category can be interpreted as increasing the probability that the desired result will occur.

The correlation coefficient calculating unit 140 calculates the sum of the relative frequency coefficients of each variable for each of a plurality of data in the training data set, and calculates the correlation coefficient between the sum of the relative frequency coefficients of the variable and the predicted probability to determine the relative frequency coefficient. Accuracy can be verified. In another embodiment, the correlation coefficient between the sum of the product of the relative frequency coefficient of each variable and the influence degree and the predicted probability may be calculated. The correlation coefficient may include, for example, a Pearson correlation codfficient. The minimum value of the sum of the relative frequency coefficients for each variable and the Pearson correlation coefficient of the prediction probability may be -1, and closer to -1 indicates a strong negative linear relationship. The correlation coefficient of the sum of the relative frequency coefficients for each variable and the predicted probability or the sum of the product of the relative frequency coefficients and the influence of each variable and the predicted probability is close to the minimum value. It means that the impact is well explained. This will be described later with reference to FIG. 4.

Therefore, the first ratio and the second ratio when the sum of the relative frequency coefficients for each variable and the Pearson correlation coefficient of the predicted probability converge to the minimum value can be determined as a ratio for selecting the upper group and the lower group, respectively. Meanwhile, the first ratio and the second ratio may be the same value, or 50% or less, but different values.

In some cases, the correlation coefficient calculator 140 may calculate a sum of a product of a relative frequency coefficient of each variable and an influence degree of each variable for each of a plurality of data. Here, the degree of influence of the variable means the degree of influence of each variable on the prediction model calculated by the model learning unit 110. The degree of influence indicates the importance of a variable, which can be calculated for each variable. The high degree of influence of a variable means that it acts as an important factor in predicting the outcome.

Therefore, when the product of the relative frequency coefficient of each variable and the influence of each variable is used, the upper group and the lower group can be selected by considering the relative frequency coefficient and the influence degree together. When the product of the relative frequency coefficient and the influence of each variable is used, the properties of the indicator indicating the same direction as the relative frequency coefficient and the properties of the index indicating the same magnitude as the influence factor are considered together to improve the accuracy of the result to be predicted. have. The correlation coefficient calculating unit 140 may calculate a correlation coefficient of a sum of a product of the calculated relative frequency coefficient and an influence degree and a prediction probability.

The control unit 150 selects the upper group and the lower group by the relative frequency coefficient calculating unit 130 and calculates the relative frequency coefficient and the operation of calculating the correlation coefficient by the correlation coefficient calculating unit 140, the correlation coefficient By controlling it to be performed repeatedly until it converges to the minimum value, it is possible to calculate the relative frequency coefficient for each category of each categorization variable that enables more accurate interpretation of the predictive model.

Meanwhile, the relative frequency coefficient when the correlation coefficient converges to the minimum value may be used as a weight of a variable corresponding to the relative frequency coefficient in relation to the machine learning algorithm of the model learning unit 110. More specifically, for the machine learning algorithm, learning can be continuously performed. When re-learning for the predictive model is performed, the calculated relative frequency coefficient is used as an initial weight of the machine learning algorithm, so It can improve performance and shorten learning time. For example, when the machine learning algorithm is a neural network algorithm having multiple layers, re-learning may be performed by applying a relative frequency coefficient as a weight for the first hidden layer.

The output unit 160 may provide, as visual information, a product of a relative frequency coefficient and an influence degree of each variable when the correlation coefficient converges to a minimum value for newly input data. More specifically, the output unit 160 may provide, as visual information, a product of a relative frequency coefficient and an influence degree of each variable when the correlation coefficient converges to a minimum value for data for a new transaction that is not a training data set. The output unit 160 may provide various types of visual information, and examples of visual information may refer to FIGS. 5 to 7. 5 to 7 will be described later in detail.

The scoring unit 170 may provide, as a scoring value, a product of a relative frequency coefficient of each variable and an influence degree of each variable when the correlation coefficient converges to a minimum value for newly input data. More specifically, the scoring unit 170 may provide, as a scoring value, a product of a relative frequency coefficient of each variable and an influence degree of each variable when the correlation coefficient converges to a minimum value for newly input data rather than a training data set. have. The scoring unit 170 may provide a scoring value in various forms, and refer to FIG. 7 for an example of a form in which the scoring value is provided.

Hereinafter, detailed operations of the method for analyzing a machine learning model according to an embodiment of the present invention will be described with reference to FIGS. 2 and 3.

Figure 2 shows the flow of each step of the machine learning model analysis method according to an embodiment of the present invention.

According to the embodiment shown in FIG. 2, a result of trying to predict by applying a predetermined machine learning algorithm to a training data set (see [Table 1]) by the model learning unit 110 (insurance fraud or excessive insurance claim) Step S110 of generating a prediction model that provides a probability of occurrence is performed.

The probability prediction unit 120 applies a prediction model to the training data set to calculate a prediction probability that a result to be predicted will occur (S120), for example, for each insurance payment transaction, as shown in [Table 2]. The application will calculate the probability of insurance fraud or overpayment.

The training data set is sorted in descending order based on the predicted probability by the relative frequency coefficient calculating unit 130 to select the upper first ratio group and the lower second ratio group (S130), and configure the training data set. For each category of each variable to be calculated, a relative frequency coefficient, which is a ratio of a ratio of data belonging to each category among data belonging to a lower group to a ratio of data belonging to each category, is calculated (S140).

The relative frequency coefficient may be calculated based on substituting the ratio of the ratio (a%) of data belonging to the corresponding category in the upper group to the ratio (b%) of data belonging to the corresponding category in the lower group in various equations. The formula that calculates the relative frequency coefficient can be 0 if a=b, positive if a<b, and a value of the same size and different sign when a and b are inversely related to each other. have. According to the formula that satisfies these conditions, the higher the relative frequency coefficient of a specific category ('Occupation') of a variable ('Occupation'), the higher the result of trying to predict a transaction belonging to that specific category ('Public Employee') (insurance fraud) ), and the lower the relative frequency coefficient of a specific category ('Housewife') of a variable ('Occupation'), the more the result of attempting to predict a transaction belonging to that specific category ('Housewife') (insurance fraud) There is a high probability of corresponding to.

인 경우의 상대빈도계수를 나타낸다. Equation f(x) for calculating the relative frequency coefficient may be one of Equations 1 to 4 below, but is not limited thereto, and any equation that satisfies the above condition may be used. Where f(x) is x

Relative frequency coefficient in case of.

In Equation 1, k may be any integer.

Subsequently, the correlation coefficient calculating unit 140 calculates the sum of the relative frequency coefficients of each variable for each data of the training data set (S150), and calculates the correlation coefficient of the sum of the relative frequency coefficients for each data and the predicted probability. Calculate (S160). At this time, steps S130 to S160 are repeatedly performed until the correlation coefficient converges to the minimum value (Yes in S170) (No in S170). The correlation coefficient calculated by the correlation coefficient calculation unit 140 may be a Pearson correlation coefficient. When the Pearson correlation coefficient converges to a minimum value, the larger the sum of the relative frequency coefficients of each variable, the less likely the result to be predicted. It means losing. In other words, it means that the first ratio and the second ratio when the correlation coefficient converges to the minimum value are ratios representing the optimal upper and lower group sizes for calculating the relative frequency coefficient for each category of each variable.

For example, for training data sets such as [Table 1] for generating a predictive model for predicting insurance fraud, the relative frequency coefficient values for each category for each variable when the correlation coefficient converges to the minimum value. Accordingly, when the relative frequency coefficient value corresponding to the variable value of each data constituting the training data set is calculated, it can be expressed as [Table 3] below. As shown in [Table 3], the relative frequency coefficient of the'public official' category is 1.923, the relative frequency coefficient of the'sales business' category is 0.061, and the relative of the'housewife' category for the'occupation' variable of the person who applied for insurance payments. It can be seen that the frequency coefficient is -0.5865. This means that if the job of the person applying for insurance payments among the three job groups is'housewife', the probability of insurance fraud is the highest, and if it is'public official', it is the lowest probability of insurance fraud. do.

job age area hospital One 1.923 4.121 0.303 1.998 2 0.061 4.062 1.858 -0.703 3 -0.5865 -0.919 -1.013 -0.001

The sum of the relative frequency coefficients corresponding to each variable may be calculated for each data for a plurality of insurance payment application transactions of the training data set by the correlation coefficient calculator 140. For example, in the case of data 1 in [Table 3], the relative frequency coefficient for the'public official' category of the'occupation' variable is 1.923, and the relative frequency coefficient for the '30s' category of the'age' variable is 4.121, The sum of the relative frequency coefficients can be calculated by adding both the relative frequency coefficient for the'Jeonbuk' category of the'Region' variable, 0.303, and the relative frequency coefficient for the'Oriental Hospital' category of the'Hospital' variable, 1.998. [Table 4] shows the result of calculating the sum of the relative frequency coefficients for each data for multiple insurance payment transaction transactions.

job age area hospital Sum of relative frequency coefficients One 1.923 4.121 0.303 1.998 8.345 2 0.061 4.062 1.858 -0.703 5.278 3 -0.5865 -0.919 -1.013 -0.001 -2.5195

The correlation coefficient calculating unit 140 may calculate the correlation coefficient of the sum of the relative frequency coefficients and the predicted probability, and the first ratio of the upper group to make the correlation coefficient of the sum of the relative frequency coefficients and the predicted probability be the minimum value. A second ratio for the subgroup can be selected.

FIG. 4 is a graph showing the relationship between the sum of the relative frequency coefficients and the predicted probability for each data calculated as a result of performing the method according to FIG. 2 for the training data set for the insurance payment application history as shown in [Table 1]. In this case, it can be seen that a strong negative linear relationship appears. This means that the smaller the sum of the relative frequency coefficients, the higher the probability of insurance fraud.

Figure 3 shows the flow of each step of the machine learning model analysis method according to another embodiment of the present invention.

In the case of the embodiment shown in FIG. 3, generating a prediction model (S210) and calculating the prediction probability (S220) are the same as the embodiment illustrated in FIG. 2. However, in the case of the embodiment of FIG. 3, a step (S230) of calculating the influence degree of each variable on the predictive model by the model learning unit 110 is performed. The degree of influence is a value representing the influence of each variable in the predictive model, and can be calculated for each variable. For example, the effect of the'occupation' variable, the effect of the'age' variable, the effect of the'region' variable, and the effect of the'hospital' variable can be calculated for the predictive model.

The influence degree of the variable can be calculated using different calculation methods for each machine learning algorithm. However, the present invention is not limited thereto, and the method of calculating the influence of variables on some machine learning algorithms may be the same. An example of an impact calculation method may be Gini impurity of a decision tree, and the impact calculation method is easy for a person skilled in the art, and a detailed description will be omitted.

The relative frequency coefficient calculating unit 130 selects an upper group of the upper first ratio and a lower group of the lower second ratio among the training data sets based on the predicted probability (S240), and each of the variables constituting the training data set The relative frequency coefficient, which is the ratio of the ratio of the data belonging to each category among the data belonging to the lower group to the ratio of data belonging to each category among the data belonging to the upper group with respect to the category, is calculated (S250).

The correlation coefficient calculating unit 140 calculates a sum of a product of a relative frequency coefficient of each variable and an influence degree of each variable for each of a plurality of data in the training data set (S260), and calculates the calculated relative frequency coefficient and influence The correlation coefficient between the sum of the products and the predicted probability is calculated (S270). Steps S240 to S270 may be performed repeatedly until the correlation coefficient converges to the minimum value (Yes in S280) (No in S280).

5 to 8 illustrate examples of visually providing information about a product of a relative frequency coefficient and an influence degree of each variable of data provided by a machine learning model analysis method and apparatus according to an embodiment of the present invention.

In the case of FIG. 5, the area from the minimum value to the maximum value of the product of the relative frequency coefficient for each variable finally determined when the correlation coefficient converges to the minimum value and the influence degree for each variable is displayed as an empty bar, and newly acquired data The product of the relative frequency coefficient and the effect of each variable for is indicated by hatched bars. Each graph shown in FIG. 5 corresponds to different new transaction data, and the scoring value at the top of each graph represents the total sum of the product of the relative frequency coefficient and the impact factor for each new data. This scoring value may be provided as a single quantity representing the total sum as described above, or may be provided by displaying the product of the relative frequency coefficient and influence degree of each variable as a vector as each component.

As shown in FIG. 5, the product of the relative frequency coefficient and the influence degree based on 0 can be displayed by differently displaying colors, patterns, etc. in the case of a negative number and a positive number, which shows how each variable of the new data affects the prediction result. It has the effect of showing more clearly. If the product of the relative frequency coefficient and the impact is negative, it means that the value of the variable can affect the probability of insurance fraud, and if it is positive, it means that it can affect the probability of insurance fraud. . As described above, by making the value of the relative frequency coefficient clearly based on 0, the user making the decision can more easily grasp the relative frequency coefficient for each variable of the customer to be determined.

FIG. 6 shows the graph of FIG. 5 adjusted to a minimum value of the relative frequency coefficient and influence degree for each variable.

FIG. 7 shows time information such that the reference values are linearly displayed by summing the sequence of the order of the difference between the minimum and maximum values of the product of the relative frequency coefficient and the influence of each variable.

FIG. 8 shows the bar graph of FIG. 5 in the form of a radial graph. In this case, the larger the product of the relative frequency coefficient and the influence degree, the more the graph is displayed in a corresponding category. Therefore, at a glance, it is possible to identify a variable that has a large product of the relative frequency coefficient and the impact factor.

All of the visual information in FIGS. 5 to 8 provides a total sum of products of a relative frequency coefficient and an influence degree represented by each variable as a scoring value.

The machine learning model analysis method and apparatus according to an embodiment of the present invention is applied to various fields that derive a result to be predicted for a customer when information of the customer is input, thereby providing a reasonable basis for the result to the user and a machine learning model Can help users make decisions. For example, a method and apparatus for interpreting a machine learning model may be used to derive desired results in the financial field. More specifically, for example, by automatically calculating the result of whether or not the insurance fraud, it can effectively and easily provide a reasonable basis for the judgment result of the insurance fraud. Based on this, it can help users make decisions.

The method and apparatus for interpreting a machine learning model according to an embodiment of the present invention can allow a user to easily make a decision by providing visual information with an effect that each variable of data affects a result of a prediction model by machine learning.

Combinations of each block in the block diagrams and respective steps in the flowcharts attached to this specification may be performed by computer program instructions. These computer program instructions may be mounted on a processor of a general purpose computer, special purpose computer, or other programmable data processing equipment, so that instructions performed through a computer or other programmable data processing equipment processor may be used in each block or flowchart of the block diagram. In each step, means are created to perform the functions described. These computer program instructions can also be stored in computer readable or computer readable memory that can be oriented to a computer or other programmable data processing equipment to implement a function in a particular way, so that computer readable or computer readable memory The instructions stored in it are also possible to produce an article of manufacture containing instructions means for performing the functions described in each step of each block or flowchart of the block diagram. Computer program instructions can also be mounted on a computer or other programmable data processing equipment, so a series of operational steps are performed on a computer or other programmable data processing equipment to create a process that is executed by the computer to create a computer or other programmable data. It is also possible for instructions to perform processing equipment to provide steps for performing the functions described in each block of the block diagram and in each step of the flowchart.

Further, each block or each step can represent a module, segment, or portion of code that includes one or more executable instructions for executing the specified logical function(s). It should also be noted that in some alternative embodiments it is also possible that the functions mentioned in blocks or steps occur out of order. For example, two blocks or steps shown in succession may in fact be executed substantially simultaneously, or it is also possible that the blocks or steps are sometimes performed in reverse order depending on the corresponding function.

The above description is merely illustrative of the technical spirit of the present invention, and those skilled in the art to which the present invention pertains may make various modifications and variations without departing from the essential quality of the present invention. Therefore, the embodiments disclosed in the present specification are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted by the claims below, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

100: machine learning model analysis device
110: model learning department
120: probability prediction unit
130: relative frequency coefficient calculation unit
140: correlation coefficient calculation unit
150: control unit
160: output unit
170: scoring unit

Claims (18)

A model learning step of generating a prediction model that provides a probability that a result to be predicted is generated by applying a predetermined machine learning algorithm to a training data set consisting of a plurality of data, each representing a predetermined transaction;
A probability prediction step of calculating a prediction probability that the result will occur by applying the prediction model to the training data set;
Based on the predicted probability, from the training data set, an upper group of an upper first ratio and a lower group of a lower second ratio are selected, and among data belonging to the upper group for each category of each variable constituting the data Calculating a relative frequency coefficient that is a ratio of a ratio of data belonging to each category among data belonging to the sub-group to a ratio of data belonging to each category; And
Calculating a sum of the relative frequency coefficients of each variable for each of the plurality of data of the training data set, and calculating a correlation coefficient of the sum of the relative frequency coefficients and the prediction probability,
The step of calculating the relative frequency coefficient and the step of calculating the correlation coefficient are performed repeatedly until the correlation coefficient converges to a minimum value.
How to interpret machine learning models. According to claim 1,
The first ratio and the second ratio are the same
How to interpret machine learning models. According to claim 1,
The first ratio and the second ratio are different from each other
How to interpret machine learning models. According to claim 1,
Calculating an influence degree of each variable on the prediction model; And
Further comprising the step of providing, as visual information, a product of a relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.
How to interpret machine learning models. The method of claim 4,
The time information is information indicating a position of the product of the relative frequency coefficient and the influence of each variable of the new data among a range of a minimum value and a maximum value of the product of the relative frequency coefficient and the influence degree.
How to interpret machine learning models. According to claim 1,
Calculating an influence degree of each variable on the prediction model; And
The method further includes providing a product of a relative frequency coefficient of each variable when the correlation coefficient converges to a minimum value for new data, and a product of the influence degree of each variable as a scoring value.
How to interpret machine learning models. According to claim 1,
When the correlation coefficient converges to a minimum value, the relative frequency coefficient is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.
How to interpret machine learning models. A model learning step of generating a prediction model that provides a probability that a result to be predicted is generated by applying a predetermined machine learning algorithm to a training data set composed of a plurality of data each representing a predetermined transaction;
A probability prediction step of applying a prediction model to the training data set and calculating a prediction probability that the result will occur;
Calculating an influence degree of each variable constituting the data on the prediction model;
Based on the prediction probability, an upper first ratio upper group and a lower second lower group are selected from the training data set, and for each category of each variable, each category of data belonging to the upper group is assigned to each category. Calculating a relative frequency coefficient that is a ratio of a ratio of data belonging to each category among data belonging to the sub-group to a ratio of belonging data; And
Calculating a sum of a product of the relative frequency coefficient of each variable and an influence degree of each variable for each of the plurality of data in the training data set, and calculating a correlation coefficient of the sum and the predicted probability And
The step of calculating the relative frequency coefficient and the step of calculating the correlation coefficient are performed repeatedly until the correlation coefficient converges to a minimum value.
How to interpret machine learning models. The method of claim 8,
The first ratio and the second ratio are the same
How to interpret machine learning models. The method of claim 8,
The first ratio and the second ratio are different from each other
How to interpret machine learning models. The method of claim 8,
Further comprising the step of providing, as visual information, a product of a relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.
How to interpret machine learning models. The method of claim 11,
The time information is obtained by multiplying the product of the relative frequency coefficient and the influence of the new data among the range of the minimum and maximum values of the product of the relative frequency coefficient and the influence of each variable when the correlation coefficient converges to the minimum value. Information indicating the location
How to interpret machine learning models. The method of claim 11,
The method further includes providing a product of the relative frequency coefficient of each variable and the influence degree of each variable as a scoring value when the correlation coefficient converges to a minimum value for new data.
How to interpret machine learning models. The method of claim 11,
When the correlation coefficient converges to a minimum value, the relative frequency coefficient is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.
How to interpret machine learning models. By applying a predetermined machine learning algorithm to a training data set composed of a plurality of data each representing a predetermined transaction, a prediction model is provided that provides a probability that a desired result will occur, and the prediction of each variable constituting the data is generated. A model learning unit for calculating an influence degree on the model;
A probability prediction unit calculating a prediction probability that the result will occur by applying the prediction model to the training data set;
Based on the predicted probability, an upper first ratio upper group and a lower second ratio lower group are selected from the training data set, and for each category of each variable, among the data belonging to the upper group belonging to each category A relative frequency coefficient calculating unit that calculates a relative frequency coefficient that is a ratio of a ratio of data belonging to each category among data belonging to the subgroup with respect to the ratio of data;
For each of the plurality of data in the training data set, a sum of the relative frequency coefficients of each variable is calculated, a sum of the relative frequency coefficients and a correlation coefficient of the prediction probability are calculated, or the relative frequency of each variable A correlation coefficient calculation unit calculating a sum of a product of a coefficient and an influence degree of each variable, and calculating a correlation coefficient between the sum and the prediction probability; And
And a control unit for causing the relative frequency coefficient calculating unit to calculate the relative frequency coefficient until the correlation coefficient converges to a minimum value and controlling the correlation coefficient calculating unit to calculate the correlation coefficient.
Machine learning model analysis device. The method of claim 15,
Further comprising an output unit that provides, as visual information, a product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.
Machine learning model analysis device. The method of claim 15,
Further comprising a scoring unit for providing a product of a product of the relative frequency coefficient of each variable and the influence degree of each variable when the correlation coefficient converges to a minimum value for new data.
Machine learning model analysis device. The method of claim 15,
When the correlation coefficient converges to a minimum value, the relative frequency coefficient is used as a weight of a variable corresponding to the relative frequency coefficient in the model training step.
Machine learning model analysis device.

Images