JP2020140521A - Human determination prediction device, prediction program and prediction method - Google Patents

Abstract

To provide a human determination prediction device with improved prediction accuracy.SOLUTION: A prediction device has a processor and a memory to which the processor can access, the processor (a) generates, for a plurality of case data having values of objective variables corresponding to values of a plurality of explanatory variables, a classification type prediction model and a regression type prediction model based on a plurality of learning data obtained by replacing the values of the explanatory variables with an identification numbers of a plurality of regions based on a variable region having the plurality of regions, (b) applies the learning data of prediction target case to the classification type prediction model to predict level of the objective variable of the learning data of the prediction target case, and (c) applies the learning data of the prediction target case to the regression straight line corresponding to the predicted level to predict the value of the objective variable of the learning data of the prediction target case.SELECTED DRAWING: Figure 2

Description

The present invention relates to a human judgment prediction device, a prediction program, and a prediction method.

Human judgments, such as judgments by government agencies for violations of laws and regulations, judgments by referees in gymnastics and figure skating competitions, include judgments based on unclear judgment flowcharts and checklists. Hereinafter, a judgment on a violation of laws and regulations will be described as an example.

For example, with the development of information processing technology, the risks of attacks on information systems and information leaks are diversifying. In response, many companies are required to take difficult steps to update their security systems. The reason is that the government agencies responsible for complying with the personal information protection laws of each country tend to tighten the regulations on personal information, and the penalties for breach of the obligation to protect personal information imposed on each company are becoming stricter. ..

Japanese Unexamined Patent Publication No. 2006-252011 JP-A-2018-153901

However, the relationship between information security incidents and penalties and sanctions imposed by government agencies is unclear and difficult to predict using machine learning and artificial intelligence. In particular, since the number of cases for which sanctions have been imposed is small, sufficient learning data cannot be obtained, and the prediction accuracy of the analysis model by machine learning does not increase. In addition, as the scale of the impact of violations of the Personal Information Protection Act grows compared to past cases, government authorities may determine penalties and penalties based on new, undisclosed indicators. As a result, many companies are forced to invest heavily in security systems to avoid excessive penalties and sanctions from government agencies.

As mentioned above, it is difficult to predict the judgment of government authorities in reality, so many companies take risks and prioritize investment in security systems.

Therefore, an object of the first aspect of the present embodiment is to provide a prediction device, a prediction program, and a prediction method for human judgment with improved prediction accuracy.

The first aspect of this embodiment includes a processor and a memory accessible to the processor.
The processor
(A) For a plurality of case data having values of objective variables corresponding to values of a plurality of explanatory variables, the values of the explanatory variables are replaced with identification numbers of the plurality of regions based on the variable regions having a plurality of regions. A classification type prediction model and a regression type prediction model are generated based on the multiple training data.
The classification type prediction model has a plurality of classification type learning data in which the value of the objective variable of the plurality of training data is replaced with a plurality of levels according to the size, and among the plurality of classification type training data. The level of the objective variable of the training data of the prediction target case and the classification type learning data having the shortest norm distance between the values of the explanatory variables is determined to be the level of the objective variable of the training data of the prediction target case.
Among the plurality of training data, the regression type prediction model has a regression line close to the coordinate points of the explanatory variables of the plurality of training data for each level divided by the level of the objective variable for each level. The value of the objective variable of the training data of the prediction target case is calculated from the regression line corresponding to the level determined by the classification type prediction model.
Further, (b) the training data of the prediction target case is applied to the classification type prediction model to predict the level of the objective variable of the training data of the prediction target case.
(C) A prediction device that predicts the value of the objective variable of the training data of the prediction target case by applying the learning data of the prediction target case to the regression line corresponding to the predicted level.

According to the first aspect, the prediction accuracy of the prediction device can be improved.

It is a figure explaining an example of a decision act or a judgment act. It is a figure which shows the example of the prediction model corresponding to the determination act of FIG. It is a figure which shows the flowchart which shows the outline of the generation, prediction, and update of the prediction model in this embodiment. It is a figure which shows the configuration example of the prediction apparatus in this embodiment. It is a figure which shows the flowchart of the generation process of the learning data master of an existing case. It is a figure which shows an example of case data. It is a figure which shows the list of the variable area of the explanatory variable, and the example of a training data master. It is a figure which shows the flowchart of the process of a variable search program. It is a figure which shows the flowchart of the process of a variable search program. It is a figure which shows the example of the partial variable list. It is a figure which shows the flowchart of the process S23 which calculates the prediction accuracy of a classification type analysis model. It is a figure explaining the calculation of the prediction accuracy of a classification type analysis model. It is a figure which shows the prediction method of an unknown case by a classification type analysis model. It is a figure which shows the calculation method of the prediction accuracy of a classification type analysis model. It is a figure which shows the method of determining the optimum partial variable of a classification type analysis model. Calculation of prediction accuracy of regression analysis model It is a flowchart of the process of S34. It is a figure explaining the calculation of the prediction accuracy of a regression type analysis model. It is a figure which shows the generation method of the regression type analysis model. It is a figure which shows the example which sorted the partial variable of the regression type analysis model by the prediction accuracy. It is a figure which shows the method of determining the optimum variable area of a regression analysis model. It is a flowchart of the process C, D which predicts the generation of a prediction model and the case of the prediction target by the prediction model. It is a flowchart of the process F which updates the variable area of the explanatory variable. It is a figure which shows an example of the variable area list of the search target and the partial variable list of the search target. It is a figure which shows the example of the prediction accuracy of the analysis model of each subvariable SSV_UP_1 to SSV_UP_5 calculated in process S52. It is a figure which shows the variable area list and the partial variable list in the search of the 1st generation and the 2nd generation. It is a figure which shows the comparison result of the prediction accuracy and the judgment example. It is a figure which shows the relationship of 3 subvariables of each of the search of the 1st generation and the 2nd generation. It is a figure which shows the value applied to the variable Y of the 1st to 3rd variable areas in the variable area list of the 1st generation search. It is a figure which shows the value applied to the variable Y of the 1st to 3rd variable areas in the variable area list of the 2nd generation search. It is a figure which shows the value applied to the variable Y of the 1st to 3rd variable areas in the variable area list of the 2nd generation search.

[Glossary]
The following is a brief definition of terms in this embodiment.

Analytical model: A model that predicts the objective variable from multiple explanatory variables in the case. In this embodiment, a classification type analysis model and a regression type analysis model are used together. The analytical model is sometimes called a predictive model.

Case: The case has a plurality of explanatory variables and an objective variable to be predicted.

Training data master: Data in which variables are constantized and quantified for multiple existing cases, and each of the multiple cases is used as the quantified value (area identification number) of the explanatory variable and the value or level of the objective variable. Set of.

Variable constantization: To quantify the variables in the case.

Variable quantification: Converting a variable value into a region identification number (1, 2, 3, etc.) (quantitative value) based on a variable region (quantification standard) having multiple regions.

Variable area: Multiple areas that serve as a reference when converting the value of an explanatory variable into an area identification number. By setting the optimum variable range for the explanatory variables, the prediction accuracy of the analytical model can be improved. For example, by assigning different variable ranges Y1, Y2, Y3, etc. to the variable Y, different explanatory variables Y1, Y2, Y3, etc. can be defined for the same explanatory variable Y. In the following description, the codes of the variables Y having different variable ranges are referred to by the same codes as the variable ranges Y1 to. The variable Y1 means the variable Y in the variable area Y1.

Variable region search: Calculate the prediction accuracy of the analytical model for each and combination of multiple variable region candidates applied to the explanatory variable Y, and find a variable region that can obtain high prediction accuracy. The search for the variable range is performed in the process of initializing the explanatory variables.

Subvariables: A combination of explanatory variables that are a subset of explanatory variables with different variable ranges and explanatory variables with a single variable range. An analysis model is determined for each subvariable.

Prediction processing by analytical model: To generate an analytical model based on the training data that quantifies the explanatory variables of existing cases in the variable range selected in the search process, and predict the predicted target cases with the analytical model.

Update of variable range: When the prediction accuracy of the once generated analysis model deteriorates, it is considered that the judgment criteria have been changed in response to a new case, and there are multiple variable range candidates that can be estimated to have high prediction accuracy. Calculate the prediction accuracy of the analysis model for partial variables, update the variable area to the variable area where the prediction accuracy is higher than the existing analysis model, and update the analysis model.

In the present embodiment, firstly, a qualitative human judgment existing in the real world and a judgment based on a quantitative checklist or a flowchart are mixed, and it is difficult to predict the judgment result. Actions or judgments An analytical model is determined based on existing cases that have already been decided or judged.

Second, the analytical model is composed of a classification type analytical model and a regression type analytical model.

Thirdly, in the initialization step, the optimum variable range of the explanatory variables of both analytical models is searched.

Fourth, when the prediction accuracy of the existing analysis model deteriorates due to the emergence of a new case, the variable range of the explanatory variables is updated to determine the analysis model with improved prediction accuracy. Hereinafter, each of the above features will be described.

Decisions or judgments include, for example, (1) judgment of fines and sanctions by the administrative authorities when a company violates laws and regulations, (2) determination of bid amount for a huge project, and (3) interview test. Pass judgment by both written examinations, (4) determination of sentence for criminal acts, etc. The present embodiment may be applicable to a decision act or a judgment act other than the above.

[Judgment behavior and prediction model]
FIG. 1 is a diagram illustrating an example of a decision act or a determination act. The example of FIG. 1 is an example of determining a sanction for a violation of a law. For example, company 1 builds and operates a system that utilizes personal information. Problems such as leakage of personal information occur due to internal improprieties, external intrusions, or attacks on this system (S1). When the trouble is reported to the administrative authority 2 (S2), the authority 2 interviews the company or the customer and confirms the situation of the trouble (S3). Then, the authority 2 applies the law to the trouble situation and judges the violation of the law (S4). Finally, the authority 2 notifies the company 1 of the punishment and the sanction decided according to the judgment (S5), and publishes the judgment case 3 (S6). A company that conducts a risk assessment on a company generates a prediction model 4 that predicts a judgment using past cases as learning data, and predicts a judgment of an unknown case by the prediction model 4 (S7). Company 1 makes improvements to the system with reference to this predicted determination.

FIG. 2 is a diagram showing an example of a prediction model corresponding to the determination action of FIG. The existing case 3 includes a plurality of judgment cases published by the administrative authorities. Case 3 is, for example, a case of leakage of personal information. First, for each of the existing cases 3, learning data 5 including explanatory variables x, y, z and objective variable SA (Sanction) is generated. Variables included in the case are artificially made constant from the existing case, the variable range of the variable is set, and the value of each variable in the case is converted into an identification number indicating which area of the variable range corresponds to.

The learning data 5 of the example of FIG. 2 includes the information type x, the number of affected people y, and the leakage period z as explanatory variables, and includes the sanctions SA as the objective variable. By machine learning the training data 5, a prediction model 4 of a case whose objective variable is unknown is generated.

For example, the explanatory variable x is an information type, and the variable range includes "1" when personal information includes one type of address, name, and age, "2" when it contains two types, and three types. In the case, it is an identification number of three areas of "3". The explanatory variable y is the number of affected people indicating the number of leaked personal information, and the variable range is "1" for less than 100 people, "2" for less than 1000 people, and "3" for 1000 or more people. It is an identification number. The explanatory variable z is the leakage period, and the variable range is the identification number of three regions of "1" for less than 5 years, "2" for less than 10 years, and "3" for 10 years or more. The objective variable SA is a sanction.

The prediction model 4 is a model based on a predetermined machine learning algorithm 6. The machine learning algorithm 6 is, for example, a linear regression analysis or a polynomial regression analysis. FIG. 2 shows, as an example of the prediction model 4, a multiple regression analysis model in which the objective variable SA is calculated by a linear function of a plurality of variables x, y, and z.

Once the prediction model 4 is generated, the explanatory variables x, y, z of the case 7 whose objective variable is unknown are input to the prediction model 4, the function of the prediction model is calculated, and the prediction value 8 of the objective variable SA is calculated. Is output.

In the above example, a prediction model is generated from a judgment case in which multiple judgment indexes such as a qualitative judgment by a human being and a judgment by a quantitative checklist or a flowchart are mixed, and the objective variable of the case having an unknown objective variable is used. Predict. When constructing such a prediction model, there are the following problems.

First, the judgment mechanism such as checklists and flowcharts used for judgment in existing cases is not disclosed. For example, in a case, a plurality of people may make a discontinuous judgment by a plurality of judgment methods, and a sanction may be determined. In addition, although general safety standards are open to the public, explanations of combinations and details of the standards are not disclosed.

Second, although the cases where sanctions have been imposed in the past are open to the public, the number of cases is very small and the number of learning data that can be used for machine learning cannot be prepared. With machine learning using a small amount of training data, the search space is too large to generate a significant prediction model.

Third, the qualitative and quantitative criteria used to make decisions in existing cases may change over time. For example, when the greatest damage in history occurs, the quantification of past cases becomes meaningless, and the prediction accuracy of future cases may decrease.

[Outline of generation, prediction, and update of prediction model in this embodiment]
FIG. 3 is a diagram showing a flowchart showing an outline of generation, prediction, and update of the prediction model according to the present embodiment.

In the generation of the prediction model, the training data master of the existing case is generated (A), the optimum variable range of the explanatory variables of the case is searched for (B), and the subvariables that are the combination of the variables of the variable range selected by the search Generate a prediction model (C).

In the prediction of unknown cases by the prediction model, the value of the objective variable is predicted by the prediction model generated in step C for the case to be predicted (D). As long as the prediction accuracy of the unknown case does not decrease (NO in E), the prediction D by the prediction model determined in step C is repeated.

Then, when the prediction accuracy of the prediction model deteriorates (YES in E), such as when the predicted value of the unknown case is significantly different from the sanctions of the sanctioned case announced after that, the existing case and the newly published case Based on this, update the variable range of the explanatory variables to a more appropriate one (F). Then, the prediction model of the subvariable of the explanatory variable of the updated variable range is redetermined (C) and used for the prediction of the unknown case thereafter (D).

In the generation A of the training data master, variables are quantified and quantified for a plurality of existing cases, and each case has a quantified value (area identification number) of an explanatory variable and a value or level of an objective variable. Let it be data. The constantization and quantification of variables are as described in the definition book.

In the variable region search B, the prediction accuracy of the prediction model is calculated for each of the plurality of variable region candidates and combinations applied to the explanatory variable Y, and the variable region in which high prediction accuracy can be obtained is extracted. The variable region search B is performed in the explanatory variable initialization step, which is performed before the process C for determining the prediction model.

In the prediction model generation process C and the prediction process D using the prediction model, a prediction model is generated (C) based on the training data obtained by quantifying the explanatory variables of the existing case in the variable area extracted in the search process B, and the prediction target is predicted. Predict the value of the objective variable of the case with the analytical model (D).

In the variable area update F, if the prediction accuracy of the once generated analysis model deteriorates, it is considered that the judgment criteria have been changed in response to a new case, and each has a plurality of variable area candidates that can be estimated to have high prediction accuracy. The prediction accuracy of the analysis model is calculated for a plurality of partial variables, the variable area is updated to a variable area where the prediction accuracy is higher than that of the existing analysis model, and the analysis model is updated.

FIG. 4 is a diagram showing a configuration example of the prediction device according to the present embodiment. The prediction device 10 is an information processing device such as a server or a personal computer, or a computer. The prediction device 10 has a processor 11 such as a CPU, a main memory 12 such as RAM (Random Access Memory), a network interface 13, a bus 14, and a large capacity such as an HDD or SDD (Hard Disk Drive, Solid State Drive). It has auxiliary storage devices 20 and 30 of the above. The network interface 13 can be accessed from the client terminal devices 40 and 41 via the Internet or the like.

The auxiliary storage device 20 stores various programs such as the variable area search program 21, the predictive model generation program 22, the predictive model program 23, and the variable area update program 24. The auxiliary storage device 30 contains case data 31, training data master 32, variable area list and partial variable list 33, training data 34 for classification, training data (1) 35_1 for regression, and training data (2) for regression. Various data such as 35_2 are stored.

The processor 11 of the computer 10 executes the variable area search program 21 to execute the variable area search process B. Similarly, the processor 11 executes the prediction model generation program 22 to execute the prediction program generation process C. Further, the processor 11 executes the prediction model program 23 to execute the prediction process D of the unknown case by the prediction model. Then, the processor 11 executes the variable area update program 24 to execute the variable area update process F.

[Generation of training data master A]
FIG. 5 is a diagram showing a flowchart of a learning data master generation process of an existing case. Values that can be made constant are extracted as explanatory variables from existing cases (S10). In this extraction process, the processor 11 may execute the explanatory variable extraction program shown in FIG. 2, analyze the case report in words, and extract the explanatory variables. Alternatively, the explanatory variables may be artificially extracted from the case report.

FIG. 6 is a diagram showing an example of case data. The case of FIG. 6 is a personal information leakage case as described above. In FIG. 6, the information type X, the number of leaked people Y, and the leaked period Z are extracted as explanatory variables X, Y, and Z that can be made constant in Cases 1 to 5. The information type X indicates what kind of personal information was included in the leaked data. The number of leaked Y indicates the number of personal information contained in the leaked data. Leakage period Z indicates the period during which data was leaked.

Returning to FIG. 5, the processor 11 executes the variable range candidate determination program for the explanatory variables shown in FIG. 2, and based on the range of the values of the extracted explanatory variables, a predetermined number of regions in a predetermined range (grain size). ) Determines one or more candidates for the variable range divided by) (S11). Alternatively, the candidates for the variable range of the extracted explanatory variables may be artificially determined.

FIG. 7 is a diagram showing a list of variable ranges of explanatory variables and an example of a learning data master. Determining Variable Area Candidates for Explanatory Variables In S11, a plurality of variable area candidates that are search candidates for the variable area search process to be executed later are determined. It is preferable to determine as many variable range candidates as possible. After that, the processor executes the variable area search program 21 to search for the variable area from the variable area candidates so that the prediction accuracy of the prediction model is high.

In FIG. 7, the information type X, which is an explanatory variable, has a variable range having three regions of including one type of personal information, two types of information, and three types of personal information. Further, for the leakage period Z, a variable region having four regions shown in the figure is determined.

As an example, the variable areas Y1, Y2, Y3, Y4, and Y5 are determined as candidates for the variable area of the number of leaks Y. As shown in the figure, these variable areas Y1, Y2, Y3, Y4, and Y5 are determined so that the maximum area (cap area) of the variable area increases in order. According to the case database of FIG. 6, considering that the data range of the number of leaked people Y is 500 or more and 10000 or less, the variable areas Y1, Y2, and Y3 have cap areas of 100 or more and 1000, respectively. As mentioned above, it is different from more than 10,000 people. In addition, variable areas Y4 and Y5 (cap areas of 50,000 or more and 100,000 or more) that may be added to the candidates in future variable area updates have been added to the variable area candidates. These variable ranges are added to the candidates in anticipation of the magnitude of future leaks.

The number of leaks Y, which is an explanatory variable, becomes different explanatory variables Y1 to Y5 when the variable ranges Y1 to Y5 are applied. By optimally setting this variable range, the prediction accuracy of the prediction model can be improved. For example, if different sanctions are determined for the number of leaked cases of 1,000 or more and 10,000 or more, different sanctions can be predicted by selecting Y3 rather than the variable range Y2.

Returning to FIG. 5, the processor 11 allocates a plurality of variable ranges to predetermined explanatory variables when determining variable range candidates for the explanatory variables. In the example of FIG. 7, five types of variable ranges Y1 to Y5 are assigned to the explanatory variables Y (number of leaks).

Next, the processor 11 executes an explanatory variable quantification program (not shown) to quantify the explanatory variables of the case with reference to the variable range of the explanatory variables (S13). In the learning data master 32 in FIG. 7, for each of the explanatory variables X, Y1 to Y5, and Z of cases 1 to 5, the case data was quantified based on the variable area candidates X, Y1 to Y5, and Z, respectively. The data is shown. For example, since the number of leaked persons in Case 1 is 1000, all the quantified data of the variable areas Y1 to Y5 are “2”. On the other hand, since the number of leaked persons in Case 2 is 10,000, the quantified data of the variable areas Y1 to Y5 are 2, 3, 4, 4, 4.

Returning to FIG. 5, the processor 11 executes the objective variable quantification program to replace the analog value of the objective variable SA with a plurality of levels (a plurality of levels according to the magnitude of the objective variable value). , Generate a new objective variable SAL (Sanction Level) (S14). In the example of FIG. 7, the sanction SA of less than 100,000 is replaced with the level A sanction level SAL of 100,000 or more. The analysis model (or prediction model) described later is a model that combines a classification type analysis model and a regression type analysis model in consideration of the small amount of training data, and the value of the objective variable (sanction) for unknown cases. Predict. The new objective variable SAL is used in this taxonomic analytical model.

By the above processing, a learning data master having a set of learning data is generated from the contents of the case.

[Search for the optimum variable range of explanatory variables B]
Next, the processor 11 executes the variable area search program 21 to search for the optimum variable area from the variable area candidates. In this variable area search process, a variable area suitable for the above-mentioned classification type analysis model and a variable area suitable for the regression type analysis model are searched for, respectively.

8 and 9 are diagrams showing a flowchart of processing of the variable search program. FIG. 8 shows a variable region search process suitable for the classification type analysis model, and FIG. 9 shows a variable region search process suitable for the regression type analysis model. The search processing of the variable range of each of the classification type analysis model and the regression type analysis model will be described below.

[Searching the variable range of the classification type analysis model]
In FIG. 8, the processor 11 executes the variable area search program 21, sets a plurality of variable area candidates (Y1 to Y5) in the corresponding explanatory variables (Y), and sets the plurality of explanatory variables (X, Y1 to Y5, Generate a subvariable list 33 that preferably has all the subvariables that are a subset of the set of Z) (S21).

FIG. 10 is a diagram showing an example of a partial variable list. The subset variable SSV (SubSet Variable) in the list is a subset of the set of multiple explanatory variables (X, Y1 to Y5, Z). For example, the subvariable SSV7 is a subset containing all seven variables, and the subvariable SSV6-1 is a subset containing the explanatory variables X, Y1 to Y5 excluding the explanatory variable Z. The same applies to other subvariables.

Therefore, as shown in FIG. 8, the processor 11 extracts one subvariable from the subvariable list 33 (S22), extracts the training data corresponding to the extracted subvariable from the training data master, and determines the classification type analysis model. Calculate the prediction accuracy (S23). The processor 11 repeats the calculation process S23 of the prediction accuracy of the classification type analysis model for all the partial variables in the partial variable list (S24).

FIG. 11 is a diagram showing a flowchart of the process S23 for calculating the prediction accuracy of the classification type analysis model. FIG. 12 is a diagram for explaining the calculation of the prediction accuracy of the classification type analysis model. FIG. 13 is a diagram showing a method of predicting an unknown case by a classification type analysis model. FIG. 14 is a diagram showing a method of calculating the prediction accuracy of the classification type analysis model. FIG. 15 is a diagram showing a method of determining the optimum subvariables of the classification type analysis model. The process of calculating the prediction accuracy of the classification type analysis model will be described with reference to these figures.

The processor 11 calculates the prediction accuracy of the classification type analysis model by the cross-validation method shown in FIG. Hereinafter, the process of calculating the prediction accuracy of FIG. 11 will be described with reference to FIG.

As shown in FIG. 11, the processor 11 selects one case (for example, case 1) in the training data master 32 as the case to be evaluated (S231), and the rest excluding the case to be evaluated (case 1). The sanction level (A or B) of the objective variable of the case to be evaluated (Case 1) is predicted by the classification type prediction model based on the training data of the case (Case 2-5) (S232).

Further, the processor determines the success or failure (success or failure) of the prediction based on the match / mismatch between the prediction level of the case (case 1) to be evaluated and the true level of the case (case 1) (S233).

The processor repeats the processes S232 and S233 for all the remaining cases (cases 2 to 5) in the training data master (S234). As a result, the success or failure of the remaining evaluation target cases (cases 2 to 5) is determined by the agreement / mismatch between the prediction level and the true level of the case. Finally, the processor uses the success / failure rate (ratio of successful predictions) in all the cases to be evaluated as the prediction accuracy of the extracted partial variables (S235). For example, the correct answer rate is calculated by dividing the number of times the prediction level and the true value match by the number of cases.

In FIG. 12, the processor extracts and extracts the values of the variable values corresponding to the extracted partial variables (for example, any of the partial variables SSV in FIG. 10) from the training data master 32 for all cases 1 to 5. Generate training data for partial variables.

Then, the processor uses a classification type prediction algorithm to determine the sanction level of the evaluation target case (case 1) by the prediction model based on the learning data of the remaining cases 2 to 5 excluding the evaluation target case (for example, case 1). Predict.

A method of predicting an unknown case by a classification type analysis model will be described with reference to FIG. FIG. 13 shows the training data of the partial variables SSV7 (variables x, Y1 to Y5, Z). Since the partial variable SSV7 has all the explanatory variables X, Y1 to Y5, Z, its training data is the same as that of the training data master 32. Then, the sanction level of the predicted target case E1 is predicted by the following method based on the learning data.

In the lower half of FIG. 13, a coordinate space having a variable of a partial variable as a coordinate axis is shown, and examples 1 to 5 in the training data are plotted in the coordinate space. When the prediction target case E1 is plotted in such a coordinate space, the case with the smallest prediction target case E1 and the L2 norm distance L2_ND is selected from cases 1 to 5. Then, the sanction level of the selected case is set as the predicted sanction level of the predicted case E1.

The L2 norm distance is the distance in the coordinate space between the two cases. The distance L2_ND in this coordinate space is obtained by squaring the differences between the values of the same coordinate axes and adding them to calculate the square root of the added value. Therefore, the processor 11 executes the classification type prediction model program 23, calculates the norm distance between the prediction target case E1 and all the cases 1 to 5, and detects the case of the minimum distance from the cases 1 to 5.

According to the example of FIG. 13, since the case of the shortest distance to the case E1 is the case 4, the predicted sanction level of the case E1 is determined to be the same as the true predicted level A of the case 4.

Now that the prediction method of the classification type prediction model is understood, the calculation method S23 of the prediction accuracy of the classification type prediction model will be described. Among the examples shown in FIG. 12, a method of predicting the sanction level of the case 3 to be evaluated based on the remaining cases 1, 2, 4 and 5 will be described with reference to FIG.

In the lower half of FIG. 14, Cases 1 to 5 are plotted in the coordinate space as in FIG. Therefore, the processor detects the case having the shortest L2 norm distance with respect to the case 3 to be evaluated from the remaining cases 1, 2, 4, and 5. As shown in FIG. 14, in case 3 to be evaluated, case 2 is the shortest L2 norm distance L2_ND_2. Therefore, the processor determines the predicted sanction level of the case 3 to be evaluated as the sanction level B of the detected case 2 (S232). Then, the processor compares the predicted sanctions level B of the case 3 to be evaluated with the true sanctions level A of the case 3, and determines that the predicted sanctions level is incorrect because they do not match (S233).

As shown in FIG. 12, the processor predicts the sanction level by the classification type prediction model based on the remaining cases for each of all cases 1 to 5, and whether or not each prediction level matches the true sanction level. That is, it is determined whether or not the prediction is correct (success or failure). A table of result verification is shown in the lower right of FIG. 12, and according to it, cases 1, 2, 4 and 5 are correct answers, and case 3 is incorrect. As a result, the processor determines the prediction accuracy of the classification type prediction model of the partial variables extracted in S22 of FIG. 8 as the correct answer rate (0.88) in the result verification table (S235).

Returning to FIG. 8, the processor calculates the prediction accuracy of the corresponding classification type prediction model for all partial variables SSV (S24) (S23). FIG. 15 shows the results of sorting by prediction accuracy (correct answer rate) for all partial variables SSV. The processor determines the optimum variable range from the top N (for example, top 5) subvariables SSV with high prediction accuracy (S25). In the example of FIG. 15, the variable range Y3 that is most contained in the top N (top 5) subvariables SSV sorted in descending order of prediction accuracy is determined as the optimum variable range. Then, the subvariables of the classification type prediction model are determined as X, Y3, Z (S25 in FIG. 15). In FIG. 15, a subvariable SSV3_ # (a subvariable having explanatory variables X, Y3, Z) having the optimum variable range Y3 as the variable Y is determined.

Further, the processor extracts the partial variable data from the training data master based on the determined classification type subvariable, and generates the classification type training data (S26).

[Search for variable range of regression analysis model]
Next, the search for the variable range of the regression analysis model will be described. FIG. 9 is a flowchart of the process of searching the variable region of the regression analysis model. Further, FIG. 16 is a flowchart of the process of calculating the prediction accuracy of the regression analysis model S34. FIG. 17 is a diagram for explaining the calculation of the prediction accuracy of the regression analysis model. FIG. 18 is a diagram showing a method of generating a regression analysis model. FIG. 19 is a diagram showing an example in which the partial variables of the regression analysis model are sorted by the prediction accuracy. FIG. 20 is a diagram showing a method of determining the optimum variable range of the regression analysis model. The process of calculating the prediction accuracy of the regression analysis model will be described with reference to these figures.

In FIG. 9, the processor 11 executes the variable area search program 21 and selects the level of the objective variable (S31). The search for the variable range of the regression analysis model is performed for each level of the objective variable. There are A and B levels of the objective variable, and it is assumed that level A is selected here.

Next, the processor extracts the training data of the case belonging to the selected level A from the training data master (S32). The learning data of the case belonging to level A extracted here becomes the learning data master belonging to level A. As shown in FIG. 17, cases 1, 3 and 4 belonging to level A in the learning data master are extracted as evaluation targets.

Next, the processor generates a recurrent analysis model for all the partial variables in the partial variable list 33 (FIG. 10) generated in S21 of FIG. 8 based on the cases 1, 3 and 4 belonging to level A, and the generation thereof. Calculate the prediction accuracy (S33-S35). That is, the processor extracts one subvariable from the subvariable list 33 (FIG. 10) (S33), extracts the training data corresponding to the extracted subvariable from the above-mentioned level A training data master, and performs regression analysis. Generate a model and calculate its prediction accuracy (S34). The processor 11 repeats the calculation process S34 of the prediction accuracy of the regression analysis model for all the partial variables in the partial variable list (S35).

The processor 11 calculates the prediction accuracy of the regression analysis model S34 by the cross-validation method shown in FIG. Hereinafter, the process of calculating the prediction accuracy of FIG. 16 will be described with reference to FIG.

As shown in FIGS. 16 and 17, the processor 11 selects one case (for example, case 1) in the level A learning data master as the case to be evaluated (S341). Further, the processor generates a regression prediction model Pmodel_1 from the training data of the remaining cases (cases 3 and 4) excluding the case to be evaluated (case 1) (S342). Further, the processor predicts the predicted value and the sanction value of the objective variable of the case (Case 1) to be evaluated by the recurrent prediction model Pmodel_1 (S342).

Then, the processor obtains the absolute ratio of the predicted value of 40000 and the true value of 60000 in Case 1 to be evaluated (S343). The absolute ratio of both values is obtained by dividing the smaller value by the larger value. The above processes S341, S342, and S343 are repeated for all the cases to be evaluated (cases 1, 3, and 4) (S344). Finally, the processor outputs the average of the absolute ratios of all the cases to be evaluated as the prediction accuracy of the partial variables (S345). In the example of FIG. 17, the prediction accuracy (average absolute ratio) of level A is 0.71.

At level A in FIG. 17, for case 1 to be evaluated, a regression prediction model Pmodel_1 is generated from the training data of the remaining cases 3 and 4, a prediction value of 40000 is calculated, and the predicted value and the true value of case 1 are calculated. An absolute ratio of 0.67 has been calculated. Recurrent prediction models Pmodel_3 and Pmodel_4 are generated in the same manner for cases 3 and 4 to be evaluated, prediction values 100000 and 12000 are calculated, and absolute ratios 0.7 and 0.75 are calculated.

The processor also generates the prediction models Pmodel_2 and Pmodel_5, calculates the prediction value, and calculates the prediction accuracy (average absolute ratio) (S342, S343) for the remaining Level B cases 2 and 5. As a result, in the example of FIG. 17, the prediction accuracy (average absolute ratio) of level B is 0.67.

FIG. 18 shows a method of generating a regression type analytical model. FIG. 18 shows the training data of the selected partial variable SSV7. Of this training data, based on the cases 1, 3 and 4 of the sanctions level A, which is the objective variable, the 7-dimensional multiple regression of level A is performed by the least squares method, which is the maximum likelihood estimation that minimizes the sum of squares error. The type analysis model RG_ANL_A is generated. This multiple regression analysis model RG_ANL_A corresponds to the regression prediction model Pmodel of FIG. Similarly, in the training data, a level B 7-dimensional multiple regression analysis model RG_ANL_B is generated by the least squares method based on the cases 2 and 5 of the sanctions level B, which is the objective variable. In the coordinate space of FIG. 18, the axes of the explanatory variables of the multiple regression analysis model are shown in a simplified manner.

In the regression-type cross-validation method of FIG. 17, a regression-type analysis model is generated for the evaluation target case 1 based on the remaining cases 3 and 4. This point is different from the fact that the level A regression analysis model is generated based on the cases 1, 3 and 4 of the example of FIG.

FIG. 19 shows the subvariables, and shows an example of the prediction formula of the regression analysis model for each subvariable and the prediction accuracy for each subvariable calculated by FIGS. 9 and 16. When this subvariable list is sorted with predictive accuracy, the sorted subvariable list shown in FIG. 20 is obtained.

Returning to FIG. 9, the processor determines the optimum variable range from the sorted partial variable list of FIG. 20, that is, the variable ranges Y1 to Y5 of the variables Y of the top N partial variables with high prediction accuracy (S36). .. In the example of FIG. 20, for example, among the variable ranges Y1 to Y5 of the variables Y of the upper five subvariates SSV4- # to SSC6-4, the variable range Y3 that appears most often is determined to be the optimum variable range. .. As a result, the processor determines the partial variable 36_A of the level A multiple regression analysis model as SSV3- # (variables X, Y3, Z). The partial variable 36_B of the level B multiple regression analysis model is determined in the same way.

Further, the variable range of the variable Y of the levels A and B is determined, and as each subvariable is determined, the processor is based on the determined regression type subvariable, and the level A and B are determined from the training data master 32. The respective training data 35_A and 35_B are extracted (S37). The process of FIG. 9 is repeated for all levels (S38).

With the above, the optimum variable range is determined from the variable range candidates Y1 to Y5 of the explanatory variable Y. The optimal variable range is determined for each of the taxonomic analytical model and the regression analytical model at each level (A, B). By optimizing this variable range, the prediction accuracy of the prediction model corresponding to the subvariable having the optimum variable range can be improved.

[Generation of prediction model and prediction of cases to be predicted by prediction model C, D]
Next, as shown in FIG. 3, the processor executes the prediction model generation program 22 and the prediction model program 23 to generate the prediction model and predict the objective variable by the prediction model for the prediction target case.

FIG. 21 is a flowchart of processes C and D for generating a prediction model and predicting a case to be predicted by the prediction model. The processor executes the prediction model generation program 22 to generate a classification type prediction model from the classification type learning data 26 (S41). The generation of the classification type prediction model is completed by acquiring the classification type learning data 26 as described with reference to FIG. Further, the processor executes the prediction model generation program 22 to generate a level A regression type prediction model and a level B regression type prediction model from the regression type training data 35_A and 35_B, respectively (S42, S43). The generation of the regression prediction model is as described in FIG.

This completes the generation of three types of prediction models based on past cases. That is, a classification type prediction model, a level A regression type prediction model, and a level B regression type prediction model. The processor uses these prediction models to predict the objective variable sanctions for the case to be predicted (D).

As shown in FIG. 21, the processor inputs the training data of the prediction target case (S44), applies the training data of the prediction target case to the classification type prediction model, and sets the sanction level (level A or B) of the objective variable. Predict (S45). The prediction method using the classification type prediction model has been described with reference to FIG.

The processor then selects a regression prediction model that corresponds to the predicted sanctions level (S46). Then, the processor applies the training data of the prediction target case to the selected recurrent prediction model and predicts the value of the sanction, which is the objective variable (S47). In the prediction of the sanctions of the case by the regression prediction model, the value of the variable of the training data of the prediction target case is input to the multiple regression line RG_ANL_A or RG_ANL_B (Fig. 18) of the regression prediction model, and the sanctions which are the objective variables. Calculate SA.

In this embodiment, since the number of existing cases is small, the level of sanctions (levels A and B) is predicted by the classification type prediction model (classification type analysis model), and the learning data of the existing cases of each level A and B is predicted. Predict the value of the objective variable in the case where the objective variable is unknown by the level A recurrent prediction model or the level B recurrent prediction model generated in. As shown in FIG. 18, the regression prediction model corresponding to each level is generated by the least squares method based on the learning data of a plurality of cases at each level. That is, the sanction level is predicted by a classification type prediction model with relatively high prediction accuracy even with a small number of cases. Then, the sanction value of the predicted target case is predicted by the regression type prediction model generated for each predicted sanction level. Since the recurrent prediction model is generated from the learning data of the cases divided based on the predicted sanction level, the prediction accuracy of the recurrent prediction model is high. Therefore, the accuracy of predicting the sanction value can be improved. In this way, by predicting the sanction value of an unknown case with a model that combines a separation type prediction model and a regression type prediction model, the prediction accuracy can be improved.

[Update the variable area of the explanatory variable F]
As shown in FIG. 3, in the present embodiment, when the prediction accuracy of the predicted case decreases (YES in E), the processor determines the classification type analysis model and the regression type analysis model used for the prediction. The variable range of the explanatory variables is updated to improve the prediction accuracy of each analytical model (F). Decreased prediction accuracy of predicted cases can be detected by comparing the prediction results of cases predicted by the prediction model with the sanction results of sanctions cases newly announced by the government authorities.

In the variable area update F, the optimum variable area search process B described with reference to FIGS. 3 and 8 may be executed by the learning data master in which the data of the new case is added to the existing learning data master.

As another method F for updating the variable range, the prediction accuracy of the analysis model candidate of the partial variable having a new variable range candidate is compared with the prediction accuracy of the existing analysis model for each of the classification type analysis model and the regression type analysis model. , If the prediction accuracy of the analysis model candidate is high, the existing variable area may be updated to the variable area candidate. The prediction accuracy in this case is calculated based on the learning data master in which a new case is added to the existing case.

Further, by searching for the variable area having the highest prediction accuracy between the two variable area candidates having high prediction accuracy, the variable area in which the highest prediction accuracy can be obtained can be detected with a small number of man-hours. For example, a third having a first and second analytical model between two (first and second) variable region candidates with different cap variable regions and a composite variable region of the first and second variable region candidates. The analysis model compares the prediction accuracy of each, and repeatedly determines which side of the two variable region candidates has the variable region with high prediction accuracy.

For example, in each search phase, the first and second variable ranges that define the search range of the next search phase are determined based on the order of prediction accuracy among these three analytical models, and the same comparison and determination as above are performed. Repeat to find the variable range with the highest prediction accuracy by drilling down. The third analysis model having the composite variable range of the first and second variable range candidates is that the subvariable has both the first variable range candidate Y_old and the second variable range candidate Y_new, as described later. Means.

The most typical example of reduced prediction accuracy is, for example, when the variable Y of the number of people affected by a new sanctions case becomes a larger number than in existing cases. When the number of people is large, which was not expected in the existing cases, the prediction accuracy of the prediction model may decrease in the variable range (any of Y1 to Y5) determined by the initialization process. In such a case, the processor detects the variable range of the optimum variable Y from the training data master by adding the new case published after creating the prediction model to the existing case, and updates the variable range of the prediction model. To do.

FIG. 22 is a flowchart of the process F for updating the variable area of the explanatory variable. The update of the variable range described below with reference to FIG. 22 is an example applied to the regression analysis model. However, it can also be applied to a classification type analysis model.

The variable area update process F in FIG. 22 is repeated for all levels of the objective variable (sanctions). That is, the variable range of the regression analysis model corresponding to each level of the objective variable is updated.

First, the processor executes the variable area update program 24, selects the level of the objective variable (sanction), and repeats the processing of S51 to S61 for all levels (S51, S62).

After selecting the level, the processor selects a variable range to search for a variable range with higher prediction accuracy (S52). Here, the processor calculates the prediction accuracy of the analysis model performed in the search B for the optimum variable range of the explanatory variables described above for a plurality of variable range candidates of the variable to be updated of the partial variables of the prediction model. Select the first and second variable range candidates that are predicted to have variable ranges with high prediction accuracy in between.

FIG. 23 is a diagram showing an example of a variable area list to be searched and a partial variable list to be searched. In FIG. 23, the variable area list to be searched includes a plurality of variable areas Y_UP_1 to Y_UP_5 to be searched. The variable area to be searched has a variable area Y_UP_1 having the same maximum cap area as the variable area of the existing analysis model of "10,000 or more people". In other words, based on the estimation that the maximum cap area of the variable Y "number of leaks" was larger than the existing "10,000 or more" because the sanctions for the new case were high, the variable area Y3 of the existing analysis model Variable area candidate Y_UP_2 ~ with the maximum cap area "10,000 or more" (see FIGS. 20 and 7) set to "50,000 or more", "100,000 or more", "200,000 or more", and "500,000 or more" Y_UP_5 is selected.

In FIG. 23, the partial variables SSV_UP_1 to SSV_UP_5 having the variable Y of the variable areas Y_UP_1 to Y_UP_5 to be searched are shown as search targets. In the above processing S52, based on the training data master 32_UP to which the above-mentioned new case is added, the prediction accuracy of each analysis model of the partial variables SSV_UP_1 to SSV_UP_5 is set to the same regression type cross-validation method as the optimum variable region search processing F. Calculate with.

FIG. 24 shows an example of the prediction accuracy of the analysis model of each partial variable SSV_UP_1 to SSV_UP_5 calculated in the process S52. According to this, the order of prediction accuracy is as follows.
SSV_UP_3 ＞ SSV_UP_1 ＞ SSV_UP_2 ＞ SSV_UP_4 ＞ SSV_UP_5
Y_UP_3 ＞ Y _UP_1 ＞ Y _UP_2 ＞ Y _UP_4 ＞ Y _UP_5
100,000 ＞ 50,000 ＞ 10,000 ＞ 200,000 ＞ 500,000

In this case, the variable range with the highest prediction accuracy can be predicted to exist between 100,000 and 200,000. Of course, the possibility that it exists between 50,000 and 100,000 cannot be denied, but it is assumed that the prediction can be made as described above from the rate of increase in prediction accuracy between variable region candidates. In this case, the processor selects the variable area Y_UP_3 and the variable area Y_UP_4 as the variable area to be searched (S52), and exists between the first variable area Y_UP_3 and the second variable area Y_UP_4 by the dichotomy described below. Search for the variable range with the maximum prediction accuracy. The variable area selection process S52 to be searched corresponds to the coarse search process of the variable area.

Returning to FIG. 22, the processor generates a list of partial variables in the first generation search based on the selected first and second variable regions Y_UP_3 and Y_UP_4 (S53). From here, the dense search process of the variable area is started.

FIG. 25 is a diagram showing a variable range list and a partial variable list in the first-generation and second-generation searches. Here, in the first-generation search, a partial variable SS_I_old of the first variable area Y_I_old, a partial variable SS_I_new of the second variable area Y_I_new, and a partial variable having the first and second variable areas (composite variable area). (Subvariable of) SS_I_mid Calculate the prediction accuracy of each. Then, from the magnitude relation of the three prediction accuracy, whether it is the first variable region side from the middle between the first variable region and the second variable region, the second variable region side from the middle, or the intermediate variable region. It is determined which of the above has the highest prediction accuracy.

In the subsequent second-generation search, the first-generation search is performed for either the first variable region side from the middle between the first and second variable regions, or the second variable region from the middle. Do the same. In this way, the search area is gradually narrowed by the dichotomy method, and the variable area with the highest prediction accuracy is detected with a small number of man-hours.

In the example of FIG. 25, the variable area candidate list 33_I (VR) in the first generation search includes the above-mentioned first variable area Y_I_old, the second variable area Y_I_new, and the third variable area Y_I_mid. The first variable area Y_I_old is a variable area having a maximum cap area of 100,000 or more people, and the second variable area Y_I_new is a variable area having the same area of 200,000 people or more. The third variable area Y_I_mid is a composite variable area having both the first and second variable areas.

Then, the subvariable list 33_I (SSV) in the first generation search includes the first variable area Y_I_old (= Yo), the first subvariable SS_I_old having the variables X and Y, and the second variable area Y_I_new (=). It has a second sub-variable SS_I_new having Yn) and variables X and Y, and a third sub-variable SS_I_mid having first and second variable ranges and variables X and Z.

As shown in FIG. 22, the processor then selects one subvariable in the subvariable list 33_UP (33_I (SSV) in FIG. 25), and the training data of the level selected in S51 and the selected subvariable. Is extracted from the training data master 32_UP, and the regression type analysis model is generated and the prediction accuracy is calculated (S54). This process S54 is the same process as the cross-validation method described with reference to FIGS. 16 and 17. The processor repeats the process S54 for all three subvariables (first to third subvariables SS_I_old, SS_I_new, SS_I_mid) in the subvariable list SS_I (SSV) of FIG. 25 (S55).

Further, the processor evaluates the prediction accuracy of the analysis model of all three partial variables (S56), and makes a judgment based on the comparison of the prediction accuracy of the processes S57 and S59. When the processor determines that the prediction accuracy of the partial variable SS_I_mid in the composite variable area is the highest in the first-generation search (YES in S57), the new variable area and the new part are further based on the comparison result of the prediction accuracy. Select a variable (S58), return to processing S53, and search the variable range of the second generation (processing below S53). By performing the second generation search, variable areas and subvariables with higher prediction accuracy are searched.

On the other hand, when the prediction accuracy of the analysis model of the partial variable SS_I_new of the variable area Y_I_new is the highest (YES in S59), the processor determines the variable area Y_I_new as the new variable area (S61). Conversely, the processor determines the variable region Y_I_old as the new variable region when the prediction accuracy of the analysis model of the partial variable SS_I_old of the variable region Y_I_old is the highest (NO in S59).

FIG. 28 is a diagram showing values applied to the variables Y of the first to third variable ranges in the variable range list of the first generation search. In the first variable area Y_I_old, the maximum cap area is 100,000, and in the case where the number of leaks is 100,000 or more, the value of the variable Y increases from 3 to 4. On the other hand, in the second variable area Y_I_new, the maximum cap area is 200,000, and the value of the variable Y increases from 3 to 4 in the case where the number of leaks is 200,000 or more. Then, since the third variable area I_I_mid becomes a subvariable having the first and second variable areas, the area of the maximum cap is substantially equivalent to the subvariable having the variable area of 150,000 people. ..

As described above, an analytical model having variables in different variable ranges will have different quantitative values depending on the scale of the number of leaked cases, and the predicted sanctions will also be different. Therefore, if government authorities impose unprecedented sanctions on unprecedented cases of maximum leaks, the variable range of the leaked variable Y may have changed.

In FIG. 25, an example of the prediction formula of the prediction model and an example of the calculated prediction accuracy are shown in the subvariable list 33_I (SSV) of the first-generation search. The following two examples are examples of prediction accuracy shown in the partial variable list 33_I (SSV).
SS_I_mid (= 0.85) ＞ SS_I_old (= 0.75) ＞ SS_I_new (= 0.70): C>A> B (Y_I_old ~ Y_I_mid side)
SS_I_mid (= 0.85) ＞ SS_I_new (= 0.75) ＞ SS_I_old (= 0.70): C>B> A (Y_I_mid ~ Y_I_new side)
Here, A means SS_I_old, B means SS_I_new, and C means SS_I_mid.

FIG. 26 is a diagram showing a comparison result of prediction accuracy and a determination example thereof. Further, FIG. 27 is a diagram showing the relationship between the three subvariables of the first-generation and second-generation searches. FIG. 27 conceptually shows a function of an analytical model in which the horizontal axis is the axis of the explanatory variable and the vertical axis is the axis of the objective variable.

Two comparative examples of prediction accuracy shown in FIG. 25 are C> A> B and C> B> A as described above. Then, C> A> B and C> B> A are cases where the partial variable C in the composite variable region has the highest prediction accuracy (YES in S57).

The determination contents corresponding to the prediction accuracy comparisons C> A> B and C> B> A in FIG. 26 are as follows. In the case of C> A> B (case a), the area of 10% to 50% between A: SS_I_old and B: SS_I_new is set as the range of the new variable area search. Further, in the case of C> B> A (case b), the area of 50% to 90% between A: SS_I_old and B: SS_I_new is set as the range of the new variable area search. The content of this determination will be described below.

FIG. 27 shows the conceptual relationship between the functions of the first, second, and third partial variables SS_I_old (= A), SS_I_new (= B), and SS_I_mid (= C) in the first generation search. ..

Then, in the case of C> A> B (case a), the prediction model of the partial variable including the variable range searched (a) in the second generation is between SS_II_old_a (= SS_I_mid) and SS_II_new_a. In this case, the subvariable (variable area) with the highest prediction accuracy is presumed to be somewhere on the SS_I_old (Y_old) side near the first generation SS_I_mid (Y_mid).

FIG. 25 shows a list 33_II (VR) of the first and second variable areas to be searched in the second generation. When C> A> B, the first variable area is Y_II_old_a and the second variable area. Becomes Y_II_new_a. Further, the second, first, and third partial variables searched (a) in the second generation in the case of C> A> B are SS_II_new_a, SS_II_old_a, and SS_II_mid_a in the partial variable list 33_II_a (SSV).

On the other hand, when C> B> A (case b), the first and second variable areas searched (b) in the second generation are between SS_II_old_b (= SS_I_mid) and SS_II_new_b. In this case, the variable range with the highest prediction accuracy is presumed to be somewhere on the SS_I_new (Y_new) side near the first generation SS_I_mid (Y_mid).

In the list 33_II (VR) of the first and second variable areas searched in the second generation in FIG. 25, when C> B> A, the first variable area is Y_II_old_b and the second variable area is Y_II_new_b. It becomes. Further, the first, second, and third subvariables to be searched (b) in the second generation in the case of C> B> A are SS_II_old_b, SS_II_new_b, SS_II_mid_b in the subvariable list 33_II_b (SSV).

FIG. 27 shows the first to third partial variables SS_II_old_a, SS_II_mid_a, SS_II_new_a in the second generation search (a) in the case of C> A> B (case a). The area of the second generation search is close to the third subvariable SS_I_mid of the first generation search and is in the area on the side of the first subvariable SS_I_old.

Further, FIG. 27 shows the first to third partial variables SS_II_old_b, SS_II_mid_b, SS_II_new_b in the second generation search (b) in the case of C> B> A (case b). The area of the second generation search is close to the third subvariable SS_I_mid of the first generation search and is in the area on the second subvariable SS_I_new side.

FIG. 29 is a diagram showing values applied to the variables Y of the first to third variable ranges in the variable range list of the second generation search. FIG. 29 is an example in the case of C> A> B (case a). In the first variable area Y_II_old_a, the value of the variable Y increases from 3 to 4 in the case where the maximum cap area is 150,000 and the number of leaks is 150,000 or more. On the other hand, in the second variable area Y_II_new_a, the maximum cap area is 120,000, and the value of the variable Y increases from 3 to 4 in the case where the number of leaks is 120,000 or more. Then, since the third variable region Y_II_mid_a becomes a partial variable having the first and second variable regions, the region of the maximum cap has a variable region between 120,000 and 150,000 people. Equivalent to a variable.

FIG. 30 is a diagram showing values applied to the variables Y of the first to third variable ranges in the variable range list of the second generation search. FIG. 29 is an example in the case of C> B> A (case b). In the first variable area Y_II_old_b, the maximum cap area is 150,000 people. On the other hand, in the second variable area Y_II_new_b, the maximum cap area is 170,000 people. Then, since the third variable region Y_II_mid_b becomes a partial variable having the first and second variable regions, the region of the maximum cap has a variable region between 150,000 and 170,000 people. Equivalent to a variable.

Comparison of four prediction accuracy in the remaining first-generation search in FIG. 26 The judgment contents corresponding to A> B> C, A> C> B, B> A> C, B> C> A are as follows. Is.

When A> B> C, the prediction accuracy of the first subvariable SS_I_old (first variable area Y_I_old) of A is the highest, so the processor changes it to the first variable area Y_I_old (S60) and changes the variable area. Ends the update process of.

When A> C> B, the prediction accuracy of the first subvariable SS_I_old (first variable range Y_I_old) of A is the highest, but the prediction accuracy of the third subvariable SS_I_mid (third variable range Y_I_mid) of C is also high. Therefore, the processor changes to the fine-tuned first variable area Y_I_old and ends the update process of the variable area.

When B> A> C, the prediction accuracy of the second subvariable SS_I_new (second variable area Y_I_new) of B is the highest, so the processor changes to the second variable area Y_I_new (S61) and changes the variable area. Ends the update process of.

When B> C> A, the prediction accuracy of B's second sub-variable SS_I_new (second variable range Y_I_new) is the highest, but C's third sub-variable SS_I_mid (third variable range Y_I_mid) is also high. Therefore, the processor changes to the fine-tuned second variable area Y_I_new and ends the update process of the variable area.

Then, in the case of C> A> B, C> B> A, the processor continues to search the variable range of the second generation. In each search generation, when the prediction accuracy comparison is A> B> C, A> C> B, B> A> C, B> C> A, the processor is the first generation, as in the first generation above. Change to the variable area or the second variable area, and end the update process of the variable area.

Furthermore, in the second and subsequent generations, in the case of C> A> B, C> B> A, if the search is continued up to a predetermined generation, the variable area of C is changed at the next point, and the search process is terminated. .. This is because the disadvantage of increasing the search man-hours outweighs the advantage of increasing the prediction accuracy.

According to the process F for updating the variable area of the explanatory variable shown in FIG. 22 above, the man-hours for the variable area update process can be reduced by combining the coarse variable area search S52 and the dense variable area search S53 to S61. .. Of course, the variable area of the explanatory variable is updated, and the prediction accuracy of the analysis model is comprehensively set by setting a large number of variable area candidates in the same way as the search process B of the optimum variable area of the explanatory variable at the time of initialization. May be calculated and updated to the variable range with the best prediction accuracy.

10: Predictor, computer 21: Variable area search program 22: Prediction model generation program 23: Prediction program 24: Variable area update program 31: Case data 32: Learning data master 33: Variable area list and partial variable list 34: For classification Training data 35: Training data for regression
X, Y, Z: Explanatory variables
Y1 to Y5: Variable area, variable area candidate
SA: Objective variable

Claims (10)

With the processor
The processor has accessible memory and
The processor
(A) For a plurality of case data having values of objective variables corresponding to values of a plurality of explanatory variables, the values of the explanatory variables are replaced with identification numbers of the plurality of regions based on the variable regions having a plurality of regions. A classification type prediction model and a regression type prediction model are generated based on the multiple training data.
The classification type prediction model has a plurality of classification type learning data in which the value of the objective variable of the plurality of training data is replaced with a plurality of levels according to the size, and among the plurality of classification type training data. The level of the objective variable of the training data of the prediction target case and the classification type learning data having the shortest norm distance between the values of the explanatory variables is determined to be the level of the objective variable of the training data of the prediction target case.
Among the plurality of training data, the regression type prediction model has a regression line close to the coordinate points of the explanatory variables of the plurality of training data for each level divided by the level of the objective variable for each level. The value of the objective variable of the training data of the prediction target case is calculated from the regression line corresponding to the level determined by the classification type prediction model.
Further, (b) the training data of the prediction target case is applied to the classification type prediction model to predict the level of the objective variable of the training data of the prediction target case.
(C) A prediction device that predicts the value of the objective variable of the training data of the prediction target case by applying the training data of the prediction target case to the regression line corresponding to the predicted level. The processor
A plurality of subvariables that are subsets of a set having the plurality of explanatory variables and a plurality of variable range candidates set in at least one explanatory variable are generated.
For the training data of the plurality of partial variables obtained by extracting the values of the explanatory variables and the variable range candidates included in each of the plurality of partial variables from the plurality of training data, the prediction accuracy of the classification type prediction model is determined for each of the partial variables. Calculate and
The variable range candidate that is most contained in the N subvariates with high prediction accuracy is determined as the variable range of the classification type prediction model, and the N pieces are positive integers.
The prediction device according to claim 1, wherein the classification type prediction model is generated after the variable range of the classification type prediction model is determined. The processor
(D) The regression type prediction for each of the partial variables with respect to the training data of the plurality of partial variables obtained by extracting the values of the explanatory variables and the variable range candidates included in each of the plurality of partial variables from the plurality of training data for each level. Calculate the prediction accuracy of the model,
(E) The variable region candidate contained most in the M subvariates with high prediction accuracy is determined as the variable region of the regression prediction model, and the M are positive integers.
The processes (d) and (e) are repeated for the plurality of levels.
The prediction device according to claim 2, wherein the regression type prediction model is generated after the variable range of the regression type prediction model is determined. The predictor according to claim 2 or 3, wherein the plurality of variable region candidates have different at least maximum regions of the plurality of regions of each variable region candidate. In the calculation of the prediction accuracy of the classification type prediction model for each partial variable,
The processor
The level of the objective variable of the training data of the evaluation target case is predicted by a classification type prediction model having the training data of the remaining cases excluding the evaluation target case among the training data of the partial variables, and the prediction is made. Judging the success or failure of the prediction of the evaluation target case based on the match or disagreement between the level and the learning data level of the evaluation target case is repeated with all the cases as the evaluation target cases, and the success or failure of the evaluation target cases is determined. The prediction device according to claim 2, which outputs the ratio as prediction accuracy. In the calculation of the prediction accuracy of the regression type prediction model for each partial variable,
The processor
(F) A regression-type prediction model generated for the training data of the remaining cases excluding the cases to be evaluated among the training data of the partial variables for each level in which the training data of the partial variables are divided according to the level of the objective variable. , The value of the objective variable of the learning data of the case to be evaluated is predicted, and the ratio of the predicted value to the value of the objective variable of the learning data of the case to be evaluated is calculated. Repeat as an example,
(G) The average value of the ratio to all the cases is output as the prediction accuracy.
The predictor according to claim 3, wherein the processes (f) and (g) are repeated for the plurality of levels. The prediction accuracy obtained by comparing the value of the objective variable of the training data of the prediction target case predicted using the regression prediction model with the definite value of the objective variable of the training data of the new case is lower than the reference accuracy. if you did this,
The processor
A plurality of subvariables that are subsets of a set having the plurality of explanatory variables and a plurality of variable range candidates set in at least one explanatory variable are generated.
(H) Explanatory variables included in each of the plurality of partial variables from the new plurality of training data for each level, in which the new plurality of training data obtained by adding the training data of the new case to the plurality of training data are divided according to the level. And, for the training data of the plurality of partial variables extracted from the values of the variable range candidates, the prediction accuracy of the regression type prediction model is calculated for each of the partial variables.
(I) The variable range of the regression type prediction model is updated with the variable range candidate included in the subvariable with relatively high prediction accuracy.
The predictor according to claim 1, wherein the processes (h) and (i) are repeated for the plurality of levels. The plurality of variable area candidates differ in at least the maximum area of each variable area candidate.
The plurality of variable region candidates include a first variable region candidate having a first maximum region larger than the maximum region of the variable candidate having the maximum prediction accuracy, and a second variable having a small second maximum region. It has a region candidate and a third variable region candidate having a composite variable region of the first variable region candidate and the second variable region candidate.
The processor
In the process (h),
In the first generation search process
The first prediction accuracy of the regression type prediction model of the first subvariable having the first variable range candidate and the second of the regression type prediction model of the second partial variable having the second variable range candidate. The prediction accuracy of 2 is compared with the third prediction accuracy of the regression prediction model of the third subvariable having the third variable range candidate.
When the third prediction accuracy is higher than the first and second prediction accuracy, in the search process of the variable range of the second generation,
A variable region having a third maximum region between the first maximum region and the second maximum region is set as the first variable region candidate.
A variable region having a fourth maximum region between the third maximum region and the first maximum region, or a fifth maximum region between the third maximum region and the second maximum region. As the second variable range candidate, the variable range having
The prediction device according to claim 7, which executes the first-generation search process. A computer-readable prediction program that causes a computer to perform prediction processing.
The prediction process is
(A) For a plurality of case data having values of objective variables corresponding to values of a plurality of explanatory variables, the values of the explanatory variables are replaced with identification numbers of the plurality of regions based on the variable regions having a plurality of regions. A classification type prediction model and a regression type prediction model are generated based on the multiple training data.
The classification type prediction model has a plurality of classification type learning data in which the value of the objective variable of the plurality of training data is replaced with a plurality of levels according to the size, and among the plurality of classification type training data. The level of the objective variable of the training data of the prediction target case and the classification type learning data having the shortest norm distance between the values of the explanatory variables is determined to be the level of the objective variable of the training data of the prediction target case.
Among the plurality of training data, the regression type prediction model has a regression line close to the coordinate points of the explanatory variables of the plurality of training data for each level divided by the level of the objective variable for each level. The value of the objective variable of the training data of the prediction target case is calculated from the regression line corresponding to the level determined by the classification type prediction model.
Further, (b) the training data of the prediction target case is applied to the classification type prediction model to predict the level of the objective variable of the training data of the prediction target case.
(C) A prediction program that predicts the value of the objective variable of the training data of the prediction target case by applying the training data of the prediction target case to the regression line corresponding to the predicted level. The processor
(A) For a plurality of case data having values of objective variables corresponding to values of a plurality of explanatory variables, the values of the explanatory variables are replaced with identification numbers of the plurality of regions based on the variable regions having a plurality of regions. A classification type prediction model and a regression type prediction model are generated based on the multiple training data.
The classification type prediction model has a plurality of classification type learning data in which the value of the objective variable of the plurality of training data is replaced with a plurality of levels according to the size, and among the plurality of classification type training data. The level of the objective variable of the training data of the prediction target case and the classification type learning data having the shortest norm distance between the values of the explanatory variables is determined to be the level of the objective variable of the training data of the prediction target case.
Among the plurality of training data, the regression type prediction model has a regression line close to the coordinate points of the explanatory variables of the plurality of training data for each level divided by the level of the objective variable for each level. The value of the objective variable of the training data of the prediction target case is calculated from the regression line corresponding to the level determined by the classification type prediction model.
Further, (b) the training data of the prediction target case is applied to the classification type prediction model to predict the level of the objective variable of the training data of the prediction target case.
(C) A prediction method in which the training data of a prediction target case is applied to the regression line corresponding to the predicted level to predict the value of the objective variable of the training data of the prediction target case.