JP7645462B2 - Model generation device, model generation method and program - Google Patents

Description

The present disclosure relates to a model generation device that generates a model that shows relationships between data.

A model generation device has been proposed that generates a model for estimating data. For example, the model generation device generates a model by selecting a response variable and an explanatory variable from a data set including data for each of a plurality of variables, and deriving a correlation coefficient between those variables or a regression model using those variables. The data set includes, for example, a plurality of manufacturing data related to manufacturing. The response variable indicates, for example, the quality characteristics of a manufactured product as data, and the explanatory variables indicate data on parameters used in the manufacturing process. Therefore, by using the generated model, it is possible to estimate the quality characteristics of a product from the manufacturing process.

Also, a related analysis device that creates a local quality model has been proposed (see, for example, Patent Document 1). This related analysis device can be said to be a model generation device because it creates a local quality model as the above-mentioned model. In addition, this related analysis device divides the space of operational factors corresponding to explanatory variables into multiple local regions, and creates a local quality model for each of the local regions.

Patent No. 4653547

However, the model generation device, which is the related analysis device in Patent Document 1, has the problem that it is difficult to improve the accuracy of the model.

Therefore, the present disclosure provides a model generation device that can easily improve the accuracy of a model.

A model generation device according to one aspect of the present disclosure is a model generation device that generates a model showing a relationship between one or more objective variables and one or more explanatory variables, and includes a receiving means for receiving a dataset including three or more variables, a variable identification means for identifying one or more objective variables and one or more explanatory variables from the dataset, and identifying one or more variables as stratification variables from among the three or more variables included in the dataset other than the identified objective variables and explanatory variables, a stratification means for classifying the dataset into a plurality of strata based on the tendency of the relationship between the stratification variables and the objective variables, and a generation means for generating the model for each of the plurality of strata.

These comprehensive or specific aspects may be realized by a system, a method, an integrated circuit, a computer program, or a recording medium such as a computer-readable CD-ROM, or may be realized by any combination of a system, a method, an integrated circuit, a computer program, and a recording medium. The recording medium may also be a non-transitory recording medium.

The model generation device disclosed herein can easily improve the accuracy of the model.

Further advantages and benefits of certain aspects of the present disclosure will become apparent from the specification and drawings. Such advantages and/or benefits may be provided by some of the embodiments and features described in the specification and drawings, respectively, but not necessarily all of them need be provided to obtain one or more identical features.

FIG. 1 is a diagram illustrating an example of a model generation system according to an embodiment. FIG. 2 is a diagram showing a configuration of a model generating device according to an embodiment. FIG. 3A is a diagram illustrating an example of a data set according to the embodiment. FIG. 3B shows the first and second rows of the data set of FIG. 3A. FIG. 3C is a diagram illustrating a response variable and an explanatory variable selected from a dataset in the embodiment. FIG. 4 is a diagram showing another example of a data set according to the embodiment. FIG. 5 is a diagram showing a simplified view of variable names of a data set according to the embodiment. FIG. 6 is a block diagram showing a functional configuration of the model generating device according to the embodiment. FIG. 7 is a diagram illustrating an example of a stratified data set according to the embodiment. FIG. 8 is a diagram showing the distribution of coordinate points indicated by each record included in each stratified data set in the embodiment. FIG. 9 is a flowchart showing an example of the overall processing operation of the model generating device according to the embodiment. FIG. 10 is a flow chart showing a specific example of the process of calculating the influence degree of the qualitative variables in step S8 of FIG. FIG. 11 is a flow chart showing a specific example of the process of calculating the influence of the quantitative variables in step S9 of FIG.

(Findings that formed the basis of this disclosure)
The present inventors have found that the model generating device of Patent Document 1 described in the "Background Art" section has the following problems.

In the above Patent Document 1, the space of the operation factors corresponding to the explanatory variables is divided into multiple local regions, and a local quality model is constructed for each local region. Therefore, multiple models are generated from a dataset. However, only the objective variables and explanatory variables included in the dataset are used to construct these models, and variables other than the objective variables and explanatory variables are not used. Specifically, the dataset is divided into multiple local regions based only on the distribution of data indicated by the explanatory variables included in the dataset, and models are generated for these local regions. In other words, in the above Patent Document 1, since the influence of the explanatory variables and variables other than the objective variables on the correlation between the explanatory variables and the objective variables is unknown, the explanatory variables and variables other than the objective variables are not used to construct the model. Therefore, when data of the explanatory variables and variables other than the objective variables affect the correlation between the explanatory variables and the objective variables, it is difficult to generate a model with high accuracy.

In this regard, a model generation device according to one aspect of the present disclosure is a model generation device that generates a model showing a relationship between one or more objective variables and one or more explanatory variables, and includes a receiving means for receiving a dataset including three or more variables, a variable identification means for identifying one or more objective variables and one or more explanatory variables from the dataset, and identifying one or more variables as stratification variables from among the three or more variables included in the dataset other than the identified objective variables and the explanatory variables, a stratification means for classifying the dataset into a plurality of strata based on the tendency of the relationship between the stratification variables and the objective variables, and a generation means for generating the model for each of the plurality of strata. For example, the stratification means may classify data of the stratification variables into a plurality of groups for each of the stratification variables based on the identity or similarity of the data of the stratification variables, and classify the dataset for each combination of the plurality of groups.

As a result, stratification is performed according to the stratification variables, which are variables other than the objective variable and the explanatory variable, and a model is generated for each of the multiple strata. That is, each data of the stratification variables is classified into multiple groups according to the commonality or similarity between the data. Then, the dataset is stratified into strata corresponding to those groups. By this stratification, each strata includes one or more records each having data of the stratification variables belonging to the same group. Note that a group is a collection of data having commonality or similarity, and this group includes a category, which is a collection of common data, and a cluster, which is a collection of similar numerical data. In addition, the stratification variables are non-utilized variables that are not adopted as variables included in the model, but are used in generating the model. In this way, in the model generating device according to one aspect of the present disclosure, stratification is performed on the dataset by non-utilized variables, which are variables other than the explanatory variable and the objective variable, among the variables included in the dataset, so that even if the non-utilized variables affect the correlation between the explanatory variable and the objective variable, a highly accurate model based on the non-utilized variables can be generated. In other words, the accuracy of the model can be easily improved.

For example, a first stratification variable and a second stratification variable, each of which is an unexploited variable, are identified. Then, two or more data of the first stratification variable are classified, for example, into a first group and a second group, and two or more data of the second stratification variable are classified, for example, into a third group and a fourth group. All data included in these groups have commonality or high similarity. In this case, for example, a first layer, a second layer, a third layer, and a fourth layer are determined as the multiple layers. The first layer corresponds to a combination of the first group of the first stratification variable and the third group of the second stratification variable. The second layer corresponds to a combination of the first group of the first stratification variable and the fourth group of the second stratification variable. The third layer corresponds to a combination of the second group of the first stratification variable and the third group of the second stratification variable. The fourth layer corresponds to a combination of the second group of the first stratification variable and the fourth group of the second stratification variable. In this way, multiple layers are determined according to the combination of the groups of the N stratification variables. Thus, in stratified classification, a record containing data belonging to the first group of the first stratification variable and data belonging to the third group of the second stratification variable is classified into the first stratum. A record containing data belonging to the first group of the first stratification variable and data belonging to the fourth group of the second stratification variable is classified into the second stratum. A record containing data belonging to the second group of the first stratification variable and data belonging to the third group of the second stratification variable is classified into the third stratum. A record containing data belonging to the second group of the first stratification variable and data belonging to the fourth group of the second stratification variable is classified into the fourth stratum.

In this way, even if there are two or more stratification variables, optimal stratification classification can be performed for the dataset, and a highly accurate model can be generated for each of the multiple strata according to the data for those stratification variables, i.e., the N non-utilized variables.

The data set may also include qualitative variables that represent data that includes characters and quantitative variables that represent data that includes numbers.

This allows the identification of multiple stratification variables that include not just one of qualitative and quantitative variables, but both, and increases the degree of freedom in the variable types of the identified stratification variables.

The data set may further include an influence calculation means for calculating the influence of each of the three or more variables included in the data set, other than the identified objective variable and explanatory variables, on the objective variable, and the variable identification means may identify the stratification variables based on the influence.

This makes it possible to identify stratification variables that have a large influence on the target variable data. As a result, by using these stratification variables, more optimal stratification classification can be performed for the dataset. As a result, a more accurate model can be generated for each of the multiple strata.

The system may further include an influence calculation means for calculating the influence of each of the three or more variables included in the dataset other than the identified objective variable and explanatory variables on the objective variable, the receiving means receiving priority information indicating whether the qualitative variable or the quantitative variable is to be prioritized, and the variable identification means identifying the stratification variable based on the influence and the priority information.

Thereby, for example, if the priority information is received by the receiving means in response to a user's input operation, the priority of the qualitative and quantitative variables can be set arbitrarily according to the user's intention, and many of the variables with higher priorities can be identified as stratification variables. As a result, appropriate stratification classification can be performed depending on whether the data set is analyzed qualitatively or quantitatively, and a model can be generated according to the analysis mode.

The influence calculation means may also calculate the influence of the qualitative variables using a random forest or a gradient boosting decision tree.

This makes it possible to calculate the appropriate influence of a qualitative variable, for example by calculating a value corresponding to the Gini coefficient of a random forest as the influence.

The influence calculation means may also use a Gaussian mixture model or a k-means method to perform clustering on two or more pieces of data of the quantitative variables included in the dataset, and calculate the influence of the quantitative variables by a simple regression analysis between each cluster obtained by the clustering and the objective variable.

This allows the influence of one or more quantitative variables, for example for each cluster, to be calculated by simple regression analysis, and the influence of the quantitative variable can be appropriately calculated from the influence for each cluster.

The generating means may further calculate an index indicating the likelihood of each of the multiple models generated, and the model generating device may further include a result output means for outputting the calculated index for each of the multiple models.

As a result, for example, the coefficient of determination adjusted for the degrees of freedom for each of multiple models is calculated and output as an index indicating the reliability of the model. Therefore, the user can easily decide whether or not to use the generated model based on the index.

The generating means may generate the model showing the relationship between the data of each of the two or more explanatory variables and the data of the objective variable as a multiple regression equation.

This allows us to generate an appropriate model regardless of the number of explanatory variables.

The following describes the embodiment in detail with reference to the drawings.

The embodiments described below are all comprehensive or specific examples. The numerical values, shapes, materials, components, component placement and connection forms, steps, and order of steps shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Furthermore, among the components in the following embodiments, components that are not described in an independent claim that indicates a superordinate concept are described as optional components.

The figures are schematic diagrams and are not necessarily precise illustrations. In each figure, the same components are given the same reference numerals.

(Embodiment 1)
[Hardware configuration]
FIG. 1 is a diagram showing an example of a model generation system according to the present embodiment.

The model generation system 1 in this embodiment includes a model generation device 100 and a manufacturing management device 500.

The manufacturing management device 500 is installed, for example, in a manufacturing factory, and is a device that manages a manufacturing system that manufactures products. This manufacturing management device 500 transmits a data set Ds obtained by the manufacturing system to the model generating device 100 via a network such as the Internet. Details of the data set Ds will be described later using Figures 3A to 5.

The model generating device 100 is configured with a personal computer or the like, and receives the data set Ds from the above-mentioned manufacturing management device 500. Then, the model generating device 100 in this embodiment generates multiple models showing the relationship between the data of the explanatory variables and the data of the objective variables based on the data set Ds.

Figure 2 shows the configuration of the model generation device 100 in this embodiment.

The model generating device 100 includes an input unit 101, an arithmetic circuit 102, a memory 103, an output unit 104, a storage unit 105, a database 106, and a communication unit 107.

The communication unit 107 communicates with devices external to the model generating device 100. The communication may be wireless or wired. The wireless communication method may be Wi-Fi (registered trademark), Bluetooth (registered trademark), or ZigBee (registered trademark), or may be another method. For example, the communication unit 107 communicates with the manufacturing management device 500 and receives a data set Ds from the manufacturing management device 500.

The input unit 101 functions as an HMI (Human Machine Interface) that accepts input operations by a user, and includes, for example, a keyboard, a mouse, a touch sensor, a touch pad, etc.

The output unit 104 has a display for displaying images or characters, and the display is, for example, a liquid crystal display, a plasma display, or an organic EL (Electro-Luminescence) display. The output unit 104 may have a printer for printing images or characters, and may have a function for storing data output from the arithmetic circuit 102 in a file format in the memory unit 105.

The storage unit 105 stores a program (i.e., a computer program) 105a in which each command to the arithmetic circuit 102 is written. The storage unit 105 may also store each temporary data 105b temporarily generated by the processing of the arithmetic circuit 102. The storage unit 105 is a non-volatile recording medium, for example, a magnetic storage device such as a hard disk, an optical disk, or a semiconductor memory. The program 105a is provided to the model generating device 100 via, for example, a removable medium or a network, and is stored in the storage unit 105. The removable medium is, for example, a CD-ROM (Compact Disc Read Only Memory), a flash memory, or the like. For this reason, the communication unit 107 may be provided with an interface for reading the program 105a from the removable medium.

The memory 103 temporarily stores the program 105a that is read and expanded by the arithmetic circuit 102. Such memory 103 is, for example, a volatile RAM (Random Access Memory).

The arithmetic circuit 102 is a circuit that executes the program 105a expanded in the memory 103, and is, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). When executing the program 105a, the arithmetic circuit 102 may use each temporary data 105b stored in the storage unit 105.

Like the storage unit 105, the database 106 is a non-volatile recording medium, for example, a magnetic storage device such as a hard disk, an optical disk, or a semiconductor memory. For example, the arithmetic circuit 102 acquires a data set Ds from the manufacturing management device 500 via the network and the communication unit 107, and stores the data set Ds in the database 106.

In this embodiment, the storage unit 105 and the database 106 are different recording media, but the storage unit 105 and the database 106 may be configured as a single recording medium that includes them.

[Dataset]
Fig. 3A is a diagram showing an example of a data set Ds in this embodiment, and Fig. 3B is a diagram showing the first row and the second row of the data set Ds.

The data set Ds is a raw data set transmitted from the manufacturing management device 500, and is a structured data set consisting of multiple manufacturing data indicating, for example, the manufacturing process in the above-mentioned manufacturing system and the quality of the product manufactured by the manufacturing process. As shown in FIG. 3A, such a data set Ds indicates the variable names of multiple variables and the data of those variables. Note that the data may be any type as long as it indicates at least one of letters and numbers. The first row of the data set Ds contains the variable names of the multiple variables, and the second and subsequent rows of the data set Ds contain the data of the multiple variables. Each of these rows from the second row onwards is treated as a record containing the data of the multiple variables. The leftmost column of the data set Ds is a linking information column, and indicates an ID, which is identification information for identifying those records. The ID links the data of each variable contained in the record.

Specifically, as shown in FIG. 3B, the first row of the data set Ds contains the variable names "Voltage," "Speed," "Resistance," "Worker," "Facility No.," "Material Mix," "Material Temperature Difference," "Auxiliary Voltage," and "Jig Temperature." The second row of records contains data d1 to d9 of variables identified by those variable names. Data d1 is data of a variable identified by the variable name "Voltage," for example, "5.488135." Data d2 is data of a variable identified by the variable name "Speed," for example, "7.151894." Data d3 is data of a variable identified by the variable name "Resistance," for example, "44.69831." Data d4 is data of a variable identified by the variable name "Worker," for example, "Suzuki." Data d5 is data of a variable identified by the variable name "Facility No.," for example, "Unit C." Data d6 is data for a variable identified by the variable name "material composition", for example "0". Data d7 is data for a variable identified by the variable name "material temperature difference", for example "8.815673". Data d8 is data for a variable identified by the variable name "auxiliary voltage", for example "3". Data d9 is data for a variable identified by the variable name "jig temperature", for example "9.298481". The record containing data d1 to d9 for each of these variables is identified by the ID "ID200901". In other words, the ID "ID200901" links the data d1 to d9 contained in the record identified by that ID.

As shown in FIG. 3A, the dataset Ds includes a plurality of such records. For example, the dataset Ds includes a record identified by the above-mentioned ID "ID200901", a record identified by the ID "ID200902", and a record identified by the ID "ID200903". In this way, the dataset Ds in this embodiment includes two or more records having data for each of a plurality of variables.

Also, as shown in FIG. 3A, the second column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Voltage". For example, the data are "5.488135", "6.027634", and "4.236548". Similarly, the third column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Speed". The fourth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Resistance". The fifth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Worker". The sixth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Equipment Number". The seventh column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Material Mixture". The eighth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "Material Temperature Difference". The ninth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "auxiliary voltage". The tenth column from the left of the data set Ds shows data for each record of the variable identified by the variable name "jig temperature".

Here, there are quantitative and qualitative variables as types of variables. Each data of a quantitative variable is data expressed only by numbers, and each data of a qualitative variable is data expressed including characters. In the example of FIG. 3A and FIG. 3B, the variables identified by the variable names "worker" and "equipment number" are qualitative variables. For example, as shown in FIG. 3B, data d4 of the variable identified by the variable name "worker" is "Suzuki", which includes characters. Therefore, the variable identified by the variable name "worker" is a qualitative variable. Also, in the example of FIG. 3A and FIG. 3B, the variables identified by the variable names "material composition", "material temperature difference", "auxiliary voltage", and "jig temperature" are quantitative variables. For example, as shown in FIG. 3B, data d6 of the variable identified by the variable name "material composition" is "0", which is expressed only by numbers. Therefore, the variable identified by the variable name "material composition" is a quantitative variable.

The variables are not limited to the examples shown in FIG. 3A and FIG. 3B, and may be any variables. The variables may be, for example, variables related to people, variables related to materials, variables related to equipment, etc. The variables related to people may be variables such as "worker" or "work crew". The variables related to materials may be variables such as "source material lot" or "intermediate process material lot". The variables related to equipment may be variables such as "production equipment type, generation", "production equipment number", "by lane in equipment, by spindle", "mold", "jig", "mold temperature", "drying temperature", "before and after equipment maintenance", etc. The variables may also be variables such as "ambient temperature", "ambient humidity", "production period, time", etc. The variables may also be variables such as "production type, part number" or "product size" of the product.

Figure 3C shows the objective variables and explanatory variables selected from the dataset Ds.

Each variable shown in the data set Ds is classified into an utilized variable and a non-utilized variable according to the input operation by the user, and each utilized variable is classified into an explanatory variable and a target variable. The utilized variable is a variable adopted in the model, and the non-utilized variable is a variable not adopted in the model. In addition, the non-utilized variable is a variable that was not adopted in the model because, conventionally, all variables other than the target variable included in the data set are not adopted as explanatory variables, and only variables that contribute greatly to the target variable are adopted as explanatory variables. Furthermore, the utilized variables include explanatory variables and target variables. In the example shown in FIG. 3C, the user selects the variable with the variable name "resistance value" as the target variable, and selects the variable with the variable name "voltage" and the variable with the variable name "speed" as the explanatory variables. As a result, the variables with the variable names "operator", "equipment number", "material composition", "material temperature difference", "auxiliary voltage", and "jig temperature" are determined as non-utilized variables. These non-utilized variables include the above-mentioned qualitative and quantitative variables. Therefore, in this embodiment, the M non-utilized variables shown in the data set Ds include one or more qualitative variables each indicating data containing characters, and one or more quantitative variables each indicating data consisting of numbers. Note that M is the number of non-utilized variables included in the data set Ds, which is six in the above example.

Figure 4 shows another example of the data set Ds in this embodiment.

The arithmetic circuit 102 replaces the qualitative variables included in the data set Ds with dummy variables. In other words, the arithmetic circuit 102 performs One Hot Encoding to convert the data of the qualitative variable with the variable name "Worker" shown in Figures 3A to 3C into a flag string consisting of data of three variables, namely, the variable names "Worker Suzuki", "Worker Sato", and "Worker Takahashi". For example, the data "Suzuki" of the variable name "Worker" shown in Figure 3A is converted into a flag string consisting of the data "1" of the variable name "Worker Suzuki", the data "0" of the variable name "Worker Sato", and the data "0" of the variable name "Worker Takahashi". Also, the data of the variable "Sato" with the variable name "Worker" shown in FIG. 3A is converted into a flag string consisting of the data of the variable "Worker Suzuki" with the variable name "Worker Sato", the data of the variable "Worker Takahashi", and the data of the variable "Worker Takahashi". Similarly, the arithmetic circuit 102 converts the data of the qualitative variable with the variable name "Facility No. " shown in FIG. 3A to FIG. 3C into a flag string consisting of the data of three variables with the variable names "Facility No. C", "Facility No. D", and "Facility No. E". For example, the data of the variable "Facility No. C" with the variable name "Facility No. " shown in FIG. 3A is converted into a flag string consisting of the data of the variable name "Facility No. C", the data of the variable name "Facility No. D", and the data of the variable name "Facility No. E". Furthermore, the data of the variable "Facility Unit D" shown in FIG. 3A is converted into a flag string consisting of the data of the variable "Facility Unit C" (0), the data of the variable "Facility Unit D" (1), and the data of the variable "Facility Unit E" (0). When handling qualitative variables in machine learning such as random forest described below, the arithmetic circuit 102 replaces the qualitative variables with dummy variables.

Figure 5 shows a simplified view of the variable names of the data set Ds in this embodiment.

For ease of understanding, the variable names "Voltage", "Speed", "Resistance", "Worker", "Facility No.", "Material Mixture", "Material Temperature Difference", "Auxiliary Voltage", and "Jig Temperature" of the data set Ds shown in Figures 3A to 3C are replaced with the variable names "X0", "X1", "Y", "Z0", "Z1", "D1", "D2", "D3", and "D4" as shown in Figure 5. In addition, the variable data "Suzuki" and "Sato" identified by the variable name "Worker" are replaced with "A" and "B", and the variable data "Unit C" and "Unit D" identified by the variable name "Facility No." are replaced with "C" and "D".

In the following, when it is necessary to identify each variable, the variable is identified by adding the variable name after the variable, such as variable X1. In the example shown in FIG. 5, variables X0, X1, and Y are quantitative variables. Variables Z0 and Z1 are qualitative variables, and variables D1, D2, D3, and D4 are quantitative variables.

[Functional configuration]
FIG. 6 is a block diagram showing the functional configuration of the arithmetic circuit 102. As shown in FIG.

The arithmetic circuit 102 executes the program 105a to realize a plurality of functions for generating a model. Specifically, the arithmetic circuit 102 includes a receiving unit (receiving means) 130, a first variable identification unit (variable identification means) 121, a stratification condition setting unit 122, a non-utilized variable extraction unit 123, a variable type determination unit 124, an influence calculation unit (influence calculation means) 125, a second variable identification unit (variable identification means) 126, a stratification unit (stratification means) 127, a generation unit (generation means) 128, and a result output unit (result output means) 129. The influence calculation unit 125 also includes a qualitative influence calculation unit 125a and a quantitative influence calculation unit 125b. These components are realized by the execution of the program 105a by the arithmetic circuit 102.

The receiving unit 130 receives a data set Ds including three or more variables. For example, the receiving unit 130 acquires the data set Ds by reading the data set Ds from the database 106. Then, the user performs an input operation on the input unit 101 to select an explanatory variable and a target variable from among the multiple variables of the data set Ds shown in FIG. 5. The first variable identification unit 121 identifies, for example, the variable X0 and the variable X1 as explanatory variables from among the multiple variables of the data set Ds shown in FIG. 5 in response to the user's input operation accepted by the input unit 101. Furthermore, the first variable identification unit 121 identifies, for example, the variable Y as a target variable from among the multiple variables. As a result, two variables are set as explanatory variables and one variable is set as a target variable.

In this manner, the first variable identification unit 121 in this embodiment identifies one or more objective variables and one or more explanatory variables from the dataset Ds. Note that, in this embodiment, two explanatory variables and one objective variable are set, but the respective numbers of the explanatory variables and objective variables are not limited to these examples and may be any number. For example, the first variable identification unit 121 may set one variable of the multiple variables of the dataset Ds as the explanatory variable, or may set each of three or more variables as the explanatory variable. In addition, the first variable identification unit 121 may identify each of two or more variables of the multiple variables of the dataset Ds as the objective variable.

The first variable identification unit 121 also stores in the memory 103 or the storage unit 105 first setting information indicating that the variables set as explanatory variables are the variable X0 and the variable X1, and that the variable set as the objective variable is the variable Y. When the first setting information is stored in the storage unit 105, the first setting information may be stored as temporary data 105b. The first variable identification unit 121 may also output the first setting information to the non-utilized variable extraction unit 123.

The stratification condition setting unit 122 sets the total number of non-utilized variables used for stratification of the data set Ds to N (N is an integer of 2 or more) in response to the input operation of the user received by the input unit 101. Note that N is the total number of non-utilized variables used for stratification, but in this embodiment, an example in which the total number N is 2 or more will be described. The total number N is not limited to 2 or more as in this embodiment, and may be 1. In addition, in this embodiment, as a specific example, the total number N is set to N=2. Note that the non-utilized variables used for stratification are also referred to as stratification variables hereinafter. In other words, the stratification condition setting unit 122 in this embodiment sets the total number of stratification variables in response to the input operation by the user. Note that, for the above-mentioned stratification, each data of the stratification variables included in the data set Ds needs to be classified into not only one group but multiple groups according to the commonality or similarity between the data.

Furthermore, the stratification condition setting unit 122 determines the priority order to be set for the stratification variables for each of the qualitative variables and the quantitative variables in response to the user's input operation accepted by the input unit 101. For example, the stratification condition setting unit 122 determines the priority order of the qualitative variables and the quantitative variables in that order. In this case, the qualitative variables have a higher priority than the quantitative variables. In this manner, the stratification condition setting unit 122 in this embodiment sets the priority order of the qualitative variables and the quantitative variables in response to the user's input operation. In other words, the stratification condition setting unit 122 receives priority information indicating whether the qualitative variables or the quantitative variables are to be prioritized in response to the user's input operation accepted by the input unit 101. In this embodiment, the stratification condition setting unit 122 receives the priority information, but the receiving unit 130 may receive the priority information. Hereinafter, the qualitative variables and the quantitative variables with a higher priority are also called upper variables, and the variables with a lower priority are also called lower variables. The stratification condition setting unit 122 stores second setting information indicating the total number N of stratification variables and the priority order in the memory 103 or the storage unit 105. When the second setting information is stored in the storage unit 105, the second setting information may be stored as temporary data 105b. In addition, the stratification condition setting unit 122 may output the second setting information to the second variable identification unit 126.

The non-utilized variable extraction unit 123 extracts M non-utilized variables from the multiple variables of the data set Ds shown in FIG. 5 read by the first variable identification unit 121. Specifically, the non-utilized variable extraction unit 123 acquires the first setting information from the first variable identification unit 121, the memory 103, or the storage unit 105. Then, the non-utilized variable extraction unit 123 extracts all variables other than the explanatory variables and the objective variable indicated by the first setting information as non-utilized variables from the multiple variables. For example, the non-utilized variable extraction unit 123 extracts the variables Z0, Z1, D1, D2, D3, and D4 as non-utilized variables from the multiple variables of the data set Ds. As a result, in this embodiment, M non-utilized variables are extracted from the multiple variables indicated by the data set Ds. Then, the non-utilized variable extraction unit 123 stores extraction information indicating the extracted M non-utilized variables in the memory 103 or the storage unit 105. When the extracted information is stored in the storage unit 105, the extracted information may be stored as temporary data 105b. In addition, the non-utilized variable extraction unit 123 may output the extracted information to the variable type determination unit 124.

The variable type determination unit 124 acquires the extraction information from the non-utilized variable extraction unit 123, the memory 103, or the storage unit 105, and sequentially determines the variable type of each of the M non-utilized variables indicated by the extraction information. There are two types of variables: the qualitative variable type and the quantitative variable type. That is, the variable type determination unit 124 determines whether the non-utilized variable is a qualitative variable or a quantitative variable based on the data of the non-utilized variable. Specifically, if the data of the non-utilized variable contains characters, the variable type determination unit 124 determines that the non-utilized variable is a qualitative variable. On the other hand, if the data of the non-utilized variable does not contain characters but contains only numbers, the variable type determination unit 124 determines that the non-utilized variable is a quantitative variable. As a result, each of the M non-utilized variables is classified into a qualitative variable and a quantitative variable. For example, in this embodiment, the variable type determination unit 124 determines that the non-utilized variables Z0 and Z1 are each qualitative variables, and that the non-utilized variables D1, D2, D3, and D4 are each quantitative variables. Then, the variable type determination unit 124 stores variable type information indicating the variable type of each of the M non-utilized variables in the memory 103 or the storage unit 105. When the variable type information is stored in the storage unit 105, the variable type information may be stored as temporary data 105b. The variable type determination unit 124 may also output the variable type information to the impact calculation unit 125.

The influence calculation unit 125 calculates the influence of each of the three or more variables included in the dataset Ds, other than the identified objective variable and explanatory variables, on the objective variable. In other words, the influence calculation unit 125 calculates the influence of each of the M non-utilized variables. That is, the influence calculation unit 125 calculates the influence of each of the M non-utilized variables, other than the identified objective variable and explanatory variables, included in the three or more variables of the dataset Ds, on the objective variable data, that is, the influence of the data of the non-utilized variables on the objective variable data. The influence is a numerical value indicating the magnitude of the influence that the data of the non-utilized variables has on the objective variable data, and the greater the influence, the larger the value indicated.

Specifically, the influence calculation unit 125 includes a qualitative influence calculation unit 125a and a quantitative influence calculation unit 125b, as described above. The qualitative influence calculation unit 125a calculates the influence of the non-utilized variable, which is a qualitative variable, and the quantitative influence calculation unit 125b calculates the influence of the non-utilized variable, which is a quantitative variable. More specifically, the qualitative influence calculation unit 125a acquires variable type information from the variable type determination unit 124, the memory 103, or the storage unit 105, and identifies the variable type of the non-utilized variable indicated in the variable type information. Then, if the variable type of the non-utilized variable is a qualitative variable, the qualitative influence calculation unit 125a calculates the influence of the non-utilized variable by an influence calculation process corresponding to the qualitative variable. On the other hand, the quantitative influence calculation unit 125b identifies the variable type of the non-utilized variable indicated in the variable type information. If the variable type of the non-utilized variable is a quantitative variable, the quantitative influence calculation unit 125b calculates the influence of the non-utilized variable by an influence calculation process corresponding to the quantitative variable.

Then, the impact calculation unit 125 stores impact information indicating the impact calculated for each of the M non-utilized variables in the memory 103 or the storage unit 105. When the impact information is stored in the storage unit 105, the impact information may be stored as temporary data 105b. In addition, the impact calculation unit 125 may output the impact information to the second variable identification unit 126.

The second variable identification unit 126 acquires the impact information of each of the M non-utilized variables from the impact calculation unit 125, the memory 103, or the storage unit 105. Furthermore, the second variable identification unit 126 acquires the second setting information from the stratification condition setting unit 122, the memory 103, or the storage unit 105. Then, the second variable identification unit 126 uses the impact information and the second setting information to identify N non-utilized variables as stratification variables from among the M non-utilized variables indicated by the data set Ds. The stratification variables are variables used to stratify the records of the data set Ds. For example, the second variable identification unit 126 attempts to identify N top variables in order of the greatest impact from all top variables in the data set Ds. If the N top variables are identified, the second variable identification unit 126 treats the N top variables as N stratification variables. On the other hand, if the number of all top variables in the data set Ds is less than N and the N top variables cannot be identified, the second variable identification unit 126 identifies the remaining number of lower-level variables from all the lower-level variables in the data set Ds in order of the degree of influence. This identifies N stratification variables that combine the less-than-N top variables and the remaining number of lower-level variables.

In this manner, the second variable identification unit 126 in this embodiment identifies, from among multiple variables, a non-utilized variable that is different from the objective variable and explanatory variables and is not utilized in the model as a stratification variable used for stratifying records. That is, the second variable identification unit 126 identifies one or more variables as stratification variables from among the three or more variables included in the data set Ds, other than the identified objective variable and explanatory variables. In this embodiment, the second variable identification unit 126 identifies N stratification variables (N is an integer of 2 or more) from among multiple variables. That is, the second variable identification unit 126 identifies only N stratification variables, which is the total number set by the stratification condition setting unit 122. At this time, the second variable identification unit 126 identifies the stratification variables based on the degree of influence. Specifically, the second variable identification unit 126 identifies N stratification variables from the M non-utilized variables based on the degree of influence of each of the M non-utilized variables calculated by the influence calculation unit 125. More specifically, the second variable identification unit 126 identifies stratification variables based on the influence and priority information. That is, the second variable identification unit 126 treats the qualitative and quantitative variables with higher priority as higher-order variables, and treats the variables with lower priority as lower-order variables. The second variable identification unit 126 then identifies the higher-order variables as stratification variables in descending order of influence from the M non-utilized variables until all K higher-order variables (K is an integer less than N) have been identified, or until N higher-order variables have been identified. When the number of identified higher-order variables is K, the second variable identification unit 126 then identifies (N-K) lower-order variables as stratification variables in descending order of influence from the M non-utilized variables.

Then, the second variable identification unit 126 stores stratification variable information indicating the identified N stratification variables in the memory 103 or the storage unit 105. When the stratification variable information is stored in the storage unit 105, the stratification variable information may be stored as temporary data 105b. In addition, the second variable identification unit 126 may output the stratification variable information to the stratification unit 127.

The stratification unit 127 performs stratification based on the commonality or similarity between two or more data of the stratification variables included in the data set Ds. In this stratification, the stratification unit 127 performs stratification, which is a process of including one or more records in each of the multiple layers by classifying two or more records included in the data set Ds into multiple layers. That is, the stratification unit 127 classifies the data set Ds into multiple layers based on the tendency of the relationship between the stratification variables and the objective variable. Specifically, the stratification unit 127 classifies the data of the stratification variables into multiple groups for each stratification variable based on the identity or similarity of the data of the stratification variable, and classifies the data set Ds for each combination of the multiple groups. Here, the stratification variables are indicated by the stratification variable information described above. Therefore, the stratification unit 127 acquires the stratification variable information from the second variable identification unit 126, the memory 103, or the storage unit 105. Then, the stratification unit 127 performs stratification for the data set Ds based on the stratification variable information.

Specifically, for each of the N stratification variables indicated by the stratification variable information, the stratification unit 127 classifies two or more data of the stratification variables included in the dataset Ds into a plurality of groups based on the commonality or similarity between the two or more data of the stratification variables. Then, the stratification unit 127 determines a plurality of strata according to the combination of the groups of the N stratification variables, and classifies two or more records included in the dataset Ds into the determined plurality of strata. As a result, a stratification dataset consisting of one or more records is generated for each of the plurality of strata. The stratification dataset for each of the plurality of strata includes one or more records classified from the dataset Ds. Each of the one or more records includes data belonging to the same group for each of the N stratification variables. Furthermore, each of the one or more records includes data for each of the objective variables and explanatory variables. Details of this stratification dataset will be described later with reference to FIG. 7.

The generation unit 128 generates a model showing the relationship between one or more objective variables and one or more explanatory variables for each of the multiple layers. That is, the generation unit 128 generates a model showing the relationship between the explanatory variable data and the objective variable data for each of the multiple layers, using one or more records included in that layer, i.e., a stratified data set. Here, in the above example, the variable X0 and the variable X1 are each an explanatory variable, but the number of explanatory variables may be one or more. Therefore, in this case, the generation unit 128 generates a model showing the relationship between the data of each of the two or more explanatory variables and the objective variable data as a multiple regression equation. For example, the generation unit 128 generates a model showing the relationship between the data of each of the explanatory variables X0 and the explanatory variable X1 and the objective variable Y by performing a multiple regression analysis on the explanatory variables X0 and the explanatory variables X1 and the objective variable Y.

The result output unit 129 outputs the multiple models that have been generated. In other words, the result output unit 129 obtains the models generated for each layer by the generation unit 128 from the generation unit 128 and outputs them to the output unit 104.

[Stratified Dataset]
FIG. 7 is a diagram showing an example of a stratified data set in this embodiment.

For example, the second variable identification unit 126 identifies the non-utilized variable Z0 and the non-utilized variable Z1, which are qualitative variables, as stratification variables. The data of the stratification variable Z0 included in each record of the data set Ds shown in FIG. 5 indicates "A" or "B". The data of the stratification variable Z1 included in each record of the data set Ds indicates "C" or "D". Therefore, as shown in FIG. 7(a), the stratification unit 127 classifies each record including the data "A" of the stratification variable Z0 and the data "C" of the stratification variable Z1 into the first layer. This generates the stratification data set Ds1. The stratification data set Ds1 consists of a record identified by the ID "ID200901", a record identified by the ID "ID200902", and a record identified by the ID "ID200903".

Similarly, as shown in FIG. 7B, the stratification unit 127 classifies each record that includes data "B" of the stratification variable Z0 and data "C" of the stratification variable Z1 into the second tier. This generates a stratification dataset Ds2. The stratification dataset Ds2 consists of a record identified by the ID "ID200904", a record identified by the ID "ID200905", and a record identified by the ID "ID200906".

Similarly, as shown in FIG. 7C, the stratification unit 127 classifies each record that includes data "A" of the stratification variable Z0 and data "D" of the stratification variable Z1 into the third tier. This generates a stratification dataset Ds3. The stratification dataset Ds3 consists of a record identified by the ID "ID200907", a record identified by the ID "ID200908", and a record identified by the ID "ID200909".

Similarly, as shown in (d) of FIG. 7, the stratification unit 127 classifies each record that includes data "B" of the stratification variable Z0 and data "D" of the stratification variable Z1 into the fourth tier. This generates a stratification dataset Ds4. The stratification dataset Ds4 consists of a record identified by the ID "ID200910", a record identified by the ID "ID200911", and a record identified by the ID "ID200912".

In other words, two or more data of the stratification variable Z0 are classified into groups "A" and "B", and two or more data of the stratification variable Z1 are classified into groups "C" and "D". The first layer corresponds to the combination of group "A" of the stratification variable Z0 and group "C" of the stratification variable Z1. The second layer corresponds to the combination of group "B" of the stratification variable Z0 and group "C" of the stratification variable Z1. The third layer corresponds to the combination of group "A" of the stratification variable Z0 and group "D" of the stratification variable Z1. The fourth layer corresponds to the combination of group "B" of the stratification variable Z0 and group "D" of the stratification variable Z1. In this way, multiple layers are determined according to the combination of the groups of the stratification variables Z0 and Z1. Therefore, in the stratification classification, a record including data belonging to group "A" of the stratification variable Z0 and data belonging to group "C" of the stratification variable Z1 is classified into the first layer. Records that contain data belonging to group "B" of stratification variable Z0 and data belonging to group "C" of stratification variable Z1 are classified into the second tier. Records that contain data belonging to group "A" of stratification variable Z0 and data belonging to group "D" of stratification variable Z1 are classified into the third tier. Records that contain data belonging to group "B" of stratification variable Z0 and data belonging to group "D" of stratification variable Z1 are classified into the fourth tier.

Note that records classified into each stratum may contain data on non-utilized variables other than the stratification variables, or may contain only data on the utilization variables and the stratification variables, as in the example shown in Figure 7.

Figure 8 shows the distribution of coordinate points indicated by each record contained in each stratified dataset Ds1 to Ds4.

Each of the stratified data sets Ds1 to Ds4 includes multiple records. Each of the multiple records includes data for explanatory variable X0, explanatory variable X1, and objective variable Y, and is shown as a coordinate point (X0, X1, Y). In other words, a record is shown as a coordinate point in a three-dimensional coordinate system consisting of explanatory variable X0, explanatory variable X1, and objective variable Y.

It is difficult to find correlations between records from the coordinate points of all the records contained in dataset Ds. However, as shown in FIG. 8, in each of stratified datasets Ds1 to Ds4, the coordinate points of multiple records contained in that stratified dataset are mutually correlated. Therefore, in each of stratified datasets Ds1 to Ds4, it is possible to find correlations between those records from the coordinate points of all the records contained in that stratified dataset.

For each of these stratified data sets Ds1 to D4, the generation unit 128 uses one or more records contained in the stratified data set to generate a model showing the relationship between the data for each of the explanatory variables X0 and X1 and the data for the objective variable Y.

In this way, in this embodiment, the stratification variables are non-utilized variables, and stratification is performed according to the non-utilized variables, and a model is generated for each of the multiple layers. Therefore, the non-utilized variables, which are variables other than the explanatory variables, can be used to optimally perform stratification for the data set Ds. As a result, even if the correlation between the explanatory variables and the objective variable changes according to the non-utilized variables, a highly accurate model can be generated according to the group of the non-utilized variables. In other words, the accuracy of the model can be easily improved. Furthermore, in this embodiment, even if there are two or more stratification variables, optimal stratification can be performed for the data set Ds, and a highly accurate model can be generated for each of the multiple layers according to the data of the stratification variables, i.e., the N non-utilized variables.

[Processing Operation]
FIG. 9 is a flowchart showing an example of the overall processing operation of the model generating device 100 in this embodiment.

First, the receiving unit 130 of the model generating device 100 performs a data receiving process (step S1). In this data receiving process, the first variable identifying unit 121 receives the data set Ds by reading the data set Ds from the database 106. Then, the first variable identifying unit 121 identifies an explanatory variable and a target variable from the multiple variables indicated by the data set Ds (step S2). In this way, the explanatory variable and the target variable are set. For example, as described above, the variables X0 and X1 are set as explanatory variables, and the variable Y is set as the target variable.

Next, the stratification condition setting unit 122 sets the total number N of stratification variables in response to the user's input operation (step S3). For example, the total number N is set to N=2. Then, the stratification condition setting unit 122 determines the priority of the variable types of the stratification variables (step S4). That is, the priority of qualitative variables and quantitative variables is determined. For example, the priority of these variables is determined in the order of qualitative variables, then quantitative variables.

Next, the non-utilized variable extraction unit 123 extracts variables other than the explanatory variables and the objective variable as non-utilized variables from the multiple variables in the dataset Ds (step S5).

Then, the model generating device 100 executes the first loop process including steps S6 to S9 for each of all the non-utilized variables extracted in step S5 in turn. That is, the first loop process is executed for each of the M non-utilized variables shown in the data set Ds in turn.

Specifically, first, the variable type determination unit 124 determines the variable type of the non-utilized variable to be processed (step S6). Then, the variable type determination unit 124 determines whether the variable type is a qualitative variable type (step S7). That is, the variable type determination unit 124 determines whether the non-utilized variable to be processed is a qualitative variable. Then, when the variable type determination unit 124 determines that the non-utilized variable to be processed is a qualitative variable (Yes in step S7), the qualitative influence calculation unit 125a calculates the influence of the non-utilized variable, which is a qualitative variable, on the objective variable (step S8). That is, the qualitative influence calculation unit 125a calculates the influence of the non-utilized variable on the objective variable by the influence calculation process of the qualitative variable. On the other hand, when the variable type determination unit 124 determines that the non-utilized variable to be processed is not a qualitative variable (No in step S7), the quantitative influence calculation unit 125b calculates the influence of the non-utilized variable, which is a quantitative variable, on the objective variable (step S9). In other words, the quantitative influence calculation unit 125b calculates the influence of the non-utilized variable on the objective variable through the influence calculation process of the quantitative variable.

The first loop process including steps S6 to S9 is executed for each of the non-utilized variables in turn, thereby calculating the degree of influence of each of the non-utilized variables.

Then, the second variable identification unit 126 sorts all non-utilized variables belonging to each variable type in order of their influence based on the influence of each of the non-utilized variables (step S10). That is, the second variable identification unit 126 rearranges all the extracted qualitative variables in order of their influence. Specifically, the second variable identification unit 126 rearranges the qualitative variables such that the greater the influence of the qualitative variable, the earlier the qualitative variable is placed. Similarly, the second variable identification unit 126 rearranges all the extracted quantitative variables in order of their influence. Specifically, the second variable identification unit 126 rearranges the quantitative variables such that the greater the influence of the quantitative variable, the earlier the quantitative variable is placed.

Next, the second variable identification unit 126 identifies N non-utilized variables as stratification variables from the sorted M non-utilized variables based on the total number N of stratification variables set in step S3 and the priority order determined in step S4 (step S11). If the qualitative variables have a higher priority than the quantitative variables, the second variable identification unit 126 attempts to identify up to N qualitative variables in descending order of influence from the sorted multiple qualitative variables. If the number of extracted qualitative variables is less than N, the second variable identification unit 126 identifies quantitative variables in descending order of influence from the sorted multiple quantitative variables. The number of quantitative variables identified at this time is the number obtained by subtracting the number of previously identified qualitative variables from N. As a result, a total of N stratification variables are identified. In the above example, N=2, and the qualitative variables have a higher priority than the quantitative variables. Therefore, in this case, the qualitative variable Z0 and the qualitative variable Z1 are identified as stratification variables.

Next, the stratification unit 127 generates a plurality of stratified datasets by performing stratification classification on the dataset Ds using the identified N stratification variables. For example, as shown in FIG. 7, stratification datasets Ds1 to Ds4 are generated. Then, the generation unit 128 calculates a multiple regression equation by performing a multiple regression analysis on the explanatory variables and the objective variable for each stratification dataset (step S12). As a result, a model consisting of a multiple regression equation is generated for each stratification dataset.

The generation unit 128 further calculates the coefficient of determination adjusted for the degrees of freedom of the explanatory variables for the multiple regression equation calculated for each of the multiple stratified data sets (step S13).

The result output unit 129 outputs each multiple regression equation calculated in step S12 and each coefficient of determination calculated in step S13 to the output unit 104. As a result, the output unit 104 displays each multiple regression equation and each coefficient of determination on a display, prints them on paper, or stores a file showing them in the memory unit 105 (step S14).

Figure 10 is a flowchart showing a specific example of the process of calculating the influence of qualitative variables in step S8 of Figure 9. Note that the non-utilized variables to be processed in this influence calculation process are qualitative variables.

The quality influence calculation unit 125a determines whether the number of categories of the non-utilized variable to be processed is equal to or less than a first threshold (step S81). The first threshold for the number of categories is, for example, 20. The number of categories is the number of groups consisting of multiple identical data represented by the non-utilized variable to be processed, which are included in the data set Ds. For example, in the data set Ds shown in FIG. 5, the multiple data represented by the non-utilized variable Z0, which is a qualitative variable, includes data representing "A" and data representing "B". Therefore, the number of categories of the non-utilized variable Z0 is 2. Similarly, in the data set Ds shown in FIG. 5, the multiple data represented by the non-utilized variable Z1, which is a qualitative variable, includes data representing "C" and data representing "D". Therefore, the number of categories of the non-utilized variable Z1 is 2.

Next, when the qualitative influence calculation unit 125a determines that the number of categories of the non-utilized variable to be processed is not equal to or less than the first threshold (No in step S81), it excludes the non-utilized variable from the calculation of the influence (step S82). For example, if a non-utilized variable with a relatively large number of categories is used as a stratification variable, many stratification data sets are generated. As a result, it is assumed that it is difficult to expect improvement in the accuracy of each model and improvement in the ease of use of the models due to the generation of many models. Therefore, in step S82, by excluding such a non-utilized variable with a large number of categories from the calculation of the influence, it is possible to suppress the non-utilized variable from being used as a stratification variable.

On the other hand, when the quality influence calculation unit 125a determines that the number of categories of the non-utilized variable to be processed is equal to or less than the first threshold (Yes in step S81), it calculates the influence of the non-utilized variable to be processed by supervised machine learning (step S83). The supervised machine learning is, for example, learning using a random forest. A random forest is a method using multiple decision trees. For example, the quality influence calculation unit 125a executes machine learning of a random forest using each data of the objective variable included in the data set Ds and each data of the non-utilized variable to be processed included in the data set Ds as teacher data. This random forest is, for example, a learning model for estimating the data of the non-utilized variable to be processed from the data of the objective variable. More specifically, the non-utilized variable to be processed is the non-utilized variable Z0. In this case, the quality influence calculation unit 125a executes machine learning so that the data of the objective variable is input to the random forest, and the data of the non-utilized variable Z0 corresponding to the data of the objective variable is output from the random forest. At this time, the data of the non-utilized variable Z0 output from the random forest is "A" or "B".

The qualitative influence calculation unit 125a calculates the influence of the non-utilized variables to be processed based on the Gini coefficient G, which is an index representing the impurity of multiple decision trees included in the random forest. The Gini coefficient G is defined by the formula (1) for each node of the decision tree.

Here, in formula (1), C is the number of categories. Furthermore, Pi is the number of data belonging to category i divided by the total number of data. In other words, Pi is the quotient obtained by dividing the number of data of category i classified in the node corresponding to the Gini coefficient G by the total number of data classified in the node. For example, when two data indicating "A" and one data indicating "B" are classified in the node, G=1-(2/3) ² -(1/3) ² .

The quality influence calculation unit 125a performs learning so that the Gini coefficient in the decision tree is as small as possible. The quality influence calculation unit 125a then calculates an influence that indicates a larger value as the average value of all the Gini coefficients of the multiple decision trees used in the random forest is smaller. For example, the quality influence calculation unit 125a calculates the inverse of the average value as the influence.

In this manner, the qualitative influence calculation unit 125a in this embodiment uses a random forest to calculate the influence of each of one or more qualitative variables. Therefore, even if the non-utilized variable is a qualitative variable, the influence of the non-utilized variable on the objective variable can be appropriately calculated.

Figure 11 is a flowchart showing a specific example of the process of calculating the impact of quantitative variables in step S9 of Figure 9. Note that the non-utilized variables to be processed in this impact calculation process are quantitative variables.

The quantitative influence calculation unit 125b performs clustering on each data of the non-utilized variables to be processed that are included in the data set Ds by unsupervised machine learning (step S91). The unsupervised machine learning is, for example, a Gaussian Mixture Model (GMM).

The Gaussian mixture model is a method for approximating a given probability distribution with a linear combination of multiple Gaussian functions (i.e., normal distributions). In linear combination, multiple Gaussian functions are combined using a weight πk. The weight πk is the weight of the kth Gaussian function and is also called a mixing coefficient. When considered in two dimensions of (a, b), the kth Gaussian function has the mean value μa_k of a, the mean value μb_k of b, the variance Σa_k of a, the variance Σb_k of b, and the covariance Σab_k of a and b. The magnitude of each normal distribution can be simply handled as Σb_k + Σa_k. In this embodiment, (a, b) are (objective variables, non-utilized variables that are quantitative variables).

The quantitative influence calculation unit 125b analyzes the Gaussian mixture model while changing the number of clusters, which is a hyperparameter in the Gaussian mixture model. The quantitative influence calculation unit 125b then adopts the number of clusters that minimizes, for example, the Akaike's Information Criterion (AIC) or the Bayesian Information Criterion (BIC). This generates as many clusters as the number of clusters. The number of clusters is one or more. The clusters correspond to the above-mentioned categories or groups.

Then, the quantitative influence calculation unit 125b executes a second loop process including steps S92 to S94 for each of all the clusters generated in step S91 in order.

Specifically, the quantitative influence calculation unit 125b first determines whether the number of data in the cluster to be processed is equal to or greater than the second threshold (step S92). Here, if the quantitative influence calculation unit 125b determines that the number of data is equal to or greater than the second threshold (Yes in step S92), it calculates the influence of the data in the cluster on the data of the objective variable as the cluster influence (step S93). At this time, the quantitative influence calculation unit 125b adopts the cluster as a cluster for which calculation processing has been completed. For example, the quantitative influence calculation unit 125b calculates the regression coefficient of the simple regression equation as the cluster influence by performing a simple regression analysis on the cluster to be processed and the objective variable. On the other hand, if the quantitative influence calculation unit 125b determines that the number of data is less than the second threshold (No in step S92), it excludes the cluster from the calculation target of the cluster influence (step S94). This makes it possible to prevent the influence of the non-utilized variable to be processed from being calculated using a cluster with low reliability.

The second loop process including steps S92 to S94 is executed in order for each of all the clusters generated in step 91. As a result, the cluster influence of each of the remaining clusters, other than the cluster excluded in step S94, is calculated.

Then, the quantitative influence calculation unit 125b judges whether the number of clusters for which the cluster influences have been calculated in step S93, i.e., the number of clusters for which the calculation process has been completed, is two or more (step S95). Here, if the quantitative influence calculation unit 125b judges that the number of clusters is two or more (Yes in step S95), it selects the maximum influence of the two or more influences calculated in the second loop process (step S97). That is, the quantitative influence calculation unit 125b selects the maximum influence as the influence of the non-utilized variable to be processed on the objective variable. On the other hand, if the quantitative influence calculation unit 125b judges that the number of clusters for which the calculation process has been completed is not two or more (No in step S95), it excludes the non-utilized variable from the calculation target of the influence (step S96). This makes it possible to prevent a non-utilized variable with a low correlation with the objective variable from being identified as a stratification variable.

In this manner, the quantitative influence calculation unit 125b in this embodiment uses a Gaussian mixture model to perform clustering on two or more data of one or more quantitative variables included in the data set Ds, and calculates the influence of the quantitative variable by a simple regression analysis between each cluster obtained by clustering and the objective variable. Therefore, even if the non-utilized variable is a quantitative variable, the influence of the non-utilized variable on the objective variable can be appropriately calculated.

As described above, in this embodiment, when the stratification variables are qualitative variables, identical data are grouped together, and when the stratification variables are quantitative variables, identical or similar data are grouped together.

In this embodiment, the stratification unit 127 classifies the data of each stratification variable into a plurality of groups based on the identity or similarity of the data of the stratification variable, and classifies the data set Ds for each combination of the plurality of groups. Here, when the stratification variables are quantitative variables, the data of the quantitative variables being similar means that the tendency of the influence of the quantitative variables on the objective variable is similar. Therefore, it can be said that the stratification unit 127 classifies the data set Ds into a plurality of strata based on the tendency of the relationship between the stratification variables and the objective variable.

In addition, in a Gaussian mixture model, each piece of data has a probability of belonging to each group, and of the multiple groups, the data belongs to the group with the highest probability. When the k-means method described below is used instead of a Gaussian mixture model, each piece of data has a distance from the center of gravity of each group, and belongs to the group with the closest distance among the multiple groups. Therefore, specifically, data of a quantitative variable is said to be similar when the probability corresponding to that data is equal to or greater than a certain value, or when the distance from the center of gravity of the group to the position corresponding to that data is equal to or less than a certain value.

[Model example]
As described above, in this embodiment, stratified classification is performed on the data set Ds. For example, when the non-utilized variable Z0 and the non-utilized variable Z1 are each identified as a stratification variable, four stratified data sets Ds1 to Ds4 are generated as shown in FIG. 7. Then, a model is generated from each of the four stratified data sets Ds1 to Ds4. This makes it possible to improve the accuracy of the model.

Specifically, when stratified classification is not performed, the model generated from the data set Ds is expressed as shown in the following formula (2). In formula (2), _x0 and _x1 correspond to the explanatory variables X0 and X1, respectively, and f corresponds to the objective variable Y.

On the other hand, in this embodiment, since stratified classification is performed as described above, four models shown in the following formulas (3) to (6) are generated as multiple regression equations. In formulas (3) to (6), x ₀ and x ₁ correspond to the explanatory variables X0 and X1, respectively, and f ₀₀ , f ₀₁ , f ₁₀ , and f ₁₁ correspond to the objective variable Y. Specifically, formula (3) is a model generated from the stratified data set Ds4 shown in FIG. 7(d) and FIG. 8, and the stratified data set Ds4 includes records including the data "B" of the stratified variable Z0 and the data "D" of the stratified variable Z1. f ₀₀ in formula (3) corresponds to the objective variable Y of this stratified data set Ds4. Equation (4) is a model generated from the stratified data set Ds3 shown in FIG. 7(c) and FIG. 8, and the stratified data set Ds3 includes each record including the data "A" of the stratified variable Z0 and the data "D" of the stratified variable Z1. f ₀₁ in Equation (4) corresponds to the objective variable Y of the stratified data set Ds3. Equation (5) is a model generated from the stratified data set Ds2 shown in FIG. 7(b) and FIG. 8, and the stratified data set Ds2 includes each record including the data "B" of the stratified variable Z0 and the data "C" of the stratified variable Z1. f ₁₀ in Equation (5) corresponds to the objective variable Y of the stratified data set Ds2. Equation (6) is a model generated from the stratified data set Ds1 shown in FIG. 7(a) and FIG. 8, and the stratified data set Ds1 includes each record including the data "A" of the stratified variable Z0 and the data "C" of the stratified variable Z1. f ₁₁ in equation (6) corresponds to the objective variable Y of this stratified data set Ds1.

In this embodiment, two stratification variables Z0 and Z1 are specified, and each data of the stratification variable Z0 is classified into two groups, and each data of the stratification variable Z1 is classified into two groups. Therefore, the number of group combinations is four, and four models are generated. Here, if three stratification variables are specified and each data of these stratification variables is classified into two groups, the number of group combinations is eight, and eight models are generated. Also, if two stratification variables are specified and each data of these stratification variables is classified into three groups, the number of group combinations is nine, and nine models are generated.

Here, R2 ^* is the coefficient of determination after the degrees of freedom adjustment. This coefficient of determination after the degrees of freedom adjustment is an index indicating the reliability of the model. In this embodiment, as shown in the above formulas (2) to (6), the coefficient of determination after the degrees of freedom adjustment can be increased from 0.273 to a range of 0.503 to 0.969, thereby improving the accuracy of the model. Then, each model and each coefficient of determination after the degrees of freedom adjustment are output by the result output unit 129.

In this manner, the generation unit 128 in this embodiment calculates an index indicating the likelihood of each of the multiple models generated. Then, the result output unit 129 outputs the calculated index for each of the multiple models. Therefore, the user can easily determine whether or not to use the generated model based on the index.

[Effects, etc.]
As described above, in this embodiment, stratification is performed according to the stratification variables, which are variables other than the objective variable and the explanatory variables, and a model is generated for each of the multiple strata. In addition, the stratification variables are non-utilized variables that are not adopted as variables included in the model but are used to generate the model. Therefore, the non-utilized variables can optimally perform stratification for the data set Ds. As a result, even if the correlation between the explanatory variable and the objective variable changes depending on the non-utilized variables, a highly accurate model can be generated according to the data of the non-utilized variables. In other words, the accuracy of the model can be easily improved.

In addition, in this embodiment, from among the M non-utilized variables in the data set Ds, effective variables are automatically identified as stratification variables. Therefore, it is possible to utilize the intention of a user, such as a factory expert, when selecting the utilized variables (i.e., the objective variable and explanatory variables), and it is possible to generate a model that is easy for the user to understand and to improve the accuracy of the model.

In addition, in this embodiment, the second variable identification unit 126 identifies N stratification variables (N is an integer equal to or greater than 2) from the multiple variables. Then, the stratification unit 127 classifies, for each of the N stratification variables, two or more pieces of data for the stratification variable included in the dataset Ds into multiple groups based on the commonality or similarity between the two or more pieces of data for the stratification variable. Furthermore, the stratification unit 127 determines multiple strata according to the combination of the groups for the N stratification variables, and classifies two or more records included in the dataset Ds into the determined multiple strata. When the stratification variables are qualitative variables, the groups correspond to the above-mentioned categories, and when the stratification variables are quantitative variables, they correspond to the above-mentioned clusters.

As a result, even if there are two or more stratification variables, optimal stratification classification can be performed for the dataset Ds, and a highly accurate model can be generated for each of the multiple strata according to the data for those stratification variables, i.e., the N non-utilized variables.

In addition, in this embodiment, the influence calculation unit 125 calculates the influence of each non-utilized variable, and the second variable identification unit 126 identifies N stratification variables from the non-utilized variables based on the influence of each non-utilized variable.

This makes it possible to identify N stratification variables that have a large influence on the target variable data. As a result, by using these stratification variables, more optimal stratification classification can be performed for the data set Ds. As a result, a more accurate model can be generated for each of the multiple strata.

In addition, in this embodiment, the stratification condition setting unit 122 sets the total number of stratification variables in response to an input operation by the user, and the second variable identification unit 126 identifies N stratification variables, which is the set total number.

This allows the total number of stratification variables to be identified to be set as desired by the user, allowing the number or accuracy of models to be generated to be adjusted.

In addition, in this embodiment, the M non-utilized variables include one or more qualitative variables each representing data that includes characters, and one or more quantitative variables each representing data that consists of numbers.

This allows us to identify N stratification variables that include not just one of qualitative variables and quantitative variables, but both, and increases the degree of freedom in the variable types of the identified stratification variables.

In this embodiment, the stratification condition setting unit 122 sets the priority order of qualitative and quantitative variables in response to an input operation by the user. The second variable identification unit 126 identifies the top variables as stratification variables in descending order of influence from the M non-utilized variables until all K top variables (K is an integer less than N) are identified or until N top variables are identified. Then, when the number of identified top variables is K, (N-K) lower variables are identified as stratification variables in descending order of influence from the M non-utilized variables.

This allows the user to arbitrarily set the priority of qualitative and quantitative variables as intended, and many of the variables with higher priorities can be identified as stratification variables. As a result, appropriate stratification can be performed depending on whether the data set Ds is analyzed qualitatively or quantitatively, and a model can be generated according to the analysis mode.

(Variations, etc.)
Although the model generating device according to one aspect of the present disclosure has been described based on the above embodiment, the present disclosure is not limited to the embodiment. As long as it does not deviate from the spirit of the present disclosure, various modifications conceivable by a person skilled in the art to the above embodiment may also be included in the present disclosure.

For example, in the present embodiment, a random forest is used as an example of supervised machine learning in the influence calculation process of the qualitative variables, but the supervised machine learning is not limited to a random forest, and other supervised machine learning may be used. For example, a gradient boosting decision tree (GBDT) may be used instead of a random forest. When the gradient boosting decision tree is used, machine learning is performed so that the error or loss coefficient is small. Then, the qualitative influence calculation unit 125a calculates an influence that indicates a larger value as the error or loss coefficient is smaller. In addition, a random forest and a gradient boosting decision tree may be combined. For example, the influence of the first non-utilized variable, which is a qualitative variable, may be calculated using a random forest, and the influence of the second non-utilized variable, which is a qualitative variable, may be calculated using a gradient boosting decision tree. At this time, the influences calculated by the two different machine learning methods may be normalized to make those influences comparable.

In addition, in this embodiment, a Gaussian mixture model is used as an example of unsupervised machine learning in the process of calculating the influence of quantitative variables, but the unsupervised machine learning is not limited to the Gaussian mixture model, and other unsupervised machine learning may be used. For example, a k-means method may be used instead of a Gaussian mixture model. In this case, each data represented by the non-utilized variable to be processed is clustered by the k-means method. The Gaussian mixture model and the k-means method may be combined. For example, the clustering of the first non-utilized variable, which is a quantitative variable, may be performed using a Gaussian mixture model, and the clustering of the second non-utilized variable, which is a quantitative variable, may be performed using the k-means method.

In addition, in this embodiment, a multiple regression equation is generated as the model, but a simple regression equation may be generated as the model, or a model other than a regression equation may be generated. For example, a neural network may be generated as the model.

In addition, the data set Ds in this embodiment shows manufacturing-related variables and data on those variables, but is not limited to manufacturing-related variables and may show variables and data on those variables in fields other than manufacturing-related fields.

The data included in the data set Ds in this embodiment may be separated into operation data and quality data. For example, the operation data is data related to the manufacturing process, and may be data for each of the variables X0, X1, Z0, Z1, D1, D2, D3, and D4 shown in FIG. 5. For example, the quality data is data related to the quality of the product, and may be variable Y shown in FIG. 5.

In addition, in this embodiment, the total number N of stratification variables and the priority of the variable types are set, but items other than these may be set. For example, the total number of qualitative variables and the total number of quantitative variables may be set separately. In this embodiment, the total number N is an integer of 2 or more, but it may also be 1.

In addition, in this embodiment, the data set Ds includes data on variables belonging to each of two variable types, but the number of variable types is not limited to two, and may be only one, or may be three or more.

In addition, in this embodiment, the data set Ds is transmitted from the manufacturing management device 500 via a network and stored in the database 106, but it may also be output to the database 106 from another device or recording medium and stored therein. The data set Ds may also be stored in the database 106 without going through a network.

The following cases are also included in this disclosure:

(1) The at least one device is specifically a computer system consisting of a microprocessor, a ROM (Read Only Memory), a RAM (Random Access Memory), a hard disk unit, a display unit, a keyboard, a mouse, etc. A computer program is stored in the RAM or hard disk unit. The at least one device achieves its functions by the microprocessor operating in accordance with the computer program. Here, a computer program is composed of a combination of multiple instruction codes that indicate commands for a computer to achieve a specified function.

(2) Some or all of the components constituting at least one of the above devices may be composed of a single system LSI (Large Scale Integration). A system LSI is an ultra-multifunctional LSI manufactured by integrating multiple components on a single chip, and specifically, is a computer system composed of a microprocessor, ROM, RAM, etc. A computer program is stored in the RAM. The system LSI achieves its functions when the microprocessor operates in accordance with the computer program.

(3) Some or all of the components constituting at least one of the above devices may be composed of an IC card or a standalone module that is detachable from the device. The IC card or module is a computer system composed of a microprocessor, ROM, RAM, etc. The IC card or module may include the above-mentioned ultra-multifunction LSI. The IC card or module achieves its functions when the microprocessor operates according to a computer program. This IC card or module may be tamper-resistant.

(4) The present disclosure may be the methods described above. It may also be a computer program that realizes these methods by a computer, or a digital signal that is a computer program.

The present disclosure may also be a computer program or a digital signal recorded on a computer-readable recording medium, such as a flexible disk, a hard disk, a CD (Compact Disc)-ROM, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray (registered trademark) Disc), a semiconductor memory, or the like. It may also be a digital signal recorded on such a recording medium.

The present disclosure may also involve the transmission of computer programs or digital signals via telecommunications lines, wireless or wired communication lines, networks such as the Internet, data broadcasting, etc.

The program or digital signal may also be implemented by another independent computer system by recording it on a recording medium and transferring it, or by transferring the program or digital signal via a network, etc.

The present disclosure has the effect of easily improving the accuracy of a model, and can be applied, for example, to an apparatus or system that generates a model that estimates the quality of a product manufactured in a manufacturing process from data on variables used in the manufacturing process.

1 Model generation system 100 Model generation device 101 Input unit 102 Arithmetic circuit 103 Memory 104 Output unit 105 Storage unit 105a Program 105b Temporary data 106 Database 121 First variable identification unit 122 Stratification condition setting unit 123 Non-utilized variable extraction unit 124 Variable type determination unit 125 Influence calculation unit 125a Qualitative influence calculation unit 125b Quantitative influence calculation unit 126 Second variable identification unit 127 Stratification unit 128 Generation unit 129 Result output unit 130 Reception unit 500 Manufacturing management device Ds Data set Ds1 to Ds4 Stratified data set

Claims (11)

A model generation device that generates a model indicating a relationship between one or more objective variables and one or more explanatory variables,
receiving means for receiving a data set including three or more variables;
a variable specifying means for specifying one or more dependent variables and one or more explanatory variables from the data set, and specifying one or more variables as stratification variables from among the three or more variables included in the data set, other than the determined dependent variables and the determined explanatory variables;
A stratification means for classifying a data set into a plurality of strata based on a tendency of a relationship between the stratification variable and the objective variable;
A generation means for generating the model for each of the plurality of layers;
A model generating device comprising: The stratification means classifies the data of the stratification variables into a plurality of groups based on identity or similarity of the data of the stratification variables, and classifies the data set for each combination of the plurality of groups.
The model generating device according to claim 1 . The data set includes qualitative variables representing data including characters and quantitative variables representing data consisting of numbers.
3. The model generating device according to claim 1 or 2. The method further includes an influence calculation means for calculating an influence of each of the three or more variables included in the data set on the objective variable, the influence being calculated for each of the variables other than the objective variable and the explanatory variable,
The variable identification means identifies the stratification variables based on the influence degree.
The model generating device according to any one of claims 1 to 3. The method further includes an influence calculation means for calculating an influence of each of the three or more variables included in the data set on the objective variable, the influence being calculated for each of the variables other than the objective variable and the explanatory variable,
the receiving means receives priority information indicating whether the qualitative variables or the quantitative variables are to be prioritized;
The variable identification means identifies the stratification variables based on the influence degree and the priority information.
4. The model generating device of claim 3. The influence degree calculation means
Calculating the influence of the qualitative variables using a random forest or a gradient boosting decision tree.
6. The model generating device according to claim 5. The influence degree calculation means
Using a Gaussian mixture model or a k-means method, clustering is performed on two or more pieces of data of the quantitative variables included in the data set, and the influence degree of the quantitative variables is calculated by a simple regression analysis between each cluster obtained by the clustering and the objective variable.
7. A model generating device according to claim 5 or 6. The generating means further comprises:
Calculating an index indicating the likelihood of each of the generated models;
The model generating device further comprises:
A result output means for outputting the index calculated for each of the plurality of models is provided.
A model generating device according to any one of claims 1 to 7. The generating means includes:
generating the model indicating a relationship between data of each of the two or more explanatory variables and data of the objective variable as a multiple regression equation;
A model generating device according to any one of claims 1 to 8. A model generation method in which a computer generates a model showing a relationship between one or more objective variables and one or more explanatory variables, comprising the steps of:
receiving a dataset containing three or more variables;
Identifying one or more objective variables and one or more explanatory variables from the data set, and identifying one or more variables as stratification variables from among the three or more variables included in the data set, other than the identified objective variables and the identified explanatory variables;
Classifying the data set into a plurality of strata based on a trend of a relationship between the stratification variables and the objective variable;
generating the model for each of the plurality of layers;
Model generation method. A program for causing a computer to generate a model showing a relationship between one or more objective variables and one or more explanatory variables,
receiving a dataset containing three or more variables;
Identifying one or more objective variables and one or more explanatory variables from the data set, and identifying one or more variables as stratification variables from among the three or more variables included in the data set, other than the identified objective variables and the identified explanatory variables;
Classifying the data set into a plurality of strata based on a trend of a relationship between the stratification variables and the objective variable;
generating the model for each of the plurality of layers;
A program that causes a computer to do something.

Images