Information processing apparatus, information processing method, and program

Abstract

To more suitably express a causal relation between multiple variables.SOLUTION: The causal relationship estimation unit estimates a causal relationship between multiple variables, and the expression model generation unit generates an expression model according to a type of a variable included in the causal relationship to be expressed when all or a part of the estimated causal relationship is expressed by the expression model. The present disclosure can be applied to an information processing device that estimates a causal relationship between multiple variables.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a program, and more particularly to an information processing device, an information processing method, and a program that enable more appropriate representation of causal relationships between multiple variables. [Background technology]

[0002] For example, Patent Document 1 discloses a method for estimating the causal relationship between multiple variables based on a test of conditional independence, even when categorical variables and numerical variables are mixed among the multiple variables. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2014-228991 Summary of the Invention [Problem to be solved by the invention]

[0004] Conventionally, when generating a statistical multivariate model from data, it was necessary to specify in advance how the model would be expressed.

[0005] The present disclosure has been made in light of such circumstances, and makes it possible to more appropriately express the causal relationships among multiple variables. [Means for solving the problem]

[0006] The information processing device of the present disclosure is an information processing device that includes a causal relationship estimation unit that estimates causal relationships between multiple variables, and an expression model generation unit that generates an expression model according to the type of variables included in the causal relationship to be expressed when expressing all or part of the estimated causal relationship using an expression model.

[0007] The information processing method disclosed herein is an information processing method that includes estimating a causal relationship between multiple variables, and when expressing all or part of the estimated causal relationship using an expression model, generating the expression model according to the types of variables included in the causal relationship to be expressed.

[0008] The program disclosed herein is a program for causing a computer to execute processing including estimating causal relationships between multiple variables and, when expressing all or part of the estimated causal relationships using an expression model, generating the expression model according to the types of variables included in the causal relationships to be expressed.

[0009] In the present disclosure, when a causal relationship between multiple variables is estimated and all or part of the estimated causal relationship is represented by an expression model, the expression model is generated according to the type of variables included in the causal relationship to be represented. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 2 is a block diagram illustrating an example of a hardware configuration of an information processing device. [Figure 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of an information processing device. [Figure 3] 10 is a flowchart illustrating an operation of the information processing device. [Figure 4] 10 is a flowchart illustrating an example of generating a local representation model. [Figure 5] FIG. 10 is a diagram illustrating an example of a relationship graph. [Figure 6] FIG. 10 is a diagram illustrating a representation of a local representation model. [Figure 7] FIG. 10 is a diagram illustrating a representation of a local representation model. [Figure 8] FIG. 10 is a diagram illustrating an example of a conditional probability table. [Figure 9] FIG. 10 is a diagram illustrating a representation of a local representation model. [Figure 10]10 is a flowchart illustrating an example of generating a global representation model. [Figure 11] FIG. 10 is a diagram illustrating another example of a relationship graph. [Figure 12] 1 is a block diagram illustrating an example of a functional configuration of a network system according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0011] Modes for carrying out the present disclosure (hereinafter referred to as embodiments) will be described below in the following order.

[0012] 1. Conventional technology and its problems 2. Overview of the Technology Disclosed 3. Configuration and operation of information processing device 4. Generating local representation models 5. Generating a global representation model 6.UI example 7.Network System Configuration 8. Application Examples 9. Effects of the Technology According to the Present Disclosure

[0013] <1. Conventional technology and its problems> Conventionally, when generating a multivariate model from data, it was necessary to specify in advance how to represent the model, such as multiple regression models, structural equation models, joint probability representations, conditional probability representations, Markov network models, and Bayesian network models.

[0014] On the other hand, actual data generally consists of a mixture of numerical data treated as continuous random variables and categorical data treated as discrete random variables. The above-mentioned representation models usually assume one of the variable types, so it is necessary to convert the data into either continuous or discrete random variables. In this case, there are problems such as the loss of original information due to variable conversion, difficulty in handling many variables, and excessive calculation time.

[0015] Furthermore, Patent Document 1 proposes a method for locally transforming variables to model the relationships between multiple variables, but it is difficult to generate an expression model because there is no guideline for what kind of expression model should ultimately be adopted.

[0016] 2. Overview of the Technology Disclosed The technology disclosed herein is characterized in that, particularly in causal models, there is no need to predefine an expression model when estimating structures and relationships, and after estimating structures and relationships, an expression model can be locally specified as needed.

[0017] Furthermore, the technology disclosed herein converts all variables globally into categorical variables by sequentially discretizing them, and enables the conversion into categorical variables using a representation model corresponding to the categorical variables. Furthermore, when all variables are continuous values, the technology disclosed herein enables the conversion into categorical variables using a continuous variable model of the same type globally, and also enables the conversion into categorical variables using a collection of continuous variable models of different types locally.

[0018] <3. Configuration and operation of information processing device> (Example of hardware configuration of information processing device) FIG. 1 is a diagram illustrating an example of a hardware configuration of an information processing device 11 according to an embodiment of the present disclosure.

[0019] When the information processing device 11 estimates causal relationships between multiple variables and expresses all or part of the causal relationships using an expression model, the information processing device 11 generates an expression model according to the types of variables included in the causal relationships to be expressed. Note that in the technology according to the present disclosure, the causal relationships between multiple variables are estimated based on the method disclosed in Patent Document 1, but a detailed description of the estimation method will be omitted in this specification.

[0020] The information processing device 11 can be configured, for example, by a computer such as a PC (Personal Computer) or a large / small computer system.

[0021] The information processing device 11 is configured to include a CPU (Central Processing Unit) 21, a ROM (Read Only Memory) 22, a RAM (Random Access Memory) 23, a bus 24, an input / output interface 25, an input unit 26, an output unit 27, a memory unit 28, a communication unit 29, and a drive 30.

[0022] In the information processing device 11, the CPU 21, ROM 22, and RAM 23 are interconnected by a bus 24. An input / output interface 25 is further connected to the bus 24. Connected to the input / output interface 25 are an input unit 26 including a keyboard, mouse, touch panel, etc., an output unit 27 including a display, speaker, etc., a storage unit 28 including a hard disk, non-volatile memory, etc., and a communication unit 29 including a network interface, etc.

[0023] A drive 30 is also connected to the input / output interface 25 as needed, and removable media 31 consisting of a magnetic disk, optical disk, magneto-optical disk, or semiconductor memory is appropriately attached, and programs read from them are installed in the memory unit 28 as needed.

[0024] The program can be received by the communication unit 29 via a wired or wireless transmission medium and installed in the storage unit 28. Alternatively, the program can be installed in the ROM 22 or the storage unit 28 in advance.

[0025] The program executed by the information processing device 11 may be a process that is performed in chronological order according to the order described in this specification, or may be a program that is performed in parallel or at the required timing, such as when a call is made.

[0026] (Example of functional configuration of information processing device) FIG. 2 shows an example of the functional configuration of an information processing device 11 to which the technology according to the present disclosure is applied.

[0027] The information processing device 11 in FIG. 2 is composed of a UI unit 51 and a processing unit 52.

[0028] The UI (User Interface) unit 51 functions as an interface between the information processing device 11 and a user who operates the information processing device 11. The UI unit 51 is configured to include an input unit 71 and a presentation unit 72.

[0029] The input unit 71 corresponds to the input unit 26 in Fig. 1, accepts the input of multiple variables and the selection of an expression model by the user, and supplies information corresponding to the contents of the input to the processing unit 52. The presentation unit 72 corresponds to the output unit 27 in Fig. 1, and, under the control of the processing unit 52, presents to the user an expression model according to the estimation result of the causal relationships between multiple variables.

[0030] The processing unit 52 corresponds to the CPU 21 in FIG. 1 and operates according to a program stored in a storage unit (not shown) to estimate the causal relationships among multiple variables input by the user and generate an expression model that expresses all or part of the causal relationships.

[0031] Specifically, the processing unit 52 executes a program stored in a memory unit (not shown) to realize functional blocks including a causal relationship estimation unit 81, an expression model generation unit 82, and a causal inference unit 83.

[0032] The causal relationship estimation unit 81 estimates the causal relationships between multiple variables input by the user based on the method disclosed in Patent Document 1 or the like.

[0033] When expressing all or part of the causal relationships between multiple variables estimated by the causal relationship estimation unit 81 using an expression model, the expression model generation unit 82 generates an expression model according to the types of variables included in all or part of the causal relationships to be expressed.

[0034] For example, when the object to be expressed includes at least one categorical variable, the expression model generation unit 82 generates a discrete probability model.

[0035] Specifically, when the representation target includes only categorical variables, the representation model generation unit 82 generates a discrete probability model using those categorical variables. When the representation target includes both categorical and numerical variables, the representation model generation unit 82 generates a discrete probability model by discretizing the numerical variables based on a reference categorical variable. The representation model generation unit 82 discretizes the numerical variables using Markov property. When the representation target is a part (local) of a causal relationship, the reference categorical variable may be selected by the user or automatically. The representation model generation unit 82 can also discretize the numerical variables by using at least one of a graph search algorithm and the discretization method described in Patent Document 1. Examples of graph search algorithms include the Kruskal algorithm, Prim's algorithm, Brufka algorithm, and reverse-delete algorithm. The edge weights used in the algorithm execution can be values ​​corresponding to the magnitude of correlation between the corresponding two variables. The magnitude of correlation can be measured using a correlation coefficient, mutual information, or a p-value in an independent test.

[0036] Furthermore, when the expression target includes only variables that can be expressed as functions, the expression model generation unit 82 generates a continuous probability model based on the goodness of fit.

[0037] In this case, the expression model generation unit 82 calculates an index value of the goodness of fit of each candidate continuous probability model to be generated. The index value may include a coefficient of determination or a multiple correlation coefficient.

[0038] At this time, the presentation unit 72 of the UI unit 51 presents to the user suggestion information suggesting whether a continuous probability model or a discrete probability model is appropriate as the expression model, based on the index value calculated by the expression model generation unit 82. The expression model generation unit 82 generates either a continuous probability model or a discrete probability model in accordance with the user's selection of the presented suggestion information.

[0039] Furthermore, the expression model generating unit 82 can automatically generate either a continuous probability model or a discrete probability model as the expression model based on the calculated index value.

[0040] The representation model that represents the entire causal relationship can include a discrete probability model generated as a representation model that represents a part of the causal relationship, and a continuous probability model generated as a representation model that represents another part of the causal relationship. In other words, the global representation model can be represented as the sum of discrete probability models and continuous probability models that are appropriately selected as local representation models.

[0041] Furthermore, the user can select a part of the causal relationship to be expressed via the UI unit 51 (input unit 71). As a result, the expression model generation unit 82 generates a local expression model that expresses the part of the causal relationship selected by the user via the UI unit 51.

[0042] The causal inference unit 83 can execute, to the extent possible, causal inference by intervention for the causal relationships for which expression models have been generated as described above.

[0043] In this case, the expression model generation unit 82 generates an expression model that expresses the causal relationship of the adjustment variables searched for according to the intervening variables, and the causal inference unit 83 performs causal inference based on the generated expression model.

[0044] (Operation of information processing device) The operation of the information processing device 11 will be described with reference to the flowchart of FIG.

[0045] First, in step S1, the input unit 71 inputs multivariate data in response to a user operation.

[0046] In step S2, the causal relationship estimation unit 81 estimates the causal relationships among the multiple variables input by the input unit 71.

[0047] In step S3, the expression model generation unit 82 generates an expression model (global expression model or local expression model) according to the type of variables included in all or part of the causal relationship to be expressed.

[0048] Then, in step S4, the presentation unit 72 presents the expression model generated by the expression model generation unit 82 to the user.

[0049] As described above, according to the technology disclosed herein, when generating a statistical model from data, it is possible to globally or locally select an expression model that is appropriate for the estimated relationship without specifying the expression model in advance.

[0050] For example, when a discrete probability model is selected as a global representation model, a model in which the entire model is discretized based on a reference variable can be generated. In this case, starting from the reference variable, a sequential discretization process is performed using the Markov property to create a model in which the entire model is discretized. Methods for performing the sequential discretization process include graph search algorithms such as the Prim method and the discretization method described in Patent Document 1. The edge weight used in executing the algorithm can be a value corresponding to the magnitude of correlation between the corresponding two variables. The magnitude of correlation can be measured using a correlation coefficient, mutual information, a p-value in an independent test, or the like.

[0051] On the other hand, when selecting a continuous probability model (multiple regression model, structural equation model, etc.) as the representation model from a global perspective, the representation model can be determined by selecting a model that fits well from linear models, probit models, logistic models, etc. At this time, an index value of the goodness of fit can be calculated and presented to the user in a list. Furthermore, based on the index value, it is possible to suggest to the user or automatically determine whether a continuous probability model or a discrete probability model is more suitable from a global perspective.

[0052] Furthermore, when selecting a discrete probability model in locally selecting a representation model, the discrete probability model can be applied by performing local discretization processing. In this case, a globally discretized model can be obtained by performing sequential discretization processing using Markov property with a certain variable as a reference. The reference variable may be selected by the user or automatically by the system. Methods for performing sequential discretization processing include graph search algorithms such as the Prim method and the discretization method described in Patent Document 1. The edge weights used in algorithm execution can be values ​​corresponding to the magnitude of correlation between the corresponding two variables. The magnitude of correlation can be measured using correlation coefficients, mutual information, p-values ​​in independent testing, etc.

[0053] On the other hand, when selecting a representation model locally, if a continuous probability model is selected, the representation model can be determined by selecting a model that fits well from among a linear model, a probit model, a logistic model, etc. In this case, if a function display is selected along with the representation model, a list of goodness-of-fit index values ​​can be presented to guide the user's selection. The goodness-of-fit index value can be a coefficient of determination or a multiple correlation coefficient. Furthermore, based on the index value, it is possible to suggest to the user, or automatically determine, for each local area whether a continuous probability model or a discrete probability model is more suitable.

[0054] Furthermore, according to the technology disclosed herein, when performing causal inference of an intervention, a set of variables necessary for inference calculation can be determined by defining the intervening variables and the variables for which the impact of the intervention is desired. For the set of variables to be determined, the user can select whether to calculate using a function representation or a discrete probability model. If the function representation is selected, a goodness-of-fit index value serving as a guideline for selection is calculated, allowing the user to select which functional representation to use, or the best-fit model to be automatically determined. The user can also decide to abandon the determined model and change to another model. Then, the necessary model estimation can be performed according to these decisions, and the calculation of causal inference of the intervention can be performed.

[0055] <4. Generation of local representation models> Here, the generation of a local representation model will be specifically described.

[0056] 4 is a flowchart illustrating an example of generating a local expression model. The process in FIG. 4 is executed when generating a local expression model in step S3 in FIG.

[0057] In step S11, the expression model generation unit 82 selects a variable of interest in the local area of ​​the causal relationship estimated by the causal relationship estimation unit 81, based on information from the input unit 71 in response to a user operation. At this time, the explanatory variables of the selected variable of interest are also automatically selected as parent variables.

[0058] In step S12, the expression model generation unit 82 determines whether the selected variables are categorical variables only.

[0059] If it is determined in step S12 that the selected variables are categorical variables only, the process proceeds to step S13, where the expression model generation unit 82 generates a discrete probability model. Here, for example, a joint probability table or a conditional probability table is generated as the discrete probability model.

[0060] On the other hand, if it is determined in step S12 that the selected variables are not only categorical variables, the process proceeds to step S14, where the expression model generation unit 82 determines whether the selected variables include categorical variables and numerical variables.

[0061] If it is determined in step S14 that the selected variables include categorical variables and numerical variables, the process proceeds to step S15, where the expression model generation unit 82 converts the numerical variables into categorical variables by discretizing the numerical variables based on one categorical variable. Thereafter, the expression model generation unit 82 generates a discrete probability model in step S13.

[0062] On the other hand, if it is determined in step S14 that the selected variables do not include categorical variables and numerical variables, the process proceeds to step S16, where the expression model generation unit 82 determines whether the selected variables are all variables that can be expressed as functions.

[0063] If it is determined in step S16 that the selected variables are not all variables that can be expressed as functions, the process proceeds to step S17, where the expression model generation unit 82 selects one of the variables of interest based on information from the input unit 71 in response to a user operation, and discretizes the variable to convert it into a categorical variable. Here, instead of the variable of interest selected in response to a user operation, an important variable selected based on importance may be selected and discretized. Thereafter, in step S15, the expression model generation unit 82 discretizes the numerical variables based on the categorical variable, and generates a discrete probability model in step S13.

[0064] If it is determined in step S16 that the selected variables are all variables that can be expressed as functions, the process proceeds to step S18, where the expression model generation unit 82 generates a continuous probability model. Here, for example, a functional causal model is generated as the continuous probability model.

[0065] For example, as shown in Figure 5, we have random variables X0, X1, X2, . . ., X 10 According to the technique disclosed in Patent Document 1, a relationship graph (graphical model) RG can be constructed by estimating the causal relationships between these variables.

[0066] In the relationship graph RG shown in Figure 5, the causal relationships between variables are shown using a technique called a path diagram, with variables connected by directed edges from cause to effect. For example, for variable X0, variables X3, X4, and X7 are presumed to be the causes. However, a quantitative method for expressing these causal relationships has not yet been determined.

[0067] An example of locally determining a representation method for such a relationship graph RG will be described.

[0068] First, as shown in FIG. 6, in the relationship graph RG, the variables X5, X6, and X 10 We know that all variables are numerical. Here, the variable X 10 If it is desirable to express the probabilistic decision formula using the variables X5 and X6 that are the causes, these variables can be described using a structural equation model or a multiple regression model because they are all numerical variables.

[0069] For example, variable X 10 can be expressed in the same format as a linear structural equation model as follows:

[0070]

number

[0071] In equation (1), a5, a6 are coefficients, ε 10 is a noise term.

[0072] Next, assume that the variables X2, X7, and X9 in the relationship graph RG, which are enclosed in a dashed frame L2, are all known to be categorical variables, as shown in Figure 7. In this case, if variable X7 is expressed using variables X2 and X9 that are estimated to be its causes, a joint probability table or conditional probability table is appropriate.

[0073] FIG. 8 is a diagram showing a conditional probability table of the variable X7 with the variables X2 and X9 as conditions.

[0074] In FIG. 8, variables X2 and X9 are binary, and variable X7 is a ternary categorical variable. Each column containing values ​​a1 to a4 and b1 to b4 (all numbers between 0 and 1) represents a condition, and the probability values ​​represent conditional probabilities. For example, a1, b1, and 1-a1-b1 represent the probability that X7=0, 1, and 2, respectively, given that X2=0 and X9=0. Values ​​a1 to a4 and b1 to b4 are parameters determined from data. These parameters can be easily estimated using cross-tabulation techniques in statistical analysis. Bayesian estimation may also be used to estimate these parameters. In other words, quantitative relationships can be determined from data using the conditional probability table shown in FIG. 8.

[0075] As shown in Figure 9, in the relationship graph RG, the causal relationships between the variables X2, X3, X5, and X9 enclosed in the dashed-line frame L3 are estimated by combining the categorical variables X2, X3, and X9 with the numerical variable X5. Here, the numerical variable X5 can be converted into a categorical variable by discretizing it using one categorical variable as a reference, using the method disclosed in Patent Document 1. Therefore, as in the example of Figure 7, all related variables can be treated as categorical variables and can be expressed as a joint probability table or a conditional probability table.

[0076] When there are multiple categorical variables to be used as a reference, the discretization can be performed using the relationship with the variable that has the strongest correlation when discretized.

[0077] <5. Generating a global representation model> Next, the generation of a global representation model will be specifically explained.

[0078] 10 is a flowchart illustrating an example of generating a global representation model, which is executed when generating a global representation model in step S3 of FIG.

[0079] In step S21, the expression model generation unit 82 determines whether the variables included in the entire causal relationship estimated by the causal relationship estimation unit 81 are only categorical variables, based on information from the input unit 71 in response to user operation.

[0080] If it is determined in step S21 that the variables included in the entire causal relationships are only categorical variables, the process proceeds to step S22, where the expression model generation unit 82 generates a discrete probability model. Here, for example, a joint probability table or a conditional probability table is generated as the discrete probability model.

[0081] On the other hand, if it is determined in step S21 that the variables included in the entire causal relationship are not only categorical variables, the process proceeds to step S23, where the expression model generation unit 82 determines whether the variables included in the entire causal relationship include categorical variables and numerical variables.

[0082] If it is determined in step S23 that the variables included in the entire causal relationships include categorical variables and numerical variables, the process proceeds to step S24, where the expression model generation unit 82 selects one categorical variable from the variables included in the entire causal relationships as a target variable or important variable. Hereinafter, the selected categorical variable will be referred to as the target categorical variable.

[0083] In step S25, the expression model generation unit 82 converts the numerical variables into categorical variables by discretizing the numerical variables, starting from the categorical variable of interest. Then, in step S22, the expression model generation unit 82 generates a discrete probability model.

[0084] On the other hand, if it is determined in step S23 that the variables included in the entire causal relationship do not include categorical variables and numerical variables, the process proceeds to step S27, where the expression model generation unit 82 determines whether the variables included in the entire causal relationship are only variables that can be expressed as functions.

[0085] If it is determined in step S27 that the variables included in the entire causal relationships are not all variables that can be expressed as functions, the process proceeds to step S28. In step S28, the expression model generation unit 82 selects one of the variables included in the entire causal relationships as a variable of interest or an important variable based on information from the input unit 71 in response to a user operation, and discretizes the selected variable, thereby converting the selected variable into a categorical variable of interest. Thereafter, in step S25, the expression model generation unit 82 discretizes the numerical variables starting from the categorical variable of interest, and generates a discrete probability model in step S22.

[0086] If it is determined in step S27 that the variables included in the entire causal relationship are all variables that can be expressed as functions, the process proceeds to step S29, where the expression model generation unit 82 generates a continuous probability model. Here, for example, a functional causal model is generated as the continuous probability model.

[0087] According to the above processing, all of the causal variables can be converted into categorical variables and expressed in a discrete probability model.

[0088] Structural equation models and multiple regression models are often expressed as linear or additive models, and it is difficult to determine how to construct more general nonlinear or nonadditive models as functions. In contrast, probability tables, which are a form of representation of categorical variables, can easily represent more realistic and flexible nonlinear and nonadditive models, although it is difficult to express detailed functions.

[0089] In the technology disclosed herein, if at least one categorical variable can be defined in a multivariate model in an algorithm for converting to categorical variables, the conversion to categorical variables can be performed in a chain reaction. This can be achieved by using a graph search algorithm for any point in discrete mathematics, such as the Kruskal method or Prim method, along with the discretization method described in Patent Document 1. Specifically, the entire relationship graph is converted into an undirected graph, and categorization (discretization) is performed while tracing the connecting links, thereby enabling the conversion to categorical variables in a chain reaction. The edge weights used in the algorithm execution can be values ​​corresponding to the magnitude of correlation between the corresponding two variables. The magnitude of correlation can be measured using a correlation coefficient, mutual information, a p-value in an independent test, or the like.

[0090] For example, in the relationship graph RG (Figure 5) described above, suppose variables X2 and X9 are categorical variables, and the other variables are numerical. In this case, starting with either variable X2 or X9, adjacent variables are sequentially categorized. If there are multiple categorical variables among the adjacent variables, the variable with the greatest correlation is selected first. By using methods such as the Kruskal method or Prim's method, all variables can be categorized based on the variable with the greatest average correlation. The edge weight used in executing the algorithm can be a value corresponding to the magnitude of correlation between the corresponding two variables. The magnitude of correlation can be measured using the correlation coefficient, mutual information, or the p-value in an independent test.

[0091] <6.UI example> The information processing device 11 can generate an expression model for the estimated causal relationship in response to a user's operation on the UI unit 51 and present the result. For example, when the user specifies "detailed modeling" of the estimated causal relationship via the input unit 71, the generated expression model and selection candidates are presented on the presentation unit 72.

[0092] In the following, variables X0 to X 10 An example of a UI when the relationship graph RG2 shown in Figure 11 is obtained as a result of inferring causal relationships between the two is explained below. The relationship graph RG2 can be a directed / undirected mixed graph, or a latent variable model (maximum / partial Markov ancestor model).

[0093] (Example 1) Regarding the relationship graph RG2 in Figure 11, when variable X6 is specified as the target variable and "detailed modeling" is selected, variables X5, X8, and X 10 are selected, and the relationship between these variables is expressed by an expression model. At this time, the expression model generation unit 82 determines whether it can be expressed by an equation model of four variables, and if it can be expressed, in addition to the equation model, a joint probability table and a conditional probability table are presented as selection candidates for the expression model.

[0094] If all four related variables are numerical and can be expressed by an equation model, equation fitting is performed to obtain the relationship between the variables as an equation.

[0095] (Example 2) Regarding the relationship graph RG2 in Figure 11, when variable X6 is specified as the target variable and "detailed modeling" is selected, variables X5, X8, and X 10 are selected, and the relationship between these variables is expressed by an expression model. At this time, the expression model generation unit 82 determines whether it can be expressed by an equation model of four variables, and if it can be expressed, in addition to the equation model, a joint probability table and a conditional probability table are presented as selection candidates for the expression model.

[0096] If all four related variables are numerical and difficult to express using an equation model, or if no equation model is selected, the user is prompted to select any one categorical variable from all variables in the relationship graph RG2. After one categorical variable is selected, the numerical variables are sequentially categorized using the method disclosed herein, and the joint probability table or conditional probability table described above is presented as a representation model. If no categorical variables exist, the user is prompted to select any one variable, and its distribution diagram is presented. The variable is converted into a categorical variable based on a threshold determined by the user, allowing the numerical variables to be sequentially categorized.

[0097] (Example 3) Regarding the relationship graph RG2 in Figure 11, when variable X6 is specified as the target variable and "detailed modeling" is selected, variables X5, X8, and X 10 are selected, and the relationship between these variables is expressed by an expression model. At this time, the expression model generation unit 82 determines whether it can be expressed by an equation model of four variables, and if it can be expressed, in addition to the equation model, a joint probability table and a conditional probability table are presented as selection candidates for the expression model.

[0098] If some of the four related variables are numerical and difficult to express using an equation model, the user is prompted to select any one categorical variable from the four variables. The selected categorical variable may be selected automatically. After the categorical variable is selected, the numerical variables are sequentially categorized using the method disclosed herein, and the joint probability table or conditional probability table described above is presented as a representation model.

[0099] (Example 4) For the relationship graph RG2 in FIG. 11, when variable X5 is specified as the variable of interest and "detailed modeling" is selected, it is unclear whether variables X3, X7, and X8, which are connected to variable X5 by undirected edges, are direct cause variables. Therefore, the user is prompted to determine the direction of those edges. Alternatively, the user is prompted to correct the direction of directed edges connecting variable X5 to other variables. After the direction of edges between variables is determined, the relationship between variable X5 and the direct cause variables is expressed in the representation model. At this time, the direct cause variables of variable X5 may be determined by user selection. For example, only variables X7 and X8 may be selected for variable X5, and only the relationship between these variables may be determined.

[0100] When the relationship graph RG2 is a latent variable model, after the directions between the variables are determined, it is determined whether or not the calculation is possible, and if the calculation is possible, the above-described processing is executed.

[0101] Thereafter, the expression model is determined in the same manner as in Examples 1 to 3.

[0102] (Example 5) For the relationship graph RG2 in FIG. 11, if no variables are designated as the target variables and "detailed modeling" is selected, the direct cause variables for each variable are identified to determine the details of all variables. Variables that cannot be identified are highlighted, and the user is prompted to determine the direction of the edge related to that variable, as in Example 4. Since the edge direction can be determined locally, the methods of Examples 1 to 4 can be repeatedly applied. Furthermore, by selecting any one categorical variable using the methods of Example 2 or Example 3, all variables can be converted into categorical variables using the method of the present disclosure.

[0103] <7. Network system configuration> In the above, it has been assumed that the estimation of causal relationships and the generation of expression models are performed within the information processing device 11, but it is also possible to perform the estimation of causal relationships and the generation of expression models between devices connected via a network.

[0104] FIG. 12 shows an example of the configuration of a network system to which the technology according to the present disclosure is applied.

[0105] The network system shown in FIG. 12 is configured by a client 111 such as a mobile terminal operated by a user, and a server 113 connected via a network 112 such as the Internet.

[0106] The client 111 includes an input unit 121 and a presentation unit 122 corresponding to the input unit 71 and the presentation unit 72 (UI unit 51) in FIG. 2, respectively.

[0107] 2. The server 113 also includes a processing unit 131 that corresponds to the processing unit 52 in FIG.

[0108] In such a configuration, the server 113 can infer causal relationships and generate expression models in response to instructions from the client 111 , and output the results to the client 111 .

[0109] <8. Application Examples> The technology according to the present disclosure can be applied to the following examples.

[0110] (1) Analysis of the causes of manufacturing quality defects In the manufacturing industry, the technology disclosed herein can be applied to a system that analyzes the cause of defects in the manufacturing quality of a device.

[0111] In this case, various values ​​and conditions are recorded as variables linked to device IDs. The variables include numerical and categorical data, and causal relationships are estimated without assuming distribution functions or functional equations. If the estimation results can be expressed as a directed graph, a partially directed graph, or an ancestral graph that includes potential confounding variables and potential selection biases, qualitative relationships can be expressed in these graphs or in text.

[0112] Next, a method for quantitatively expressing the expressed causal relationship is determined. At this time, the technology according to the present disclosure makes it possible to generate a quantitative expression model for each local area. For example, suppose a relationship is estimated in which variables A and B are the cause and variable C is the result. For example, variables A and B may be the current, voltage, or pressure in the device manufacturing process, and variable C may be the output characteristics (output value) of the device or the film thickness of a semiconductor. In other words, variables A, B, and C are all numerical variables, and it is desired to obtain a quantitative expression model expressed in function form. In this case, the technology according to the present disclosure makes it possible to select a functional expression that is a candidate for fit, and also to refer to the degree of fit. For example, if variables A and B are close to a linear relationship, the degree of fit of the functional expression C = aA + bB + e is high, and C / C max The function expression =1 / [1+exp{-(a0+a1A+a2B)}] is shown to have a low degree of fit, allowing the user to understand that selecting the former function expression is a better choice.

[0113] On the other hand, when the same data and analysis results are obtained, it is assumed that variables D, E, and F are estimated to have a causal relationship apart from variables A, B, and C. Variable D is a numerical variable, and variables E and F are categorical variables that are not binary, and the system determines that it is difficult to express them as functions. In this case, according to the technology disclosed herein, by converting variable D into a categorical variable, it is possible to quantitatively express the local relationship between these variables using a conditional probability table or a joint probability table instead of a functional expression.

[0114] In this way, the causal relationships between multiple variables can be presented by selecting the most appropriate representation format for each local area, or by having the system automatically determine and present the format, allowing the user to better understand the results of the data analysis.

[0115] Furthermore, the causal effects of intervention can be inferred using the analysis results described above. In particular, the technology disclosed herein allows causal inference to be performed using expressions appropriate for each locality, as follows:

[0116] For example, consider the case where, in the relationship between variables A, B, and C described above, if a hypothetical intervention were made with variable A to shift the overall distribution, i.e., to shift the mean value by a certain value, and then analyze how variable C would be affected as a result. In this case, the system searches for adjustment variables according to the theory of causal inference, selects variables G and H as suitable adjustment variables, and the variables required for the causal inference calculation are variables A, B, C, G, and H. Assume that variables G and H are all categorical variables. In this case, according to the technology disclosed herein, the system determines that it is difficult to express the relationship between these variables as a function and recommends to the user that they be calculated as all categorical variables. That is, all variables A, B, C, G, and H can be converted into categorical variables, and the causal inference calculation can be performed in the form of a probability table, and the results can be presented.

[0117] (2) Marketing data analysis In analyzing marketing data, the technology disclosed herein can be applied to a system that analyzes various data linked to customer IDs.

[0118] The variables here include attribute data such as whether or not a certain product has been purchased, gender, age, and residential area, purchase history of other products, and subscription status and billing status of other services. Age and billing status are numerical variables, while gender, residential area, and purchase status are categorical variables. When it is desired to estimate the causal relationships of this data and quantitatively confirm the estimation results, it is assumed that the numerical and categorical variables are intricately related, and the system determines that it is difficult to express the relationships surrounding several variables of interest as functions. According to the technology disclosed herein, a user can determine, or the system can automatically determine, that it is appropriate to express the data using a conditional probability table or a joint probability table, thereby converting all variables into categorical variables and presenting the causal relationships between any number of variables.

[0119] (3) Analysis of medical databases The technology disclosed herein can be applied to a system that analyzes medical databases.

[0120] Suppose a certain disease, various other diseases thought to be related to that disease, measurement results, medical examination results, prescriptions, treatment results, and other data are linked to an individual's ID. In this case, for example, blood pressure and uric acid levels contained in the measurement results are considered numerical variables, while gender and whether or not medication is administered are considered categorical variables. After analyzing the causal relationship between these variables and estimating whether or not there is a relationship, if you want to quantitatively confirm that relationship, assume that variables A, B, and C are all numerical variables of measurement results, and variable D is a binary categorical variable.

[0121] According to the technology disclosed herein, by searching for a method for expressing the relationship between these variables, it is possible to express it using a linear multiple regression equation, and by presenting the calculated degree of fit, if the user judges that the expression is good, it is possible to express the causal relationship between variables A, B, C, and D. Furthermore, according to the technology disclosed herein, when it is estimated that there is a causal relationship between variables B, E, F, and G, which are a mixture of numerical and categorical variables, and it is judged that a functional expression is difficult, the user can instruct, or the system can automatically decide, to convert all variables into categorical variables and present the results. Furthermore, according to the technology disclosed herein, when converting numerical variables into categorical variables, the user can select a reference categorical variable, or the system can automatically set an appropriate categorical variable.

[0122] 9. Effects of the Technology According to the Present Disclosure Traditionally, estimating causal relationships and constructing a minimum necessary relationship model from correlations have been important data analysis methods for understanding data, making predictions based on that data, estimating causal effects, etc. However, for multivariate data that contains a mixture of numerical and categorical variables, it was necessary to predefine a representation model that assumed that the variables were of one type before conducting the analysis.

[0123] The technology disclosed herein makes it possible to estimate relationships between multiple variables without specifying a representation model, and then generate a representation model that optimally represents the relationships locally after estimating the relationships. This allows users to select a representation format that is easy for them to understand, learn a suitable quantitative relationship graph, and obtain information contained in the data with minimal loss. As a result, users can understand predictions and causal effect estimates using a model that is closer to reality with fewer assumptions, enabling them to effectively utilize data.

[0124] It should be noted that the embodiments of the present disclosure are not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the technology according to the present disclosure.

[0125] For example, the technology according to the present disclosure can be configured as a cloud computing system in which a single function is shared and processed jointly by multiple devices via a network.

[0126] Furthermore, each step described in the above flowchart can be executed by one device, or can be shared and executed by multiple devices.

[0127] Furthermore, when one step includes multiple processes, the multiple processes included in that one step can be executed by one device or can be shared and executed by multiple devices.

[0128] Furthermore, the technology according to the present disclosure can have the following configuration. (1) a causal relationship estimation unit that estimates causal relationships between multiple variables; an expression model generation unit that generates an expression model according to a type of a variable included in the causal relationship to be expressed when the whole or part of the estimated causal relationship is expressed by an expression model; An information processing device comprising: (2) The expression model generation unit generates a discrete probability model when the expression target includes at least one categorical variable. The information processing device described in (1). (3) When the object to be expressed includes the categorical variables and the numerical variables, the expression model generation unit generates the discrete probability model by discretizing the numerical variables based on the categorical variables that serve as a reference. (2) An information processing device according to the present invention. (4) The representation model generation unit discretizes the numerical variables by utilizing the Markov property. (3) An information processing device according to the present invention. (5) When the object to be expressed is part of the causal relationship, the categorical variable to be used as a reference is selected by the user or automatically selected. An information processing device according to (3) or (4). (6) The representation model generation unit discretizes the numerical variables by using a graph search algorithm. An information processing device according to any one of (3) to (5). (7) The graph search algorithm includes at least one of Kruskal's algorithm, Prim's algorithm, Brufka's algorithm, and inverse reduction algorithm, in which the weight of an edge is a value corresponding to the magnitude of correlation between the corresponding two variables. (6) An information processing device according to (6). (8) The expression model generation unit generates a continuous probability model based on goodness of fit when the expression target includes only variables that can be expressed as functions. The information processing device described in (1). (9) The representation model generation unit calculates an index value of the goodness of fit for each of the continuous probability models that are candidates for generation. (8) An information processing device according to (8). (10) The index value includes a coefficient of determination or a multiple correlation coefficient. (9) An information processing device according to (9). (11) a presentation unit that presents to a user suggestion information that suggests whether the continuous probability model or the discrete probability model is appropriate as the expression model based on the index value; The expression model generation unit generates either the continuous probability model or the discrete probability model in response to the user's selection of the suggestive information. The information processing device according to (9) or (10). (12) The expression model generation unit automatically generates either the continuous probability model or the discrete probability model as the expression model based on the index value. The information processing device according to (9) or (10). (13) The representation model that represents the entire causal relationship includes a discrete probability model that is generated as the representation model that represents a part of the causal relationship, and a continuous probability model that is generated as the representation model that represents another part of the causal relationship. An information processing device according to any one of (1) to (12). (14) The expression model generation unit generates the expression model that expresses a part of the causal relationship selected by a user via a UI unit. An information processing device according to any one of (1) to (13). (15) The present invention further includes a causal inference unit that performs causal inference by intervention on the causal relationship for which the expression model has been generated. An information processing device according to any one of (1) to (14). (16) the expression model generation unit generates the expression model that expresses the causal relationship of the adjustment variables searched for depending on the intervening variables; The causal inference unit executes the causal inference based on the generated expression model. (15) An information processing device according to (15). (17) Estimating causal relationships between multiple variables and When expressing the whole or part of the estimated causal relationship by an expression model, the expression model is generated according to the type of variables included in the causal relationship to be expressed. An information processing method including: (18) Estimating causal relationships between multiple variables and When expressing the whole or part of the estimated causal relationship by an expression model, the expression model is generated according to the type of variables included in the causal relationship to be expressed. A program for causing a computer to execute a process including the above. [Explanation of symbols]

[0129] 11 information processing device, 51 UI unit, 52 processing unit, 71 input unit, 72 presentation unit, 81 causal relationship estimation unit, 82 expression model generation unit, 83 causal inference unit

Claims

1. a causal relationship estimation unit that estimates causal relationships between multiple variables; an expression model generation unit that generates an expression model according to a type of a variable included in the causal relationship to be expressed when the whole or part of the estimated causal relationship is expressed by an expression model; An information processing device comprising:

2. The expression model generation unit generates a discrete probability model when the expression target includes at least one categorical variable. The information processing device according to claim 1 .

3. When the object to be expressed includes the categorical variables and the numerical variables, the expression model generation unit generates the discrete probability model by discretizing the numerical variables based on the categorical variables that serve as a reference. The information processing device according to claim 2 .

4. The representation model generation unit discretizes the numerical variables by utilizing the Markov property. The information processing device according to claim 3 .

5. When the object to be expressed is part of the causal relationship, the categorical variable to be used as a reference is selected by the user or automatically selected. The information processing device according to claim 3 .

6. The representation model generation unit discretizes the numerical variables by using a graph search algorithm. The information processing device according to claim 3 .

7. The graph search algorithm includes at least one of Kruskal's algorithm, Prim's algorithm, Brufka's algorithm, and inverse reduction algorithm, in which the weight of an edge is a value corresponding to the magnitude of correlation between the corresponding two variables. The information processing device according to claim 6 .

8. The expression model generation unit generates a continuous probability model based on goodness of fit when the expression target includes only variables that can be expressed as functions. The information processing device according to claim 1 .

9. The representation model generation unit calculates an index value of the goodness of fit for each of the continuous probability models that are candidates for generation. The information processing device according to claim 8 .

10. The index value includes a coefficient of determination or a multiple correlation coefficient. The information processing device according to claim 9 .

11. a presentation unit that presents to a user suggestion information that suggests whether the continuous probability model or the discrete probability model is appropriate as the expression model based on the index value; The expression model generation unit generates either the continuous probability model or the discrete probability model in response to the user's selection of the suggestive information. The information processing device according to claim 9 .

12. The expression model generation unit automatically generates either the continuous probability model or the discrete probability model as the expression model based on the index value. The information processing device according to claim 9 .

13. The representation model that represents the entire causal relationship includes a discrete probability model that is generated as the representation model that represents a part of the causal relationship, and a continuous probability model that is generated as the representation model that represents another part of the causal relationship. The information processing device according to claim 1 .

14. The expression model generation unit generates the expression model that expresses a part of the causal relationship selected by a user via a UI unit. The information processing device according to claim 1 .

15. The present invention further includes a causal inference unit that performs causal inference by intervention on the causal relationship for which the expression model has been generated. The information processing device according to claim 1 .

16. the expression model generation unit generates the expression model that expresses the causal relationship of the adjustment variables searched for depending on the intervening variables; The causal inference unit executes the causal inference based on the generated expression model. The information processing device according to claim 15.

17. Estimating causal relationships between multiple variables and When expressing the whole or part of the estimated causal relationship by an expression model, the expression model is generated according to the type of variables included in the causal relationship to be expressed. An information processing method including:

18. Estimating causal relationships between multiple variables and When expressing the whole or part of the estimated causal relationship by an expression model, the expression model is generated according to the type of variables included in the causal relationship to be expressed. A program for causing a computer to execute a process including the above.

Images