Analysis system, analysis method, and analysis program

Abstract

To provide an evaluation system that increases reliability of newly estimated causal structure and can efficiently evaluate new causal structure.SOLUTION: An analysis system includes: an acquisition section that uses a value of an additional variable of additional data as correct answer data with a value of a first variable of additional data that is the value of a first variable appearing in first causal structure as input data, creates a machine learning model for outputting a value of an additional variable with a first variable as input, inputs observation data to the created machine learning model, and acquires a prediction value of an additional variable; an estimation section for estimating second causal structure by using data obtained by connecting observation data, a prediction value, and additional data; and a display section for displaying a structure display screen displaying at least second causal structure in a display mode allowing a difference between first causal structure and the second causal structure to be identifiable.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to an analysis system, an analysis method, and an analysis program for causal analysis. [Background technology]

[0002] Conventionally, the causal structure of a process has been estimated based on observed data of the process (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-194560 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, there are cases where it is desired to estimate a more plausible causal structure using newly acquired additional data after a causal analysis has already been performed. In such cases, there is a demand for a technique for increasing the reliability of the newly estimated causal structure and efficiently evaluating the new causal structure. [Means for solving the problem]

[0005] According to one embodiment of the present disclosure, there is provided an analysis system for causal analysis, the analysis system including: a receiving unit that receives input of a first causal structure determined using observation data having values ​​of observed variables, and additional data having values ​​of additional variables different from the observed variables and the observed variables, an acquisition unit that creates a machine learning model that uses the value of the first variable in the additional data as input data and the value of the additional variable in the additional data as ground truth data, and outputs the value of the additional variable using the first variable as an input, inputs the observed data into the created machine learning model, and acquires a predicted value of the additional variable, an estimation unit that estimates a second causal structure using data obtained by combining the observed data, the predicted value, and the additional data, and a display unit that displays at least the second causal structure in a display mode in which a difference between the first causal structure and the second causal structure can be identified. [Brief description of the drawings]

[0006] [Figure 1] FIG. 1 is a block diagram showing a configuration of an analysis system. [Diagram 2] 11 is a flowchart showing the procedure of a causal analysis process. [Diagram 3] Causal diagram for the first causal structure. [Figure 4] FIG. 1 is a diagram illustrating a comparison between observed data, additional data, and combined data. [Diagram 5] Diagram explaining machine learning in a machine learning model. [Figure 6] FIG. 4 is a diagram for explaining the relationship between a first variable and an observed variable. [Figure 7] Causal diagram for the second causal structure. [Figure 8] FIG. [Figure 9] FIG. 11 is a diagram for explaining another embodiment of the structure display screen. [Figure 10] 10 is a flowchart showing the procedure of a causal analysis process according to the second embodiment. [Figure 11]FIG. 13 is a diagram for explaining a structure display screen that is displayed when an edge connecting an important variable and an additional variable does not appear. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0007] A. First embodiment: A1. System Configuration: FIG. 1 is a block diagram showing a configuration of an analysis system 1 for causal analysis. The analysis system 1 includes an information processing device 100, an input device 140, and a display unit 150. The information processing device 100 can be realized by, for example, a personal computer. The information processing device 100 has a processor 110, a memory 120, an interface circuit 130, and a communication interface 132. The input device 140 and the display unit 150 are connected to the interface circuit 130. In addition, a storage medium 160 can be connected to the interface circuit 130. The input device 140 is specifically a keyboard or a mouse. The display unit 150 is specifically a display such as a liquid crystal display. The storage medium 160 is specifically a flash memory. The communication interface 132 communicates with an external device through a network such as a LAN (Local Area Network).

[0008] The processor 110 has a reception unit 112, an acquisition unit 114, an estimation unit 116, and an output unit 118 as functional units. The reception unit 112, the acquisition unit 114, the estimation unit 116, and the output unit 118 are realized by the processor 110 executing a computer program stored in the memory 120. However, each function may be realized by a dedicated electronic circuit. The term processor in this specification is a term that also includes such a dedicated electronic circuit. In addition, the processor that executes the processing of each functional unit may be a processor included in a remote computer connected to the information processing device 100 via a network. In addition, the processing performed by each functional unit may be executed by multiple processors. The memory 120 stores a machine learning model 124, a first causal structure CAa, and additional data 126.

[0009] A2.Causal analysis processing: Fig. 2 is a flowchart showing the procedure of the causal analysis process of this embodiment. Fig. 3 is a causal diagram of the first causal structure CAa. Fig. 4 is a diagram for explaining a comparison between the observed data, the additional data 126, and the combined data. Fig. 5 is a diagram for explaining the machine learning of the machine learning model 124. Fig. 6 is a diagram for explaining the relationship between the first variable EVa and the observed variable ma. Fig. 7 is a causal diagram of the second causal structure CAb.

[0010] The causal analysis process is a process for estimating a new causal structure using newly acquired additional data 126 in order to obtain a more plausible causal structure after the causal analysis using the observed data has been performed. The following describes an example of a case where causal analysis is performed for factor analysis aimed at improving a coating rate, which is the ratio of the amount of paint actually attached to an object to the amount of paint used in painting, in a painting process in which paint is sprayed onto an object.

[0011] The causal analysis process is started when an instruction to start the causal analysis process is received, for example, via the input device 140 shown in FIG. 1. In step S10 as the reception process and reception process in FIG. 2, the reception unit 112 receives input of the first causal structure CAa obtained using the observation data having the value of the observation variable ma, the additional data 126, and a threshold value. Specifically, the reception unit 112 displays on the display unit 150 a GUI (Graphical User Interface) for receiving designation of the respective storage locations of the first causal structure CAa and the additional data 126, and a GUI for receiving the value of the threshold value in step S16 described later. Then, the reception unit 112 acquires and receives the first causal structure CAa and the additional data 126 from the storage locations input to the GUI. Similarly, the reception unit 112 stores the value of the threshold value input to the GUI in the memory 120.

[0012] As shown in FIG. 4, the observation data is data having a value of an observation variable ma and a sample size of "na". The observation variable ma has a plurality of variables. In the illustrated example, the observation variable ma includes "spray distance", "paint temperature", "discharge amount", "adhesion rate", and other variables shown in FIG. 3. In this embodiment, the first causal structure CAa is a causal structure obtained using a causal analysis algorithm. As the causal analysis algorithm, for example, GFCI (Greedy Fast Causal Inference) or FGES (Fast Greedy Equivalence Search) can be used. The first causal structure CAa is a causal structure obtained by inputting the observation data into the causal analysis model 122. In the illustrated example, as shown in FIG. 3, the first causal structure CAa includes the variables "spray distance", "paint temperature", "discharge amount", and "adhesion rate". Here, the variable that appears in the first causal structure CAa is called a first variable EVa.

[0013] 4, the additional data 126 is data having a value of an observed variable ma and a value of an additional variable mb, and a sample size of "nb." In this embodiment, the sample size nb of the additional data 126 is smaller than the sample size na of the observed data.

[0014] The additional variable mb is a variable extracted by the user based on the user's knowledge. The additional variable mb is a variable different from the observed variable ma. In the example, the additional variable mb includes "viscosity", which is the viscosity of the paint, and other variables, as shown in FIG. 7.

[0015] In this embodiment, before the causal analysis process shown in FIG. 2 is performed, the first causal structure CAa is obtained using the observation data and stored in the memory 120. In addition, in this embodiment, the additional data 126 is stored in the memory 120 before the causal analysis process shown in FIG. 2 is performed. Then, as described above, the user inputs the respective storage locations of the first causal structure CAa and the additional data 126 into the GUI displayed on the display unit 150. Note that, as another embodiment, the respective storage locations of the first causal structure CAa and the additional data 126 may be set in advance, and the reception unit 112 may acquire the first causal structure CAa and the additional data 126 from the preset storage locations. Also, the storage location of the first causal structure CAa and the additional data 126 is not limited to the memory 120. The storage location may be, for example, a server or other information processing device connected to the network to which the communication interface 132 is connected. Also, the storage location may be the storage medium 160.

[0016] 2, the acquisition unit 114 uses the value of the first variable EVa of the additional data 126 as input data and the value of the additional variable mb of the additional data 126 as correct answer data to create a machine learning model 124 that takes the first variable EVa as input and outputs the value of the additional variable mb. The created machine learning model 124 is a machine learning model that predicts the additional variable mb using the first variable EVa as input. As the machine learning model 124, various neural networks, support vector machines (SVMs), gradient boosting trees (GBDTs), random forests, etc. can be used.

[0017] As described above, the first variable EVa is a variable that appears in the first causal structure CAa. As shown in FIG. 5, in the illustrated example, the acquisition unit 114 creates a machine learning model 124 that takes the first variable EVa, namely, "spray distance," "paint temperature," "discharge amount," and "adhesion rate," as inputs and outputs the additional variable "viscosity." Here, as shown in FIG. 6, the first variable EVa is a part of the observation variable ma. By creating the machine learning model 124 using the first variable EVa, which is a part of the observation variable ma, rather than the observation variable ma, it is expected that overlearning can be suppressed. Furthermore, by using the first variable EVa that appears in the first causal structure CAa as an input, more direct feature selection can be realized.

[0018] In the process of creating the machine learning model 124, the acquisition unit 114 obtains a numerical value that indicates the magnitude of an error between the value of the additional variable mb of the additional data 126 and the predicted value of the additional variable that is output by inputting the additional data 126 to the machine learning model 124. Then, the acquisition unit 114 creates the machine learning model 124 so that the numerical value that indicates the magnitude of the error becomes small. As the numerical value that indicates the magnitude of the error, a training error or a generalization error can be used. The training error or the generalization error can be obtained using a loss function. When the generalization error is used, a method of creating the machine learning model 124 is performed by performing test data such as a cross-validation method. In this specification, the training error and the generalization error are collectively referred to as a learning error.

[0019] In step S14 as the acquisition step and acquisition process in FIG. 2, the acquisition unit 114 inputs the observed data to the machine learning model 124 created in step S12, and acquires a predicted value of the additional variable mb.

[0020] In step S16, the estimation unit 116 determines whether the learning error of the machine learning model 124 is equal to or smaller than a predetermined threshold. In this embodiment, the threshold used in step S16 is the value received in step S10. As another embodiment of the threshold used in step S16, a predetermined fixed value may be used.

[0021] If it is determined in step S16 that the learning error is equal to or smaller than the threshold, the estimation unit 116 estimates a second causal structure CAb, which is a new causal structure, using data obtained by combining the observed data, the predicted value, and the additional data 126 in step S18 as an estimation process and estimation process. The case where the learning error is equal to or smaller than the threshold means that the additional variable mb can be obtained with high accuracy using the first variable EVa. In this case, the additional variable mb is highly likely to appear in the second causal structure CAb, and it is meaningful to estimate the second causal structure CAb. On the other hand, if it is determined in step S16 that the learning error is not equal to or smaller than the threshold, that is, if it is determined that the learning error is larger than the threshold, the estimation unit 116 proceeds to the process in step S28. The case where the learning error is larger than the threshold means that the additional variable mb cannot be obtained with high accuracy using the first variable EVa. In this case, it is unlikely that the additional variable mb will appear in the second causal structure CAb, or the predicted additional variable mb is unlikely to be reliable, making it unlikely that a more plausible causal structure than the first causal structure CAa can be inferred using the additional variable mb.

[0022] As shown in "combined data" in FIG. 4, the combined data obtained by combining the observed data, the predicted value obtained in step S14, and the additional data 126 is data in which the value of the additional variable mb, which is not included in the observed data, is substituted with the predicted value. In "combined data" in FIG. 4, the predicted value obtained in step S14 is displayed by cross-hatching, and the additional data 126 is displayed by single hatching. The causal analysis requires input of complete data including the values ​​of all variables. Therefore, when estimating the second causal structure CAb, the observed data cannot be used as it is because it does not include the value of the additional variable mb. Therefore, by combining the predicted value of the value of the additional variable mb with the observed data, the observed data can be effectively utilized. In addition, when the sample size of the additional data 126 is small, data with a large sample size can be used by using the combined data, so that the causal analysis is less likely to become unstable when the second causal structure CAb is obtained.

[0023] 7, in the illustrated example, the variable "viscosity" included in the additional variable mb appears in the second causal structure CAb. Here, the variable appearing in the second causal structure CAb is referred to as the second variable EVb.

[0024] In step S26 of Fig. 2, the display unit 150 displays a structure display screen DSa that displays at least the second causal structure CAb in a display mode that allows the difference between the first causal structure CAa and the second causal structure CAb to be identified. In this embodiment, a causal diagram of the second causal structure CAb is displayed on the structure display screen DSa shown in Fig. 8. In step S26, more specifically, the output unit 118 performs an output process and an output step of creating structure display screen data for displaying the structure display screen DSa and outputting the data to the display unit 150. As a result, the structure display screen DSa is displayed on the display unit 150.

[0025] FIG. 8 is a diagram for explaining the structure display screen DSa. FIG. 9 is a diagram for explaining another embodiment of the structure display screen DSa. As shown in FIG. 8, in the structure display screen DSa, each variable of the second variable EVb is displayed as a node, and a side connecting two nodes corresponding to any two variables of the second variable EVb that have a causal relationship is displayed as an edge in a diagram. In addition, in the structure display screen DSa, a node corresponding to "viscosity", which is a variable different between the first variable EVa and the second variable EVb, is displayed in a display mode different from that of the nodes corresponding to "spray distance", "paint temperature", and "discharge amount", which are variables common to the first variable EVa and the second variable EVb. In this embodiment, the nodes corresponding to "spray distance", "paint temperature", and "discharge amount" are displayed in the same color. And the node corresponding to "viscosity" is displayed in a color different from that corresponding to "spray distance". As a result, the user can immediately see that the different part between the first causal structure CAa and the second causal structure CAb is "viscosity" by looking at the structure display screen DSa, and therefore can efficiently evaluate the second causal structure CAb.

[0026] The difference in the display mode of the nodes displayed on the structure display screen DSa is not limited to the color of the nodes. The difference in the display mode may be the shape of the nodes or the display mode. Specifically, the difference in the display mode refers to a mode in which only the node corresponding to "viscosity" is flashed, or a mode in which only the node corresponding to "viscosity" is displayed with a graphic such as an arrow. The display mode in which the difference between the first causal structure CAa and the second causal structure CAb can be identified is not limited to FIG. 8, and for example, the first causal structure CAa and the second causal structure CAb may be displayed side by side. As shown in FIG. 9, the display unit 150 may display text information 201 that identifies the different variables between the first variable EVa and the second variable EVb.

[0027] In this embodiment, in step S26, the display unit 150 displays a first evaluation screen DSb and a second evaluation screen DSc in addition to the structure display screen DSa. In detail, the output unit 118 creates first evaluation screen data for displaying the first evaluation screen DSb and second evaluation screen data for displaying the second evaluation screen DSc, and outputs them to the display unit 150. As a result, the first evaluation screen DSb and the second evaluation screen DSc are displayed on the display unit 150.

[0028] As shown in FIG. 8, the first evaluation screen DSb displays a first evaluation graph 202 and an evaluation index 203. The first evaluation graph 202 is a graph showing the correlation between the value of the additional variable mb of the additional data 126 and the predicted value of the additional variable mb outputted when the additional data 126 is inputted to the machine learning model 124. When creating the first evaluation graph 202, the value of the additional variable mb of the additional data 126, which is an actual measurement value, is inputted as the value of the horizontal axis of the first evaluation graph 202. When creating the first evaluation graph 202, the predicted value of the additional variable mb of the additional data 126, which is an output of the machine learning model 124, is inputted as the value of the vertical axis of the first evaluation graph 202.

[0029] The evaluation index 203 is an evaluation index of the machine learning model 124. The evaluation index 203 includes a correlation coefficient (R), a coefficient of determination (R 2 ), root mean square error (RMSE), mean square error (MAE), MIC, etc. can be used.

[0030] For example, in regression analysis, the coefficient of determination (R 2 The closer to 1, the better the fit of the regression equation. For example, the coefficient of determination (R 2 Even if the value of the additional data 126 is close to 1, such as "0.8," the regression equation may not fit well in a portion of the range of values ​​of the additional data 126. Therefore, the first evaluation graph 202 is displayed, allowing the user to check the degree of variation in the predicted values ​​for the regression equation.

[0031] The second evaluation screen DSc displays a second evaluation graph 204 that shows the relationship between the value of the first variable EVa of the observation data and the predicted value of the additional variable mb obtained by inputting the value of the first variable EVa of the observation data to the machine learning model 124. When creating the second evaluation graph 204, the value of the additional variable mb of the observation data, which is the input of the machine learning model 124, is input as the value of the horizontal axis of the second evaluation graph 204. When creating the second evaluation graph 204, the predicted value of the additional variable mb of the additional data 126, which is the output of the machine learning model 124, is input as the value of the vertical axis of the second evaluation graph 204. In the illustrated example, in response to a user's selection, the second evaluation screen DSc displays, as the second evaluation graph 204, any one of a graph showing the relationship between "spray distance" and "viscosity", a graph showing the relationship between "paint temperature" and "viscosity", a graph showing the relationship between "discharge amount" and "viscosity", and a graph showing the relationship between "application rate" and "viscosity". By displaying the second evaluation graph 204, the user can check the correlation between the explanatory variables and the objective variable in the machine learning model 124.

[0032] In step S28 of FIG. 2, the display unit 150 displays information indicating that the second causal structure CAb has not been estimated. In this embodiment, the display unit 150 displays a message stating "estimation of the second causal structure has been stopped" as information indicating that the second causal structure CAb has not been estimated. This allows the user to know that the learning error of the machine learning model is large and the second causal structure CAb has not been created. When the message in step S28 is displayed, the user can take measures such as setting another variable as an additional variable or judging that the first causal structure CAa is a sufficiently plausible causal structure. After the completion of step S26 and the completion of step S22, this processing routine ends.

[0033] Although the painting process has been described as an example, the field to which the causal analysis process of the present embodiment is applied is not limited to the manufacturing industry. For example, the causal analysis of the present embodiment can be applied to other fields such as marketing. Furthermore, in step S12, the learning method applied in the process of creating the machine learning model 124 is not limited to regression, but may be classification. In the case of classification, for example, a confusion matrix may be used to verify the classification accuracy of the machine learning model 124. Then, when the confusion matrix is ​​used, a table of the confusion matrix may be displayed on the first evaluation screen DSb in step S26. Furthermore, when the confusion matrix is ​​used, the AUC (Area Under the Curve) may be displayed as the evaluation index 203 on the first evaluation screen DSb in step S26.

[0034] According to the first embodiment described above, the analysis system 1 includes the reception unit 112, the acquisition unit 114, the estimation unit 116, and the display unit 150. The acquisition unit 114 creates a machine learning model 124, inputs observation data to the created machine learning model 124, and acquires a predicted value of the additional variable mb. The estimation unit 116 estimates the second causal structure CAb using combined data obtained by combining the observation data, the predicted value, and the additional data. The display unit 150 displays a structure display screen DSa that displays the second causal structure CAb in a display form that allows the difference between the first causal structure CAa and the second causal structure CAb to be identified. As a result, even if the amount of additional data is small, when the second causal structure CAb is obtained using the additional data, the reliability of the causal analysis when the second causal structure CAb is obtained can be improved by using the observation data and the predicted value obtained using the created machine learning model 124 in addition to the additional data 126. The second causal structure CAb thus obtained is displayed in a manner that allows the difference between the first causal structure CAa and the second causal structure CAb to be identified, thereby enabling the user to efficiently evaluate the second causal structure CAb.

[0035] Furthermore, the display unit 150 displays the second causal structure CAb in a causal diagram on the structure display screen DSa shown in Fig. 8, and displays nodes corresponding to variables that differ between the first variable EVa and the second variable EVb in a display mode different from that of nodes corresponding to variables that are common to the first variable EVa and the second variable EVb. This allows the user to immediately recognize the difference between the first causal structure CAa and the second causal structure CAb by looking at the structure display screen DSa, and therefore allows the user to efficiently evaluate the second causal structure CAb.

[0036] Furthermore, the display unit 150 displays the second causal structure CAb in a causal diagram on the structure display screen DSa shown in Fig. 9, and also displays text information 201 that identifies the different variables between the first variable EVa and the second variable EVb. This allows the user to immediately recognize the difference between the first causal structure CAa and the second causal structure CAb by looking at the structure display screen DSa, and therefore allows the user to efficiently evaluate the second causal structure CAb.

[0037] Furthermore, the estimation unit 116 estimates the second causal structure CAb when the learning error obtained by the acquisition unit 114 is equal to or less than the threshold value, and does not estimate the second causal structure CAb when the learning error is greater than the threshold value. Then, in step S28, the display unit 150 displays information indicating that the second causal structure CAb has not been estimated when the estimation unit 116 has not estimated the second causal structure CAb. By estimating the second causal structure CAb when the learning error obtained by the acquisition unit 114 is equal to or less than the threshold value, the second causal structure CAb can be estimated when the additional variable mb is highly likely to appear in the second causal structure CAb and the significance of estimating the second causal structure CAb is great. On the other hand, when the additional variable mb is less likely to appear in the second causal structure CAb and the significance of estimating the second causal structure CAb is small, the estimation of the second causal structure CAb is not performed, thereby reducing the processing load performed by the estimation unit 116. Furthermore, the display unit 150 displays information indicating that the second causal structure CAb has not been inferred, thereby enabling the user to recognize that the second causal structure CAb has not been inferred.

[0038] The display unit 150 also displays a first evaluation screen DSb that displays the first evaluation graph 202 and the evaluation index 203. This allows the user to evaluate the prediction accuracy of the machine learning model 124. The display unit 150 also displays a second evaluation screen DSc that displays the second evaluation graph 204. This allows the user to check the correlation between the explanatory variables and the objective variables of the machine learning model 124.

[0039] B. Second embodiment: FIG. 10 is a flowchart showing the procedure of the causal analysis process according to the second embodiment. In this embodiment, when creating the machine learning model 124, the importance of the explanatory variables in the machine learning model 124 is obtained. Then, when the first edge, which is an edge connecting the first node, which is a node corresponding to the important variable, which is a variable with high importance, and the second node, which is a node corresponding to the additional variable mb, does not appear in the second causal structure CAb, the first edge is displayed in a display mode different from the display mode of the edges excluding the first edge among the multiple edges displayed on the structure display screen DSa. This allows the user to recognize that the variable with high importance does not appear in the second causal structure CAb. The same processing steps as those in the first embodiment are given the same reference numerals, and detailed explanations are omitted as appropriate.

[0040] 10, the receiving unit 112 receives an input of a first causal structure CAa calculated using observation data having a value of an observation variable ma, and additional data 126. In this embodiment, in step S11, the value of a parameter N used in step S20, which will be described later, is also received. Step S11 is a processing step similar to step S10 in the first embodiment, and the receiving unit 112 receives the first causal structure CAa and the like, for example, via a GUI.

[0041] In step S12, the acquisition unit 114 creates a machine learning model 124 that receives the first variable EVa as an input and outputs an additional variable mb. When creating the machine learning model 124, the acquisition unit 114 obtains a numerical value that indicates the importance of the first variable EVa, i.e., the variable input to the machine learning model 124, in the machine learning model 124. In this embodiment, a permutation feature importance (PFI) is used as the numerical value that indicates the importance. The larger the PFI numerical value, the higher the importance.

[0042] In step S14, the acquisition unit 114 inputs the observation data to the machine learning model 124 created in step S14, and acquires a predicted value of the additional variable mb. In step S18, the estimation unit 116 estimates a second causal structure CAb, which is a new causal structure, by using combined data that combines the observation data, the predicted value, and the additional data 126. Note that the causal analysis algorithm used to estimate the second causal structure CAb may be the same as or different from the causal analysis algorithm used to estimate the first causal structure CAa.

[0043] In step S20 as the display step and display process, the output unit 118 judges whether or not there is an edge connecting the node of the important variable and the node of the additional variable mb in the second causal structure CAb. Here, the important variables are explanatory variables up to Nth in order of the importance of PFI of each explanatory variable output in the process of creating the machine learning model 124 in step S12. "N" is an integer of 1 or more. In this embodiment, the value of the parameter N accepted in step S11 is used as the value of "N". In another embodiment, a predetermined fixed value may be used as the value of "N". When the value of "N" is "1", if there is no edge connecting the node of the important variable and the node of the additional variable mb, it is determined in step S20 that there is no edge connecting the node of the important variable and the node of the additional variable mb. When the value of "N" is an integer of 2 or more, if there is one or more nodes of important variables that are not connected to the node of the additional variable mb among the multiple important variables, it is determined in step S20 that there is no edge connecting the node of the important variable and the node of the additional variable mb.

[0044] If it is determined in step S20 that there is an edge connecting the node of the important variable and the node of the added variable mb, then in step S26, the display unit 150 displays a structure display screen DSa displaying the second causal structure CAb.

[0045] If it is determined in step S20 that there is no edge connecting the node of the important variable and the node of the added variable mb, the display unit 150 adds an edge connecting the node of the important variable and the node of the added variable mb to the second causal structure CAb displayed on the structure display screen DSa and displays it in step S30. Regarding step S30, in detail, the output unit 118 creates structure display screen data for displaying the structure display screen DSa and the nodes and edges to be added, and outputs it to the display unit 150. As a result, the structure display screen DSa is displayed on the display unit 150.

[0046] In the structure display screen DSa displayed in step S30, the edges connecting the nodes of the important variables and the nodes of the added variables mb are displayed in a display mode different from that of the edges appearing in the second causal structure CAb.

[0047] Step S30 will be described with reference to FIG. 11. FIG. 11 is a diagram for explaining a structure display screen DSa displayed when an edge connecting an important variable and an additional variable does not appear when the value of "N" is "1". The upper part of FIG. 11 is a causal diagram of the second causal structure CAb in which an edge connecting an important variable and an additional variable does not appear. The lower part of FIG. 11 is a structure display screen DSa displayed in step S30. Here, a case in which "variable b", which is a variable of the second causal structure CAb, is an important variable will be described as an example. As shown in the upper part of FIG. 11, in the second causal structure CAb, an edge connecting the node of "variable b", which is an important variable, and the node of "additional variable a" does not appear. Therefore, in step S30, a causal diagram of the second causal structure CAb in which the node of "variable b" and the node of "additional variable a" are connected by a dashed edge is displayed.

[0048] 11, the node corresponding to the important variable "variable b" is also called the first node. Also, the node corresponding to the additional variable "additional variable a" which is the additional variable mb is also called the second node.

[0049] Furthermore, in the structure display screen DSa displayed in step S30, the display mode of the edge connecting "variable b" and "additional variable a" is different from the display mode of the other edges. Here, the edge connecting the important variable and the additional variable is referred to as the first edge EG1. The first edge EG1 is displayed in a display mode different from the display mode of the other edges other than the first edge EG1. In this embodiment, in the structure display screen DSa, edges other than the first edge EG1 are displayed with a solid line, and the first edge EG1 is displayed with a dashed line. The difference in display mode is not limited to the difference in line type, and for example, the colors of the edges may be different from each other.

[0050] The important variables are strongly correlated with the additional variables. Therefore, if an edge connecting the important variables and the additional variables does not appear in the second causal structure CAb, there is a high possibility that the second causal structure CAb is invalid. Therefore, if the first edge EG1 does not appear in the second causal structure CAb, the first edge EG1 is displayed in a different display mode from the display mode of edges other than the first edge EG1, so that the user can recognize that the important variables do not appear in the second causal structure CAb. After the completion of step S26 and the completion of step S30, this processing routine ends.

[0051] According to the second embodiment described above, when the first edge EG1 does not appear in the second causal structure CAb, the display unit 150 displays the first edge EG1 on the structure display screen DSa in a display mode different from the display mode of edges other than the first edge EG1. This allows the user to recognize that the important variable does not appear in the second causal structure CAb. Therefore, the user can efficiently evaluate the validity of the second causal structure CAb.

[0052] C. Other embodiments: (C1) In the above first embodiment, in step S18, the estimation unit 116 estimates the second causal structure CAb, which is a new causal structure, by using data obtained by combining the observed data, the predicted value, and the additional data 126. As another embodiment, the second causal structure CAb, which is a new causal structure, may be estimated by using data obtained by combining only the value of the first variable EVa, of the observed data, the predicted value, and the additional data 126. According to this embodiment, the processing load can be reduced.

[0053] (C2) In the first embodiment, in step S16, when the learning error is equal to or smaller than the threshold, the second causal structure CAb is estimated. In another embodiment, the second causal structure CAb may be estimated regardless of whether the learning error is equal to or smaller than the threshold.

[0054] (C3) In the first embodiment, the machine learning model 124 used in step S12 is created using one algorithm. In another embodiment, the machine learning model 124 may be created using a plurality of algorithms, and step S14 may be performed using the machine learning model 124 created using the algorithm with the best performance. The performance is, specifically, the coefficient of determination (R 2 ), root mean square error (RMSE), AUC, F-measure, etc.

[0055] (C4) In the above-mentioned first embodiment, the structure display screen DSa, the first evaluation screen DSb, and the second evaluation screen DSc are displayed on the display unit 150. The present invention is not limited to this, and may be configured such that only the structure display screen DSa is displayed on the display unit 150. By displaying at least the structure display screen DSa, the user can evaluate the second causal structure CAb. In addition, in the above-mentioned first embodiment, both the first evaluation graph 202 and the evaluation index 203 are displayed on the first evaluation screen DSb. The present invention is not limited to this, and only one of the first evaluation graph 202 and the evaluation index 203 may be displayed on the first evaluation screen DSb. Even when only one of the first evaluation graph 202 and the evaluation index 203 is displayed, the user can evaluate the prediction accuracy of the machine learning model 124.

[0056] (C5) In the above first embodiment, the structure display screen DSa is displayed on the display unit 150 connected to the information processing device 100 via the interface circuit 130. As another form of the display unit 150 on which the structure display screen DSa is displayed, it may be a display unit included in an information processing device or display device that is capable of communicating through a network to which the communication interface 132 is connected.

[0057] D. Other forms: The present disclosure is not limited to the above-mentioned embodiment, and can be realized in various configurations without departing from the spirit of the present disclosure. For example, the technical features of the embodiments corresponding to the technical features in each form described below can be appropriately replaced or combined in order to solve some or all of the above-mentioned problems or to achieve some or all of the above-mentioned effects. Furthermore, if a technical feature is not described as essential in this specification, it can be appropriately deleted.

[0058] (1) According to a first aspect of the present disclosure, there is provided an analysis system for causal analysis, comprising: a receiving unit that receives input of a first causal structure determined using observation data having values ​​of observed variables, and additional data having values ​​of additional variables different from the observed variables and the observed variables, an acquisition unit that creates a machine learning model that uses the value of the first variable in the additional data as input data and the value of the additional variable in the additional data as ground truth data, and outputs the value of the additional variable using the first variable as an input, inputs the observed data into the created machine learning model, and acquires a predicted value of the additional variable, an estimation unit that estimates a second causal structure using data obtained by combining the observed data, the predicted value, and the additional data, and a display unit that displays at least the second causal structure in a display mode in which a difference between the first causal structure and the second causal structure can be identified. According to this embodiment, even if the amount of additional data is small, when the additional data is used to obtain a causal structure, the reliability of the causal analysis when obtaining the second causal structure can be improved by using the additional data, in addition to the observation data, and the predicted value obtained using the created machine learning model. In addition, when creating a machine learning model, overlearning can be suppressed by using the first variable, which is a part of the observation variables, as an input. Then, the second causal structure obtained in this way is displayed in a display mode that allows the difference between the first causal structure and the second causal structure to be distinguished, so that the user can efficiently evaluate the second causal structure.

[0059] (2) In the analysis system of the above embodiment, each of the second variable and the first variable appearing in the second causal structure may include a plurality of variables, and the display unit may display, on the structure display screen, a diagram in which each of the second variables is a node and a side connecting two of the nodes corresponding to any two of the second variables having a causal relationship is an edge, and the nodes corresponding to the variables that are different between the first variable and the second variable are displayed in a different display mode from the nodes corresponding to the variables that are common to the first variable and the second variable. According to this embodiment, a user can immediately recognize the difference between the first causal structure and the second causal structure by looking at the structure display screen, and therefore the second causal structure can be efficiently evaluated.

[0060] (3) In the analysis system of the above form, each of the second variables and the first variables appearing in the second causal structure includes a plurality of variables, and the display unit may display, on the structure display screen, a diagram in which each variable of the second variables is a node and a side connecting two of the nodes corresponding to any two of the second variables having a causal relationship is an edge, and may also display text information identifying the variables that are different between the first variable and the second variable. According to this form, a user can immediately recognize the difference between the first causal structure and the second causal structure by looking at the structure display screen, and therefore the second causal structure can be efficiently evaluated.

[0061] (4) In the analysis system of the above embodiment, the acquisition unit may obtain a numerical value indicating the magnitude of an error between a value of the additional variable in the additional data and a predicted value of the additional variable outputted by inputting the additional data to the machine learning model, and the estimation unit may estimate the second causal structure when the numerical value obtained by the acquisition unit is equal to or less than a predetermined threshold value, and may not estimate the second causal structure when the error is greater than the threshold value, and the display unit may display information indicating that the second causal structure has not been estimated when the estimation unit has not estimated the second causal structure. By estimating the second causal structure when the learning error obtained by the acquisition unit is equal to or less than a threshold value, the second causal structure can be estimated when the possibility that an additional variable appears in the second causal structure is high and the significance of estimating the second causal structure is great. On the other hand, when the possibility that an additional variable appears in the second causal structure is low and the significance of estimating the second causal structure is small, the estimation of the second causal structure is not performed, thereby reducing the processing load on the estimation unit. Furthermore, the display unit displays information indicating that the second causal structure has not been inferred, thereby enabling the user to recognize that the second causal structure has not been inferred.

[0062] (5) In the analysis system of the above aspect, the display unit may further display a first evaluation screen that displays at least one of (i) a graph showing a correlation between the value of the additional variable of the additional data and a predicted value of the additional variable outputted by inputting the additional data into the machine learning model, and (ii) an evaluation index of the machine learning model. According to this aspect, a user can evaluate the prediction accuracy of the machine learning model.

[0063] (6) In the analysis system of the above aspect, the display unit may further display a second evaluation screen that displays a graph showing a relationship between the value of the first variable of the observation data and a predicted value of the additional variable obtained by inputting the value of the first variable of the observation data to the machine learning model. According to this aspect, a user can check the correlation between the variable of the machine learning model and a target variable.

[0064] (7) In the analysis system of the above embodiment, each of the second variables and the first variables appearing in the second causal structure may include a plurality of variables, the acquisition unit may obtain a numerical value indicating the importance of each of the first variables in the machine learning model, and the display unit may display, in the structure display screen, a diagram in which each of the second variables is a node and an edge connecting two of the nodes corresponding to any two of the second variables having a causal relationship is an edge, and there are a plurality of edges, and when a first edge connecting a first node that is a node corresponding to an important variable that is one of the first variables up to N (N is a predetermined integer of 1 or more) in descending order of importance among the first variables and a second node that is a node corresponding to the additional variable does not appear in the second causal structure, the display unit may display the first edge in a display mode different from the display mode of the edges other than the first edge among the plurality of edges displayed on the structure display screen. According to this embodiment, a user can recognize that an important variable does not appear in the second causal structure. Thus, the user can efficiently evaluate the validity of the second causal structure.

[0065] In addition to the above aspects, the present disclosure can be realized in the form of an analysis method, an analysis program, and the like. [Explanation of symbols]

[0066] 1...analysis system, 100...information processing device, 110...processor, 112...reception unit, 114...acquisition unit, 116...estimation unit, 118...output unit, 120...memory, 124...machine learning model, 126...additional data, 130...interface circuit, 132...communication interface, 140...input device, 150...display unit, 160...storage medium, 201...text information, 202...first evaluation graph, 203...evaluation index, 204...second evaluation graph, CAa...first causal structure, CAb...second causal structure, DSa...structure display screen, DSb...first evaluation screen, DSc...second evaluation screen, EG1...first edge, EVa...first variable, EVb...second variable, ma...observation variable, mb...additional variable

Claims

1. 1. An analysis system for causal analysis, comprising: a receiving unit that receives an input of a first causal structure determined using observation data having values ​​of observation variables, and additional data having values ​​of additional variables different from the observation variables and the values ​​of the observation variables; an acquisition unit that uses the value of the first variable in the additional data as input data and the value of the additional variable in the additional data as correct answer data, creates a machine learning model that takes the first variable as an input and outputs the value of the additional variable, the value being a first variable that appears in the first causal structure, inputs the observation data into the created machine learning model, and acquires a predicted value of the additional variable; an estimation unit that estimates a second causal structure using data obtained by combining the observed data, the predicted value, and the additional data; an analysis system comprising: a display unit that displays a structure display screen that displays at least the second causal structure in a display manner that allows a difference between the first causal structure and the second causal structure to be identified.

2. The analysis system according to claim 1 , each of the second variable and the first variable appearing in the second causal structure includes a plurality of variables; The display unit displays, on the structure display screen, a diagram in which each variable of the second variables is a node and a side connecting two of the nodes corresponding to any two of the second variables that are causally related is an edge, and displays the nodes corresponding to the variables that are different between the first variable and the second variable in a display manner different from that of the nodes corresponding to the variables that are common to the first variable and the second variable.

3. The analysis system according to claim 1 , each of the second variable and the first variable appearing in the second causal structure includes a plurality of variables; The display unit displays, on the structure display screen, a diagram in which each variable of the second variables is a node and each side connecting two of the nodes corresponding to any two of the second variables that are causally related is an edge, and also displays text information that identifies the variables that are different between the first variable and the second variable.

4. The analysis system according to any one of claims 1 to 3, the acquiring unit calculates a numerical value indicating a magnitude of an error between a value of the additional variable of the additional data and a predicted value of the additional variable outputted by inputting the additional data into the machine learning model; the estimation unit estimates the second causal structure when the numerical value obtained by the acquisition unit is equal to or smaller than a predetermined threshold, and does not estimate the second causal structure when the error is greater than the threshold; The display unit, when the estimation unit does not estimate the second causal structure, displays information indicating that the second causal structure has not been estimated.

5. The analysis system according to any one of claims 1 to 3, The display unit further displays a first evaluation screen that displays at least one of (i) a graph showing the correlation between the value of the additional variable in the additional data and the predicted value of the additional variable outputted when the additional data is inputted into the machine learning model, and (ii) an evaluation index of the machine learning model.

6. The analysis system according to any one of claims 1 to 3, The display unit further displays a second evaluation screen that displays a graph showing the relationship between the value of the first variable of the observation data and the predicted value of the additional variable obtained by inputting the value of the first variable of the observation data into the machine learning model.

7. The analysis system according to claim 1 , each of the second variable and the first variable appearing in the second causal structure includes a plurality of variables; The acquisition unit obtains a numerical value indicating a degree of importance of each variable of the first variables in the machine learning model, The display unit is on the structure display screen, each variable of the second variables is displayed as a node, and a side connecting two of the nodes corresponding to any two variables having a causal relationship among the second variables is displayed as an edge in a diagram, the number of the edges being plural; In the case where a first edge that connects a first node that is the node corresponding to an important variable that is one of the first variables up to Nth (N is a predetermined integer of 1 or more) in order of importance, and a second node that is the node corresponding to the additional variable, does not appear in the second causal structure, the analysis system displays the first edge in a display mode different from the display mode of the edges excluding the first edge among the multiple edges displayed on the structure display screen.

8. 1. An analysis method for causal analysis, comprising: a receiving step of receiving an input of a first causal structure determined using observation data having values ​​of observed variables, and additional data having values ​​of additional variables different from the observed variables and the values ​​of the observed variables; an acquisition step of acquiring a predicted value of the additional variable by using the value of the first variable in the additional data as input data and the value of the additional variable in the additional data as correct answer data to create a machine learning model that takes the first variable as an input and outputs the additional variable, the observed data being a value of the first variable that appears in the first causal structure; an estimation step of estimating a second causal structure using data obtained by combining the observed data, the predicted value, and the additional data; an output step of outputting structure display screen data for displaying at least the second causal structure in a display form in which a difference between the first causal structure and the second causal structure can be identified.

9. An analysis program for causal analysis, a receiving process for receiving an input of a first causal structure determined using observation data having values ​​of observation variables, and additional data having values ​​of additional variables different from the observation variables and the values ​​of the observation variables; an acquisition process of creating a machine learning model that uses the value of the first variable in the additional data as input data and the value of the additional variable in the additional data as correct answer data, the value of the first variable being a value of a first variable that appears in the first causal structure, and that outputs the additional variable using the value of the first variable as an input, inputting the observation data into the created machine learning model, and acquiring a predicted value of the additional variable; an estimation process of estimating a second causal structure using data obtained by combining the observed data, the predicted value, and the additional data; and an output process for outputting structure display screen data for displaying at least the second causal structure in a display manner in which the difference between the first causal structure and the second causal structure can be identified.

Images