Analysis system, analysis method, and program

Abstract

This enables proper analysis of the relationship between the inputs and outputs of a product performance prediction model. [Solution] The analysis system 1 includes a tire performance prediction model storage unit 200 that stores a tire performance prediction model, an analysis specification data acquisition unit 202 that acquires multiple analysis specification data, a group setting unit 203 that sets up groups consisting of mutually related features from among multiple features, and a contribution calculation unit 2040 that calculates a SHAP value for each combination of feature values ​​included in each group based on the tire performance prediction model and the multiple analysis specification data.

Description

Technical Field

[0001] The present invention relates to an analysis system, an analysis method, and a program.

Background Art

[0002] As a technique for analyzing the relationship between the input (i.e., feature amount) and output (i.e., predicted value) of a machine learning model, XAI (eXplainable AI: Explainable AI) is known. According to XAI, it is possible to clarify the relationship between the feature amount and the predicted value, such as that when the value of feature amount A increases, the predicted value also increases, or when the value of feature amount A fluctuates, the predicted value also fluctuates greatly, while the predicted value hardly changes even when the value of feature amount B fluctuates. As XAI methods, for example, SHAP (SHapley Additive exPlanations) and PFI (Permutation Feature Importance) are known.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In conventional XAI, it may be difficult to appropriately analyze the relationship between the input and output of a machine learning model.

[0004] The present invention has been made in view of the above problems, and one of its objects is to provide an analysis system, an analysis method, and a program that enable appropriate analysis of the relationship between the input and output of a machine learning model.

Means for Solving the Problems

[0005] An analysis system according to one embodiment of the present invention includes: a product performance prediction model storage means that stores a product performance prediction model that is trained to accept specification data having a plurality of features as input and output a first performance index value indicating the performance of the product; an analysis specification data acquisition means that acquires a plurality of analysis specification data which are the specification data to be used for analysis; a group setting means that sets a first group consisting of mutually related features from the plurality of features; and a contribution calculation means that calculates a value indicating the contribution of each combination of feature values ​​included in the first group to the output of the first performance index value by the product performance prediction model, based on the product performance prediction model and the plurality of analysis specification data. [Brief explanation of the drawing]

[0006] [Figure 1] This figure shows an example of the hardware configuration of an analysis system according to an embodiment of the present invention. [Figure 2] A functional block diagram showing an example of the functions realized in the analysis system according to an embodiment of the present invention. [Figure 3] This diagram shows an example of a tire structure. [Figure 4] This diagram shows an example of the hierarchical relationship between the multiple components of a tire. [Figure 5] This figure shows an example of a component grouping table. [Figure 6] This figure shows an example of a feature grouping table. [Figure 7] This figure shows an example of multiple analysis specification data. [Figure 8] This figure shows an example of a group trend graph. [Figure 9] This figure shows an example of a group trend graph created for each performance indicator value. [Figure 10] This figure shows an example of a feature trend graph. [Figure 11] This figure shows an example of how an importance graph can be displayed. [Figure 12A] A flowchart illustrating an example of the processing performed by the analysis system according to an embodiment of the present invention. [Figure 12B] A flowchart illustrating an example of the processing performed by the analysis system according to an embodiment of the present invention. [Figure 12C] A flowchart illustrating an example of the processing performed by the analysis system according to an embodiment of the present invention. [Figure 12D] A flowchart illustrating an example of the processing performed by the analysis system according to an embodiment of the present invention. [Figure 13] This diagram illustrates how to set up groups in a modified example. [Modes for carrying out the invention]

[0007] Embodiments of the present invention will be described below with reference to the drawings.

[0008] [1. Hardware configuration of the analysis system] Figure 1 shows an example of the hardware configuration of an analysis system 1 according to an embodiment of the present invention. In this embodiment, the case in which the analysis system 1 is implemented by a single computer is illustrated, but the analysis system 1 may be implemented by multiple computers connected to each other via a network such as the Internet or a LAN.

[0009] As shown in Figure 1, the analysis system 1 includes a control unit 10, a storage unit 12, a communication unit 14, a display unit 16, and an operation unit 18. The control unit 10 is a program control device such as a CPU that operates according to a program stored in the storage unit 12. The storage unit 12 is a storage element such as ROM or RAM or a hard disk drive. The storage unit 12 stores programs executed by the control unit 10. The communication unit 14 is a communication interface such as a network board or a wireless LAN module. The display unit 16 is a display such as a liquid crystal display or an organic EL display. The operation unit 18 is an input device such as a keyboard, mouse, or touch panel.

[0010] The program stored in the storage unit 12 may be supplied via a network. Alternatively, the storage unit 12 may include a reading unit (e.g., a memory card slot) for reading computer-readable information storage media or an input / output unit (e.g., a USB terminal) for connecting to external devices. In this case, the program stored on the information storage media may be supplied via the reading unit or input / output unit.

[0011] [2. Functions implemented by the analysis system] Figure 2 is a functional block diagram showing an example of the functions realized in the analysis system 1 according to an embodiment of the present invention. As shown in Figure 2, the analysis system 1 functionally includes a tire performance prediction model storage unit 200, a component information storage unit 201, an analysis specification data acquisition unit 202, a group setting unit 203, a SHAP value calculation unit 204, a group contribution data acquisition unit 205, a group trend graph creation unit 206, a feature quantity contribution data acquisition unit 207, a feature quantity trend graph creation unit 208, an importance graph creation unit 209, and an importance graph display unit 210.

[0012] The tire performance prediction model storage unit 200 stores a tire performance prediction model (product performance prediction model). The tire performance prediction model is a trained model that accepts specification data, which indicates the specifications of a tire and has multiple feature quantities, as input, and outputs a first performance index value S and a second performance index value T that indicate the performance of the tire.

[0013] Each of the multiple features in the specification data may include, for example, the dimensions of the tire (nominal width, aspect ratio, inner diameter, etc.), the physical properties of the materials constituting each component of the tire, and the manufacturing conditions of the tire (vulcanization temperature, molding machine drum width, etc.). Note that the multiple features may also include tire test conditions (air pressure, load, rim width, road surface condition, etc.).

[0014] In this embodiment, as an example, a case where the first performance index value S is the rolling resistance (RRC) and the second performance index value T is the cornering power is illustrated. However, the first performance index value S and the second performance index value T are not limited to the above example and may be any values indicating the performance of the tire. For example, the first performance index value S and the second performance index value T may be the longitudinal (lateral) spring or the endurance time under predetermined conditions, etc.

[0015] In this embodiment, as an example, a case where the tire performance prediction model is a neural network is illustrated. However, the tire performance prediction model is not limited to the above example and may be any machine learning model. For example, the tire performance prediction model may be a model based on non-linear regression other than neural networks such as regression trees. Alternatively, the tire performance prediction model may be a model based on linear regression such as ridge regression, lasso regression, elastic net, etc.

[0016] Here, before explaining the details of the member information storage unit 201, the structure of the tire will be explained using FIGS. 3 and 4. FIG. 3 is a diagram showing an example of the structure of the tire 3. FIG. 4 is a diagram showing an example of the hierarchical relationship between a plurality of members included in the tire 3.

[0017] As shown in FIG. 3, the tire 3 includes a tread rubber 30, a belt cover 32, a belt 34, and a carcass 36. The tread rubber 30 is a rubber layer that contacts the road surface. The belt cover 32 is a member for tightening and reinforcing the belt 34. The belt 34 is a member for tightening the carcass. The carcass 36 is a member that forms the skeleton of the tire and is made of a material such as polyester or rayon, for example. Note that the members shown in FIG. 3 are only a part of the plurality of members included in the tire 3, and the tire 3 has a number of other members in addition to those shown in FIG. 3, such as bead fillers, bead wires, and sidewalls.

[0018] The multiple components of the tire 3 shown in Figure 3 have a hierarchical relationship, as shown in Figure 4. For example, the belt cover 32 includes four types of components: JE1, JE2, JF1, and JF2. The belt 34 includes two types of components: 1B and 2B. The carcass 36 also includes two types of components: 1P and 2P.

[0019] Note that the belt cover 32 is not necessarily made up of all of JE1, JE2, JF1, and JF2. In reality, there are fixed patterns for the combinations of the constituent members JE1, JE2, JF1, and JF2, such as patterns consisting of JE1, patterns consisting of JE1 and JE2, patterns consisting of JF1 and JE1, and patterns consisting of JE1, JF1, and JF2.

[0020] Furthermore, if JE1 is not included in the belt cover 32, then JE2 is not included in the belt cover 32, and if JF1 is not included in the belt cover 32, then JF2 is not included in the belt cover 32. In addition, there are patterns in which the carcass 36 includes only 1P, and patterns in which the carcass 36 includes both 1P and 2P.

[0021] Returning to Figure 2, the component information storage unit 201 stores the relationship between each of the multiple components of the tire and the multiple constituent components. Specifically, the component information storage unit 201 stores the constituent component grouping table T1 shown in Figure 5 and the feature quantity grouping table T2 shown in Figure 6. Figure 5 is a diagram showing an example of the constituent component grouping table T1. Figure 6 is a diagram showing an example of the feature quantity grouping table T2.

[0022] The component grouping table T1 is a table that stores the relationships between multiple members and multiple components. For example, as shown in Figure 5, the component grouping table T1 stores the relationship between the belt cover and the components that make up the belt cover, JE1, JE2, JF1, and JF2 (see Figures 3 and 4). The same applies to other members and components.

[0023] The feature grouping table T2 is a table that stores multiple components in association with multiple feature quantities related to each component. For example, as shown in Figure 6, the feature grouping table T2 stores JE1 in association with multiple feature quantities related to JE1 (hardness, TB, width, length, etc.). The same applies to other components and feature quantities.

[0024] The analysis specification data acquisition unit 202 acquires multiple analysis specification data, each of which is specification data to be used for analysis. Figure 7 shows an example of multiple analysis specification data.

[0025] As shown in Figure 7, the analysis specification data has multiple features. These multiple features include multiple features related to the belt cover. Specifically, as shown in Figure 7, the multiple features include multiple features related to JE1 (JE1 hardness, JE1TB, etc.), multiple features related to JE2 (JE2 hardness, JE2TB, etc.), multiple features related to JF1 (JF1 hardness, JF1TB, etc.), and multiple features related to JF2 (JF2 hardness, JF2TB, etc.). Note that the multiple features related to the belt cover shown in Figure 7 are only a part of the multiple features in the analysis specification data. In Figure 7, some other features (for example, features related to the belt, features related to the carcass, etc.) are not shown, but please note that the analysis specification data actually includes some of these other features.

[0026] Each analysis specification data represents one of several product specifications (Specification 1 to Specification n). For example, Specification 1 is the specification for a tire with a belt cover consisting of JE1 and JF1. Specification 2 is the specification for a tire without a belt cover. Specification n is the specification for a tire with a belt cover consisting of JE1, JF1, and JF2.

[0027] In each of the multiple analysis specification data sets, all of the multiple feature quantities related to each of the multiple component members are set to values ​​that indicate the existence of that component member in the specification shown in the analysis specification data. For example, as shown in Figure 7, in the analysis specification data showing specification 1, all of the multiple feature quantities related to JE1 are set to significant values ​​such as "63", "18", and "24", that is, values ​​that indicate the existence of JE1 in specification 1. Also, in the analysis specification data showing specification 1, all of the multiple feature quantities related to JF1 are set to significant values ​​such as "63", "18", and "215", that is, values ​​that indicate the existence of JF1 in specification 1.

[0028] Furthermore, in each of the multiple analysis specification data sets, all of the multiple feature quantities related to each of the multiple component members are set to a value indicating that the component member does not exist in the specification shown in the analysis specification data. For example, as shown in Figure 7, in the analysis specification data showing specification 1, all of the multiple feature quantities related to JE2 are set to a value ("none") indicating that JE2 does not exist in specification 1. The value indicating that JE2 does not exist in specification 1 may be, for example, 0 or a missing value (null value).

[0029] The group setting unit 203 refers to the contents stored in the member information storage unit 201 and sets up multiple groups corresponding to each of the multiple members. Each of the multiple groups set up consists of multiple feature quantities relating to multiple components associated with the member corresponding to that group. For example, the group setting unit 203 refers to the contents stored in the member information storage unit 201 and sets up group A (first group) corresponding to the belt cover. Group A consists of multiple feature quantities (JE1 hardness to JF2 length; see Figure 7) relating to multiple components JE1, JE2, JF1 and JF2 associated with the belt cover. The group setting unit 203 similarly sets up group B (second group), group C, ... corresponding to other members not shown. Note that there may be only one group.

[0030] The SHAP value calculation unit 204 includes a contribution calculation unit 2040 and a group importance calculation unit 2042.

[0031] The contribution calculation unit 2040 calculates a SHAP value for each combination of feature values ​​included in Group A, based on the tire performance prediction model and multiple analysis specification data. More specifically, the contribution calculation unit 2040 calculates a SHAP value for each combination of feature values ​​that make up Group A, based on the tire performance prediction model and multiple analysis specification data. The calculated SHAP value indicates the contribution of the combination of feature values ​​to the output of the performance index value by the tire performance prediction model. Each combination of feature values ​​included in Group A is specifically the combination of feature values ​​included in Group A in each specification. For example, the contribution calculation unit 2040 calculates a SHAP value for each of the combinations of feature values ​​included in Group A in Specification 1 (63,...,18,24,1985,None,...), the combination of feature values ​​included in Group A in Specification 2 (None,None,...), and the combination of feature values ​​included in Group A in Specification n (50,...,15,22,1975,None,...). The contribution calculation unit 2040 can also calculate the SHAP value for each analysis specification data (for each specification).

[0032] The contribution calculation unit 2040 may calculate SHAP values ​​based on a tire performance prediction model and multiple analysis specification data using a known SHAP algorithm. In known SHAP algorithms, the SHAP value of a feature is calculated by varying the value of each feature, but in this embodiment, the SHAP value of a group can be calculated by varying the values ​​of all the features included in that group.

[0033] The contribution calculation unit 2040 similarly calculates SHAP values ​​for each combination of feature values ​​included in the other groups (group B, group C, ...).

[0034] In this embodiment, the tire performance prediction model outputs a first performance index value S and a second performance index value T. Therefore, the contribution calculation unit 2040 calculates SHAP values ​​for each combination of feature values ​​included in each group, corresponding to the first performance index value S and the second performance index value T.

[0035] The group importance calculation unit 2042 calculates the importance of group A in the output of performance index values ​​by the tire performance prediction model, based on the SHAP values ​​calculated for each combination of feature values ​​included in group A. Specifically, the group importance calculation unit 2042 calculates the importance of group A as the average of the absolute values ​​of the SHAP values ​​calculated for each combination of feature values ​​included in group A.

[0036] The group importance calculation unit 2042 similarly calculates the importance of each group in the output of performance index values ​​by the tire performance prediction model for the other groups (Group B, Group C, ...).

[0037] In this embodiment, the tire performance prediction model outputs a first performance index value S and a second performance index value T, so the group importance calculation unit 2042 calculates the importance of each group for the first performance index value S and the second performance index value T, respectively.

[0038] The group contribution data acquisition unit 205 acquires group contribution data that associates each combination of feature values ​​included in group A with the SHAP value calculated for that combination of feature values. Since each combination of feature values ​​included in group A is a combination of feature values ​​included in group A in each specification, the number of group contribution data is the same as the number of specifications. The group contribution data acquisition unit 205 can also acquire group contribution data for each analysis specification data (each specification). The group contribution data acquisition unit 205 similarly acquires group contribution data for other groups.

[0039] The group trend graph creation unit 206 creates a group trend graph GG based on the group contribution data for each combination of feature values ​​included in group A. Figure 8 shows an example of the group trend graph GG. As shown in Figure 8, the group trend graph shows how the performance index values ​​change when the combination of feature values ​​included in group A changes. Specifically, the group trend graph GG is a graph in which group contribution data is plotted with the combination of feature values ​​included in group A (A6, A8, A4, ...) on the horizontal axis (first axis) and the SHAP value for each combination of feature values ​​included in group A on the vertical axis (second axis). The group trend graph creation unit 206 creates the group trend graph GG by plotting the group contribution data in ascending order of SHAP values. Alternatively, the group trend graph creation unit 206 may create the group trend graph GG by plotting the group contribution data in descending order of SHAP values.

[0040] In this embodiment, the tire performance prediction model outputs a first performance index value S and a second performance index value T, so the group trend graph creation unit 206 creates a group trend graph GG for each of the first performance index value S and the second performance index value T.

[0041] Specifically, the group trend graph creation unit 206 creates a first group trend graph GG1 corresponding to the first performance index value S and a second group trend graph GG2 corresponding to the second performance index value T, as shown in Figure 9. The first group trend graph GG1 is a scatter plot of SHAP values ​​for the first performance index value S for each combination of feature values ​​for each group. The second group trend graph GG2 is a scatter plot of SHAP values ​​for the second performance index value T for each combination of feature values ​​included in each group. Note that in Figure 9, only graphs relating to groups A and B are shown for both the first group trend graph GG1 and the second group trend graph GG2, and graphs relating to other groups are omitted.

[0042] In the example in Figure 9, the smaller the value of the features included in Group A, the larger both the SHAP value of Group A in the output of the first performance index S and the SHAP value of Group A in the output of the second performance index T become. From this, it can be seen that if you want to create a tire in which both the first performance index S and the second performance index T are large, you should reduce the value of Group A in the tire specifications.

[0043] On the other hand, the smaller the value of the features included in group A, the larger the SHAP value of group A in the output of the first performance index value S, while the smaller the SHAP value of group A in the output of the second performance index value T. From this, it can be seen that if you want to create a tire in which both the first performance index value S and the second performance index value T are large, it is not desirable to make the value of group B in the tire specifications extremely large or small.

[0044] The feature contribution data acquisition unit 207 acquires feature contribution data that associates each feature value included in group A with the SHAP value calculated for the combination of feature values ​​that includes the value of that feature. For example, the feature contribution data acquisition unit 207 acquires feature contribution data that associates the JE1 hardness value "63" (the JE1 hardness value in specification 1) included in group A with the SHAP value calculated for the combination of feature values ​​that includes that value (combination of feature values ​​included in group A in specification 1 (63,..., 18, 24, 1985, none,...)). The feature contribution data acquisition unit 207 can also acquire feature contribution data that associates each value of each feature in each analysis specification data (each specification) included in group A with the SHAP value calculated for that analysis specification data (each specification). The feature contribution data acquisition unit 207 similarly acquires feature contribution data for other groups.

[0045] The feature trend graph creation unit 208 creates a feature trend graph FG based on the feature contribution data for each feature value included in group A. Figure 10 shows an example of the feature trend graph FG. As shown in Figure 10, the feature trend graph FG shows how the performance index value changes when the value of the feature included in group A changes. Specifically, the feature trend graph FG is a graph in which the feature contribution data is plotted with the values ​​of the features included in group A (feature a, feature b) on the horizontal axis (first axis) and the SHAP value of the combination of feature values ​​included in group A on the second axis. The feature trend graph creation unit 208 creates the feature trend graph FG by plotting the feature contribution data in ascending order of feature values. Alternatively, the feature trend graph creation unit 208 may create the feature trend graph FG by plotting the feature contribution data in descending order of feature values.

[0046] In this embodiment, the tire performance prediction model outputs a first performance index value S and a second performance index value T, so the feature trend graph creation unit 208 creates a feature trend graph FG for each of the first performance index value S and the second performance index value T (not shown).

[0047] The importance graph creation unit 209 creates the importance graphs (first bar graph BG1 and second bar graph BG2) shown in Figure 11. Figure 11 is a diagram showing an example of how the importance graphs are displayed. The first bar graph BG1 corresponds to the first performance index value S. As shown in Figure 11, the first bar graph BG1 shows the importance of each of the multiple groups (group A, group B, ...) in the output of the first performance index value S by the tire performance prediction model. The second bar graph BG2 corresponds to the second performance index value T. As shown in Figure 11, the second bar graph BG2 shows the importance of each of the multiple groups (group A, group B, ...) in the output of the second performance index value T by the tire performance prediction model.

[0048] The importance graph display unit 210 displays the first bar graph BG1 and the second bar graph BG2 in such a way that the importance displayed for each of the multiple groups in the first bar graph BG1 corresponds to the importance displayed for the same group in the second bar graph BG2. For example, as shown in Figure 11, the importance graph display unit 210 displays the first bar graph BG1 and the second bar graph BG2 in such a way that the bar graph of group A in the first bar graph BG1 and the bar graph of group A in the second bar graph BG2 are aligned on the same axis.

[0049] In the example in Figure 11, Group A is highly important for both the first performance metric S and the second performance metric T, indicating that Group A is important for both the output of the first performance metric S and the output of the second performance metric T. On the other hand, Group B is highly important for the first performance metric S, but less important for the second performance metric T. From this, it can be seen that Group B is important only for the output of the first performance metric S.

[0050] [3. Processes executed by the analysis system] Figures 12A to 12D are flowcharts showing an example of processing performed by the analysis system 1 according to an embodiment of the present invention. The processing shown in Figures 12A to 12D is performed by the control unit 10 operating according to a program stored in the storage unit 12.

[0051] As shown in Figure 12A, first the control unit 10 acquires multiple analysis specification data (S1200). Next, the control unit 10 refers to the contents of the component information storage unit 201 and sets up multiple groups corresponding to each of the multiple components (S1202). Based on the tire performance prediction model and the multiple analysis specification data, the control unit 10 calculates a SHAP value for each combination of feature values ​​included in each group (S1204). Furthermore, based on the SHAP value calculated for each combination of feature values ​​included in each group, the control unit 10 calculates the importance of each group in the output of performance index values ​​by the tire performance prediction model (S1206) and terminates the process.

[0052] Moving to Figure 12B, the control unit 10 acquires group contribution data (S1208). The control unit 10 sorts the group contribution data in ascending order of SHAP values ​​(S1210), creates a group trend graph GG by plotting the group contribution data (S1212), and then terminates the process. The group trend graph GG created in S1212 is displayed on the display unit 16.

[0053] Moving to Figure 12C, the control unit 10 acquires feature contribution data (S1214). The control unit 10 sorts the feature contribution data in ascending order of feature values ​​(S1216), creates a feature trend graph FG by plotting the feature contribution data (S1218), and then terminates the process. The feature trend graph FG created in S1218 is displayed on the display unit 16.

[0054] Moving to Figure 12D, the control unit 10 creates the first bar graph BG1 and the second bar graph BG2 (S1220). The control unit 10 displays the created first bar graph BG1 and the second bar graph BG2 on the display unit 16 (S1222) and terminates the process. In S222, the control unit 10 displays the first bar graph BG1 and the second bar graph BG2 in such a way that the display of SHAP values ​​for each of the multiple groups in the first bar graph BG1 corresponds to the display of SHAP values ​​for the same groups in the second bar graph BG2 (see Figure 8).

[0055] [4. Summary] Grouped Permutation Feature Importance (GPFI) is a technique for calculating the importance of each group of related features in predictions made by a machine learning model. Conventional PFI does not consider the relationships between features and calculates the importance of each individual feature in predictions made by the machine learning model, which can lead to inaccurate calculations. GPFI takes the relationships between features into account, allowing for more accurate calculation of importance.

[0056] However, GPFI only allows us to understand how important each group is in the predictions made by the machine learning model; it does not allow us to understand the relationship between the feature values ​​included in each group and the predicted values ​​output by the machine learning model.

[0057] In this embodiment, for each combination of feature values ​​included in a group, a value is calculated that indicates the contribution of that combination of feature values ​​to the output of the performance index value by the tire performance prediction model. This makes it possible to understand the relationship between the feature values ​​included in a group and the performance prediction value output by the tire performance prediction model.

[0058] Furthermore, in this embodiment, a group trend graph is created that shows how the performance index value changes when the combination of feature values ​​included in the group changes, making it possible to identify combinations of feature values ​​that increase or decrease the performance index value.

[0059] Furthermore, in this embodiment, a feature trend graph is created for at least one feature included in the group, showing how the performance index value changes when the value of that feature changes. This makes it possible to understand the tendency for the performance index value to increase or decrease when the value of each feature included in the group is increased or decreased.

[0060] Incidentally, generally speaking, the components of products such as tires are made up by combining multiple components, and there are fixed patterns in how these components can be combined. For example, as mentioned above, the belt cover of a tire is made up by combining four types of components: JE1, JE2, JF1, and JF2. If JE1 is not included in the belt cover, then JE2 is not included, and if JF1 is not included in the belt cover, then JF2 is not included. In short, there are fixed patterns in how these components can be combined.

[0061] In XAI, including SHAP, a value indicating the contribution of a feature to the predicted output of a machine learning model is calculated by varying the value of that feature. However, when conventional XAI methods are applied to the analysis of a tire performance prediction model, when the features are varied, non-existent combinations of constituent elements may occur (for example, the value of JE2 is set even though the value of JE1 is not set). Based on data containing such non-existent combinations of feature values, it is difficult to accurately calculate the value indicating the contribution of the above features, and consequently, to accurately analyze the relationship between the input and output of the tire performance prediction model.

[0062] Therefore, in this embodiment, the contents of the component information storage unit 201, which stores the association between each of the multiple members and the multiple constituent members, are referenced to set up multiple groups corresponding to each of the multiple members, each consisting of multiple feature quantities related to the multiple constituent members associated with that member. As a result, non-existent combinations of constituent members as described above will not occur, and the relationship between the input and output of the tire performance prediction model can be accurately analyzed.

[0063] Furthermore, in this embodiment, the importance of each of the multiple groups in the first bar graph BG1 corresponding to the first performance index value S is displayed in correspondence with the importance of the same group in the second bar graph BG2 corresponding to the second performance index value T, and the first and second graphs are displayed accordingly. This makes it easy to understand whether each group is important to both the output of the first performance index value S and the output of the second performance index value T, or whether it is important to only the output of the first performance index value S or the output of the second performance index value T.

[0064] As described above, this embodiment makes it possible to appropriately analyze the relationship between the input and output of a machine learning model.

[0065] [5. Variant] Furthermore, the present invention is not limited to the embodiments described above. Also, the specific strings and numbers shown above and in the drawings are illustrative examples and the invention is not limited to these.

[0066] For example, in this embodiment, the product is shown as a tire, but the product is not limited to this example and may be any product having multiple components (e.g., an engine, a smartphone, a golf club, a bicycle, etc.).

[0067] Furthermore, while this embodiment illustrates a case where the tire performance prediction model outputs two values ​​(a first performance index value S and a second performance index value T), the tire performance prediction model may output only one performance index value. Alternatively, the tire performance prediction model may output three or more performance index values.

[0068] Furthermore, although this embodiment illustrates a case where the group setting unit 203 sets up multiple groups, the group setting unit 203 may set up only one group corresponding to a single component.

[0069] Furthermore, in this embodiment, we have illustrated the case where the SHAP value is the value that shows the contribution of each combination of feature values ​​to the output of the performance index value by the tire performance prediction model (i.e., the value that shows the contribution is the SHAP value). However, we are not limited to this example, and other algorithms used in XAI, such as PD (Partial Dependence) and ICE (Individual Conditional Expectation), may be used to calculate the value that shows the contribution of each combination of feature values ​​to the output of the performance index value by the tire performance prediction model.

[0070] When using PD and ICE, the value indicating the contribution of each combination of feature values ​​to the performance index value output by the tire performance prediction model is the performance index value itself. In this case, the group trend graph is a graph plotting the performance evaluation value output by the tire performance prediction model when analysis specification data including that combination of feature values ​​is input, for each combination of feature values ​​included in the group. The feature trend graph is a graph plotting the performance evaluation value output by the tire performance prediction model when analysis specification data including that feature value is input, for each value of each feature included in the group.

[0071] Furthermore, in this embodiment, the relationship between each of the multiple members and the multiple constituent members is stored in a table. However, for each feature of each analysis specification data, the constituent member related to that feature and the member related to that constituent member may be associated and stored. In this case, it is not necessary to prepare a table for setting up groups.

[0072] Furthermore, the group setting unit 203 may set groups without referring to the contents stored in the component information storage unit 201. As an example, the group setting unit 203 may set a group consisting of multiple feature quantities whose values ​​have a certain relationship with each other. Figure 13 is a diagram illustrating how to set groups in a modified example. As shown in Figure 13, for example, the group setting unit 203 may set a group consisting of 1B angle and 2B angle. The values ​​of 1B angle and 2B angle will be the same. Also, as shown in Figure 13, the group setting unit 203 may set a group consisting of 1B width and 2B width. The value of 1B width will be the value of 2B width plus 10.

[0073] The importance graph creation unit 209 and the importance graph display unit 210 may be configured as follows.

[0074] In other words, the importance graph creation unit 209 may create a first graph corresponding to each of the first performance index values, showing the contribution of each of the multiple features to the output of the first performance index value by the product performance prediction model, and may also create a second graph corresponding to each of the second performance index values, showing the contribution of each of the multiple features to the output of the second performance index value by the product performance prediction model. In this case, the importance graph display unit 210 may display the first graph and the second graph such that the display of the contribution of each of the multiple features in the first graph corresponds to the display of the contribution of each of the multiple features in the second graph. [Explanation of Symbols]

[0075] 1 Analysis system, 10 Control unit, 12 Storage unit, 14 Communication unit, 16 Display unit, 18 Operation unit, 200 Tire performance prediction model storage unit, 201 Component information storage unit, 202 Analysis specification data acquisition unit, 203 Group setting unit, 204 SHAP value calculation unit, 2040 Contribution calculation unit, 2042 Group importance calculation unit, 205 Group contribution data acquisition unit, 206 Group trend graph creation unit, 207 Feature quantity contribution data acquisition unit, 208 Feature quantity trend graph creation unit, 209 Importance graph creation unit, 210 Importance graph display unit, 3 Tire, 30 Tread rubber, 32 Belt cover, 34 Belt, 36 Carcass, T1 Component grouping table, T2 Feature quantity grouping table, FG Feature quantity trend graph, GG Group trend graph, BG1 First bar graph, BG2 Second bar graph.

Claims

1. A product performance prediction model storage means stores a product performance prediction model that has been trained to indicate the product specifications, accept specification data having multiple features as input, and output a first performance index value indicating the performance of the product. An analysis specification data acquisition means for acquiring multiple analysis specification data, each of which is specification data to be used for analysis, A group setting means for setting a first group consisting of mutually related features from the aforementioned plurality of features, Based on the product performance prediction model and the plurality of analysis specification data, a contribution calculation means calculates a value for each combination of feature value values ​​included in the first group that indicates the contribution of the combination of feature value values ​​to the output of the first performance index value by the product performance prediction model, An analytical system having the following features.

2. A group contribution data acquisition means for acquiring group contribution data, which associates each combination of feature values ​​included in the first group with a value indicating the contribution calculated for that combination of feature values; A group trend graph creation means creates a group trend graph that shows how the first performance index value changes when the combination of feature values ​​included in the first group changes, based on the group contribution data for each combination of feature values ​​included in the first group. It further has, The analysis system according to claim 1.

3. The group trend graph is a graph in which the group contribution data is plotted with the combination of feature values ​​included in the first group as the first axis and the value indicating the contribution of the combination of feature values ​​to the output of the first performance index value by the product performance prediction model as the second axis. The group trend graph creation means creates the group trend graph by plotting the group contribution data in ascending or descending order of the values ​​indicating the contribution. The analysis system according to claim 2.

4. A feature contribution data acquisition means for acquiring feature contribution data associated with each of the values ​​of at least one of the feature quantities included in the first group and a value indicating the contribution calculated for a combination of the feature values ​​including the value of the feature quantity, The system further includes a feature trend graph creation means that, based on the feature contribution data relating to each of the values ​​of at least one of the features included in the first group, creates a feature trend graph showing how the first performance index value changes when the value of at least one of the features included in the first group changes. The analysis system according to claim 1 or 2.

5. The feature trend graph is a graph in which the feature contribution data is plotted with the value of at least one feature on the first axis and the value showing the contribution of the combination of feature values ​​included in the first group to the output of the first performance index value by the product performance prediction model on the second axis. The feature trend graph creation means creates the feature trend graph by plotting the feature contribution data in ascending or descending order of the values ​​of at least one feature. The analysis system according to claim 4.

6. The value indicating the aforementioned contribution is the SHAP value. The analysis system according to claim 1 or 2.

7. The aforementioned product has multiple specifications, The aforementioned analysis system, Member information storage means for storing the association between each of the multiple members and the multiple constituent members. It further possesses, The group setting means refers to the contents of the member information storage means and sets the first group consisting of a plurality of feature quantities relating to a plurality of constituent members associated with one or more of the members. The analysis system according to claim 1 or 2.

8. The product performance prediction model is trained to further output a second performance index value indicating the performance of the product, The group setting means sets a second group which is different from the first group and consists of mutually related features from the plurality of features, The contribution calculation means is, Based on the product performance prediction model and the plurality of analysis specification data, for each combination of feature value values ​​included in the first group, a value is calculated that shows the contribution of the combination of feature value values ​​to the output of the first performance index value and the second performance index value, respectively, by the product performance prediction model. Based on the product performance prediction model and the plurality of analysis specification data, for each combination of feature value values ​​included in the second group, a value is calculated that indicates the contribution of the combination of feature value values ​​to the output of the first performance index value and the second performance index value, respectively, by the product performance prediction model. The aforementioned analysis system, A group importance calculation means calculates the importance of the first group in the output of the first and second performance indicators by the product performance prediction model, based on the values ​​indicating the contribution to the first and second performance indicators, respectively, calculated for each combination of feature values ​​included in the first group, and calculates the importance of the second group in the output of the first and second performance indicators by the product performance prediction model, based on the values ​​indicating the contribution to the first and second performance indicators, respectively, calculated for each combination of feature values ​​included in the second group, An importance graph creation means that creates a first importance graph corresponding to the first performance index value and showing the importance of the first group and the second group in the output of the first performance index value by the product performance prediction model, and creates a second importance graph corresponding to the second performance index value and showing the importance of the first group and the second group in the output of the second performance index value by the product performance prediction model, An importance graph display means for displaying the first importance graph and the second importance graph such that the display of the importance for the first group in the first importance graph corresponds to the display of the contribution for the first group in the second importance graph, and the display of the importance for the second group in the first importance graph corresponds to the display of the contribution for the second group in the second importance graph, It further has, The analysis system according to claim 1 or 2.

9. Product performance prediction model storage means that stores a product performance prediction model that has been trained to indicate the product specifications, accept specification data having multiple features as input, and output a first performance index value indicating the performance of the product. An analysis method performed by a computer having, An analysis specification data acquisition step involves acquiring multiple analysis specification data, each of which is specification data to be used for analysis, A group setting step of setting a first group consisting of mutually related features from the aforementioned plurality of features, A contribution calculation step, based on the product performance prediction model and the plurality of analysis specification data, calculates a value for each combination of feature value values ​​included in the first group that indicates the contribution of the combination of feature value values ​​to the output of the first performance index value by the product performance prediction model. An analytical method having the following characteristics.

10. Product performance prediction model storage means that stores a product performance prediction model that is trained to indicate the specifications of a product, accept specification data having multiple features as input, and output a first performance index value indicating the performance of the product. An analysis specification data acquisition means for acquiring multiple analysis specification data, each of which is specification data to be used for analysis, A group setting means for setting a first group consisting of mutually related features from among the aforementioned plurality of features, Based on the product performance prediction model and the plurality of analysis specification data, a contribution calculation means calculates a value for each combination of feature value values ​​included in the first group that indicates the contribution of the combination of feature value values ​​to the output of the first performance index value by the product performance prediction model. A program that makes a computer function.

11. A product performance prediction model storage means stores a product performance prediction model that has been trained to indicate the specifications of a product having multiple components, accept specification data having multiple feature quantities as input, and output performance index values ​​indicating the performance of the product. A component information storage means that stores the association between each of the aforementioned multiple components and the multiple constituent components, An analysis specification data acquisition means for acquiring multiple analysis specification data, each of which is specification data to be used for analysis, A group setting means that, by referring to the contents of the member information storage means, sets a group consisting of a plurality of feature quantities relating to a plurality of constituent members associated with one or more of the members, A group importance calculation means that calculates the importance of the group in the output of the performance index value by the product performance prediction model based on the product performance prediction model and the plurality of analysis specification data, An analytical system having the following features.

12. In each of the aforementioned multiple analysis specification data, For each of the multiple feature quantities relating to each of the multiple components, a value is set to indicate that the component exists in the specifications shown in the analysis specification data. Or, For each of the multiple feature quantities relating to each of the multiple components, a value is set to indicate that the component does not exist in the specifications shown in the analysis specification data. The analysis system according to claim 11.

13. A product performance prediction model storage means stores a product performance prediction model that has been trained to indicate the specifications of a product, accept specification data having multiple features as input, and output a first performance index value and a second performance index value indicating the performance of the product, respectively. An analysis specification data acquisition means for acquiring multiple analysis specification data, each of which is specification data to be used for analysis, Importance calculation means that calculates the importance of each of the multiple feature quantities in the output of the first performance index value and the second performance index value by the product performance prediction model, based on the product performance prediction model and the multiple analysis specification data, A graph creation means that creates a first graph corresponding to the first performance index value, showing the importance of each of the multiple features relative to the output of the first performance index value by the product performance prediction model, and a second graph corresponding to the second performance index value, showing the importance of each of the multiple features relative to the output of the second performance index value by the product performance prediction model, A graph display means for displaying the first graph and the second graph such that the display of the importance of each of the plurality of features in the first graph corresponds to the display of the importance of the feature in the second graph, An analytical system having the following features.

Images