Information processing device and information processing method

Abstract

An information processing device is provided that can calculate an interpretation index for each explanatory variable in the process of calculating a risk score, thereby reducing the calculation load. [Solution] An accounting audit support device 1 detects risks that can be detected from accounting-related data using a machine learning model. The system is equipped with a learning model generation unit 32 that uses a plurality of training records that combine a plurality of explanatory variables and a dependent variable selected from the financial data to generate, as a machine learning model, a shape function that indicates the relationship between each explanatory variable and dependent variable; a learning model storage unit 41 that stores the shape function for each explanatory variable generated by the learning model generation unit; and a risk calculation unit 34 that calculates an interpretation index for each explanatory variable using the shape function for a new record that is the target of risk detection, and calculates a risk score from the plurality of interpretation indexes.

Description

[Technical Field]

[0001] The present invention relates to an information processing device and an information processing method that use a machine learning model to detect risks such as accounting fraud that can be detected from accounting-related data. [Background technology]

[0002] In accounting audit work such as internal audits and external audits, the contents of accounting documents such as financial statements are analyzed and checked to see if there are any fraudulent transactions, such as fictitious transactions (see Patent Documents 1 and 2, etc.).

[0003] There have also been attempts to use the results of machine learning to evaluate the risk of accounting fraud through computer processing, as disclosed in Patent Document 3. Here, the calculations performed in the learning process of machine learning are becoming more complex, and while this has the advantage of enabling highly accurate evaluations, it is often difficult for humans to easily understand the basis for the evaluations.

[0004] Therefore, the information processing device of Patent Document 3 is able to identify which account items, which are explanatory variables used in machine learning, are abnormal for the evaluation output by the machine learning model by calculating the SHAP value. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent No. 7052135 [Patent Document 2] Japanese Patent Application Publication No. 2019-179531 [Patent Document 3] Patent No. 7216854 Summary of the Invention [Problem to be solved by the invention]

[0006] However, the information processing device described in Patent Document 3 requires a process of calculating a SHAP value to interpret which account items are abnormal, in addition to a process of bagging outputs from multiple weak learners to calculate an index (risk score) indicating inappropriateness. Furthermore, the amount of calculation required for the theoretical value of this SHAP value increases exponentially depending on the number of explanatory variables, so in practice an approximate calculation is required.

[0007] Therefore, an object of the present invention is to provide an information processing device and an information processing method that can calculate an interpretation index for each explanatory variable in the process of calculating a risk score, thereby reducing the calculation load. [Means for solving the problem]

[0008] In order to achieve the above object, the information processing device of the present invention is an information processing device that detects risks detectable from accounting-related data using a machine learning model, and is characterized by comprising: a learning model generation unit that uses a plurality of training records that combine a plurality of explanatory variables and a target variable selected from the accounting-related data to generate, as a machine learning model, a shape function that indicates the relationship between each of the explanatory variables and the target variable; a learning model memory unit that stores the shape function for each of the explanatory variables generated by the learning model generation unit; and a risk calculation unit that calculates an interpretation index for each of the explanatory variables for a new record that is to be subject to risk detection using the shape function stored in the learning model memory unit, and calculates a risk score from the plurality of interpretation indexes.

[0009] Here, the risk score can be calculated based on an intercept value set when the machine learning model was generated and the value of the interpretation index for each of the explanatory variables. The explanatory variables used to generate the shape function may include a group of accounting-related data items that are affected by the content of the risk. The accounting-related data may be financial data, and the risk to be detected may be accounting fraud.

[0010] The invention also relates to an information processing method, which is an information processing method for detecting risks detectable from accounting-related data using a machine learning model through computer processing, and is characterized by comprising the steps of: using a plurality of training records in which a plurality of explanatory variables are selected from the accounting-related data and combined with a target variable, generating shape functions indicating the relationship between each of the explanatory variables and the target variable through machine learning; calculating an interpretation index for each of the explanatory variables using the generated shape functions for a new record that is to be subject to risk detection; and calculating a risk score based on the calculated plurality of interpretation indexes. [Effects of the Invention]

[0011] The information processing device and information processing method of the present invention configured as described above use a plurality of training records each combining a plurality of explanatory variables and a target variable selected from accounting-related data to generate a machine learning model represented by a shape function for each explanatory variable. Then, for a new record to be subjected to risk detection, an interpretation index for each explanatory variable is calculated using the shape function, and a risk score is calculated from the values ​​of the plurality of interpretation indexes.

[0012] In this way, since the interpretation index for each explanatory variable is calculated in the process of calculating the risk score, there is no need to perform a separate calculation after calculating the risk score, and the calculation load can be reduced. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a block diagram illustrating the configuration of an accounting audit support device according to an embodiment of the present invention. [Figure 2] 1 is a flowchart illustrating the process flow of an accounting audit support method according to an embodiment of the present invention. [Figure 3] 1A and 1B are diagrams for explaining an outline of an EBM model, in which FIG. 1A is an explanatory diagram schematically showing the flow up to score calculation, and FIG. 1B is an explanatory diagram illustrating an example of a shape function. [Figure 4] FIG. 1 is an explanatory diagram illustrating an example data set of training records. [Figure 5] FIG. 1 is an explanatory diagram illustrating explanatory variables of training records used in machine learning. [Figure 6] 1A is a diagram illustrating the shape function and output of explanatory variable x1; FIG. 1B is a diagram illustrating the shape function and output of explanatory variable x2; and FIG. 1C is a diagram illustrating the shape function and output of explanatory variable x3. [Figure 7] FIG. 10 is an explanatory diagram illustrating an example of calculation results of interpretation indices for new records. [Figure 8] FIG. 10 is an explanatory diagram illustrating a process of calculating a risk score from an interpretation index. [Figure 9] FIG. 1 is an explanatory diagram illustrating a data set of grouped training records in the first embodiment. [Figure 10] FIG. 10 is an explanatory diagram illustrating calculation results of interpretation indexes and risk scores of new records in Example 1. [Figure 11] 10A and 10B are diagrams illustrating the process of checking interpretation indices within a variable group of interest, where FIG. 10A is a diagram illustrating an example of explanatory variables included in the variable group, and FIG. 10B is a diagram illustrating an example of the calculation results of interpretation indices for each explanatory variable. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a block diagram illustrating the configuration of an accounting audit support device 1, which is an information processing device according to this embodiment. Fig. 2 is a flowchart illustrating the processing flow of an accounting audit support method, which is an information processing method according to this embodiment.

[0015] The accounting audit support device 1 of this embodiment is an information processing device that uses a machine learning model to detect the risk of accounting fraud for new records (for example, by fiscal year and month) of a company that is the target of an accounting audit such as an internal audit or an external audit. Here, detecting the risk of accounting fraud includes both detecting fraud at the present time and predicting future fraud.

[0016] In the accounting audit support device 1 of this embodiment, as shown in FIG. 1, a processing unit 3 performs computer processing based on data input from an input device 2, a memory unit 4, etc., and displays detection results, etc. on a display device 5, etc.

[0017] The calculation processing unit 3 of the accounting audit support device 1 is mainly composed of a training data setting unit 31 that sets training data to be used for machine learning, a learning model generation unit 32 that generates a machine learning model using training records included in the training data, an interpretation index calculation unit 33 that calculates interpretation indexes for new records, a risk score calculation unit 34 that calculates risk scores from the interpretation indexes, and a display control unit 35 that controls the display content on the display device 5. Here, the interpretation index calculation unit 33 and the risk score calculation unit 34 correspond to the risk calculation unit.

[0018] The input device 2 for inputting various data is a data input means connected to or equipped on a personal computer (PC), laptop computer, tablet terminal, wearable terminal, smartphone, etc. Examples of the input device 2 include a keyboard, mouse, touch panel, touch pad, scanner, microphone for voice input, camera, etc.

[0019] On the other hand, the display device 5 may be a liquid crystal display, an organic EL (Electro-Luminescence) display, a printer, or the like.

[0020] Furthermore, the storage unit 4 is a storage medium that records training data used in processing in the arithmetic processing unit 3, data generated by the arithmetic processing, etc., and corresponds to a hard disk, a solid state drive (SSD), a flash memory (such as an SD memory card), a magnetic disk, an optical disk, etc. External online storage (cloud storage) such as a server connected via a network can also be used as the storage unit 4. In the following, the area of ​​the storage medium that stores the machine learning model generated by the learning model generation unit 32 will be distinguished and described by assigning it the reference symbol "learning model storage unit 41."

[0021] The arithmetic processing unit 3 is configured as computer hardware such as a CPU (Central Processing Unit), MPU (Micro-processing unit), GPU (Graphics Processing Unit), etc., and is equipped with memories such as RAM (Random Access Memory) and ROM (Read Only Memory).

[0022] The arithmetic processing unit 3 can execute various functions by applications (programs) installed on a computer. Also, a server connected via a network such as the Internet can execute part or all of the arithmetic processing unit 3. The network can be configured from part or all of the Internet, a wide area network (WAN), a wired local area network (LAN), a wireless LAN (Wi-Fi), a provider device, a wireless base station, a dedicated line, etc.

[0023] As described above, the calculations performed in the learning process of machine learning are becoming more complex, and it is often difficult for humans to easily understand the basis for results such as risk assessment values ​​(scores) detected using machine learning models. Therefore, the accounting audit support device 1 and accounting audit support method of this embodiment calculate interpretation indices for each explanatory variable, making it possible to understand the impact of each explanatory variable on the score and use the results in business decision-making.

[0024] Specifically, we apply a method called Explainable Boosting Machine (EBM). In this method, we first generate an estimation model called an EBM model using machine learning, and then use this EBM model to calculate interpretation indices and scores (risk scores) for each explanatory variable. Figure 3 is a diagram that explains the overview of the EBM model.

[0025] To create an EBM model, as shown on the left side of Figure 3(a), multiple explanatory variables (x1, x2, x3) are selected to be used to predict the dependent variable, and multiple training records that combine the selected explanatory variables with the dependent variable are used to generate a machine learning model with a shape function that indicates the relationship between each explanatory variable and the dependent variable. Specifically, the intercept (β0), which will be described later, and the shape function of each explanatory variable are estimated. Figure 3(b) is an explanatory diagram illustrating an example of the shape function f1(x1) estimated for the dependent variable x1.

[0026] The shape function is nonlinear and is generated for each explanatory variable, resulting in a model (f1(x1), f2(x2), f3(x3)) that combines shallow decision trees that use only one relevant explanatory variable. It is also possible to create a shape function that incorporates an interaction term between explanatory variables. For example, when using an interaction term between two explanatory variables, a model (f 1,2 (x1,x2)).

[0027] Once an EBM model has been created, it can be used to calculate the interpretation index and risk score of a new record that is the target of risk detection. For example, if the data corresponding to each explanatory variable of a new record (a) (x1 (a) ,x2 (a) ,x3 (a) ) into the EBM model, the interpretation index (f1(x1 (a) ), f2(x2 (a) ), f3(x3 (a) )) can be calculated. The risk score is calculated by the intercept (β0) and the output of the shape function of each explanatory variable (interpretation index (f1(x1 (a) ), f2(x2 (a) ), f3(x3 (a) ))) can be obtained by, for example, performing a logistic transformation.

[0028] The method for interpreting the results output using the EBM model is to output the shape function of each explanatory variable for each record, and to regard this output as the interpretation index for that explanatory variable for that record. The interpretation index is given in terms of how much the explanatory variable increases or decreases the logarithmic odds (logit) of the score (fraud probability). As mentioned above, the interpretation index (f 1,2 (x1,x2)) can also be calculated.

[0029] A feature of the EBM method is that the interpretation index is calculated in advance during the score calculation process, so there are no additional development or maintenance costs for calculating the interpretation index. Also, no additional calculation time is required just to calculate the interpretation index. Furthermore, the impact of the explanatory variable value on the score is given by a nonlinear shape function, so the model is highly flexible. And the impact on the score when the explanatory variable value is 0 is determined by the specific impact based on the shape function (for example, f1(x1 (a) =0), and the impact does not become 0.

[0030] Furthermore, with the EBM method, even if outliers exist among the explanatory variables, they are not adversely affected when creating the model or calculating the score. For unknown data (such as new categorical values ​​that do not exist at the time of model training) for categorical variables (such as the manager's address or educational background), the influence of the explanatory variable on the score can be eliminated by setting the output value of the explanatory variable to 0 (e.g., f1(x1=U)=0). Furthermore, for missing data in numerical variables, the influence of the explanatory variable on the score can be eliminated by setting the output value of the explanatory variable to 0 (e.g., f1(x1=NA)=0).

[0031] To apply such an EBM method, the training data setting unit 31 of the calculation processing unit 3 extracts and uses a plurality of training records, each combining a plurality of explanatory variables and a target variable, from the accounting-related data (records) usable for machine learning and stored in the storage unit 4. For example, if the accounting-related data is financial data, some of the items listed in the financial statements are selected as explanatory variables. Then, the record combining the selected explanatory variables and target variable is set as the training record to be used to generate a machine learning model.

[0032] Figure 4 is an explanatory diagram illustrating an example of a training record dataset. Records from the same company can be used as separate training records if their fiscal year and month are different. A training record is a combination of the values ​​of explanatory variables (x1-x3) selected from the items in the financial statements and the fraud label y assigned to that training record. In short, it is training data in which the solution when the values ​​of the explanatory variables (x1-x3) are input is the fraud label y. Here, the fraud label y assigned to records where accounting fraud occurred is set to 1, and the fraud label y assigned to records where there was no accounting fraud is set to 0.

[0033] Figure 5 shows examples of items that are considered effective as explanatory variables for training records used in machine learning when detecting accounting fraud risks. In Figure 5, items listed in financial statements that can be used as explanatory variables are grouped according to the following criteria:

[0034] First, when creating groups, we decided to select groups from the perspective of grouping explanatory variables at the financial indicator level that are affected by common fraudulent content.

[0035] Specifically, when inventory-related fraud occurs, the inventory turnover period, inventory net asset ratio, and inventory total asset ratio are all affected, so these explanatory variables were combined into one group called "inventory valuation."

[0036] Additionally, the "Sales Overstatement Evaluation" group includes items such as the rate of change in accounts receivable, sales growth rate, and the ratio of the rate of change in accounts receivable divided by the sales growth rate. Furthermore, the "Goodwill Evaluation" group includes items such as the ratio of net goodwill assets and the balance of goodwill exceeding operating profit.

[0037] Each item listed in these financial statements can be used individually as an explanatory variable, and as will be described later, the above-mentioned group units can also be selected as explanatory variables.

[0038] The training data setting unit 31 can set records containing selected explanatory variables and target variables from the records that are accounting-related data stored in the storage unit 4 as training records to be used for machine learning. For example, the data set to be used for training is set by sampling from all available records (data set). Bootstrap sampling or other sampling methods can be used.

[0039] The learning model generation unit 32 creates an EBM model using the training records set by sampling. In the following explanation, we will use a data set (see Figure 4) that combines three selected explanatory variables X (x1, x2, x3) and a target variable (fraud label y) as an example.

[0040] First, using a dataset of sampled training records, a base model F(X) = β is created to model the relationship between explanatory variables X (x1, x2, x3) and the objective variable (fraudulent label y). This β becomes the intercept of the EBM model. Examples of β include the mean value of the objective variable in the case of a regression model, and the logarithm of the odds of the objective variable in the case of a classification model.

[0041] Then, the model F(X) based on the intercept β is updated based on the information (training records) of each explanatory variable. In detail, the following steps (a) to (c) are performed for each explanatory variable in order.

[0042] (a) Creating a decision tree for x1 Set the explanatory variables and target variables as follows to create a decision tree for x1. Explanatory variable: x1 only Response variable: Residual y - F(X) between the actual response variable y and the model's predicted result F(X) This decision tree is a model that expresses the relationship between explanatory variable x1 and target variable y. The created decision tree is multiplied by an arbitrary learning rate to obtain h1(x1), and F(X) is updated by adding h1(x1) to the original model F(X). F(X) := F(X) + h1(x1)

[0043] (b) Creating a decision tree for x2 A decision tree is created by setting the explanatory variable to x2 and the objective variable to the residual y - F(X), and multiplying this by an arbitrary learning rate to obtain h2(x2). F(X) is then updated by adding h2(x2) to the original model F(X). F(X) := F(X) + h2(x2)

[0044] (c) Creating a x3 decision tree A decision tree is created by setting the explanatory variable to x3 and the objective variable to the residual y - F(X), and multiplying this by an arbitrary learning rate to obtain h3(x3). F(X) is then updated by adding h3(x3) to the original model F(X). F(X) := F(X) + h3(x3)

[0045] The above steps (a) to (c) are repeated until the value of the loss function between the model prediction result F(X) and the objective variable y converges. Examples of loss functions include mean squared error for regression models and cross entropy for classification models. In this way, the decision tree model h for each explanatory variable is iterated for the number of times. n (x n ) is created, where n=1,2,3.

[0046] Therefore, for each explanatory variable, a decision tree model h n (x n) are integrated. One way to integrate them is to average all the decision trees created. The integrated result becomes the shape function of each explanatory variable generated using multiple training records.

[0047] The above-mentioned process of setting training records by sampling and generating shape functions from the training records is repeated M times. As a result, M shape functions h n (x n ) will be created.

[0048] Therefore, M shape functions h n (x n ) to obtain the integrated shape function f n (x n ) The integration method here can also be the averaging of all the created decision trees. The integrated shape function (integrated shape function f n (x n )) can be said to represent the relationship between each explanatory variable and the objective variable (see Figure 6).

[0049] As described above, the shape function (integrated shape function f n (x n )) are stored in the learning model storage unit 41. They are then read and used during calculations by the interpretation index calculation unit 33, which will be described later.

[0050] The interpretation index calculation unit 33 calculates the shape function (integrated shape function f n (x n )) to calculate the interpretation index for each explanatory variable. Figure 6 shows the integrated shape function f of the explanatory variables X(x1, x2, x3). n (x n ) is an explanatory diagram illustrating an example.

[0051] Fig. 6(a) is an explanatory diagram illustrating the shape function f1(x1) of explanatory variable x1 and the value 0.3 output for a new record. Fig. 6(b) is an explanatory diagram illustrating the shape function f2(x2) of explanatory variable x2 and the value -0.2 output for a new record. Fig. 6(c) is an explanatory diagram illustrating the shape function f3(x3) of explanatory variable x3 and the value 0.1 output for a new record.

[0052] The values ​​output by the shape functions for each of these explanatory variables are the interpretation index values ​​for each explanatory variable. Figure 7 is an explanatory diagram showing an example of the calculation results of the interpretation index for a new record. By calculating the interpretation index for each explanatory variable in this way, it becomes possible to understand the degree of influence of each explanatory variable.

[0053] The risk score calculation unit 34 calculates a risk score based on the interpretation index of each explanatory variable calculated by the interpretation index calculation unit 33 and the value of the intercept (β0: see FIG. 7). FIG. 8 is an explanatory diagram illustrating the process of calculating a risk score from an interpretation index. The fraud score, which is the risk score, can be obtained by, for example, logistic transformation using a logistic function.

[0054] For example, the transformation can be performed using the following logistic function σ(z): σ(z) = 1 / (1+exp(-z)) Risk of fraud score = 1 / (1+exp(-(-0.5+0.3-0.2+0.1)) = 0.426

[0055] The display control unit 35 controls the output format (see Figure 8) of the interpretation index calculated by the interpretation index calculation unit 33 and the fraud score (risk score) calculated by the risk score calculation unit 34 to the display device 5.

[0056] The processing flow of the accounting audit support method of this embodiment will be described below with reference to the flowchart shown in FIG.

[0057] First, in step S1, several items are selected as explanatory variables from among the items listed in financial statements, which are financial data. Then, records that combine the selected explanatory variables with the objective variable are set as training records to be used to generate a machine learning model (step S2).

[0058] Specifically, the training records are set by reading all data sets that can be used for machine learning from the data stored in the storage unit 4 and setting the training records by sampling.

[0059] In the next step S3, an EBM model is created using the training records set by sampling. As described above, the setting of training records by sampling in step S2 and the generation of an EBM model in step S3 are repeated a predetermined number of times (M times in the above example).

[0060] Then, in step S4, each interpretation index is calculated for each explanatory variable of the new record to be audited. The example in Figure 7 shows that the new record to be audited is financial data extracted from the financial statements of Company Z, whose fiscal year ends in March 2023.

[0061] For the new record, the value of the interpretation index output from the shape function f1(x1) of the explanatory variable x1 was calculated to be 0.3, the value of the interpretation index output from the shape function f2(x2) of the explanatory variable x2 was calculated to be -0.2, and the value of the interpretation index output from the shape function f3(x3) of the explanatory variable x3 was calculated to be 0.1. In addition, the value of the intercept β0 was calculated to be -0.5.

[0062] Therefore, in step S5, a risk score is calculated based on the values ​​of the interpretation index and the intercept value of each explanatory variable calculated in this way. As shown in Fig. 8, the fraud score (risk score) was calculated as 0.426 by logistic transformation. The analysis results obtained by computer processing as shown in Fig. 8 are displayed on the display device 5.

[0063] This fraud score can be said to be an assessment of the risk of accounting fraud, with 1 being the case if accounting fraud is present, but it is difficult to make specific considerations based on this number alone, such as what fraudulent methods may be used to commit accounting fraud.

[0064] Therefore, accounting audit personnel can refer to the interpretation indicators for each explanatory variable calculated in the process of calculating the fraud score to understand which items of financial data they should focus on, which will enable them to take specific audit actions for new records.

[0065] Next, the operation of the accounting audit support device 1 and the accounting audit support method according to this embodiment will be described. The accounting audit support device 1 and accounting audit support method of this embodiment, configured as described above, use multiple training records that combine multiple explanatory variables and target variables selected from accounting-related data, and generate a machine learning model represented by a shape function for each explanatory variable in the learning model generation unit 32.

[0066] Then, for new records that are to be subject to risk detection, the interpretation index calculation unit 33 calculates interpretation indexes for each explanatory variable using a shape function, and the risk score calculation unit 34 calculates a risk score from the values ​​of the multiple interpretation indexes and the intercept values.

[0067] In this way, since the interpretation index for each explanatory variable is calculated in the process of calculating the risk score, there is no need to perform a separate calculation after calculating the risk score, and the calculation load can be reduced.

[0068] Furthermore, even if the fraud score (risk score value) alone does not allow the person performing the accounting audit to easily understand the basis for the prediction, the value of the interpretation index can be used to understand the influence of the financial data items that have become explanatory variables, making it possible to use it in business decision-making.

[0069] Furthermore, with the accounting audit support device 1 and accounting audit support method of this embodiment, the computational load can be reduced, so that it is possible to create a machine learning model using many variables to improve the accuracy of the model, and then instantly identify variables that have a significant effect on the fraud score, thereby improving the efficiency of the audit procedure.

[0070] Furthermore, by applying EBM techniques, the negative impact of outliers in the training records can be reduced, making it possible to create models and calculate fraud scores without being influenced by the financial figures of a small number of very large companies. Furthermore, the model is also unaffected by extreme outliers in division indicators (such as turnover period). [Example]

[0071] Another embodiment of the accounting audit support device 1 according to the embodiment described above will be described below with reference to Figures 9 to 11. Note that the same terms or the same reference numerals will be used to describe the same or equivalent parts as those described in the embodiment described above.

[0072] In the above embodiment, one explanatory variable is associated with one item of financial data. However, as described in the above embodiment, multiple items of financial data can be grouped together, and one variable group can be treated as one explanatory variable.

[0073] In the explanation given in the above embodiment with reference to FIG. 5, a variable group was illustrated that was created from the perspective of grouping explanatory variables that are affected by a common fraudulent content, among explanatory variables at the financial indicator level.

[0074] Fig. 9 is an explanatory diagram illustrating a data set of grouped training records in Example 1. The groups (a, b, c) in Fig. 9 correspond to groups such as "inventory valuation," "sales overstatement valuation," and "goodwill valuation" as illustrated in Fig. 5, and the explanatory variables (x1-x6) correspond to items of financial data belonging to each group.

[0075] The process for creating an EBM model for each variable group and calculating interpretation indicators is basically the same as the process for each item described above. In other words, multiple financial data items are grouped together into variable groups, and an EBM model is created for each variable group. Then, interpretation indicators are calculated for each variable group using shape functions.

[0076] By grouping in this way, when a fraud occurs, it becomes possible to avoid the importance of the group that should be the focus being reduced due to the dispersion of importance among explanatory variables related to the fraud.In other words, it becomes possible to create and interpret an EBM model by eliminating multicollinearity between similar feature quantities (financial indicators) without excluding one of them, and without overlapping explanatory variable information between variable groups.

[0077] FIG. 10 is an explanatory diagram illustrating the calculation results of the interpretation index and risk score for a new record in Example 1. That is, for a new record of Company Z, which is the subject of an audit and whose fiscal year ends in March 2023, the interpretation index and fraud score that are the output of the shape function for each variable group are shown. As can be seen from this result, the interpretation index is output for each variable group, and the impact on the fraud score can be understood for each variable group.

[0078] When interpretation indices are calculated for each variable group in this way, it is also possible to further specify a variable group of interest and create and calculate an EBM model for each individual variable within the group. Figure 11 is a diagram illustrating the process of checking interpretation indices within a variable group of interest. Figure 11(a) is an explanatory diagram illustrating an example of a dataset of training records for explanatory variables (x3, x4, x5) included in the group b of interest. Figure 11(b) is an explanatory diagram illustrating an example of the calculation results of interpretation indices output by the shape function of each explanatory variable included in group b.

[0079] When grouping is performed in this way, it becomes possible to identify the variable group to be focused on based on the output of the shape function of the variable group, and further to create an EBM model using only the items that make up the variable group of interest and to calculate interpretation indices, etc.

[0080] Specifically, in the event of inventory-related fraud, the significance of the inventory valuation group, which includes the inventory turnover period, the inventory net asset ratio, and the inventory total asset ratio, can be evaluated, and further, interpretation indicators for the inventory turnover period, the inventory net asset ratio, and the inventory total asset ratio can be calculated, respectively.

[0081] The other configurations and effects are substantially the same as those of the above embodiment, and therefore the description thereof will be omitted.

[0082] The embodiments of the present invention have been described above in detail with reference to the drawings, but the specific configurations are not limited to these embodiments and examples, and design changes that do not deviate from the gist of the present invention are included in the present invention.

[0083] For example, in the above embodiment, the case where financial data is used as accounting-related data to detect the risk of accounting fraud has been described, but the present invention is not limited to this. In addition to financial data, accounting-related data also includes corporate governance data, bank account transaction history, credit card transaction history, etc. Risks that can be detected from this data include money laundering, fraudulent card use, and deterioration of business performance (bankruptcy, default, etc.).

[0084] In addition, in the above embodiment, the items listed in financial statements are used as explanatory variables, but this is not limited to this, and various data items or groups listed in various accounting documents can be used as explanatory variables.

[0085] Furthermore, in the above embodiment, a method for calculating a risk score by performing a general logistic transformation on the sum of the intercept and the interpretation index of each explanatory variable has been described, but this is not limited to this, and any method can be applied to calculate a risk score from multiple interpretation indexes, such as a method for further adjusting the logistic transformation. [Explanation of symbols]

[0086] 1: Accounting audit support device (information processing device) 32: Learning model generation unit 33: Interpretation index calculation unit 34: Risk score calculation section 41: Learning model memory unit

Claims

1. An information processing device that detects risks detectable from accounting-related data using a machine learning model, a learning model generation unit that uses a plurality of training records each combining a plurality of explanatory variables and a target variable selected from the accounting-related data to generate a shape function indicating the relationship between each of the explanatory variables and the target variable as a machine learning model; a learning model storage unit in which the shape function for each of the explanatory variables generated by the learning model generation unit is stored; an information processing device comprising: a risk calculation unit that calculates an interpretation index for each explanatory variable using the shape function stored in the learning model storage unit for a new record that is a risk detection target; and that calculates a risk score from a plurality of the interpretation indexes.

2. The information processing device according to claim 1 , wherein the risk score is calculated based on an intercept value set when the machine learning model is generated and a value of the interpretation index for each of the explanatory variables.

3. 3. The information processing device according to claim 1, wherein the explanatory variables used to generate the shape function include a group of items of the accounting-related data that are affected by the content of the risk.

4. 4. The information processing device according to claim 3, wherein the accounting-related data is financial data, and the risk to be detected is accounting fraud.

5. An information processing method for detecting risks detectable from accounting-related data using a machine learning model by computer processing, comprising: generating, by machine learning, a shape function indicating the relationship between each of the explanatory variables and the objective variable, using a plurality of training records in which a plurality of explanatory variables are selected from the accounting-related data and combined with the objective variable; calculating an interpretation index for each of the explanatory variables using the generated shape function for a new record that is a risk detection target; and calculating a risk score based on the calculated plurality of interpretation indicators.

6. The information processing method according to claim 5, wherein the explanatory variables used to generate the shape function include a grouping of items of the accounting-related data that are affected by the content of the risk.

Images