Fraud detection system, fraud detection method, and program

Abstract

To provide a fraud detection system capable of appropriately detecting fraudulent product inflating.SOLUTION: A fraud detection system according to the present disclosure comprises: a handling history acquisition unit for acquiring a plurality of handling histories including qualitative data on a product that is the target of fraud detection and a quantity of the product before and after processing; and a distribution fluctuation detection unit for comparing, for each piece of qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detecting fraudulent product inflating based on the comparison results.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a fraud detection system, a fraud detection method, and a program. [Background technology]

[0002] Food traceability systems are systems for detecting fraud in the food distribution process. These systems require businesses to record the history of how food is handled (hereinafter referred to as "handling history"). These systems also detect fraud by automatically verifying the handling history.

[0003] As a related technique, Patent Document 1 discloses a method for detecting fraud when the difference between the weight loss of food at a specific business and the average weight loss exceeds a threshold. In this method, the average weight loss of food is calculated in advance, and warehouses and ports with a weight loss higher than the average are detected as fraudulent. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2023-508188 Summary of the Invention [Problem to be solved by the invention]

[0005] Food fraud is one type of fraud that occurs during the food distribution process. Food fraud involves the act of "padding," which involves mixing cheaper food (including foreign substances) with more expensive food to increase the quantity, or replacing some or all of the more expensive food with cheaper food. For example, when padding occurs during the pork distribution process, cheaper branded pork meat is mixed with more expensive branded pork meat.

[0006] When the weight loss of an individual inflated food product deviates significantly from the average weight loss of past food products, it is possible to detect inflating using a method such as that disclosed in Patent Document 1. For example, suppose the average weight loss of expensive brand X pork over a given period in the past is 30 kg. In this case, if an individual X pork, weighing 70 kg before butchering and 40 kg after butchering, is inflated with 10 kg of inexpensive brand Y pork, the weight loss will be 20 kg, which is a large difference from the average and will be detected as fraud.

[0007] However, if a business calculates the amount of padding required to avoid fraud detection based on past weight loss of food products and pads the amount within that range, detection will be missed. For example, a business can inflate the weight of food products to avoid detection by working backward from the weight loss of food products in the handling history when no fraudulent activity occurred. For example, suppose that among the X pigs handled in the past by a fraudulent business, there was an individual that weighed approximately 70 kg before butchering, which is close to the weight of newly processed animals, and weighed 45 kg after butchering, which is relatively large. In this case, if 5 kg of Y pig meat is mixed with 40 kg of newly processed X pig meat, the weight loss will be 25 kg, which is a small difference from the average past weight loss of 30 kg, and therefore will not be detected.

[0008] In view of the above-mentioned problems, an object of the present disclosure is to provide a fraud detection system, a fraud detection method, and a program that can appropriately detect fraudulent inflating of products. [Means for solving the problem]

[0009] The fraud detection system according to the present disclosure comprises: a handling history acquisition unit that acquires a plurality of handling histories including qualitative data of the product that is the subject of fraud detection and the quantity of the product before and after processing; The system is equipped with a distribution fluctuation detection unit that compares, for each qualitative data, a current probability distribution that indicates the probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution that indicates the probability distribution of the quantity of the product before and after processing in the past handling history, and detects fraudulent padding of the product based on the comparison results.

[0010] The fraud detection method according to the present disclosure includes: a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; The method includes a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution showing the probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution showing the probability distribution of the quantity of the product before and after processing in the past handling history, and detecting fraudulent padding of the product based on the comparison results.

[0011] The program according to the present disclosure is a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; The computer is caused to execute a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution showing the probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution showing the probability distribution of the quantity of the product before and after processing in the past handling history, and detecting fraudulent padding of the product based on the comparison results. [Effects of the Invention]

[0012] The fraud detection system, fraud detection method, and program according to the present disclosure can appropriately detect fraudulent inflating of merchandise. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a block diagram showing the configuration of a fraud detection system according to the present disclosure. [Figure 2] FIG. 2 is a flowchart showing the processing performed by the fraud detection system. [Figure 3] FIG. 3 is a diagram showing the probability distribution of weight data of food before and after processing when padding has occurred. [Figure 4] FIG. 4 is a block diagram showing the configuration of the traceability system according to the present disclosure. [Figure 5] FIG. 5 is a diagram illustrating an example of a handling history stored in the handling history storage unit. [Figure 6] FIG. 6 is a diagram showing an example of input / output data in the distribution creating unit. [Figure 7] FIG. 7 is a diagram illustrating an example of parameters stored in the distribution storage unit. [Figure 8] FIG. 8 is a diagram illustrating an example of input / output data in the distribution fluctuation detection unit. [Figure 9] FIG. 9 is a diagram illustrating an example of input / output data in the distribution fluctuation detection unit. [Figure 10] FIG. 10 is a diagram showing an example of the probability distribution of weight data of food before and after processing. [Figure 11] FIG. 11 is a flowchart showing the processing of the traceability system in the learning phase. [Figure 12] FIG. 12 is an explanatory diagram of the steps. [Figure 13] FIG. 13 is an explanatory diagram of the steps. [Figure 14] FIG. 14 is an explanatory diagram of the steps. [Figure 15] FIG. 15 is a flowchart showing the processing of the traceability system in the fraud detection phase. [Figure 16] FIG. 16 is an explanatory diagram of the steps. [Figure 17] FIG. 17 is an explanatory diagram of the steps. [Figure 18] FIG. 18 is an explanatory diagram of the steps. [Figure 19] FIG. 19 is an explanatory diagram of the steps. [Figure 20] FIG. 20 is a block diagram showing the configuration of the traceability system according to the present disclosure. [Figure 21] FIG. 21 is a diagram showing an example of input / output data in the moving-window distribution creating unit. [Figure 22] FIG. 22 is a diagram illustrating an example of parameters stored in the moving-window distribution storage unit. [Figure 23]FIG. 23 is a diagram illustrating an example of input / output data in the distribution fluctuation detection unit. [Figure 24] FIG. 24 is a diagram illustrating an example of input / output data in the distribution fluctuation detection unit. [Figure 25] FIG. 25 is a diagram illustrating an example of input / output data in the qualitative data correlation analysis unit. [Figure 26] FIG. 26 is a diagram showing an example of a combination of uncorrelated qualitative data stored in the qualitative data uncorrelated combination storage unit. [Figure 27] FIG. 27 is a diagram illustrating an example of input / output data in the qualitative data correlation variation detection unit. [Figure 28] FIG. 28 is a diagram illustrating an example of input / output data in the qualitative data correlation variation detection unit. [Figure 29] FIG. 29 is a diagram illustrating the processing of the traceability system. [Figure 30] FIG. 30 is a flowchart showing the processing of the traceability system in the learning phase. [Figure 31] FIG. 31 is an explanatory diagram of the steps. [Figure 32] FIG. 32 is an explanatory diagram of the steps. [Figure 33] FIG. 33 is an explanatory diagram of the steps. [Figure 34] FIG. 34 is an explanatory diagram of the steps. [Figure 35] FIG. 35 is an explanatory diagram of the steps. [Figure 36] FIG. 36 is a flowchart showing the processing of the traceability system in the fraud detection phase. [Figure 37] FIG. 37 is an explanatory diagram of the steps. [Figure 38] FIG. 38 is an explanatory diagram of the steps. [Figure 39] FIG. 39 is an explanatory diagram of the steps. [Figure 40] FIG. 40 is an explanatory diagram of the steps. [Figure 41] FIG. 41 is an explanatory diagram of the steps. [Figure 42] FIG. 42 is an explanatory diagram of the steps. [Figure 43] FIG. 43 is an explanatory diagram of the steps. [Figure 44] FIG. 44 is an explanatory diagram of the steps. [Figure 45] FIG. 45 is an explanatory diagram of the steps. [Figure 46] FIG. 46 is an explanatory diagram of the steps. [Figure 47] FIG. 47 is an explanatory diagram of the steps. [Figure 48] FIG. 48 is a block diagram showing the configuration of the traceability system according to the present disclosure. [Figure 49] FIG. 49 is a diagram showing an example of input / output data in the per-enterprise correlation analysis unit. [Figure 50] FIG. 50 is a diagram showing an example of input / output data in the per-carrier correlation fluctuation detection unit. [Figure 51] FIG. 51 is a diagram for explaining the processing of the traceability system. [Figure 52] FIG. 52 is a flowchart showing the processing of the traceability system in the learning phase. [Figure 53] FIG. 53 is an explanatory diagram of the steps. [Figure 54] FIG. 54 is an explanatory diagram of the steps. [Figure 55] FIG. 55 is an explanatory diagram of the steps. [Figure 56] FIG. 56 is an explanatory diagram of the steps. [Figure 57] FIG. 57 is an explanatory diagram of the steps. [Figure 58] FIG. 58 is an explanatory diagram of the steps. [Figure 59] FIG. 59 is an explanatory diagram of the steps. [Figure 60] FIG. 60 is a flowchart showing the processing of the traceability system in the fraud detection phase. [Figure 61] FIG. 61 is an explanatory diagram of the steps. [Figure 62] FIG. 62 is an explanatory diagram of the steps. [Figure 63] FIG. 63 is an explanatory diagram of the steps. [Figure 64] FIG. 64 is an explanatory diagram of the steps. [Figure 65] FIG. 65 is an explanatory diagram of the steps. [Figure 66] FIG. 66 is an explanatory diagram of the steps. [Figure 67] FIG. 67 is an explanatory diagram of the steps. [Figure 68] FIG. 68 is an explanatory diagram of the steps. [Figure 69] FIG. 69 is an explanatory diagram of the steps. [Figure 70] FIG. 70 is an explanatory diagram of the steps. [Figure 71] FIG. 71 is an explanatory diagram of the steps. [Figure 72] FIG. 72 is an explanatory diagram of the steps. [Figure 73] FIG. 73 is an explanatory diagram of the steps. [Figure 74] FIG. 74 is an explanatory diagram of the steps. [Figure 75] FIG. 75 is an explanatory diagram of the steps. [Figure 76] FIG. 76 is an explanatory diagram of the steps. [Figure 77] FIG. 77 is a block diagram showing the configuration of the traceability system according to the present disclosure. [Figure 78] FIG. 78 is a diagram showing an example of fluctuations in the parameters of the probability distribution of weight data of food before and after processing due to factors other than padding, which is stored in the past case storage unit. [Figure 79] FIG. 79 is a diagram showing an example of input / output data in the fluctuation pattern matching detection unit. [Figure 80] FIG. 80 is a diagram showing an example of input / output data in the fluctuation pattern matching detection unit. [Figure 81] FIG. 81 is a diagram for explaining the processing of the traceability system. [Figure 82]FIG. 82 is a flowchart showing the processing of the traceability system in the fraud detection phase. [Figure 83] FIG. 83 is an explanatory diagram of the steps. [Figure 84] FIG. 84 is an explanatory diagram of the steps. [Figure 85] FIG. 85 is an explanatory diagram of the steps. [Figure 86] FIG. 86 is an explanatory diagram of the steps. [Figure 87] FIG. 87 is an explanatory diagram of the steps. [Figure 88] FIG. 88 is an explanatory diagram of the steps. [Figure 89] FIG. 89 is an explanatory diagram of the steps. [Figure 90] FIG. 90 is an explanatory diagram of the steps. [Figure 91] FIG. 91 is an explanatory diagram of the steps. [Figure 92] FIG. 92 is an explanatory diagram of the steps. [Figure 93] FIG. 93 is a block diagram illustrating an example of the hardware configuration of a computer that realizes a fraud detection system or the like. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In each drawing, the same or corresponding elements are designated by the same reference numerals. For clarity of explanation, duplicated explanations will be omitted as necessary.

[0015] <Common embodiment> First, an embodiment common to the fraud detection systems described in embodiments 1 to 4 will be described. The users of the fraud detection system according to the present disclosure are food business operators (hereinafter simply referred to as "business operators") and auditors. Each time a business operator handles food, the business operator records the handling history in the system. The auditor analyzes the handling history data of multiple business operators accumulated in the system, and identifies the fraudulent business operator and qualitative data.

[0016] (Configuration of fraud detection system 100) 1 is a block diagram showing the configuration of a fraud detection system 100 according to the present disclosure. The fraud detection system 100 includes a transaction history acquisition unit 101 and a distribution fluctuation detection unit 102.

[0017] The handling history acquisition unit 101 acquires a plurality of handling histories, each including qualitative data of the product that is the subject of fraud detection and the quantity of the product before and after processing. The handling history is information relating to the history of handling of the product.

[0018] Qualitative data is data that represents the quality or characteristics of a product, such as information indicating the product's place of origin, brand, sex, breeding method, farming method, feed, grade, award history, certification, or authentication.

[0019] The distribution fluctuation detection unit 102 compares, for each qualitative data item, a current probability distribution that indicates the probability distribution of the quantity of a product before and after processing in the acquired handling history with a past probability distribution that indicates the probability distribution of the quantity of a product before and after processing in past handling history.The distribution fluctuation detection unit 102 detects fraudulent product padding based on the comparison results.Note that, hereinafter, the probability distribution may be simply referred to as "distribution."

[0020] The fraud detection system 100 includes a processor, memory, and storage device (not shown). The storage device stores a computer program that implements the processing described herein. The processor can load the computer program from the storage device into the memory and execute the computer program. In this way, the processor realizes the functions of the transaction history acquisition unit 101 and the distribution fluctuation detection unit 102.

[0021] The handling history acquisition unit 101 and the distribution fluctuation detection unit 102 may each be realized by dedicated hardware. Furthermore, some or all of the components may be realized by general-purpose or dedicated circuits, processors, etc., or a combination of these. These may be configured by a single chip, or by multiple chips connected via a bus. Some or all of the components may be realized by a combination of the above-mentioned circuits, etc., and a program.

[0022] (Processing of fraud detection system 100) Next, the processing performed by the fraud detection system 100 will be described with reference to Fig. 2. Fig. 2 is a flowchart showing the processing performed by the fraud detection system 100.

[0023] First, the transaction history acquisition unit 101 acquires multiple transaction histories (S1). Next, the distribution fluctuation detection unit 102 compares the current probability distribution with the past probability distribution for each qualitative data item (S2). Then, the distribution fluctuation detection unit 102 detects fraudulent product inflating based on the comparison results (S3).

[0024] As described above, the fraud detection system 100 according to the present disclosure can appropriately detect fraudulent padding.

[0025] <Embodiment 1> Next, a first embodiment will be described. The first embodiment is a specific example of the fraud detection system 100 according to the above-mentioned common embodiment. Generally, fraudulent businesses pad their prices multiple times to make a profit. As a result, the probability distribution of weight data before and after food processing changes before and after the fraud. Therefore, the traceability system 11 according to the present disclosure detects fraud based on the fluctuation in the probability distribution of weight data before and after food processing.

[0026] First, referring to Figure 3, we will explain the relationship between water padding and the fluctuation in the probability distribution of weight data before and after food processing. Figure 3 is a diagram showing the probability distribution of weight data before and after food processing when water padding has occurred. The vertical axis represents the probability density, and the horizontal axis represents weight loss. The left side of Figure 3 shows expensive foods, and the right side shows inexpensive foods.

[0027] Even if the weight of a food item is inflated based on the weight loss of the food item in the past handling history, the probability distribution of the weight data before and after processing the food will fluctuate, as shown in the figure. Specifically, for foods processed by a specific business, the weight loss of expensive foods will decrease while the weight loss of cheap foods will increase.

[0028] The traceability system 11 according to the present disclosure detects such padding. A traceability system is a system in which multiple businesses record the handling history of food products and verify that history to detect fraud. For example, in the distribution process of pigs, the handling history is registered in the traceability system each time each process, such as production, slaughtering, butchering, and sales, is completed.

[0029] The traceability system 11 detects food fraud by utilizing changes in the statistical characteristics of weight loss across the entire food product of a specific business. The traceability system 11 detects fraud by performing predetermined processing based on the probability distribution of the entire food product, rather than the weight loss of individual foods. For example, the traceability system 11 collects the weight loss of multiple foods handled over a certain period of time and compares the probability distribution of weight data for current and past foods before and after processing. This allows the traceability system 11 to reduce the possibility of missing detections.

[0030] Specifically, the traceability system 11 uses a new set of handling histories and parameters of the probability distribution for each piece of qualitative data to calculate the likelihood for each piece of qualitative data, and detects padding by determining that the probability distribution has changed for qualitative data for which the likelihood falls below a threshold. The likelihood indicates the degree to which the past probability distribution matches the current probability distribution. The traceability system 11 calculates the likelihood for a new set of handling histories using parameters of the probability distribution, and detects a change in the probability distribution if the likelihood is below a threshold. The configuration of the traceability system 11 is described in detail below.

[0031] Below, an example will be described in which the traceability system 11 is used as a system for detecting fraud in the distribution process of pork. Since the processing details are the same for each business operator, an example of slaughterman A will be described. Note that the traceability system 11 may be used for any meat, not just pork. For example, the traceability system 11 may be used for beef, mutton, chicken, or other meat. Furthermore, the traceability system 11 is not limited to these, and may be used for any product. For example, the traceability system 11 may be used to track various foods, medicines, etc.

[0032] (Configuration of Traceability System 11) Fig. 4 is a block diagram showing the configuration of the traceability system 11 according to the present disclosure. Figs. 5 to 9 are diagrams explaining in detail each component of the traceability system 11. The traceability system 11 is an example of the fraud detection system 100 described above. The traceability system 11 includes a handling history creation unit 1, a handling history storage unit 2, a learning unit 3, a handling history tracking unit 4, and a fraud detection unit 5. The handling history creation unit 1 and the handling history tracking unit 4 may be implemented as smart contracts.

[0033] The handling history creation unit 1 creates a handling history. A handling history is a record of who handled which food, when, and how. For example, the handling history creation unit 1 stores information about the processing registered via a terminal by a food processing business operator as a handling history in the handling history storage unit 2.

[0034] For example, a pig producer registers information about the birth of pigs in the traceability system 11. A processing business also registers information about the processing in the traceability system 11. The handling history includes, for example, information indicating the food ID, place of origin, brand, sex, business, business division, name of person in charge, handling details, weight before processing, weight after processing, and work date and time. The business, business division, and name of person in charge are automatically entered based on the authentication information when the handling history was created, and other information is entered manually.

[0035] A food ID is identification information for identifying a food. The place of origin, brand, and sex are examples of qualitative data that indicate the quality or properties of the food. For example, if the food is pork, the place of origin, brand, and sex correspond to the place of origin, brand, and sex of the pig.

[0036] The business operator is information that indicates the business operator that performed the processing. In this case, the business operator indicates the business operator that processed the pig. The business operator may be represented by a business operator ID that identifies the business operator, a business operator name, etc. The processing content indicates the processing content of the food. Examples of processing content include slaughtering and butchering.

[0037] The weight before processing and the weight after processing indicate the weight of the food at each timing. The weight before processing may be measured before processing, or the weight after the most recent processing may be used. Note that although the weight of a pig is used in the explanation here, other quantitative data may be used instead of weight. The quantitative data may be, for example, the volume, length, number, or apparent area of ​​the food.

[0038] The handling history storage unit 2 stores the handling history. Figure 5 is a diagram showing an example of a handling history stored in the handling history storage unit 2. The handling history is created for each individual livestock, and includes qualitative data of the food such as the brand and sex, weight data of the food such as the weight before processing and the weight after processing, and the name of the business that processed the food. The handling history is managed in a table for each business. One handling history storage unit 2 may store the handling history of only one business. Alternatively, the handling history may be stored in a separate handling history storage unit 2 for each business. The method of storing the handling history in the handling history storage unit 2 is not limited to these.

[0039] The learning unit 3 performs learning according to the present disclosure. The learning unit 3 includes a distribution creation unit 31 and a distribution storage unit 32. The distribution creation unit 31 may be implemented as a smart contract.

[0040] The distribution creation unit 31 uses the acquired multiple handling histories to learn the parameters of each probability distribution. Fig. 6 is a diagram showing an example of input / output data in the distribution creation unit 31. The internal processing of the distribution creation unit 31 will be described with reference to this figure.

[0041] The distribution creation unit 31 acquires all transaction histories in the table of businesses that are subject to fraud detection from the transaction history storage unit 2. The distribution creation unit 31 separates the acquired transaction history group into qualitative data. The distribution creation unit 31 extracts pairs of weight before and after processing from the transaction history group for each qualitative data. The distribution creation unit 31 uses a machine learning method on the weight data before and after processing for each qualitative data, and learns the parameters of the probability distribution for each. The distribution creation unit 31 saves the learned parameters in the distribution storage unit 32.

[0042] The distribution storage unit 32 stores parameters of the probability distribution. Fig. 7 is a diagram showing an example of parameters stored in the distribution storage unit 32. As shown in the figure, the distribution storage unit 32 stores the mean vector μ→ and the covariance matrix Σ, which are parameters of the bivariate normal distribution of the weight data of food before and after processing. Note that the symbol "→" indicates a vector.

[0043] Here, two random variables, the weight before processing and the weight after processing, are used for the explanation, but other combinations of random variables may be used instead of the weights before and after processing. For example, one variable representing the ratio of the weights before and after processing, or a combination of four random variables, the weights before and after processing and the volumes before and after processing, may be used. Also, here, a bivariate normal distribution is used for the explanation, but other probability distributions may be used instead of the bivariate normal distribution. For example, a mixed normal distribution may be used.

[0044] The handling history tracking unit 4 tracks the handling history. For example, the handling history tracking unit 4 receives a handling history tracking request from a user terminal used by a user. The handling history tracking unit 4 tracks the handling history of the food product and returns the tracking result to the user terminal.

[0045] The fraud detection unit 5 detects padding of products. The fraud detection unit 5 includes a distribution fluctuation detection unit 51. The distribution fluctuation detection unit 51 may be implemented as a smart contract. The distribution fluctuation detection unit 51 is an example of the handling history acquisition unit 101 and the distribution fluctuation detection unit 102 described above.

[0046] The distribution fluctuation detection unit 51 functions as a handling history acquisition unit that acquires multiple handling histories including qualitative data of the product that is the target of fraud detection and the quantity of the product before and after processing. For example, the distribution fluctuation detection unit 51 acquires multiple handling histories of foods that have been processed within a predetermined period from the present.

[0047] Furthermore, the distribution fluctuation detection unit 51 compares, for each qualitative data item, a current probability distribution indicating the probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating the probability distribution of the quantity of the product before and after processing in the past handling history. The distribution fluctuation detection unit 51 detects fraudulent product inflating based on the comparison result. For example, the distribution fluctuation detection unit 51 calculates a likelihood indicating the degree to which the past probability distribution matches the current probability distribution, and detects inflating based on the calculation result.

[0048] 8 and 9 are diagrams showing an example of input / output data in the distribution fluctuation detection unit 51. The internal processing of the distribution fluctuation detection unit 51 will be described with reference to the figures. The distribution fluctuation detection unit 51 acquires a new handling history group from the handling history storage unit 2. The distribution fluctuation detection unit 51 also acquires probability distribution parameters of weight data of food before and after processing for each qualitative data item from the distribution storage unit 32.

[0049] The distribution change detection unit 51 divides the new handling history group into each qualitative data of the food, such as brand and gender. The distribution change detection unit 51 extracts a set of weights before and after processing Y={y1, y2, . . . , y m} are extracted respectively.

[0050] The distribution change detection unit 51 calculates the likelihood L for each piece of qualitative data of food using the following formula (1).

number

[0051] (Example of Embodiment 1) Next, a description will be given of an example of embodiment 1. The processing performed by the traceability system 11 is divided into a learning phase and a fraud detection phase.

[0052] In the learning phase, the traceability system 11 uses past handling history to learn the parameters of the probability distribution of weight data of food before and after processing for each combination of brand and gender, rather than learning the average weight loss of food.

[0053] In the fraud detection phase, the traceability system 11 uses the new handling history to detect fraud. The traceability system 11 does not detect fraud by comparing the weight loss of each individual food product with the average weight loss, but rather detects fraud for each combination of brand and gender. The traceability system 11 uses likelihood to check whether a dataset aggregating the weight data of each food product before and after processing follows the probability distribution of the weight data of the food product before and after processing learned in the learning phase. If the probability distributions differ between the learning phase and the fraud detection phase, the traceability system 11 detects fraud.

[0054] FIG. 10 shows an example of the probability distribution of food weight data before and after processing. The vertical axis indicates the weight after processing, and the horizontal axis indicates the weight before processing. The ellipses indicate the probability distribution, and the dots indicate the weight data of individual foods. The outer ellipse indicates the probability distribution of food weight data in the learning phase. The inner ellipse indicates the probability distribution of food in the fraud detection phase.

[0055] In this example, the weight loss (points) of individual foods can take values ​​within the outer ellipse, but as shown by the inner ellipse, the distribution fluctuates. When looking at a single point, it does not deviate from the past distribution (outer ellipse), but when looking at multiple points, fluctuations in the distribution between the past and present are apparent. The traceability system 11 detects padding by detecting such fluctuations in distribution. This allows the traceability system 11 to reduce the number of times padding goes undetected.

[0056] (Traceability System 11 Processing) Next, a description will be given of the processing performed by the traceability system 11. First, a scenario of this embodiment will be described.

[0057] Slaughterer A processes three types of brands (P pigs, Q pigs, and R pigs), and each brand has two types of sex (male and female). The traceability system 11 stores the handling history. The handling history includes weight data before and after processing. The traceability system 11 detects padding by monitoring the handling history.

[0058] In the learning phase, the traceability system 11 learns the probability distribution of weight data before and after processing for each combination (six types) of brand and gender. In the fraud detection phase, the traceability system 11 calculates the likelihood that the weight data before and after processing conforms to specific parameters, and detects padding based on the calculation results.

[0059] (Learning phase) The processing performed by the traceability system 11 in the learning phase will be described with reference to Fig. 11 to Fig. 14. Fig. 11 is a flowchart showing the processing performed by the traceability system 11 in the learning phase. Fig. 12 to Fig. 14 are diagrams for explaining each step shown in Fig. 11 in detail.

[0060] Steps S101 and S102 shown in Figure 11 are steps for learning a probability distribution for each combination of brand and gender. The distribution creation unit 31 divides all handling histories into each combination of brand and gender (S101). The distribution creation unit 31 uses the handling histories for each combination of brand and gender to learn parameters for the probability distribution of weight data before and after food processing (S102). The distribution creation unit 31 may store the learned parameters in the distribution storage unit 32.

[0061] 12 and 13 are explanatory diagrams of step S101. FIG. 12 shows an example of a handling history that is the premise for explaining this process. The handling history consists of columns for brand, sex, weight before processing, weight after processing, business name, and work date and time. There are three types of brand (P pork, Q pork, R pork) and two types of sex (male, female). The weight of each individual food item is recorded before and after processing. The handling history is recorded for the past 10 days (11 / 01 to 11 / 10). The following explanation will use this example.

[0062] As shown in Figure 13, the distribution creation unit 31 divides all handling histories of slaughterer A into combinations of brand and sex. The distribution creation unit 31 extracts records for each combination of brand and sex (3 x 2 = 6 types). The distribution creation unit 31 extracts records based on the values ​​of the brand column and the sex column. The distribution creation unit 31 divides the histories into combinations of brand and sex. From the handling histories for each qualitative data of food, the distribution creation unit 31 extracts a set of weights before and after processing, X = {x1, x2,..., x n} are extracted respectively.

[0063] 14 is an explanatory diagram of step S102. The distribution creation unit 31 learns the parameters of the bivariate normal distribution of the weight data before and after food processing based on the likelihood L for the handling history. The distribution creation unit 31 learns two parameters, the mean vector μ→ and the covariance matrix Σ.

[0064] The likelihood L indicates the degree to which the weight data before and after processing conforms to a specific parameter. The likelihood L can be calculated using the following formula (2).

number

[0065] The distribution creation unit 31 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L. Specifically, the distribution creation unit 31 substitutes the values ​​of the columns of weight before processing and weight after processing into the equation. The parameters are variable.

[0066] (Fraud detection phase) The processing performed by the traceability system 11 in the fraud detection phase will be described with reference to Figures 15 to 19. Figure 15 is a flowchart showing the processing performed by the traceability system 11 in the fraud detection phase. Figures 16 to 19 are diagrams for explaining each step shown in Figure 11 in detail.

[0067] Steps S111 to S114 shown in Fig. 15 are steps for detecting padding based on fluctuations in the probability distribution. The distribution fluctuation detection unit 51 divides the new transaction history into combinations of brand and gender (S111). The distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand and gender from the transaction history and probability distribution parameters for each combination of brand and gender (S112).

[0068] The distribution fluctuation detection unit 51 determines whether there is a combination of brand and gender for which the likelihood L is less than the threshold (S113). If it is determined that there is a combination of brand and gender for which the likelihood L is less than the threshold (Yes in S113), the distribution fluctuation detection unit 51 detects the combination as inflated (S114). If it is determined that there is no combination of brand and gender for which the likelihood L is less than the threshold (No in S113), the distribution fluctuation detection unit 51 ends the process.

[0069] 16 and 17 are explanatory diagrams of step S111. FIG. 16 shows an example of a new handling history that is the premise for explaining this process. Here, it is assumed that 10 days have passed since the learning phase, and 10 days' worth of new handling history (11 / 11 to 11 / 20) has been recorded. The following explanation will be based on this example.

[0070] 17, the distribution fluctuation detection unit 51 divides the new handling history of slaughterer A into combinations of brand and sex. The distribution fluctuation detection unit 51 extracts records for each combination of brand and sex.

[0071] 18 is an explanatory diagram of step S112. The distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand and gender using the mean vector μ→ and covariance matrix Σ learned from the transaction history for each combination of brand and gender. In the fraud detection phase, the likelihood L indicates the degree to which the probability distribution learned in the learning phase fits the new transaction history. In the learning phase, the parameters of the probability distribution were variable, but in the fraud detection phase, the distribution fluctuation detection unit 51 substitutes pre-learned values ​​into the formula.

[0072] 19 is an explanatory diagram of step S113. The distribution fluctuation detection unit 51 searches for a combination of brand and gender for which the likelihood is less than the threshold. If there is a combination of brand and gender for which the likelihood is less than the threshold, the distribution fluctuation detection unit 51 determines that the combination is inflated.

[0073] As described above, the traceability system 11 according to the present disclosure acquires multiple handling histories containing qualitative data of the product to be detected for fraud and the quantity of the product before and after processing, and uses the acquired handling histories to compare the current probability distribution with the past probability distribution for each piece of qualitative data. The traceability system 11 detects fraudulent inflating of the product based on the comparison results. With this configuration, the traceability system 11 can detect inflated food weight from fluctuations in the probability distribution of weight data before and after food processing, based on weight loss of the food in the past handling histories.

[0074] <Modification of the First Embodiment> A modified example of the first embodiment will be described. In the related art described above, even if fraud is detected, it is not possible to identify the business operator who committed the fraud. The traceability system 11 according to this modified example is a system that is capable of identifying the business operator who committed the fraud.

[0075] The configuration of the traceability system 11 according to this modification is omitted from the illustration because it is generally the same as the configuration shown in Fig. 4. The following describes the differences from the first embodiment.

[0076] As explained in the first embodiment, the transaction history includes qualitative data of the product that is the target of fraud detection, the quantity of the product before and after processing, and data on the business that processes the product. The distribution fluctuation detection unit 51 compares the current probability distribution with the past probability distribution for each combination of qualitative data and business, and detects padding by the business based on the comparison results.

[0077] Specifically, the distribution fluctuation detection unit 51 repeatedly performs the fraud detection process described in embodiment 1 while changing the business operator to be processed, thereby enabling the distribution fluctuation detection unit 51 to identify the business operator that has committed fraud from among multiple business operators.

[0078] Furthermore, if the transaction history includes data such as business divisions and names of personnel, the distribution fluctuation detection unit 51 may repeatedly perform fraud detection processing while changing the business divisions and personnel to be processed. This allows the distribution fluctuation detection unit 51 to identify the business division or personnel who committed fraud from among multiple business divisions and personnel.

[0079] As described above, the traceability system 11 of this modified example compares the current probability distribution with the past probability distribution for each combination of qualitative data and business operator, and detects padding by the business operator based on the comparison results, thereby making it possible to identify businesses that have committed fraud.

[0080] <Regarding Embodiments 2 to 4> During the food distribution process, food weight loss can fluctuate due to factors other than fraudulent padding. Examples of factors other than padding include seasonal fluctuations and replacement of processing equipment. For example, temperature fluctuations affect the thickness of pork back fat, causing weight loss to fluctuate. Furthermore, weight loss can fluctuate if processing equipment such as cutters is replaced.

[0081] Therefore, when padding is detected using related technology, there is a possibility of false positives. In the second to fourth embodiments, a traceability system that detects padding more accurately and reduces the possibility of false positives will be described.

[0082] In the traceability systems according to the second to fourth embodiments, qualitative data corresponding to the current probability distribution is identified as a padding candidate based on the comparison result between the current probability distribution and the past probability distribution. The traceability system also detects padding from among the padding candidates based on the fluctuation over time of the probability distribution corresponding to the padding candidate.

[0083] <Embodiment 2> Next, a description will be given of embodiment 2. First, in order to explain this embodiment, the correlation of parameter fluctuations of probability distributions will be described. For example, if a parameter value of probability distribution P1 changes and the parameter value of probability distribution P2 also changes in conjunction with the change at the same time, it can be said that the parameter values ​​of probability distributions P1 and P2 are correlated.

[0084] Furthermore, if an increase in the value of one parameter coincides with an increase in the value of the other parameter, then it can be said that there is a positive correlation between the parameter values ​​of probability distributions P1 and P2. If an increase in the value of one parameter coincides with a decrease in the value of the other parameter, then it can be said that there is a negative correlation between the parameter values ​​of probability distributions P1 and P2.

[0085] When food is not padded, foods with different qualitative data (such as food brand and gender) are processed independently, so the fluctuations in the probability distribution for each qualitative data are uncorrelated. On the other hand, when food is padded, foods with different qualitative data are mixed, so the fluctuations in the probability distribution are negatively correlated.

[0086] When data is inflated, the probability distribution of the weight data before and after processing for expensive and inexpensive foods handled by a particular business fluctuates at the same time, and the correlation between the fluctuations in each probability distribution changes from no correlation to a negative correlation. The traceability system 12 according to the present disclosure reduces false positives of inflated data by detecting when the fluctuations in the probability distribution for each piece of qualitative data change from no correlation to a negative correlation.

[0087] (Traceability System 12 Configuration) Fig. 20 is a block diagram showing the configuration of the traceability system 12 according to the present disclosure. Figs. 21 to 28 are diagrams explaining in detail each component of the traceability system 12. Similar to the above-described traceability system 11, the traceability system 12 includes a handling history creation unit 1, a handling history storage unit 2, a learning unit 3, a handling history tracking unit 4, and a fraud detection unit 5.

[0088] The learning unit 3 and the fraud detection unit 5 each have functions different from those of the traceability system 11. Specifically, the learning unit 3 has a moving window distribution creation unit 34, a moving window distribution storage unit 35, and a qualitative data correlation analysis unit 36. The moving window distribution creation unit 34 and the qualitative data correlation analysis unit 36 ​​may be implemented as smart contracts. The learning unit 3 performs sliding window (moving window) processing on the handling history of a specific business operator, and obtains the probability distribution parameters of the weight data of food before and after processing for each qualitative data (such as brand) and window. The learning unit 3 performs correlation analysis on the parameter fluctuations of the probability distribution for all combinations of qualitative data, and identifies uncorrelated combinations of qualitative data.

[0089] The fraud detection unit 5 also includes a qualitative data correlation variation detection unit 52 and a qualitative data uncorrelated combination storage unit 53. The qualitative data correlation variation detection unit 52 may be implemented as a smart contract. If there is a combination that has changed from uncorrelated to negatively correlated among all the detected combinations of qualitative data, the fraud detection unit 5 determines that this is a distribution variation due to padding.

[0090] The traceability system 12 according to the present disclosure achieves the following two things. (i) Detect fluctuations in the probability distribution of each qualitative data item of food related to a specific business operator. (ii) Among the probability distributions in which fluctuations are detected, combinations that have changed from uncorrelated to negatively correlated are detected as inflated.

[0091] The moving-window distribution creation unit 34 and the distribution fluctuation detection unit 51 realize the above (i). Also, the qualitative data correlation analysis unit 36 ​​and the qualitative data correlation fluctuation detection unit 52 realize the above (ii).

[0092] In the following, differences from the traceability system 11 will be mainly explained, and overlapping points will be omitted as appropriate.

[0093] Moving-window distribution creation unit 34 acquires all handling histories in the table of businesses that are subject to fraud detection from handling history storage unit 2. Moving-window distribution creation unit 34 performs sliding window processing on the acquired data and learns the parameters of the probability distribution of weight data of food before and after processing for each window. Moving-window distribution creation unit 34 saves the learning results in moving-window distribution storage unit 35.

[0094] FIG. 21 is a diagram showing an example of input / output data in the moving-window distribution creation unit 34. The internal processing of the moving-window distribution creation unit 34 will be described with reference to the same figure. The moving-window distribution creation unit 34 acquires all the handling history sets of a specific business operator from the handling history storage unit 2. The moving-window distribution creation unit 34 separates the acquired handling history sets into qualitative data sets. The moving-window distribution creation unit 34 creates a set of weights before processing and weights after processing, X={x1, x2, . . . , x n} are extracted respectively. The moving-window distribution creation unit 34 sets a specific window (for example, 7 days) and uses a machine learning technique on the weight data before and after processing of each qualitative data while shifting the window, to learn the parameters of the probability distribution for each. The moving-window distribution creation unit 34 stores the learned parameters in the moving-window distribution storage unit 35.

[0095] The moving-window distribution storage unit 35 sets a window, shifts the window along the time axis, and records the probability distribution parameters learned based on the handling history within the window for each qualitative data of the food (brand, gender, etc.). The weight before and after processing for each handling history is used to learn the parameters. Figure 22 is a diagram showing an example of parameters stored in the moving-window distribution storage unit 35.

[0096] The distribution fluctuation detection unit 51 detects a fluctuation in the probability distribution when the likelihood of the parameters of the probability distribution for each qualitative data item is less than a threshold. Specifically, the distribution fluctuation detection unit 51 acquires a new handling history group from the handling history storage unit 2 and acquires the parameters of the probability distribution of the weight data of the food before and after processing for the latest window for each qualitative data item from the moving window distribution storage unit 35. The distribution fluctuation detection unit 51 outputs a list storing the detected pairs of qualitative data items and business operators to the qualitative data correlation fluctuation detection unit 52.

[0097] 23 and 24 are diagrams showing examples of input and output data in the distribution fluctuation detection unit 51. The internal processing of the distribution fluctuation detection unit 51 will be described with reference to these figures. The distribution fluctuation detection unit 51 acquires a new handling history group from the handling history storage unit 2 and acquires probability distribution parameters of the weight data before and after processing of food in the latest window for each qualitative data item from the moving window distribution storage unit 35. The distribution fluctuation detection unit 51 separates the new handling history group into each qualitative data item of food, such as brand and gender.

[0098] The distribution fluctuation detection unit 51 detects a set of weights before and after processing Y={y1, y2, . . . , y m} are extracted respectively. The distribution variation detection unit 51 calculates the likelihood L for each piece of qualitative data of the food using the above-mentioned formula (1).

[0099] The distribution fluctuation detection unit 51 detects that the likelihood L is below a threshold as being padded, and outputs the detected qualitative data and a set Y of the weight before and after processing.

[0100] The qualitative data correlation analysis unit 36 ​​analyzes the correlation between past probability distributions of different qualitative data and identifies combinations of qualitative data that correspond to uncorrelated past probability distributions. Specifically, the qualitative data correlation analysis unit 36 ​​acquires the parameters of the probability distribution learned for each window from the moving window distribution storage unit 35. The qualitative data correlation analysis unit 36 ​​calculates the correlation coefficient of the parameter fluctuations of the probability distribution for all combinations of qualitative data and compares it with a threshold to identify uncorrelated combinations of qualitative data. The qualitative data correlation analysis unit 36 ​​stores the uncorrelated combinations in the qualitative data uncorrelated combination storage unit 53.

[0101] FIG. 25 is a diagram showing an example of input / output data in the qualitative data correlation analysis unit 36. The internal processing of the qualitative data correlation analysis unit 36 ​​will be described with reference to the same figure. The qualitative data correlation analysis unit 36 ​​acquires the parameters of the probability distribution learned for each window from the moving window distribution storage unit 35. The qualitative data correlation analysis unit 36 ​​performs correlation analysis of fluctuations in the parameters of the probability distribution over time for all combinations of qualitative data to determine correlation coefficients. If the correlation coefficient is less than a threshold, the qualitative data correlation analysis unit 36 ​​determines that the combination of qualitative data is uncorrelated. The qualitative data correlation analysis unit 36 ​​transmits the uncorrelated qualitative data combination to the qualitative data uncorrelated combination storage unit 53.

[0102] The qualitative data uncorrelated combination storage unit 53 stores uncorrelated qualitative data combinations transmitted from the qualitative data correlation analysis unit 36. Fig. 26 is a diagram showing an example of uncorrelated qualitative data combinations stored in the qualitative data uncorrelated combination storage unit 53. The qualitative data uncorrelated combination storage unit 53 holds, for each piece of qualitative data, qualitative data whose fluctuations in probability distribution over time are uncorrelated.

[0103] The qualitative data correlation change detection unit 52 identifies, from among the identified combinations of uncorrelated qualitative data, combinations of qualitative data whose current probability distribution has changed to a negative correlation, and detects padding based on the identification results.

[0104] Specifically, in response to the input of a list of pairs of qualitative data and business operators from the distribution fluctuation detection unit 51, the qualitative data correlation fluctuation detection unit 52 acquires the parameters of the probability distribution corresponding to each of the qualitative data in the list from the moving-window distribution storage unit 35. The qualitative data correlation fluctuation detection unit 52 acquires uncorrelated qualitative data combinations from the qualitative data uncorrelated combination storage unit 53.

[0105] The qualitative data correlation variation detection unit 52 calculates the correlation coefficient of the parameter variation of the probability distribution for all combinations of qualitative data in the list and compares it with a threshold to identify combinations of qualitative data with negative correlations. The qualitative data correlation variation detection unit 52 detects qualitative data combinations that have changed from no correlation to a negative correlation as being inflated. The qualitative data correlation variation detection unit 52 outputs the detected combinations of qualitative data to the auditor.

[0106] 27 and 28 are diagrams showing examples of input and output data in the qualitative data correlation variation detection unit 52. The internal processing of the qualitative data correlation variation detection unit 52 will be described with reference to these figures.

[0107] The qualitative data correlation variation detection unit 52 receives a list of qualitative data of food products and corresponding pairs of weights before and after processing from the distribution variation detection unit 51. The qualitative data correlation variation detection unit 52 acquires uncorrelated qualitative data corresponding to the received qualitative data from the qualitative data uncorrelated combination storage unit 53. The qualitative data correlation variation detection unit 52 learns probability distribution parameters from pairs of weights before and after processing of the detected qualitative data. The qualitative data correlation variation detection unit 52 calculates correlation coefficients between the fluctuated probability distributions and identifies combinations with negative correlations that fall below a threshold. If the identified combination is included in combinations of uncorrelated qualitative data, the qualitative data correlation variation detection unit 52 detects it as padding.

[0108] (Example of Embodiment 2) Next, a description will be given of an example of embodiment 2. The processing performed by the traceability system 12 is divided into a learning phase and a fraud detection phase.

[0109] In the learning phase, the traceability system 12 learns past handling history for each combination of brand and gender over a shifting period, thereby obtaining fluctuations in the parameters of the probability distribution over time. Specifically, the traceability system 12 performs sliding window processing in addition to the processing of embodiment 1. The traceability system 12 calculates correlation coefficients between combinations of brand and gender from the fluctuations in the parameters of the probability distribution over time, and finds uncorrelated combinations of brand and gender.

[0110] In the fraud detection phase, the traceability system 12 uses the new handling history to find combinations of brand and gender for which the probability distribution of weight data before and after food processing differs between the learning phase and the fraud detection phase for each combination of brand and gender. Based on the correlation coefficient between the combinations of brand and gender, the traceability system 12 detects fraud when there is a change from no correlation to a negative correlation.

[0111] Figure 29 is a diagram explaining the processing of the traceability system 12. The vertical axis indicates the weight after processing, and the horizontal axis indicates the weight before processing. The ellipses represent the probability distribution. In this figure, the ellipses show the probability distributions learned for brands a1 and b1 while shifting the period. The probability distribution for brand a1 shifts downward in the order of periods T1, T2, and T3. Conversely, the probability distribution for brand b1 shifts upward in the order of periods T1, T2, and T3. The traceability system 12 detects padding only when fluctuations in the probability distribution result in a negative correlation in this way.

[0112] (Traceability System 12 Processing) Next, a description will be given of the processing performed by the traceability system 12. First, a scenario of this embodiment will be described.

[0113] Slaughterer A processes three types of brands (P pigs, Q pigs, and R pigs), and each brand has two types of sex (male and female). The traceability system 12 stores the handling history. The handling history includes weight data before and after processing. The traceability system 12 detects padding by monitoring the handling history.

[0114] In the learning phase, the traceability system 12 learns the probability distribution of weight data before and after processing for each combination (six types) of brand and sex. The traceability system 12 finds P male pigs and R male pigs as combinations of brand and sex that are uncorrelated.

[0115] In the fraud detection phase, the traceability system 12 detects P male pigs and R male pigs, whose weight loss has decreased due to seasonal fluctuations, as a combination of brand and sex that may be inflated based on fluctuations in probability distribution. However, since there is a positive correlation between the fluctuations in these two probability distributions, it detects that this is not inflated.

[0116] The traceability system 12 detects the P male pigs whose weight loss has decreased due to padding and the R male pigs whose weight loss has increased as combinations of brand and sex that may be padded. Since there is a negative correlation between the fluctuations of these two probability distributions, it detects them as padding.

[0117] (Learning phase) The processing performed by the traceability system 12 in the learning phase will be described with reference to Fig. 30 to Fig. 35. Fig. 30 is a flowchart showing the processing performed by the traceability system 12 in the learning phase. Fig. 31 to Fig. 35 are diagrams for explaining each step shown in Fig. 30 in detail.

[0118] Steps S201 to S203 shown in Fig. 30 are steps for learning a probability distribution for each combination of brand and gender and for each window. Steps S204 and S205 are steps for identifying uncorrelated combinations of brand and gender.

[0119] The moving-window distribution creation unit 34 divides all handling histories into combinations of brand and gender (S201). The moving-window distribution creation unit 34 divides the handling histories for each combination of brand and gender using sliding window processing (S202). The moving-window distribution creation unit 34 uses the handling histories for each combination of brand and gender and for each window to learn parameters for the probability distribution of weight data before and after food processing (S203).

[0120] The qualitative data correlation analysis unit 36 ​​uses the parameters of the probability distribution to calculate the correlation coefficient between combinations of stock and gender (S204).The qualitative data correlation analysis unit 36 ​​identifies and saves uncorrelated combinations of stock and gender based on the correlation coefficient between combinations of stock and gender (S205).

[0121] Step S201 is the same as step S101 in Fig. 11, and therefore a detailed description thereof will be omitted. In this embodiment as well, an example of handling history shown in Figs.

[0122] 31 is an explanatory diagram of step S202. The moving-window distribution creation unit 34 divides the trading history for each combination of brand and gender using sliding window processing. If the window width is 7 days and the sliding width is 1 day, the 10-day trading history is divided into four windows t1 to t4.

[0123] 32 is an explanatory diagram of step S203. The moving-window distribution creation unit 34 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood for the handling history. The moving-window distribution creation unit 34 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0124] 33 and 34 are explanatory diagrams of step S204. As preprocessing for calculating correlation coefficients, the qualitative data correlation analysis unit 36 ​​generates a vector combining the mean vector μ→ and the covariance matrix Σ for each brand and gender combination and for each window. The moving window distribution creation unit 34 combines the mean vector μ→ with the one-dimensional covariance matrix Σ. As a result, the moving window distribution creation unit 34 creates a six-dimensional vector.

[0125] The qualitative data correlation analysis unit 36 ​​calculates a correlation coefficient r for each dimension from the vector for each combination of brand and gender and for each window. i Calculate the correlation coefficient r i The correlation coefficient r between the combination of brand and gender can be obtained by averaging the above. The qualitative data correlation analysis unit 36 ​​can obtain the correlation coefficient r using the following formulas (3) to (6).

[0126]

number

number

number

number

[0127] 35 is an explanatory diagram of step S205. The qualitative data correlation analysis unit 36 ​​searches for a combination of stock and gender for which the correlation coefficient r falls within a threshold range. The qualitative data correlation analysis unit 36 ​​uses a threshold for the correlation coefficient r and determines that combinations of stock and gender within the threshold range are uncorrelated.

[0128] (Fraud detection phase) The processing performed by the traceability system 12 in the fraud detection phase will be described with reference to Fig. 36. Fig. 36 is a flowchart showing the processing performed by the traceability system 12 in the fraud detection phase.

[0129] Steps S211 to S213 shown in Fig. 36 are steps for detecting padding candidates based on fluctuations in probability distribution, and steps S214 to S218 are steps for detecting padding based on fluctuations from no correlation to negative correlation.

[0130] The distribution fluctuation detection unit 51 divides the new transaction history into combinations of brand and gender (S211). The distribution fluctuation detection unit 51 calculates likelihood L for each combination of brand and gender from the transaction history and probability distribution parameters for each combination of brand and gender (S212). The distribution fluctuation detection unit 51 determines whether there is a combination of brand and gender for which likelihood L is less than a threshold (S213).

[0131] If it is determined that there is a combination of brand and gender for which the likelihood L is less than the threshold (Yes in S213), the distribution fluctuation detection unit 51 identifies the combination as a padding candidate and proceeds to step S214. If it is determined that there is no combination of brand and gender for which the likelihood L is less than the threshold (No in S213), the distribution fluctuation detection unit 51 ends the process.

[0132] The qualitative data correlation fluctuation detection unit 52 divides the transaction history of the combination of stock and gender that is a candidate for padding using sliding window processing (S214).The qualitative data correlation fluctuation detection unit 52 uses the transaction history for each combination of stock and gender and for each window to learn the parameters of the probability distribution (S215).

[0133] The qualitative data correlation variation detection unit 52 uses the parameters of the probability distribution to calculate the correlation coefficient r between the combinations of stocks and genders that are candidates for padding (S216).The qualitative data correlation variation detection unit 52 determines whether there are any combinations of stocks and genders that have changed from no correlation to a negative correlation (S217).

[0134] If it is determined that there is a combination of stock and gender that has changed from no correlation to a negative correlation (Yes in S217), the qualitative data correlation variation detection unit 52 detects the combination of stock and gender that has changed to a negative correlation as inflated (S218).If it is determined that there is no combination of stock and gender that has changed from no correlation to a negative correlation (No in S217), the qualitative data correlation variation detection unit 52 ends the process.

[0135] Figures 37 to 47 are diagrams for explaining in detail each step shown in Figure 36. Of these, Figures 37 to 42 show an example in which the traceability system 12 detects seasonal fluctuations, and Figures 43 to 47 show an example in which the traceability system 12 detects padding.

[0136] (Detects seasonal fluctuations) First, an example of detecting seasonal variations will be described with reference to Figures 37 to 42. Steps S211 to S213 are generally similar to steps S111 to S113 described with reference to Figures 16 to 19, and therefore detailed description thereof will be omitted.

[0137] Here, it is assumed that 10 days have passed since the learning phase, and that 10 days' worth of new handling history (11 / 11 to 11 / 20) has been recorded. It is also assumed that the weight loss of both the P male pigs and the R male pigs has decreased due to seasonal fluctuations. Note that the "probability distribution parameters for the P male pigs" shown in Figure 18 should be read as "probability distribution parameters for the P male pigs and the latest window." Furthermore, the "inflated brand and gender combination" shown in Figure 19 should be read as "candidate brand and gender combination for inflating."

[0138] 37 is an explanatory diagram of step S214. The qualitative data correlation variation detection unit 52 divides each of the handling histories of the padding candidates using sliding window processing. If the window width is 7 days and the sliding width is 1 day, the handling history for 10 days is divided into four windows t1 to t4.

[0139] 38 is an explanatory diagram of step S215. The qualitative data correlation variation detection unit 52 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood for the handling history. The qualitative data correlation variation detection unit 52 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0140] 39 and 40 are explanatory diagrams of step S216. The qualitative data correlation variation detection unit 52 performs preprocessing in the same manner as in step S204 of the learning phase. The qualitative data correlation variation detection unit 52 combines the mean vector μ→ and the one-dimensional covariance matrix Σ.

[0141] The qualitative data correlation variation detection unit 52 calculates the correlation coefficient r between the combinations of brand and gender in the same dimension as in step S204 of the learning phase. i The qualitative data correlation variation detection unit 52 calculates the correlation coefficient r between the probability distributions by calculating the correlation coefficient r for each dimension. i is calculated by averaging.

[0142] Due to seasonal fluctuations, the weight loss of both P and R male pigs decreased. For example, in windows t1 and t2, the post-processing weights of P and R male pigs increased simultaneously. Therefore, there is a positive correlation between them.

[0143] 41 and 42 are explanatory diagrams of step S217. The qualitative data correlation variation detection unit 52 searches for a combination of brand and gender for which the correlation coefficient r is less than a threshold. The qualitative data correlation variation detection unit 52 determines that a probability distribution for which the correlation coefficient r is less than the threshold is a negative correlation.

[0144] As shown in Figure 41, the combination of brand and gender that has a negative correlation here is a null value. The qualitative data correlation fluctuation detection unit 52 detects a combination that is included in the pre-specified non-correlated combinations and is determined to have a negative correlation as being inflated. Here, the qualitative data correlation fluctuation detection unit 52 does not detect inflated data.

[0145] (Inflation detected) Next, an example of detecting padding will be described with reference to Figures 43 to 47. Content that overlaps with the example of detecting seasonal fluctuations described above will be omitted as appropriate.

[0146] Here, it is assumed that 10 days have passed since the learning phase, and new handling history for 10 days (11 / 11 to 11 / 20) has been recorded. Also, it is assumed that the weight of the P male pig has been inflated with the weight of the R male pig.

[0147] The new handling history is the same as in Figure 16, so it is not shown here. The average weight of P male pigs after processing is 71 kg, and the weight is inflated to 72 kg and registered. The average weight of R male pigs after processing is 81 kg, and the weight is reduced to 78 kg and registered. This example will be used below.

[0148] 43 is an explanatory diagram of step S215. The qualitative data correlation variation detection unit 52 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood for the handling history. The qualitative data correlation variation detection unit 52 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0149] 44 and 45 are explanatory diagrams of step S216. The qualitative data correlation variation detection unit 52 performs preprocessing in the same manner as in step S204 of the learning phase. The qualitative data correlation variation detection unit 52 combines the mean vector μ→ and the one-dimensional covariance matrix Σ.

[0150] The qualitative data correlation variation detection unit 52 calculates the correlation coefficient r between the combinations of brand and gender in the same dimension as in step S204 of the learning phase. i The qualitative data correlation variation detection unit 52 calculates the correlation coefficient r between the probability distributions by calculating the correlation coefficient r for each dimension. i is calculated by averaging.

[0151] Focusing on windows t1 and t2, the post-processing weight of P male pigs increases while the post-processing weight of R male pigs decreases, indicating a negative correlation between them.

[0152] 46 and 47 are explanatory diagrams of step S217. The qualitative data correlation variation detection unit 52 searches for a combination of brand and gender for which the correlation coefficient r is less than a threshold. The qualitative data correlation variation detection unit 52 determines that a probability distribution for which the correlation coefficient r is less than the threshold is a negative correlation.

[0153] As shown in Figure 46, the combination of brand and sex that has a negative correlation here is P male pigs and R male pigs. The qualitative data correlation variation detection unit 52 detects as inflated combinations that are included in the pre-specified non-correlated combinations and are determined to be negatively correlated. Here, the qualitative data correlation variation detection unit 52 detects as inflated the processing of P male pigs and R male pigs.

[0154] As described above, the traceability system 12 according to the present disclosure identifies qualitative data corresponding to a current probability distribution as a padding candidate, which may have been padded, based on the comparison result between the current probability distribution and the past probability distribution. The traceability system 12 also detects padding from among the padding candidates based on the temporal fluctuations of the probability distribution corresponding to the padding candidates. The traceability system 12 analyzes the correlation between the past probability distributions of different qualitative data, identifies combinations of qualitative data corresponding to uncorrelated past probability distributions, and identifies, among these combinations, combinations of qualitative data for which the current probability distribution has shifted to a negative correlation.

[0155] With this configuration, the traceability system 12 monitors the correlation between the fluctuations in the probability distribution of weight data of expensive and inexpensive foods before and after processing for each food business. In this way, the traceability system 12 can identify fluctuations in the probability distribution that are specific to padding. This allows the traceability system 12 to reduce false positives.

[0156] <Embodiment 3> Next, a description will be given of embodiment 3. First, in order to explain this embodiment, correlation between businesses in fluctuations in probability distribution will be described.

[0157] Factors other than padding that cause fluctuations in the probability distribution of weight data before and after processing of food include those that simultaneously affect the weight loss of food with the same qualitative data (such as brand) handled by multiple businesses. When fluctuations in probability distribution are caused by such factors, the fluctuations in the probability distribution of weight data before and after processing for the same qualitative data are strongly correlated among multiple businesses.

[0158] However, if padding is the cause of fluctuations in the probability distribution, the fluctuations in the probability distribution of a particular operator will be weakly correlated with the fluctuations in the probability distribution of other operators. In most cases, padding is not carried out by multiple operators simultaneously. Therefore, when padding occurs, one operator will cause fluctuations in the probability distribution.

[0159] For example, among businesses a to c that have a strong correlation, suppose business a no longer correlates with other businesses b and c. In such a case, the traceability system can detect fraud against business a.

[0160] The traceability system 13 according to the present disclosure reduces false positives of padding by finding strong correlations between multiple businesses in fluctuations in probability distribution. Specifically, the traceability system 13 analyzes the correlation between business A and other businesses, and identifies other businesses that correlate with business A. In cases other than padding, there is a correlation with other businesses, and in cases of padding, there is no correlation with other businesses. By detecting this, the traceability system 13 detects padding by a specific business.

[0161] (Structure of the traceability system 13) Fig. 48 is a block diagram showing the configuration of the traceability system 13 according to the present disclosure. Figs. 49 to 50 are diagrams explaining in detail each component of the traceability system 13. Similar to the traceability systems 11 and 12 described above, the traceability system 13 includes a handling history creation unit 1, a handling history storage unit 2, a learning unit 3, a handling history tracking unit 4, and a fraud detection unit 5.

[0162] The learning unit 3 and the fraud detection unit 5 each have functions different from those of the traceability system 12. Specifically, the learning unit 3 has a per-business correlation analysis unit 37 instead of the per-qualitative data correlation analysis unit 36. The per-business correlation analysis unit 37 may be implemented as a smart contract. In addition, the moving-window distribution creation unit 34 and the moving-window distribution storage unit 35 have functions different from those of the traceability system 12. The learning unit 3 analyzes the correlation between businesses for each qualitative data based on the parameters of the probability distribution of qualitative data (such as brand) and weight data of food before and after processing for each business, and identifies combinations of businesses that are strongly correlated.

[0163] Furthermore, the fraud detection unit 5 includes a per-business operator correlation variation detection unit 54 and a business operator correlation combination storage unit 55, instead of the per-qualitative data correlation variation detection unit 52 and the qualitative data uncorrelated combination storage unit 53. The per-business operator correlation variation detection unit 54 may be implemented as a smart contract. The fraud detection unit 5 considers that, among pairs of businesses and qualitative data whose probability distributions have changed, pairs that maintain a pre-specified strong correlation are not inflated.

[0164] The traceability system 13 according to the present disclosure achieves the following two things. (i) Detect fluctuations in the qualitative data of food products (such as brands) and the probability distribution for each business operator. (ii) When the correlation between the parameter fluctuations of the probability distributions of one business operator and another business operator among the probability distributions for the same qualitative data changes from a strong correlation to a weak correlation, this is detected as padding.

[0165] The per-enterprise correlation analysis unit 37 and per-enterprise correlation fluctuation detection unit 54 realize the above (ii). The following will mainly explain the differences from the traceability system 12, and will omit overlapping points as appropriate.

[0166] The moving-window distribution creation unit 34 acquires the handling histories of multiple businesses from the handling history storage unit 2. The moving-window distribution creation unit 34 performs sliding window processing on the acquired data and learns the parameters of the probability distribution of the weight data of the food before and after processing for each window. The moving-window distribution creation unit 34 saves the learning results in the moving-window distribution storage unit 35.

[0167] The moving window distribution memory unit 35 sets a window, shifts the window along the time axis, and records the parameters of the probability distribution learned based on the handling history within the window for each combination of qualitative data of the food (brand, gender, etc.) and business operator.

[0168] The per-business correlation analysis unit 37 analyzes the correlation between past probability distributions of qualitative data for each business that processes the product. The per-business correlation analysis unit 37 analyzes the correlation for all combinations of businesses with respect to parameter fluctuations in the probability distribution for each business related to the same qualitative data. The per-business correlation analysis unit 37 identifies combinations of businesses that correspond to past probability distributions with a correlation stronger than a predetermined value. As a result, the per-business correlation analysis unit 37 identifies combinations of businesses that are strongly correlated.

[0169] Specifically, the per-carrier correlation analysis unit 37 acquires qualitative data and parameters of probability distribution for each window from the per-moving-window distribution storage unit 35 prepared for each carrier. The per-carrier correlation analysis unit 37 calculates a correlation coefficient for each combination of carriers in response to parameter fluctuations in the probability distribution for each carrier related to the same qualitative data. The per-carrier correlation analysis unit 37 compares the correlation coefficient with a threshold to identify combinations of carriers that are strongly correlated. The per-carrier correlation analysis unit 37 stores the identified combinations of carriers in the carrier correlation combination storage unit 55.

[0170] 49 is a diagram showing an example of input / output data in the per-carrier correlation analysis unit 37. The internal processing of the per-carrier correlation analysis unit 37 will be described with reference to the same figure. The per-carrier correlation analysis unit 37 acquires qualitative data and parameters of the probability distribution learned for each window from the per-carrier distribution storage unit 35 created for each carrier. The per-carrier correlation analysis unit 37 groups the acquired parameters of the probability distribution by carrier and qualitative data.

[0171] The per-business correlation analysis unit 37 performs correlation analysis of the fluctuations in probability distribution parameters over time between groups with the same qualitative data but different businesses to determine a correlation coefficient. If the correlation coefficient exceeds a threshold, the per-business correlation analysis unit 37 identifies the combination of businesses as correlated. The per-business correlation analysis unit 37 transmits the correlated combination of businesses to the business correlation combination storage unit 55.

[0172] Furthermore, the per-business correlation analysis unit 37 may identify a combination of businesses that corresponds to an uncorrelated past probability distribution, and store the combination in the business correlation combination storage unit 55 .

[0173] The per-service provider correlation fluctuation detection unit 54 detects padding based on the analysis results of the per-service provider correlation analysis unit 37. The per-service provider correlation fluctuation detection unit 54 identifies, from among the identified combinations of services with strong correlations, combinations of services for which the current probability distribution has changed to a weak correlation below a predetermined level, and detects padding based on the identification results. In this way, the per-service provider correlation fluctuation detection unit 54 identifies services for which the correlation has changed from strong to weak.

[0174] Specifically, in response to input of a list of pairs of operators and qualitative data from the distribution fluctuation detection unit 51, the per-operator correlation fluctuation detection unit 54 acquires parameters of the probability distribution corresponding to the pairs of operators and qualitative data in the list from the moving-window distribution storage unit 35. In addition, the per-operator correlation fluctuation detection unit 54 acquires combinations of operators that have a strong correlation from the operator correlation combination storage unit 55.

[0175] The per-enterprise correlation variation detection unit 54 calculates the correlation coefficient of the parameter variation of the probability distribution for all combinations of qualitative data and enterprises in the list. The per-enterprise correlation variation detection unit 54 identifies combinations of enterprises with weak correlations by comparing the correlation coefficients with a threshold. The per-enterprise correlation variation detection unit 54 detects enterprises whose correlation has changed from strong to weak as being inflated. The per-enterprise correlation variation detection unit 54 outputs the detected enterprises to the auditor.

[0176] Figure 50 is a diagram showing an example of input / output data in the per-business correlation variation detection unit 54. The internal processing of the per-business correlation variation detection unit 54 will be described with reference to the same figure. The per-business correlation variation detection unit 54 receives food qualitative data, a list of businesses, and corresponding sets of pre-processing weight and post-processing weight from the distribution variation detection unit 51. The per-business correlation variation detection unit 54 obtains from the business correlation combination storage unit 55 combinations of businesses that are strongly correlated with the business for which a distribution variation has been detected.

[0177] The per-enterprise correlation fluctuation detection unit 54 learns probability distribution parameters from pairs of weights before and after processing of the detected qualitative data. The per-enterprise correlation fluctuation detection unit 54 calculates the correlation coefficient between the changed probability distributions and identifies combinations with weak correlations that fall below a threshold. If the identified combination is included in a combination of strongly correlated operators, the per-enterprise correlation fluctuation detection unit 54 detects it as padding.

[0178] The per-enterprise correlation fluctuation detection unit 54 may identify combinations of entities whose current probability distribution has changed to correlated among the identified combinations of uncorrelated entities, and detect padding based on the identification results. The correlation may be positive or negative. For example, if multiple entities collude to pad data, a positive correlation may occur between entities that were previously uncorrelated.

[0179] The business operator correlation combination storage unit 55 stores data transmitted from the business operator correlation analysis unit 37. For example, the business operator correlation combination storage unit 55 stores a combination of business operators that are strongly correlated. The business operator correlation combination storage unit 55 may also store a combination of business operators that are uncorrelated.

[0180] (Example of Embodiment 3) Next, a description will be given of an example of embodiment 3. The processing performed by the traceability system 13 is divided into a learning phase and a fraud detection phase.

[0181] In the learning phase, the traceability system 13 learns past handling histories for each combination of brand, gender, and business operator over a shifting period, and obtains temporal fluctuations in the parameters of the probability distribution. The traceability system 13 finds strongly correlated combinations for each combination of brand and gender based on the correlation coefficient between businesses in the temporal fluctuations.

[0182] In the fraud detection phase, the traceability system 13 uses the new handling history to compare the probability distribution of the weight data before and after food processing in the learning phase and the fraud detection phase, and finds combinations of brands, genders, and business operators for which the distribution has changed. The traceability system 13 detects fraud when a combination of business operators with a strong correlation has changed to a weak correlation.

[0183] Figure 51 is a diagram explaining the processing of the traceability system 13. The vertical axis indicates the weight after processing, and the horizontal axis indicates the weight before processing. The ellipses represent the probability distribution. In this figure, the ellipses show the probability distributions learned for brand a1 of business operator a and brand b1 of business operator b, each learned over a different period. The probability distribution for brand a1 of business operator a shifts upward in the order of periods T1, T2, and T3. The probability distribution for brand b1 of business operator b also shifts upward in the order of periods T1, T2, and T3. The traceability system 13 detects this as normal when there is a strong correlation between the fluctuations in the probability distributions for each business operator.

[0184] (Traceability System 13 Processing) Next, a description will be given of the processing performed by the traceability system 13. First, a scenario of this embodiment will be described.

[0185] Slaughterer A, Slaughterer B, and Slaughterer C process three types of brands (P pork, Q pork, and R pork), and each brand has two types of sex (male and female). The traceability system 13 stores the handling history. The handling history includes weight data before and after processing. The traceability system 13 detects padding by monitoring the handling history.

[0186] In the learning phase, the traceability system 13 learns the probability distribution of weight data before and after processing for each combination (18 types) of brand, sex, and business operator. The traceability system 13 identifies business operator combinations that are strongly correlated for each combination of brand and sex, and for P male pigs and R male pigs, there was a strong correlation between slaughterman A, slaughterman B, and slaughterman C, respectively.

[0187] In the fraud detection phase, the traceability system 13 detects P male pigs and R male pigs, whose weight loss has decreased due to seasonal fluctuations, as a combination of brand and sex that may be inflated based on fluctuations in probability distribution. However, because the probability distributions of slaughterers A, B, and C maintained a strong correlation, the traceability system 13 determines that this is not inflated.

[0188] Furthermore, suppose that slaughterers A, B, and C processed P male pigs and R male pigs whose weight loss has decreased due to seasonal fluctuations. Slaughterer A inflated the weight of the P male pigs with R male pigs to bring the weight of the P male pigs closer to the average weight loss. Based on fluctuations in the probability distribution, the traceability system 13 identifies P male pigs and R male pigs as combinations of brand and sex that may have been inflated. Because the probability distribution of slaughterer A fluctuated from a strong correlation to a weak correlation with slaughterers B and C, the traceability system 13 determines that the weight has been inflated.

[0189] (Learning phase) The processing performed by the traceability system 13 in the learning phase will be described with reference to Fig. 52 to Fig. 59. Fig. 52 is a flowchart showing the processing performed by the traceability system 13 in the learning phase. Fig. 53 to Fig. 59 are diagrams for explaining each step shown in Fig. 52 in detail.

[0190] Steps S301 to S303 shown in Fig. 52 are steps for learning a probability distribution for each combination of brand, gender, and company and for each window. Steps S304 and S305 are steps for identifying combinations of companies that are strongly correlated.

[0191] The moving-window distribution creation unit 34 divides all handling histories into combinations of brand, gender, and business operator (S301). The moving-window distribution creation unit 34 divides the handling histories for each combination of brand, gender, and business operator using sliding window processing (S302). The moving-window distribution creation unit 34 uses the handling histories for each combination of brand, gender, and business operator and for each window to learn parameters for the probability distribution of weight data before and after food processing (S303).

[0192] The per-business correlation analysis unit 37 uses the probability distribution parameters to calculate the correlation coefficient between businesses for each combination of brand and gender (S304).The per-business correlation analysis unit 37 uses the correlation coefficient to identify and save combinations of businesses that are strongly correlated for each combination of brand and gender (S305).

[0193] Figures 53 and 54 are explanatory diagrams of step S301. Figure 53 shows an example of a handling history that is the premise for explaining this process. The handling history consists of columns for brand, sex, weight before processing, weight after processing, business name, and work date and time. There are three types of businesses (slaughterhouse A, slaughterhouse B, and slaughterhouse C), three types of brands (P pigs, Q pigs, and R pigs), and two types of sex (male and female). The weight of each individual food item is recorded before and after processing. The handling history covers the past 10 days (11 / 01 to 11 / 10). This example will be used for the following explanation.

[0194] The moving-window distribution creation unit 34 divides all handling histories of slaughterers A, B, and C into combinations of brand and sex. The moving-window distribution creation unit 34 extracts records for each combination of brand and sex.

[0195] 55 is an explanatory diagram of step S302. The moving-window distribution creation unit 34 divides the transaction history for each combination of brand, gender, and business operator using sliding window processing. If the window width is 7 days and the sliding width is 1 day, the transaction history for 10 days is divided into four windows t1 to t4.

[0196] 56 is an explanatory diagram of step S303. The moving-window distribution creation unit 34 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood L for the handling history. The moving-window distribution creation unit 34 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0197] 57 and 58 are explanatory diagrams of step S304. As a preprocessing for obtaining the correlation coefficient, the per-company correlation analysis unit 37 generates a vector combining the mean vector μ→ and the one-dimensional covariance matrix Σ for each combination of brand, gender, and company, and for each window.

[0198] The company-specific correlation analysis unit 37 calculates a correlation coefficient r for each dimension from the vectors for each combination of brand, gender, and company and for each window. i Calculate the correlation coefficient r i By averaging these, we can obtain the correlation coefficient r between businesses for each combination of brand and gender. As the window becomes newer, the processed weights of Slaughterer A's P male pigs and Slaughterer B's P male pigs increase simultaneously. Therefore, there is a positive correlation between them.

[0199] 59 is an explanatory diagram of step S305. The per-business correlation analysis unit 37 uses a threshold value for the correlation coefficient r, and determines that a combination of businesses within the threshold range has a strong correlation. The per-business correlation analysis unit 37 searches for a combination of businesses whose correlation coefficient r falls within the threshold range.

[0200] (Fraud detection phase) The processing performed by the traceability system 13 in the fraud detection phase will be described with reference to Fig. 60. Fig. 60 is a flowchart showing the processing performed by the traceability system 13 in the fraud detection phase.

[0201] Steps S311 to S313 shown in Fig. 60 are steps for detecting padding candidates based on fluctuations in probability distribution, and steps S314 to S318 are steps for detecting padding based on fluctuations from strong correlation to weak correlation.

[0202] The distribution fluctuation detection unit 51 divides the new transaction history into combinations of brand, gender, and business operator (S311). The distribution fluctuation detection unit 51 calculates the likelihood for each combination of brand, gender, and business operator from the transaction history and probability distribution parameters for each combination of brand, gender, and business operator (S312). The distribution fluctuation detection unit 51 determines whether there is a combination of brand, gender, and business operator whose likelihood is less than a threshold value (S313).

[0203] If it is determined that there is a combination of brand and gender for which the likelihood L is less than the threshold (Yes in S313), the distribution fluctuation detection unit 51 identifies the combination as a padding candidate and proceeds to step S314. If it is determined that there is no combination of brand and gender for which the likelihood L is less than the threshold (No in S313), the distribution fluctuation detection unit 51 ends the process.

[0204] The per-business correlation fluctuation detection unit 54 divides the transaction history of the combination of brand, gender, and business operator that is a candidate for padding using sliding window processing (S314). The per-business correlation fluctuation detection unit 54 learns probability distribution parameters from the transaction history for each combination of brand, gender, and business operator and for each window (S315).

[0205] The per-business correlation fluctuation detection unit 54 calculates the correlation coefficient r between businesses for each combination of stock and gender of padding candidates (S316). The per-business correlation fluctuation detection unit 54 determines whether there is a business whose correlation has changed from strong to weak (S317).

[0206] If it is determined that there is a business operator whose correlation has changed from strong to weak (Yes in S317), the per-business correlation change detection unit 54 detects the combination of brand, gender, and business operator whose correlation has changed to weak as inflated (S318).If it is determined that there is no business operator whose correlation has changed from strong to weak (No in S317), the per-business correlation change detection unit 54 ends the processing.

[0207] Here, the per-service provider correlation fluctuation detection unit 54 determines whether there is a service provider whose correlation has changed from strong to weak, but this is not limiting. The per-service provider correlation fluctuation detection unit 54 may also identify a combination of service providers whose correlation has changed from uncorrelated to correlated, and detect padding based on the identification result.

[0208] Figures 61 to 76 are diagrams for explaining in detail each step shown in Figure 60. Of these, Figures 61 to 70 show an example in which the traceability system 13 detects seasonal fluctuations, and Figures 71 to 76 show an example in which the traceability system 13 detects padding.

[0209] (Detects seasonal fluctuations) First, an example of detecting seasonal fluctuations will be described with reference to Figures 61 to 70. Figures 61 and 62 are explanatory diagrams of step S311. Figure 61 shows an example of new handling history that is the premise for the explanation of this process. Here, it is assumed that 10 days have passed since the learning phase, and new handling history for 10 days (11 / 11 to 11 / 20) has been recorded. Due to seasonal fluctuations, weight loss has decreased for P male pigs and R male pigs at slaughterhouses A, B, and C. This example will be used hereafter.

[0210] The distribution fluctuation detection unit 51 divides the new handling history of slaughterers A, B, and C into each combination of brand and sex. The distribution fluctuation detection unit 51 extracts records for each combination of brand and sex.

[0211] 63 is an explanatory diagram of step S312. The distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand, gender, and business operator from the transaction history and probability distribution parameters for each combination of brand, gender, and business operator.

[0212] Specifically, the distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand, gender, and business operator using the transaction history for each combination of brand, gender, and business operator, and the mean vector μ→ and covariance matrix Σ learned from the most recent window. In this way, the distribution fluctuation detection unit 51 measures the degree to which the probability distribution learned in the learning phase fits the new transaction history.

[0213] 64 is an explanatory diagram of step S313. The distribution fluctuation detection unit 51 searches for a combination of brand, gender, and business operator for which the likelihood L is less than the threshold. The distribution fluctuation detection unit 51 determines that a combination of brand, gender, and business operator for which the likelihood L is less than the threshold is a padding candidate.

[0214] 65 is an explanatory diagram of step S314. The per-carrier correlation fluctuation detection unit 54 divides each of the padding candidate handling histories using sliding window processing. If the window width is 7 days and the sliding width is 1 day, then 10 days of handling history is divided into four windows t1 to t4.

[0215] 66 is an explanatory diagram of step S315. The per-business correlation variation detection unit 54 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood L for the handling history. The per-business correlation variation detection unit 54 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0216] 67 and 68 are explanatory diagrams of step S316. The per-operator correlation fluctuation detection unit 54 performs preprocessing in the same way as in step S304 in the learning phase. The per-operator correlation fluctuation detection unit 54 combines the mean vector μ→ and the one-dimensional covariance matrix Σ.

[0217] The per-company correlation fluctuation detection unit 54 calculates the correlation coefficient r between companies for each combination of brand and gender, similar to step S304 in the learning phase. Specifically, the per-company correlation fluctuation detection unit 54 calculates the correlation coefficient r between companies for the same dimension for different combinations of brand and gender.i The per-carrier correlation fluctuation detection unit 54 calculates the correlation coefficient r i The correlation coefficient r between the probability distributions is calculated by averaging the values.

[0218] The weight of the processed meat increased significantly for both Slaughterhouse A and Slaughterhouse B. Therefore, there is a positive correlation between them.

[0219] 69 and 70 are explanatory diagrams of step S317. The per-service provider correlation fluctuation detection unit 54 searches for a combination of services for which the correlation coefficient r falls within the threshold range. The per-service provider correlation fluctuation detection unit 54 determines that a combination of services for which the correlation coefficient r is less than the threshold has a weak correlation.

[0220] As shown in Figure 69, here, a combination of operators for which the correlation coefficient r is less than the threshold is a null value. When a combination of operators that are strongly correlated and a combination of operators that are weakly correlated match, the per-operator correlation fluctuation detection unit 54 detects this as padding. Here, the per-operator correlation fluctuation detection unit 54 does not detect padding.

[0221] (Inflation detected) Next, an example of detecting padding will be described with reference to Figures 71 to 76. Content that overlaps with the example of detecting seasonal fluctuations described above will be omitted as appropriate.

[0222] Figure 71 is an explanatory diagram of step S311. Figure 71 shows an example of a new handling history that is the premise for explaining this process. Here, it is assumed that 10 days have passed since the learning phase, and new handling history for 10 days (11 / 11 to 11 / 20) has been recorded. Suppose that slaughterer A inflates the weight of P male pigs with R male pigs. The average post-processing weight of P male pigs is 71 kg, and the registered weight is inflated to 72 kg. The average post-processing weight of R male pigs is 81 kg, and the registered weight is reduced to 78 kg. This example will be used hereafter.

[0223] 72 is an explanatory diagram of step S315. The business operator correlation variation detection unit 54 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood L for the handling history. The business operator correlation variation detection unit 54 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0224] 73 and 74 are explanatory diagrams of step S316. The per-operator correlation fluctuation detection unit 54 performs preprocessing in the same way as in step S304 in the learning phase. The per-operator correlation fluctuation detection unit 54 combines the mean vector μ→ and the one-dimensional covariance matrix Σ.

[0225] The per-company correlation fluctuation detection unit 54 calculates the correlation coefficient r between companies for each combination of brand and gender, similar to step S304 in the learning phase. Specifically, the per-company correlation fluctuation detection unit 54 calculates the correlation coefficient r between companies for the same dimension for different combinations of brand and gender. i The per-carrier correlation fluctuation detection unit 54 calculates the correlation coefficient r i The correlation coefficient r between the probability distributions is calculated by averaging the values.

[0226] While there is a large increase in processed weight at Slaughterhouse A, there is a small decrease at Slaughterhouse B. Therefore, there is a weak negative correlation between them.

[0227] 75 and 76 are explanatory diagrams of step S317. The per-service provider correlation fluctuation detection unit 54 searches for a combination of services for which the correlation coefficient r falls within the threshold range. The per-service provider correlation fluctuation detection unit 54 determines that a combination of services for which the correlation coefficient r is less than the threshold has a weak correlation.

[0228] As shown in Figure 75, the combinations of businesses for which the correlation coefficient r is less than the threshold are the combination of slaughterers A and B, and the combination of slaughterers A and C. The per-business correlation fluctuation detection unit 54 detects this as padding when a combination of businesses that is strongly correlated and a combination of businesses that is weakly correlated, which have been specified in advance, match. Here, the per-business correlation fluctuation detection unit 54 detects padding for slaughterer A.

[0229] As described above, the traceability system 13 according to the present disclosure analyzes the correlation between past probability distributions of qualitative data for each business that processes products and detects padding based on the analysis results. For example, the traceability system 13 identifies combinations of businesses that correspond to past probability distributions with a strong correlation greater than a predetermined value, and among the identified combinations of businesses with strong correlations, identifies combinations of businesses whose current probability distribution has changed to a weak correlation less than a predetermined value. The traceability system 13 detects padding based on the identification results.

[0230] With this configuration, the traceability system 13 manages fluctuations in the probability distribution of weight data before and after processing for foods with the same qualitative data using correlations between businesses. The traceability system 13 can identify fluctuations due to factors other than padding that simultaneously affect the weight loss of foods handled by multiple businesses. This allows the traceability system 13 to reduce false positives.

[0231] <Embodiment 4> Next, a fourth embodiment will be described. Factors other than padding cause similar fluctuations in probability distribution over time each time they occur. The traceability system 14 according to the present disclosure utilizes this to detect fluctuations in probability distribution and then finds past cases with similar fluctuations in probability distribution over time, thereby reducing false positives.

[0232] Specifically, when the traceability system 14 detects a fluctuation in the probability distribution as being inflated, it manually examines and records the fluctuation pattern of the probability distribution caused by factors other than the inflated value. The traceability system 14 assumes that when the cause of the fluctuation in the probability distribution is the same, the accompanying fluctuation pattern will be similar.

[0233] (Configuration of Traceability System 14) Fig. 77 is a block diagram showing the configuration of the traceability system 14 according to the present disclosure. Figs. 78 to 80 are diagrams for explaining each component of the traceability system 14 in detail.

[0234] The traceability system 14 includes a handling history creation unit 1, a handling history storage unit 2, a learning unit 3, a handling history tracking unit 4, and a fraud detection unit 5, similar to the above-described traceability systems 11 to 13.

[0235] The learning unit 3 and fraud detection unit 5 each have functions different from those of the traceability systems 11 to 13. Specifically, the learning unit 3 only has a moving-window distribution creation unit 34 and a moving-window distribution storage unit 35. However, this is not limited to this, and the learning unit 3 may have other functional units. Furthermore, the fraud detection unit 5 has a fluctuation pattern matching detection unit 56 and a past case storage unit 57 in addition to the distribution fluctuation detection unit 51. The distribution fluctuation detection unit 51 may be implemented as a smart contract.

[0236] The fraud detection unit 5 registers fluctuation patterns of the probability distribution of weight data of food before and after processing due to factors other than padding that have occurred in the past, and determines that the data is normal if there is a past case with a similar fluctuation pattern.

[0237] The traceability system 14 according to the present disclosure achieves the following two things. (i) Detect fluctuations in the probability distribution for each qualitative data (brand, etc.) of food products for a specific business operator. (ii) The fluctuation pattern of the probability distribution is compared with past cases, and if the fluctuation pattern is similar to past cases, it is determined that a factor other than padding is the cause, and if there are no similar cases, it is determined that padding is the cause.

[0238] The fluctuation pattern matching detection unit 56 realizes (ii). In the following, differences from the above-described traceability systems 11 to 13 will be mainly explained, and overlapping points will be omitted as appropriate.

[0239] The fluctuation pattern matching detection unit 56 identifies a fluctuation pattern of the probability distribution caused by factors other than padding based on multiple past probability distributions, compares the identified fluctuation pattern with the fluctuation pattern of the current probability distribution, and detects padding based on the comparison result. Specifically, in response to input of a list of pairs of qualitative data and business operators from the distribution fluctuation detection unit 51, the fluctuation pattern matching detection unit 56 acquires from the moving-window distribution storage unit 35 the parameters of the probability distribution corresponding to the qualitative data in the list.

[0240] The fluctuation pattern matching detection unit 56 acquires fluctuation patterns of probability distribution due to factors other than padding from the past case storage unit 57. The fluctuation pattern matching detection unit 56 calculates the similarity between the parameter fluctuation of the probability distribution for each qualitative data in the list and the fluctuation pattern of the probability distribution due to factors other than padding. The fluctuation pattern matching detection unit 56 compares the similarity with a threshold to determine whether there are any past cases with similar fluctuation patterns. If there are no similar cases, the fluctuation pattern matching detection unit 56 detects it as padding and reports the detected qualitative data to the auditor.

[0241] The past case storage unit 57 stores fluctuations in the parameters of the probability distribution of weight data of food before and after processing due to factors other than padding.

[0242] FIG. 78 is a diagram showing an example of fluctuations in the parameters of the probability distribution of weight data of food before and after processing due to factors other than padding, stored in the past case storage unit 57.

[0243] 79 and 80 are diagrams showing examples of input and output data in the variation pattern matching detection unit 56. The internal processing of the variation pattern matching detection unit 56 will be described with reference to these figures. The variation pattern matching detection unit 56 receives a list of food quality data and corresponding weight data before and after processing from the distribution variation detection unit 51.

[0244] The fluctuation pattern matching detection unit 56 acquires the parameters of the probability distribution corresponding to the qualitative data from the moving window distribution storage unit 35. The fluctuation pattern matching detection unit 56 learns the parameters of the probability distribution from the weight data before and after processing corresponding to the detected qualitative data.

[0245] The fluctuation pattern matching detection unit 56 creates time series data for each qualitative data from the probability distribution parameters obtained from [1] and [2]. The fluctuation pattern matching detection unit 56 acquires fluctuation patterns of the probability distribution due to factors other than padding from the past case storage unit 57. The fluctuation pattern matching detection unit 56 calculates the similarity between the created time series data and past fluctuation patterns using DTW (Dynamic Time Warping) or the like. The fluctuation pattern matching detection unit 56 determines that the data is normal if there is a past case showing a similarity exceeding a threshold, and detects padding otherwise.

[0246] Figure 81 is a diagram explaining the processing of the traceability system 14. The vertical axis indicates the weight after processing, and the horizontal axis indicates the weight before processing. The ellipses represent the probability distribution. In this figure, the ellipses show the probability distributions learned by shifting the time periods for the fluctuations in the probability distribution of weight data due to seasonal fluctuations and the fluctuations in the probability distribution of weight data for brand a1. Both probability distributions are shifted upward in the order of periods T1, T2, and T3. The traceability system 14 detects this as normal because there are past cases in which the fluctuations in the probability distribution over time are similar.

[0247] In the learning phase, the traceability system 14 learns the past handling history for each combination of brand and gender over different periods, thereby obtaining the temporal fluctuations of the parameters of the probability distribution.

[0248] (Example of Embodiment 4) Next, an example of embodiment 4 will be described. In the fraud detection phase, the traceability system 14 uses the new handling history to find combinations of brand and gender for which the probability distribution of weight data before and after food processing has changed between the learning phase and the fraud detection phase for each combination of brand and gender. The traceability system 14 detects fraud for combinations of brand and gender for which there are no past cases with similar temporal changes in the probability distribution.

[0249] (Traceability System 14 Processing) Next, we will explain the processing performed by the traceability system 14. First, we will explain the scenario of this embodiment.

[0250] Slaughterer A processes three types of brands (P pigs, Q pigs, and R pigs), and each brand has two types of sex (male and female). The traceability system 14 stores the handling history. The handling history includes weight data before and after processing. The traceability system 14 detects padding by monitoring the handling history.

[0251] In the learning phase, the traceability system 14 learns the probability distribution of weight data before and after processing for each combination (six types) of brand and gender. For example, two past cases, seasonal fluctuations and processing equipment updates, are registered.

[0252] In the fraud detection phase, the traceability system 14 detects P male pigs and R male pigs, whose weight loss has increased due to the update of processing equipment, as a combination of brand and sex that may be inflated based on fluctuations in probability distribution. However, because the fluctuations in these two probability distributions are similar to those caused by the update of processing equipment, the traceability system 14 detects that they are not inflated.

[0253] The traceability system 14 detects the P male pig whose weight loss has decreased due to padding and the R male pig whose weight loss has increased as a combination of brand and sex that may be padded based on the fluctuations in probability distribution. Since there are no similar past cases of the fluctuations in these two probability distributions, the traceability system 14 detects them as padding.

[0254] (Fraud detection phase) The processing performed by the traceability system 14 in the fraud detection phase will be described with reference to Fig. 82. Fig. 82 is a flowchart showing the processing performed by the traceability system 14 in the fraud detection phase.

[0255] Steps S411 to S413 shown in Fig. 82 are steps for detecting padding candidates based on fluctuations in probability distribution, and steps S414 to S418 are steps for detecting padding based on the similarity with past cases.

[0256] The distribution fluctuation detection unit 51 divides the new transaction history into combinations of brand and gender (S411). The distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand and gender from the transaction history and probability distribution parameters for each combination of brand and gender (S412). The distribution fluctuation detection unit 51 determines whether there is a combination of brand and gender for which the likelihood L is less than a threshold (S413).

[0257] If it is determined that there is a combination of brand and gender for which the likelihood L is less than the threshold (Yes in S413), the distribution fluctuation detection unit 51 identifies the combination as a padding candidate and proceeds to step S414. If it is determined that there is no combination of brand and gender for which the likelihood L is less than the threshold (No in S413), the distribution fluctuation detection unit 51 ends the process.

[0258] The fluctuation pattern matching detection unit 56 divides the transaction history of the combination of stock and gender that is a candidate for padding by sliding window processing (S414). The fluctuation pattern matching detection unit 56 uses the transaction history for each combination of stock and gender and for each window to learn the parameters of the probability distribution (S415).

[0259] The fluctuation pattern matching detection unit 56 calculates the similarity between the transition of the probability distribution of the padding candidate and the transition of the probability distribution of past cases (S416). The fluctuation pattern matching detection unit 56 determines whether there is a case showing a similarity exceeding the threshold (S417). If it is determined that there is a case showing a similarity exceeding the threshold (Yes in S417), the fluctuation pattern matching detection unit 56 ends the processing. If it is determined that there is no case showing a similarity exceeding the threshold (No in S417), the fluctuation pattern matching detection unit 56 detects the stock and gender for which there is no similar past case as padding (S418).

[0260] Figures 83 to 92 are diagrams for explaining in detail each step shown in Figure 82. Of these, Figures 83 to 87 show examples in which the traceability system 14 detects seasonal fluctuations, and Figures 88 to 92 show examples in which the traceability system 14 detects padding.

[0261] (Detects seasonal fluctuations) First, an example of detecting seasonal variations will be described with reference to Figures 83 to 87. Steps S411 to S415 are generally similar to Figures 16 to 19, 37, and 38, and therefore will not be shown.

[0262] Here, we assume that 10 days have passed since the learning phase, and new handling history for 10 days (11 / 11 to 11 / 20) has been recorded. The new handling history is the same as in Figure 16. Slaughterhouse A has updated its processing equipment, and the weight loss of P male pigs and R male pigs has decreased. We will use this example below for explanation.

[0263] In step S411, the distribution fluctuation detection unit 51 performs processing similar to that in Fig. 17. The distribution fluctuation detection unit 51 divides the new handling history of Slaughterer A into combinations of brand and sex. The distribution fluctuation detection unit 51 extracts records for each combination of brand and sex.

[0264] In step S412, the distribution fluctuation detection unit 51 performs processing in the same manner as in Figure 18. The distribution fluctuation detection unit 51 calculates the likelihood L for each combination of brand and gender using the mean vector μ→ and covariance matrix Σ learned from the handling history for each combination of brand and gender. In this way, the distribution fluctuation detection unit 51 measures the degree to which the probability distribution learned in the learning phase matches the new handling history. Note that the "parameters of the probability distribution of P male pigs" shown in Figure 18 should be read as "parameters of the probability distribution of P male pigs and the latest window."

[0265] In step S413, the distribution fluctuation detection unit 51 performs processing in the same manner as in FIG. 19. The distribution fluctuation detection unit 51 searches for a combination of stock and gender for which the likelihood is less than the threshold. If there is a combination of stock and gender for which the likelihood is less than the threshold, the distribution fluctuation detection unit 51 determines that the combination is a candidate for padding. Note that the "combination of stock and gender that is padded" shown in FIG. 19 should be read as a "combination of stock and gender that is a candidate for padding."

[0266] In step S414, the fluctuation pattern matching detection unit 56 performs processing similar to that in Fig. 37. The fluctuation pattern matching detection unit 56 divides each of the handling histories of the padding candidates using sliding window processing. If the window width is 7 days and the sliding width is 1 day, the handling history for 10 days is divided into four windows t1 to t4.

[0267] In step S415, the fluctuation pattern matching detection unit 56 performs processing similar to that shown in Fig. 38. The fluctuation pattern matching detection unit 56 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood for the handling history. The qualitative data correlation fluctuation detection unit 52 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0268] 83 to 86 are explanatory diagrams of step S416. FIG. 83 shows an example of a past case that serves as a premise for the explanation of this process. As shown in FIG. 83, it is assumed here that two past cases of fluctuations in probability distribution due to reasons other than padding have been recorded: processing equipment updates and seasonal fluctuations. The handling history storage unit 2 stores the mean vector μ→ and covariance matrix Σ for each window in the past cases.

[0269] As shown in FIGS. 84 and 85, the fluctuation pattern matching detection unit 56 generates a vector by combining the mean vector μ→ and the one-dimensional covariance matrix Σ as a preprocessing step for determining the similarity.

[0270] As shown in FIG. 86, the fluctuation pattern matching detection unit 56 calculates the similarity D for each dimension from the vector for each combination of brand and gender and for each window. i The variation pattern matching detection unit 56 calculates the similarity D i By averaging these, we obtain the similarity D with past cases.

[0271] The variation pattern matching detection unit 56 can obtain the similarity D using the following equations (7) and (8).

[0272]

number

number

[0273] In the vector of P male pigs for each window and the vector of each window for updating the processing equipment, the weight after processing both increases monotonically. Therefore, the similarity between the two is high.

[0274] 87 is an explanatory diagram of step S417. The fluctuation pattern matching detection unit 56 searches for past cases in which the similarity between the combination of stock and gender of the padding candidate and the past cases is equal to or greater than a threshold. The fluctuation pattern matching detection unit 56 determines that a probability distribution in which the similarity is equal to or greater than a threshold is similar to the past cases. If there is a similar past case, the fluctuation pattern matching detection unit 56 detects it as normal, and otherwise detects it as padding.

[0275] In this example, it is assumed that the similarity between the updates of the processing equipment for the P male pigs and the R male pigs is equal to or greater than the threshold value, and therefore the fluctuation pattern matching detection unit 56 detects them as normal.

[0276] (Inflation detected) Next, an example of detecting padding will be described with reference to Figures 88 to 92. Details that overlap with the example of detecting the update of processing equipment described above will be omitted as appropriate.

[0277] Figure 88 is an explanatory diagram of step S411. Figure 88 shows an example of a new handling history that is the premise for explaining this process. Here, it is assumed that 10 days have passed since the learning phase, and new handling history for 10 days (11 / 11 to 11 / 20) has been recorded. It is also assumed here that the weight of P male pigs has been inflated with that of R male pigs. The average weight of P male pigs after processing is 71 kg, and this weight is inflated to 72 kg and registered. The average weight of R male pigs after processing is 81 kg, and this weight is reduced to 78 kg and registered. This example will be used hereafter.

[0278] In step S415, the fluctuation pattern matching detection unit 56 performs processing similar to that shown in Fig. 38. The fluctuation pattern matching detection unit 56 learns the parameters (mean vector μ→ and covariance matrix Σ) of the bivariate normal distribution of the weight data before and after food processing based on the likelihood for the handling history. The qualitative data correlation fluctuation detection unit 52 finds the mean vector μ→ and covariance matrix Σ that maximize the likelihood L.

[0279] 89 to 91 are explanatory diagrams of step S416. An example of a past case that is the premise for explaining this process is the same as that in FIG. 83, and is therefore not shown in the figures. As shown in FIG. 83, it is assumed here that two past cases of fluctuations in probability distribution due to reasons other than padding have been recorded: processing equipment updates and seasonal fluctuations. The handling history storage unit 2 stores the mean vector μ→ and covariance matrix Σ for each window in the past cases.

[0280] As shown in FIGS. 89 and 90, the fluctuation pattern matching detection unit 56 generates a vector by combining the mean vector μ→ and the one-dimensional covariance matrix Σ as a preprocessing step for determining the similarity.

[0281] As shown in FIG. 91, the fluctuation pattern matching detection unit 56 calculates the similarity D for each dimension from the vector for each combination of brand and gender and for each window. i The variation pattern matching detection unit 56 calculates the similarity D i By averaging these, we obtain the similarity D with past cases.

[0282] In the vector of P male pigs, the weight after processing alternates between increasing and decreasing. On the other hand, in the vector for each window of processing equipment updates, the weight after processing increases monotonically. Therefore, the similarity between the two is low.

[0283] 92 is an explanatory diagram of step S417. The fluctuation pattern matching detection unit 56 searches for past cases in which the similarity between the combination of stock and gender of the padding candidate and the past cases is equal to or greater than a threshold. The fluctuation pattern matching detection unit 56 determines that a probability distribution in which the similarity is equal to or greater than a threshold is similar to the past cases. If there is a similar past case, the fluctuation pattern matching detection unit 56 detects it as normal, and otherwise detects it as padding.

[0284] In this example, the similarity between the updates of the processing equipment for the P male pigs and the R male pigs is less than the threshold value. Therefore, the fluctuation pattern matching detection unit 56 detects the combination of the P male pigs and the R male pigs as being inflated.

[0285] As described above, the traceability system 14 according to the present disclosure identifies a fluctuation pattern of the probability distribution caused by factors other than padding based on multiple past probability distributions, and compares the identified fluctuation pattern with the fluctuation pattern of the current probability distribution. The traceability system 14 detects padding based on the comparison result.

[0286] With this configuration, the traceability system 14 can detect food padding based on the fluctuation pattern of the probability distribution in past cases, thereby reducing false positives.

[0287] The above has described the configurations of fraud detection system 100 and traceability systems 11 to 14 and the processes performed by each system. Note that the configurations of fraud detection system 100 and traceability systems 11 to 14 described above are merely examples and can be changed as appropriate.

[0288] For example, when some or all of the components of the fraud detection system 100 and the traceability systems 11-14 are realized by multiple information processing devices, circuits, etc., the multiple information processing devices, circuits, etc. may be centrally or decentralized. For example, the information processing devices, circuits, etc. may be realized as a client-server system, cloud computing system, etc., in a form in which each is connected via a communication network. Furthermore, the functions of the fraud detection system 100 and the traceability systems 11-14 may be provided in a SaaS (Software as a Service) format.

[0289] For example, in the first embodiment, the traceability system 11 has been described as having a configuration including a handling history creation unit 1, a handling history storage unit 2, a learning unit 3, a handling history tracking unit 4, and a fraud detection unit 5, but this is not limited to this. For example, the fraud detection unit 5 may be configured independently as a fraud detection system. For example, some or all of the handling history creation unit 1, the handling history storage unit 2, the learning unit 3, and the handling history tracking unit 4 may be provided in a device other than the traceability system 11. The same applies to the traceability systems 12 to 14.

[0290] <Hardware configuration example> Each functional component of fraud detection system 100 and traceability systems 11-14 (hereinafter referred to as "fraud detection system 100, etc.") may be realized by hardware (e.g., hardwired electronic circuits, etc.) that realizes each functional component, or may be realized by a combination of hardware and software (e.g., a combination of an electronic circuit and a program that controls it). Below, we will further explain the case where each functional component of fraud detection system 100, etc. is realized by a combination of hardware and software.

[0291] 93 is a block diagram illustrating an example of the hardware configuration of a computer 900 that realizes the fraud detection system 100, etc. The computer 900 may be a dedicated computer designed to realize the fraud detection system 100, etc., or may be a general-purpose computer. The computer 900 may also be a portable computer such as a smartphone or tablet terminal.

[0292] For example, by installing a predetermined application on computer 900, the functions of fraud detection system 100 and the like are realized on computer 900. The application is configured as a program for realizing the functional components of fraud detection system 100 and the like.

[0293] The computer 900 includes a bus 902, a processor 904, a memory 906, a storage device 908, an input / output interface 910, and a network interface 912. The bus 902 is a data transmission path that allows the processor 904, the memory 906, the storage device 908, the input / output interface 910, and the network interface 912 to transmit and receive data to and from each other. However, the method of connecting the processor 904 and other components to each other is not limited to a bus connection.

[0294] The processor 904 is a variety of processors such as a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), or a quantum processor (quantum computer control chip). The memory 906 is a main storage device realized using a random access memory (RAM) or the like. The storage device 908 is an auxiliary storage device realized using a hard disk, a solid state drive (SSD), a memory card, a read only memory (ROM), or the like.

[0295] The input / output interface 910 is an interface for connecting the computer 900 to an input / output device. For example, the input / output interface 910 is connected to an input device such as a keyboard and an output device such as a display device.

[0296] The network interface 912 is an interface for connecting the computer 900 to a network. This network may be a LAN (Local Area Network) or a WAN (Wide Area Network).

[0297] The storage device 908 stores programs (programs that realize the above-mentioned applications) that realize the various functional components of the fraud detection system 100, etc. The processor 904 reads these programs into the memory 906 and executes them, thereby realizing the various functional components of the fraud detection system 100, etc.

[0298] Each processor executes one or more programs containing instructions for causing a computer to perform an algorithm. The programs contain instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The programs may be stored on various types of non-transitory computer-readable media or tangible storage media. By way of example and not limitation, non-transitory computer-readable media or tangible storage media include RAM, ROM, flash memory, SSD or other memory technologies, CD-ROM, DVD (Digital Versatile Disc), Blu-ray® disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices. The programs may also be transmitted over various types of transitory computer-readable media or communication media. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.

[0299] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that would be understood by a person skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be appropriately combined with other embodiments. For example, combinations such as embodiment 1 and embodiment 2, embodiment 1 and embodiment 3, embodiment 1 and embodiment 4, embodiment 1 and embodiment 2 and embodiment 3, embodiment 1 and embodiment 2 and embodiment 4, etc. may be realized.

[0300] Each drawing is merely an example for describing one or more embodiments. Each drawing may relate not only to one particular embodiment, but also to one or more other embodiments. As will be understood by those skilled in the art, various features or steps described with reference to any one drawing can be combined with features or steps shown in one or more other drawings to create, for example, an embodiment not explicitly shown or described. Not all features or steps shown in any one drawing are necessary to describe an exemplary embodiment, and some features or steps may be omitted. The order of steps described in any drawing may be changed as appropriate.

[0301] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes. (Appendix 1) a handling history acquisition unit that acquires a plurality of handling histories including qualitative data of the product that is the subject of fraud detection and the quantity of the product before and after processing; a distribution fluctuation detection unit that compares, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detects fraudulent padding of the product based on the comparison result. Fraud detection system. (Appendix 2) The handling history further includes data of a business operator that processes the product, The distribution change detection unit compares the current probability distribution with the past probability distribution for each combination of the qualitative data and the business operator, and detects the padding at the business operator based on the comparison result. 1. A fraud detection system as described in Appendix 1. (Appendix 3) The distribution change detection unit calculates a likelihood indicating the degree to which the past probability distribution matches the current probability distribution, and detects the padding based on the calculation result. 3. The fraud detection system of claim 1 or 2. (Appendix 4) Based on a comparison result between the current probability distribution and the past probability distribution, the qualitative data corresponding to the current probability distribution is identified as a padding candidate that may have been padded; Detecting the padding from the padding candidates based on a time-dependent change in a probability distribution corresponding to the padding candidates. 4. A fraud detection system according to any one of appendices 1 to 3. (Appendix 5) a qualitative data correlation analysis unit that analyzes correlations between the past probability distributions of different qualitative data and identifies combinations of the qualitative data corresponding to uncorrelated past probability distributions; and a qualitative data correlation change detection unit that identifies a combination of the qualitative data in which the current probability distribution has changed to a negative correlation among the identified combinations of uncorrelated qualitative data, and detects the padding based on the identification result. 10. The fraud detection system described in Appendix 4. (Appendix 6) a per-enterprise correlation analysis unit that analyzes the correlation between the past probability distributions of the qualitative data for each enterprise that processes the product; and a per-carrier correlation fluctuation detection unit that detects the padding based on the analysis result. 6. The fraud detection system of claim 4 or 5. (Appendix 7) the per-business correlation analysis unit identifies a combination of businesses corresponding to the past probability distributions having a correlation stronger than a predetermined value, The per-enterprise correlation change detection unit identifies a combination of the entities for which the current probability distribution has changed to a weak correlation less than a predetermined value among the combinations of the entities for which a strong correlation has been identified, and detects the padding based on the identification result. 6. A fraud detection system as described in Appendix 6. (Appendix 8) the per-enterprise correlation analysis unit identifies a combination of enterprises corresponding to the uncorrelated past probability distribution, The per-enterprise correlation change detection unit identifies a combination of the entities for which the current probability distribution has changed to correlated among the identified combinations of the entities that are uncorrelated, and detects the padding based on the identification result. 8. The fraud detection system of claim 6 or 7. (Appendix 9) The system further includes a fluctuation pattern matching detection unit that identifies a fluctuation pattern of the probability distribution caused by a factor other than the padding based on a plurality of the past probability distributions, compares the identified fluctuation pattern with a fluctuation pattern of the current probability distribution, and detects the padding based on the comparison result. A fraud detection system according to any one of appendices 4 to 8. (Appendix 10) a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detecting fraudulent padding of the product based on the comparison result. Fraud detection methods. (Appendix 11) a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detecting fraudulent padding of the product based on the comparison result. program.

[0302] Some or all of the elements (e.g., configurations and functions) described in Supplements 2 to 8 that are dependent on Supplement 1 may also be dependent on Supplements 9 and 10 in the same dependency relationship as Supplements 2 to 8. Some or all of the elements described in any Supplement may be applied to various hardware, software, recording means for recording software, systems, and methods. [Explanation of symbols]

[0303] 1. Transaction History Creation Department 2. Transaction history memory section 3. Learning Department 4. Transaction History Tracking Section 5. Fraud detection unit 11~14 Traceability System 31 Distribution Creation Department 32 Distribution storage section 33 Qualitative Data Distribution Analysis Section 34 Moving window distribution creation section 35 Moving window distribution memory 36 Qualitative Data Correlation Analysis Section 37 Business Correlation Analysis Department 51 Distribution change detection unit 52 Qualitative data correlation fluctuation detection unit 53 Qualitative data uncorrelated combination memory section 54 Correlation fluctuation detection unit for each operator 55 Business correlation combination memory unit 56 Fluctuation pattern matching detection unit 57 Past case memory section 100 Fraud Detection System 101 Transaction History Acquisition Department 102 Distribution change detection unit 900 Computers 902 Bus 904 processor 906 memory 908 Storage Devices 910 Input / Output Interface 912 Network Interface A~C Slaughterman D, D i Similarity L likelihood r, r i Correlation coefficient t1~t4 window μ→ mean vector Σ covariance matrix

Claims

1. a handling history acquisition unit that acquires a plurality of handling histories including qualitative data of the product that is the subject of fraud detection and the quantity of the product before and after processing; a distribution fluctuation detection unit that compares, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detects fraudulent padding of the product based on the comparison result. Fraud detection system.

2. The handling history further includes data of a business operator that processes the product, The distribution change detection unit compares the current probability distribution with the past probability distribution for each combination of the qualitative data and the business operator, and detects the padding at the business operator based on the comparison result. The fraud detection system of claim 1 .

3. The distribution change detection unit calculates a likelihood indicating the degree to which the past probability distribution matches the current probability distribution, and detects the padding based on the calculation result. The fraud detection system according to claim 1 or 2.

4. Based on a comparison result between the current probability distribution and the past probability distribution, the qualitative data corresponding to the current probability distribution is identified as a padding candidate that may have been padded; Detecting the padding from the padding candidates based on a time-dependent change in a probability distribution corresponding to the padding candidates. The fraud detection system according to claim 1 or 2.

5. a qualitative data correlation analysis unit that analyzes correlations between the past probability distributions of different qualitative data and identifies combinations of the qualitative data corresponding to uncorrelated past probability distributions; and a qualitative data correlation change detection unit that identifies a combination of the qualitative data in which the current probability distribution has changed to a negative correlation among the identified combinations of uncorrelated qualitative data, and detects the padding based on the identification result. The fraud detection system of claim 4 .

6. a per-enterprise correlation analysis unit that analyzes the correlation between the past probability distributions of the qualitative data for each enterprise that processes the product; and a per-carrier correlation fluctuation detection unit that detects the padding based on the analysis result. The fraud detection system of claim 4 .

7. the per-business correlation analysis unit identifies a combination of businesses corresponding to the past probability distributions having a correlation stronger than a predetermined value, The per-enterprise correlation change detection unit identifies a combination of the entities for which the current probability distribution has changed to a weak correlation less than a predetermined value among the combinations of the entities for which a strong correlation has been identified, and detects the padding based on the identification result. The fraud detection system of claim 6.

8. The system further includes a fluctuation pattern matching detection unit that identifies a fluctuation pattern of the probability distribution caused by a factor other than the padding based on a plurality of the past probability distributions, compares the identified fluctuation pattern with a fluctuation pattern of the current probability distribution, and detects the padding based on the comparison result. The fraud detection system of claim 4 .

9. a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detecting fraudulent padding of the product based on the comparison result. Fraud detection methods.

10. a handling history acquisition step of acquiring a plurality of handling histories including qualitative data of the product that is the fraud detection target and the amount of the product before and after processing; a distribution fluctuation detection step of comparing, for each of the qualitative data, a current probability distribution indicating a probability distribution of the quantity of the product before and after processing in the acquired handling history with a past probability distribution indicating a probability distribution of the quantity of the product before and after processing in a past handling history, and detecting fraudulent padding of the product based on the comparison result. program.

Images