JP6881203B2 - Classification program, classification method, and classification device - Google Patents

Description

The present invention relates to a classification program, a classification method, and a classification device.

Conventionally, a technique of generating a decision tree model including a node representing a condition for classifying target data and a leaf node representing an attribute for classifying target data based on training data, and classifying the target data by the generated decision tree model. There is. For example, the decision tree model is used to classify social security benefit billing data into normal billing data or anomalous billing data for fraudulent billing.

International Publication No. 2016/189606 Japanese Unexamined Patent Publication No. 2017-62713 Special Table 2001-516107

However, in the prior art, it is difficult for the user to grasp the reason why the target data is classified as normal or abnormal. For example, it is conceivable to enumerate the judgment results of the conditions judged in the process of classifying the target data and notify the user, but as the number of the judged conditions increases, the user normalizes the target data. Or, it becomes difficult to intuitively understand the reason for the abnormal classification.

In one aspect, it is an object of the present invention to provide a classification program, a classification method, and a classification device that generate information for a user to understand why the target data is classified as normal or abnormal. ..

According to one embodiment, a decision tree model that includes a node that represents a condition for classifying the target data and a leaf node that represents that the target data is normal or that the target data is abnormal. When the target data is classified and the target data is classified into the first leaf node whose name has not been set, the name has already been set in the decision tree model based on the positional relationship between the nodes. A classification program, a classification method, and a classification device that select a second leaf node and generate a name for the first leaf node based on the name of the selected second leaf node are proposed.

According to one aspect, the user can generate information for grasping the reason why the target data is classified as normal or abnormal.

FIG. 1 is an explanatory diagram showing an embodiment of a classification method according to an embodiment. FIG. 2 is an explanatory diagram showing an example of the classification system 200. FIG. 3 is a block diagram showing a hardware configuration example of the classification device 100. FIG. 4 is an explanatory diagram showing an example of the stored contents of the data structure of the billing data 400. FIG. 5 is an explanatory diagram showing an example of the stored contents of the billing table 500. FIG. 6 is a block diagram showing a hardware configuration example of the terminal device 201. FIG. 7 is a block diagram showing a functional configuration example of the classification device 100. FIG. 8 is an explanatory diagram showing a specific example of a situation in which the classification device 100 is used. FIG. 9 is an explanatory diagram showing a flow for generating the decision tree model 805. FIG. 10 is an explanatory diagram (No. 1) showing a flow of setting a name for a leaf node. FIG. 11 is an explanatory diagram (No. 2) showing a flow of setting a name for a leaf node. FIG. 12 is an explanatory diagram (No. 1) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 13 is an explanatory diagram (No. 2) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 14 is an explanatory diagram (No. 3) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 15 is an explanatory diagram (No. 4) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 16 is an explanatory diagram (No. 5) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 17 is an explanatory diagram (No. 6) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 18 is an explanatory diagram (No. 7) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 19 is an explanatory diagram (No. 8) showing an operation example 1 in which the classification device 100 sets a name for the leaf node. FIG. 20 is an explanatory diagram (No. 1) showing an operation example 2 in which the classification device 100 sets a name for the leaf node. FIG. 21 is an explanatory diagram (No. 2) showing an operation example 2 in which the classification device 100 sets a name for the leaf node. FIG. 22 is an explanatory diagram (No. 3) showing an operation example 2 in which the classification device 100 sets a name for the leaf node. FIG. 23 is an explanatory diagram showing an output example 1 in the terminal device 201. FIG. 24 is an explanatory diagram showing an output example 2 in the terminal device 201. FIG. 25 is a flowchart showing an example of the overall processing procedure. FIG. 26 is a flowchart showing an example of the selection processing procedure. FIG. 27 is a flowchart showing an example of the search processing procedure.

Hereinafter, embodiments of a classification program, a classification method, and a classification device according to the present invention will be described in detail with reference to the drawings.

(An example of a classification method according to an embodiment)
FIG. 1 is an explanatory diagram showing an embodiment of a classification method according to an embodiment. The classification device 100 is a computer that classifies target data based on a decision tree model.

The decision tree model is a model including a node representing a condition for classifying the target data and a leaf node representing an attribute for classifying the target data. The leaf node represents, for example, that the target data is normal or that the target data is abnormal.

In the following drawings, the child node on the left side of the node representing the condition displays the child node corresponding to the case where the condition judgment result is True, and the child node on the right side corresponds to the case where the condition judgment result is False. The child node may be displayed and the display of True and False may be omitted.

Here, ministries and local governments may accept billing data corresponding to social security benefit bills and check whether there is abnormal billing data. The larger the billing data, the more ministries and local governments The burden on the staff will increase. In addition, the larger the billing data, the easier it is for staff such as ministries and local governments to overlook abnormal billing data. The abnormal billing data is, for example, fraudulent billing billing data. The anomalous billing data is specifically billing data for inflated billing. The abnormal billing data is specifically billing data for fictitious billing.

Therefore, it is desired to automatically classify the billing data into normal billing data or abnormal billing data as target data. For example, it is desirable to use a decision tree model to classify billing data into normal billing data or abnormal billing data. The decision tree model is generated, for example, based on historical billing data. A specific situation in which the decision tree model is used will be described later with reference to, for example, FIG.

In addition, hospitals and pharmacies sometimes manage medical data corresponding to medical records, prescriptions, medical fee statements, etc., and check whether there is abnormal medical data, resulting in a huge amount of medical data. Indeed, the burden on staff such as hospitals and pharmacies will increase. In addition, staff at hospitals and pharmacies are more likely to overlook abnormal medical data. Abnormal medical data is, specifically, medical data that includes information on a specific disease but does not include information on drugs or tests that are essential for a specific disease. The abnormal medical data is, specifically, medical data in which a request can be made only once a month and the request is made twice a month.

Therefore, it is desired to automatically classify medical data into normal medical data or abnormal medical data as target data. For example, it is desirable to use a decision tree model to classify medical data into normal or abnormal medical data. The decision tree model is generated, for example, based on historical medical data.

However, it is difficult for a user who uses the decision tree model to grasp the detailed type of the result of classifying the target data as normal or abnormal, and it is difficult to grasp the reason why the target data is classified as normal or abnormal. .. Specifically, when the target data is billing data and the target data is classified as fraudulent billing, it is difficult for the user to grasp the type of fraudulent billing in which the target data is classified, and the target data is classified. It is difficult to understand the reason.

On the other hand, it is conceivable that the judgment results of the conditions judged in the process of classifying the target data are listed and notified to the user. However, in this case, as the number of determined conditions increases, it becomes more difficult for the user to intuitively grasp the reason why the target data is classified as normal or abnormal.

Also, when the user is made to set the name for all the leaf nodes of the decision tree model, and when the target data is classified into the leaf nodes, the user is notified of the name set for the leaf node in association with the target data. Can be considered. However, in this case, it is difficult to let the user set the name for all the leaf nodes of the decision tree model. For example, as the number of leaf nodes increases, the burden on the user increases.

Also, in situations where unsupervised learning is used or combined with supervised learning and unsupervised learning to generate a decision tree model, the user is a leaf node added to the decision tree model by unsupervised learning. In addition, it is difficult to grasp in advance what kind of target data is classified. Therefore, it is difficult for the user to set a name for the leaf node added by unsupervised learning.

Therefore, in the present embodiment, when the target data is classified into the first leaf node whose name has not been set on the decision tree model, the first leaf is based on the second leaf node whose name has been set. A classification method that can generate node names will be described. According to this classification method, the name of the leaf node that classified the target data can be presented to the user, and the user can easily understand the reason why the target data is classified as normal or abnormal. ..

In FIG. 1, the classification device 100 stores the decision tree model 110. The classification device 100 generates and stores the decision tree model 110 based on the learning data, for example. The classification device 100 may receive and store the decision tree model 110 from another device that generates the decision tree model 110 based on the learning data, for example.

The decision tree model 110 includes a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is submitted, for example, by some business operator. The target data is specifically the billing data corresponding to the invoice for social security benefits. In the example of FIG. 1, the decision tree model 110 has, for example, nodes 111 to 116 representing conditions for classifying the target data, and leaf nodes 121 indicating that the target data is normal or the target data is abnormal. Includes ~ 127.

The classification device 100 classifies the target data according to the decision tree model 110. The classification device 100 determines, for example, whether or not the target data satisfies the conditions represented by the nodes in order from the root node of the decision tree model 110, and classifies the target data into any leaf node based on the determination result. To do. The classification into leaf nodes corresponds to the classification into the attributes represented by the leaf node, for example, the classification into normal or abnormal represented by the leaf node. In the example of FIG. 1, the classification device 100 specifically classifies the target data into leaf nodes 127.

When the target data is classified into the first leaf node whose name has not been set, the classification device 100 has the name set in the second decision tree model 110 based on the positional relationship between the nodes. Select a leaf node. The positional relationship between nodes is represented by, for example, the parent-child relationship between nodes. The classification device 100 selects the second leaf node from the leaf nodes in the vicinity of the first leaf node, for example, based on the positional relationship between the nodes. The distance between nodes is represented, for example, by the number of edges connecting the nodes. The neighborhood corresponds, for example, that the number of edges connecting the nodes is less than or equal to a predetermined number.

In the example of FIG. 1, the classification device 100 specifically classifies the target data into the leaf nodes 127 whose names have not been set, so that the leaves are among the leaf nodes 123 and 126 in the vicinity of the leaf nodes 127. Select node 123. A specific example of selecting the second leaf node from the leaf nodes in the vicinity of the first leaf node will be described later with reference to FIGS. 17 and 18.

The classification device 100 generates the name of the first leaf node based on the name of the selected second leaf node. The name of the second leaf node corresponds to the detailed type of the result of classifying the target data, and is a name indicating the reason why the target data is classified. The name of the second leaf node is, for example, XX service or B fraud. Specifically, the name of the second leaf node is inflated billing or fictitious billing.

In the example of FIG. 1, the classification device 100 generates, for example, the name “Tendency close to B fraud” of the leaf node 127 based on the name “B fraud” of the selected leaf node 123. Further, for example, when the leaf node 126 is selected, the classification device 100 is based on the name "XX service" of the leaf node 126, and the name of the leaf node 127 "is considered to be fraudulent billing related to the XX service". To generate.

The classification device 100 further determines a determination result of a condition determined when the target data is classified into the selected second leaf node, and a condition determined when the target data is classified into the first leaf node. The difference from the determination result of may be reflected in the name of the first leaf node. The difference is, for example, the judgment result of the condition represented by each node on the route from the root node to the first leaf node and the judgment result of the condition represented by each node on the route from the root node to the second leaf node. Is the difference.

The classification device 100 uses, for example, the name "XX service" of the selected leaf node 126 and the name "XX service" of the leaf node 127 based on the difference in the determination conditions "(compared to the XX service) D> 100". Related to the service, D> not 100 "is generated.

As a result, the classification device 100 can output the name of the generated first leaf node in association with the target data. Therefore, the classification device 100 can enable the user to grasp the detailed type of the result of classifying the target data and to grasp the reason why the target data is classified.

As a result, even if the target data is enormous, the user can efficiently check whether or not there is abnormal target data by referring to the reason why each target data is classified, and efficiently. It is possible to carry out business and reduce the burden. In addition, the user is less likely to overlook abnormal target data and can perform business efficiently.

In addition, the user can grasp which target data is classified as a relatively serious abnormality and determine which target data is preferable to be scrutinized. In addition, the user can determine which of the multiple businesses that submitted the target data submitted the target data with a relatively serious abnormality, and it is preferable to audit the business. Can be identified.

Further, the classification device 100 can generate the name of the first leaf node based on the name of the second leaf node, and can easily generate a name that is intuitively easy for the user to understand. .. For this reason, the user can easily understand the reason why the target data is classified, and can perform the work efficiently, as compared with the case of referring to the information listing the judgment results of the conditions judged in the process of classifying the target data. It can be carried out.

In addition, the classification device 100 can generate a name for a leaf node for which the user has not set a name. Therefore, the user does not have to set the names for all the leaf nodes of the decision tree model 110, and the burden can be reduced. The classification device 100 can also set the name of the leaf node added to the decision tree model 110 by unsupervised learning. Therefore, the user can understand the reason why the target data is classified into the leaf nodes added by unsupervised learning.

Here, the case where the classification device 100 generates the name of the first leaf node even if the first leaf node in which the target data is classified is a leaf node indicating that the target data is normal will be described. However, it is not limited to this. For example, if the classification device 100 is a leaf node indicating that the target data is normal, the first leaf node into which the target data is classified may not generate the name of the first leaf node. Good. As a result, the classification device 100 can reduce the processing amount. For example, when the user requests to find abnormal target data, the classification device 100 does not reduce the business efficiency of the user even if the leaf node indicating that the target data is normal is not named. In addition, the amount of processing can be reduced.

Here, the case where the classification device 100 selects the second leaf node from the leaf nodes in the vicinity of the first leaf node based on the positional relationship between the nodes has been described, but the present invention is not limited to this. .. For example, the classification device 100 includes leaf nodes in a second subtree that is the same as or similar to the first subtree that includes the first leaf node in the decision tree model 110, based on the positional relationship between the nodes. The second leaf node may be selected from the above. A specific example of selecting the second leaf node from the leaf nodes included in the second subtree will be described later with reference to FIGS. 19 to 21.

(Example of classification system 200)
Next, an example of the classification system 200 to which the classification device 100 shown in FIG. 1 is applied will be described with reference to FIG.

FIG. 2 is an explanatory diagram showing an example of the classification system 200. In FIG. 2, the classification system 200 includes a classification device 100 and a terminal device 201.

In the classification system 200, the classification device 100 and the terminal device 201 are connected via a wired or wireless network 210. The network 210 is, for example, a LAN (Local Area Network), a WAN (Wide Area Network), the Internet, or the like.

The classification device 100 receives billing data as learning data from the terminal device 201 and generates a decision tree model. The classification device 100 receives the billing data to be the target data, and classifies the billing data into the leaf nodes of the generated decision tree model. The classification device 100 generates a name if the name is not set for the leaf node that has classified the billing data.

The classification device 100 causes the terminal device 201 to output information based on the result of classifying the billing data, and notifies the user of the terminal device 201. The classification device 100, for example, associates the billing data with the name of the leaf node that classifies the billing data, and displays the billing data on the terminal device 201. The classification device 100 is, for example, a server, a PC (Personal Computer), or the like.

The terminal device 201 is a computer used by the user of the classification system 200. The user is, for example, an employee of a ministry or local government, or a hospital or pharmacy. The terminal device 201 is provided in, for example, a ministry, a local government, a hospital, a pharmacy, or the like. The terminal device 201 transmits billing data, which is learning data, to the classification device 100. The terminal device 201 transmits billing data as target data to the classification device 100, and receives and outputs information based on the result of classifying the billing data from the classification device 100. The terminal device 201 is, for example, a PC, a tablet terminal, a smartphone, or the like.

Here, the case where the classification device 100 and the terminal device 201 are different devices has been described, but the present invention is not limited to this. For example, the classification device 100 may be integrated with the terminal device 201. In this case, the classification device 100 receives billing data as learning data, billing data as target data, or the like based on the operation input of the user.

Here, the case where the classification device 100 generates a decision tree model and classifies the target data has been described, but the present invention is not limited to this. For example, another device different from the classification device 100 may generate a decision tree model and send it to the classification device 100. Further, for example, another device different from the classification device 100 may transmit the result of classifying the target data to the classification device 100. Another device is, for example, a terminal device 201.

(Example of hardware configuration of classification device 100)
Next, a hardware configuration example of the classification device 100 will be described with reference to FIG.

FIG. 3 is a block diagram showing a hardware configuration example of the classification device 100. In FIG. 3, the classification device 100 includes a CPU (Central Processing Unit) 301, a memory 302, a network I / F (Interface) 303, a recording medium I / F 304, and a recording medium 305. Further, each component is connected by a bus 300.

Here, the CPU 301 controls the entire classification device 100. The memory 302 includes, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), a flash ROM, and the like. Specifically, for example, a flash ROM or ROM stores various programs, and the RAM is used as a work area of the CPU 301. The program stored in the memory 302 is loaded into the CPU 301 to cause the CPU 301 to execute the coded process.

The network I / F 303 is connected to the network 210 through a communication line, and is connected to another computer via the network 210. Then, the network I / F 303 controls the internal interface with the network 210 and controls the input / output of data from another computer. For the network I / F 303, for example, a modem, a LAN adapter, or the like can be adopted.

The recording medium I / F 304 controls read / write of data to the recording medium 305 according to the control of the CPU 301. The recording medium I / F 304 is, for example, a disk drive, an SSD (Solid State Drive), a USB (Universal Bus) port, or the like. The recording medium 305 is a non-volatile memory that stores data written under the control of the recording medium I / F 304. The recording medium 305 is, for example, a disk, a semiconductor memory, a USB memory, or the like. The recording medium 305 may be detachable from the sorting device 100. The memory 302 or the recording medium 305 may store, for example, billing data or a decision tree model.

The sorting device 100 may include, for example, a keyboard, a mouse, a display, a printer, a scanner, a microphone, a speaker, and the like, in addition to the above-described components. Further, the classification device 100 may have a plurality of recording media I / F 304 and recording media 305. Further, the classification device 100 does not have to have the recording medium I / F 304 or the recording medium 305.

(Data structure of billing data 400)
Next, the data structure of the billing data 400 acquired by the classification device 100 will be described with reference to FIG. The billing data 400 is transmitted from the terminal device 201 to the classification device 100, for example.

FIG. 4 is an explanatory diagram showing an example of the stored contents of the data structure of the billing data 400. As shown in FIG. 4, the billing data 400 has one or more item fields. By setting information in each field of the billing data 400, the value of the item of the invoice for social security benefit expenses is stored.

In the item field, the value of the item on the invoice for social security benefits is set. The items are, for example, item A, item B, and item C. Specifically, the items are a date, a service category, an amount of money, a usage amount, additional information, and the like. Additional information is a condition that allows you to apply by increasing the amount.

(Memory contents of billing table 500)
Next, the stored contents of the billing table 500 that manages the billing data 400 acquired by the classification device 100 will be described with reference to FIG. The billing table 500 is realized, for example, by a storage area such as a memory 302 or a recording medium 305 of the classification device 100 shown in FIG.

FIG. 5 is an explanatory diagram showing an example of the stored contents of the billing table 500. As shown in FIG. 5, the billing table 500 has fields for a key, one or more items, a result, a reason for fraud, and a reason for normality. By setting information in each field, the billing table 500 stores the billing data 400 and the result data regarding the result of classifying the billing data 400 as a record.

A key for identifying the billing data 400 is set in the key field. In the item field, the value set in the item field of the billing data 400 identified by the key corresponding to the value of the item of the social security benefit bill is set.

In the result field, the result of classifying the billing data 400 identified by the key into normal or fraudulent billing is set. When the billing data 400 is classified as fraudulent billing, the fraudulent reason for which the billing data 400 is classified as fraudulent billing is set in the fraudulent reason field. Reasons for fraud include, for example, inflated billing and fictitious billing. When the billing data 400 is normally classified, the normal reason for which the billing data 400 is normally classified is set in the normal reason field. The normal reason is, for example, home care.

(Data structure of training data)
Since the training data includes the past billing data 400, the data structure of the training data is, for example, the same as the data structure of the billing data 400 shown in FIG. 4, but the training data further has a field of results. You may. In the result field, the result that the user determines that the past billing data 400 is normal or fraudulent billing is set.

The learning data is stored using, for example, a learning table. An example of the learning table is shown in FIG. 12, for example. The stored contents of the learning table are, for example, the same as the stored contents of the billing table 500 shown in FIG. 5, but the learning table does not have to have fields for the reason for fraud and the reason for normal.

(Hardware configuration example of terminal device 201)
Next, a hardware configuration example of the terminal device 201 will be described with reference to FIG.

FIG. 6 is a block diagram showing a hardware configuration example of the terminal device 201. In FIG. 5, the terminal device 201 includes a CPU 601, a memory 602, a network I / F 603, a recording medium I / F 604, a recording medium 605, a display 606, and an input device 607. Further, each component is connected by a bus 600.

Here, the CPU 601 controls the entire terminal device 201. The memory 602 includes, for example, a ROM, a RAM, a flash ROM, and the like. Specifically, for example, a flash ROM or ROM stores various programs, and RAM is used as a work area of CPU 601. The program stored in the memory 602 is loaded into the CPU 601 to cause the CPU 601 to execute the coded process.

The network I / F 603 is connected to the network 210 through a communication line, and is connected to another computer via the network 210. Then, the network I / F 603 controls the internal interface with the network 210 and controls the input / output of data from another computer. For the network I / F 603, for example, a modem, a LAN adapter, or the like can be adopted.

The recording medium I / F 604 controls read / write of data to the recording medium 605 according to the control of the CPU 601. The recording medium I / F 604 is, for example, a disk drive, an SSD, a USB port, or the like. The recording medium 605 is a non-volatile memory that stores data written under the control of the recording medium I / F 604. The recording medium 605 is, for example, a disk, a semiconductor memory, a USB memory, or the like. The recording medium 605 may be detachable from the terminal device 201.

The display 606 displays data such as a cursor, an icon, a toolbox, a document, an image, and functional information. As the display 606, for example, a CRT (Cathode Ray Tube), a liquid crystal display, an organic EL (Electroluminescence) display and the like can be adopted.

The input device 607 has keys for inputting characters, numbers, various instructions, and the like, and inputs data. The input device 607 may be a keyboard, a mouse, or the like, or may be a touch panel type input pad, a numeric keypad, or the like.

In addition to the above-described components, the terminal device 201 may include, for example, a printer, a scanner, a microphone, a speaker, and the like. Further, the terminal device 201 may have a plurality of recording media I / F 604 and recording media 605. Further, the terminal device 201 does not have to have the recording medium I / F 604 or the recording medium 605.

(Example of functional configuration of classification device 100)
Next, a functional configuration example of the classification device 100 will be described with reference to FIG. 7.

FIG. 7 is a block diagram showing a functional configuration example of the classification device 100. The classification device 100 includes a storage unit 700, an acquisition unit 701, a learning unit 702, a classification unit 703, a selection unit 704, a generation unit 705, and an output unit 706.

The storage unit 700 is realized by, for example, a storage area such as the memory 302 or the recording medium 305 shown in FIG. Hereinafter, the case where the storage unit 700 is included in the classification device 100 will be described, but the present invention is not limited to this. For example, the storage unit 700 may be included in a device different from the classification device 100, and the stored contents of the storage unit 700 may be referred to by the classification device 100.

The acquisition units 701 to the output unit 706 function as an example of the control unit. Specifically, the acquisition units 701 to the output unit 706 are made by causing the CPU 301 to execute a program stored in a storage area such as the memory 302 or the recording medium 305 shown in FIG. 3, or the network I / F 303. To realize the function. The processing result of each functional unit is stored in a storage area such as the memory 302 or the recording medium 305 shown in FIG. 3, for example.

The storage unit 700 stores various information referred to or updated in the processing of each functional unit. The storage unit 700 may store the learning data, for example. The learning data is, for example, data in which the past data is associated with the result of the user's determination as to whether the past data is normal or abnormal. The past data is, for example, billing data 400. The anomaly is, for example, fraudulent billing.

The storage unit 700 may store, for example, a generation rule that generates a decision tree model from training data. The decision tree model is a model in which a plurality of nodes are included and the nodes are connected by an edge. The decision tree model includes, for example, a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal, and the decision tree model includes at least one leaf node. The name is set. Specifically, the storage unit 700 may store a machine learning FW (FirmWare) having a generation rule.

The storage unit 700 may store, for example, a decision tree model. The storage unit 700 may store, for example, a verbalization rule or a name template used when generating a name of a leaf node of a decision tree model. The verbalization rule is, for example, to rewrite the judgment result of the condition using the inequality sign by using words such as "greater than or equal to", "less than or equal to", "greater than", "less than", "less than", and "range". Is the rule of. The name template is, for example, a sentence such as "a tendency close to a known A is seen" or "a fraudulent claim related to a known A". For A, for example, a name set in another leaf node is inserted.

The storage unit 700 may store, for example, a screen template. The screen template is used when generating a screen to be displayed on the terminal device 201. The screen template is used, for example, when generating a screen for causing the user to set the reason for fraud or the reason for normality of the billing table 500. The screen template is used, for example, when generating a screen for outputting the billing data 400 in association with the name of the leaf node that classifies the billing data 400.

The acquisition unit 701 acquires various information used for processing of each functional unit and outputs the information to each functional unit. For example, the acquisition unit 701 may acquire various information used for processing of each functional unit from the storage unit 700 and output the information to each functional unit. For example, the acquisition unit 701 may acquire various information used for processing of each functional unit from a device different from the classification device 100 and output the information to each functional unit.

Specifically, the acquisition unit 701 receives the learning data from the terminal device 201, stores it in the storage unit 700, and outputs it to each function unit. Specifically, the acquisition unit 701 may accept the input of the learning data via the operation input of the user or the recording medium I / F 304. As a result, the acquisition unit 701 can output the learning data to the learning unit 702 and enable the learning unit 702 to generate the decision tree model.

Specifically, the acquisition unit 701 receives the target data from the terminal device 201, stores it in the storage unit 700, and outputs it to each function unit. Specifically, the acquisition unit 701 may accept the input of the target data via the operation input of the user or the recording medium I / F 304. As a result, the acquisition unit 701 can output the target data to the classification unit 703 and start the classification of the target data in the classification unit 703.

The learning unit 702 refers to the generation rule and generates a decision tree model based on the learning data. The learning unit 702 generates a decision tree model based on the learning data by using, for example, a machine learning FW. As a result, in the learning unit 702, the classification unit 703 can refer to the decision tree model, and the classification unit 703 can classify the target data.

The classification unit 703 classifies the target data by the decision tree model. The classification unit 703 traces child nodes based on the result of determining whether or not the target data satisfies the conditions represented by the nodes in order from the root node of the decision tree model, and sets the target data in one of the leaf nodes. Classify. As a result, the classification unit 703 can classify the target data into any leaf node and classify the target data as normal or abnormal.

When the target data is classified into the first leaf node whose name has not been set, the selection unit 704 uses the second leaf node whose name has been set in the decision tree model based on the positional relationship between the nodes. Select. The positional relationship between nodes is represented by, for example, the parent-child relationship between nodes.

For example, when the first leaf node is a leaf node indicating that the target data is abnormal, the selection unit 704 selects the second leaf node. As a result, the selection unit 704 can generate the name of the leaf node indicating that the target data is abnormal, which the user tends to attach importance to.

For example, when the first leaf node is a leaf node indicating that the target data is normal, the selection unit 704 does not have to select the second leaf node. As a result, when the user is required to grasp the reason why the target data is abnormally classified, the selection unit 704 does not have to generate a name corresponding to the reason why the target data is normally classified. Can be. Therefore, the selection unit 704 can reduce the processing amount of the generation unit 705.

The selection unit 704 selects the second leaf node from the leaf nodes in the vicinity of the first leaf node, for example, based on the positional relationship between the nodes. The distance between nodes is represented, for example, by the number of edges connecting the nodes. The neighborhood corresponds, for example, that the number of edges connecting the nodes is less than or equal to a predetermined number. The leaf node in the vicinity of the first leaf node is, for example, a leaf node that can be reached from the first leaf node via a predetermined number or less of edges.

Specifically, the selection unit 704 selects a second leaf node from among the leaf nodes that can be reached from the first leaf node via a predetermined number or less of edges. This makes it easier for the selection unit 704 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node. Therefore, the selection unit 704 can easily generate a name for the first leaf node that is intuitively easy for the user to understand.

A leaf node in the vicinity of the first leaf node is, for example, a leaf node above the first leaf node and below a higher node that can be reached from the first leaf node via a predetermined number of edges or less. Is. The higher level is the one closer to the root node. The lower part is closer to the leaf node.

Specifically, the selection unit 704 is the first leaf node among the leaf nodes above the first leaf node and below the upper node that can be reached from the first leaf node via a predetermined number of edges or less. Select 2 leaf nodes. This makes it easier for the selection unit 704 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node. Therefore, the selection unit 704 can easily generate a name for the first leaf node that is intuitively easy for the user to understand.

The selection unit 704 sets, for example, a predetermined number based on the depth of the decision tree model. Specifically, when the depth of the decision tree model exceeds the threshold value, the selection unit 704 sets a predetermined number to the depth * 20% of the decision tree model. The threshold is, for example, 20. Specifically, the selection unit 704 may set a predetermined number to a fixed value as long as the depth of the decision tree model is equal to or less than the threshold value. The fixed value is, for example, 2. This makes it easier for the selection unit 704 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node, depending on the depth of the decision tree model. Further, the selection unit 704 can limit the range of selection targets according to the depth of the decision tree model and reduce the processing amount.

The selection unit 704 is a second leaf node from among the leaf nodes included in the second subtree at a position different from the first subtree in the decision tree model based on the positional relationship between the nodes. Select. The first subtree is a subtree containing the first leaf node. The second subtree is located at a different position from the first subtree and is the same subtree as the first subtree. The subtree that is the same as the first subtree includes, for example, a node that represents the same conditions as the node included in the first subtree, and the subtree that has the same positional relationship between the first subtree and the node. Is.

The second subtree is located at a different position from the first subtree and may be a subtree similar to the first subtree. A subtree similar to the first subtree includes, for example, a node representing the same conditions as a node included in a part of the first subtree, and a positional relationship between the part of the first subtree and the node. It is a subtree including the part where is the same. This makes it easier for the selection unit 704 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node.

The generation unit 705 generates the name of the first leaf node based on the name of the second leaf node selected by the selection unit 704. The generation unit 705 further determines that the condition of each node on the first path is different from the judgment result of the condition of each node on the second path of the first leaf node. You may generate a name.

The first path is, for example, the path from the root node of the decision tree model to the first leaf node. The second path is, for example, the path from the root node of the decision tree model to the second leaf node. The difference is, for example, a determination result that does not overlap with the determination result regarding the condition represented by each node on the second route among the determination results regarding the condition represented by each node on the first route.

For example, the generation unit 705 refers to the storage unit 700, and sets the judgment result that makes a difference as words such as "greater than or equal to", "less than or equal to", "greater than", "less than", "less than", and "range". Generate a rewritten sentence using. For example, the generation unit 705 refers to the storage unit 700 and generates a sentence in which the name of the selected second leaf node is inserted in the "A" of the sentence "a tendency close to the known A is seen".

Then, the generation unit 705 combines the generated sentences to generate the name of the first leaf node. Specifically, the generation unit 705 generates a name such as "a tendency close to an inflated request is seen. The amount is 100 or less and the usage fee is 10 or more." As a result, the generation unit 705 can generate the name of the first leaf node based on the name of the second leaf node, which makes it easier for the user to generate a name that is intuitively easy to understand. it can.

In addition, the generation unit 705 can make the sentence in which the determination result that is different is rewritten can be referred to by the user. Therefore, the generation unit 705 can make it possible to grasp what kind of condition the determination result is different, and can generate a name that makes it easy to grasp the reason why the target data is classified as normal or abnormal.

The generation unit 705 sets the name of the generated first leaf node in the first leaf node. As a result, the generation unit 705 does not have to regenerate the name when the target data is classified into the first leaf node next time, and the processing amount can be reduced.

The output unit 706 outputs the target data in association with the name of the generated first leaf node. The output format is, for example, display on a display, print output to a printer, transmission to an external device by the network I / F 303, or storage in a storage area such as a memory 302 or a recording medium 305. As a result, the output unit 706 can enable the user to grasp the detailed type of the result of classifying the target data, and can grasp the reason why the target data is classified.

The output unit 706 may output the processing result of each functional unit. As a result, the output unit 706 can notify the user of the processing result of each functional unit, and can support the management and operation of the classification device 100, for example, the update of the set value of the classification device 100, and the classification device. It is possible to improve the convenience of 100.

(Specific example of the situation where the classification device 100 is used)
Next, a specific example of the situation in which the classification device 100 is used will be described with reference to FIG.

FIG. 8 is an explanatory diagram showing a specific example of a situation in which the classification device 100 is used. As shown in FIG. 8, in the classification device 100, for example, an employee of a ministry or a local government collects social security benefit billing data 806 from a business operator such as a long-term care facility or a nursing home, and provides guidance and audit work for the business operator. It is used in situations where The number of billing data 806 collected is, for example, one million units. In the following explanation, employees of ministries and local governments may be referred to as "users".

Here, it is desired that the user can efficiently perform the guidance and audit work of the business operator in a limited time. For example, it is desired that the user can efficiently find the billing data 806 for fraudulent billing from the collected billing data 806.

Further, for example, it is desired that the user can determine which business operator is preferable for guidance and auditing. Therefore, it is desired that the user can grasp the number of billing data 806 for fraudulent billing for each business operator. Further, it is desired that the user can grasp the number of fraudulent billing billing data 806 for each business operator and for each type of fraudulent billing.

Also, for example, it is desirable that a user prepares materials related to billing data 806 before visiting the location of any business in order to guide and audit any business. Therefore, it is desired that the user can use the type of fraudulent billing of the billing data 806 of the business operator to create the material related to the billing data 806.

Further, for example, it is desired that the user can determine what kind of fraudulent claim should be prioritized and considered when instructing and auditing any business operator. Therefore, the user can grasp what type of fraudulent billing billing data 806 is in the business operator's billing data 806, and the number of fraudulent billing billing data 806 is calculated for each fraudulent billing type. It is desirable to be able to grasp. On the other hand, the classification device 100 operates as shown below.

(8-1) The classification device 100 collects past billing data 801 in cooperation with a failure system such as a ministry or a local government. In addition, the classification device 100 collects past fraudulent billing case data 802 from employees of ministries and local governments. Next, the classification device 100 generates and stores the learning data set 803 by processing and combining the past billing data 801 and the past fraudulent billing case data 802. Then, the classification device 100 uses the machine learning FW804 to generate a decision tree model 805 based on the learning data set 803. The flow of generating the decision tree model 805 will be described later with reference to, for example, FIG.

(8-2) The classification device 100 classifies the new billing data 806 into leaf nodes according to the decision tree model 805, and classifies the new billing data 806 into normal or fraudulent billing. Here, if the name of the leaf node that classifies the new billing data 806 is not set, the classification device 100 generates the name of the leaf node that classifies the new billing data 806 and sets the name of the leaf node. .. The flow of setting the name of the leaf node will be described later with reference to, for example, FIGS. 10 and 11. Then, the classification device 100 stores the new billing data 806 and the name of the leaf node that classifies the new billing data 806 in association with each other.

(8-3) The classification device 100 presents to the user the corresponding data 807 in which the new billing data 806 and the name of the leaf node that classifies the new billing data 806 are associated with each other. Further, the classification device 100 may calculate the number of billing data 806 for fraudulent billing for each business operator and present it to the user. Further, the classification device 100 may calculate the number of fraudulent billing billing data 806 for each business operator and each type of fraudulent billing and present it to the user.

As a result, the classification device 100 can support the user to efficiently perform the guidance and audit work of the business operator in a limited time. Since the classification device 100 classifies new billing data 806 by, for example, the decision tree model 805, the user can efficiently find the billing data 806 for fraudulent billing, which imposes a burden on the user. It can be reduced.

Further, in order to present to the user the number of billing data 806 for fraudulent billing for each business operator, for example, it is preferable for the user to give priority to which business operator to give guidance and audit. Can be judged.

Further, in the classification device 100, for example, in order to present the number of fraudulent billing billing data 806 to the user for each business operator and for each type of fraudulent billing, the user gives priority to any business operator. It can be made possible to judge whether it is preferable to provide guidance and audit. For example, the user can distinguish between 10 businesses with fraudulent billing data 806 for date errors and 10 businesses with fraudulent billing data 806 for inflated bills. It can make it easier to judge whether it is preferable to give priority to guidance and auditing.

Further, since the classification device 100 presents to the user, for example, the name of the leaf node that classifies the new billing data 806, the user can easily understand the reason why the new billing data 806 is classified. Can be done. Then, the user can easily create a document regarding the billing data 806 for instructing and auditing the business operator.

Further, the classification device 100 presents to the user, for example, the name of the leaf node that classifies the new billing data 806, so that the user can use the billing data 806 of the business operator for any type of fraudulent billing. It is possible to grasp whether or not there is billing data 806. Next, the explanation shifts to FIG. 9, and the flow of generating the decision tree model 805 will be described.

(Flow to generate decision tree model 805)
FIG. 9 is an explanatory diagram showing a flow for generating the decision tree model 805. Here, the decision tree model 805 is preferably relatively small in scale. If the scale of the decision tree model 805 is relatively small, the number of conditions to be determined when classifying the target data tends to be small, and it tends to be easy to reduce the processing amount and the required time. Therefore, it is preferable.

Further, in the decision tree model 805, when the training data set 803 is divided into two under the conditions represented by any of the nodes, the number of normal training data and the abnormal training data included in the divided learning data subsets are obtained. It is preferable that there is a bias in the number of. Bias is preferable because the number of conditions used to classify normal training data and abnormal training data tends to be small.

Further, when one node represents a condition in which the value of one item is used, the condition in which the combination of the values of a plurality of items is used is represented by the combination of a plurality of nodes. Therefore, in the upper part of the decision tree model 805, since the number of nodes passing through is relatively small, a plurality of training data are divided into two based on the values of relatively few items. On the other hand, in the lower part of the decision tree model 805, since the number of nodes passing through is relatively large, a plurality of training data are divided into two based on the values of a relatively large number of items.

Therefore, first, the classification device 100 is a node representing a condition in which a plurality of training data can be relatively preferably divided into two based on the value of one item of the training data in the upper part of the decision tree model 805. It is preferable to generate. Then, the classification device 100 lowers the decision tree model 805, and each condition of the plurality of conditions capable of relatively preferably dividing the plurality of training data into two based on the values of the plurality of items of the training data. It is preferable to generate a node representing.

As shown in FIG. 9, first, the classification device 100 sets a node 901 representing a condition for classifying the learning data set 803 into a learning data subset 921 in which normal and abnormal are mixed and a learning data subset of abnormality 1. Generate. Further, since the classification device 100 can classify the learning data subset of the abnormality 1 without mixing the normal and the abnormality, the leaf node 911 representing the abnormality 1 is generated.

Similarly, the classification device 100 generates a node 902 representing a condition for classifying the learning data subset 921 into a learning data subset 922, which is a mixture of normal and abnormal, and a normal learning data subset 922. Further, since the classification device 100 can classify the normal learning data subset without mixing the normal and the abnormal, the leaf node 912 representing the normal is generated.

As a result, the classification device 100 can generate a decision tree model 805 including nodes 901 to 904 and leaf nodes 911 to 915. Next, the description shifts to the description of FIGS. 10 and 11, and the flow of setting the name to the leaf node to which any of the billing data 806 is classified and whose name has not been set will be described.

(Flow of setting the name of the leaf node)
10 and 11 are explanatory views showing a flow of setting a name for a leaf node. In FIG. 10, the classification device 100 stores the decision tree model 1000 and sets a name for the leaf node when classifying the billing data 806 of No. 1 and the billing data 806 of No2. The decision tree model 1000 includes nodes 101 to 1006 and leaf nodes 101 to 1017.

As shown in FIG. 10, the classification device 100 determines whether or not the billing data 806 of No. 1 satisfies the condition represented by the root node 1001. Since the classification device 100 satisfies the condition represented by the root node 1001, it is determined whether or not the billing data 806 of No. 1 satisfies the condition represented by the node 1002 which is a child node on the True side of the root node 1001.

Since the classification device 100 satisfies the condition represented by the node 1002, it is determined whether or not the billing data 806 of No. 1 satisfies the condition represented by the node 1004 which is a child node on the True side of the node 1002. Since the classification device 100 satisfies the condition represented by the node 1004, it is determined whether or not the billing data 806 of No. 1 satisfies the condition represented by the node 1006 which is a child node on the True side of the node 1004.

Since the classification device 100 satisfies the condition represented by the node 1006, the billing data 806 of No. 1 is classified into the leaf node 1016 on the True side of the node 1006. Here, the classification device 100 does not have to generate a name because the leaf node 1016 has a name set and is a named normal indicating normality.

The classification device 100 determines whether or not the billing data 806 of No. 2 satisfies the condition represented by the root node 1001. Since the classification device 100 satisfies the condition represented by the root node 1001, it is determined whether or not the billing data 806 of No. 2 satisfies the condition represented by the node 1002 which is a child node on the True side of the root node 1001.

Since the classification device 100 satisfies the condition represented by the node 1002, it is determined whether or not the billing data 806 of No. 2 satisfies the condition represented by the node 1004 which is a child node on the True side of the node 1002. Since the classification device 100 satisfies the condition represented by the node 1004, it is determined whether or not the billing data 806 of No. 2 satisfies the condition represented by the node 1006 which is a child node on the True side of the node 1004.

Since the classification device 100 does not satisfy the condition represented by the node 1006, the billing data 806 of No. 2 is classified into the leaf node 1017 on the False side of the node 1006. Here, the classification device 100 generates a name because the leaf node 1017 represents an unknown abnormality because the name has not been set. The classification device 100 selects, for example, a leaf node in the vicinity of the leaf node 1017. Here, the explanation shifts to FIG. 11, and the flow of selecting the leaf node will be described.

As shown in FIG. 11, the classification device 100 searches for leaf nodes whose names have been set in order from the lower leaf nodes, and selects the leaf nodes whose names have been set. The classification device 100 selects the leaf node 1016 from, for example, the leaf node 1013 or the leaf node 1016 which is in the vicinity of the leaf node 1017 and whose name has been set. The classification device 100 generates the name of the leaf node 1017 based on the name of the selected leaf node 1016.

Here, for example, in the decision tree model, the leaf node whose name has not been set and the leaf node in the vicinity of the leaf node whose name has not been set tend to include overlapping portions in the route from the root node. There is. In other words, the conditions that overlap between the case where the target data is classified into the leaf node whose name has not been set and the case where the target data is classified into the leaf node near the leaf node whose name has not been set. Tends to be used. Therefore, there is a relatively high possibility that the leaf node whose name has not been set and the leaf node in the vicinity of the leaf node whose name has not been set represent similar contents.

Further, for example, in the decision tree model, the leaf node at a relatively lower level is a leaf node in which the target data tends to be classified using a relatively large number of conditions. For this reason, the leaf nodes that are relatively inferior tend to represent relatively important anomalies for the user, which are difficult to find without referring to the values of a plurality of items included in the target data.

On the other hand, as shown in FIG. 11, the classification device 100 is a leaf whose name has been set in order from the leaf node which is in the vicinity of the leaf node whose name has not been set and which is relatively lower. Search for a node.

Therefore, the classification device 100 may use the name of the leaf node, which is relatively likely to represent a content similar to that of the leaf node whose name has not been set, for the generation of the name of the leaf node whose name has not been set. it can. As a result, the classification device 100 can easily generate a name that is relatively likely to correctly represent the content represented by the leaf node whose name has not been set, and the name is intuitively easy for the user to understand. Can be easily generated.

Further, the classification device 100 can use the name of the leaf node that tends to represent an abnormality that is relatively important to the user in generating the name of the leaf node for which the name has not been set. As a result, the classification device 100 can make it easier for the user to grasp the relationship between the abnormality that is relatively important to the user and the content represented by the leaf node whose name has not been set.

(Operation example 1 in which the classification device 100 sets a name for a leaf node)
Next, an operation example 1 in which the classification device 100 sets a name for the leaf node will be described with reference to FIGS. 12 to 19. The operation example 1 is an example in which the leaf node whose name has not been set and which is in the vicinity of the leaf node whose name has not been set is selected and the name of the leaf node whose name has not been set is generated.

12 to 19 are explanatory views showing an operation example 1 in which the classification device 100 sets a name for the leaf node. In FIG. 12, the classification device 100 receives the learning data set from the terminal device 201 and stores it using the learning table 1200. Since the data structure of the learning table 1200 is the same as the data structure of the billing table 500 shown in FIG. 5, the description thereof will be omitted. Next, the description proceeds to FIG.

In FIG. 13, the classification device 100 uses machine learning FW804 to generate a decision tree model 1300 based on a learning table. The decision tree model 1300 includes, for example, nodes 1301-1303 and leaf nodes 1311-1313. Next, the description shifts to FIG.

In FIG. 14, the classification device 100 receives billing data from the terminal device 201 and stores it using the billing table 1400. Here, the result of billing table 1400, fraudulent reason, and normal reason fields are blank. Next, the description shifts to FIG.

In FIG. 15, the classification device 100 classifies the target data as normal or abnormal according to the decision tree model 1300. The classification device 100 sets the result of classifying the target data as normal or abnormal in the result field of the billing table 1400. Next, the description shifts to FIG.

In FIG. 16, the classification device 100 causes the display 606 of the terminal device 201 to display a reception screen 1600 that accepts inputs of an illegal reason and a normal reason. The terminal device 201 receives the input of the reason for fraud and the reason for normal based on the user's operation input by the input device 607, and when the save button is clicked, transmits the input to the classification device 100. The classification device 100 receives the reason for fraud and the reason for normal from the terminal device 201.

Specifically, the terminal device 201 transmits the search conditions entered in the search condition field of the reception screen 1600 to the classification device 100, and details the records of the billing table 1400 based on the search conditions according to the control of the classification device 100. Display in the information column. The terminal device 201 reflects the fraudulent reason or normal reason selected from the list in the update content column in the record displayed in the detailed information column and checked in the selection box, and has been reflected in response to the click of the save button. The record is transmitted to the sorting device 100. Next, the description shifts to FIG.

In FIG. 17, the classification device 100 stores the received fraudulent reason and normal reason using the billing table 1400. Here, the normal reason associated with the key 1 is set. Further, the reason for fraud associated with the key 2 is set. On the other hand, the reason for fraud associated with the key 3 is not input to the user and is unknown. Next, the description shifts to FIG.

In FIG. 18, the classification device 100 sets a name for the leaf nodes 1311-1313. The classification device 100 associates a key with the leaf nodes 131 to 1313 and sets a name. For example, the classification device 100 associates the key 1 with the leaf node 1311 and sets the normal reason corresponding to the key 1 as a name.

For example, the classification device 100 associates the key 2 with the leaf node 1312 and sets the reason for fraud corresponding to the key 2 as a name. The classification device 100 associates the key 3 with the leaf node 1313, for example, but the reason for the fraud corresponding to the key 3 is unknown and the name is not set. Next, the description shifts to FIG.

In FIG. 19, the classification device 100 generates the name of the leaf node 1313 whose name has not been set. The classification device 100 selects, for example, a leaf node in the vicinity of the leaf node 1313. Then, the classification device 100 generates the name of the leaf node 1313 based on the name of the selected leaf node.

The classification device 100 selects one of the leaf nodes from, for example, the leaf nodes in the range from the hierarchy of the leaf node 1313 to the hierarchy higher than a predetermined number of the hierarchy of the leaf node 1313. The predetermined number is, for example, the depth * 20% of the decision tree model 1300.

Specifically, the classification device 100 reaches from the leaf node 1313 by the smallest number of edges among the leaf nodes in the range from the hierarchy of the leaf node 1313 to the hierarchy higher than a predetermined number of the hierarchy of the leaf node 1313. Select the leaf node to do.

In the example of FIG. 19, the classification device 100 identifies the leaf nodes 1311, 1312 in the range from the layer of the leaf node 1313 to the layer three steps higher. Next, the classification device 100 selects the leaf node 1312 closest to the leaf node 1313 among the leaf nodes 1311, 1312. Then, the classification device 100 generates the name of the leaf node 1313 based on the name of the selected leaf node 1312.

Here, the case where the classification device 100 generates the name of the leaf node when classifying the billing data different from the learning data has been described, but the present invention is not limited to this. For example, after the classification device 100 generates a decision tree model based on the training data, the name of the leaf node may be generated while trying to classify the training data by the decision tree model.

(Operation example 2 in which the classification device 100 sets a name for the leaf node)
Next, operation example 2 in which the classification device 100 sets a name for the leaf node will be described with reference to FIGS. 20 to 22. The operation example 1 is an example in which the leaf node whose name has not been set and which is in the vicinity of the leaf node whose name has not been set is selected and the name of the leaf node whose name has not been set is generated. On the other hand, the operation example 2 is an example in which a leaf node whose name has not been set and which is not in the vicinity of the leaf node whose name has not been set is selected to generate the name of the leaf node whose name has not been set.

20 to 22 are explanatory views showing an operation example 2 in which the classification device 100 sets a name for the leaf node. In FIG. 20, it is assumed that the classification device 100 classifies the target data into the leaf nodes 2001 by the decision tree model 2000. It is assumed that the leaf node 2001 has not been named.

Here, the classification device 100 identifies the subtree 2010 containing the leaf node 2001. The classification device 100 identifies, for example, the subtree 2010 in the range from the hierarchy of the leaf node 2001 to the hierarchy higher than a predetermined number of the hierarchy of the leaf node 2001. The predetermined number is, for example, the depth * 20% of the decision tree model. Next, the description shifts to FIG. 21, and an example of the subtree 2010 will be described.

As shown in FIG. 21, the subtree 2010 includes nodes 2101 to 2103 representing conditions and leaf nodes 211 to 2114. Further, the parent-child relationship between the nodes of the subtree 2010 is, for example, the relationship that the child nodes of the node 2101 have the leaf nodes 2111 and the nodes 2102. Next, the description shifts to FIG. 22.

In FIG. 22, the classification device 100 searches the decision tree model for the same subtree as the subtree 2010. The subtree that is the same as the subtree 2010 is, for example, a subtree that includes nodes that represent the same conditions as the nodes 2101 to 2103, and has the same parent-child relationship between the subtree 2010 and the nodes.

The classification device 100 identifies the same subtree 2202, 2204 as the subtree 2010 from, for example, the subtrees 2201 to 2204 of the decision tree model. Next, the classification device 100 selects the leaf nodes included in the subtrees 2202 and 2204 and whose names have been set. Then, the classification device 100 generates the name of the leaf node 2001 based on the name of the selected leaf node.

The classification device 100 may select leaf nodes from relatively lower subtrees, for example. Specifically, the classification device 100 selects a leaf node from the relatively lower subtree 2202 among the subtrees 2202 and 2204. As a result, the classification device 100 can select another leaf node even if there is no leaf node whose name has been set in the vicinity of the leaf node 2001.

Further, since the classification device 100 selects a leaf node from a subtree that is relatively inferior, it tends to represent an abnormality that is relatively important to the user in generating the name of the leaf node for which the name has not been set. You can use the name of the leaf node. As a result, the classification device 100 can make it easier for the user to grasp the relationship between the abnormality that is relatively important to the user and the content represented by the leaf node whose name has not been set.

Here, the case where the classification device 100 executes either the operation example 1 or the operation example 2 has been described, but the present invention is not limited to this. For example, the classification device 100 may combine the operation example 1 and the operation example 2. Specifically, there may be a case where the classification device 100 executes the operation example 2 depending on the fact that the name could not be generated by the operation example 1. The various flowcharts described in FIGS. 25 to 27 correspond to the case where the operation example 1 and the operation example 2 are combined.

(Output example 1 in the terminal device 201)
Next, an output example 1 in the terminal device 201 will be described with reference to FIG. 23.

FIG. 23 is an explanatory diagram showing an output example 1 in the terminal device 201. In FIG. 23, the classification device 100 sets the name of the leaf node according to the operation example 1 or the operation example 2, and sets the detail number of the invoice for the social security benefit expense as the source of the billing data classified into the leaf nodes. Output in association with each other.

The classification device 100 displays, for example, the display content 2300 on the display 606 of the terminal device 201. The name of the leaf node is set in the content field of the display content 2300. As a result, the classification device 100 can support the user to efficiently perform the guidance and audit work of the business operator in a limited time.

The classification device 100 can make it easy to understand, for example, the reason why the billing data is classified. Further, the classification device 100 can, for example, make it possible to grasp what kind of fraudulent billing billing data is included in the billing data of the business operator. Therefore, the user can easily create materials related to billing data for instructing and auditing the business operator.

(Output example 2 in the terminal device 201)
Next, an output example 2 in the terminal device 201 will be described with reference to FIG. 24.

FIG. 24 is an explanatory diagram showing an output example 2 in the terminal device 201. In FIG. 24, the classification device 100 calculates and outputs the number of billing data classified as fraudulent billing for each business operator according to the operation example 1 or the operation example 2.

The classification device 100 displays, for example, the display content 2400 on the display 606 of the terminal device 201. In the field of the number of display contents 2400, the number of billing data classified as fraudulent billing is set. As a result, the classification device 100 can support the user to efficiently perform the guidance and audit work of the business operator in a limited time. The user can refer to, for example, the number of billing data classified as fraudulent billing and determine which business operator is preferable for guidance and auditing.

Further, the classification device 100 may display the number of billing data for fraudulent billing, for example, for each business operator and for each type of fraudulent billing. As a result, the user can determine which business operator is preferable for guidance and auditing.

(Overall processing procedure)
Next, an example of the overall processing procedure executed by the classification device 100 will be described with reference to FIG. 25. The entire processing is realized by, for example, the CPU 301 shown in FIG. 3, a storage area such as a memory 302 or a recording medium 305, and a network I / F 303.

FIG. 25 is a flowchart showing an example of the overall processing procedure. In FIG. 25, first, the classification device 100 classifies each target data of the plurality of target data into leaf nodes representing normal or abnormal based on the learned decision tree model (step S2501).

Next, the classification device 100 reflects the classification result in the billing table 500 (step S2502). Then, the classification device 100 receives the input of the reason why each target data of at least one target data is classified and reflects it in the billing table 500 (step S2503).

Next, the classification device 100 sets the reason why each target data of the received at least one target data is classified in the name of the leaf node in which each target data is classified (step S2504). Then, if there is target data classified into the leaf nodes whose names have not been set, the classification device 100 selects the leaf nodes whose names have been set by executing the selection process described later in FIG. 26 (step S2505). ).

Next, the classification device 100 acquires the name of the selected leaf node (step S2506). Then, the classification device 100 identifies the difference between the condition represented by the node on the route from the root node to the unnamed leaf node and the condition represented by the node on the route from the root node to the selected leaf node. (Step S2507).

Next, the classification device 100 sets the name of the leaf node whose name has not been set based on the name of the selected leaf node and the specified difference (step S2508). Then, the classification device 100 ends the entire processing. As a result, the classification device 100 can set the name of the leaf node and make it possible to present the name of the leaf node to the user.

(Selection processing procedure)
Next, an example of the selection processing procedure executed in step S2505 will be described with reference to FIG. 26. The selection process is realized, for example, by the CPU 301 shown in FIG. 3, a storage area such as a memory 302 or a recording medium 305, and a network I / F 303.

FIG. 26 is a flowchart showing an example of the selection processing procedure. In FIG. 26, first, the classification device 100 sets the parent node of the leaf node whose name has not been set as the reference node (step S2601).

Next, the classification device 100 determines whether or not a name is set for the child node on the right side of the reference node (step S2602). Here, when a name is set for the child node on the right side (step S2602: Yes), the classification device 100 shifts to the process of step S2604. On the other hand, when the name is not set for the child node on the right side (step S2602: No), the classification device 100 shifts to the process of step S2603.

In step S2603, the classification device 100 determines whether or not a name is set for the child node on the left side of the reference node (step S2603). Here, when a name is set for the child node on the left side (step S2603: Yes), the classification device 100 shifts to the process of step S2604. On the other hand, when the name is not set for the child node on the left side (step S2603: No), the classification device 100 shifts to the process of step S2605.

In step S2604, the classification device 100 selects a leaf node whose name has been set from the child nodes of the reference node (step S2604). Then, the classification device 100 ends the selection process.

In step S2605, the classification device 100 determines whether or not the number of changes of the reference node> the depth of the decision tree * 10% (step S2605). Here, when the number of changes of the reference node> the depth of the decision tree * 10% is not satisfied (step S2605: No), the classification device 100 shifts to the process of step S2606. On the other hand, when the number of changes of the reference node> the depth of the decision tree * 10% (step S2605: Yes), the classification device 100 shifts to the process of step S2607.

In step S2606, the classification device 100 changes the parent node of the reference node to the reference node (step S2606). Then, the classification device 100 shifts to the process of step S2602.

In step S2607, the classification device 100 does not change the reference node and executes the search process shown in FIG. 27 to resemble the subtree in the vicinity of the unnamed leaf node from the entire decision tree model. (Step S2607).

Next, the classification device 100 selects a leaf node whose name has been set from among the other searched leaf nodes of the subtree (step S2608). Then, the classification device 100 ends the selection process. Thereby, the classification device 100 can select the leaf node used when generating the name of the leaf node.

(Search processing procedure)
Next, an example of the search processing procedure executed in step S2505 will be described with reference to FIG. 27. The search process is realized, for example, by the CPU 301 shown in FIG. 3, a storage area such as a memory 302 or a recording medium 305, and a network I / F 303.

FIG. 27 is a flowchart showing an example of the search processing procedure. In FIG. 27, first, the classification device 100 includes a leaf node whose name has not been set, and the name has not been set from a hierarchy higher than the hierarchy of the leaf node whose name has not been set by the depth of the decision tree * 10%. Identify subtrees up to the leaf node hierarchy (step S2701).

Next, the classification device 100 searches the entire decision tree model for one or more subtrees including nodes representing conditions that are the same as or similar to the conditions represented by the nodes included in the specified subtree (step S2702). Then, the classification device 100 searches for the subtree in the deepest hierarchy among the searched subtrees of one or more (step S2703). After that, the classification device 100 ends the search process. As a result, the classification device 100 can search for subtrees that fall within the range for selecting leaf nodes.

As described above, the classification device 100 can classify the target data by the decision tree model. According to the classification device 100, when the target data is classified into the first leaf node whose name has not been set, the second of the decision tree models whose name has been set is set based on the positional relationship between the nodes. Leaf nodes can be selected. According to the classification device 100, the name of the first leaf node can be generated based on the name of the selected second leaf node. As a result, the classification device 100 makes it possible for the user to grasp the type of the result of classifying the target data and to grasp the reason why the target data is classified.

According to the classification device 100, when the first leaf node is a leaf node indicating that the target data is abnormal, the second leaf node can be selected. As a result, the classification device 100 can reduce the processing amount.

According to the classification device 100, the second leaf node can be selected from the leaf nodes in the vicinity of the first leaf node based on the positional relationship between the nodes. This makes it easier for the classifier 100 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node. Therefore, the classification device 100 can easily generate a name that is easy for the user to intuitively understand for the first leaf node.

According to the classification device 100, the second leaf node can be selected from the leaf nodes that can be reached from the first leaf node via a predetermined number or less of edges. As a result, the classification device 100 can easily generate a name that is intuitively easy for the user to understand for the first leaf node.

According to the classification device 100, the second leaf is among the leaf nodes above the first leaf node and below the upper node that can be reached from the first leaf node via a predetermined number of edges or less. You can select a node. As a result, the classification device 100 can easily generate a name that is intuitively easy for the user to understand for the first leaf node.

According to the classification device 100, a predetermined number used when selecting the second leaf node can be set based on the depth of the decision tree model. As a result, the classification device 100 can limit the range of selection targets according to the depth of the decision tree model and reduce the processing amount.

According to the classification device 100, the leaf nodes included in the second subtree that is the same as or similar to the first subtree containing the first leaf node in the decision tree model based on the positional relationship between the nodes. A second leaf node can be selected from among them. This makes it easier for the classifier 100 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node. Further, the classification device 100 can select the second leaf node even if there is no leaf node for which the name has been set in the vicinity of the first leaf node.

According to the classification device 100, the second subtree includes a node representing the same conditions as the node included in the first subtree, and the subtree has the same positional relationship between the first subtree and the node. Can be identified. This makes it easier for the classifier 100 to select a second leaf node that represents normal or abnormal, which is relatively similar to the first leaf node.

According to the classification device 100, further, the determination result regarding the condition represented by each node on the route from the root node to the first leaf node and the condition represented by each node on the route from the root node to the second leaf node. It is possible to identify the difference from the judgment result regarding. According to the classification device 100, the name of the first leaf node can be generated based on the identified difference. As a result, the classification device 100 can grasp what kind of condition the determination result is different, and can generate a name that makes it easy to grasp the reason why the target data is classified as normal or abnormal. it can.

According to the classification device 100, the name of the generated first leaf node can be associated with the target data and output. Thereby, the classification device 100 can enable the user to grasp the name of the first leaf node in which the target data is classified.

According to the classification device 100, the name of the generated first leaf node can be set in the first leaf node. As a result, the classification device 100 does not have to regenerate the name when the target data is classified into the first leaf node next time, and the processing amount can be reduced.

The classification method described in the present embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. The classification program described in this embodiment is recorded on a computer-readable recording medium such as a hard disk, flexible disk, CD-ROM, MO, or DVD, and is executed by being read from the recording medium by the computer. Further, the classification program described in the present embodiment may be distributed via a network such as the Internet.

The following additional notes are further disclosed with respect to the above-described embodiment.

(Appendix 1) To the computer
A name is set for at least one of the leaf nodes, including a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is classified by the decision tree model, and
When the target data is classified into the first leaf node whose name has not been set, the second leaf node whose name has been set is selected from the decision tree models based on the positional relationship between the nodes. And
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification program characterized by executing processing.

(Appendix 2) The classification program according to Appendix 1, wherein the first leaf node is a leaf node indicating that the target data is abnormal.

(Appendix 3) The selected process is
The classification program according to Appendix 1 or 2, wherein the second leaf node is selected from the leaf nodes in the vicinity of the first leaf node based on the positional relationship between the nodes. ..

(Supplementary note 4) The leaf node in the vicinity of the first leaf node is a leaf node that can be reached from the first leaf node via a predetermined number or less of edges. Described classification program.

(Appendix 5) The leaf node in the vicinity of the first leaf node is a higher node that is above the first leaf node and can be reached from the first leaf node via a predetermined number or less of edges. The classification program according to Appendix 3, characterized in that it is a lower leaf node.

(Supplementary note 6) The classification program according to Supplementary note 5, wherein the predetermined number is set based on the depth of the decision tree model.

(Appendix 7) The selected process is
From among the leaf nodes included in the second subtree that is the same as or similar to the first subtree including the first leaf node in the decision tree model based on the positional relationship between the nodes. The classification program according to any one of Supplementary notes 1 to 6, characterized in that a second leaf node is selected.

(Appendix 8) The second subtree includes a node representing the same conditions as the node included in the first subtree, and the positional relationship between the first subtree and the node is the same. The classification program according to Appendix 7, which comprises the above.

(Appendix 9) The generated process further includes a determination result regarding the conditions represented by each node on the path from the root node of the determination tree model to the first leaf node, and the second leaf from the root node. The description in any one of Supplementary notes 1 to 8, wherein the name of the first leaf node is generated based on the difference from the determination result regarding the condition represented by each node on the route to the node. Classification program.

(Appendix 10) To the computer
The classification program according to any one of Supplementary note 1 to 9, wherein the target data is associated with the name of the generated first leaf node and output, and the process is executed.

(Appendix 11) To the computer
The classification program according to any one of Supplementary note 1 to 10, wherein a process is executed in which the name of the generated first leaf node is set in the first leaf node.

(Appendix 12) The computer
A name is set for at least one of the leaf nodes, including a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is classified by the decision tree model, and
When the target data is classified into the first leaf node whose name has not been set, the second leaf node whose name has been set is selected from the decision tree models based on the positional relationship between the nodes. And
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification method characterized by performing processing.

(Appendix 13) A node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal are included, and at least one of the leaf nodes is named. According to the decision tree model in which is set, the target data is classified and
When the target data is classified into the first leaf node whose name has not been set, the second leaf node whose name has been set is selected from the decision tree models based on the positional relationship between the nodes. And
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification device having a control unit.

100 Sorting device 110,805,1000,1300 Decision tree model 111-116,901-904,1001-1006,1301-1303,2101-2103 Nodes 121-127,911-915,101-11-1017,1311-11332001 , 2111-2114 Leaf Node 200 Classification System 201 Terminal Equipment 210 Network 300,600 Bus 301,601 CPU
302,602 Memory 303,603 Network I / F
304,604 Recording medium I / F
305,605 Recording medium 400,801,806 Billing data 500,1400 Billing table 606 Display 607 Input device 700 Storage unit 701 Acquisition unit 702 Learning unit 703 Classification unit 704 Selection unit 705 Generation unit 706 Output unit 802 Case data 803 Learning data set 804 Machine Learning FW
807 Corresponding data 921,922 Learning data Subset 1200 Learning table 2010, 2201-2204 Subtree 2300, 2400 Display contents

Claims (8)

On the computer
A name is set for at least one of the leaf nodes, including a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is classified by the decision tree model, and
When the target data is classified into the first leaf node whose name has not been set, the first leaf node and the other leaf nodes in the determination tree model are selected based on the positional relationship between the nodes. Based on the number of connected edges, select the second leaf node whose name has been set from the leaf nodes determined to be in the vicinity of the first leaf node.
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification program characterized by executing processing. The classification program according to claim 1, wherein the first leaf node is a leaf node indicating that the target data is abnormal. The selected process is
Based on the positional relationship between the nodes, the first portion of the determination tree model includes nodes representing the same conditions as each node included in the first subtree including the first leaf node. A second subtree in which the positional relationship between the tree and the node is the same, or a node representing the same conditions as each node included in a part of the first subtree, and the first subtree includes a node. A second subtree including a part having the same positional relationship between the node and the part is specified, and the second leaf node is selected from the leaf nodes included in the specified second subtree. , The classification program according to claim 1 or 2. In the selection process, one of the second leaf nodes is selected.
The generated process further includes a determination result regarding the conditions represented by each node on the path from the root node of the determination tree model to the first leaf node, and the second leaf node selected from the root node. The classification according to any one of claims 1 to 3, wherein the name of the first leaf node is generated based on the difference from the determination result regarding the condition represented by each node on the path of. program. On the computer
The classification program according to any one of claims 1 to 4, wherein the target data is associated with the name of the generated first leaf node and output, and the process is executed. On the computer
The classification program according to any one of claims 1 to 5, wherein a process is executed in which the name of the generated first leaf node is set in the first leaf node. The computer
A name is set for at least one of the leaf nodes, including a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is classified by the decision tree model, and
When the target data is classified into the first leaf node whose name has not been set, the first leaf node and the other leaf nodes in the determination tree model are selected based on the positional relationship between the nodes. Based on the number of connected edges, select the second leaf node whose name has been set from the leaf nodes determined to be in the vicinity of the first leaf node.
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification method characterized by performing processing. A name is set for at least one of the leaf nodes, including a node representing a condition for classifying the target data and a leaf node indicating that the target data is normal or the target data is abnormal. The target data is classified by the decision tree model, and
When the target data is classified into the first leaf node whose name has not been set, the first leaf node and the other leaf nodes in the determination tree model are selected based on the positional relationship between the nodes. Based on the number of connected edges, select the second leaf node whose name has been set from the leaf nodes determined to be in the vicinity of the first leaf node.
Generates the name of the first leaf node based on the name of the selected second leaf node.
A classification device having a control unit.

Images