KR20210029111A - Method and apparatus for explaining the reason of the evaluation resulted from credit evaluation models based on machine learning - Google Patents

Abstract

According to one aspect of the present invention, a method for explaining a reason of evaluation of a credit rating classification model based on machine learning performed by a computer device includes the steps of: (a) receiving first customer information including credit information composed of a plurality of items and a target class to be derived as an evaluation result of the credit rating classification model; (b) selecting a specific item affecting a change of the class determined by the credit rating classification model to the target class with respect to first customer information; (c) changing a value of the selected specific item and generating second customer information including a changed item value; (d) calculating a result probability that the target class is derived as an evaluation result of a credit rating classification model with respect to an input of second customer information; and (e) comparing whether the second customer information is determined as the target class.

Description

A device and method to explain the evaluation reason of a machine learning-based credit rating classification model {METHOD AND APPARATUS FOR EXPLAINING THE REASON OF THE EVALUATION RESULTED FROM CREDIT EVALUATION MODELS BASED ON MACHINE LEARNING}

The present invention relates to an apparatus and method for explaining the evaluation reason for a machine learning-based credit rating classification model.

Recently, as machine learning technology has developed greatly, attempts to apply machine learning models for credit evaluation of individuals and companies have increased in the financial field related to loans. The application of the machine learning-based credit rating classification model in the financial field means that credit ratings are not performed by experts in the financial field as in the past, but that the computer can learn by itself and automatically classify the data by analyzing the characteristics of the data. In addition, the classification model based on machine learning can find out the relationship between items, and because inference from complex data is possible, it is possible to derive high performance.

However, as both the number of parameters to be trained at the same time and the complexity of the model increase dramatically, there is a fatal disadvantage that humans cannot understand the reason for determining the classification of the model. Incomprehensible to humans means that they cannot be used in fields or companies where reliability is important. In particular, the issue of credit rating decisions in the financial sector is directly related to corporate interests, so it is difficult to apply if the model's decisions are unreliable regardless of the high performance of the credit rating model.

Therefore, research on explainable artificial intelligence that enables understanding of machine learning-based model decisions has recently attracted attention. Mainly, in these studies, a method of finding items that have a significant influence on deriving model decisions for a specific class is used.

Korean Registered Patent Publication No. 10-2005734 (Title of invention: Comprehensive financial service providing system reflecting the results of quantitative evaluation of tangible assets based on machine learning)

The present invention is to solve the above-described problem, a machine learning-based credit rating classification model that finds effective customer credit information to be determined as a target class and determines how to change the value of the information to change the credit rating result It is a technical problem to provide an apparatus and a method for explaining the reason for the evaluation.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for solving the above-described technical problem, a method for explaining the evaluation reason of the machine learning-based credit rating classification model performed by the computer device according to an aspect of the present invention is (a) credit consisting of a plurality of items. Receiving first customer information including information and a target class to be derived as an evaluation result of the credit rating classification model; (b) selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information; (c) changing the value of the selected specific item and generating second customer information including the changed item value; (d) calculating a result probability of deriving a target class as an evaluation result of the credit rating classification model for input of the second customer information; And (e) comparing whether the second customer information is determined to be a target class.

If the second customer information is not determined as the target class in step (e), it is sequentially repeated from the process of selecting a specific item, but each time it is repeated, the number of specific items is added one by one, and the second customer information When is determined to be the target class, the step of detecting and providing changed items of the second customer information is further included.

In step (b), specific items are selected in the order in which the absolute value of the derivative for the second customer information is large, based on the probability of the result from which the target class is derived, but indexing is performed in the order in which the absolute value of the derivative is as large as the number of items. do.

In step (c), the value of the first customer information for the indexed item is changed to the item value of the second customer information, but the changed item value by minimizing the value of the preset loss function based on the first and second customer information. To decide.

An apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model according to another aspect of the present invention includes a memory storing a program for performing a method of explaining the evaluation reason for the machine learning-based credit rating classification model; And a processor that performs a program, wherein the processor inputs first customer information including credit information composed of a plurality of items and a target class to be derived as an evaluation result of the credit rating classification model according to the execution of the program. 2nd customer information including receiving, selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information, changing the value of the selected specific item, and including the changed item value And, as a result of evaluation of the credit rating classification model for input of the second customer information, a result probability of deriving a target class is calculated, and whether the second customer information is determined as the target class is compared.

If the second customer information is not determined to be the target class, the processor sequentially repeats from the process of selecting a specific item, but adds one specific number of items each time it is repeated, and the second customer information is the target class. If determined to be, the changed items of the second customer information are detected and provided.

The processor selects specific items in the order in which the absolute value of the derivative for the second customer information is large, based on the result probability of deriving the target class, and performs indexing in the order in which the absolute value of the derivative is as large as the number of items.

The processor changes the value of the first customer information to the item value of the second customer information for the indexed item, but determines the changed item value by minimizing the value of a preset loss function based on the first and second customer information. .

According to one of the above-described problem solving means of the present application, the machine learning-based credit rating classification model not only understands the behavior of the model, but also intensively considers the model's behavior by finding sensible items between classes that can or cannot be borrowed. Since it is possible to check what the item is, it is possible to know the features that should be considered important in the development process of a new credit rating model in the future.

In addition, if there is a common factor used to determine the target class in a large number of customer data, even when modeling afterwards, a method that restricts the item to be considered as an important feature by the credit rating classification model is used to create an improved model. It can help you build.

In addition, a technology that changes the value of a corresponding item among credit information to be determined as a target class can benefit both companies that make loans and customers who make loans. For example, if the credit rating classification model for one customer's information is that it is not possible to borrow, the position of a company that makes a loan can explain that a loan is possible if the value of a certain item is appropriately changed to the customer. In other words, the more loans to high-credit customers are, the more profitable the company is, this can be a good effect. In addition, from the viewpoint of a customer who wants to lend, it is possible to obtain information on the items to be considered in order to borrow because it is possible to know which item of their credit information is negatively acting on the loan.

1 is a block diagram showing the configuration of an apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.
2 is a block diagram illustrating a detailed module of an apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.
FIG. 3 is a diagram showing credit information of customer A according to an embodiment of the present invention and a result of a credit rating classification model therefor.
FIG. 4 is a diagram illustrating an item to be changed among credit information in order to derive a target class of credit information of customer A according to a credit rating classification model according to an embodiment of the present invention.
5 is a flowchart illustrating a method of explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" with another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

In the present specification, the term "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized by using two or more hardware, or two or more units may be realized by one piece of hardware. Meanwhile,'~ unit' is not meant to be limited to software or hardware, and'~ unit' may be configured to be in an addressable storage medium or configured to reproduce one or more processors. Thus, as an example,'~ unit' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, and procedures. , Subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays and variables. Components and functions provided in the'~ units' may be combined into a smaller number of elements and'~ units', or may be further separated into additional elements and'~ units'. In addition, components and'~ units' may be implemented to play one or more CPUs in a device or a security multimedia card.

The "device for explaining the reason for evaluating the machine learning-based credit rating classification model" mentioned below may be implemented as a computer or portable terminal that can access a server or other terminal through a network. Here, the computer includes, for example, a notebook equipped with a web browser, a desktop, a laptop, and the like, and the portable terminal is, for example, a wireless communication device that guarantees portability and mobility. , Various types of handheld-based wireless communication devices such as various mobile communication-based terminals, smart phones, and tablet PCs may be included. In addition, “network” refers to a wired network such as a local area network (LAN), a wide area network (WAN), or a value added network (VAN), or a mobile radio communication network or satellite. It can be implemented with any type of wireless network such as a communication network.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing the configuration of an apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.

Referring to FIG. 1, the apparatus 100 for explaining the reason for the evaluation of the machine learning-based credit rating classification model includes a memory 103 and a program storing a program for performing a method of explaining the evaluation reason for the machine learning-based credit rating classification model. Including a processor 101 to perform, the processor 101 is to derive as an evaluation result of the first customer information including credit information composed of a plurality of items and the credit rating classification model 102 according to the execution of the program. A target class is input, selects a specific item that affects the change of the class determined by the credit rating classification model 102 to the target class for the first customer information, changes the value of the selected specific item, and changes the item Generates second customer information including a value, calculates a result probability of deriving a target class as the evaluation result of the credit rating classification model 102 for input of the second customer information, and calculates the result probability that the second customer information is the target class. Compare whether it is judged or not.

Therefore, the apparatus 100 for explaining the evaluation reason for the machine learning-based credit rating classification model 102 of the present invention is given some kind of customer information (first customer information), and the corresponding information (for example, annual salary, past bankruptcy experience) , FICO score, loan application amount, etc.)When the credit rating classification model is determined to be class A, the original customer information (first customer information) is determined to be the target class B. ), you can provide customer information (second customer information) that needs to be changed. Through this, the credit rating classification model of the present invention can provide an interpretation of what are distinctive and characteristic credit information items between the existing class A and the target class B.

In other words, the present invention is intended to provide an explanation of the result of the model (target class) calculated by inputting given customer information (a plurality of credit information items) rather than an overall interpretation of the credit rating classification model. At this time, the purpose of the explanation is to visualize the items of credit information that are important in determining as the target class, and to propose a method to determine the target class by changing the value of the target item.

That is, the present invention is an effective technique in understanding and interpreting the behavior of a model by explaining exactly what items should be changed for a specific customer to be determined as a target class.

As shown in FIG. 1, the apparatus 100 for explaining the evaluation reason for the machine learning-based credit rating classification model includes a memory 103, a processor 101, a database 104, and a credit rating classification model 102. I can.

The memory 103 stores a program for performing a method of explaining the evaluation reason of the machine learning-based credit rating classification model. In the program to perform the method of explaining the evaluation reason of the machine learning-based credit rating classification model, input the first customer information including credit information consisting of a plurality of items and the target class to be derived as the evaluation result of the credit rating classification model. 2nd customer information including receiving, selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information, changing the value of the selected specific item, and including the changed item value And, as a result of evaluation of the credit rating classification model for input of the second customer information, a result probability of deriving a target class is calculated, and whether the second customer information is determined as the target class is compared.

The credit rating classification model 102 predicts whether or not a loan is made by learning through an artificial neural network by inputting customer credit information.

The memory 103 contains an operating system for driving the apparatus 100 for explaining the evaluation reason for the machine learning-based credit rating classification model or a program for performing a method for explaining the evaluation reason for the machine learning-based credit rating classification model. Several types of data that occur during execution are stored.

In this case, the memory 103 collectively refers to a nonvolatile storage device that continuously maintains stored information even when power is not supplied, and a volatile storage device that requires power to maintain the stored information.

In addition, the memory 103 may perform a function of temporarily or permanently storing data processed by the processor 101. Here, the memory 103 may include a magnetic storage media or a flash storage media in addition to a volatile storage device requiring power to maintain stored information, but the scope of the present invention is limited thereto. It does not become.

The processor 101 executes the program stored in the memory 103, but controls the entire process according to the execution of the program for performing a method of explaining the evaluation reason of the machine learning-based credit rating classification model. Each operation performed by the processor 101 will be described in more detail later.

The processor 101 may include all kinds of devices capable of processing data. For example, it may refer to a data processing device embedded in hardware having a circuit that is physically structured to perform a function represented by a code or command included in a program. As an example of the data processing device built into the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, and an application-specific application-specific device (ASIC) Integrated circuit) and processing devices such as a field programmable gate array (FPGA) may be covered, but the scope of the present invention is not limited thereto.

The database 104 stores or provides data necessary for the device 100 explaining the reason for evaluating the machine learning-based credit rating classification model under the control of the processor 101. The database 104 may be included as a separate component from the memory 103 or may be built in a partial area of the memory 103.

2 is a block diagram illustrating a detailed module of an apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.

Meanwhile, the processor 101 may perform various functions according to the execution of the program stored in the memory 103, and the detailed modules included in the processor 101 according to each function are input unit 110 and the item selection unit 120 ), an item value change unit 130, a classification model result unit 140, a comparison unit 150, and a description unit 160.

The input unit 110 may receive first customer information including credit information composed of a plurality of items and a target class to be derived as an evaluation result of the credit rating classification model from the user. For example, the user may input first customer information (original credit information of an individual or corporate customer) and a target class (loanable or non-loanable) through the input unit 110. At this time, the credit information may be composed of a plurality of items such as an annual salary, past bankruptcy experience, FICO score, and loan application amount.

The item selection unit 120 may select a specific item that affects changing the class determined by the credit rating classification model 102 to the target class for the first customer information.

The item value change unit 130 may change a value of a selected specific item and generate second customer information including the changed item value.

Specifically, the item selection unit 120 selects specific items in the order in which the absolute value of the derivative for the second customer information is large, based on the result probability of deriving the target class, but in the order in which the absolute value of the derivative is as large as the number of items. Indexing can be performed as it is. Here, the meaning of the item with the largest differential absolute value means that the change in the item value gives the greatest change in the probability of the target class compared to other items. That is, in determining that the first customer information is a target class, the item selection unit 120 plays a role of selecting and increasing the effective credit information items that have an effect on one by one.

In addition, the item value change unit 130 changes the value of the first customer information to the item value of the second customer information for the indexed item, but minimizes the value of the preset loss function based on the first and second customer information. The changed item value can be determined. Here, the loss function is as shown in [Equation 1] below.

[Equation 1]

f _k is the logic score for class k (the value before the soft max function), K is the number of classes, N is the number of training data as a hyper parameter, training data i, t is the training data determined by t as the target class, It means hyper parameter. In the present invention, N is set to 100 and λ is set to 0.3. To minimize the value of the front side of the loss function, the logic score of the second customer information (the changed customer information of A) becomes similar to the logic score of the training data determined as the target class. Minimizing the value of the back side is to make the second customer information (the changed customer information of A) become similar to the first customer information (the original customer information).

The classification model result unit 140 may calculate a result probability of deriving a target class as an evaluation result of the credit rating classification model 102 for input of the second customer information. That is, the classification model result unit 140 calculates the result probability value of the credit rating classification model 102 that inputs the second customer information (changed customer A information) in order to derive the target class (eg, loanable). Can be calculated.

The comparison unit 150 may compare whether the second customer information is determined to be a target class. If the second customer information is not determined to be the target class, the comparison unit 150 sequentially repeats from the process of selecting a specific item, but may add the number of specific items one by one each time it is repeated. In addition, when it is determined that the second customer information is the target class, the comparison unit 150 may detect and provide changed items of the second customer information through the description unit 160.

The description unit 160 determines that the second customer information is a target class, and may visualize and provide the changed item of the second customer information. In addition, the description unit 160 is a device for displaying the evaluation reason of the machine learning-based credit rating classification model, and as shown in FIGS. 3 and 4 to be described later, it is also possible to display the customer's credit information and whether a loan is possible. Do. Due to this, the user can check from one customer information what distinguishable characteristics (items of changed credit information) between the original class (results according to the first customer information) and the target class (results according to the second customer information). have.

FIG. 3 is a diagram showing credit information of customer A according to an embodiment of the present invention and a result of a credit rating classification model therefor. 4 is a diagram illustrating an item to be changed among credit information in order to derive a target class of credit information of customer A according to a classification model according to an embodiment of the present invention.

Exemplarily, as shown in FIG. 3, the classification result of the credit rating classification model 102 for the first customer information 10 will be derived as a loan impossible state (loan available: 40%, loan not allowed: 60%). I can. At this time, the first customer information is the original customer A's credit information including annual salary: 30 million won, absence: 20 million won, loan amount: 70 million won, and five credit transactions.

For example, referring to FIG. 4, the processor 101 is a specific item that affects the change to the target class (loanable) when the first customer information 10 and the target class are input as loanable, and credit information Select the annual salary and debt, and change the value of each item (annual salary: 60 million won, debt: 10 million won) as a result of deriving the target class (loan available) with the second customer information (20) (loan available: 90% ) Is calculated, and when it is determined that the second customer information 20 is a target class, the changed items (annual salary, debt) of the second customer information 20 may be provided.

That is, the user may grasp the reason for evaluation of the credit rating classification model through the changed item of the second customer information 20. In the case of customer A, in the past, the annual salary was 30 million won and the amount of debt was 20 million won (first customer information), which prevented it from being judged as the target class (loan available). However, if the annual salary of customer A's credit information is raised to 60 million won and the amount of debt is reduced by 10 million won, it means that the result of the credit rating classification model 102 determining that a loan is possible changes to 90%.

As a result, the lending company can present the conditions for lending to the customer, motivating them to increase the loan performance, and the effect that the customer can accurately know which part of their credit information is insufficient for lending. There is.

Hereinafter, a description of a configuration that performs the same function among the configurations illustrated in FIGS. 1 to 4 will be omitted.

5 is a flowchart illustrating a method of explaining an evaluation reason for a machine learning-based credit rating classification model according to an embodiment of the present invention.

Referring to FIG. 5, a method of explaining the reason for evaluation of a machine learning-based credit rating classification model according to an embodiment of the present invention is to evaluate first customer information including credit information composed of a plurality of items and a credit rating classification model. Step of receiving a target class to be derived as a result (S110), selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information (S120), selected specific Step of changing the value of the item and generating second customer information including the changed item value (S130), calculating the result probability of deriving the target class as a result of the evaluation of the credit rating classification model for the input of the second customer information Step S140 and comparing whether the second customer information is determined as a target class (S150).

If the second customer information is not determined to be the target class in step S150, it is sequentially repeated from the process of selecting a specific item, but each time it is repeated, the number of specific items is added one by one, and the second customer information is the target. When it is determined as the class, the step of detecting and providing changed items of the second customer information may be further included.

In step S120, specific items are selected in the order in which the absolute value of the derivative for the second customer information is large, based on the result probability of deriving the target class, but indexing can be performed in the order in which the absolute value of the derivative is as large as the number of items. have.

In step S130, the value of the first customer information is changed to the item value of the second customer information for the indexed item, but the changed item value is determined by minimizing the value of the preset loss function based on the first and second customer information. I can.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include a computer storage medium. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

Although the methods and systems of the present invention have been described in connection with specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. do.

100: Device explaining the evaluation reason of the machine learning-based credit rating classification model
101: processor
102: credit rating classification model
103: memory
104: database
110: input unit
120: item selection unit
130: item value change unit
140: classification model result unit
150: comparison unit
160: description

Claims (8)

In a method for explaining the evaluation reason of a machine learning-based credit rating classification model performed by a computer device,
(a) receiving first customer information including credit information composed of a plurality of items and a target class to be derived as an evaluation result of the credit rating classification model;
(b) selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information;
(c) changing the value of the selected specific item and generating second customer information including the changed item value;
(d) calculating a result probability of deriving the target class as an evaluation result of the credit rating classification model for the input of the second customer information; And
(e) comparing whether the second customer information is determined to be the target class; and a method for explaining the evaluation reason of the machine learning-based credit rating classification model.
The method of claim 1,
In step (e)
When the second customer information is not determined to be the target class, it is sequentially repeated from the process of selecting the specific item, but each time it is repeated, the number of the specific item is added one by one,
If the second customer information is determined to be the target class, the method further comprising the step of detecting and providing changed items of the second customer information.
The method of claim 2,
The step (b) is
Selecting the specific items in the order of the largest absolute value of the derivative for the second customer information based on the result probability of the target class being derived,
The method of explaining the evaluation reason for a machine learning-based credit rating classification model, wherein indexing is performed in an order in which the absolute value of the derivative is as large as the number of the items.
The method of claim 3,
Step (c) is
For the indexed item, the value of the first customer information is changed to the item value of the second customer information, but the changed item value by minimizing the value of a preset loss function based on the first and second customer information A method of explaining the evaluation reason of the machine learning-based credit rating classification model.
In the device for explaining the evaluation reason of the machine learning-based credit rating classification model,
A memory storing a program for performing a method of explaining the evaluation reason of the machine learning-based credit rating classification model; And
Includes; a processor that executes the program,
The processor, according to the execution of the program,
Receive first customer information including credit information composed of a plurality of items and a target class to be derived as an evaluation result of the credit rating classification model,
Selecting a specific item that affects changing the class determined by the credit rating classification model to the target class for the first customer information,
Change the value of the selected specific item, create second customer information including the changed item value,
Calculating a result probability of deriving the target class as a result of evaluation of the credit rating classification model for the input of the second customer information,
An apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model for comparing whether the second customer information is determined to be the target class.
The method of claim 5,
The processor is
When the second customer information is not determined to be the target class, it is sequentially repeated from the process of selecting the specific item, but each time it is repeated, the number of the specific item is added one by one,
When the second customer information is determined to be the target class, changed items of the second customer information are detected and provided.
The method of claim 6,
The processor is
Selecting the specific items in the order of the largest absolute value of the derivative for the second customer information based on the result probability of the target class being derived,
The apparatus for explaining an evaluation reason for a machine learning-based credit rating classification model, wherein indexing is performed in an order in which the absolute value of the derivative is as large as the number of the items.
The method of claim 7,
The processor is
For the indexed item, the value of the first customer information is changed to the item value of the second customer information, but the changed item value by minimizing the value of a preset loss function based on the first and second customer information The device for explaining the evaluation reason of the machine learning-based credit rating classification model to determine the.

Images