Credit card risk category determination method, apparatus, device, medium, and program product

Abstract

Embodiments of the present application provide a credit card risk category determination method, device, equipment, medium and program product. The method comprises: obtaining a sample set, the sample set comprising data information of n credit cards, the n credit cards being marked with respective risk categories, the data information of the n credit cards corresponding to m data categories, any credit card comprising data information under the m data categories, m and n being integers greater than 1; determining whether each data category in the sample set has relevance with the risk category by a fuzzy rough set model, and determining a data category having relevance with the risk category as a target data category, the fuzzy rough set model being a model for determining relevance according to a fuzzy approximation operator; in the case where at least one target data category is determined, inputting data information of a to-be-tested credit card under the at least one target data category into a classification model for classification prediction, and determining a risk category of the to-be-tested credit card. The method can improve determination efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a credit card risk category determination method, device, equipment, medium and program product. BACKGROUND

[0002] With the development of credit card business, the number of users using credit cards is increasing. In order to better adapt to the card use needs of users, the use scenarios of credit cards are also developing in a diversified manner. For example, credit cards can be applied to online payment, offline payment, or online transfer, etc.

[0003] Some users may have the behavior of using credit cards irregularly when using credit cards. For example, forging credit cards, using others' credit cards, or malicious overdraft, etc. Therefore, it is necessary for the card issuing institution of the credit card to determine whether there is a card use risk in the use process of the credit card, so as to identify the credit card that may have risks in time and facilitate card use management.

[0004] In related technologies, when determining whether a credit card has risks, the data information of the credit card user collected is input into a pre-trained prediction model for processing to obtain a prediction result representing whether the credit card has risks. When predicting in this way, the required operation amount is large and the time consumption is large, resulting in low efficiency in determining the risk category. If the prediction model is changed to reduce the operation amount in order to improve the efficiency, the accuracy of the prediction result may be reduced. Therefore, there is a problem in the prior art that the determination efficiency cannot be improved while the prediction accuracy is taken into account. SUMMARY

[0005] The embodiments of the present application provide a credit card risk category determination method, device, equipment, medium and program product to achieve the effect of improving the determination efficiency while taking into account the accuracy of determining the credit card risk category.

[0006] In a first aspect, the embodiments of the present application provide a credit card risk category determination method, comprising: obtaining a sample set, the sample set comprising data information of n credit cards, the n credit cards being marked with respective risk categories, the data information of the n credit cards corresponding to m data categories, any one of the credit cards comprising data information under the m data categories, the m and the n being integers greater than 1; determining whether each data category in the sample set has relevance with the risk category by a fuzzy rough set model, and determining a data category having relevance with the risk category as a target data category, the fuzzy rough set model being a model for determining relevance according to a fuzzy approximation operator; in the case where at least one target data category is determined, inputting data information of a to-be-tested credit card under the at least one target data category into a classification model for classification prediction to determine the risk category of the to-be-tested credit card.

[0007] In a possible implementation, the determining, by the fuzzy rough set model, whether each data category in the sample set is associated with the risk category, and determining the data category associated with the risk category as the target data category, comprises: selecting one or more data categories from each data category in the sample set to obtain a first category subset; adding one or more data categories not selected to the first category subset or removing one or more data categories from the first category subset to obtain a second category subset; calculating, by the fuzzy rough set model, the fuzzy dependency of the data information of each credit card in the sample set under the first category subset on the risk category to obtain a first fuzzy dependency, and calculating the fuzzy dependency of the data information of each credit card in the sample set under the second category subset on the risk category to obtain a second fuzzy dependency; determining, according to the first fuzzy dependency and the second fuzzy dependency, whether the data category added or removed when obtaining the second category subset is associated with the risk category, and determining the data category associated with the risk category as the target data category.

[0008] In a possible implementation, the fuzzy dependency calculated by the fuzzy rough set model satisfies the following formula: wherein, U represents the sample set; x i represents the data information of the i th credit card in the sample set; L represents each risk category; M represents a distance metric matrix; B represents a category subset corresponding to the calculated fuzzy dependency; represents the fuzzy lower approximation operator of x i based on B on each risk category; represents a fuzzy positive region; | | represents a modulo operation; represents the fuzzy dependency.

[0009] In a possible implementation, the distance metric matrix is determined after iterative learning by the large margin nearest neighbor algorithm according to the data information in the sample set.

[0010] In a possible implementation, the distance metric matrix is a Markov matrix, and the Markov matrix satisfies the following formula: wherein, M represents the Markov matrix; x j represents the data information of the j th credit card in the sample set; D M (x i , x j ) represents the Markov distance between x i and x j ; and the loss function used in the iterative learning by the large margin nearest neighbor algorithm satisfies the following formula: ε(M) = (1-μ)εpull (M) + με(M) push (M) + με(M) wherein, ε(M) represents a value of the loss function; μ represents a preset coefficient; SKNN(x i ) represents a set of k nearest neighbors with x i having the same risk category, k being an arbitrary positive integer; x l represents data information of the lth credit card belonging to the set of KNN(x i ) and not belonging to the set of SKNN(x i ), KNN(x i ) being a set of k nearest neighbors most proximate to x i , k being an arbitrary positive integer; [Z] + =max(Z, 0) represents a value in the brackets compared with 0 and taking the maximum; y il represents a value corresponding to x i and x j having the same or different risk categories, the value being 1 when x i and x j have the same risk category, and the value being 0 when x i and x j have different risk categories.

[0011] In a possible implementation, the inputting of the data information of the credit card to be tested under the at least one target data category into the classification model for classification prediction and the determination of the risk category of the credit card to be tested comprise: inputting the data information of the credit card to be tested under the at least one target data category into the classification model to calculate a function value of the credit card to be tested corresponding to a target function; and performing classification prediction according to the function value and a preset determination threshold to determine the risk category of the credit card to be tested; wherein the target function is used to calculate a probability of the risk category of the credit card to be tested, the target function is obtained by presetting a support vector distance according to the risk category of the credit card in the sample set and a high-dimensional discriminant function, the high-dimensional discriminant function is obtained by transforming a linear discriminant function through a Gaussian kernel, and the linear discriminant function is determined by using a support vector machine model under the at least one target data category according to the data information in the sample set.

[0012] In a second aspect, the embodiments of the present application provide a credit card risk category determination apparatus, comprising: an acquisition module configured to acquire a sample set, wherein the sample set comprises data information of n credit cards, each of the n credit cards is marked with a respective risk category, the data information of the n credit cards corresponds to m data categories, each of the credit cards comprises data information under the m data categories, and the m and the n are integers greater than 1; a determination module configured to determine, by a fuzzy rough set model, whether each data category in the sample set has relevance with the risk category, and determine a data category having relevance with the risk category as a target data category, wherein the fuzzy rough set model is a model for determining relevance according to a fuzzy approximation operator; and a processing module configured to, in a case where at least one target data category is determined, input data information of a to-be-tested credit card under the at least one target data category into a classification model for classification prediction, and determine a risk category of the to-be-tested credit card.

[0013] In a possible implementation, the determination module is specifically configured to: select one or more data categories from each data category in the sample set to obtain a first category subset; add one or more data categories not selected in the first category subset or remove one or more data categories in the first category subset to obtain a second category subset; calculate, by the fuzzy rough set model, a fuzzy dependency degree of the data information of each credit card in the sample set under the first category subset with respect to the risk category to obtain a first fuzzy dependency degree, and calculate a fuzzy dependency degree of the data information of each credit card in the sample set under the second category subset with respect to the risk category to obtain a second fuzzy dependency degree; determine, according to the first fuzzy dependency degree and the second fuzzy dependency degree, whether the data category added or removed when obtaining the second category subset has relevance with the risk category, and determine a data category having relevance with the risk category as a target data category.

[0014] In a possible implementation, the fuzzy dependency degree is calculated by the fuzzy rough set model to satisfy the following formula: wherein U represents the sample set; x i represents data information of an i th credit card in the sample set; L represents each risk category; M represents a distance metric matrix; and B represents a category subset corresponding to the calculated fuzzy dependency degree. represents a fuzzy lower approximation operator of x i based on B with respect to each risk category. represents a fuzzy positive region; and | | represents a modulo operation. represents the fuzzy dependency degree.

[0015] In a possible implementation, the distance metric matrix is determined according to the data information in the sample set by iterative learning through a large-margin nearest neighbor algorithm.

[0016] In a possible implementation, the distance metric matrix is a Markov matrix, which satisfies the following formula: D M (x i ,x j )=(x i -x j ) T M(x i -x j ); wherein M represents the Markov matrix; x j represents the data information of the jth credit card in the sample set; D M (x i ,x j ) represents the Markov distance between x i and x j ; and the loss function used in the iterative learning by the large margin nearest neighbor algorithm satisfies the following formula: ε(M)=(1-μ)ε pull (M)+με push (M); wherein ε(M) represents the value of the loss function; μ represents a preset coefficient; SKNN(x i ) represents a set of k nearest neighbors having the same risk category as x i ; x l represents the data information of the lth credit card belonging to the set KNN(x i ) but not belonging to the set SKNN(x i ), KNN(x i ) being a set of k nearest neighbors closest to x i , k being any positive integer; [Z] + =max(Z,0) represents the maximum of the value in the brackets and 0; y il represents a value corresponding to the same or different risk category of x i and x j , the value being 1 when the risk categories of x i and x j are the same and 0 when the risk categories of x i and x j are different.

[0017] In a possible implementation, the processing module is specifically configured to: input data information of the to-be-tested credit card in the at least one target data category into a classification model to calculate a function value of the to-be-tested credit card corresponding to a target function; and perform classification prediction according to the function value and a preset judgment threshold to determine the risk category of the to-be-tested credit card, wherein the target function is used to calculate a probability of a risk category corresponding to the to-be-tested credit card, the target function is obtained according to risk categories of credit cards in the sample set and a high-dimensional discriminant function through a support vector distance presetting, the high-dimensional discriminant function is obtained by transforming a linear discriminant function through a Gaussian kernel, and the linear discriminant function is determined according to data information in the sample set under the at least one target data category by using a support vector machine model.

[0018] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a memory and a processor; the memory stores computer execution instructions; and the processor executes the computer execution instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect.

[0019] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the first aspect and / or various possible implementation manners of the first aspect.

[0020] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the first aspect and / or various possible implementation manners of the first aspect.

[0021] The credit card risk category determination method, device, equipment, medium and program product provided by the embodiment of the application, the method obtains a sample set through the data information of the credit card with the marked risk category, determines which data categories in the sample set have relevance with the risk category through the fuzzy rough set model, and determines the target data categories having relevance with the risk category. Since the fuzzy rough set model can determine the relevance according to the fuzzy approximation operator, the data categories that are helpful to determine the risk category can be accurately screened out from the numerous data categories, the purpose of filtering out the data categories irrelevant to the determination of the risk category can be achieved, the number of data categories can be reduced, and then the information amount of the data information can be reduced. In the case of determining at least one target data category, the data information of the credit card to be tested under at least one target data category is input into the classification model for classification prediction. Compared with the case that the data information of the credit card to be tested under all data categories is input, the method can reduce the data amount input into the classification model, reduce the operation amount during the operation of the classification model, and then the time required for operation can be shortened, and the efficiency of classification prediction can be improved. Since the filtered data categories are irrelevant data categories, the irrelevant data information is excluded, the accuracy of the prediction result is high, and therefore, based on the method, the effect of improving the determination efficiency while ensuring the accuracy of determining the risk category of the credit card can be achieved. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0023] Figure 1 A schematic diagram of an application scenario provided by the embodiment of the application;

[0024] Figure 2 A flowchart of the credit card risk category determination method provided by the embodiment of the application Figure One ;

[0025] Figure 3 A flowchart of the credit card risk category determination method provided by the embodiment of the application Figure Two ;

[0026] Figure 4 A flowchart of the credit card risk category determination method provided by the embodiment of the application Figure Three ;

[0027] Figure 5 A structural diagram of the credit card risk category determination device provided by the embodiment of the application;

[0028] Figure 6 A structural diagram of an electronic device provided by the embodiment of the application.

[0029] The specific embodiments of the application have been shown by way of example in the above figures, and will be described in greater detail below. These figures and this written description are not intended to limit the scope of the inventive concept in any way, but rather to illustrate the inventive concept by reference to specific embodiments. DETAILED DESCRIPTION

[0030] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The same reference numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments are not meant to represent all embodiments consistent with the application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the application as detailed in the appended claims.

[0031] In the embodiments of the present application, the same or similar items or elements with substantially the same function and role are distinguished by using "first", "second", etc. For example, the first electronic device and the second electronic device are only used to distinguish different electronic devices, and do not limit the order. Those skilled in the art can understand that "first", "second", etc. do not limit the number and execution order, and "first", "second", etc. do not necessarily mean different.

[0032] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design described herein as "exemplary" or "for example" should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the exemplary or example embodiments are presented to best explain the inventive concept in detail.

[0033] In the embodiments of the present application, "at least one" means one or more, and "multiple" means two or more. The association relationship between the associated objects is described by "and / or", which means that there can be three relationships, for example, A and / or B, which can represent the following three cases: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after it.

[0034] In the technical solutions of the embodiments of the present application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the technical solutions comply with relevant laws and regulations and do not violate public order and good customs.

[0035] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

[0036] With the rapid development of the credit card industry, banks and other credit card issuing institutions are constantly launching new credit card products to expand market share, broadening card issuance channels, actively cooperating with Internet companies, using big data and Internet technology to deeply explore user needs, and exploring consumption patterns. In terms of product innovation, credit card services are becoming more personalized and specialized to meet the needs of different consumer groups. At the same time, card issuing institutions are also improving service quality and focusing on humanized services to enhance user stickiness and satisfaction.

[0037] With the popularity of mobile payments, the use of credit cards is becoming more diversified, and consumers can make online and offline payments more conveniently. However, with the growth of credit card business, problems such as credit card fraud and other illegal use of credit cards have become increasingly prominent. Credit card fraud includes behaviors such as forging credit cards, using others' credit cards, and malicious overdrafts, which can bring risks and losses to banks and consumers. To address these challenges, banks and other financial institutions are taking various measures to strengthen risk management, improve credit card security, and promote financial knowledge education.

[0038] In this process, banks and credit card companies are also using financial technology such as machine learning algorithms to predict and prevent credit card fraud, reducing losses. The use of these technologies not only improves the security of credit card business, but also provides banks with more accurate risk assessment and decision support.

[0039] Figure 1 An application scenario provided by an embodiment of the present application is shown in a schematic diagram as shown in Figure 1 Among the many credit cards, there may be risky credit cards with credit card fraud behaviors. The credit card management party can collect data information of each of the plurality of credit cards to be tested, and construct a data table including the plurality of credit card data information. Through machine algorithms such as prediction models, the data table is processed to identify risky credit cards and obtain prediction results of the risk categories of each credit card. Further, corresponding management measures can be taken for risky credit cards to better manage card use behaviors.

[0040] In the related art, there is a method for predicting credit card fraud risk, which first collects cardholder data information, including basic data information, financial operation behavior data and credit card transaction data information. The basic data information, financial operation behavior data and credit card transaction data information of the cardholder are embedded and represented by a feedforward neural network model, a convolutional neural network model and a bidirectional long short-term memory network model (Bi-LSTM) respectively. A sample feature representation vector is obtained by splicing. A generative adversarial network model is constructed and trained, a generator model is used to generate simulated fraud samples, a new training set is generated by merging, and multiple detection base models are constructed and trained. The trained multiple detection base models and the discriminator model are combined to obtain a credit card fraud detection integrated model. When a new credit card transaction occurs, the credit card fraud detection integrated model determines the credit card risk category as having fraud risk or no fraud risk.

[0041] Using a feedforward neural network model, a convolutional neural network model and a Bi-LSTM model for feature embedding representation, and combining a generative adversarial network model, such a model structure is relatively complex, which may cause low computational efficiency, especially when processing large-scale data sets, the demand for computing resources may become a limiting factor.

[0042] It can be seen that in this method, the model used for determining credit card fraud risk is obtained by combining multiple relatively complex models. Such a model has high complexity, usually including complex model structure and large number of model parameters. The model needs to consume a large amount of computing resources and the operation time is usually high when performing inference operation. In addition, the collected data information includes basic data information, financial operation behavior data and credit card transaction data information. These data information corresponds to multiple data categories, which can be understood as high dimension of data information. If these high-dimensional data information is not effectively reduced, the amount of data input into the model will be large, which will also increase the operation time and affect the efficiency. Furthermore, as the number of credit card users increases, the number of credit cards that need to be detected for risk increases. If a large amount of credit card data information is processed in batches, the efficiency of credit card risk determination will be further reduced.

[0043] The inventors have found that if the conventional improvement idea is followed, the structure or parameters of the model are usually improved to simplify the model and reduce the complexity of the model. In this way, the operation steps or operation difficulty of the model during inference can be reduced, the operation time consumption can be reduced, and the efficiency of risk category determination can be improved.

[0044] However, if the structure or parameters of the simplified model are reduced, the accuracy of data processing may be affected. For example, if the number of convolution layers in the convolutional neural network model is reduced, although the amount of convolution operation can be reduced, the representation of the output data of the convolutional neural network model may be reduced, which may eventually reduce the accuracy of the prediction result. For example, a credit card with fraud risk may be predicted as a credit card without fraud risk, or a credit card without fraud risk may be predicted as a credit card with fraud risk.

[0045] Therefore, in order to improve the efficiency of determining the risk category of the credit card while ensuring the accuracy of determining the risk category of the credit card, an embodiment of the present application provides a method for determining the risk category of a credit card. The method can filter a plurality of data categories corresponding to the data information of the credit card based on a sample set, and determine a data category associated with the determination of the risk category of the credit card as a target data category. In this way, the dimension of the data information is reduced, and the amount of data information is reduced. Moreover, the data categories that are filtered out are data categories that are not associated with the risk category. Therefore, when predicting, the amount of data information is reduced, and irrelevant data information is excluded. Therefore, the accuracy of the prediction is not reduced, and the accuracy of the prediction may be improved due to the removal of irrelevant data information. In addition, based on the method, the amount of calculation during the inference of the model can be reduced without changing the structure or parameters of the model, and the prediction efficiency of the model can be improved. Therefore, the accuracy and efficiency of determining the risk category of the credit card can be considered by using the method, and the efficiency of determining the risk category of the credit card can be effectively improved.

[0046] The technical solutions of the present application will be described in detail below with specific embodiments. The specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0047] Figure 2 Flowchart of the method for determining the risk category of a credit card provided by the embodiments of the present application Figure One The execution subject of the method can be an electronic device with corresponding data storage capability and computing capability, for example, a computer, a server or a server cluster. As shown in Figure 2 The method comprises the following steps.

[0048] S201, obtaining a sample set, the sample set comprising data information of n credit cards, the n credit cards each being marked with a respective risk category, the data information of the n credit cards corresponding to m data categories, each credit card comprising data information under the m data categories, and m and n each being an integer greater than 1.

[0049] Exemplarily, the data information of the credit card can be any data information associated with the credit card. The data information can include basic information of a cardholder of the credit card, transaction information of the credit card, card information of the credit card, and financial operation behavior information of the cardholder of the credit card, and the like. For example, the card number, the validity period, the account balance, the asset evaluation information, the security code, the issuing institution information, the credit limit, the billing address, the transaction history data, and the associated bank card information of the cardholder of the credit card, and the like.

[0050] The data category can be understood as a category corresponding to specific data in the data information. For example, the card number of the credit card is a data category, and the data information of the first credit card in the n credit cards under the data category is the specific card number of the credit card, such as “XXXX XXXX XXXX” and the like. For another example, the account balance of the credit card is a data category, and the data information of the first credit card in the n credit cards under the data category is the specific balance value of the credit card, such as “200.01” and the like.

[0051] It should be understood that the data category can also be understood as the dimension of the data information. For example, m is 100, that is, there are 100 data categories in the sample set, and the dimension of the data information is 100. The data information of any credit card can be represented as a 100-dimensional vector. Assuming that the first dimension is the card number of the credit card, the second dimension is the account balance of the credit card, and the like, the data information of the credit card can be represented as (XXXX XXXX XXXX, 200.01, …) by a multi-dimensional vector.

[0052] For ease of description and understanding, the data information corresponding to any credit card in the sample set can be understood as a sample, and the risk category of the sample can be understood as the label of the sample. When the data information is processed, the sample can be represented by a multi-dimensional vector, for example, a sample is (XXXX XXXX XXXX, 200.01, …), and the multi-dimensional vector can be processed when the sample is processed.

[0053] The data information of the n credit cards is obtained by data acquisition. For example, the data information is obtained from a database storing the data information of the plurality of credit cards by data query. When the data information is obtained, the risk category representing that the corresponding credit card is risky or non-risky can also be obtained. For example, 0 can represent non-risky, and 1 can represent risky; or -1 can represent non-risky, and 1 can represent risky. The risk category can also be three risk categories of low risk, medium risk, and high risk, or other risk categories. The method of the embodiments of the present application can be understood as a two-classification task of two classification categories, or a multi-classification task of multiple classification categories.

[0054] Alternatively, the n credit cards obtained can be marked with risk categories by manual marking and / or machine marking, etc. For example, after obtaining the data information of 1000 credit cards, the risk categories of the 1000 credit cards are predicted by using an existing classification model, and the risk categories of the credit cards are obtained. Then, the risk categories are verified and adjusted by manual verification, and then 1000 credit cards marked with risk categories and their data information are obtained, and a sample set is obtained. The classification model can be a machine model obtained by training, for example, a model obtained by supervised training of an initial deep learning model for classification prediction.

[0055] In S202, whether each data category in the sample set has relevance with the risk category is determined by using a fuzzy rough set model, and a data category having relevance with the risk category is determined as a target data category. The fuzzy rough set model is a model for determining relevance according to a fuzzy approximation operator.

[0056] For example, the fuzzy rough set model can be a mathematical calculation model for determining relevance according to a fuzzy approximation operator. The fuzzy approximation operator in the fuzzy rough set model can be a fuzzy lower approximation operator or a fuzzy upper approximation operator. The fuzzy rough set model can be constructed based on a preset distance measurement method.

[0057] For example, the distance measurement method is preset to calculate the distance between samples in the Euclidean distance. Then, the fuzzy equivalence relation can be defined in advance under the distance measurement of the Euclidean distance. Based on the fuzzy equivalence relation, the fuzzy lower approximation operator or the fuzzy upper approximation operator can be defined. Taking the fuzzy lower approximation operator as an example, the fuzzy positive region can be defined based on the fuzzy lower approximation operator. The fuzzy lower approximation operator and the fuzzy positive region are the constructed fuzzy rough set model. Based on the fuzzy positive region in the fuzzy rough set model, the data information of the plurality of credit cards in the sample set can be calculated. Whether each data category corresponding to the data information has relevance with the risk category can be determined according to the calculated value.

[0058] For example, the data category having relevance with the risk category can be understood as a data category that can affect the determination of the risk category of the credit card. The data category can be determined as the target data category. For example, after calculating by using the fuzzy rough set model, if it is determined that the transaction amount of the credit card can affect the determination of the risk category of the credit card, the transaction amount of the credit card can be determined as a target data category.

[0059] Exemplarily, a fuzzy rough set model based on metric learning (ML-FRS) can be constructed. Since the distance metric in the fuzzy rough set model is obtained after metric learning, the distance metric can adjust the parameters of the metric according to the distribution characteristics of each sample in the sample set during learning, so that the distance metric obtained after learning can more accurately measure the distance between any two samples, thereby improving the ability of the fuzzy rough set model to determine relevance, making the determined target data category have a greater probability of being associated with the determined risk category, and improving the reliability of the target data category.

[0060] In a possible implementation, a distance metric matrix can be determined by performing iterative learning on a large margin nearest neighbor (LMNN) algorithm according to data information in a sample set. Based on the obtained distance metric matrix, a fuzzy rough set model based on metric learning can be constructed.

[0061] Exemplarily, in the sample set, the data information of any credit card can be a sample for learning a distance metric matrix. The LMNN algorithm minimizes the distance between a sample and its nearest neighbor sample, while maximizing the distance between samples with different labels, thereby establishing a more appropriate boundary and learning a distance metric matrix. The label is the risk category corresponding to the sample. For any sample, the nearest neighbor sample of the sample can be understood as a target sample, the label of the target sample is the same as the label of the any sample, and the distance is the smallest, which can be the minimum value of the distance calculated under any distance metric such as Euclidean distance, cosine similarity or Mahalanobis distance.

[0062] Exemplarily, when determining the risk category of a credit card, the collected data categories are usually more, and each data information of a credit card corresponds to multiple data categories, which can be understood as a higher vector dimension corresponding to the sample, and the number of samples of different risk categories is relatively unbalanced, for example, the number of credit cards marked as risky and the number of credit cards marked as non-risky have a large difference, so the sample set of the credit card has the characteristic of uneven distribution of each sample.

[0063] The LMNN algorithm mainly learns a linear transformation, so that in the transformed space, the probability that the k-nearest neighbors of the data points belong to the same category is maximized, and the distance between data points of different categories is maximized. Since the Mahalanobis distance can be represented by a linear transformation and can capture the covariance structure of the data, the Mahalanobis distance is usually used as a metric in LMNN. However, the Mahalanobis distance is not suitable for the distribution characteristics of the samples in the sample set of the credit card.

[0064] For the unevenly distributed or sparse sample set, the Markov distance measure can better deal with the sparse data information by focusing on the state transition rather than the frequency of individual state occurrence, so a Markov matrix can be learned through the LMNN algorithm, and the learned Markov matrix can adapt to the distribution characteristics of the samples in the credit card sample set, and then the distance between any two samples can be more accurately measured, which helps to improve the accuracy when using the fuzzy rough set model to calculate the model under the sample set.

[0065] In a possible implementation, the distance measure matrix is a Markov matrix, and the Markov matrix satisfies the following formula (1):

[0066]

[0067] Wherein, M represents the Markov matrix; x j represents the data information of the jth credit card in the sample set; D M (x i ,x j ) represents the Markov distance between x i and x j .

[0068] When the large-margin nearest neighbor algorithm is used for iterative learning, the loss function used satisfies the following formula (2) to formula (4):

[0069] ε(M)=(1-μ)ε pull (M)+με push (M) (2).

[0070]

[0071] Wherein, ε(M) represents the value of the loss function; μ represents a preset coefficient; SKNN(x i ) represents a set of k nearest neighbors with the same risk category as x i , and k is any positive integer; x l represents the data information of the lth credit card belonging to the KNN(x i ) set and not belonging to the SKNN(x i ) set, and KNN(x i ) is a set of k nearest neighbors closest to x i , and k is any positive integer; [Z] + =max(Z,0) represents the value in the brackets compared with 0 and taking the maximum; y il represents the value corresponding to the same or different risk categories of x i and x j . i j ​When the risk categories are the same, the value is 1, x i With x j The value is 0 when the risk categories are different.

[0072] Specifically, an iterative learning process is used to determine the appropriate mapping matrix M representing the original data space in the form of Markov distances, thus obtaining the Markov matrix M. Assume the sample set x = {x1, x2, ..., x...} n}, where x n This represents the data information of the nth credit card in the sample set, i.e., the nth sample. The goal of iterative learning is to learn a Markov distance metric based on n samples with known labels using the LMNN algorithm, as shown in formula (1) above.

[0073] A Markov matrix M can be a positive semidefinite symmetric matrix, which can be represented as: Here, Q represents one matrix and P represents another matrix. The Markov matrix M can be represented by the product of the transpose of matrix Q and matrix P. Matrix Q and matrix P can be obtained through learning.

[0074] For example, when constructing the loss function, for a risk category L i Sample x i The sample x i The k-nearest neighbors of a target are defined as those that are closest to the target x in the sample set. i Having the same risk category L i And the k nearest samples, then we obtain the sample x. i The set of k nearest neighbors of SKNN(x) i ), where k can be any positive integer, for example, k can be 10. The set SKNN(x i The sample in ) can be understood as sample x i The target sample. If x j It is x i The target sample can be represented as x. j ∈SKNN(x i ).

[0075] For example, a y can be constructed ij A binary matrix ∈ {0,1} is used to represent the relationship between whether all samples in a sample set have the same risk category. For example, in this binary matrix, the rows and columns are n samples arranged in the same order, and the elements in the binary matrix are y. ij y ij =1 indicates that sample x i and sample x j The risk categories are the same, that is, the labels are the same, y ij =0 indicates that sample x i and sample xj The risk categories of the two are different, i.e., the labels are different.

[0076] When the iterative learning is performed by the large margin nearest neighbor algorithm, the distance metric suitable for the sample set can be learned by the loss function shown in the above formula (2) to formula (4). Formula (3) and formula (4) can be substituted into formula (2), and can be understood as two components of formula (2). The part shown in formula (3) ε pull (M) has the effect of punishing the large distance between the sample and its target sample; and the part shown in formula (4) ε push (M) has the effect of punishing the small distance between the input samples without the same label.

[0077] Exemplarily, a convergence condition for determining whether the iterative learning is ended can be set. When the iterative learning is performed, the parameters in the Markov matrix M are adjusted, for example, the Q matrix and the P matrix are adjusted, so that the value of the loss function ε(M) calculated after the adjustment meets the convergence condition, it can be determined that the iterative learning is ended, and at this time, the Markov matrix M learned by the adjusted Q matrix and the P matrix can be obtained.

[0078] It should be noted that μ represents a preset coefficient, which can be understood as a coefficient for weight distribution of the two parts of formula (3) and formula (4), and the value of μ can be preset by an empirical value or a calculated value. + =max(Z,0), where Z represents the part in the brackets in formula (4), i.e., Z=1+D M (x i ,x j )-D M (x i ,x l ), after the value of Z is calculated, the value of Z is compared with 0 in size, if the value of Z is greater than 0, the value of this part is taken as the value of Z, if the value of Z is less than or equal to 0, the value of this part is taken as 0. l x i represents the data information of the lth credit card belonging to the KNN(x i ) set and not belonging to the SKNN(x l ) set, which can also be expressed as x i ∈KNN(x i )\SKNN(x i ), where KNN(x i ) represents a set of k nearest neighbors of sample x i , and the samples in the set are the k samples closest to sample x i , and these nearest neighbor samples can be samples with the same label as sample x ilThe value of y ij can be quickly obtained in the binary matrix of ∈{0, 1}.

[0079] Based on this, the learning-based distance metric matrix can be obtained through several rounds of iterative learning by using the LMNN algorithm. Through the LMNN algorithm, clustering of high-dimensional data information can be applied, and then a distance metric matrix suitable for the distribution characteristics of credit card data information can be obtained. The distance metric matrix obtained by this method helps to improve the accuracy of fuzzy calculation of the sample set based on the fuzzy rough set model.

[0080] After obtaining a distance metric matrix based on metric learning, the distance between any two samples in the sample set can be calculated based on the distance metric matrix. This distance can also be understood as the similarity between the two samples. For example, if the calculated distance is smaller, the similarity between the two is higher, and if the calculated distance is larger, the similarity between the two is lower. Based on the learning distance metric matrix, a fuzzy rough set model based on metric learning can be constructed.

[0081] For example, an information decision system based on metric learning can be denoted as IDS = <U, A, L, Δ M >, where U represents the sample set, A represents a set composed of all data categories included in the sample set, L represents the risk categories, and Δ M represents the distance metric of the learning-based distance metric matrix M. For example, F(U) can be used to represent the fuzzy power set of the sample set U, which is defined based on Δ M on U, where B represents a subset of A. For any sample x i , δ M (x i ) can be used to represent the neighborhood of sample x i about the label L i under the distance metric matrix M.

[0082] For any subset B of A, i.e., for the fuzzy lower approximation operator can be defined as: The fuzzy upper approximation operator can be defined as: wherein, The fuzzy lower approximation operator can also be understood as fuzzy lower approximation or lower approximation, and the fuzzy upper approximation operator can also be understood as fuzzy upper approximation or upper approximation.

[0083] wherein a represents an element in the set A, which can be understood as a data category in A. x i(a) represents sample x i Data information under data category a, x j (a) represents sample x j Data information under data category a, It can be understood as sample x i and sample x j Each data information under data category a, the distance obtained after calculation based on distance measurement matrix M, which can represent the dissimilarity of sample x i and sample x j . Based on It can be obtained Where ∩ represents the intersection or minimum operation.

[0084] In the above formula, δ M (x i ) represents the neighborhood of sample x i based on M, if x j ∈ δ M (x i ), then D M (x i , x j ) ≤ δ, that is Where δ can be understood as a preset field radius. The fuzzy lower approximation operator of the sample can represent the degree of certainty of the sample belonging to the risk category it belongs to, and the fuzzy upper approximation operator of the sample can represent the degree of possibility of the sample belonging to the risk category it belongs to. Where the degree of certainty can be understood as the degree of affirmation, and the degree of possibility can be understood as the degree of possibility, therefore, in order to determine the relevance between the data category and the risk category, the fuzzy lower approximation operator can be used for related calculation when selecting the target data category for each data category.

[0085] Further, the fuzzy positive region can be defined. For It represents the lower approximation of x i ∈ U to each risk category based on B under distance measurement matrix M. The fuzzy positive region of B for all samples based on distance measurement matrix M Is defined as:

[0086] Based on the distance measurement matrix M obtained after metric learning, after the fuzzy lower approximation operator and the fuzzy positive region are defined, the fuzzy rough set model based on metric learning is constructed.

[0087] Figure 3 Flowchart of the credit card risk category determination method provided by the embodiment of the present application Figure Two The embodiment is a further refinement of step S202 based on the Figure 2 Embodiment, for example,Figure 3 As shown, step S202 can specifically include the following steps:

[0088] S301, selecting one or more data categories in each data category of the sample set to obtain a first category subset.

[0089] For example, all data categories included in the sample set can constitute a set of data categories, which can be understood as a universal set of data categories, and any data category can be an element in the universal set, for example, the universal set is represented by set A. The set A can be determined as the first category subset, or any subset of the set A can be determined as the first category subset.

[0090] S302, adding one or more data categories not selected in the first category subset or removing one or more data categories in the first category subset to obtain a second category subset.

[0091] For example, after determining the first category subset, one or more elements belonging to the set A but not belonging to the first category subset can be added to the first category subset to obtain the second category subset; or one or more elements in the first category subset can be removed to obtain the second category subset.

[0092] S303, calculating the fuzzy dependence degree of the data information of each credit card in the sample set under the first category subset on the risk category through the fuzzy rough set model to obtain a first fuzzy dependence degree, and calculating the fuzzy dependence degree of the data information of each credit card in the sample set under the second category subset on the risk category to obtain a second fuzzy dependence degree.

[0093] For the elements included in the first category subset and the second category subset, the data information corresponding to each sample in the sample set is calculated through the pre-constructed fuzzy rough set model, and the first fuzzy dependence degree corresponding to the first category subset is obtained, and the second fuzzy dependence degree corresponding to the second category subset is obtained. The fuzzy dependence degree can be understood as an indication information capable of representing the correlation between the data category and the risk category.

[0094] S304, determining whether the data category added or removed when obtaining the second category subset has correlation with the risk category according to the first fuzzy dependence degree and the second fuzzy dependence degree, and determining the data category having correlation with the risk category as a target data category.

[0095] For example, the first category subset and the second category subset can be two sets having an intersection, and both of the two sets belong to the set of all data categories included in the sample set.

[0096] For example, Wherein, A represents a set consisting of all data categories included in the sample set, a represents a data category in the set A, i.e. an element, B is a subset of A; B-a represents a set obtained by removing a from the set B, which is also a subset of A. The first category subset can be B, and the second category subset can be B-a; or, the first category subset can be B-a, and the second category subset can be B. a is a data category, and after calculating the fuzzy dependency degree, it can be determined whether the data category a is associated with the risk category.

[0097] For example, the association of two or more data categories can also be determined simultaneously, for example, removing data categories a and b from the first category subset B to obtain the second category subset B-a-b. After calculating the fuzzy dependency degree, it can be determined whether the combination of a and b is associated with the risk category.

[0098] In a possible implementation, the fuzzy dependency degree is calculated by the fuzzy rough set model, and the following formula (5) and formula (6) are satisfied:

[0099]

[0100] Wherein, U represents the sample set; x i represents the data information of the i th credit card in the sample set; L represents each risk category; M represents the distance measure matrix; B represents the category subset corresponding to the calculated fuzzy dependency degree; represents x i represents the fuzzy lower approximation operator of B for each risk category based on U; represents the fuzzy positive region; | | represents the modulus; represents the fuzzy dependency degree.

[0101] By substituting formula (6) into formula (5), the fuzzy dependency degree function for calculating the fuzzy dependency degree can be obtained |U| can be understood as the number of samples in the sample set obtained by taking the modulus of the sample set. The multi-label fuzzy dependency degree function reflects the approximation ability of the fuzzy equivalence class induced by the set including all data categories under the new measure to the label set.

[0102] The first fuzzy dependency degree is obtained by calculating the fuzzy dependency of the data information of each credit card in the sample set under the first category subset on the risk category through the fuzzy rough set model, and the second fuzzy dependency degree is obtained by calculating the fuzzy dependency of the data information of each credit card in the sample set under the second category subset on the risk category. This process can be understood as calculating the corresponding fuzzy dependency degrees of the first category subset and the second category subset respectively. Then, the relevance of the data categories not in the intersection of the two category subsets can be determined according to the fuzzy dependency degrees. If the data categories are regarded as features, this process can also be understood as judging redundant features and non-redundant features. Redundant features can be understood as data categories that have no relevance to the risk category, and non-redundant features can be understood as data categories that have relevance to the risk category.

[0103] When determining whether the data categories added or removed when obtaining the second category subset have relevance to the risk category according to the first fuzzy dependency degree and the second fuzzy dependency degree, the difference between the first fuzzy dependency degree and the second fuzzy dependency degree can be compared with a preset threshold to determine whether the data categories added or removed when obtaining the second category subset have relevance to the risk category.

[0104] For example, the preset threshold can be an arbitrary preset value, such as 0. If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree is greater than 0, it can be determined that the data categories added or removed when obtaining the second category subset have relevance to the risk category. If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree is less than or equal to 0, it can be determined that the data categories added or removed when obtaining the second category subset have no relevance to the risk category.

[0105] It should be understood that the preset threshold can also be a preset threshold range, such as (0, 1). If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree belongs to (0, 1), it can be determined that the data categories added or removed when obtaining the second category subset have relevance to the risk category. If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree does not belong to (0, 1), it can be determined that the data categories added or removed when obtaining the second category subset have no relevance to the risk category.

[0106] For example, when determining whether the data categories added or removed when obtaining the second category subset have relevance to the risk category according to the first fuzzy dependency degree and the second fuzzy dependency degree, it can be determined by whether the difference between the first fuzzy dependency degree and the second fuzzy dependency degree is equal to the preset threshold. For example, the preset threshold is 0. If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree is not equal to 0, it can be determined that the data categories added or removed when obtaining the second category subset have relevance to the risk category. If the difference between the first fuzzy dependency degree and the second fuzzy dependency degree is equal to 0, it can be determined that the data categories added or removed when obtaining the second category subset have no relevance to the risk category.

[0107] Specifically, let IDS = <U, A, L, Δ M be a decision system based on metric learning, Δ M be a distance metric matrix M, For any data class a, let a e A and a e B, if i.e. then a is indispensable to B. It can be understood that the first category subset B-a is obtained after removing the data class a from the first category subset B, at this time, if the first fuzzy dependence degree of the first category subset is not equal to the second fuzzy dependence degree of the second category subset, it indicates that the data information of each sample in the sample set under the data class a has an influence on the calculation of the approximation to the label set, then a is indispensable to B. Therefore, this data class a can affect the determination of the risk class, and further indicates that the data class a and the risk class have relevance. The data class a can be understood as a non-redundant feature. Similarly, when the difference between the first fuzzy dependence degree and the second fuzzy dependence degree is equal to 0, i.e. it indicates that the data information of each sample in the sample set under the data class a has no influence on the calculation of the approximation to the label set, then a is not indispensable to B, and then the data class a can be understood as a redundant feature.

[0108] The indispensable non-redundant feature can provide necessary information support for the risk class classification of the credit card. However, the redundant feature not only cannot provide effective information for the risk class classification, but also may reduce the performance of the classifier, and also prolongs the classification time. Therefore, by determining the relevance between the data class and the risk class through the embodiments of the present application, the screening of each data class is realized, and when the target data class is determined, the data class useful for the risk class classification can be retained, which is helpful for better dividing the sample space in the classification, and further constructing a machine learning model with better performance for classification prediction.

[0109] For example, the decision system IDS based on metric learning = <U, A, L, Δ M , the feature dependence of the data class can be defined, which can represent the strength of the relevance between the corresponding data class and the risk class. If B is a category subset of A, for any a e A-B, i.e. for any data class not in B in A, the feature dependence of a can be represented as: wherein Sig(a, B) represents the feature dependence of the data class a, represents the fuzzy dependence degree calculated under the category subset B u {a}, represents the fuzzy dependence degree calculated under the category subset B.

[0110] From the expression of feature dependency, if 0≤Sig(a, B)≤1, when Sig(a, B)=0, it indicates that a is a redundant feature for B, when Sig(a, B)≠0, it indicates that a is a non-redundant feature for B, and when Sig(a, B) is larger, it indicates that the importance of data category a is higher, that is, the relevance of data category a to the risk category is stronger, and the distinguishing ability is stronger, so when data category a is used to distinguish the risk category of the credit card, the accuracy of the determined risk category result is higher.

[0111] For example, for a given IDS=<U, A, L, Δ M >, if B can be understood as a subset of C, if condition (1) is met: and condition (2) is met: for any a∈B, there is When the number of data categories in set B is less than the number of data categories in set C, and the distinguishing ability of set C and set B is the same, set B can be called the reduction of set C. At this time, set B can be understood as a category subset including several target data categories with higher distinguishing ability, and the number of target data categories in the set is minimized. Based on the reduced category subset, when determining the risk category of the credit card to be tested, the data information corresponding to the data category with higher relevance can be selected from the data information of the credit card with more dimensions, and then the amount of input data for calculation can be effectively reduced.

[0112] In the embodiments of the present application, since the fuzzy rough set model is a mathematical model combining fuzzy set theory and rough set theory, it can be used to process uncertainty and inaccuracy. Based on the combination of the specific application scenario of credit card risk category determination, by constructing a fuzzy rough set model based on metric learning, the distribution characteristics of the data information in the credit card sample set can be better adapted, effective distance measurement and relevance judgment can be realized, and then the accuracy of determining the target data category can be improved.

[0113] S203, in the case of determining at least one target data category, inputting the data information of the credit card to be tested under at least one target data category into the classification model for classification prediction to determine the risk category of the credit card to be tested.

[0114] For example, the classification model can be any machine model that can perform classification tasks after pre-training, such as linear model, support vector machine (SVM) model, and / or decision tree, etc. For the credit card to be determined, the risk category of the credit card to be determined, the target data category after the data category is simplified, each data information under the target data category is input into the classification model, and the risk category of the credit card to be determined is determined.

[0115] In a possible implementation, data information of the to-be-tested credit card under at least one target data category can be input into the classification model to calculate a function value of the to-be-tested credit card corresponding to a target function; and a classification prediction is performed according to the function value and a preset determination threshold, to determine the risk category of the to-be-tested credit card; wherein the target function is used to calculate a probability of the to-be-tested credit card corresponding to the risk category, the target function is obtained according to the risk category of the credit card in the sample set and a high-dimensional discriminant function through a support vector distance preset, the high-dimensional discriminant function is obtained by transforming a linear discriminant function through a Gaussian kernel, and the linear discriminant function is determined by using a support vector machine model according to data information in the sample set under at least one target data category.

[0116] Exemplarily, the SVM model can be trained by using the reduced category subset, to determine the risk category of the to-be-tested credit card through the trained SVM model.

[0117] For the obtained sample set x={x1,x2,…,x n}, any sample therein can be represented as (x i , L i ), i=1,2,3,……n, L i represents the risk category of the i th sample, L∈{+1,-1}, for example, L i =1 represents a risk. x∈ wherein, represents that the data information of the sample belongs to data information under d dimensions, and the d dimensions correspond to the number of the determined at least one target data category, for example, if 10 target data categories are determined, then d is 10.

[0118] Based on the data information of each sample in the sample set and the label corresponding to the sample, the SVM model is trained, and a hyperplane (w, b) used for classification can be determined. In the d-dimensional space, the linear discriminant function can be represented as g(x)=w·x+b. By transforming x through a kernel function, the linear discriminant function in the d-dimensional space can be transformed into a high-dimensional discriminant function: g(x)=w·φ(x)+b.

[0119] Since the dimensions of the samples in the sample set are high, if the selected kernel function does not adapt to the high-dimensional characteristics of the sample set, there can be a problem of insufficient dimensions. Therefore, the kernel function parameter can be limited to a Gaussian kernel in a radial basis function kernel (RBF). Since the Gaussian kernel can map the original feature space to infinite dimensions, it can better adapt to the high-dimensional characteristics of the data information in the sample set. The expression of the Gaussian kernel can be: wherein, represents a high-dimensional feature distribution parameter, determines the dispersion degree of the sample after the original data is mapped to a high-dimensional space, and a larger will cause the sample dispersion to be relatively sparse, and the farther apart the samples are, the easier they are to separate, but extreme samples are more likely to be separated, and the model is also prone to overfitting. x and y represent the same sample before and after transformation, respectively.

[0120] For example, the high-dimensional discriminant function is normalized, and the samples of the two risk categories satisfy the absolute value |g(x)|≥1, the distance of the nearest sample |g(x)|=1, the distance of the support vector from the discriminant surface is 2 / ∥w∥, and the maximum classification distance is equivalent to:

[0121] For example, in the case of linear non-separability, a slack variable ξ and a penalty coefficient c can be introduced on the basis of the above embodiment, and c represents the tolerance for errors. If a sample point falls on the correct side of the maximum margin hyperplane, the slack variable is 0<ξ i <1; for the sample point j that is misclassified, the slack variable ξ j＝1 The slack variable is multiplied by the penalty coefficient c, and when c is large, even a small slack variable will have a large impact on the final result. Further, the objective function can be transformed as: Based on the objective function and the data information of the credit card to be tested, the probability representing the risk category corresponding to the credit card to be tested can be obtained. The probability can be the function value of the objective function.

[0122] According to the function value and the preset determination threshold, the risk category of the credit card to be tested is determined. The determination threshold can be a preset value for determining the function value, for example, it can be 0.5 or any other value.

[0123] For example, if the determination condition is defined as the function value being greater than or equal to the determination threshold, the risk category is risky, and if the function value is less than the determination threshold, the risk category is non-risky. For example, the function value of the objective function obtained by solving the data information of a credit card to be tested is 0.7, which is greater than the determination threshold 0.5, so it can be determined that the risk category corresponding to the credit card to be tested is risky. Risky can be understood as the credit card may have the risk of credit card fraud and other behaviors.

[0124] Therefore, by inputting the data information of the credit card to be tested under at least one target data category into the classification model, the function value of the objective function corresponding to the credit card to be tested is calculated; according to the function value and the preset determination threshold, the risk category of the credit card to be tested is determined. Based on this, the risk category of the credit card to be tested can be determined more accurately and efficiently.

[0125] The credit card risk category determination method provided in the embodiments of the present application obtains a sample set from the data information of the credit cards with labeled risk categories, and determines which data categories in the sample set have relevance with the risk categories by using a fuzzy rough set model, and determines the target data categories having relevance with the risk categories. Since the fuzzy rough set model can determine the relevance according to a fuzzy approximation operator, the data categories that are helpful to determine the risk categories can be accurately filtered from a large number of data categories, the purpose of filtering out the data categories irrelevant to the determination of the risk categories can be achieved, the number of data categories can be reduced, and then the information amount of the data information can be reduced. In the case where at least one target data category is determined, the data information of the credit card to be tested under the at least one target data category is input into a classification model for classification prediction. Compared with the case where the data information of the credit card to be tested under all data categories is input, the data amount input into the classification model can be reduced, the calculation amount during the operation of the classification model can be reduced, and then the time required for the operation can be shortened, and the efficiency of the classification prediction can be improved. Since the filtered data categories are irrelevant data categories, the irrelevant data information is excluded, the accuracy of the prediction result is high, and therefore, based on the method, the effect of improving the determination efficiency while ensuring the accuracy of the determination of the credit card risk category can be achieved.

[0126] In addition, in the financial field, the interpretability of the model is very important, and the decision basis of the model needs to be explained to the regulatory agencies and users. In the related art described above, since the deep learning model is generally considered as a black box model, the decision process lacks transparency, and the interpretability of the model in the prior art is poor. By constructing a fuzzy rough set model to determine the target data category, the mathematical expression of the fuzzy rough set model is strongly explainable, and therefore, the method based on the embodiments of the present application can improve the interpretability of the risk category determination process.

[0127] The following will be described in combination with Figure 4 A specific implementation manner of the credit card risk category determination method provided in the embodiments of the present application is described, Figure 4 The flowchart of the credit card risk category determination method provided in the embodiments of the present application is shown in Figure Three As shown in the figure, the method comprises the following steps: Figure 4

[0128] S401, data information of a plurality of credit cards is collected, each data information is preprocessed, and is converted into a feature vector.

[0129] For example, when the data information is preprocessed, data cleaning operations such as data completion and error correction can be performed, and normalization processing of the data information can also be included to eliminate the dimensional differences of the data information between different data categories, etc.

[0130] ​For discrete data in the data information, a binary number such as 0 or 1 can be assigned to make it data-based. The discrete data can be information representing yes or no. For example, when using the distance metric Δ M , when the data information a i of the i-th data category is discrete, when x(a i )≠y(a i ), i.e., the data information of sample x and sample y in the i-th data category is not equal, x(a i )-y(a i )=1; when x(a i )=y(a i ), i.e., the data information of sample x and sample y in the i-th data category is equal, x(a i )-y(a i )=0.

[0131] For example, the data information of all data categories of a credit card can be connected to form a vector, which can be converted into a feature vector of the credit card. The feature vector can be understood as the above-mentioned multi-dimensional vector, such as (XXXX XXXX XXXX, 200.01, …) and the like.

[0132] S402, constructing a distance metric function suitable for the collected sample set by the LMNN algorithm.

[0133] For example, the distance metric function can be constructed by the method of the above-mentioned embodiments, and the distance metric function can be, for example, formula (1) in the above-mentioned embodiments.

[0134] S403, constructing an ML-FRS model.

[0135] For example, by constructing the ML-FRS model, a fuzzy upper approximation operator, a fuzzy lower approximation operator, and / or a fuzzy positive region can be defined, and then a function for calculating the fuzzy dependence degree can be obtained, such as formula (5) and formula (6) in the above-mentioned embodiments.

[0136] S404, determining the target data category based on the ML-FRS model and the sample set.

[0137] For example, for each data category included in the sample set, a traversal calculation can be performed to determine whether each data category is a data category having relevance to the risk category, and then the target data category can be determined comprehensively. For example, for any data category a∈B, by constructing a first category subset B and a second category subset B-a, the fuzzy dependence degree is calculated by the ML-FRS model, and then it is determined whether the data category a has relevance to the risk category to determine whether it is the target data category.

[0138] S405, according to the target data category, the SVM model is used for classifying the to-be-tested credit card.

[0139] After the target data category is determined, the data information of the to-be-tested credit card in each target data category can be processed by the SVM model to predict the risk category of the to-be-tested credit card.

[0140] The method provided by the embodiments of the present application can select features in the data categories in the sample set, filter out features that are more suitable for identifying whether the credit card use is risky, that is, determine the target data category, and based on this, the dimension of the data information can be reduced when there are many data categories of the credit card, and the operation efficiency of the model prediction can be improved. In actual application, the method can identify credit card fraud risks such as false application and theft. By verifying the applicant information, analyzing the transaction mode, and other means, abnormal situations can be found in time, and the safety of the funds of the bank and the cardholder can be protected.

[0141] Figure 5 The structure diagram of the credit card risk category determination device provided by the embodiments of the present application is shown in FIG. 1, and the credit card risk category determination device includes: Figure 5

[0142] The acquisition module 501 is configured to acquire a sample set, the sample set including data information of n credit cards, the n credit cards being marked with respective risk categories, the data information of the n credit cards corresponding to m data categories, any one of the credit cards including data information in the m data categories, and m and n being integers greater than 1.

[0143] The determination module 502 is configured to determine, by using a fuzzy rough set model, whether each data category in the sample set has relevance with the risk category, and determine a data category having relevance with the risk category as a target data category, the fuzzy rough set model being a model for determining relevance according to a fuzzy approximation operator.

[0144] The processing module 503 is configured to, in a case where at least one target data category is determined, input data information of a to-be-tested credit card in the at least one target data category into a classification model to perform classification prediction, and determine a risk category of the to-be-tested credit card.

[0145] ​In a possible implementation, the determining module 502 is specifically configured to: select one or more data categories in each data category of the sample set to obtain a first category subset; add or remove one or more data categories in the first category subset to obtain a second category subset; calculate the fuzzy dependency of the data information of each credit card in the sample set under the first category subset on the risk category by using the fuzzy rough set model to obtain a first fuzzy dependency, and calculate the fuzzy dependency of the data information of each credit card in the sample set under the second category subset on the risk category to obtain a second fuzzy dependency; determine whether the data category added or removed when the second category subset is obtained has relevance with the risk category according to the first fuzzy dependency and the second fuzzy dependency, and determine the data category having relevance with the risk category as the target data category.

[0146] In a possible implementation, the fuzzy dependency is calculated by using the fuzzy rough set model, and the following formula is satisfied: wherein, U represents the sample set; x i represents the data information of the i th credit card in the sample set; L represents each risk category; M represents a distance metric matrix; B represents a category subset corresponding to the calculated fuzzy dependency; represents the data information of the i th credit card in the sample set; L represents each risk category; M represents a distance metric matrix; B represents a category subset corresponding to the calculated fuzzy dependency; i represents the fuzzy lower approximation operator of B on each risk category based on U; represents a fuzzy positive region; | | represents a modulo operation; represents the fuzzy dependency.

[0147] In a possible implementation, the distance metric matrix is determined by iterative learning according to the data information in the sample set by using a large-margin nearest neighbor algorithm.

[0148] In a possible implementation, the distance metric matrix is a Markov matrix, and the Markov matrix satisfies the following formula: wherein, M represents the Markov matrix; x j represents the data information of the j th credit card in the sample set; D M (x i ,x j ) represents the Markov distance between x i and x j ; when the iterative learning is performed by using the large-margin nearest neighbor algorithm, a loss function used satisfies the following formula: ε(M)=(1-μ)ε pull (M)+με push (M); wherein, ε(M) represents the value of the loss function; μ represents a preset coefficient; SKNN(x i ) represents the distance between x ia set of k nearest neighbors with the same risk category, k being any positive integer; x l data information of the lth credit card belonging to the KNN(x i ) set and not belonging to the SKNN(x i ) set, KNN(x i ) being a set of k nearest neighbors most proximate to x i , k being any positive integer;[Z] + =max(Z,0) represents the value in the parentheses compared with 0 and taking the maximum; y il represents the value corresponding to x i and x j having the same or different risk categories, x i and x j having the same risk category taking 1, and x i and x j having different risk categories taking 0.

[0149] In a possible implementation, the processing module 503 is specifically configured to: input data information of the to-be-tested credit card in at least one target data category into the classification model to calculate a function value of the to-be-tested credit card corresponding to a target function; and perform classification prediction according to the function value and a preset determination threshold to determine the risk category of the to-be-tested credit card, wherein the target function is used to calculate a probability of the risk category of the to-be-tested credit card, the target function is obtained by presetting a support vector distance according to risk categories of credit cards in the sample set and a high-dimensional discriminant function, the high-dimensional discriminant function is obtained by transforming a linear discriminant function by using a Gaussian kernel, and the linear discriminant function is determined by using a support vector machine model in at least one target data category according to data information in the sample set.

[0150] The credit card risk category determination apparatus provided in the embodiments of the present application can be used to execute the technical solutions of the credit card risk category determination method in any of the embodiments of the present application, and has similar implementation principles and technical effects, which will not be described herein again.

[0151] Figure 6 A structural schematic diagram of an electronic device provided in the embodiments of the present application is shown in FIG. 1, which can include at least one processor 601 and a memory 602 in communication connection with the at least one processor 601, wherein the memory 602 stores instructions executable by the at least one processor 601, and the instructions are executed by the at least one processor 601 to enable the electronic device to perform the method in any of the above embodiments. Figure 6

[0152] Optionally, the memory 602 can be independent or integrated with the processor 601.

[0153] ​The implementation principles and technical effects of the electronic device provided in this embodiment can be referred to the foregoing embodiments, and will not be described here.

[0154] The application further provides a computer readable storage medium, wherein computer execution instructions are stored in the computer readable storage medium, and when a processor executes the computer execution instructions, the method in any of the foregoing embodiments is implemented.

[0155] The application further provides a computer program product, comprising a computer program, and when the computer program is executed by a processor, the method in any of the foregoing embodiments is implemented.

[0156] In several embodiments provided in the application, it should be understood that the disclosed devices and methods can be implemented in other manners. For example, the device embodiments described above are merely schematic; for example, the division of the modules is merely a logical function division, and there can be another division manner in actual implementation; for example, a plurality of modules or features can be combined or integrated into another system, or some features can be ignored or not executed.

[0157] The integrated modules in the form of software functional modules above can be stored in a computer readable storage medium. The software functional modules stored in the storage medium include a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the methods described in the embodiments of the application.

[0158] It should be understood that the processor above can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the application can be directly embodied in the form of hardware processor execution, or be executed by a combination of hardware and software modules in the processor. The memory can include random access memory (RAM), and can also include non-volatile memory (NVM), for example, at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0159] The storage medium can be realized by any type of volatile or nonvolatile storage devices or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0160] An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. Of course, the storage medium can be part of the processor. Consistent with the teachings provided herein, the processor and the storage medium can be located in a special purpose computing device. The processor and the storage medium can also be located in a general purpose computing device.

[0161] It should be noted that, in the present document, the terms "comprising", "containing" or any other similar term are intended to encompass non-exclusive inclusions, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0162] The above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0163] From the above description of the embodiments, one skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, and of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk), and includes a plurality of instructions for causing a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the methods described in the various embodiments of the present application.

[0164] The preferred embodiments of the present application are described above with the specific details. However, it should be evident that the application does not inciide to limit the application to these details alone, but rather other modifications will occur to those skilled in the art, once given the present disclosure. Hence, numerous changes and modifications can be made to the embodiments described above and illustrated in the drawings without departing from the scope of the present application.

[0165] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the present application cover any and all variations of the embodiments of the present application described herein and any and all equivalents thereof and be limited only by the scope of the claims.

[0166] It is to be understood that the application is not limited to the specific details described above and illustrated in the drawings. Rather, various modifications can be made to the embodiments described above and illustrated in the drawings without departing from the scope of the present application. The scope of the application is limited only by the claims.

Claims

1. A credit card risk category determination method, characterized by, The method comprises the following steps: obtaining a sample set comprising data information of n credit cards, each of the n credit cards being marked with a respective risk category, the data information of the n credit cards corresponding to m data categories, each of the credit cards comprising data information in the m data categories, the m and the n being integers greater than 1; determining, by a fuzzy rough set model, whether each data category in the sample set has relevance with the risk category, and determining a data category having relevance with the risk category as a target data category, the fuzzy rough set model being a model for determining relevance according to a fuzzy approximation operator; in the case where at least one target data category is determined, inputting data information of a to-be-tested credit card in the at least one target data category into a classification model for classification prediction to determine a risk category of the to-be-tested credit card.

2. The method of claim 1, wherein, The step of determining, by a fuzzy rough set model, whether each data category in the sample set has relevance with the risk category, and determining a data category having relevance with the risk category as a target data category comprises the following steps: selecting one or more data categories in each data category of the sample set to obtain a first category subset; adding one or more data categories not selected in the first category subset or removing one or more data categories in the first category subset to obtain a second category subset; calculating, by a fuzzy rough set model, a fuzzy dependency degree of data information of each credit card in the sample set in the first category subset with respect to the risk category to obtain a first fuzzy dependency degree, and calculating a fuzzy dependency degree of data information of each credit card in the sample set in the second category subset with respect to the risk category to obtain a second fuzzy dependency degree; determining, according to the first fuzzy dependency degree and the second fuzzy dependency degree, whether the data category added or removed when obtaining the second category subset has relevance with the risk category, and determining a data category having relevance with the risk category as a target data category.

3. The method of claim 2, wherein, The fuzzy dependency degree calculated by the fuzzy rough set model satisfies the following formula: where U denotes the sample set; x i denotes the data information of the i-th credit card in the sample set; L denotes each risk category; M denotes the distance metric matrix; B denotes the category subset corresponding to the calculated fuzzy dependence degree; denotes x i based on B for the fuzzy lower approximation operator of each risk category; denotes the fuzzy positive region; | | denotes the modulus; denotes the fuzzy dependence degree.

4. The method of claim 3, wherein, The distance metric matrix is determined according to data information in the sample set by iterative learning through a large-margin nearest neighbor algorithm.

5. The method of claim 4, wherein, The distance metric matrix is a Markov matrix, and the Markov matrix satisfies the following formula: wherein M represents the Markov matrix; x j represents the data information of the jth credit card in the sample set; D M (x i ,x j ) represents the Markov distance between x i and x j ; When the iterative learning is performed through the large-margin nearest neighbor algorithm, a loss function used satisfies the following formula: ε(M) = (1 - μ)ε pull (M) + με push (M); wherein ε(M) represents the value of the loss function; μ represents a preset coefficient; SKNN(x i ) represents a set of k nearest neighbors with the same risk category as x i ; x l represents the data information of the lth credit card belonging to the set of KNN(x i ) and not belonging to the set of SKNN(x i ); KNN(x i ) is a set of k nearest neighbors most proximate to x i ; k is an arbitrary positive integer; [Z] + =max(Z,0) represents the value in the brackets compared with 0 and taking the maximum; y il represents the value corresponding to x i and x j with the same or different risk categories; when x i and x j have the same risk category, the value is 1; when x i and x j have different risk categories, the value is 0.

6. The method according to any one of claims 1 to 5, characterized in that, The step of inputting data information of a to-be-tested credit card in the at least one target data category into a classification model for classification prediction to determine a risk category of the to-be-tested credit card comprises the following steps: inputting data information of the to-be-tested credit card in the at least one target data category into the classification model to calculate a function value of the to-be-tested credit card corresponding to a target function; performing classification prediction according to the function value and a preset determination threshold to determine a risk category of the to-be-tested credit card. The target function is used to calculate the probability of the risk category of the credit card to be tested, and the target function is obtained according to the risk category of the credit card in the sample set and the high-dimensional discriminant function through the support vector distance presetting. The high-dimensional discriminant function is obtained by transforming the linear discriminant function through the Gaussian kernel. The linear discriminant function is determined by the support vector machine model according to the data information in the sample set under the at least one target data category.

7. A device for determining the risk category of a credit card, characterized in that, Comprise: An acquisition module is configured to acquire a sample set, wherein the sample set comprises data information of n credit cards, each of the n credit cards is marked with a respective risk category, the data information of the n credit cards corresponds to m data categories, any one of the credit cards comprises data information under the m data categories, and m and n are integers greater than 1. A determination module is configured to determine whether each data category in the sample set has relevance with the risk category by using a fuzzy rough set model, and determine the data category having relevance with the risk category as a target data category, wherein the fuzzy rough set model is a model for determining relevance according to a fuzzy approximation operator. A processing module is configured to input data information of a credit card to be tested under at least one target data category into a classification model for classification prediction to determine the risk category of the credit card to be tested.

8. An electronic device, comprising: Comprise: A memory and a processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to implement the method of any one of claims 1-6.

10. A computer program product, characterised in that, Comprise a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-6.

Images