Transaction risk identification method and device, equipment, storage medium and product

Abstract

The invention relates to the technical field of big data, and discloses a transaction risk identification method, device and equipment, a storage medium and a computer program product, and the method comprises the steps: obtaining a transaction request, inputting the transaction request into a transaction risk identification model, the transaction risk identification model comprises a feature extraction network, a fine feature extraction and selection network and a transaction risk identification network, initial features are extracted from a transaction request through the feature extraction network, and key fine features are extracted from the initial features through the fine feature extraction and selection network; performing risk identification on the transaction request based on the key fine features through a transaction risk identification network to obtain a transaction risk identification result; according to the invention, the network structure of the transaction risk identification model is improved, and two layers of feature extraction networks are designed, so that the accuracy of transaction risk identification can be improved, and the transaction risk can be reduced.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a method, apparatus, device, storage medium, and computer program product for identifying transaction risks. Background Technology

[0002] With social progress and economic development, daily transactions have increased. Consequently, many abnormal transaction situations have emerged. For example, some entities use malicious consumer credit installment products to seek illegal profits; specifically, end-users may maliciously default on payments after using consumer credit installment products.

[0003] In existing technologies, the accuracy of transaction risk detection is low, and failure to accurately identify transaction risks of end users can have a significant impact on business operations. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, storage medium, and computer program product for identifying transaction risks, aiming to solve the technical problem that the accuracy of transaction risk detection in the prior art is low, and failure to accurately identify the transaction risks of end users will have a significant impact on business.

[0005] To achieve the above objectives, the present invention provides a method for identifying transaction risks, the method comprising:

[0006] A transaction request is obtained and input into a transaction risk identification model, which includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network.

[0007] Initial features are extracted from the transaction request through the feature extraction network, and key fine features are extracted from the initial features through the fine feature extraction and selection network.

[0008] The transaction risk identification network identifies the risks of the transaction request based on the key fine features, and obtains the transaction risk identification result.

[0009] Optionally, the fine feature extraction and selection network includes multiple fine feature extraction and selection blocks, each with a different number of neurons. These blocks form multiple feature extraction and selection branches. Extracting key fine features from the initial features through the fine feature extraction and selection network includes:

[0010] Key fine features are extracted from the initial features through the feature extraction and selection branch, wherein different operations are performed on the output data of different fine feature extraction selection blocks in the feature extraction and selection branch, and the different operations include at least one of concatenation operation, averaging operation, convolution operation, and information transmission.

[0011] Optionally, the feature extraction network includes multiple stacked residual blocks, each residual block having a different number of convolutional channels. Each residual block includes a deep feature extraction branch and a basic feature extraction branch. Extracting initial features from the transaction request through the feature extraction network includes:

[0012] Deep features are extracted from the transaction request through the deep feature extraction branch, wherein the rectified linear unit is used as the activation function in the deep feature extraction branch;

[0013] Basic features are extracted from the transaction request through the basic feature extraction branch;

[0014] The deep features are combined with the basic features to obtain the initial features.

[0015] Optionally, the transaction risk identification network includes an encoder, a generator, a discriminator, and a classifier. Before acquiring the transaction request and inputting it into the transaction risk identification model, the network further includes:

[0016] Obtain business information samples and input them into the transaction risk identification model;

[0017] Initial features are extracted from the business information sample through the feature extraction network, and key fine features are extracted from the initial features through the fine feature extraction and selection network.

[0018] The encoder encodes the key fine features, and the generator recreates the risk-free transaction data in the encoded features to obtain risky transaction samples.

[0019] The business information sample and the risk transaction sample are input into the discriminator to identify the risk transaction and obtain the risk transaction identification result.

[0020] The transaction risk identification model is trained based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples to obtain the trained transaction risk identification model.

[0021] Accordingly, the step of using the transaction risk identification network to identify the transaction request based on the key refined features and obtaining the transaction risk identification result includes:

[0022] The classifier in the trained transaction risk identification model identifies the risk of the transaction request based on the key fine features, and obtains the transaction risk identification result.

[0023] Optionally, the step of recreating the risk-free transaction data from the encoded features to obtain risky transaction samples includes:

[0024] Based on the risk-free transaction data in the encoded features, identify end users without transaction risk and obtain the user information of the end users.

[0025] Based on the user information, query the fuzzy average return, risk-free rate of return, and standard deviation of the fuzzy return of the main revenue sources of the terminal user during the business period;

[0026] The business delinquency risk of the end user is determined based on the fuzzy average return, the risk-free rate of return, and the standard deviation of the fuzzy returns of the main revenue source;

[0027] The risk of overdue credit purchase transactions of the end users is adjusted based on the risk-free rate of return, the actual number of business days of the end users, and the risk premium.

[0028] A risk transaction sample is constructed based on the adjusted business delinquency risk and the business information of the end users.

[0029] Optionally, training the transaction risk identification model based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples to obtain the trained transaction risk identification model includes:

[0030] The modified cross-entropy loss function is calculated based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples. In the modified cross-entropy loss function, the weight value of the sample category is inversely proportional to the number of samples of the sample category.

[0031] The trading risk identification model is trained based on the modified cross-entropy loss function to obtain the trained trading risk identification model.

[0032] Furthermore, to achieve the above objectives, the present invention also proposes a transaction risk identification device, the transaction risk identification device comprising:

[0033] An input module is used to acquire transaction requests and input the transaction requests into a transaction risk identification model, which includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network.

[0034] The extraction module is used to extract initial features from the transaction request through the feature extraction network, and to extract key fine features from the initial features through the fine feature extraction and selection network;

[0035] The identification module is used to identify the risk of the transaction request based on the key fine features through the transaction risk identification network, and obtain the transaction risk identification result.

[0036] Furthermore, to achieve the above objectives, the present invention also proposes a transaction risk identification device, which includes a memory, a processor, and a transaction risk identification program stored in the memory and executable on the processor. The transaction risk identification program is configured to implement the transaction risk identification method described above.

[0037] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method as described above.

[0038] In addition, to achieve the above objectives, the present invention also provides a computer program product, the computer program product including a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method as described above.

[0039] This invention discloses a method for acquiring transaction requests and inputting them into a transaction risk identification model. The transaction risk identification model includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network. The feature extraction network extracts initial features from the transaction request, and the fine feature extraction and selection network extracts key fine features from the initial features. The transaction risk identification network identifies the risk of the transaction request based on the key fine features, thereby obtaining the transaction risk identification result. Because this invention improves the network structure of the transaction risk identification model by designing a two-layer feature extraction network, it can improve the accuracy of transaction risk identification and thus reduce transaction risk. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the structure of a transaction risk identification device in the hardware operating environment involved in the embodiments of the present invention;

[0041] Figure 2 This is a flowchart illustrating the first embodiment of the transaction risk identification method of the present invention;

[0042] Figure 3 This is an interactive diagram of an embodiment of the transaction risk identification method of the present invention;

[0043] Figure 4This is a flowchart illustrating the second embodiment of the transaction risk identification method of the present invention;

[0044] Figure 5 This is a structural diagram of a transaction risk identification model according to an embodiment of the transaction risk identification method of the present invention;

[0045] Figure 6 This is a detailed network structure diagram of each fine feature extraction selection block in an embodiment of the transaction risk identification method of the present invention;

[0046] Figure 7 This is a flowchart illustrating the third embodiment of the transaction risk identification method of the present invention;

[0047] Figure 8 This is a schematic diagram of the network structure of a transaction risk identification network according to an embodiment of the transaction risk identification method of the present invention;

[0048] Figure 9 This is a structural block diagram of the first embodiment of the transaction risk identification device of the present invention.

[0049] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] Reference Figure 1 , Figure 1 This is a schematic diagram of the transaction risk identification device structure in the hardware operating environment involved in the embodiments of the present invention.

[0052] like Figure 1As shown, the transaction risk identification device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. In this invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0053] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the transaction risk identification device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0054] like Figure 1 As shown, the memory 1005, which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a transaction risk identification program.

[0055] exist Figure 1 In the transaction risk identification device shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the user equipment; the transaction risk identification device calls the transaction risk identification program stored in the memory 1005 through the processor 1001 and executes the transaction risk identification method provided in the embodiment of the present invention.

[0056] Based on the above hardware structure, an embodiment of the transaction risk identification method of the present invention is proposed.

[0057] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the transaction risk identification method of the present invention, which presents the first embodiment of the transaction risk identification method of the present invention.

[0058] It should be understood that the transaction risk identification method of the present invention can be applied to credit-based mobile phone purchase business. Of course, the transaction risk identification method of the present invention can also be applied to other businesses. This embodiment uses credit-based mobile phone purchase business as a scenario for explanation.

[0059] Based on its credit system and payment capabilities, China Mobile Fintech utilizes user communication, consumption, credit, and social data, along with big data risk control technology, and combines this with the call charge policies of various provincial companies to lower the barriers to handset purchase for users. This allows users to enjoy more affordable terminals, a wider range of terminals, and more convenient communication services through consumer credit installment products and services.

[0060] The credit-based handset purchase service is applicable to various consumption scenarios within the operator's system, such as offline sales of terminals and contracts, online electronic channel consumption, installment payments for unlocked handsets, and sales of general smart terminals. It supports installment methods such as pre-approval, real-time approval, and asset pledging, including installment payments for debit cards, Alipay pre-authorization, WeBank installment payments, credit card installment payments, and installment payments secured by debit cards. It can meet the needs of different user groups and provide explicit, flexible, and rich marketing forms and methods for provincial companies' package promotions.

[0061] However, with social progress and economic development, daily transactions have increased. This has also led to many abnormal transaction situations, such as entities using malicious consumer credit installment products to seek illegal profits. Specifically, for example, end-users may maliciously default on payments after using consumer credit installment products.

[0062] In existing technologies, the accuracy of identifying transaction risks of end users who maliciously default on payments is low. If the transaction risks of end users are not accurately identified, it will have a significant impact on the credit purchase business and cause great losses to the company.

[0063] Therefore, in order to overcome the above-mentioned defects, this embodiment improves the network structure of the transaction risk identification model by designing a two-layer feature extraction network, which can improve the accuracy of transaction risk identification and thus reduce transaction risk.

[0064] In the first embodiment, the transaction risk identification method includes:

[0065] Step S10: Obtain a transaction request and input the transaction request into a transaction risk identification model, which includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network.

[0066] It is understood that the execution subject of this embodiment may be a transaction risk identification device with data processing, network communication and program running functions, such as a server, or other electronic devices that can achieve the same or similar functions. This embodiment does not limit this.

[0067] Step S20: Extract initial features from the transaction request through the feature extraction network, and extract key fine features from the initial features through the fine feature extraction and selection network.

[0068] Step S30: The transaction risk identification network identifies the transaction request based on the key fine features to obtain the transaction risk identification result.

[0069] For ease of understanding, please refer to Figure 3 This description is provided but does not limit the scope of the invention. Figure 3 This is an interactive diagram of an embodiment of the transaction risk identification method of the present invention. The server is equipped with a transaction risk identification system (i.e., the Yutong Risk Control System in the diagram), and the transaction risk identification system is equipped with a transaction risk identification module.

[0070] To address the technical problems existing in related technologies, the transaction risk identification method of this invention includes the following steps:

[0071] Step 1: After an end user sends a credit purchase request to the business hall terminal, the business hall terminal will forward the end user's credit purchase request to the Yutong risk control system.

[0072] A credit purchase request (i.e., a business request) includes information on the installment loan product the end user wants to purchase, the installment loan method, and the end user's identity information.

[0073] Step 2: After receiving the credit purchase request from the end user, the Yutong Risk Control System calls the transaction risk identification module in the Yutong Risk Control System so that the transaction risk identification module can identify the transaction risk of the end user's credit purchase.

[0074] The YuTong Risk Control System (i.e., Transaction Risk Identification System) is a risk control platform that can simultaneously identify business risks for services such as points malls, marketing activities, aggregated payments, and credit-based mobile phone purchases. Since the YuTong Risk Control System has built different transaction risk identification models for different services, in this invention, after receiving a credit-based mobile phone purchase request from an end user, the YuTong Risk Control System calls the transaction risk identification module corresponding to the credit-based mobile phone purchase service. This allows the transaction risk identification module to assess the end user based on big data credit scoring capabilities.

[0075] Step 3: The transaction risk identification module responds to the request from the Yutong risk control system and uses the trained transaction risk identification model to identify transaction risks in the credit-based mobile phone purchase business.

[0076] Among them, the feature extraction network is used to extract feature information vectors (i.e. initial features) based on the credit installment product information, credit installment method and end user identity information of the end user in the credit purchase business request.

[0077] The fine feature extraction and selection network is used to extract more refined features based on the output of the feature extraction network, and to filter the more critical features among the refined features;

[0078] The transaction risk identification network is used to identify risks in transaction requests based on key, refined features.

[0079] Step 4: After the transaction risk identification module obtains the transaction risk identification result of the credit purchase business, it returns the transaction risk identification result to the Yutong Risk Control System, which then sends the transaction risk identification result to the business hall terminal.

[0080] Step 5: After receiving the transaction risk identification results returned by the Yutong Risk Control System, the business hall terminal can determine whether to continue selling credit installment products to the end user based on the transaction risk identification results.

[0081] For example, if the transaction risk identification result returned by the Yutong risk control system indicates that the end user's credit-based mobile phone purchase transaction involves transaction risk, the sales hall terminal can refuse to provide the end user with the credit-based mobile phone purchase transaction. Conversely, if the transaction risk identification result returned by the Yutong risk control system indicates that the end user's credit-based mobile phone purchase transaction does not involve transaction risk, the sales hall terminal can provide the end user with the credit-based mobile phone purchase transaction.

[0082] This embodiment provides a transaction risk identification model. Due to the innovative adjustment of the network structure of the transaction risk identification model, the accuracy of the transaction risk identification results identified by the model can be guaranteed to be high.

[0083] Furthermore, because the transaction risk identification model in this embodiment employs a two-layer feature extraction network (including a feature extraction network and a fine-grained feature extraction and selection network), it can learn more complex feature information from credit-based mobile phone purchase requests. This allows for the gradual combination of simple features into higher-level features, ensuring more accurate identification results. On the other hand, it reduces reliance on specific details (because different overdue repayment situations correspond to different specific details; focusing too much on details may cause the transaction risk identification model to only identify risky transactions with certain specific characteristics), thereby improving the model's generalization ability and robustness.

[0084] In this embodiment, a method for acquiring transaction requests and inputting them into a transaction risk identification model is disclosed. The transaction risk identification model includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network. The feature extraction network extracts initial features from the transaction request, and the fine feature extraction and selection network extracts key fine features from the initial features. The transaction risk identification network identifies the risk of the transaction request based on the key fine features to obtain the transaction risk identification result. Because this embodiment improves the network structure of the transaction risk identification model by designing a two-layer feature extraction network, the accuracy of transaction risk identification can be improved, thereby reducing transaction risk.

[0085] Reference Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the transaction risk identification method of the present invention, based on the above. Figure 2 The first embodiment shown presents a second embodiment of the transaction risk identification method of the present invention.

[0086] In the second embodiment, the feature extraction network includes multiple stacked residual blocks, each residual block having a different number of convolutional channels. Each residual block includes a deep feature extraction branch and a basic feature extraction branch. Step S20 includes:

[0087] Step S201: Extract deep features from the transaction request through the deep feature extraction branch, wherein the rectified linear unit is used as the activation function in the deep feature extraction branch.

[0088] Step S202: Extract basic features from the transaction request through the basic feature extraction branch.

[0089] Step S203: Combine the deep features with the basic features to obtain the initial features.

[0090] It is understandable that this embodiment uses rectified linear units as activation functions after the deep feature extraction branch, thereby increasing the nonlinear expressive power of the feature extraction network.

[0091] For ease of understanding, please refer to Figure 5 This description is provided but does not limit the scope of the invention. Figure 5 This is a structural diagram of a transaction risk identification model according to an embodiment of the transaction risk identification method of the present invention. In the diagram, the transaction risk identification model of the present invention mainly includes a feature extraction network (i.e.,...) Figure 5 The part within the red dashed box), fine-grained feature extraction and selection network (i.e. Figure 5 (the part within the green dashed box) and the transaction risk identification network ( Figure 5 (The detailed structure is not shown here; a detailed description of this part will follow later.) It consists of three parts.

[0092] Among them, the feature extraction network is mainly used to extract feature information vectors based on the credit installment product information, credit installment method, and end-user identity information of the end-user in the credit purchase business request.

[0093] The feature extraction network consists of four stacked residual blocks ( Figure 5 The content within the blue dashed box represents only the network structure of one residual block. The network structures of the following three residual blocks are the same as those within the blue dashed box (detailed structures omitted in the figure). The number of convolution channels in these four residual blocks are set to 8, 16, 32, and 64, respectively.

[0094] Each residual block consists of a deep feature extraction branch (i.e., Figure 5 The blue dashed box contains two branches of one-dimensional convolutional layers on the right (each convolutional kernel is set to a size of 1*1) and a basic feature extraction branch (i.e., ...). Figure 5 The input data is composed of the left branch within the blue dashed box. The deep feature extraction branch, as the name suggests, is used to extract deep features from the input data. The basic feature extraction branch is used to extract the basic features from the input data.

[0095] After the deep feature extraction branch, a rectified linear unit (ReLU) is used as the activation function to increase the non-linear expressive power of the feature extraction network.

[0096] The mapping process of a set of convolutional layers in the deep feature extraction branch is as follows:

[0097]

[0098] Where m is the size of the convolution kernel, k j Let represent the j-th weight coefficient of the convolution kernel, b be the bias, and δ be the non-linear activation function.

[0099] The output of the residual block is obtained by adding the outputs of the deep feature extraction branch and the basic feature extraction branch. To ensure consistency between the outputs of the two branches, a one-dimensional convolutional layer with a 1*1 kernel can be used to adjust the output dimension in the feature extraction branch. At the output of each residual block, a max pooling layer can be used to reduce the feature dimension. The kernel size of the pooling layer is set to 3. Furthermore, the max pooling layer can reduce the feature dimension while also performing feature selection.

[0100] In this embodiment, after the deep feature extraction branch, a rectified linear unit is used as the activation function, thereby increasing the nonlinear expressive power of the feature extraction network.

[0101] In the second embodiment, the fine feature extraction and selection network includes multiple fine feature extraction and selection blocks, each with a different number of neurons. These fine feature extraction and selection blocks form multiple feature extraction and selection branches. Step S20 further includes:

[0102] Step S204: Extract key fine features from the initial features through the feature extraction and selection branch, wherein different operations are performed on the output data of different fine feature extraction selection blocks in the feature extraction and selection branch, and the different operations include at least one of concatenation operation, averaging operation, convolution operation, and information transmission.

[0103] It should be understood that after receiving the initial feature output result from the feature extraction network, the fine feature extraction and selection network in this embodiment will further extract fine features from the initial features through multiple feature extraction and selection branches composed of fine feature extraction and selection blocks, and perform information filtering on the features in the fine features, thereby obtaining more critical fine features.

[0104] For ease of understanding, please refer to Figure 5 This description is provided but does not limit the scope of the invention. Figure 5 This is a structural diagram of a transaction risk identification model according to an embodiment of the transaction risk identification method of the present invention. In the diagram, the fine feature extraction and selection network, namely... Figure 5 The area within the green dashed box is used to extract more refined features based on the output of the feature extraction network, and to filter out the more critical features among the refined features.

[0105] The reason this invention first extracts features using a feature extraction network, and then further refines and selects more critical features from these refined features is twofold. Firstly, it allows learning more complex feature information from credit-based mobile phone purchase requests, enabling the gradual combination of simple features into higher-level features. Secondly, it reduces reliance on specific details (because different overdue repayment situations correspond to different specific details; focusing too much on details might cause the transaction risk identification model to only recognize risky transactions with certain specific characteristics), thereby improving the model's generalization ability and robustness.

[0106] from Figure 5 As can be seen, the fine feature extraction and selection network consists of 16 fine feature extraction and selection blocks, with different numbers of neurons in each block: 64, 128, 256, and 512.

[0107] Furthermore, from Figure 5 It can also be seen that the connecting lines between different fine feature extraction selection blocks are different colors, for example... Figure 5The color scheme includes red, green, orange, and black. This indicates different operations performed on the output data of different fine-grained feature extraction selection blocks. For example, red indicates a concatenation operation; green indicates an averaging operation; orange indicates a convolution operation; and black indicates no operation but only information transmission.

[0108] After receiving the initial feature output from the feature extraction network, the fine feature extraction and selection network further extracts fine features from the initial features through four feature extraction and selection branches composed of its 16 fine feature extraction and selection blocks, and filters information from the fine features to obtain more critical fine features.

[0109] For ease of understanding, please also refer to Figure 6 This description is provided but does not limit the scope of the invention. Figure 6 This is a detailed network structure diagram of each fine feature extraction selection block in an embodiment of the transaction risk identification method of the present invention. In the diagram, it is mainly achieved through 4 cascaded blocks ( Figure 6 (block1-block4), 4 max pooling layers) Figure 6 It consists of a maxpool and four dropout layers. Each cascaded block (e.g., maxpool) is composed of four dropout layers. Figure 6 The content within the purple dashed box is a cascaded block (the detailed network structure of blocks 2 to 4 is consistent with this), consisting of 15 convolutional layers and 3 fusion layers.

[0110] After receiving the initial feature output from the feature extraction network, the fine feature extraction and selection network in this embodiment further extracts fine features from the initial features through multiple feature extraction and selection branches composed of fine feature extraction and selection blocks, and filters information from the fine features to obtain more critical fine features.

[0111] Reference Figure 7 , Figure 7 This is a flowchart illustrating the third embodiment of the transaction risk identification method of the present invention. Based on the above embodiments, the third embodiment of the transaction risk identification method of the present invention is proposed.

[0112] In the third embodiment, the transaction risk identification network includes an encoder, a generator, a discriminator, and a classifier. Before step S10, it further includes:

[0113] Step S01: Obtain a business information sample and input the business information sample into the transaction risk identification model.

[0114] Step S02: Extract initial features from the business information sample through the feature extraction network, and extract key fine features from the initial features through the fine feature extraction and selection network.

[0115] Step S03: The key fine features are encoded by the encoder, and the risk-free transaction data in the encoded features are recreated by the generator to obtain risky transaction samples.

[0116] Furthermore, to improve the reliability of risky transaction samples, the process of recreating risk-free transaction data from the encoded features to obtain risky transaction samples includes: identifying end users without transaction risk based on the risk-free transaction data from the encoded features, and obtaining the user information of the end users; querying the fuzzy average return, risk-free rate of return, and standard deviation of the fuzzy return of the main revenue sources of the end users during the business period based on the user information; determining the business delinquency risk of the end users based on the fuzzy average return, the risk-free rate of return, and the standard deviation of the fuzzy return of the main revenue sources; adjusting the delinquency risk of the credit purchase business of the end users based on the risk-free rate of return, the actual number of business days of the end users, and the risk premium; and constructing risky transaction samples based on the adjusted business delinquency risk and the business information of the end users.

[0117] Step S04: Input the business information sample and the risk transaction sample into the discriminator to identify the risk transaction and obtain the risk transaction identification result.

[0118] Step S05: Train the transaction risk identification model based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples to obtain the trained transaction risk identification model.

[0119] Furthermore, considering the potential imbalance in the training sample set during the training of the transaction risk identification network, which could lead to the trained model favoring categories with more samples, this invention modifies the cross-entropy loss function to address this issue. Higher weights are assigned to categories with fewer samples, while lower weights are assigned to categories with more samples. This compensates for misclassification of the data. Step S05 includes: calculating the modified cross-entropy loss function based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples. In the modified cross-entropy loss function, the weight value of the sample category is inversely proportional to the number of samples in that category. The transaction risk identification model is then trained based on the modified cross-entropy loss function to obtain the trained transaction risk identification model.

[0120] Accordingly, step S30 includes:

[0121] Step S30': The classifier in the trained transaction risk identification model identifies the transaction request based on the key fine features to obtain the transaction risk identification result.

[0122] It should be understood that the transaction risk identification network can construct risky transaction samples based on the overdue risk of credit-based mobile phone purchases by end users without transaction risk and the credit-based mobile phone purchase information of end users without transaction risk. Based on the actual situation of the credit-based mobile phone purchase information samples and the newly constructed risky transaction samples, the update generator and discriminator network are updated. In this way, since these risky transaction samples are obtained based on changes in the credit-based mobile phone purchase information of end users without risk, the transaction risk identification model trained in this way can identify end users who appear to be without risk but actually have transaction risks, thus improving the accuracy of transaction risk identification.

[0123] For ease of understanding, please refer to Figure 8 This description is provided but does not limit the scope of the invention. Figure 8 This is a schematic diagram of the network structure of a transaction risk identification network according to an embodiment of the transaction risk identification method of the present invention. Figure 8 As can be seen from this, the transaction risk identification network includes an encoder (i.e., Figure 8 (blue part), generator (i.e.) Figure 8 (red part in the middle), classifier (i.e.) Figure 8 (green part) and discriminator (i.e.) Figure 8 (The purple part) consists of four parts.

[0124] The generator and discriminator can form a generator-discriminator network, which in turn contains a generator network. The generator network includes an encoder and a decoder network. The encoder encodes the output data from the fine-grained feature extraction and selection network, while the decoder can reconstruct risky transaction samples from the input risk-free data, leaving data that is inherently risky unprocessed. The classifier then uses these extracted features to classify and identify transaction risks.

[0125] The encoder can use an improved one-dimensional feature vector converter, and the encoder has two layers and two heads. The generated vector sequence is then fed into the encoder.

[0126] The classifier consists of a normalization layer, a fully connected multilayer perceptron (MLP), a linear layer, and a softmax activation function. Furthermore, to improve the accuracy of transaction risk identification, the encoder, classifier, generator, and discriminator networks can be integrated. The generator consists of a normalization layer, a fully connected neural network, and a tanh activation function. The generator generates risky transaction samples, which are then fed into the discriminator.

[0127] Secondly, from Figure 8 As can be seen, since the transaction risk identification network contains several fully connected multilayer neural networks, dropout technology is used to prevent overfitting.

[0128] The discriminator network contains three convolutional blocks, three normalization layers, one average pooling layer, and one sigmoid activation function. Figure 8 (LeakyReLU). The discriminator network distinguishes between actual samples and generated sample data. During training, these inputs are fed into the discriminator, and a gradient penalty is used to update both the generator and discriminator networks. If two inputs are classified as similar, the output should be 1; if they are not classified as similar, the output should be zero.

[0129] In addition, from Figure 8 As can be seen, a normalization layer is used in the encoder, generator, discriminator, and classifier. Figure 8 In this way, all received information can be normalized using layer normalization before being fed into the specific network module.

[0130] For any vector, the normalization is calculated as follows:

[0131]

[0132] Where v represents the feature vector; d represents the number of different feature vectors; and γ, μ, β and σ represent preset normalization parameters.

[0133] Considering the potential imbalance in the training sample set during the training of the transaction risk identification network, which could lead to a biased model towards categories with more samples, this invention modifies the cross-entropy loss function to address this issue. Higher weights are assigned to categories with fewer samples, while lower weights are assigned to categories with more samples. This compensates for misclassification of our data. The modified cross-entropy loss function can be calculated as follows:

[0134]

[0135] Where L1 represents the modified cross-entropy loss function; wk Indicates the preset weights; M k Let K represent the maximum number of samples in one class, and M represent the maximum number of samples in the other class; y i k represents the sample label.

[0136] The following describes the transaction risk identification process of the present invention in detail, based on the transaction risk identification model described above.

[0137] First, the trading risk identification model is pre-trained using the following method to obtain a trained trading risk identification model:

[0138] (1) Collect historical credit purchase business information and construct a credit purchase business information sample.

[0139] Information on credit-based mobile phone purchase services includes information on installment loan products, installment loan methods, end-user identity information, and end-user carrier data. The carrier data here is mainly used to reflect the end-user's creditworthiness.

[0140] (2) Input the credit purchase business information sample into the transaction risk identification model to train the model.

[0141] After the credit-based mobile phone purchase information sample is input into the transaction risk identification model, the feature extraction network of the transaction risk identification model first extracts the initial feature information of the credit-based mobile phone purchase information sample. Then, the extracted initial feature information is output to the fine feature extraction and selection network.

[0142] (3) After receiving the initial feature information, the fine-grained feature extraction and selection network further extracts fine features from the initial features through four feature extraction and selection branches composed of its 16 fine-grained feature extraction and selection blocks. This allows for the gradual combination of simple features into higher-level features, and reduces reliance on specific details, thereby improving the generalization ability and robustness of the transaction risk identification model. The network also filters information from the fine-grained features to obtain more critical fine-grained features. These critical fine-grained features are then input into the transaction risk identification network.

[0143] (4) After receiving key fine features, the transaction risk identification network first uses an improved one-dimensional feature vector converter to feed the generated vector sequence into the encoder. The decoder recreates the risky transaction samples based on the risk-free transaction data represented in the received key fine features.

[0144] The method for recreating risky transaction samples based on risk-free transaction data is as follows:

[0145] A. Based on the identity information in the credit purchase business information of end users with no transaction risk, query the fuzzy average return, risk-free rate of return, and standard deviation of the fuzzy return of the main income sources of the end users during the credit purchase period.

[0146] B. Based on the fuzzy average return, risk-free rate of return (which can be determined based on the bank interest rate during this period), and standard deviation of the fuzzy returns of the main income sources for each end user without transaction risk, as well as credit installment product information and credit installment methods, the overdue risk of the user's credit purchase business is calculated as follows (since end users without transaction risk are selected, the calculated value is usually less than or equal to 0):

[0147]

[0148] Where η represents the overdue risk of credit-based mobile phone purchases by end users with no transaction risk; S represents the total value of the installment loan product; N represents the number of installment periods for the installment loan product; E(R) p R represents the fuzzy average return for end users without transaction risk; f This represents the risk-free rate of return; The standard deviation of fuzzy returns representing the main source of income;

[0149] C. Based on the overdue risk of the credit-based mobile phone purchase business for each end user without transaction risk as determined in step B above, adjust the overdue risk of the credit-based mobile phone purchase business for end users without transaction risk as follows:

[0150]

[0151] Where ζ represents the adjusted overdue risk of credit-based mobile phone purchases; n represents the actual installment days for end users without transaction risk; M 2 This represents the risk premium after adjusting for total overdue risk.

[0152] D. Based on the adjusted overdue risk of credit purchase transactions by end users without transaction risk and the credit purchase transaction information of end users without transaction risk, a risky transaction sample is constructed.

[0153] (5) Input the credit purchase business information sample and the newly constructed risk transaction sample into the detector to identify the risk transaction.

[0154] (6) Based on the risk transaction identification results of the discriminator and the actual situation of the credit purchase business information sample and the newly constructed risk transaction sample, update the update generator and discriminator network to obtain the trained transaction risk identification model.

[0155] Secondly, the risk of credit-based mobile phone purchase is identified through a well-trained risk identification model.

[0156] The identification process here is actually roughly the same as the process described above of training a transaction risk identification model using credit purchase business information samples. The only difference is that after the transaction risk identification model is trained, the transaction risk identification network, after receiving the new key fine features of the credit purchase business output by the fine feature extraction and selection network, directly performs classification and identification based on the classifier of the risk identification model, without needing to construct new samples as described in the steps above.

[0157] In addition, refer to Figure 9 This invention also proposes a transaction risk identification device, which includes:

[0158] Input module 10 is used to acquire transaction requests and input the transaction requests into a transaction risk identification model, wherein the transaction risk identification model includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network;

[0159] Extraction module 20 is used to extract initial features from the transaction request through the feature extraction network, and extract key fine features from the initial features through the fine feature extraction and selection network;

[0160] The identification module 30 is used to identify the risk of the transaction request based on the key fine features through the transaction risk identification network, and obtain the transaction risk identification result.

[0161] In this embodiment, a method for acquiring transaction requests and inputting them into a transaction risk identification model is disclosed. The transaction risk identification model includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network. The feature extraction network extracts initial features from the transaction request, and the fine feature extraction and selection network extracts key fine features from the initial features. The transaction risk identification network identifies the risk of the transaction request based on the key fine features to obtain the transaction risk identification result. Because this embodiment improves the network structure of the transaction risk identification model by designing a two-layer feature extraction network, the accuracy of transaction risk identification can be improved, thereby reducing transaction risk.

[0162] In one embodiment, the fine feature extraction and selection network includes multiple fine feature extraction and selection blocks, each with a different number of neurons. The fine feature extraction and selection blocks form multiple feature extraction and selection branches. The extraction module 20 is further configured to extract key fine features from the initial features through the feature extraction and selection branches. In the feature extraction and selection branches, different operations are performed on the output data of different fine feature extraction and selection blocks. The different operations include at least one of concatenation, averaging, convolution, and information transmission.

[0163] In one embodiment, the feature extraction network includes multiple stacked residual blocks, each residual block having a different number of convolutional channels. Each residual block includes a deep feature extraction branch and a basic feature extraction branch. The extraction module 20 is further configured to extract deep features from the transaction request through the deep feature extraction branch, wherein the deep feature extraction branch uses rectified linear units as activation functions; extract basic features from the transaction request through the basic feature extraction branch; and combine the deep features with the basic features to obtain initial features.

[0164] In one embodiment, the transaction risk identification network includes an encoder, a generator, a discriminator, and a classifier; the transaction risk identification device further includes:

[0165] The training module is used to acquire business information samples and input them into a transaction risk identification model; extract initial features from the business information samples through the feature extraction network, and extract key fine features from the initial features through the fine feature extraction and selection network; encode the key fine features through the encoder, and recreate risk-free transaction data from the encoded features through the generator to obtain risky transaction samples; input the business information samples and the risky transaction samples into the discriminator for risky transaction identification to obtain risky transaction identification results; and train the transaction risk identification model based on the risky transaction identification results, the business information samples, and the actual situation corresponding to the risky transaction samples to obtain the trained transaction risk identification model.

[0166] Accordingly, the identification module 30 is also used to identify the risk of the transaction request based on the key fine features by the classifier in the trained transaction risk identification model, and obtain the transaction risk identification result.

[0167] In one embodiment, the training module is further configured to: identify end users without transaction risk based on risk-free transaction data in the encoded features; obtain user information of the end users; query the fuzzy average return, risk-free rate of return, and standard deviation of the fuzzy return of the main revenue sources of the end users during the business period based on the user information; determine the business delinquency risk of the end users based on the fuzzy average return, the risk-free rate of return, and the standard deviation of the fuzzy return of the main revenue sources; adjust the delinquency risk of the credit purchase business of the end users based on the risk-free rate of return, the actual number of business days of the end users, and the risk premium; and construct risky transaction samples based on the adjusted business delinquency risk and the business information of the end users.

[0168] In one embodiment, the training module is further configured to calculate a modified cross-entropy loss function based on the risk transaction identification result, the business information sample, and the actual situation corresponding to the risk transaction sample. In the modified cross-entropy loss function, the weight value of the sample category is inversely proportional to the number of samples in the sample category. The transaction risk identification model is trained based on the modified cross-entropy loss function to obtain the trained transaction risk identification model.

[0169] Other embodiments or specific implementations of the transaction risk identification device described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0170] Furthermore, this embodiment of the invention also proposes a storage medium storing a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method as described above.

[0171] Furthermore, this invention also proposes a computer program product, including a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method described above.

[0172] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-described transaction risk identification method, and will not be repeated here.

[0173] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0174] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0175] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0176] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for identifying transaction risks, characterized in that, The method for identifying transaction risks includes: A transaction request is obtained and input into a transaction risk identification model, which includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network. Initial features are extracted from the transaction request through the feature extraction network, and key fine features are extracted from the initial features through the fine feature extraction and selection network. The transaction risk identification network identifies the risks of the transaction request based on the key fine features, and obtains the transaction risk identification result.

2. The transaction risk identification method as described in claim 1, characterized in that, The fine-grained feature extraction and selection network includes multiple fine-grained feature extraction and selection blocks, each with a different number of neurons. These blocks form multiple feature extraction and selection branches. The extraction of key fine-grained features from the initial features through the fine-grained feature extraction and selection network includes: Key fine features are extracted from the initial features through the feature extraction and selection branch, wherein different operations are performed on the output data of different fine feature extraction selection blocks in the feature extraction and selection branch, and the different operations include at least one of concatenation operation, averaging operation, convolution operation, and information transmission.

3. The transaction risk identification method as described in claim 1, characterized in that, The feature extraction network comprises multiple stacked residual blocks, each with a different number of convolutional channels. Each residual block includes a deep feature extraction branch and a basic feature extraction branch. Extracting initial features from the transaction request using the feature extraction network includes: Deep features are extracted from the transaction request through the deep feature extraction branch, wherein the rectified linear unit is used as the activation function in the deep feature extraction branch; Basic features are extracted from the transaction request through the basic feature extraction branch; The deep features are combined with the basic features to obtain the initial features.

4. The transaction risk identification method as described in any one of claims 1 to 3, characterized in that, The transaction risk identification network includes an encoder, a generator, a discriminator, and a classifier. Before acquiring the transaction request and inputting it into the transaction risk identification model, the network further includes: Obtain business information samples and input them into the transaction risk identification model; Initial features are extracted from the business information sample through the feature extraction network, and key fine features are extracted from the initial features through the fine feature extraction and selection network. The encoder encodes the key fine features, and the generator recreates the risk-free transaction data in the encoded features to obtain risky transaction samples. The business information sample and the risk transaction sample are input into the discriminator to identify the risk transaction and obtain the risk transaction identification result. The transaction risk identification model is trained based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples to obtain the trained transaction risk identification model. Accordingly, the step of using the transaction risk identification network to identify the transaction request based on the key refined features and obtaining the transaction risk identification result includes: The classifier in the trained transaction risk identification model identifies the risk of the transaction request based on the key fine features, and obtains the transaction risk identification result.

5. The transaction risk identification method as described in claim 4, characterized in that, The process of recreating risk-free transaction data from the encoded features to obtain risky transaction samples includes: Based on the risk-free transaction data in the encoded features, identify end users without transaction risk and obtain the user information of the end users. Based on the user information, query the fuzzy average return, risk-free rate of return, and standard deviation of the fuzzy return of the main revenue sources of the terminal user during the business period; The business delinquency risk of the end user is determined based on the fuzzy average return, the risk-free rate of return, and the standard deviation of the fuzzy returns of the main revenue source; The risk of overdue credit purchase transactions of the end users is adjusted based on the risk-free rate of return, the actual number of business days of the end users, and the risk premium. A risk transaction sample is constructed based on the adjusted business delinquency risk and the business information of the end users.

6. The transaction risk identification method as described in claim 4, characterized in that, The step of training the transaction risk identification model based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples to obtain the trained transaction risk identification model includes: The modified cross-entropy loss function is calculated based on the risk transaction identification results, the business information samples, and the actual situation corresponding to the risk transaction samples. In the modified cross-entropy loss function, the weight value of the sample category is inversely proportional to the number of samples of the sample category. The trading risk identification model is trained based on the modified cross-entropy loss function to obtain the trained trading risk identification model.

7. A transaction risk identification device, characterized in that, The transaction risk identification device includes: An input module is used to acquire transaction requests and input the transaction requests into a transaction risk identification model, which includes a feature extraction network, a fine feature extraction and selection network, and a transaction risk identification network. The extraction module is used to extract initial features from the transaction request through the feature extraction network, and to extract key fine features from the initial features through the fine feature extraction and selection network; The identification module is used to identify the risk of the transaction request based on the key fine features through the transaction risk identification network, and obtain the transaction risk identification result.

8. A transaction risk identification device, characterized in that, The transaction risk identification device includes: a memory, a processor, and a transaction risk identification program stored in the memory and executable on the processor. When the transaction risk identification program is executed by the processor, it implements the transaction risk identification method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a transaction risk identification program, which, when executed by a processor, implements the transaction risk identification method as described in any one of claims 1 to 6.

Images