Abnormal object recognition method, device and equipment, computer program product and computer readable storage medium

Abstract

The invention provides an abnormal object identification method, device and equipment, a computer program product and a computer readable storage medium. The method comprises the steps that electronic payment record data of a target object are acquired, and the electronic payment record data comprise payment variables of multiple dimensions; obtaining a first vector representation and a first feature weight respectively corresponding to the payment variables of the multiple dimensions; performing feature combination based on the first vector representation and the first feature weight corresponding to the payment variables of the multiple dimensions to obtain a first feature combination result; performing feature extraction on the first feature combination result to obtain a second feature combination result; performing first linear transformation processing on the second feature combination result to obtain a linear transformation processing result; and performing nonlinear transformation processing on a linear transformation processing result to obtain a probability that the target object is an abnormal object. Through the method and the device, the feature combination and representation capability can be enhanced, and the accuracy of identifying the abnormal object is improved.

Description

Technical Field

[0001] This application relates to the technical field of electronic payment, and in particular, to a method, apparatus, device, computer program product, and computer-readable storage medium for identifying abnormal objects. Background Art

[0002] In the related art, there have been various technical solutions for identifying abnormal objects in the electronic payment scenario. For example, in the electronic payment scenario, objects performing illegal acts (such as money laundering, etc.) are intercepted to achieve anti-money laundering, anti-fraud, etc. operations, so as to perform financial risk control. These technical solutions are usually based on traditional statistical methods or machine learning algorithms, such as logistic regression, decision tree, random forest, etc., and usually use a set of manually selected features for classification. However, the ability of this feature combination is limited and cannot accurately represent the complex associations between features in the electronic payment scenario. Therefore, there is a problem of low accuracy in identifying abnormal objects. Summary of the Invention

[0003] Embodiments of this application provide a method, apparatus, device, computer program product, and computer-readable storage medium for identifying abnormal objects, which can enhance the combination and representation ability of features and improve the accuracy of identifying abnormal objects.

[0004] The technical solution of the embodiments of this application is implemented as follows:

[0005] Embodiments of this application provide a method for identifying abnormal objects, the method includes:

[0006] Obtain the electronic payment record data of a target object, where the electronic payment record data includes payment variables in multiple dimensions;

[0007] Obtain the first vector representation and the first feature weight corresponding to each of the payment variables in the multiple dimensions;

[0008] Perform feature combination based on the first vector representation and the first feature weight corresponding to each of the payment variables in the multiple dimensions to obtain a first feature combination result;

[0009] Perform feature extraction on the first feature combination result to obtain a second feature combination result;

[0010] Perform a first linear transformation process on the second feature combination result to obtain a linear transformation process result;

[0011] Perform a non-linear transformation process on the linear transformation process result to obtain the probability that the target object is an abnormal object.

[0012] Embodiments of this application provide an apparatus for identifying abnormal objects, including:

[0013] An acquisition module, configured to acquire electronic payment record data of a target object, where the electronic payment record data includes payment variables in multiple dimensions;

[0014] A processing module, configured to acquire first vector representations and first feature weights respectively corresponding to the payment variables in the multiple dimensions;

[0015] The processing module is further configured to perform feature combination based on the first vector representations and the first feature weights respectively corresponding to the payment variables in the multiple dimensions to obtain a first feature combination result;

[0016] The processing module is further configured to perform feature extraction on the first feature combination result to obtain a second feature combination result;

[0017] A generation module, configured to perform a first linear transformation process on the second feature combination result to obtain a linear transformation process result;

[0018] The generation module is further configured to perform a non - linear transformation process on the linear transformation process result to obtain the probability that the target object is an abnormal object.

[0019] An embodiment of the present application provides an electronic device, including:

[0020] A memory, configured to store computer - executable instructions;

[0021] A processor, configured to implement the abnormal object recognition method provided by the embodiment of the present application when executing the computer - executable instructions stored in the memory.

[0022] An embodiment of the present application provides a computer - readable storage medium, storing a computer program or computer - executable instructions, which are configured to implement the abnormal object recognition method provided by the embodiment of the present application when being executed by a processor.

[0023] An embodiment of the present application provides a computer program product, including a computer program or computer - executable instructions, where when the computer program or computer - executable instructions are executed by a processor, the abnormal object recognition method provided by the embodiment of the present application is implemented.

[0024] The embodiments of the present application have the following beneficial effects:

[0025] Through the feature combination processing based on payment variables in multiple dimensions, the automatic selection of core features is realized in the electronic payment scenario. Compared with the related technology where feature extraction is carried out through manual feature engineering, it is difficult to make full use of all available data and potential key features are easily overlooked. In the embodiments of the present application, through selective feature extraction from the first feature combination result, further feature extraction is realized on the basis of the first feature combination result, thereby enhancing the depth representation ability and feature combination ability of the model to better capture the complexity of payment variables in multiple dimensions of the target object. The probability of an abnormal object (such as an object engaging in illegal activities such as money laundering and fraud) obtained through the linear transformation and non-linear transformation processing of the second feature combination result can accurately reflect the abnormal degree of the target object, achieving the beneficial effect of improving the recognition accuracy of abnormal objects and providing accurate reference data for financial risk control in the electronic payment scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 FIG. is a schematic structural diagram of an abnormal object recognition system architecture provided by an embodiment of the present application;

[0027] Figure 2 FIG. is a schematic structural diagram of a server provided by an embodiment of the present application;

[0028] Figure 3 FIG. is a schematic diagram of the structure of an abnormal object recognition model provided by an embodiment of the present application;

[0029] Figure 4A FIG. is a first flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0030] Figure 4B FIG. is a second flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0031] Figure 4C FIG. is a third flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0032] Figure 4D FIG. is a fourth flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0033] Figure 4E FIG. is a fifth flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0034] Figure 4F FIG. is a sixth flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0035] Figure 4G FIG. is a seventh flowchart of an abnormal object recognition method provided by an embodiment of the present application;

[0036] Figure 4H It is the eighth process schematic diagram of the abnormal object recognition method provided by the embodiments of the present application;

[0037] Figure 4I It is the ninth process schematic diagram of the abnormal object recognition method provided by the embodiments of the present application;

[0038] Figure 4J It is the tenth process schematic diagram of the abnormal object recognition method provided by the embodiments of the present application;

[0039] Figure 5 It is the interaction process schematic diagram of the abnormal object recognition method provided by the embodiments of the present application;

[0040] Figure 6 It is the schematic diagram of the structure of the feature selection network provided by the embodiments of the present application;

[0041] Figure 7 It is the schematic diagram of the structure of the gated residual network provided by the embodiments of the present application;

[0042] Figure 8 It is the schematic diagram of the structure of the deep neural network provided by the embodiments of the present application. Detailed implementation manners

[0043] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application.

[0044] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments. However, it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0045] In the following description, the terms "first / second / third" are only used to distinguish similar objects and do not represent a specific order for the objects. It can be understood that "first / second / third" can be interchanged with a specific order or sequence when allowed, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0046] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of that module or unit.

[0047] In the embodiments of the present application, when collecting and processing relevant data (such as electronic payment record data) in an example application, it should strictly comply with the requirements of relevant national laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing behaviors within the scope authorized by laws and regulations and the personal information subject.

[0048] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as commonly understood by those skilled in the technical field to which the present application belongs. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0049] Before further elaborating on the embodiments of the present application, the nouns and terms involved in the embodiments of the present application are described. The nouns and terms involved in the embodiments of the present application are applicable to the following explanations.

[0050] 1) Electronic payment, a mobile payment service that allows users to make payments and transfers through instant messaging applications.

[0051] 2) Social payment scenario refers to payment activities carried out in social media or communication applications, such as payments between friends and sending red envelopes.

[0052] 3) Variable Selection Network (VSN), which is used to select the most relevant input features.

[0053] 4) Gated Residual Network (GRN), a neural network structure used to enhance the expressive power of the model.

[0054] 5) Deep Neural Networks (DNN), a machine learning model used to handle complex non-linear data modeling tasks.

[0055] 6) Dense Layer, also known as the fully connected layer, is a common layer type in neural networks. It establishes a full connection relationship between the previous layer and the next layer, and each node is connected to all nodes of the previous layer.

[0056] The abnormal object recognition method in the related art, for example, intercepting objects that perform illegal acts (such as money laundering, etc.) in the electronic payment scenario to implement anti-money laundering, anti-fraud, etc. operations, thereby performing financial risk control, usually relies on developers to manually select features, it is difficult to make full use of all available data, it is easy to ignore potential key features, and there is a problem of feature selection limitation. In addition, the expression capabilities of the machine learning method and the simple deep neural network model in the related art are limited, it is difficult to process complex non-linear relationships, the model lacks interpretability, and it is difficult to understand the decision basis of the model, resulting in a low recognition accuracy of abnormal objects.

[0057] The embodiments of the present application provide a method, device, equipment, computer-readable storage medium and computer program product for recognizing abnormal objects, which can enhance the combination and representation capabilities of features and improve the accuracy of recognizing abnormal objects.

[0058] The equipment provided by the embodiments of the present application can be implemented as various types of user terminals such as laptop computers, tablet computers, desktop computers, set-top boxes, mobile devices (for example, mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable game devices), smart phones, smart speakers, smart watches, smart TVs, in-vehicle terminals, etc., or can also be implemented as a server. Below, the exemplary application when the equipment is implemented as a server will be described.

[0059] See Figure 1 , Figure 1 is a schematic structural diagram of the abnormal object recognition system architecture provided by the embodiments of the present application. By way of example, Figure 1 involves a server 100, a terminal device 200 and a network 300. The terminal device 200 is connected to the server 100 through the network 300, where the network 300 can be a wide area network or a local area network, or a combination of the two.

[0060] In some embodiments, the abnormal object recognition system provided by the embodiments of the present application can be implemented collaboratively by a server and a terminal. For example, the terminal device 200 runs a client, such as an instant messaging client (including an electronic payment function) or an electronic payment client, to support a user in making electronic payments, such as transferring money, online shopping, etc. When the client receives an electronic payment operation of a target object (for example, it can be the user account currently logged in the instant messaging client), the terminal device 200 sends the electronic payment record data of the target object to the server 100. The server 100 receives the electronic payment record data including payment variables in multiple dimensions, and obtains the probability or label that the target object is an abnormal object through the abnormal object recognition method provided by the embodiments of the present application as the recognition result. If the target object is an abnormal object, secondary verification is performed on the target object (for example, the target object is secondarily verified by obtaining a payment password, biometric verification, etc., and after the verification is passed, the electronic payment request of the target object is processed) or the payment request of the target object is rejected, etc. If it is not an abnormal object, the terminal device 200 sends an electronic payment request to the server 100 so that the server 100 processes the electronic payment request.

[0061] In some embodiments, the terminal device or the server can implement the abnormal object recognition method provided by the embodiments of the present application by running various computer-executable instructions or computer programs. For example, the computer-executable instructions can be commands at the microprogram level, machine instructions, or software instructions. The computer program can be a native program or software module in an operating system; it can be a local (Native) application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as an application program with an electronic payment function; it can also be a game applet embedded in any APP, that is, a program that only needs to be downloaded to the browser environment to run. In short, the above computer-executable instructions can be instructions in any form, and the above computer programs can be application programs, modules, or plug-ins in any form.

[0062] In some embodiments, the server 100 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms. Among them, the cloud service can be an interactive processing service for the terminal to call.

[0063] In some embodiments, multiple servers can be grouped into a blockchain network, and server 100 is a node on the blockchain network. There can be information connections between each node in the blockchain network, and information can be transmitted between nodes through the above-mentioned information connections. Among them, the data related to the method for identifying abnormal objects provided in the embodiments of the present application can be stored on the blockchain.

[0064] The embodiments of the present application can be implemented with the help of Artificial Intelligence (AI) technology. Artificial Intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, Artificial Intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial Intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machines to have the functions of perception, reasoning, and decision-making.

[0065] Artificial Intelligence technology is an interdisciplinary subject, involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of Artificial Intelligence generally include sensors, dedicated Artificial Intelligence chips, cloud computing, distributed storage, big data processing technology, pre-trained model technology, operation / interaction systems, mechatronics, etc. Among them, the pre-trained model, also known as the large model or the foundation model, can be widely applied to downstream tasks in various major directions of Artificial Intelligence after fine-tuning. The software technologies of Artificial Intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0066] Taking the above-mentioned server 100 for identifying abnormal objects as an example, refer to Figure 2 , Figure 2 which is a schematic structural diagram of the server provided by the embodiments of the present application. Figure 2 The server 100 shown includes at least one processor 110, a memory 130, and at least one network interface 120. Each component in the server 100 is coupled together through a bus system 140. It can be understood that the bus system 140 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 140 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clear illustration, in Figure 2 all kinds of buses are labeled as the bus system 140.

[0067] The processor 110 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or any conventional processor, etc.

[0068] The memory 130 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disc drives, etc. Optionally, the memory 130 includes one or more storage devices that are physically located far from the processor 110.

[0069] The memory 130 includes volatile memory or non-volatile memory, and may also include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 130 described in the embodiments of the present application is intended to include any suitable type of memory.

[0070] In some embodiments, the memory 130 is capable of storing data to support various operations. Examples of such data include programs, modules, and data structures, or subsets or supersets thereof, which are illustrated below.

[0071] The operating system 131 includes system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, the core library layer, the driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0072] The network communication module 132 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 120. Exemplary network interfaces 120-1 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.;

[0073] In some embodiments, the device provided by the embodiments of the present application may be implemented in software. Figure 2 Shows an identification device 133 for abnormal objects stored in the memory 130, which may be software in the form of programs and plugins, etc., including the following software modules: an acquisition module 1331, a processing module 1332, and a generation module 1333. These modules are logical, so they can be combined arbitrarily or further split according to the functions implemented. The functions of each module will be described below.

[0074] In some other embodiments, the device provided in the embodiments of the present application may be implemented in a hardware manner. As an example, the device provided in the embodiments of the present application may be a processor in the form of a hardware decoding processor, which is programmed to execute the method for identifying an abnormal object provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor may employ one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), or other electronic components.

[0075] Next, in combination with the exemplary applications and implementations of the server provided in the embodiments of the present application, taking the server as the execution subject, the method for identifying an abnormal object provided in the embodiments of the present application will be described.

[0076] See Figure 3 , Figure 3 which is a schematic diagram of the structure of the abnormal object recognition model provided in the embodiments of the present application. The method for identifying an abnormal object provided in the embodiments of the present application is implemented through the abnormal object recognition model. The abnormal object recognition model may include a feature selection network (VSN), a gated residual network (GRN), and a deep neural network (DNN). After inputting the electronic payment record data of the target object into the abnormal object recognition model, the method for identifying an abnormal object provided in the embodiments of the present application is processed through the above network structure. Finally, the output result of the abnormal object recognition model is the probability or label that the target object is an abnormal object. Among them, the feature selection network can be stacked multiple times to mine high-order features. For example Figure 3 shows 3 groups of feature selection networks stacked (feature selection network VSN-1, feature selection network VSN-2, and feature selection network VSN-3). The embodiments of the present application do not limit the stacking combination times of each network in the abnormal object recognition model. The principle of the object recognition model shown below will be described in combination with a specific flowchart. Figure 3 shows the principle of the object recognition model.

[0077] See Figure 4A , Figure 4A which is the first flowchart of the method for identifying an abnormal object provided in the embodiments of the present application. It will be described in combination with the steps shown in Figure 4A below.

[0078] In step 101, obtain the electronic payment record data of the target object, where the electronic payment record data includes payment variables in multiple dimensions.

[0079] In some embodiments, the electronic payment record data includes payment variables in multiple dimensions, and the payment variables in multiple dimensions include, but are not limited to, the age range of the target object, account type, transaction method, transaction amount, transaction time interval, transaction amount change rate, etc.

[0080] As an example, the target object can be the user account currently logged in to the instant messaging client.

[0081] In some embodiments, before proceeding to step 102, the payment variables in multiple dimensions included in the electronic payment record data can also be divided into categorical features and numerical features, and data preprocessing operations are performed separately. For example, the payment variables are divided into categorical features (such as the age range of the target object, account type, etc.) and numerical features (such as transaction time interval, transaction amount, transaction amount change rate, etc.). One-Hot Encoding is performed on the categorical features (for example, by obtaining all the categories of the categorical features, creating a zero matrix of size [number of samples, number of categories], traversing the categorical features to determine the category of each feature, and setting the corresponding position in the zero matrix to 1. For example, when the One-Hot Encoding result is a two-dimensional array, each row represents a categorical feature, each column represents a category, and if a certain categorical feature belongs to a certain category, the corresponding position is 1, otherwise 0), so as to perform classification feature representation. For the numerical features, linear projection representation of the features is performed through a Dense Layer (see the second linear transformation process in step 1021 below). By performing One-Hot Encoding on the categorical features, each category is converted into a binary feature, which can better represent the relationship between categories, increase the interpretability of the features. By performing linear mapping on the numerical features, the combination and interaction relationship between features can be captured, the non-linear fitting ability can be improved, and the dimension of the features can be reduced.

[0082] In step 102, obtain the first vector representation and the first feature weight corresponding to each of the payment variables in multiple dimensions.

[0083] In some embodiments, obtaining the first vector representation and the first feature weight corresponding to each of the payment variables in multiple dimensions can be implemented through a feature selection network. See Figure 6 , Figure 6 is a schematic diagram of the structure of the feature selection network (VSN) provided by the embodiments of the present application. VSN is a structure composed of N (here, the value of N is equal to the number of dimensions of the payment variables) Dense layers and a Gated Residual Network (GRN). For the electronic payment record data of each target object, there are N-dimensional features. As Figure 6As shown in the left structure in the figure, for each dimension of features {x1, …, x n}, first, a dense layer is used for linear projection of the features, and then the linear projection results are input into the feature selection network for combined mapping to generate the first vector representations {y1, …, y n} of each dimension. At the same time, GRN can be applied to the concatenation result of all the feature linear projections (the output result of the Dense layer), that is, the concatenation result of the first vector representations of multiple dimensions, and then a normalization function (such as softmax) is used for normalization processing to obtain the first feature weight V of each dimension.

[0084] Here, the first vector representation is the representation of the payment variable in a unified vector space, and the first feature weight characterizes the importance degree of the first vector representation of the corresponding dimension.

[0085] In some embodiments, referring to Figure 4B , Figure 4A shown in step 102, it can be implemented through the following steps 1021 to 1023, which are specifically described below.

[0086] In step 1021, second linear transformation processing is respectively performed on the payment variables of multiple dimensions to obtain the first linear payment variable of each dimension.

[0087] In some embodiments, second linear transformation processing is respectively performed on the payment variables of multiple dimensions through a dense layer, that is, in the dense layer, the payment variables of multiple dimensions are respectively linearly transformed through the weight matrix of the dense layer and a bias term is added, and then a non-linear transformation is performed through an activation function (such as sigmoid, ReLU, etc.). Finally, the first linear payment variable of each dimension is obtained. After the second linear transformation processing of the dense layer, the payment variables of multiple dimensions can obtain a new dimension (for example, if the input is an N-dimensional payment variable, after the second linear transformation through a dense layer with M neurons, the output will be an M-dimensional vector), so that the first linear payment variable has a higher-level feature combination ability. By using an activation function after the linear transformation, the first linear payment variable has a certain non-linear expression ability, thus achieving the beneficial effect of better learning the features of the input payment variables of multiple dimensions.

[0088] As an example, the second linear transformation processing can be represented by the following formula:

[0089] Dense 2nd = WX1 + b (1)

[0090] where, Dense 2nd represents the second linear transformation processing, X1 = {x1, …, x n}{represents the payment variables of multiple dimensions of the input, n represents the number of dimensions of the payment variables, W represents the weight matrix of the dense layer, and b represents the bias term. For the sake of convenient expression, all the weight matrices and bias terms of the dense layers in the text are uniformly represented as W and b here. During the training process of the abnormal object recognition model, the weight matrix and bias term of the dense layer are continuously optimized and updated. During the network inference process after the training of the abnormal object recognition model is completed, the parameters of the weight matrix and bias term of the dense layer are fixed and will not be elaborated further below.

[0091] Through step 1021, it realizes the number of dimensions of the numerical features in the payment variables of multiple dimensions specified through the dense layer (see the description in step 101) after being processed by the second linear transformation, so as to perform the linear mapping of the payment variables of multiple dimensions while reducing the number of dimensions and achieve the beneficial effect of information extraction.

[0092] In step 1022, the first linear payment variables of each dimension are subjected to a combined mapping process to obtain the first vector representation corresponding to each dimension.

[0093] In some embodiments, the combined mapping process of the first linear payment variables of each dimension to obtain the first vector representation corresponding to each dimension can be implemented through a gated residual network. See Figure 7 , Figure 7 is a schematic diagram of the structure of the gated residual network provided by the embodiment of the present application. The gated residual network mainly consists of a Dense layer, an ELU activation function (corresponding to the first activation function below), and a gated linear unit GLU (corresponding to the first gated linear model below).

[0094] In some embodiments, see Figure 4C , Figure 4B shown in step 1022, can be implemented by performing the following steps 10221 to 10226 for the first linear payment variables of each dimension. The following is a specific description.

[0095] In step 10221, the first linear payment variables are subjected to a third linear transformation process to obtain the second linear payment variables corresponding to each dimension.

[0096] In some embodiments, the first linear payment variables of each dimension are subjected to a third linear transformation process through the dense layer to obtain the second linear payment variables corresponding to each dimension. As an example, the third linear transformation process can be represented by the following formula:

[0097] Dense 3rd = WX2 + b (2)

[0098] where, Dense 3rdIt represents the third linear transformation process. X2 represents the first linear payment variable of multiple dimensions of the input, W represents the weight matrix of the dense layer, and b represents the bias term.

[0099] In step 10222, the second linear payment variable is subjected to a first activation process through a preset first activation function to obtain the non-linear payment variable corresponding to each dimension.

[0100] In some embodiments, the second linear payment variable is subjected to a first activation process through a preset first activation function (such as the Exponential Linear Unit activation function (ELU)), that is, the second linear payment variable is subjected to a non-linear transformation through the first activation function to obtain the non-linear payment variable corresponding to each dimension. Here, the first activation process can be represented by the following formula:

[0101]

[0102] where ELU 1st represents the first activation process, X3 represents the input second linear payment variable, α is a hyperparameter used to control the slope of the negative region. When the input X3 is greater than or equal to 0, the output of the ELU activation function is equal to the input X3 itself, which maintains the linear characteristic and enables the neural network to learn linear relationships. When the input X3 is less than 0, the output of the ELU activation function is equal to This part of the negative region introduces non-linear characteristics, enabling the neural network to learn non-linear relationships. For the convenience of description, all the hyperparameters of the first activation function in the text are uniformly represented as α here, and will not be elaborated below.

[0103] In step 10223, the non-linear payment variable is subjected to a fourth linear transformation process to obtain the third linear payment variable corresponding to each dimension.

[0104] In some embodiments, the non-linear payment variable is subjected to a fourth linear transformation process through the dense layer to obtain the third linear payment variable corresponding to each dimension. As an example, the fourth linear transformation process can be represented by the following formula:

[0105] Dense 4th = WX4 + b (4)

[0106] where Dense 4th represents the fourth linear transformation process, X4 represents the input non-linear payment variable of multiple dimensions, W represents the weight matrix of the dense layer, and b represents the bias term.

[0107] In step 10224, the third linear payment variable is subjected to a first gated linear process to obtain the linear combination vector corresponding to the dimension.

[0108] In some embodiments, the first gated linear processing is implemented by a first gated linear model, which includes a first convolutional layer, a second convolutional layer, and a second activation function (e.g., sigmoid activation function).

[0109] In some embodiments, referring to Figure 4D , Figure 4C as shown in step 10224, it can be implemented through the following steps 102241 to 102244, which will be specifically described below.

[0110] In step 102241, the third linear payment variable is subjected to a first convolution process by the first convolutional layer to obtain a first convolutional payment vector.

[0111] In some embodiments, the first gated linear processing of the Gated Linear Unit (GLU) can be represented by the following formula:

[0112] GLU 1st = δ(W1X5 + b1) * (W2X5 + b2) (5)

[0113] Wherein, X5 represents the input third linear payment variable, W1 and W2 respectively represent the weight matrices of the first convolutional layer and the second convolutional layer, b1 and b2 respectively represent the biases of the first convolutional layer and the second convolutional layer, δ is the activation function (e.g., sigmoid), and * represents the multiplication of elements at corresponding positions (here referring to the multiplication of elements at corresponding positions of the two same-order matrices δ(W1x + b1) and (W2x + b2)), also known as the Hadamard product. As can be seen from formula (5), first, X5 is input into the first convolutional layer and the second convolutional layer to obtain two outputs, then the output of the first convolutional layer is subjected to a non-linear transformation using the activation function δ, and then the corresponding GLU output is obtained through element-wise multiplication by the Hadamard product.

[0114] Here, the third linear payment variable is subjected to a first convolution process by the first convolutional layer to obtain a first convolutional payment vector, corresponding to (W1X5 + b1) in formula (5).

[0115] In step 102242, the first convolutional payment vector is subjected to a second activation process by the second activation function to obtain an importance parameter vector of the first convolutional payment vector. Here, the importance parameter vector is used to characterize the importance degree of the feature vectors of each dimension in the first convolutional payment vector.

[0116] Continuing with the above example, the second activation function is used to perform a second activation process on the first convolutional payment vector, obtaining the importance parameter vector of the first convolutional payment vector, corresponding to δ(W1X5 + b1) in formula (5), where δ represents the second activation function (such as sigmoid).

[0117] In step 102243, the second convolutional layer is used to perform a second convolutional process on the third linear payment variable, obtaining the second convolutional payment vector.

[0118] Continuing with the above example, the second convolutional layer is used to perform a second convolutional process on the third linear payment variable, obtaining the second convolutional payment vector, corresponding to (W2X5 + b2) in formula (5).

[0119] In step 102244, the second convolutional payment vector and the importance parameter vector of the first convolutional payment vector are multiplied to obtain a linear combination vector corresponding to the dimension.

[0120] Continuing with the above example, the second convolutional payment vector and the importance parameter vector of the first convolutional payment vector are multiplied to obtain a linear combination vector, corresponding to δ(W1X5 + b1)*(W2X5 + b2) in formula (5), where * represents the multiplication process.

[0121] Through steps 102241 to 102242, the gated linear model selectively learns and transmits information about the input features. The network can more flexibly adapt to different input situations and can decide whether to transmit information based on the importance of the input. After being trained in the following steps 109 to 116, the abnormal object recognition model incorporates the processing of selectively learning the input features by the gated linear model, making the abnormal object recognition model pay more attention to the key information in the input data and ignore some unimportant information. For example, when the weight (corresponding to the importance parameter vector) of a certain dimension is relatively large, the extraction of features in this dimension is emphasized, enabling the information contained in the features of this dimension to be transmitted to the next stage, thereby improving the performance and interpretability of the trained abnormal object recognition model.

[0122] Continue to refer to Figure 4C , in step 10225, the first linear payment variable of each dimension and the linear combination vector are added to obtain a gated residual vector corresponding to each dimension.

[0123] In some embodiments, the elements at the corresponding positions of the first linear payment variable of each dimension and the linear combination vector (here referring to the elements at the corresponding positions of the two matrices of the same order of the first linear payment variable and the linear combination vector) are added to obtain a gated residual vector corresponding to each dimension.

[0124] In step 10226, the gated residual vector is normalized to obtain the first vector representation corresponding to each dimension.

[0125] In some embodiments, the gated residual vector is normalized to obtain the first vector representation corresponding to each dimension. The normalization process (Norm) can be expressed by the following formula:

[0126]

[0127] where X6 represents the input gated residual vector, represents the expected value (mean) of X6, represents the variance of X6, and N represents the number of dimensions of the input gated residual vector. The meaning of formula (6) is to subtract the expected value of the input X6 from the input X6, then divide by the square root of the variance of the input X6, and finally multiply by the weight W norm1 and add the bias term b norm1 , so that the mean of the first vector representation corresponding to each dimension is 0 and the variance is 1, thus achieving the purpose of normalization. By normalization, the characteristic value ranges of different dimensions are unified into a smaller range, avoiding excessive differences between the characteristics of different dimensions, which helps the model to converge faster when training the abnormal object recognition model, reduces the training time, and improves the stability and robustness of the model.

[0128] Continue to refer to Figure 4B , in step 1023, the first vector representation corresponding to each dimension is normalized to obtain the first feature weight corresponding to each dimension.

[0129] In some embodiments, the first vector representation corresponding to each dimension can be normalized through the softmax function to obtain the first feature weight corresponding to each dimension. The softmax function can be expressed by the following formula:

[0130]

[0131] where x i ∈X7 = {x1,…,x M} represents the input first vector representation, and M represents the number of dimensions of the first vector representation.

[0132] In some embodiments, referring to Figure 6 , GRN can also be applied to the concatenation result of the first vector representation corresponding to each dimension, and then the normalization function (such as softmax) is used to generate the first feature weight V of each dimension, where the first feature weight is used to characterize the importance of the first vector representation corresponding to each dimension.

[0133] Continue to refer toFigure 4A In step 103, a first feature combination result is obtained by performing feature combination based on the first vector representations and the first feature weights respectively corresponding to the payment variables in multiple dimensions.

[0134] In some embodiments, referring to Figure 4E , Figure 4A step 103 shown, it can be implemented through the following steps 1031 to 1032, which will be specifically described below.

[0135] In step 1031, the first vector representation and the first feature weight corresponding to each dimension are multiplied to obtain a feature vector corresponding to each dimension.

[0136] In some embodiments, the corresponding position elements of the first vector representation and the first feature weight corresponding to each dimension are multiplied to obtain a feature vector corresponding to each dimension (corresponding to Figure 6 Z in).

[0137] In step 1032, the feature vectors corresponding to each dimension are combined to obtain a first feature combination result.

[0138] In some embodiments, the feature vectors corresponding to each dimension are subjected to splicing combination processing to obtain a first feature combination result (for example, the first feature combination result can be expressed as [feature vector 1,..., feature vector N], where N represents the number of dimensions).

[0139] Through steps 102 to 103, the key features in the payment variables of multiple dimensions are automatically selected, which reduces the burden of manual feature selection and combination. In the related art, traditional abnormal object recognition methods usually rely on manual feature selection, making it difficult to fully utilize all available data, easily overlooking potential key features, and having problems with feature selection limitations. The beneficial effects of simplifying the network, reducing the curse of dimensionality, improving the network calculation efficiency, and enhancing the feature combination ability are achieved. Here, automatically selecting the key features in the payment variables of multiple dimensions is a processing method of soft feature selection that weights the feature vectors through feature weights, that is, higher weights are assigned to important features, thereby realizing a soft selection. In some other embodiments, the features corresponding to the dimensions with feature weights lower than the preset weight threshold can also be removed by the method of presetting the weight threshold to achieve the selection of key features.

[0140] Continuing to refer to Figure 4A , in step 104, feature extraction is performed on the first feature combination result to obtain a second feature combination result.

[0141] In some embodiments, referring to Figure 4F , Figure 4AThe shown step 104 can be implemented through the following steps 1041 to 1046, which will be specifically described below.

[0142] In step 1041, perform a fifth linear transformation on the first feature combination result to obtain a first linear feature combination variable.

[0143] In some embodiments, perform a fifth linear transformation on the first feature combination result through a dense layer, that is, in the dense layer, the first feature combination results of multiple dimensions are respectively linearly transformed through the weight matrix of the dense layer, and a bias term is added, and then a non-linear transformation is performed through an activation function (such as sigmoid, ReLU, etc.). Finally, a first linear feature combination variable is obtained. As an example, the fifth linear transformation can be expressed by the following formula:

[0144] Dense 5th = WX8 + b (8)

[0145] Where, Dense 5th represents the fifth linear transformation, X8 represents the first feature combination results of multiple input dimensions, W represents the weight matrix of the dense layer, and b represents the bias term.

[0146] In step 1042, perform a third activation on the first linear feature combination variable through a preset first activation function to obtain a non-linear feature combination variable.

[0147] In some embodiments, perform a third activation on the first linear feature combination variable through a preset first activation function (such as ELU), that is, perform a non-linear transformation on the first linear feature combination variable through the first activation function to obtain a non-linear feature combination variable corresponding to each dimension. Here, the third activation can be expressed by the following formula:

[0148]

[0149] Where, ELU 3rd represents the third activation, X9 represents the input first linear feature combination variable, α is a hyperparameter used to control the slope of the negative region. When the input X9 is greater than or equal to 0, the output of the ELU activation function is equal to the input X9 itself, which maintains the linear characteristic and enables the neural network to learn linear relationships. When the input X9 is less than 0, the output of the ELU activation function is equal to This introduction of non-linear characteristics in this part of the negative region enables the neural network to learn non-linear relationships.

[0150] In step 1043, perform a sixth linear transformation on the non-linear feature combination variable to obtain a second linear feature combination variable.

[0151] In some embodiments, a sixth linear transformation process is performed on the non-linear feature combination variables through a dense layer, that is, in the dense layer, the non-linear feature combination variables of multiple dimensions are respectively linearly transformed through the weight matrix of the dense layer, and a bias term is added, and then a non-linear transformation is performed through an activation function (such as sigmoid, ReLU, etc.), and finally a second linear feature combination variable is obtained. As an example, the sixth linear transformation process can be expressed by the following formula:

[0152] Dense 6th = WX 10 +b (10)

[0153] Where, Dense 6th represents the sixth linear transformation process, X 10 represents the input non-linear feature combination variables of multiple dimensions, W represents the weight matrix of the dense layer, and b represents the bias term.

[0154] In step 1044, a second gated linear process is performed on the second linear feature combination variable to obtain a gated feature combination vector.

[0155] In some embodiments, the second gated linear process is implemented through a second gated linear model, and the structure of the second gated linear model is the same as that of the first gated linear model described above, including a third convolutional layer, a fourth convolutional layer, and a second activation function.

[0156] In some embodiments, referring to Figure 4G , Figure 4F shown in step 1044, it can be implemented through the following steps 10441 to 10444, which are specifically described below.

[0157] In step 10441, a third convolution process is performed on the second linear feature combination variable through a third convolutional layer to obtain a first combination vector.

[0158] In some embodiments, the second gated linear process of the second gated linear model can be expressed by the following formula:

[0159] GLU 2nd = δ(W3X 11 +b3)*(W4X 11 +b4) (11)

[0160] Where, X 11 represents the input second linear feature combination variable, W3 and W4 respectively represent the weight matrices of the third convolutional layer and the fourth convolutional layer, b3 and b4 respectively represent the biases of the third convolutional layer and the fourth convolutional layer, δ is an activation function (such as sigmoid), and * represents the elements at the corresponding positions (here referring to δ(W3X 11+(b3) and (W4X 11 Multiply the corresponding elements of the two matrices of the same order, (b4), element-wise, which is also called the Hadamard product. As can be seen from Equation (11), first input X 11 into the third convolutional layer and the fourth convolutional layer to obtain two outputs. Then, apply a non-linear transformation to the output of the third convolutional layer using the activation function δ, and then perform element-wise multiplication through the Hadamard product to obtain the corresponding output of the GLU.

[0161] Here, the second linear feature combination variable is processed by the third convolutional layer through the third convolution to obtain the first combined vector, corresponding to (W3X 11 +b3) in Equation (11).

[0162] In step 10442, perform a fourth activation process on the first combined vector through the second activation function to obtain the importance parameter vector of the first combined vector.

[0163] Continuing with the above example, perform a fourth activation process on the first combined vector through the second activation function (such as sigmoid), that is, perform a non-linear transformation on the first combined vector through the second activation function to obtain the importance parameter vector corresponding to each dimension of the first combined vector, corresponding to δ(W3X 11 +b3) in Equation (11), where δ represents the second activation function (such as sigmoid).

[0164] In step 10443, perform a fourth convolution process on the second linear feature combination variable through the fourth convolutional layer to obtain the second combined vector.

[0165] Continuing with the above example, perform a fourth convolution process on the second linear feature combination variable through the fourth convolutional layer to obtain the second combined vector, corresponding to (W4X 11 +b4) in Equation (11).

[0166] In step 10444, multiply the second combined vector and the importance parameter vector of the first combined vector to obtain the gated feature combination vector.

[0167] Continuing with the above example, multiply the second combined vector and the importance parameter vector of the first combined vector to obtain the gated feature combination vector, corresponding to δ(W3X 11 +b3)*(W4X 11 +b4) in Equation (11), where * represents the multiplication process.

[0168] Continue to refer to Figure 4F, in step 1045, add the first feature combination result and the gated feature combination vector to obtain a gated combination residual vector.

[0169] In some embodiments, add the corresponding position elements of the first feature combination result and the gated feature combination vector to obtain a gated combination residual vector.

[0170] In step 1046, perform a normalization process on the gated combination residual vector to obtain a second feature combination result.

[0171] In some embodiments, perform a normalization process on the gated combination residual vector to obtain the second feature combination result corresponding to each dimension. The normalization process (Norm) can be represented by the following formula:

[0172]

[0173] where X 12 represents the input gated residual vector, represents X 12 's expected value (mean), represents X 12 's variance, and N represents the number of dimensions of the input gated residual vector.

[0174] Through step 104, it is realized to further perform selective feature extraction processing on the first feature combination result through the gated residual network, achieving the beneficial effect of better capturing the complexity of the payment variables in multiple dimensions of the target object.

[0175] Continue to refer to Figure 4A , in step 105, perform a first linear transformation process on the second feature combination result to obtain a linear transformation result.

[0176] In some embodiments, perform a first linear transformation process on the second feature combination result through a dense layer, that is, in the dense layer, the second feature combination results in multiple dimensions are respectively linearly transformed through the weight matrix of the dense layer, and a bias term is added, and then a non-linear transformation is performed through an activation function (such as sigmoid, ReLU, etc.). Finally, a linear transformation result is obtained. As an example, the first linear transformation process can be represented by the following formula:

[0177] Dense 1st = WX 13 +b (13)

[0178] where Dense 1st represents the first linear transformation process, X 13 represents the input second feature combination results in multiple dimensions, W represents the weight matrix of the dense layer, and b represents the bias term.

[0179] Continue to refer to Figure 4A , in step 106, perform a non-linear transformation on the linear transformation result to obtain the probability that the target object is an abnormal object.

[0180] In some embodiments, refer to Figure 4H , Figure 4A The shown step 106 can be implemented through the following steps 1061 to 1063, and the following is a specific description.

[0181] In step 1061, perform a fifth activation on the linear transformation result through a preset first activation function to obtain a first non-linear processing result.

[0182] In some embodiments, perform a fifth activation on the linear transformation result through a preset first activation function (such as the ELU activation function), that is, perform a non-linear transformation on the linear transformation result through the first activation function to obtain the first non-linear processing result corresponding to each dimension. Here, the fifth activation can be represented by the following formula:

[0183]

[0184] where ELU 5th represents the fifth activation, X 14 represents the input linear transformation result, α is a hyperparameter used to control the slope of the negative region. When the input X 14 is greater than or equal to 0, the output of the ELU activation function is equal to the input X 14 itself, which maintains the linear characteristic and enables the neural network to learn linear relationships. When the input X 14 is less than 0, the output of the ELU activation function is equal to This part of the negative region introduces non-linear characteristics, enabling the neural network to learn non-linear relationships.

[0185] In step 1062, perform a seventh linear transformation on the first non-linear processing result to obtain a second linear transformation result.

[0186] In some embodiments, perform a seventh linear transformation on the first non-linear processing result through a dense layer, that is, in the dense layer, linearly transform the first non-linear processing results of multiple dimensions through the weight matrix of the dense layer, add a bias term, and then perform a non-linear transformation through an activation function (such as sigmoid, ReLU, etc.). Finally, obtain the second linear transformation result. As an example, the seventh linear transformation can be represented by the following formula:

[0187] Dense 7th = WX 15 +b (15)

[0188] Among them, Dense 7th represents the seventh linear transformation process, and X 15 represents the first non-linear processing result of multiple dimensions of the input. W represents the weight matrix of the dense layer, and b represents the bias term. After the seventh linear transformation process of the dense layer, the first non-linear processing result of multiple dimensions can obtain a new dimension (for example, if the input is an N-dimensional first non-linear processing result, after the seventh linear transformation by a dense layer with M neurons, the output will be an M-dimensional vector), so that the finally obtained second linear transformation processing result has a higher-level feature combination ability.

[0189] In step 1063, the second linear transformation processing result is subjected to the sixth activation process through a preset second activation function to obtain the probability that the target object is an abnormal object.

[0190] In some embodiments, the second linear transformation processing result is subjected to the sixth activation process through a preset second activation function (such as sigmoid), that is, the second linear transformation processing result is non-linearly transformed through the second activation function to obtain the probability that the target object is an abnormal object.

[0191] In some embodiments, the second activation function is used for exponential function calculation. The second linear transformation processing result is subjected to exponential function calculation, and the result of the exponential function calculation is mapped to a value greater than 0 and less than 1 as the probability that the target object is an abnormal object.

[0192] Among them, the second activation function (such as the sigmoid function) can be expressed as the following formula:

[0193]

[0194] Among them, X 16 represents the second linear transformation processing result of the input.

[0195] In some other embodiments, after obtaining the probability that the target object is an abnormal object, the label of whether the target object is an abnormal object can be output through a preset threshold. For example, only when the probability value is greater than or equal to 0.7, the target object is classified as an abnormal object.

[0196] In some embodiments, steps 105 to 106 can be implemented by a deep neural network. Refer to Figure 8 , Figure 8 which is a schematic diagram of the structure of the deep neural network provided by the embodiments of the present application. Among them, the Dense layer and the ELU layer adopt the structures described above, and the Sigmoid layer is used to obtain the probability that the target object is an abnormal object.

[0197] In some embodiments, when the electronic payment record data of the target object includes payment variables of multiple dimensions corresponding to a single transaction, and there are multiple pieces of electronic payment data, the fifth activation process is performed on the linear transformation processing results corresponding to the multiple pieces of electronic payment record data through a preset first activation function (such as the ELU activation function) to obtain first non-linear processing results corresponding to the multiple pieces of electronic payment record data respectively; a seventh linear transformation process is performed on the first non-linear processing results corresponding to the multiple pieces of electronic payment record data respectively to obtain second linear transformation processing results corresponding to the multiple pieces of electronic payment record data respectively; a sixth activation process is performed on the second linear transformation processing results corresponding to the multiple pieces of electronic payment record data respectively to obtain abnormal probabilities corresponding to the multiple pieces of electronic payment record data respectively; a weighted summation process is performed on the multiple abnormal probabilities, and the weighted summation result is used as the probability that the target object is an abnormal object.

[0198] In some embodiments, refer to Figure 4I , before Figure 4A the step 102 shown, the following steps 107 to 108 may also be performed, which are specifically described below.

[0199] In step 107, multiple payment variables corresponding to the same dimension are extracted from the payment variables of multiple dimensions corresponding to multiple transactions respectively.

[0200] In some embodiments, multiple payment variables corresponding to the same dimension are extracted from the payment variables of multiple dimensions corresponding to multiple transactions respectively, for example, data such as transaction amounts and transaction methods corresponding to different transactions are obtained.

[0201] In step 108, the multiple payment variables corresponding to the same dimension are aggregated to obtain a new payment variable, where the new payment variable is used to transfer to the step of obtaining the first vector representation and the first feature weight corresponding to the payment variables of multiple dimensions respectively, that is, the corresponding first vector representation and the first feature weight are obtained based on the new payment variable. For the specific implementation manner, please refer to the above step 102.

[0202] In some embodiments, the multiple payment variables corresponding to the same dimension are aggregated (such as statistical processing methods such as accumulation, mean calculation, and variance calculation) to obtain a new payment variable, where the new payment variable is used to transfer to the step of obtaining the first vector representation and the first feature weight corresponding to the payment variables of multiple dimensions respectively (corresponding to step 102 above).

[0203] In some embodiments, the abnormal object recognition method provided by the embodiments of the present application is implemented through an abnormal object recognition model. Refer to Figure 3, the abnormal object recognition model includes a feature selection network, a gated residual network, and a deep neural network; obtaining the first vector representation and the first feature weight corresponding to the payment variables in multiple dimensions respectively is achieved through the feature selection network (corresponding to step 102 above); feature combination is achieved through the feature selection network (corresponding to step 103 above); feature extraction is achieved through the gated residual network (corresponding to step 104 above); the first linear transformation process (corresponding to step 105 above) and the non-linear transformation process (corresponding to step 106 above) are achieved through the deep neural network.

[0204] In some embodiments, before obtaining the electronic payment record data of the target object (corresponding to step 101 above), refer to Figure 4J , the following steps 109 to 116 can also be executed to train the abnormal object recognition model, which will be specifically described below.

[0205] In step 109, obtain the payment record data and labels of multiple sample objects, where the payment record data of multiple sample objects includes sample payment variables in multiple dimensions for each sample object, and the labels are used to characterize whether the sample object is an abnormal object.

[0206] In some embodiments, obtain the payment record data and labels of multiple sample objects, and divide the payment record data and labels of multiple sample objects into a training set and a test set, where the training set is used for training the abnormal object recognition model, and the test set is used for testing and evaluating the trained abnormal object recognition model to adjust the model parameters (such as hyperparameters), so as to further optimize the abnormal object recognition model.

[0207] In step 110, through the feature selection network, obtain the first sample vector representation and the first sample feature weight corresponding to the sample payment variables in multiple dimensions respectively.

[0208] Here, refer to the description of step 102 above, and the specific implementation process refers to formula (1-7).

[0209] In step 111, based on the first sample vector representation and the first sample feature weight corresponding to the sample payment variables in multiple dimensions respectively, call the feature selection network to obtain the first sample feature combination result.

[0210] Here, refer to the description of step 103 above, multiply the first sample vector representation and the first sample feature weight corresponding to each dimension, and perform a splicing and combination process on the multiplied results to obtain the first sample feature combination result.

[0211] In step 112, perform feature extraction on the first sample feature combination result through the gated residual network to obtain the second sample feature combination result.

[0212] In some embodiments, selective feature extraction is performed on the first sample feature combination result through a gated residual network to obtain a second sample feature combination result. Here, referring to the description of step 104 above, the specific implementation process can be seen in formulas (8 - 12).

[0213] In step 113, a first linear transformation process is performed on the second sample feature combination result through a deep neural network to obtain a sample linear transformation process result.

[0214] In some embodiments, a first linear transformation process is performed on the second sample feature combination result through a dense layer, that is, in the dense layer, the second sample feature combination result in multiple dimensions is linearly transformed through the weight matrix of the dense layer respectively, and a bias term is added, and then a non - linear transformation is performed through an activation function (such as sigmoid, ReLU, etc.), and finally a sample linear transformation process result is obtained. Here, the first linear transformation process refers to the description of step 105 above.

[0215] In step 114, a non - linear transformation process is performed on the sample linear transformation process result through a deep neural network to obtain the prediction probability that the sample object is an abnormal object.

[0216] Here, referring to the description of steps 1061 to 1063 above, the specific implementation process can be seen in formulas (14 - 15).

[0217] In step 115, a loss value corresponding to a preset loss function is obtained through the prediction probability and the label.

[0218] In some embodiments, a loss value of a preset loss function (such as a cross - entropy loss function) is obtained through the prediction probability and the label.

[0219] In step 116, the model parameters of the abnormal object recognition model are updated through the loss value to obtain a trained abnormal object recognition model.

[0220] In some embodiments, the gradient of the loss function is calculated through backpropagation and an optimization algorithm (such as Adam, SGD, AdamW, etc.), and the parameters of the abnormal object recognition model (such as the weight matrix, bias term, etc. of the dense layer in formula (1)) are updated according to the gradient direction to minimize the loss value of the loss function, and a trained abnormal object recognition model is obtained.

[0221] Through steps 109 to 116, the training of the abnormal object recognition model on a large - scale data is realized, achieving the beneficial effect of adapting to the behavior patterns of abnormal objects characterized by continuously changing payment variables in multiple dimensions.

[0222] In some embodiments, after obtaining the trained abnormal object recognition model, the trained abnormal object recognition model can also be evaluated using a test set (evaluation metrics may include accuracy, recall rate, precision, F1 score, etc.) to obtain an evaluation result; based on the evaluation result, the hyperparameters of the trained abnormal object recognition model are adjusted to obtain the abnormal object recognition model.

[0223] In some embodiments, after obtaining the probability that the target object is an abnormal object, a transaction risk warning process for the target object can also be performed in response to the probability that the target object is an abnormal object being higher than a preset probability value.

[0224] In some embodiments, after obtaining the probability that the target object is an abnormal object, when the probability that the target object is an abnormal object is higher than a preset probability value, the current payment operation is intercepted, and the Figure 1 terminal device 200 shown is called to display a prompt indicating that there is a transaction risk in the current payment operation. An entry for self-release can be further displayed. After the target object clicks to enter the self-release interface, risk notice, risk confirmation, and identity verification operations are performed. After confirmation, the normal payment function of the target object is restored.

[0225] Among them, the specific descriptions of the user's risk notice, risk confirmation, and identity verification operations are as follows:

[0226] 1) Risk notice: The user reads the displayed common risk types and can choose to abandon the payment or confirm the transaction risk after 5S;

[0227] 2) Risk confirmation: Ensure that the user clearly knows the true identity of the other party and complete the name of the other party; check the purpose of the current transaction; check the identity of the other party; during this process, the user can abandon the payment or enter the next identity verification after filling in the information.

[0228] 3) Identity verification: Confirm the identity of the user himself through face recognition.

[0229] Next, in combination with the exemplary applications and implementations of the server and terminal device provided in the embodiments of the present application, the method for identifying abnormal objects provided in the embodiments of the present application in the electronic payment scenario will be described. In the electronic payment scenario, the abnormal object identification strategy that combines data analysis results and business information usually directly acts on the electronic payment process to identify abnormal objects based on the user's payment situation, so as to intercept abnormal objects that perform illegal acts (such as money laundering, etc.), in order to achieve anti-money laundering, anti-fraud, etc. operations, and then conduct financial risk control to protect the personal interests and property safety of users. Usually, the identification strategy uses a series of judgment logics to distinguish the user and the transaction order situation, so as to complete the identification of abnormal objects. In this process, the probability that the user in each interval is an abnormal object is significantly different. For example, if the user's transaction score is high or the credit score is low, the probability of being considered an abnormal object is too high and should be intercepted. Therefore, before constructing the strategy, it is necessary to identify abnormal objects based on the user's electronic payment record data, so as to be used for user transaction risk early warning processing. Refer to Figure 5 , Figure 5 is a schematic interaction flow diagram of the method for identifying abnormal objects provided in the embodiments of the present application, and will be described in combination with Figure 5 the steps shown.

[0230] In step 201, the terminal device 200 receives the payment operation of the user.

[0231] Among them, the user's payment operation can be generated in the following scenarios, for example: online shopping, transfer and remittance in instant messaging, rewarding the anchor in the live broadcast interface, booking a hotel in the navigation interface, etc.

[0232] As an example, the user's payment operation can be realized through the electronic payment function of the instant messaging application program. For example, transfer and remittance are carried out through the electronic payment function in the instant messaging application program.

[0233] In step 202, the terminal device 200 sends a payment request to the server 100.

[0234] In some embodiments, the terminal device 200 generates a payment request in response to the user's payment operation and sends it to the server 100.

[0235] In step 203, the server 100 obtains historical electronic payment record data.

[0236] In some embodiments, in response to the server 100 receiving a payment request sent by the terminal device 200, the server 100 obtains historical electronic payment record data. The electronic payment record data includes payment variables in multiple dimensions, and the payment variables in multiple dimensions include, but are not limited to, user age group, account type, transaction method, transaction amount, transaction time interval, transaction amount change rate, etc. For example, after obtaining the historical electronic payment record data, it is necessary to perform data preprocessing on it, such as cleaning the data format, removing duplicate data, handling missing values, detecting and handling outliers, converting data types, data verification and consistency checking, and data backup and storage, etc.

[0237] In step 204, the server 100 determines the probability that the user is an abnormal object according to the historical electronic payment record data.

[0238] In some embodiments, referring to Figure 4A , Figure 5 The shown step 204 can be implemented by obtaining the electronic payment record data of the target object (user), where the electronic payment record data includes payment variables in multiple dimensions; obtaining the first vector representation and the first feature weight corresponding to each of the payment variables in multiple dimensions; performing feature combination based on the first vector representation and the first feature weight corresponding to each of the payment variables in multiple dimensions to obtain a first feature combination result; performing feature extraction on the first feature combination result to obtain a second feature combination result; performing a first linear transformation process on the second feature combination result to obtain a linear transformation process result; performing a non-linear transformation process on the linear transformation process result to obtain the probability that the target object (user) is an abnormal object.

[0239] In some other embodiments, in addition to using Figure 3 the network structure of the abnormal object recognition model shown in to perform the probability prediction task, at the same time, an ensemble learning method, such as random forest, gradient boosting tree, etc., can also be considered to integrate multiple different models together to improve the overall performance and robustness of the abnormal object recognition model. Different types of data, such as the sequential behavior data and geographical location data of the user, can also be fused into the abnormal object recognition model when using the electronic payment record data to provide more comprehensive information for identifying abnormal objects. Considering that the behaviors of abnormal objects may have different characteristics at different times and locations, transfer learning can be used to transfer the model trained in one domain to another domain, thereby improving the performance of the abnormal object recognition model.

[0240] For the specific description of step 204, reference can be made to the relevant descriptions of steps 101 to 106 above, which will not be elaborated here.

[0241] In step 205, the server 100 processes the current payment operation according to the probability that the user is an abnormal object.

[0242] In some embodiments, after obtaining the probability that the user is an abnormal object, when the probability that the user is an abnormal object is higher than a preset probability threshold, the user is considered an abnormal object. For example, an object that can be identified as performing illegal acts (such as money laundering and fraud) in an electronic payment scenario is intercepted for the current payment operation, and Figure 1 the terminal device 200 shown is called to display a prompt indicating that there is a transaction risk for the current payment operation. A self-relief entry can be further displayed. After the user clicks to enter the self-relief interface, the user performs risk notice, risk confirmation, and identity verification operations. After confirmation, the normal payment function of the user is restored.

[0243] Through steps 201 to 205, by performing feature combination processing based on the payment variables of multiple dimensions, the core features are automatically selected in the electronic payment scenario, reducing the need for manual feature engineering. By selectively extracting features from the first feature combination result, further feature extraction is achieved, thereby enhancing the depth representation ability and feature combination ability of the model to better capture the complexity of the payment variables of the target object in multiple dimensions, and finally achieving the beneficial effect of improving the recognition accuracy of abnormal objects, and being able to accurately identify and intercept objects performing illegal acts (such as money laundering and fraud) in the electronic payment scenario to achieve anti-money laundering and anti-fraud financial risk control.

[0244] Next, the implementation of the abnormal object recognition device 133 provided in the embodiments of the present application as a software module will be continued. In some embodiments, as Figure 2 shown, the software module in the abnormal object recognition device 133 stored in the memory 130 may include:

[0245] An acquisition module 1331, configured to acquire the electronic payment record data of the target object, where the electronic payment record data includes payment variables of multiple dimensions.

[0246] A processing module 1332, configured to acquire the first vector representation and the first feature weight corresponding to each of the payment variables of the multiple dimensions.

[0247] In some embodiments, the processing module 1332 is further configured to perform feature combination based on the first vector representation and the first feature weight corresponding to each of the payment variables of the multiple dimensions to obtain a first feature combination result.

[0248] In some embodiments, the processing module 1332 is further configured to perform feature extraction on the first feature combination result to obtain a second feature combination result.

[0249] A generation module 1333, configured to perform a first linear transformation process on the second feature combination result to obtain a linear transformation process result.

[0250] In some embodiments, the generating module 1333 is further configured to perform a non-linear transformation process on the linear transformation processing result to obtain the probability that the target object is an abnormal object.

[0251] In some embodiments, the processing module 1332 is further configured to perform a second linear transformation process on the payment variables of the multiple dimensions respectively to obtain a first linear payment variable for each dimension; perform a combined mapping process on the first linear payment variable for each dimension to obtain a first vector representation corresponding to each dimension; perform a normalization process on the first vector representation corresponding to each dimension to obtain a first feature weight corresponding to each dimension.

[0252] In some embodiments, the processing module 1332 is further configured to multiply the first vector representation corresponding to each dimension by the first feature weight to obtain a feature vector corresponding to each dimension; combine the feature vectors corresponding to each dimension to obtain the first feature combination result.

[0253] In some embodiments, the processing module 1332 is further configured to perform the following processing for the first linear payment variable of each dimension: perform a third linear transformation process on the first linear payment variable to obtain a second linear payment variable corresponding to each dimension; perform a first activation process on the second linear payment variable through a preset first activation function to obtain a non-linear payment variable corresponding to each dimension; perform a fourth linear transformation process on the non-linear payment variable to obtain a third linear payment variable corresponding to each dimension; perform a first gated linear process on the third linear payment variable to obtain a linear combination vector corresponding to the dimension; add the first linear payment variable of each dimension and the linear combination vector to obtain a gated residual vector corresponding to each dimension; perform a normalization process on the gated residual vector to obtain a first vector representation corresponding to each dimension.

[0254] In some embodiments, the first gated linear process is implemented by a first gated linear model, and the first gated linear model includes a first convolutional layer, a second convolutional layer, and a second activation function; the processing module 1332 is further configured to perform a first convolutional process on the third linear payment variable through the first convolutional layer to obtain a first convolutional payment vector; perform a second activation process on the first convolutional payment vector through the second activation function to obtain an importance parameter vector of the first convolutional payment vector; perform a second convolutional process on the third linear payment variable through the second convolutional layer to obtain a second convolutional payment vector; perform a multiplication process on the second convolutional payment vector and the importance parameter vector of the first convolutional payment vector to obtain a linear combination vector corresponding to the dimension.

[0255] In some embodiments, the processing module 1332 is further configured to perform a fifth linear transformation process on the first feature combination result to obtain a first linear feature combination variable; perform a third activation process on the first linear feature combination variable through a preset first activation function to obtain a non-linear feature combination variable; perform a sixth linear transformation process on the non-linear feature combination variable to obtain a second linear feature combination variable; perform a second gated linear process on the second linear feature combination variable to obtain a gated feature combination vector; add the first feature combination result and the gated feature combination vector to obtain a gated combination residual vector; perform a normalization process on the gated combination residual vector to obtain a second feature combination result.

[0256] In some embodiments, the second gated linear process is implemented by a second gated linear model, and the second gated linear model includes a third convolutional layer, a fourth convolutional layer, and a second activation function; the processing module 1332 is further configured to perform a third convolution process on the second linear feature combination variable through the third convolutional layer to obtain a first combination vector; perform a fourth activation process on the first combination vector through the second activation function to obtain an importance parameter vector of the first combination vector; perform a fourth convolution process on the second linear feature combination variable through the fourth convolutional layer to obtain a second combination vector; perform a multiplication process on the second combination vector and the importance parameter vector of the first combination vector to obtain a gated feature combination vector.

[0257] In some embodiments, the generating module 1333 is further configured to perform a fifth activation process on the linear transformation result through a preset first activation function to obtain a first non-linear processing result; perform a seventh linear transformation process on the first non-linear processing result to obtain a second linear transformation result; perform a sixth activation process on the second linear transformation result through a preset second activation function to obtain the probability that the target object is an abnormal object.

[0258] In some embodiments, the generating module 1333 is further configured to perform the exponential function calculation on the second linear transformation result, and map the result of the exponential function calculation to a value greater than 0 and less than 1 as the probability that the target object is an abnormal object.

[0259] In some embodiments, when the electronic payment record data of the target object includes the payment variables of the multiple dimensions corresponding to a transaction, and there are multiple pieces of such electronic payment data, the generating module 1333 is further configured to perform a fifth activation process on the linear transformation processing results corresponding to the multiple pieces of electronic payment record data through a preset first activation function to obtain first non-linear processing results respectively corresponding to the multiple pieces of electronic payment record data; perform a fifth linear transformation process on the first non-linear processing results respectively corresponding to the multiple pieces of electronic payment record data to obtain second linear transformation processing results respectively corresponding to the multiple pieces of electronic payment record data; perform a sixth activation process on the second linear transformation processing results respectively corresponding to the multiple pieces of electronic payment record data to obtain abnormal probabilities respectively corresponding to the multiple pieces of electronic payment record data; perform a weighted summation process on the multiple abnormal probabilities, and use the weighted summation result as the probability that the target object is an abnormal object.

[0260] In some embodiments, when the electronic payment record data includes the payment variables of the multiple dimensions corresponding to multiple transactions respectively, the processing module 1332 is further configured to extract multiple payment variables corresponding to the same dimension from the payment variables of the multiple dimensions corresponding to the multiple transactions respectively; perform an aggregation process on the multiple payment variables corresponding to the same dimension to obtain new payment variables, where the new payment variables are used to transfer to the step of obtaining the first vector representation and the first feature weight respectively corresponding to the payment variables of the multiple dimensions.

[0261] In some embodiments, the method for identifying the abnormal object is implemented by an abnormal object identification model, and the abnormal object identification model includes a feature selection network, a gated residual network, and a deep neural network; obtaining the first vector representation and the first feature weight corresponding to each of the plurality of dimensions of payment variables is implemented by the feature selection network; the feature combination is implemented by the feature selection network; the feature extraction is implemented by the gated residual network; the first linear transformation process and the non-linear transformation process are implemented by the deep neural network; the generation module 1333 is further configured to train the abnormal object identification model in the following manner: obtaining payment record data and labels of a plurality of sample objects, wherein the payment record data of the plurality of sample objects includes a plurality of dimensions of sample payment variables of each sample object, and the labels are used to indicate whether the sample object is the abnormal object; through the feature selection network, obtaining the first sample vector representation and the first sample feature weight corresponding to each of the plurality of dimensions of sample payment variables; based on the first sample vector representation and the first sample feature weight corresponding to each of the plurality of dimensions of sample payment variables, invoking the feature selection network to obtain a first sample feature combination result; performing feature extraction on the first sample feature combination result through the gated residual network to obtain a second sample feature combination result; performing the first linear transformation process on the second sample feature combination result through the deep neural network to obtain a sample linear transformation process result; through the deep neural network, performing the non-linear transformation process on the sample linear transformation process result to obtain a prediction probability that the sample object is an abnormal object; obtaining a loss value corresponding to a preset loss function through the prediction probability and the label; updating the model parameters of the abnormal object identification model through the loss value to obtain the trained abnormal object identification model.

[0262] In some embodiments, the generation module 1333 is further configured to, in response to the probability that the target object is an abnormal object being higher than a preset probability value, perform a transaction risk warning process on the target object.

[0263] An embodiment of the present application provides a computer program product, which includes a computer program or computer-executable instructions, and the computer program or computer-executable instructions are stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium, and the processor executes the computer-executable instructions, so that the electronic device executes the method for identifying the abnormal object in the above embodiments of the present application.

[0264] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, where computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will be caused to execute the method for identifying an abnormal object provided by the embodiment of the present application. For example, as Figure 4A shown in the method for identifying an abnormal object.

[0265] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0266] In some embodiments, the computer-executable instructions may be in the form of a program, software, software module, script, or code, and may be written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including being deployed as an independent program or being deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0267] As an example, the computer-executable instructions may or may not correspond to a file in the file system, and may be stored as part of a file that stores other programs or data. For example, they may be stored in one or more scripts in a Hyper Text Markup Language (HTML) document, stored in a single file dedicated to the program under discussion, or stored in multiple cooperating files (for example, files that store one or more modules, subroutines, or code portions).

[0268] As an example, the computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed at multiple locations and interconnected by a communication network.

[0269] In summary, through the embodiment of the present application, based on the feature combination processing of payment variables in multiple dimensions, the automatic selection of core features is realized in the electronic payment scenario, reducing the need for manual feature engineering. By selectively extracting features from the first feature combination result, further feature extraction is realized, thereby enhancing the depth representation ability and feature combination ability of the model to better capture the complexity of payment variables in multiple dimensions of the target object, and finally achieving the beneficial effect of improving the accuracy of identifying abnormal objects.

[0270] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for identifying abnormal objects, characterized in that, The method includes: Obtaining electronic payment record data of a target object, where the electronic payment record data includes payment variables in multiple dimensions; Obtaining a first vector representation and a first feature weight corresponding to each of the payment variables in the multiple dimensions; Performing feature combination based on the first vector representation and the first feature weight corresponding to each of the payment variables in the multiple dimensions to obtain a first feature combination result; Performing feature extraction on the first feature combination result to obtain a second feature combination result; Performing a first linear transformation process on the second feature combination result to obtain a linear transformation process result; Performing a non-linear transformation process on the linear transformation process result to obtain the probability that the target object is an abnormal object.

2. The method according to claim 1, wherein The obtaining the first vector representation and the first feature weight corresponding to each of the payment variables in the multiple dimensions includes: Performing a second linear transformation process on each of the payment variables in the multiple dimensions to obtain a first linear payment variable for each dimension; Performing a combined mapping process on the first linear payment variable for each dimension to obtain a first vector representation corresponding to each dimension; Performing a normalization process on the first vector representation corresponding to each dimension to obtain a first feature weight corresponding to each dimension.

3. The method according to claim 2, wherein the performing feature combination based on the first vector representation and the first feature weight corresponding to each of the payment variables in the multiple dimensions to obtain a first feature combination result includes: Multiplying the first vector representation corresponding to each dimension by the first feature weight to obtain a feature vector corresponding to each dimension; Combining the feature vectors corresponding to each dimension to obtain the first feature combination result.

4. The method according to claim 2, wherein The performing a combined mapping process on the first linear payment variable for each dimension to obtain a first vector representation corresponding to each dimension includes: Performing the following process for the first linear payment variable of each dimension: Performing a third linear transformation process on the first linear payment variable to obtain a second linear payment variable corresponding to each dimension; Performing a first activation process on the second linear payment variable through a preset first activation function to obtain a non-linear payment variable corresponding to each dimension; Performing a fourth linear transformation process on the non-linear payment variable to obtain a third linear payment variable corresponding to each dimension; Performing a first gated linear process on the third linear payment variable to obtain a linear combined vector corresponding to the dimension; Adding the first linear payment variable of each dimension and the linear combined vector to obtain a gated residual vector corresponding to each dimension; Performing a normalization process on the gated residual vector to obtain a first vector representation corresponding to each dimension.

5. The method according to claim 4, wherein The first gated linear process is implemented through a first gated linear model, and the first gated linear model includes a first convolutional layer, a second convolutional layer, and a second activation function; The performing a first gated linear process on the third linear payment variable to obtain a linear combined vector corresponding to the dimension includes: Perform a first convolution process on the third linear payment variable through the first convolution layer to obtain a first convolutional payment vector; Perform a second activation process on the first convolutional payment vector through the second activation function to obtain an importance parameter vector of the first convolutional payment vector; Perform a second convolution process on the third linear payment variable through the second convolution layer to obtain a second convolutional payment vector; Multiply the second convolutional payment vector by the importance parameter vector of the first convolutional payment vector to obtain a linear combination vector corresponding to the dimension.

6. The method according to claim 1, wherein The feature extraction of the first feature combination result to obtain a second feature combination result includes: Perform a fifth linear transformation process on the first feature combination result to obtain a first linear feature combination variable; Perform a third activation process on the first linear feature combination variable through a preset first activation function to obtain a non-linear feature combination variable; Perform a sixth linear transformation process on the non-linear feature combination variable to obtain a second linear feature combination variable; Perform a second gated linear process on the second linear feature combination variable to obtain a gated feature combination vector; Add the first feature combination result and the gated feature combination vector to obtain a gated combination residual vector; Perform a normalization process on the gated combination residual vector to obtain a second feature combination result.

7. According to the method described in claim 6, The second gated linear process is implemented through a second gated linear model, and the second gated linear model includes a third convolution layer, a fourth convolution layer, and a second activation function; The performing a second gated linear process on the second linear feature combination variable to obtain a gated feature combination vector includes: Perform a third convolution process on the second linear feature combination variable through the third convolution layer to obtain a first combination vector; Perform a fourth activation process on the first combination vector through the second activation function to obtain an importance parameter vector of the first combination vector; Perform a fourth convolution process on the second linear feature combination variable through the fourth convolution layer to obtain a second combination vector; Multiply the second combination vector by the importance parameter vector of the first combination vector to obtain a gated feature combination vector.

8. The method according to claim 1, characterized in that, The performing a non-linear transformation process on the result of the linear transformation process to obtain the probability that the target object is an abnormal object includes: Perform a fifth activation process on the result of the linear transformation process through a preset first activation function to obtain a first non-linear processing result; Perform a seventh linear transformation process on the first non-linear processing result to obtain a second linear transformation processing result; Perform a sixth activation process on the second linear transformation processing result through a preset second activation function to obtain the probability that the target object is an abnormal object.

9. The method according to claim 8, characterized in that The second activation function is used for exponential function calculation, and the performing a sixth activation process on the second linear transformation processing result through a preset second activation function to obtain the probability that the target object is an abnormal object includes: Perform the exponential function calculation on the result of the second linear transformation process, and map the result of the exponential function calculation to a value greater than 0 and less than 1 as the probability that the target object is an abnormal object.

10. The method according to claim 8, wherein when the electronic payment record data of the target object includes the payment variables of the multiple dimensions corresponding to one transaction, and there are multiple pieces of the electronic payment data, the fifth activation process is performed on the result of the linear transformation process through a preset first activation function to obtain a first non-linear processing result, including: Performing a fifth activation process on the result of the linear transformation process corresponding to the multiple pieces of electronic payment record data through a preset first activation function to obtain first non-linear processing results corresponding to the multiple pieces of electronic payment record data respectively; The seventh linear transformation process is performed on the first non-linear processing result to obtain a second linear transformation process result, including: Performing a seventh linear transformation process on the first non-linear processing results corresponding to the multiple pieces of electronic payment record data respectively to obtain second linear transformation results corresponding to the multiple pieces of electronic payment record data respectively; The sixth activation process is performed on the result of the second linear transformation process through a preset second activation function to obtain the probability that the target object is an abnormal object, including: Performing a sixth activation process on the second linear transformation results corresponding to the multiple pieces of electronic payment record data respectively to obtain abnormal probabilities corresponding to the multiple pieces of electronic payment record data respectively; Performing a weighted summation process on the multiple abnormal probabilities, and using the weighted summation result as the probability that the target object is an abnormal object.

11. The method according to any one of claims 1 to 10, wherein when the electronic payment record data includes the payment variables of the multiple dimensions corresponding to multiple transactions respectively, before obtaining the first vector representation and the first feature weight corresponding to the payment variables of the multiple dimensions respectively, the method further includes: Extracting multiple payment variables corresponding to the same dimension from the payment variables of the multiple dimensions corresponding to the multiple transactions respectively; Performing an aggregation process on the multiple payment variables corresponding to the same dimension to obtain a new payment variable, wherein the new payment variable is used to transfer to the step of obtaining the first vector representation and the first feature weight corresponding to the payment variables of the multiple dimensions respectively.

12. The method according to claim 1, wherein The method for identifying the abnormal object is implemented through an abnormal object identification model, and the abnormal object identification model includes a feature selection network, a gated residual network, and a deep neural network; obtaining the first vector representation and the first feature weight corresponding to the payment variables of the multiple dimensions respectively is implemented through the feature selection network; the feature combination is implemented through the feature selection network; the feature extraction is implemented through the gated residual network; the first linear transformation process and the non-linear transformation process are implemented through the deep neural network.

13. The method according to claim 12, wherein Before obtaining the electronic payment record data of the target object, the method further includes: Train the abnormal object recognition model in the following manner: Obtain the payment record data and labels of multiple sample objects, where the payment record data of the multiple sample objects includes sample payment variables in multiple dimensions for each sample object, and the labels are used to characterize whether the sample object is the abnormal object; Through the feature selection network, obtain the first sample vector representation and the first sample feature weight corresponding to the sample payment variables in the multiple dimensions respectively; Based on the first sample vector representation and the first sample feature weight corresponding to the sample payment variables in the multiple dimensions respectively, call the feature selection network to obtain the first sample feature combination result; Extract features from the first sample feature combination result through the gated residual network to obtain the second sample feature combination result; Perform the first linear transformation process on the second sample feature combination result through the deep neural network to obtain the sample linear transformation process result; Through the deep neural network, perform the non-linear transformation process on the sample linear transformation process result to obtain the prediction probability that the sample object is an abnormal object; Obtain the loss value corresponding to the preset loss function through the prediction probability and the label; Update the model parameters of the abnormal object recognition model through the loss value to obtain the trained abnormal object recognition model.

14. The method according to any one of claims 1 to 13, characterized in that, After obtaining the probability that the target object is an abnormal object, the method further includes: In response to the probability that the target object is an abnormal object being higher than the preset probability value, perform a transaction risk warning process on the target object.

15. An abnormal object recognition device, characterized in that The device includes: An acquisition module, configured to acquire the electronic payment record data of the target object, where the electronic payment record data includes payment variables in multiple dimensions; A processing module, configured to obtain the first vector representation and the first feature weight corresponding to the payment variables in the multiple dimensions respectively; The processing module is further configured to perform feature combination based on the first vector representation and the first feature weight corresponding to the payment variables in the multiple dimensions respectively to obtain the first feature combination result; The processing module is further configured to extract features from the first feature combination result to obtain the second feature combination result; A generation module, configured to perform the first linear transformation process on the second feature combination result to obtain the linear transformation process result; The generation module is further configured to perform the non-linear transformation process on the linear transformation process result to obtain the probability that the target object is an abnormal object.

16. An electronic device, characterized in that, The electronic device includes: A memory, configured to store computer-executable instructions; A processor, configured to implement the recognition method of the abnormal object according to any one of claims 1 to 14 when executing the computer-executable instructions stored in the memory.

17. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, The computer-executable instructions or the computer program, when executed by the processor, implement the recognition method of the abnormal object according to any one of claims 1 to 14.

18. A computer program product, comprising computer-executable instructions or a computer program, characterized in that, The computer-executable instructions or the computer program, when executed by the processor, implement the recognition method of the abnormal object according to any one of claims 1 to 14.

Images