Group identification method and device, electronic equipment, storage medium and program product

Abstract

The invention provides a group identification method and device, electronic equipment, a storage medium and a program product. The method comprises the following steps: carrying out probability prediction processing on asset transfer data among objects in a group to obtain an actual asset transfer probability of the group; obtaining a group adding sequence of each object, and executing sequence coding processing on the group adding sequence of the plurality of objects to obtain a group adding sequence code of the group; executing object coding processing on the object data of the plurality of objects to obtain object codes of the group; performing transition probability prediction processing of the corresponding risk type based on the group joining sequence code of the group and the object code of the group to obtain a template asset transition probability of the group conforming to the risk type; and performing prediction processing corresponding to the risk type based on the template asset transfer probability and the actual asset transfer probability, and determining a risk type prediction result of the group. According to the invention, the accuracy of identifying the type of the group can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular, to a group identification method, device, electronic device, storage medium, and program product. Background Art

[0002] With the rapid popularization of mobile payment scenarios, the black industry is prone to operate in groups and carry out illegal activities by means of forging in-group member relationships, manipulating transaction models, etc. In order to protect user privacy and payment security, it is necessary to identify and prevent groups with specified risks. In some high-risk application scenarios (for example: risk prevention and control in payment services), higher accuracy requirements are imposed on group identification, and discovering and understanding the associations and behavior patterns between groups are crucial for preventing fraud, identifying abnormal transactions, and ensuring network security.

[0003] Related technologies determine high-risk groups by identifying displayed text information and mine and fit the information of both parties generating transfer data to identify groups, but the information utilization rate of the entire group members is not high, and the identification effect on complex groups is limited. Summary of the Invention

[0004] Embodiments of the present application provide a group identification method, device, electronic device, storage medium, and program product, which can improve the accuracy of identifying the type to which a group belongs.

[0005] The technical solution of the embodiments of the present application is implemented as follows:

[0006] Embodiments of the present application provide a data processing group identification method, and the method includes:

[0007] Perform probability mapping processing on the asset transfer data between objects in the group to obtain the actual asset transfer probability between the objects in the group;

[0008] Obtain the order of joining the group of each object in the group, and perform order encoding processing on the order of joining the group of multiple objects in the group to obtain the group joining order encoding;

[0009] Perform object encoding processing on the object data of multiple objects in the group to obtain the object encoding of the group;

[0010] Perform transfer probability prediction processing of the corresponding risk type based on the group joining order encoding and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type between the objects in the group;

[0011] Perform prediction processing corresponding to the risk type based on the template asset transfer probability and the actual asset transfer probability, and determine the risk type prediction result of the group.

[0012] An embodiment of the present application provides a group identification device, including:

[0013] A transfer probability determination module, configured to perform probability mapping processing on asset transfer data between objects in a group to obtain the actual asset transfer probability between the objects in the group;

[0014] An encoding module, configured to obtain the joining order of each object in the group, and perform order encoding processing on the joining orders of multiple objects in the group to obtain the joining order encoding of the group; perform object encoding processing on the object data of multiple objects in the group to obtain the object encoding of the group;

[0015] A template probability prediction module, configured to perform transfer probability prediction processing corresponding to the risk type based on the joining order encoding of the group and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type between the objects in the group;

[0016] A type determination module, configured to perform prediction processing corresponding to the risk type based on the template asset transfer probability and the actual asset transfer probability, and determine the risk type prediction result of the group.

[0017] An embodiment of the present application provides an electronic device, and the electronic device includes:

[0018] A memory, configured to store computer-executable instructions or computer programs;

[0019] A processor, configured to implement the group identification method provided by the embodiment of the present application when executing the computer-executable instructions or computer programs stored in the memory.

[0020] An embodiment of the present application provides a computer-readable storage medium, storing a computer program or computer-executable instructions, which are used to implement the group identification method provided by the embodiment of the present application when being executed by a processor.

[0021] An embodiment of the present application provides a computer program product, including a computer program or computer-executable instructions, and when the computer program or computer-executable instructions are executed by a processor, the group identification method provided by the embodiment of the present application is implemented.

[0022] The embodiment of the present application has the following beneficial effects:

[0023] By performing probability mapping processing on the asset transfer data between objects in a group, the actual asset transfer probability between objects in the group is obtained. The actual data transfer characteristics of the group are dynamically quantified through the actual asset transfer probability, and the discrete transfer behavior is converted into a continuous probability distribution characteristic. The joining order of each object is obtained, and order encoding processing is performed on the joining orders of multiple objects to obtain the joining order encoding of the group. Capturing the sequential dependence relationship of joining the group according to the temporal correlation of the joining order is beneficial to improving the subsequent recognition accuracy. Object encoding processing is performed on the object data of multiple objects to obtain the object encoding of the group, and group characteristics are constructed from the dimension of the object data, thereby retaining the relevant characteristics of group members. Based on the joining order encoding of the group and the object encoding of the group, transfer probability prediction processing for the corresponding risk type is performed, fusing the characteristics of the group joining order dimension and the object data dimension, and performing prediction through multi-dimensional feature fusion to obtain the template asset transfer probability of the group conforming to the risk type, which is equivalent to determining the pattern corresponding to the transfer probability when the group conforms to the risk type, and improving the accuracy of the template asset transfer probability in learning group characteristics. Comparing the pattern characteristics that should conform when conforming to the risk type with the actual asset transfer probability, performing prediction processing for the corresponding risk type based on the template asset transfer probability and the actual asset transfer probability, and determining the risk type prediction result of the group, thereby improving the accuracy of identifying the risk type of the group. Description of the Drawings

[0024] Figure 1 is a schematic diagram of the application mode of the group identification method provided by an embodiment of the present application;

[0025] Figure 2 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application;

[0026] Figure 3A is a schematic diagram of the first process of the group identification method provided by an embodiment of the present application;

[0027] Figure 3B is a schematic diagram of the second process of the group identification method provided by an embodiment of the present application;

[0028] Figure 3C is a schematic diagram of the third process of the group identification method provided by an embodiment of the present application;

[0029] Figure 4 is a schematic diagram of the fourth process of the group identification method provided by an embodiment of the present application;

[0030] Figure 5 is a schematic diagram of the deployment application provided by an embodiment of the present application;

[0031] Figure 6 is a schematic diagram of the model structure provided by an embodiment of the present application.

[0032] It should be noted that the above-mentioned "first" and "second" are only used to distinguish different solutions, and do not represent the distinction of the advantages and disadvantages of the solutions or the priority in the implementation process. Detailed implementation manners

[0033] In order to make the purpose, 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 regarded as limitations of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0034] In the following description, "some embodiments" are involved, 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.

[0035] In the following description, the terms "first / second / third" only 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 here can be implemented in an order other than that illustrated or described here.

[0036] 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 the overall module or unit that includes the function of the module or unit.

[0037] 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.

[0038] In the embodiments of the present application, the collection and processing of relevant data (such as: object data in a group) during instance applications should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and within the scope of authorization of laws and regulations and the personal information subject, carry out subsequent data use and processing.

[0039] 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.

[0040] 1) Positional Encoding: It is a key technology in natural language processing and sequence modeling, used to input the order information of sequence elements into the model. Since self-attention models such as transformers themselves do not have the ability to capture the sequence order (their attention mechanism is insensitive to the input order), positional encoding enables the model to distinguish the absolute or relative position relationships of different elements in the sequence by explicitly adding position features to the input embedding vectors. Before each sequence element is input into the model, its original embedding vector (representing semantic or feature information) is superimposed with a positional encoding vector.

[0041] 2) Data Masking: It is an information security technology aimed at protecting sensitive information and privacy data, preventing unauthorized access or leakage. By making strategic modifications or replacements to the original data, a data copy similar to the original data but without the real sensitive details is created for safe use in non-production environments such as development, testing, analysis, or training.

[0042] 3) Transformer: It is a deep learning model that adopts the attention mechanism. This mechanism can allocate different weights according to the different importance of each part of the input data. This model is mainly used in the fields of natural language processing and computer vision.

[0043] 4) Attention Mechanism: It is a mechanism that emulates the human visual and cognitive systems. It allows the neural network to pay attention to or focus on the parts that need to be learned during the fitting process. Through this mechanism, the neural network can adaptively learn the effective information in the input data.

[0044] With the rapid popularization of mobile payment scenarios, black industries are prone to operate in groups, carrying out illegal activities by means such as forging group member relationships and manipulating transaction patterns. In some high-risk application scenarios, higher accuracy requirements for group identification are imposed. Discovering and understanding the associations and behavior patterns between groups are crucial for preventing fraud, identifying abnormal transactions, and ensuring network security.

[0045] The related art determines high-risk groups by identifying and displaying text information. It is easy to hide the group name by modifying the displayed text, resulting in inaccurate identification. Mining and fitting the information of both parties generating transfer data to identify whether a group is high-risk can analyze and aggregate the interaction information between users, but the utilization rate of the information of all members of the entire group is not high, and the identification effect for complex groups is limited. It is impossible to efficiently utilize the interaction patterns of members in the group to accurately identify the group type.

[0046] The embodiments of the present application provide a group identification method, a group identification device, an electronic device, a computer-readable storage medium, and a computer program product, which can improve the accuracy of identifying the type to which a group belongs.

[0047] The following describes the exemplary applications of the electronic device provided in the embodiments of the present application. The device provided in the embodiments of the present application can be implemented as various types of terminals such as laptop computers, tablet computers, desktop computers, set-top boxes, smart phones, smart speakers, smart watches, smart TVs, in-vehicle terminals, etc., or can also be implemented as a server. Hereinafter, the exemplary applications will be described when the device is implemented as a terminal or a server.

[0048] See Figure 1 , Figure 1 is a schematic diagram of the application mode of the group identification method provided in the embodiments of the present application. To support a group identification application, for example, Figure 1 involves a server 200, a network 300, a terminal device 400, and a database 500. The terminal device 400 is connected to the server 200 through the network 300. The network 300 can be a wide area network, a local area network, or a combination of the two.

[0049] In some embodiments, the embodiments of the present application can be jointly implemented by a server and a terminal device. For example, the user can be a person skilled in the art, the server 200 is a server for performing group identification, the terminal device 400 is a terminal operated by the user, an application program capable of performing group identification is installed in the terminal device 400, object data of objects in the group is stored in the database 500, the terminal device 400 sends a group identification request to the server 200, the server 200 receives the group identification request, and performs prediction processing of the corresponding risk type based on the template asset transfer probability and the actual asset transfer probability through the group identification method provided in the embodiments of the present application, determines the risk type prediction result of the group, and sends the determined group identification result to the terminal device 400.

[0050] The group identification method provided in the embodiments of the present application can be applied to various risk scenarios that need to be identified, such as group identification scenarios for illegal information, group identification scenarios for illegal activities, etc. The following gives examples.

[0051] 1) Group identification scenario for illegal information. For example, the terminal device receives an identification request to determine whether the group to be identified belongs to a group containing illegal information. The server uses the group identification method to determine the asset transfer data between the objects in the group to be identified. The asset transfer data can be data containing illegal information, determine the actual asset transfer probability of the group to be identified, obtain the order of joining the group for each object and perform encoding processing to obtain the group order encoding of the group to be identified. Encoding processing is performed on the object data of each object to obtain the object encoding of the group to be identified, and corresponding illegal information type transfer probability prediction processing is performed based on the group order encoding and the object encoding to obtain the template asset transfer probability of the group to be identified that conforms to the illegal information type. Further, when the difference between the template asset transfer probability and the actual asset transfer probability is less than the difference threshold, it is determined that the group to be identified belongs to a group of illegal information.

[0052] 2) Group identification scenario for illegal activities. For example, the terminal device receives an identification request to determine whether the group to be identified belongs to a group containing illegal activities. The server uses the group identification method to determine the asset transfer data between the objects in the group to be identified. The asset transfer data can be information data corresponding to illegal activities, determine the actual asset transfer probability of the group to be identified, obtain the order of joining the group for each object and perform encoding processing to obtain the group order encoding of the group to be identified. Encoding processing is performed on the object data of each object to obtain the object encoding of the group to be identified, and corresponding illegal activity type transfer probability prediction processing is performed based on the group order encoding and the object encoding to obtain the template asset transfer probability of the group to be identified that conforms to the illegal activity type. Further, when the difference between the template asset transfer probability and the actual asset transfer probability is less than the difference threshold, it is determined that the group to be identified belongs to a group of illegal activities.

[0053] See Figure 2 , Figure 2 is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Figure 2 The terminal 400 shown includes: at least one processor 410, a memory 450, and at least one network interface 420. Each component in the terminal 400 is coupled together through a bus system 440. It can be understood that the bus system 440 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 440 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 440.

[0054] The processor 410 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.

[0055] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disc drives, etc. The memory 450 optionally includes one or more storage devices that are physically remote from the processor 410.

[0056] The memory 450 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 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0057] In some embodiments, the memory 450 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 will be exemplarily described below.

[0058] The operating system 451 includes system programs for handling 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 handling hardware-based tasks;

[0059] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.

[0060] In some embodiments, the device provided by the embodiments of the present application may be implemented in software. Figure 2 Shown is the group identification device 455 stored in the memory 450, which may be software in the form of programs and plugins, etc., including the following software modules: a transition probability determination module 4551, an encoding module 4552, a template probability prediction module 4553, a type determination template 4554, and a model training module 4555. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be described below.

[0061] In some embodiments, a terminal or a server may implement the group identification method provided in the embodiments of the present application by running various computer-executable instructions or computer programs. For example, the computer-executable instructions may be commands at the microprogram level, machine instructions, or software instructions. The computer program may be a native program or a software module in an operating system; it may be a local (Native) application (APPlication, APP); or it may be a small program that can be embedded in any APP, that is, a program that only needs to be downloaded to a browser environment to run. In short, the above computer-executable instructions may be instructions in any form, and the above computer programs may be application programs, modules, or plugins in any form.

[0062] The exemplary applications and implementations of the terminal provided in the embodiments of the present application will be combined to illustrate the group identification method provided in the embodiments of the present application.

[0063] Next, the group identification method provided in the embodiments of the present application will be described. As mentioned above, the electronic device implementing the group identification method in the embodiments of the present application may be a terminal, a server, or a combination of both. Therefore, the execution subject of each step will not be repeated hereinafter.

[0064] See Figure 3A , Figure 3A is the first process schematic diagram of the group identification method provided in the embodiments of the present application, and will be described in combination with Figure 3A the steps shown, Figure 3A the execution subject of Figure 1 is the server 200 in

[0065] In step 301, probability mapping processing is performed on the asset transfer data between objects in the group to obtain the actual asset transfer probability between objects in the group.

[0066] In some embodiments, step 301 may be implemented by the following method: obtaining an object node corresponding to each object; performing transformation processing on the asset transfer data between objects to obtain the asset transfer data between object nodes; and performing probability prediction processing on the asset transfer data between object nodes to obtain the actual asset transfer probability of the group.

[0067] As an example, the group contains multiple objects. The order of joining the group for each object is desensitized, and the desensitized order data for joining the group is obtained. Each desensitized object is abstracted into a simple node representation, and an object node corresponding to each object is obtained. Data desensitization is an information security technology that creates a data copy similar to the original data but without the real sensitive details by making strategic modifications or replacements to the original data, protecting sensitive information and privacy data, and preventing unauthorized access or leakage.

[0068] There is a data flow relationship between objects. The data generated by the flow relationship is used as asset transfer data, indicating that the asset transfer data is transferred from one object node to another object node. The asset transfer data between objects is transformed to obtain the asset transfer data between object nodes. For the asset transfer data between object nodes in a group, the actual asset transfer probability of the asset transfer data between object nodes is calculated. According to the number of objects in the group, a matrix in the dimension of the number of objects is constructed. The calculated actual asset transfer probability between object nodes is used as an element in the object dimension matrix, and the transfer probability matrix of the group is obtained as the actual asset transfer probability of the group.

[0069] Through the embodiments of the present application, privacy protection is achieved through data desensitization and node abstraction. According to the asset transfer data between object nodes, the flow data between object nodes is quantified, and through matrix operations, the complex interaction relationship is quantified into an actual asset transfer probability matrix, which can effectively ensure information security and provide a quantifiable basis for group identification based on the behavior of asset transfer data.

[0070] In some embodiments, the probability prediction process is performed on the asset transfer data between object nodes, and the actual asset transfer probability of the group can be achieved through the following method: the following process is performed for each object node: the sum of the asset transfer data flowing out from the object node to all other object nodes is calculated to obtain the total asset transfer of the object node; when the total asset transfer is less than one, the asset transfer data between the object node and each other object node is determined as the asset transfer probability between the object node and each other object node; when the total asset transfer is not less than one, the ratio between the asset transfer data between the object node and each other object node and the total asset transfer is determined, and the ratio is determined as the asset transfer probability between the object node and each other object node; the asset transfer probabilities between the object node and at least one other object node are combined to form the asset transfer probability of the corresponding object node; the asset transfer probabilities corresponding to multiple object nodes are combined to form the actual asset transfer probability of the group.

[0071] As an example, for each object node, the sum of the asset transfer data flowing out from the object node to all other object nodes is calculated to obtain the total asset transfer of the object node. When the total asset transfer is less than one, the asset transfer data between the object node and each other object node is used as the asset transfer probability between the object node and each other object node. When the total asset transfer is not less than one, the ratio between the asset transfer data between the object node and each other object node and the total asset transfer is obtained, and the ratio is determined as the asset transfer probability between the object node and each other object node, which can be achieved through formula (1), and the following is a specific description.

[0072]

[0073] wherein, i outflow represents the asset transfer data transferred from node i to node j, and i outflow_total represents the total asset transfer amount transferred from node i to other nodes. Node i is defined as the node with the order of i in the joined group. max(i outflow_total , 1) represents taking the larger value between i outflow_total and 1 to ensure that the denominator is at least 1 and avoid errors in the formula caused by a zero denominator. m ij represents the transfer probability from node i to node j. The larger the transfer probability value, the greater the attention score from node i to node j in the subsequent calculation of the attention mechanism.

[0074] The asset transfer probabilities between an object node and at least one other object node are combined to form the asset transfer probability of the corresponding object node. According to the joining order of the object nodes, the asset transfer probabilities corresponding to multiple object nodes are combined to form the actual asset transfer probability of the group. For example, if a group has N members, the transfer probability matrix is an N*N-dimensional matrix, denoted as matrix M, M = [m ij N*N , where m ij represents the element in the i-th row and j-th column of the matrix. The value of the element m ij in the matrix can be calculated by formula (1). Taking the asset transfer probabilities calculated for each object node as the element values in the matrix, the transfer probability matrix filled with asset transfer probabilities is obtained as the actual asset transfer probability of the group.

[0075] For example, in a payment scenario, m ij represents the amount of payment from the user corresponding to node i to the user corresponding to node j, that is, the asset transfer data. In a group with five nodes A, B, C, D, and E, the asset transfer data (transfer values) between the nodes in this scenario are shown in Table 1 below, which will be specifically described as follows.

[0076] Table 1 Transfer values between nodes

[0077] Outflow node Inflow node Transfer value A B 1000 A C 2000 A D 3000 A E 4000 B D 1000 B E 2000 C E 500 D E 2000 E D 1000

[0078] Based on formula (1) and the asset transfer data (transfer values) in Table 1, calculate the asset transfer probabilities of each object node. Then, the probability transfer matrix corresponding to the group in Table 1 is expressed as formula (2), which will be specifically described as follows.

[0079]

[0080] ​Among them, each element value in the matrix M is calculated and determined according to the asset transfer data between nodes shown in formula (1) and Table 1, and the row and column order of the matrix M is the joining group order corresponding to the object nodes.

[0081] Through the embodiments of the present application, the probabilistic conversion of asset transfer data is realized through normalization processing, which is used for quantitative analysis of the relationship of the actual asset transfer data intensity. By judging the numerical value of the total number of asset transfers, the discrete asset transfer values are converted into a standard probability distribution, which provides a mathematical basis for subsequent attention weight allocation, and the denominator protection mechanism is used to avoid division-by-zero errors, ensuring the robustness of the algorithm in sparse data scenarios.

[0082] In step 302, obtain the joining group order of each object in the group, and perform order encoding processing on the joining group orders of multiple objects in the group to obtain the joining group order encoding of the group.

[0083] In some embodiments, step 302 can be implemented by the following method: for each object, perform the following processing: obtain the object node corresponding to the object, and use the joining group order of the object as the joining group order of the object node; perform order encoding processing on the joining group order of the object node to obtain the encoding value of the object node in each dimension; form the joining group order encoding of the object node by combining the encoding values of the object node in multiple dimensions; form the joining group order encoding of the group by combining the joining group order encodings of the object nodes corresponding to multiple objects.

[0084] As an example, the joining group order encoding of the group includes the position encodings of multiple objects, and the position encodings include multiple dimensions. Obtain the object node corresponding to the object, and use the joining group order of the object as the joining group order of the object node. For example, convert five user objects in the group to be recognized into node representations [A, B, C, D, E], and their corresponding joining group orders are [5, 4, 3, 2, 1]. Then, after arranging them in the order of joining the group, the object nodes can be represented as [E, D, C, B, A].

[0085] As an example, perform order encoding processing on the joining group order of the object node to obtain the encoding value of the object node in each dimension. Encode the joining group order of each object node into a position vector, and represent the position corresponding to each element of the joining group order with a coding vector. The order encoding processing is implemented through a transformer model. Position encoding is a key technology in natural language processing and sequence modeling, which is used to input the order information of sequence elements into the model. Position encoding enables the model to distinguish the absolute or relative position relationships of different elements in the sequence by explicitly adding position features to the input embedding vector. Before each sequence element is input into the model, its original embedding vector (representing semantic or feature information) will be superimposed with a position encoding vector.

[0086] The encoded values of an object node in multiple dimensions form the additive group order encoding of the object node. It is assumed that the dimension of the feature vector of the object node feature is d model , for a certain object node, the additive group order is pos, and for a certain position i in the vector of the additive group order encoding, the additive group order encoding is also a d model -dimensional vector. The additive group order encodings of the object nodes corresponding to multiple objects are combined according to the additive group order of the object nodes to construct the additive group order encoding matrix of the group. Filling the encoded values of the additive group order encoding into the additive group order encoding matrix of the group can be expressed by formula (4), which is specifically described below.

[0087]

[0088] Among them, P is the additive group order encoding matrix of the group, PE is the additive group order encoding of each object node, d model represents the backbone dimension of the transformer model, represents the unified dimension of each feature vector, and pos is the position index in the additive group order sequence (starting from 0).

[0089] Through the embodiments of the present application, the additive group order of the object nodes is converted into a computable vector representation through position encoding, multi-dimensional position features are generated, and the representation of the group structure is realized through the construction of the encoding matrix. There is a temporal correlation between object nodes, and through unified dimension mapping, heterogeneous data can be processed in the feature space, providing a structured input containing sequential semantics for the subsequent attention mechanism.

[0090] In some embodiments, sequential encoding processing is performed on the additive group order of the object nodes, and the encoded values of the object nodes in each dimension can be obtained through the following method: the following processing is performed for each dimension: when the dimension number is even, sine encoding processing is performed on the additive group order and encoding of the object nodes to obtain the encoded value of the object node in the dimension; when the dimension number is odd, cosine encoding processing is performed on the additive group order and number of the object nodes to obtain the encoded value of the object node in the dimension.

[0091] As an example, for an object node with an additive group order of pos, and for a certain position i in the vector of the additive group order encoding, when the encoded value of the position dimension is even, sine encoding processing is performed on the additive group order and encoding of the object nodes, and when the number is odd, cosine encoding processing is performed on the additive group order and number of the object nodes. The encoded value of the object node in the dimension can be realized through formula (3), which is specifically described below.

[0092]

[0093] Among them, d modelRepresents the backbone dimension of the transformer model, which is the unified dimension of each feature vector. All inputs will be mapped by the transformer model into the same dimensional space (i.e., d model dimension), for example, when d model takes the value of 512, it means that the feature vector and position encoding value of each node are both 512-dimensional floating-point arrays. pos is the position index in the additive group order sequence (starting from 0), and 10000 is a hyperparameter used to control the wavelength. In formula (3), when the node position i is even, sine position encoding is used; when the node position i is odd, cosine position encoding is used. By combining the sine position encoding and the cosine position encoding, the encoding value of the object node in the position dimension is obtained.

[0094] Through the embodiments of the present application, by combining the additive group order of the object node with the encoding, sine encoding processing is performed on the nodes in the even dimensions, and cosine encoding processing is performed on the nodes in the odd dimensions, so as to generate encoding values for the object node in each dimension. Effectively map the node position information to the same dimension, ensure the consistency of the feature vector and the position encoding value, and use the hyperparameter to control the wavelength to optimize the encoding effect, providing an accurate encoding representation for the object node.

[0095] In step 303, object encoding processing is performed on the object data of multiple objects in the group to obtain the object encoding of the group.

[0096] As an example, when performing object encoding processing on the object data of multiple objects in the group, the object data is the basic data information of the object. For example, the inherent attributes such as the user's age and registration time. The object node with the additive group order of pos is subjected to object encoding processing to obtain the object encoding of the group, denoted as x pos .

[0097] In step 304, based on the additive group order encoding of the group and the object encoding of the group, transfer probability prediction processing for the corresponding risk type is performed to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

[0098] As an example, the risk type here can be risky, or it can be a specific risk level, such as high risk, medium risk, low risk, etc.

[0099] In some embodiments, referring to Figure 3B , Figure 3B is the second process schematic diagram of the group identification method provided by the embodiments of the present application. Figure 3A The step 304 in Figure 3B can be implemented by executing step 3041 to step 3042 in

[0100] In step 3041, the group addition order encoding and the object encoding are fused to obtain a first fused encoding.

[0101] As an example, the group addition order encoding and the object encoding are fused to obtain a first fused encoding, which is expressed as formula (5), and the following is a specific description.

[0102]

[0103] Among them, the object encoding of the object node with the group addition order of pos is represented as x pos , and the group addition order encoding corresponding to the same group addition order is represented as p pos , and the obtained first fused encoding after fusion is represented as The first fused encodings after fusing all nodes are combined to obtain a first fused encoding matrix of the group, which is represented as Among them, The first fused encoding matrix fuses the group addition order information and the basic data information of each object node in the group.

[0104] For example, for the node representation [A, B, C, D, E], the corresponding group addition orders are [5, 4, 3, 2, 1], then the corresponding first fused encoding matrix can be represented as

[0105] In step 3042, a transfer probability prediction process for the corresponding risk type is performed on the first fused encoding to obtain the template asset transfer probability of the group conforming to the risk type.

[0106] As an example, a transfer probability prediction process for the corresponding risk type is performed on the first fused encoding. The transfer probability prediction process is implemented by a group recognition model composed of a multi-layer attention mechanism structure. In some embodiments, refer to Figure 6 , Figure 6 is a schematic diagram of the model structure provided by the embodiments of the present application. In the embodiments of the present application, taking three attention mechanism layers as an example, the first fused encoding is sequentially input into the attention mechanism layer 601, the attention mechanism layer 602, and the attention mechanism layer 603, and the output of the three attention mechanism layers is normalized by an additional attention layer 604 to determine the finally output attention score matrix as the template asset transfer probability of the group conforming to the risk type.

[0107] Through the embodiments of the present application, the group addition sequence encoding and the object encoding are fused to obtain a first fused encoding, integrating the group addition sequence information and the basic data information of each object node, and using a group recognition model with a multi-layer attention mechanism structure to perform a transfer probability prediction process on the first fused encoding to obtain a template asset transfer probability that conforms to the risk type. Combining the order of the nodes and the basic data, through the processing of group recognition, the transfer probability between the objects in the group is predicted, which helps to identify the behavior pattern corresponding to the risk type of the group.

[0108] In some embodiments, refer to Figure 3C , Figure 3C which is the third process schematic diagram of the group recognition method provided by the embodiments of the present application. Figure 3B Step 3042 in Figure 3C can be implemented by executing steps 30421 to 30422 in

[0109] Specifically, in step 30421, a mapping process based on a multi-layer attention mechanism is performed on the first fused encoding to obtain a first attention matrix.

[0110] In some embodiments, step 30421 can be implemented by the following method: through the nth attention network in the N-level cascaded attention network, a mapping process is performed on the input of the nth attention network to obtain a mapping result output by the nth attention network; the mapping result output by the nth attention network is used as the input of the (n + 1)th attention network to continue the mapping process to obtain a mapping result output by the (n + 1)th attention network; the mapping result output by the Nth attention network is used as the first attention matrix.

[0111] As an example, n is an integer variable starting from 1 and increasing. The value range of n is 1 ≤ n < N. When n takes the value of 1, the input of the nth attention network is the first fused encoding. When n takes the value of 2 ≤ n < N, the input of the nth attention network is the mapping result output by the (n - 1)th attention network. A mapping process is performed on the input of the nth attention network to obtain a mapping result of the nth attention network, and the mapping result of the nth attention network is used as the input of the (n + 1)th attention network to continue the mapping process. Refer to Figure 6 , the mapping result output by the first attention mechanism layer 601 (attention network) is used as the input of the attention mechanism layer 602, and the mapping result output by the attention mechanism layer 602 (attention network) is used as the input of the attention mechanism layer 603 (attention network) until the mapping process of the last attention mechanism is completed. The attention mechanism layer learns the transfer probability between the object nodes. The greater the transfer probability value, the greater the corresponding attention score. The mapping result of the last attention mechanism is used as the first attention matrix.

[0112] Through the embodiments of the present application, through the layer-by-layer mapping processing of the multi-layer attention network, the input first fusion encoding is gradually transformed. Each layer of the attention network maps the input and outputs a mapping result, which serves as the input for the next layer of the attention network. The first attention matrix generated by the iterative mapping of the multi-layer attention network can learn and reflect the transition probability between object nodes. Nodes with higher attention scores represent stronger correlations between nodes.

[0113] In some embodiments, the mapping process is performed on the input of the nth layer attention network in the N-level cascaded attention network, and the mapping result of the nth layer attention network can be achieved through the following method: performing a mapping process based on the attention mechanism on the input of the nth layer attention network to obtain a first attention result; performing a normalization process based on the residual mechanism on the first attention result to obtain a first normalized result; performing a feed-forward process on the first normalized result to obtain a first feed-forward result; performing a normalization process based on the residual mechanism on the first feed-forward result to obtain the mapping result of the nth layer attention network.

[0114] As an example, taking the first fusion encoding as the input of the nth layer attention network to perform the mapping process based on the attention mechanism, refer to Figure 6 , in the attention mechanism layer 601 (the nth layer attention network), the first fusion encoding is input into the attention layer 6011 for mapping processing. By calculating the similarity between the first fusion encodings (e.g., dot product attention or scaled dot product attention), an attention weight matrix is generated, and the input first fusion encoding is weighted and aggregated as the first attention result. The first attention result is input into the residual connection and normalization layer 6012, and the output of the attention layer 6011 is normalized based on the residual mechanism with its input (the first fusion encoding), and element-wise addition is performed for residual connection, and then layer normalization is performed on the added result to obtain a normalized result, outputting a stable and smooth feature representation as the first normalized result.

[0115] Performing a feed-forward process on the first normalized result, inputting the first transformed encoding into the feed-forward group recognition layer 6013, performing two linear transformations on the normalized features. The first linear transformation expands the feature dimension and introduces non-linearity through an activation function, and the second linear transformation restores the dimension to the original size, generating a high-order non-linear feature representation to obtain the first feed-forward result after non-linear transformation. Performing a second normalization process based on the residual mechanism on the first feed-forward result, inputting it into the residual connection and normalization layer 6014 for a second transformation process, adding the output of the feed-forward group recognition layer 6013 to its input again, and performing layer normalization to output a further optimized feature representation to obtain the mapping result of the nth layer attention network.

[0116] Through the embodiments of the present application, by calculating the similarity between the first fusion encodings and generating an attention weight matrix, important information in the input encodings can be identified and strengthened, enabling the model to pay more attention to key features, and improving efficiency and accuracy. In the first transformation process, the output of the attention layer is added to the input through a residual connection, ensuring the integrity of information. Through normalization processing, the feature representation becomes more stable and smooth, helping the model resist the problem of gradient vanishing. By performing a non-linear transformation process on the first transformed encoding, an activation function and a dimensional transformation are introduced to increase the complexity and diversity of the feature representation. Through an N-level cascaded attention network, the stability and non-linearity of the feature representation are improved, and the output of the attention mechanism is optimized.

[0117] Continue to refer to Figure 3C In step 30422, a probability mapping process is performed on the first attention matrix to obtain the template asset transfer probability that conforms to the risk type among the objects within the group.

[0118] As an example, in addition to including a multi-level cascaded attention network, the group recognition model further adds an additional attention layer to perform a probability mapping process on the first attention matrix. Refer to Figure 6 The output result (the first attention matrix) of the attention mechanism layer 603 is used as the input of the additional attention layer 604 to obtain the template asset transfer probability that the group conforms to the risk type.

[0119] Through the embodiments of the present application, a multi-level attention network is used to deeply analyze the first fusion encoding. Through the interaction between layers, important information in the encoding is extracted to form the first attention matrix, which represents the degree of association between different encodings. A probability mapping process is performed on the first attention matrix to convert the information in the attention matrix into a probability form, obtaining the template asset transfer probability that the group conforms to the risk type, reflecting the transfer trend of the group under the risk type, effectively classifying and predicting the group, and identifying the template asset transfer probability of the group that conforms to a specific type, providing data support for subsequent decision-making.

[0120] In some embodiments, step 30422 can be implemented by the following method: sequentially multiply the first attention matrix, the query parameter matrix, the transposed matrix of the key parameter matrix, and the transposed matrix of the first attention matrix to obtain the first matrix multiplication result; perform normalization processing on the first matrix multiplication result based on the dimension of the additive group order encoding to obtain a normalized matrix; perform maximum likelihood processing on the normalized matrix to obtain the template asset transfer probability that conforms to the risk type among the objects within the group.

[0121] As an example, the additional attention layer 604 performs a probability mapping on the first attention matrix output by the attention mechanism layer 603. For the first attention matrix, the query parameter matrix, the transposed matrix of the key parameter matrix, and the transposed matrix of the initial attention matrix are sequentially multiplied, and the multiplication result is normalized based on the dimension (d model ) of the additive group order coding, and the maximum likelihood processing is performed on the normalized result to obtain the template asset transfer probability that the group conforms to the risk type, which can be implemented by formula (6), and the following is a specific description.

[0122]

[0123] Among them, the activation function softmax is the maximum likelihood processing, which finds a set of parameter values to maximize the probability of the observed data. Y is the first attention matrix output by the multi-layer attention mechanism layer, with a dimension of N*d model , N is the number of nodes, d model is the dimension of the additive group order coding, W Q is the value vector (Query) weight matrix, is the transpose of the key vector (Key) weight matrix, scales the dot product result to prevent the dot product value from being too large and causing the gradient to disappear, and obtains the template asset transfer probability that conforms to the risk type between the objects within the group.

[0124] Through the embodiments of the present application, by sequentially multiplying the first attention matrix and other related matrices, a comprehensive information matrix is obtained. The matrix is normalized according to the dimension of the additive group order coding to ensure that the relative size relationship between each element is reasonably adjusted. Through the maximum likelihood processing, the template asset transfer probability that conforms to the risk type between the objects within the group is calculated, and the information in the attention matrix is converted into a probability form, enabling the model to quantify the transfer possibility of the group under the risk type, providing a numerical basis for group identification and classification.

[0125] Continue to refer to Figure 3A , in step 305, based on the template asset transfer probability and the actual asset transfer probability, the prediction process corresponding to the risk type is performed to determine the risk type prediction result of the group.

[0126] As an example, based on the template asset transfer probability and the actual asset transfer probability, the prediction process for the corresponding risk type is performed. The matching degree of the risk type corresponding to the group is determined by the difference between the template asset transfer probability and the actual asset transfer probability, and the risk type prediction result of the group is determined. Calculate the difference between the template asset transfer probability and the actual asset transfer probability, and take the sum of the squares of the differences between the template asset transfer probability and the actual asset transfer probability as the difference between the template asset transfer probability and the actual asset transfer probability, which can be achieved by formula (7), and the following is a specific description.

[0127]

[0128] Among them, ∈ represents the difference between the attention score matrix (template asset transfer probability) and the transfer probability matrix (actual asset transfer probability). Y is the attention score matrix (template asset transfer probability), M is the transfer probability matrix (actual asset transfer probability), and F is the operation of summing the squared differences of matrix elements, which aggregates the element-wise errors of the matrix into a scalar value to measure the overall difference. When the difference between the template asset transfer probability and the actual asset transfer probability is less than the difference threshold, the characteristics of the group and the risk type are more similar, satisfying a certain threshold comparison relationship, then it is determined that the current group belongs to the risk type, and taking the current group belonging to the risk type as the risk type prediction result of the group.

[0129] Through the embodiments of the present application, based on the template asset transfer probability and the actual asset transfer probability, the prediction process for the corresponding risk type is performed, the difference between the template asset transfer probability and the actual asset transfer probability is determined, and by comparing the difference between the template asset transfer probability and the actual asset transfer probability, it is judged whether the group belongs to the risk type. If the difference is less than the preset difference threshold, it is considered that the similarity between the group and the characteristics of the risk type is relatively high, and it is determined that the risk type prediction result of the group belongs to high risk. The judgment method based on the difference threshold provides a quantitative standard for group classification. When the difference is less than the type threshold, according to the risk level of the risk type, it is convenient to take corresponding measures for the group.

[0130] In some embodiments, the prediction process of the template asset transfer probability corresponding to the risk type is implemented by calling a group identification model. Before performing the transfer probability prediction process of the corresponding risk type based on the group addition order encoding of the group and the object encoding of the group to obtain the template asset transfer probability of the group conforming to the risk type, the group identification model is trained by the following method: performing a probability prediction process on the asset transfer data between sample objects within the sample group to obtain the historical actual asset transfer probability between sample objects within the sample group, where the sample group is a historical group conforming to the risk type; obtaining the second addition order of each sample object within the sample group, and performing an order encoding process on the second addition orders of multiple sample objects within the sample group to obtain the addition order encoding of the sample group; performing an object encoding process on the object data of multiple sample objects within the sample group to obtain the object encoding of the sample group; the group identification model performs the following processes: performing a transfer probability prediction process of the corresponding risk type based on the addition order encoding and object encoding of the sample group to obtain the historical template asset transfer probability of the sample objects within the sample group conforming to the risk type; performing a fitting process based on the difference between the historical actual asset transfer probability and the historical template asset transfer probability to obtain a first loss, and updating the group identification model based on the first loss.

[0131] As an example, the sample group is a historical group conforming to the risk type, and a historical group sample conforming to the risk type is selected to train the group identification model. Each historical group is used as a sample, and the asset transfer data between sample objects within the sample group is calculated to determine the historical transfer probability of the sample group. Obtaining the second addition order of each sample object, and performing the same encoding process on the sample objects. The principle of performing an order encoding process on the second addition orders of multiple sample objects and performing an object encoding process on the object data of multiple sample objects is the same as the principle of performing an order encoding process and an object encoding process on the above objects, which will not be elaborated here. Performing a transfer probability prediction process of the corresponding risk type based on the addition order encoding and object encoding of the sample group to obtain the historical template asset transfer probability of the sample group conforming to the risk type is the same as the principle of the prediction process of the template asset transfer probability of the group, which will not be elaborated here.

[0132] As an example, performing a fitting process based on the difference between the historical actual asset transfer probability and the historical template asset transfer probability to obtain the first loss of the group identification model. Calculating the difference between the attention score matrix Y (historical template asset transfer probability) and the transfer probability matrix M (historical actual asset transfer probability) of the sample group to perform a fitting process to obtain a first loss function L, and training the group identification model based on the first loss function L. The first loss function L can be implemented by formula (8), which is specifically described below.

[0133]

[0134] Among them, S is the number of samples in the entire sample group, and M i is the transition probability matrix of sample group i (historical actual asset transfer probability), and Y i is the attention score matrix of sample group i (historical template asset transfer probability). F is an operation of summing the squared differences of matrix elements, which aggregates the element-wise squared errors of the matrix into a scalar value, used to measure the overall difference between the network prediction and the true transfer probability. The group identification model is updated by minimizing the loss function L, and the updated group identification model is used for offline deployment to perform group identification of the corresponding risk type.

[0135] Through the embodiments of the present application, by analyzing the sample groups that meet the risk type, the asset transfer probability between the objects within the historical groups is determined. By obtaining the joining order of each object and performing encoding processing, the joining order encoding and object encoding of the historical group are obtained, and based on the joining order encoding and object encoding, the historical template asset transfer probability of the sample group that meets the risk type is predicted. By comparing the difference between the historical transfer probability and the historical template asset transfer probability, the first loss is calculated. The loss value reflects the accuracy of the prediction of the group identification model. The group identification model is updated by minimizing the loss value to improve the prediction ability, enabling the group identification model to learn the group characteristics of the risk type and accurately perform group identification in practical applications.

[0136] The group identification method provided by the embodiments of the present application has the following beneficial effects:

[0137] Privacy protection is achieved by desensitizing the data of the order in which objects are added to a group. The objects are abstracted as object nodes. Based on the asset transfer data between object nodes, the flow data between object nodes is quantified, and through matrix operations, the complex interaction relationships are quantified into an actual transfer probability matrix, which can effectively ensure information security. The asset transfer probability of object nodes is determined by the total number of asset transfers, forming the actual asset transfer probability of the group. The discrete asset transfer values are converted into a standard probability distribution, providing a mathematical basis for subsequent attention weight allocation, and the denominator protection mechanism is used to avoid division-by-zero errors, ensuring the robustness of the algorithm in sparse data scenarios. The order of adding objects to the group and the object data are encoded, and the encoded order of adding objects and the object encoding obtained from the encoding process are fused. Through the construction of the encoding matrix, the representation of the group structure is realized, effectively mapping the node position information to the same dimension and ensuring the consistency of the feature vector and the position encoding value, providing an accurate encoding representation for object nodes. For the first fused encoding, the transfer probability prediction of the corresponding risk type is performed, and the mapping process based on the multi-layer attention mechanism is executed. The first attention matrix generated by the iterative mapping of the multi-layer attention mechanism can learn and reflect the transfer probability between object nodes, forming the first attention matrix. The probability mapping process is performed on the first attention matrix to convert the information in the attention matrix into a probability form, obtaining the template asset transfer probability of the group that conforms to the risk type, reflecting the transfer trend of the group under the risk type, effectively classifying and predicting the group, identifying the template asset transfer probability of the group that conforms to a specific type, and providing data support for subsequent decision-making. Based on the difference between the template asset transfer probability and the actual asset transfer probability of the group, it is judged whether the group belongs to the risk type, improving the accuracy of group identification. The judgment method based on the difference threshold provides a quantitative standard for group classification. When the difference is less than the type threshold, according to the risk level of the risk type, corresponding measures can be taken for the group.

[0138] Next, an exemplary application of the embodiments of the present application in a practical application scenario will be described.

[0139] With the rapid popularization of mobile payment scenarios, black industries are prone to operate in groups and carry out illegal activities by forging group member relationships, manipulating transaction models, etc. In some high-risk application scenarios, higher accuracy requirements for group identification are imposed. Discovering and understanding the associations and behavior patterns between groups are crucial for preventing fraud, identifying abnormal transactions, and ensuring network security. Related technologies determine high-risk groups by identifying displayed text information, but it is easy to hide the group name by modifying the displayed text, resulting in inaccurate identification. Mining and fitting the information of both parties generating transfer data to identify whether a group has high risks can analyze and aggregate the interaction information between users, but the information utilization rate of all members of the entire group is not high, and the identification effect for complex groups is limited. It is unable to efficiently utilize the interaction patterns of members in the group to accurately identify the group type.

[0140] In the embodiments of the present application, by considering the natural structural feature of the order of joining and the corresponding transaction mode among group members, each group is taken as a sample. In some embodiments, refer to Figure 5 , Figure 5 is a schematic diagram of the deployment application provided by the embodiments of the present application. The user 503 identifies a group through the offline deployment module 501 and the risk control inference module 502. The corresponding model structure and parameter information are set in the model structure and parameter deployment module 5011 and deployed to the online system for subsequent inference and prediction. The group member joining order acquisition module 5021 acquires the desensitized joining order of each group and performs vector encoding on the order, which is fused with the basic features of the group nodes in the group member feature matrix construction module 5022 to construct a group feature matrix. The group feature matrix is predicted by a trained neural network model to obtain the attention score matrix of the group. And the group transition probability matrix construction module 5023 calculates the transfer value between each group object to determine the actual transfer probability matrix of the group. The model inference module 5024 calculates the difference between the actual transfer probability matrix and the attention score matrix for inference to accurately identify the group type, and corresponding penalty operations are executed through the group handling module 5031 according to the identification result.

[0141] The following is described in conjunction with the accompanying drawings. Refer to Figure 4 , Figure 4 is the fourth process schematic diagram of the group identification method provided by the embodiments of the present application. The execution subject can be a terminal device, a server, or a combination of both. In the embodiments of the present application, taking Figure 1 the server 200 in Figure 4 as the execution subject as an example, the following steps shown in

[0142] In step 401, according to the transfer values between objects in the group to be recognized, determine the transfer probability matrix of the group to be recognized.

[0143] As an example, the group to be recognized is a group with multiple members. Each member in the group is regarded as an object node in the group, and there is a data flow relationship between members. Then there is a transfer value of the data flow between object nodes, indicating that the data is transferred from one object node to another object node. According to the transfer values between object nodes, determine the transfer probability matrix of the group to be recognized. The transfer probability matrix is a matrix corresponding to the dimension of the number of members in the group. For example, if there are N members in a group, the transfer probability matrix is an N*N-dimensional matrix, denoted as matrix M, M = [m ij N*N , where, m ij represents the element in the i-th row and j-th column of the matrix. The value of the element m ij in the matrix can be calculated by formula (1), which is specifically described below.

[0144]

[0145] where, i outflow represents the specific value of the transfer from node i to node j, i outflow_total represents the sum of the values of the transfers from node i to other nodes. Node i is defined as the node with the order of i in the group. max(i outflow_total , 1) represents taking the larger value between i outflow_total and 1 to ensure that the denominator is at least 1 and avoid formula errors caused by a zero denominator. m ij represents the transfer probability from node i to node j. The larger the transfer probability value, the larger the attention score from node i to node j in the subsequent calculation of the attention mechanism.

[0146] For example, in a payment scenario, m ij represents the amount of payment from the user corresponding to node i to the user corresponding to node j. There are five nodes A, B, C, D, and E in a group. The transfer values between nodes in this scenario are shown in Table 1 below, which is specifically described below.

[0147] Table 1 Transfer values between nodes

[0148]

[0149]

[0150] Based on formula (1) and the transfer values in Table 1, calculate the transfer probability of each node. Then the probability transfer matrix corresponding to the group in Table 1 is expressed as formula (2), which is specifically described below. ​

[0151]

[0152] Among them, each element value in the matrix M is calculated and determined according to the transfer value between nodes shown in formula (1) and Table 1. The row and column order of the matrix M is the addition group order of the nodes.

[0153] In step 402, the addition group order of each object in the group to be recognized is encoded to obtain the position encoding matrix of the group to be recognized.

[0154] As an example, for each object in the group to be recognized, the addition group order of each object is desensitized to obtain the desensitized addition group order data. Desensitization processing is to hide some sensitive information, abstract each object into a simple node representation, and only obtain the sequence of adding groups of nodes in the group to be recognized. Data desensitization is an information security technology that creates a data copy similar to the original data but without real sensitive details by strategically modifying or replacing the original data, protecting sensitive information and privacy data from unauthorized access or leakage.

[0155] For example, converting five user objects in the group to be recognized into node representations [A, B, C, D, E], and their corresponding group entry orders are [5, 4, 3, 2, 1], then after arranging them in the order of joining the group, the object nodes can be represented as [E, D, C, B, A].

[0156] As an example, sequential encoding processing is performed on the desensitized addition group order. The addition group order of each object node is encoded into a position vector, and the position corresponding to each element in the addition group order sequence is represented by a coding vector. The sequential encoding processing is implemented through a transducer model. Position encoding is a key technology in natural language processing and sequence modeling, used to input the order information of sequence elements into the model. Position encoding enables the model to distinguish the absolute or relative position relationships of different elements in the sequence by explicitly adding position features to the input embedding vector. Before each sequence element is input into the model, its original embedding vector (representing semantic or feature information) is superimposed with a position encoding vector. It is assumed that the dimension of the feature vector of the object node feature is d model , the addition group order of a certain object node is pos, and a certain position in the vector of the addition group order encoding is i, then the addition group order encoding is also a d model vector, and the position encoding can be implemented through formula (3), which is specifically described below.

[0157]

[0158] Among them, d modelDenotes the backbone dimension of the transformer model, representing the unified dimension of each feature vector. All inputs will be mapped by the transformer model into the same dimensional space (i.e., d model dimension), for example, when d model takes the value of 512, it means that the feature vector and the position encoding vector of each node are both 512-dimensional floating-point arrays. pos is the position index in the group addition order sequence (starting from 0), and 10000 is a hyperparameter used to control the wavelength. In formula (3), when the node position i is even, sine position encoding is used; when the node position i is odd, cosine position encoding is used. Combine the position encoding vectors of the positions in the group addition order corresponding to all nodes at the pos position in the group, and obtain the position encoding matrix of the group to be recognized, which is expressed as formula (4), and the following is a specific description.

[0159]

[0160] Among them, P is the position encoding matrix of the group to be recognized (corresponding to the above group), PE is the position encoding of each object node, and d model denotes the backbone dimension of the transformer model, representing the unified dimension of each feature vector, and pos is the position index in the group addition order sequence (starting from 0).

[0161] In step 403, fuse the position encoding matrix and the basic feature matrix of the group to be recognized to construct the fused feature matrix of the group to be recognized.

[0162] As an example, perform feature encoding processing on the basic data information of the objects in the group to be recognized. The basic data information of the object is the feature of the object node itself that is not in the group addition order. For example, the inherent attributes such as the user's age and registration time. Denote the technical feature after encoding the object node with the group addition order of pos as x pos , and denote the position encoding corresponding to the same group addition order as p pos , and obtain the fused feature encoding after fusion, which is expressed as formula (5), and the following is a specific description.

[0163]

[0164] Merge the fused feature encodings of all nodes in the group to be recognized to obtain the fused feature matrix of the group to be recognized, which is expressed as Among them, The fused feature matrix fuses the group addition order information and the basic data information of each object node in the group to be recognized.

[0165] For example, for the node representation [A, B, C, D, E], whose corresponding group entry orders are [5, 4, 3, 2, 1], the corresponding fused feature matrix can be expressed as

[0166] In step 404, the fused feature matrix is processed by a neural network model to obtain an attention score matrix of the group to be recognized.

[0167] As an example, when the fused feature matrix is processed by a neural network model, the structure of the neural network model includes multiple attention mechanisms. In some embodiments, Figure 6 is a schematic diagram of the model structure provided by the embodiments of the present application. In the embodiments of the present application, taking three attention mechanism layers as an example, the fused feature matrix is sequentially input into the attention mechanism layer 601, the attention mechanism layer 602, and the attention mechanism layer 603, and the outputs of the three attention mechanism layers are normalized by an additional attention layer 604 to determine the final output result as the attention score matrix of the group to be recognized.

[0168] As an example, in the attention mechanism layer 601, the fused features in the fused feature matrix are input into the attention layer 6011 to generate an attention weight matrix by calculating the similarity between feature vectors (for example, dot product attention or scaled dot product attention), and the input features are weighted and aggregated to obtain a first mapping result. The first mapping result is input into the residual connection and normalization layer 6012 to perform a first transformation process on the output of the attention layer 6011 and its input (fused feature encoding), add them element by element for residual connection, and then perform layer normalization on the added result to obtain a normalized result, output a stable and smooth feature representation, and fuse the correction of the attention mechanism while retaining the original input information.

[0169] The normalized result is input into the feed-forward neural network layer 6013 to perform two linear transformations on the normalized features: the first linear transformation expands the feature dimension and introduces non-linearity through an activation function, and the second linear transformation restores the dimension to the original size to generate a high-order non-linear feature representation with stronger representation ability. The generated high-order non-linear feature representation is input into the residual connection and normalization layer 6014 for a second transformation process, add the output of the feed-forward neural network layer 6013 and its input again, and perform layer normalization to output a further optimized feature representation and use it as the input of the attention mechanism layer 602.

[0170] The attention layer 6021, residual connection and normalization layer 6022, feed-forward neural network layer 6023, and residual connection and normalization layer 6024 in the attention mechanism layer 602 perform the same processing as in the attention mechanism layer 601, and use the output result of the attention mechanism layer 602 as the input of the attention mechanism layer 603. The attention layer 6031, residual connection and normalization layer 6032, feed-forward neural network layer 6033, and residual connection and normalization layer 6034 in the attention mechanism layer 603 perform the same processing as in the attention mechanism layer 601, and use the output result of the attention mechanism layer 603 as the input of the additional attention layer 604.

[0171] The additional attention layer 604 performs probability mapping on the initial attention matrix output by the attention mechanism layer 603, performs sequential multiplication processing on the initial attention matrix, the transposed matrix of the query parameter matrix, the transposed matrix of the key parameter matrix, and the transposed matrix of the initial attention matrix, performs normalization processing on the multiplication result based on the dimension of the positional encoding, and processes the normalization result through an activation function to obtain the attention score matrix of the group to be recognized, which can be implemented by formula (6), and is specifically described below.

[0172]

[0173] Among them, the activation function softmax is maximum likelihood processing, which finds a set of parameter values to maximize the probability of the observed data. Y is the feature matrix output by the multi-layer attention mechanism layer, with dimensions N*d model , N is the number of nodes, d model is the dimension of the positional encoding, W Q is the value vector (Query) weight matrix, is the transpose of the key vector (Key) weight matrix, scales the dot product result to prevent the dot product value from being too large and causing the gradient to disappear.

[0174] In step 405, calculate the difference between the attention score matrix and the transition probability matrix. When the difference is less than the difference threshold, it is determined that the group to be recognized belongs to the risk type.

[0175] As an example, calculate the sum of squares of the differences between the attention score matrix and the transition probability matrix, which can be implemented by formula (7), and is specifically described below.

[0176]

[0177] Among them, ∈ represents the difference between the attention score matrix and the transition probability matrix. Y is the attention score matrix, M is the transition probability matrix, and F is the operation of summing the squared differences of matrix elements, which aggregates the element-wise error squares of the matrix into a scalar value for measuring the overall difference. When the difference ∈ is less than the type threshold, which is a pre-set high-risk threshold, that is, the characteristics of the group to be identified are more similar to those of the risk type and satisfy a certain threshold comparison relationship, it is determined that the current group to be identified belongs to the risk type, and corresponding handling operations are performed on the group to be identified according to the operations corresponding to the risk type. Multiple type thresholds can be set, corresponding to multiple levels, which is a numerical value. For example, when the difference is less than the type threshold, it is determined that the group to be identified belongs to the risk type, and if the risk type is high risk, the group to be identified is blocked.

[0178] As an example, the risk type corresponds to the type of samples for training the neural network model. Historical group samples that conform to the risk type are selected to train the neural network model. Each historical group is regarded as a sample, the transition values between the objects in each historical group are calculated, and the transition probability matrix M of the historical group is determined. Through the neural network model, the historical group is predicted, the attention score matrix Y corresponding to the historical group is obtained, the difference between the attention score matrix Y and the transition probability matrix M of the historical group is calculated and fitted to obtain the loss function L, and the neural network model is trained based on the loss function L. The loss function L can be implemented by formula (8), which is specifically described below.

[0179]

[0180] Among them, S is the number of all samples, M i is the transition probability matrix of historical group i, Y i is the attention score matrix of historical group i, F is the operation of summing the squared differences of matrix elements, which aggregates the element-wise error squares of the matrix into a scalar value for measuring the overall difference between the network prediction and the true transition probability. The neural network model is trained by minimizing the loss function L, and the trained neural network model is used for offline deployment to perform prediction processing on the group to be identified.

[0181] The group identification method provided by the embodiments of this application has the following beneficial effects:

[0182] By constructing a position encoding matrix based on the temporal features of the group member joining order, the joining order is transformed into a quantifiable position vector. Then, the trading correlation strength between members is quantified through a transition probability matrix. Combining the fusion of the basic feature matrix, a multi-dimensional feature representation is constructed. The fused features are predicted through the multi-layer attention mechanism neural network of the neural network model. Through the cascaded design of the multi-layer attention mechanism layer and the additional attention layer, local temporal feature extraction, global relationship modeling, and dynamic weight allocation are sequentially completed to generate an attention score matrix. Finally, by calculating the difference between the attention score matrix and the transition probability matrix, the dynamic determination of the risk level is realized, and corresponding operations are performed on the group according to the risk. While improving the recognition accuracy, the false positive rate and the risk of data leakage are reduced, and the efficiency of risk prevention and control is improved.

[0183] Next, the implementation of the group recognition device 455 provided in the embodiments of the present application as an exemplary structure of software modules will be further described. In some embodiments, as Figure 2 shown, the software modules stored in the group recognition device 455 in the memory 450 may include: a transition probability determination module 4551, configured to perform probability mapping processing on the asset transfer data between the objects in the group to obtain the actual asset transfer probability between the objects in the group; an encoding module 4552, configured to obtain the joining order of each object in the group, and perform order encoding processing on the joining orders of multiple objects in the group to obtain the joining order encoding of the group; perform object encoding processing on the object data of multiple objects in the group to obtain the object encoding of the group; a template probability prediction module 4553, configured to perform transfer probability prediction processing of the corresponding risk type based on the joining order encoding of the group and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type between the objects in the group; a type determination module 4554, configured to perform prediction processing of the corresponding risk type based on the template asset transfer probability and the actual asset transfer probability to determine the risk type prediction result of the group.

[0184] In some embodiments, the transition probability determination module 4551 is further configured to obtain the object nodes corresponding to each object; perform transformation processing on the asset transfer data between the objects to obtain the asset transfer data between the object nodes; perform probability prediction processing on the asset transfer data between the object nodes to obtain the actual asset transfer probability of the group.

[0185] In some embodiments, the transfer probability determination module 4551 is further configured to perform the following processing for each object node: perform a summation process on the asset transfer data flowing out from the object node to all other object nodes to obtain the total asset transfer of the object node; when the total asset transfer is less than one, determine the asset transfer data between the object node and each other object node as the asset transfer probability between the object node and each other object node; when the total asset transfer is not less than one, determine the ratio between the asset transfer data between the object node and each other object node and the total asset transfer, and determine the ratio as the asset transfer probability between the object node and each other object node; form the asset transfer probabilities between the object node and at least one other object node into the asset transfer probability of the corresponding object node; form the asset transfer probabilities corresponding to multiple object nodes into the actual asset transfer probability of the group.

[0186] In some embodiments, the group addition order encoding of the group includes the position encodings of multiple objects. The position encoding includes multiple dimensions. The encoding module 4552 is further configured to perform the following processing for each object: obtain the object node corresponding to the object, and use the addition order of the object as the addition order of the object node; perform an order encoding process on the addition order of the object node to obtain the encoding values of the object node in each dimension; form the encoding values of the object node in multiple dimensions into the addition order encoding of the object node; form the addition order encodings of the object nodes corresponding to multiple objects into the addition order encoding of the group.

[0187] In some embodiments, the encoding module 4552 is further configured to perform the following processing for each dimension: when the number of the dimension is even, perform a sine encoding process on the addition order and the encoding of the object node to obtain the encoding value of the object node in the dimension; when the number of the dimension is odd, perform a cosine encoding process on the addition order and the number of the object node to obtain the encoding value of the object node in the dimension.

[0188] In some embodiments, the template probability prediction module 4553 is further configured to perform a fusion process on the addition order encoding and the object encoding to obtain a first fusion encoding; perform a transfer probability prediction process for the corresponding risk type on the first fusion encoding to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

[0189] In some embodiments, the template probability prediction module 4553 is further configured to perform a mapping process based on a multi-layer attention mechanism on the first fusion encoding to obtain a first attention matrix; perform a probability mapping process on the first attention matrix to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

[0190] In some embodiments, the template probability prediction module 4553 is further configured to perform a mapping process on the input of the n-th layer attention network through the n-th layer attention network in the N-layer cascaded attention network to obtain a mapping result output by the n-th layer attention network; use the mapping result output by the n-th layer attention network as the input of the (n + 1)-th layer attention network to continue performing the mapping process to obtain a mapping result output by the (n + 1)-th layer attention network; use the mapping result output by the N-th layer attention network as the first attention matrix; where N is an integer greater than 1, n is an integer variable that starts from 1 and increases incrementally, and the value range of n is 1 ≤ n < N. When n takes the value of 1, the input of the n-th layer attention network is the first fusion encoding. When 2 ≤ n < N, the input of the n-th layer attention network is the mapping result output by the (n - 1)-th layer attention network.

[0191] In some embodiments, the template probability prediction module 4553 is further configured to perform an attention mechanism-based mapping process on the input of the n-th layer attention network to obtain a first attention result; perform a residual mechanism-based normalization process on the first attention result to obtain a first normalization result; perform a feed-forward process on the first normalization result to obtain a first feed-forward result; perform a residual mechanism-based normalization process on the first feed-forward result to obtain the mapping result of the n-th layer attention network.

[0192] In some embodiments, the template probability prediction module 4553 is further configured to perform sequential multiplication processing on the first attention matrix, the query parameter matrix, the transposed matrix of the key parameter matrix, and the transposed matrix of the first attention matrix to obtain a first matrix multiplication result; perform a normalization process on the first matrix multiplication result based on the dimension of the additive group order encoding to obtain a normalized matrix; perform a maximum likelihood process on the normalized matrix to obtain the template asset transfer probability that conforms to the risk type among the objects within the group.

[0193] In some embodiments, the template asset transfer probability is implemented by invoking a group identification model. Before performing the transfer probability prediction process for the corresponding risk type based on the group-based group addition order encoding and the object encoding of the group to obtain the template asset transfer probability of the group conforming to the risk type, the model training module 4555 is further configured to perform a probability prediction process on the asset transfer data between the sample objects in the sample group to obtain the historical actual asset transfer probability between the sample objects in the sample group, where the sample group is a historical group conforming to the risk type; obtain the second group addition order of each sample object in the sample group, and perform order encoding processing on the second group addition orders of multiple sample objects in the sample group to obtain the group addition order encoding of the sample group; perform object encoding processing on the object data of multiple sample objects in the sample group to obtain the object encoding of the sample group; perform the following processing through the group identification model: perform the transfer probability prediction process for the corresponding risk type based on the group addition order encoding and the object encoding of the sample group to obtain the historical template asset transfer probability conforming to the risk type between the sample objects in the sample group; perform fitting processing based on the difference between the historical actual asset transfer probability and the historical template asset transfer probability to obtain a first loss, and update the group identification model based on the first loss.

[0194] 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. The processor of the 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 group identification method described above in the embodiments of the present application.

[0195] An embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions or a computer program are stored. When the computer-executable instructions or the computer program are executed by a processor, the processor will be caused to execute the group identification method provided by the embodiments of the present application. For example, Figure 3A the group identification method shown.

[0196] 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; it may also be various devices including one or any combination of the above memories.

[0197] 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.

[0198] As an example, the computer-executable instructions may or may not correspond to a file in a file system, and may be stored as part of a file that holds 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 (e.g., files that store one or more modules, subroutines, or portions of code).

[0199] As an example, the computer-executable instructions may be deployed to execute on one electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0200] In summary, through the determination of the actual asset transfer probability and the template asset transfer probability of the group in the embodiments of the present application, considering the joining order and data characteristics of the objects in the group, the template asset transfer probability is determined through a neural network model that learns the risk type characteristics, and whether the group belongs to the risk type is determined through the difference between the actual asset transfer probability and the template asset transfer probability, improving the accuracy of group type recognition.

[0201] The above is only the embodiments 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 group identification method, characterized in that: The method includes: Performing probability mapping processing on the asset transfer data between objects in the group to obtain the actual asset transfer probability between the objects in the group; Obtaining the joining order of each object in the group, and performing sequential encoding processing on the joining orders of multiple objects in the group to obtain the joining order encoding of the group; Performing object encoding processing on the object data of multiple objects in the group to obtain the object encoding of the group; Performing transfer probability prediction processing of the corresponding risk type based on the joining order encoding of the group and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type between the objects in the group; Performing prediction processing of the corresponding risk type based on the template asset transfer probability and the actual asset transfer probability to determine the risk type prediction result of the group.

2. The method according to claim 1, characterized in that, The performing probability mapping processing on the asset transfer data between objects in the group to obtain the actual asset transfer probability between the objects in the group includes: Obtaining the object node corresponding to each object; Performing transformation processing on the asset transfer data between the objects to obtain the asset transfer data between the object nodes; Performing probability prediction processing on the asset transfer data between the object nodes to obtain the actual asset transfer probability of the group.

3. The method according to claim 2, wherein The performing probability prediction processing on the asset transfer data between the object nodes to obtain the actual asset transfer probability of the group includes: Performing the following processing for each object node: Performing summation processing on the asset transfer data flowing out from the object node to all other object nodes to obtain the total asset transfer of the object node; When the total asset transfer is less than one, determining the asset transfer data between the object node and each other object node as the asset transfer probability between the object node and each other object node; When the total asset transfer is not less than one, determining the ratio between the asset transfer data between the object node and each other object node and the total asset transfer, and determining the ratio as the asset transfer probability between the object node and each other object node; Composing the asset transfer probabilities between the object node and at least one other object node into the asset transfer probability corresponding to the object node; Composing the asset transfer probabilities corresponding to multiple object nodes into the actual asset transfer probability of the group.

4. The method according to claim 1, wherein The joining order encoding of the group includes the position encodings of multiple objects, and the position encoding includes multiple dimensions. The performing sequential encoding processing on the joining orders of multiple objects in the group to obtain the joining order encoding of the group includes: Performing the following processing for each object: Obtaining the object node corresponding to the object, and taking the joining order of the object as the joining order of the object node; Performing sequential encoding processing on the joining order of the object node to obtain the encoding value of the object node in each dimension; Composing the encoding values of the object node in multiple dimensions into the joining order encoding of the object node; Encode the group-joining order of the object nodes corresponding to multiple said objects to form the group-joining order encoding of the group.

5. The method according to claim 4, wherein Performing sequential encoding processing on the group-joining order of the object nodes to obtain the encoding values of the object nodes in each said dimension, including: Performing the following processing for each said dimension: When the number of the dimension is even, performing sine encoding processing on the group-joining order of the object nodes and the encoding to obtain the encoding value of the object nodes in the dimension; When the number of the dimension is odd, performing cosine encoding processing on the group-joining order of the object nodes and the number to obtain the encoding value of the object nodes in the dimension.

6. The method according to claim 1, wherein Performing transfer probability prediction processing of the corresponding risk type based on the group-joining order encoding of the group and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type among the objects in the group, including: Performing fusion processing on the group-joining order encoding and the object encoding to obtain a first fusion encoding; Performing transfer probability prediction processing of the corresponding risk type on the first fusion encoding to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

7. The method according to claim 6, characterized in that Performing transfer probability prediction processing of the corresponding risk type on the first fusion encoding to obtain the template asset transfer probability that conforms to the risk type among the objects in the group, including: Performing mapping processing based on a multi-layer attention mechanism on the first fusion encoding to obtain a first attention matrix; Performing probability mapping processing on the first attention matrix to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

8. The method according to claim 7, wherein Performing mapping processing based on a multi-layer attention mechanism on the first fusion encoding to obtain a first attention matrix, including: Performing mapping processing on the input of the n-th attention network in the N-level cascaded attention network to obtain the mapping result output by the n-th attention network; Using the mapping result output by the n-th attention network as the input of the (n + 1)-th attention network to continue performing mapping processing to obtain the mapping result output by the (n + 1)-th attention network; Using the mapping result output by the N-th attention network as the first attention matrix; Where N is an integer greater than 1, n is an integer variable starting from 1 and increasing, and the value range of n is 1 ≤ n < N. When n = 1, the input of the n-th attention network is the first fusion encoding. When 2 ≤ n < N, the input of the n-th attention network is the mapping result output by the (n - 1)-th attention network.

9. The method according to claim 8, characterized in that, Performing mapping processing on the input of the n-th attention network in the N-level cascaded attention network to obtain the mapping result output by the n-th attention network, including: Performing mapping processing based on the attention mechanism on the input of the n-th attention network to obtain a first attention result; Performing normalization processing based on the residual mechanism on the first attention result to obtain a first normalization result; Perform feed-forward processing on the first normalization result to obtain a first feed-forward result; Perform normalization processing based on the residual mechanism on the first feed-forward result to obtain the mapping result of the nth layer attention network.

10. The method according to claim 7, characterized in that The performing probability mapping processing on the first attention matrix to obtain the template asset transfer probability that conforms to the risk type among the objects in the group includes: Perform sequential multiplication processing on the first attention matrix, the transposed matrix of the query parameter matrix, the transposed matrix of the key parameter matrix, and the transposed matrix of the first attention matrix to obtain a first matrix multiplication result; Perform normalization processing on the first matrix multiplication result based on the dimension of the group addition order encoding to obtain a normalized matrix; Perform maximum likelihood processing on the normalized matrix to obtain the template asset transfer probability that conforms to the risk type among the objects in the group.

11. The method according to claim 1, wherein The template asset transfer probability is realized by calling a group recognition model. Before performing the transfer probability prediction processing of the corresponding risk type based on the group addition order encoding of the group and the object encoding of the group to obtain the template asset transfer probability that conforms to the risk type among the objects in the group, the method further includes: Perform probability prediction processing on the asset transfer data between sample objects in a sample group to obtain the historical actual asset transfer probability between the sample objects in the sample group, where the sample group is a historical group that conforms to the risk type; Obtain the second addition order of each sample object in the sample group, and perform sequential encoding processing on the second addition orders of multiple sample objects in the sample group to obtain the addition order encoding of the sample group; Perform object encoding processing on the object data of multiple sample objects in the sample group to obtain the object encoding of the sample group; Execute the following processing through the group recognition model: perform transfer probability prediction processing of the corresponding risk type based on the addition order encoding of the sample group and the object encoding to obtain the historical template asset transfer probability that conforms to the risk type among the sample objects in the sample group; Perform fitting processing based on the difference between the historical actual asset transfer probability and the historical template asset transfer probability to obtain a first loss, and update the group recognition model based on the first loss.

12. A group identification device, characterized in that, The device includes: A transfer probability determination module, configured to perform probability mapping processing on the asset transfer data between objects in a group to obtain the actual asset transfer probability between the objects in the group; An encoding module, configured to obtain the addition order of each object in the group, and perform sequential encoding processing on the addition orders of multiple objects in the group to obtain the addition order encoding of the group; perform object encoding processing on the object data of multiple objects in the group to obtain the object encoding of the group; A template probability prediction module, configured to perform a transfer probability prediction process for a corresponding risk type based on the group addition order encoding of the group and the object encoding of the group, so as to obtain a template asset transfer probability that conforms to the risk type among the objects in the group; A type determination module, configured to perform a prediction process for the corresponding risk type based on the template asset transfer probability and the actual asset transfer probability, and determine a risk type prediction result of the group.

13. An electronic device, characterized in that, The electronic device includes: A memory, configured to store computer-executable instructions or a computer program; A processor, configured to implement the group identification method according to any one of claims 1 to 11 when executing the computer-executable instructions or the computer program stored in the memory.

14. A computer-readable storage medium stores 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 group identification method according to any one of claims 1 to 11.

15. 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 group identification method according to any one of claims 1 to 11.

Images