Object group determination method and apparatus, device, storage medium, and program product

Abstract

The application discloses a kind of object group determination method, device, equipment, storage medium and program product, involve artificial intelligence and data processing technical field.The method comprises: obtaining the aggregation zone structure graph corresponding to reference object group, and the multiple aggregation zone structure graphs corresponding to the network to be detected;Obtain the distributed representation corresponding to reference object group, and the multiple distributed representations corresponding to the network to be detected, the distributed representation is used to represent the structural features of aggregation zone structure graph;According to the distributed representation corresponding to reference object group, determine the target aggregation zone structure graph from the multiple aggregation zone structure graphs corresponding to the network to be detected;Obtain the target object group corresponding to target aggregation zone structure graph.The application determines the object group by supporting the aggregation zone level, without determining object one by one, and considering the overall correlation of object group, so as to improve the determination efficiency and determination accuracy of object group.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and in particular to a method, apparatus, device, storage medium, and program product for determining a group of objects. Background Technology

[0002] Illegal group activities (such as fraud, gambling, pyramid schemes, and loan fraud) are key targets for risk prevention and control. With the development of artificial intelligence technology, deep learning models can be applied to identify illegal target groups. Here, "illegal target groups" refers to the groups of people involved in illegal group activities.

[0003] In related technologies, a graph embedding model based on GNN (Graph Neural Network) is used to obtain the vector representations of each object in the illegal object group and the vector representations of each object in the network to be detected (such as a payment network). Then, the vector representations of each object in the illegal object group and the vector representations of each object in the network to be detected are matched to identify suspicious objects from the network to be detected.

[0004] However, related technologies require matching objects one by one, which involves a large workload and low matching efficiency. Summary of the Invention

[0005] This application provides a method, apparatus, device, storage medium, and program product for determining a group of objects, which can improve the efficiency and accuracy of determining the group of objects. The technical solution is as follows:

[0006] According to one aspect of the embodiments of this application, a method for determining a group of objects is provided, the method comprising:

[0007] Obtain the aggregation region structure diagram corresponding to the reference object group, and obtain multiple aggregation region structure diagrams corresponding to the network to be detected; wherein, the aggregation region structure diagram is used to characterize the correlation between objects belonging to the same aggregation region;

[0008] Based on the aggregation region structure diagram corresponding to the reference object group, a distributed representation corresponding to the reference object group is obtained, and based on the multiple aggregation region structure diagrams corresponding to the network to be detected, multiple distributed representations corresponding to the network to be detected are obtained; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram;

[0009] The distributed representation corresponding to the reference object group is matched with multiple distributed representations corresponding to the network to be detected, and the target aggregation region structure diagram is determined from the multiple aggregation region structure diagrams corresponding to the network to be detected.

[0010] Obtain the target object group corresponding to the target aggregation region structure diagram.

[0011] According to one aspect of the embodiments of this application, an apparatus for determining a group of objects is provided, the apparatus comprising:

[0012] The structure graph acquisition module is used to acquire the aggregation region structure graph corresponding to the reference object group, and to acquire multiple aggregation region structure graphs corresponding to the network to be detected; wherein, the aggregation region structure graph is used to characterize the correlation between objects belonging to the same aggregation region;

[0013] The distributed representation acquisition module is used to acquire the distributed representation corresponding to the reference object group based on the aggregation region structure diagram corresponding to the reference object group, and to acquire multiple distributed representations corresponding to the network to be detected based on multiple aggregation region structure diagrams corresponding to the network to be detected; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram;

[0014] The structure graph matching module is used to match the distributed representation corresponding to the reference object group with the multiple distributed representations corresponding to the network to be detected, and to determine the target aggregation region structure graph from the multiple aggregation region structure graphs corresponding to the network to be detected.

[0015] The target group acquisition module is used to acquire the target object group corresponding to the target aggregation region structure diagram.

[0016] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the above-described method for determining the object group.

[0017] The computer device is a terminal or a server.

[0018] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored in the storage medium, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described method for determining the object group.

[0019] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the aforementioned method for determining the object group.

[0020] The technical solutions provided in this application embodiment may have the following beneficial effects:

[0021] By directly matching the distributed representations corresponding to the reference object group with multiple distributed representations corresponding to the network under test, object group determination at the aggregation region level is achieved without the need for individual object matching, reducing the matching workload and thus improving the efficiency of object group determination. Simultaneously, since the distributed representation of the aggregation region structure graph encompasses the structural features of the aggregation region structure graph, it considers the overall correlation of the object group, rather than being limited to individual objects, thereby improving the accuracy of object group determination. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the implementation environment of a solution provided in one embodiment of this application;

[0024] Figure 2 This is a flowchart of a method for determining a group of objects according to an embodiment of this application;

[0025] Figure 3 This is a flowchart of a method for obtaining an aggregation region structure diagram according to an embodiment of this application;

[0026] Figure 4 This is a schematic diagram of the distribution of the aggregation region structure provided in one embodiment of this application;

[0027] Figure 5 This is a flowchart of a method for obtaining a distributed representation provided in one embodiment of this application;

[0028] Figure 6 This is a schematic diagram of a graph embedding model provided in one embodiment of this application;

[0029] Figure 7 This is a schematic diagram of a method for obtaining a rooted subgraph according to an embodiment of this application;

[0030] Figure 8 This is a flowchart of a method for partitioning a relational structure diagram according to an embodiment of this application;

[0031] Figure 9 This is a flowchart of a training method for an association scoring model provided in one embodiment of this application;

[0032] Figure 10 This is a block diagram of an apparatus for determining a group of objects according to an embodiment of this application;

[0033] Figure 11 This is a block diagram of an apparatus for determining a group of objects according to another embodiment of this application;

[0034] Figure 12 This is a block diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0036] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0037] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0038] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0039] The technical solution provided in this application relates to natural language processing technology in artificial intelligence. It uses natural language processing technology to obtain relational structure graphs, aggregation region structure graphs, and distributed representations of aggregation region structure graphs, and uses the obtained relational structure graphs, aggregation region structure graphs, and distributed representations of aggregation region structure graphs to determine object groups.

[0040] The method provided in this application can be executed by a computer device, which refers to an electronic device with data computing, processing, and storage capabilities. This computer device can be a terminal such as a PC (Personal Computer), tablet computer, smartphone, desktop computer, wearable device, or intelligent robot; or it can be a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0041] The technical solutions provided in this application are applicable to any scenario requiring the identification of a target group, such as target group identification scenarios, risk control systems, network security systems, payment networks, and social networks. The technical solutions provided in this application can improve the efficiency and accuracy of target group identification.

[0042] In one example, such as Figure 1 As shown, taking an object group identification system as an example, the system may include a terminal 10 and a server 20.

[0043] Terminal 10 can be an electronic device such as a mobile phone, desktop computer, tablet computer, PC, or wearable device. A client application for the target application can be installed on terminal 10. The target application can be a target group identification application, a risk control application, a payment application, a social application, an entertainment application, etc., and this embodiment does not limit the scope of the application.

[0044] Server 20 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services. Server 20 is used to provide background services for the client of the target application in terminal 10. For example, server 20 can be the background server of the aforementioned target application (such as an object group identification application).

[0045] Terminal 10 and server 20 can communicate via network 30.

[0046] For example, the user selects a reference object group and a network to be detected through the client in terminal 10, and sends the reference object group and the network to be detected to server 20 through the client in terminal 10. After obtaining the reference object group and the network to be detected, server 20 determines the target object group from the object group corresponding to the network to be detected based on the reference object group. Server 20 sends the target object group to the client so as to display the target object group to the user through the client.

[0047] Of course, in some other examples, the process of determining the object group can also be performed in terminal 10, such as in the client of the target application described above. For example, after obtaining the reference object group and the network to be detected, the client determines the target object group from the object group corresponding to the network to be detected based on the reference object group and displays the target object group to the user.

[0048] Please refer to Figure 2 It illustrates a flowchart of a method for determining a group of objects according to an embodiment of this application, wherein the entity executing each step of the method may be... Figure 1 The terminal 10 or server 20 in the implementation environment of the scheme shown may include the following steps (201-204).

[0049] Step 201: Obtain the aggregation region structure diagram corresponding to the reference object group, and obtain multiple aggregation region structure diagrams corresponding to the network to be detected; wherein, the aggregation region structure diagram is used to characterize the correlation between objects belonging to the same aggregation region.

[0050] An object group refers to a group composed of multiple objects. There are relationships between the objects in an object group, such as cooperative relationships, interactive relationships, and hierarchical relationships. Objects can include users, consumers, buyers, payers, etc. In some embodiments, an object group may also be referred to as a gang. All objects in an object group belong to the same aggregation region. There are strong relationships between objects within an aggregation region. For example, if objects in an object group participate in a certain event (such as fraud, gambling, loan fraud, etc.), then the object group can be identified as corresponding to an aggregation region under that event. This application embodiment does not limit the division of aggregation regions. Optionally, for the aggregation region structure diagram, the aggregation region in this application embodiment may also refer to the set of object nodes corresponding to the aggregation region structure diagram.

[0051] In this embodiment of the application, the reference object group refers to the object group used as a comparison standard. This embodiment of the application does not limit the type of the reference object group. For example, the type of the reference object group can be an illegal object group (i.e., a criminal gang), such as a fraud target group, a gambling target group, a pyramid scheme target group, a loan fraud target group, etc. The type of the reference object group can also be a professional object group, such as a student group, a worker group, etc.

[0052] A network to be tested refers to a network that is to be tested for its target population. Such networks can be payment, risk control, or any network that requires target population testing.

[0053] A structural graph is a graph model in graph theory consisting of nodes and the edges connecting them. In this embodiment, the aggregation region structural graph is a graph model where each object in the aggregation region is a node, and the relationships between objects are edges. The nodes in the aggregation region structural graph can be referred to as object nodes.

[0054] Since the embodiments of this application aim to identify target object groups similar to reference object groups from the network to be detected, only one aggregation region structure diagram needs to be generated for one reference object group, and only multiple aggregation region structure diagrams corresponding to each reference object group need to be generated for multiple reference object groups. However, the object group division in the network to be detected is relatively complex, so it is necessary to first determine multiple object groups corresponding to the network to be detected, and then obtain multiple aggregation region structure diagrams corresponding to the network to be detected.

[0055] In one example, for the process of obtaining the structure diagrams of multiple aggregation regions corresponding to the network to be detected, step 201 may also include the following sub-steps:

[0056] Step 201a: Based on the object information of each object in the network to be detected, construct the object node set corresponding to the network to be detected.

[0057] Here, object information refers to the identification information that can be used to uniquely identify an object. The object node set includes multiple object nodes, each object node corresponds to an object, and the object node corresponding to an object can be represented by the object information corresponding to that object.

[0058] In one example, the process of obtaining object information of each object in the network to be detected can be as follows: determine the target object set in the network to be detected within the target time period, obtain the object information of each target object in the target object set, and construct the object node set of the network to be detected under the target time period based on the object information of each target object in the target object set.

[0059] The target time period is the same as the target time period mentioned below. This target time period can be adaptively set and adjusted by the designer according to the actual usage. The target time period can be in units of hours, days, etc., such as 24 hours, 2 days, etc., and this application embodiment does not limit this. The target object set refers to the collection composed of target objects. The target object refers to the object whose state changes in the network to be detected within the target time period, such as state changes caused by interactive operations, payment operations, setting operations, etc.

[0060] Step 201b: Based on the association information between various objects in the network to be detected, construct the relationship edges between various object nodes in the object node set, and generate the relationship structure graph corresponding to the network to be detected.

[0061] The aforementioned association information can be used to reflect the association between objects, which can refer to social relationships, payment relationships, intimacy, etc. This association information may include resource transfer information, such as the quantity of resource transfers (e.g., transfer amount, payment amount), the number of resource transfers (e.g., number of transfers, payment times), the quantity of resource gifts (e.g., red envelope amount), and the number of resource gifts (e.g., number of red envelopes). It may also include social relationship information, such as the number of shared friends, the number of shared social groups, and the duration of the friendship. Furthermore, it may include object usage information, such as document usage information and device usage information. This application embodiment does not limit this aspect.

[0062] Optionally, the relationship edges between the aforementioned object nodes can be used to characterize the association between objects, that is, to determine the relationship edges between object nodes based on the association between objects. For example, the weight value of the relationship edge between object nodes can be determined based on the association between objects, or the existence of a relationship edge between object nodes can be determined based on the association between objects.

[0063] In one example, the process of obtaining relational edges and relational structure graphs can be as follows: Obtain the association information between various objects in the network to be detected within the target time period; based on the association information, determine the association score between various objects in the network to be detected, where the association score characterizes the degree of association between objects; for the third and fourth object nodes in the object node set, determine the association score between the third and fourth object nodes as the weight value of the relational edge between the third and fourth object nodes; based on the weight values ​​of the relational edges between various object nodes in the object node set, generate the relational structure graph corresponding to the network to be detected.

[0064] The aforementioned association information can refer to the association information between various target objects in the target object set. The third and fourth object nodes can refer to two nodes in any pair of object nodes in the object node set. Weight values ​​can be used to characterize the distance between object nodes, and they are positively correlated with the degree of association between objects. A relationship structure graph is a structural diagram used to describe the relationships between various objects in a network. For example, the relationship structure graph corresponding to the network to be detected is used to describe the relationships between all objects in the network to be detected, while the aggregation region structure graph corresponding to the network to be detected is used to describe the relationships between some objects in the network to be detected.

[0065] A correlation score is a rating of the degree of association between objects. Optionally, the correlation score can be obtained through a correlation scoring model. The association information between objects is input into the correlation scoring model, which then calculates and processes the association information to output the correlation score. Optionally, this correlation scoring model can be constructed based on an LR (Logistic Regression) model. The specific training process of the correlation scoring model will be described in detail below and will not be repeated here.

[0066] Step 201c: Divide the relational structure graph into aggregate regions to obtain multiple aggregate region structure graphs corresponding to the network to be detected. At least two of the multiple aggregate region structure graphs corresponding to the network to be detected have overlapping regions.

[0067] In this context, dividing the relational structure graph into aggregation regions means dividing the network to be detected into multiple aggregation regions, each corresponding to a type of object group. Then, based on the aggregation regions, the object nodes in the relational structure graph are further divided to obtain multiple aggregation region structure graphs corresponding to the network to be detected.

[0068] An object node may belong to multiple aggregation region structure graphs simultaneously; such an object node can be called an overlapping object node. If two aggregation region structure graphs have overlapping areas, it indicates that there are overlapping object nodes between them; these two aggregation region structure graphs can also be called overlapping aggregation region structure graphs. For example, see reference... Figure 4 The network to be detected has aggregate region structure diagrams 401, 402, 403 and 404. There are overlapping areas between each pair of aggregate region structure diagrams 401, 402, 403 and 404.

[0069] To address this, this application employs BigCLAM (Cluster Affiliation Model for BigNetworks) to partition the relationship structure graph into aggregation regions. Since BigCLAM is an overlapping aggregation region discovery method, it can effectively capture overlapping aggregation region structure graphs, reflecting the overlapping relationships between multiple aggregation region structure graphs corresponding to the network under test. Compared to obtaining individual aggregation region structure graphs, partitioning the relationship structure graph into aggregation regions based on BigCLAM yields a more realistic and accurate aggregation region structure graph. This more realistic and accurate aggregation region structure graph allows for the determination of object groups, effectively improving the accuracy of object group identification. The method for partitioning the relationship structure graph into aggregation regions based on BigCLAM will be described in detail below and will not be repeated here.

[0070] Optionally, the process of obtaining the aggregation region structure graph corresponding to the reference object group is similar to the process of obtaining the relational structure graph corresponding to the network to be detected. The specific process can be as follows: obtain the object information of each object in the reference object group, and obtain the association information between each object in the reference object group; construct the object node set corresponding to the reference object group based on the object information of each object in the reference object group; construct the relationship edges between each object node in the object node set corresponding to the reference object group based on the association information between each object in the reference object group, and generate the aggregation region structure graph corresponding to the reference object group.

[0071] For example, refer to Figure 1 After receiving the reference object group and the network to be detected, server 20 uses the correlation scoring model 40 to determine the set of relational edges corresponding to the reference object group based on the correlation information between each object in the reference object group, and to determine the set of relational edges corresponding to the network to be detected based on the correlation information between each object in the network to be detected, thereby constructing the aggregation region structure graph corresponding to the reference object group and the relational structure graph corresponding to the network to be detected. Server 20 then uses BigCLAM (corresponding to...) Figure 150) Divide the relational structure graph corresponding to the network to be detected into aggregate regions to obtain multiple aggregate region structure graphs corresponding to the network to be detected.

[0072] Step 202: Based on the aggregation region structure diagram corresponding to the reference object group, obtain the distributed representation corresponding to the reference object group; and based on the multiple aggregation region structure diagrams corresponding to the network to be detected, obtain multiple distributed representations corresponding to the network to be detected; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram.

[0073] Distributed representation refers to a vector representation that can characterize the relationships between data. In the embodiments of this application, the distributed representation of the aggregation region structure graph can characterize the relationships between the rooted subgraphs corresponding to the aggregation region structure graph, as well as the distribution of the rooted subgraphs corresponding to the aggregation region structure graph. Since a rooted subgraph is an ordered structure and can better capture the non-linear structural features in the structure graph, the distributed representation determined based on the rooted subgraphs can more accurately characterize the structural features of the aggregation region structure graph.

[0074] In one example, distributed representations can be obtained using Graph2vec (a method for generating distributed graph representations). For example, see [reference]. Figure 1 The distributed representation of the reference object group can be obtained based on the aggregation region structure graph corresponding to the reference object group using a graph embedding model (i.e., Graph2vec) 60, and multiple distributed representations of the network to be detected can be obtained based on multiple aggregation region structure graphs corresponding to the network to be detected. Optionally, step 202 above may also include the following sub-steps:

[0075] Step 202a: Combine the aggregate region structure diagram corresponding to the reference object group and the multiple aggregate region structure diagrams corresponding to the network to be detected to obtain an aggregate region structure diagram set.

[0076] For example, the aggregate region structure diagram corresponding to the reference object group can be labeled as G. a ={G a1 G a2 , ..., G an}, where G an This represents the aggregate region structure diagram corresponding to the nth reference object group, where n is a positive integer. Multiple aggregate region structure diagrams corresponding to the network to be detected can be labeled as G. b ={G b1 G b2 , ..., G bm}, where G bm This represents the structure graph of the m-th aggregation region corresponding to the network to be detected, where m is a positive integer. For G... a and G bBy combining these elements, we obtain the aggregation region structure set G = {G1, G2, ..., G...} n+m}

[0077] Step 202b: Construct a graph embedding model, which includes a first weight matrix and a second weight matrix. The number of rows in the first weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set, and the number of columns in the second weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set.

[0078] The graph embedding model is used to obtain a distributed representation of the aggregation region structure graph. Optionally, the graph embedding model can be constructed based on Skip-Gram (a skip-word model). The number of columns in the first weight matrix is ​​the same as the number of rows in the second weight matrix, and the number of columns in the first weight matrix and the number of rows in the second weight matrix can be set based on actual usage requirements (or empirical values). When constructing the graph embedding model, the elements in the first weight matrix and the second weight matrix are randomly initialized. Each row of elements in the initialized first weight matrix can initializely represent a distributed representation of an aggregation region structure graph.

[0079] For example, refer to Figure 6 The graph embedding model 600 includes an input layer 601, a hidden layer 602, and an output layer 603. Data, such as a one-hot encoding of dimension 1*V, can be input into the input layer 601, where V is the number of aggregation region structure graphs in the aggregation region structure graph set. The hidden layer 602 corresponds to the first weight matrix W. V*N The dimension is V*N (N can be an empirical value). The output layer 603 corresponds to the second weight matrix W′. N*V The dimension is N*V.

[0080] Step 202c: Obtain the input pair set of the graph embedding model; wherein, the input pair set includes multiple input pairs corresponding to each aggregation region structure graph, and for the target aggregation region structure graph in the aggregation region structure graph set, the input pair corresponding to the target aggregation region structure graph includes a set of one-hot codes of the target aggregation region structure graph and a one-hot code of a rooted subgraph corresponding to the target aggregation region structure graph.

[0081] The input set is used to train the graph embedding model to optimize the elements in the first and second weight matrices. The target aggregation region structure graph can refer to any aggregation region structure graph in the set of aggregation region structure graphs.

[0082] In one example, the process of obtaining the input pair set can be as follows:

[0083] 1. Obtain the rooted subgraph set corresponding to each object node in each aggregation region structure graph; wherein, for the fifth object node in the target aggregation region structure graph, the rooted subgraph set of the fifth object node includes the rooted subgraph of the fifth object node in the target structure graph at each depth corresponding to the fifth object node, and the rooted subgraph of the fifth object node's adjacent points in the target structure graph at each depth corresponding to the fifth object node's adjacent points. The fifth object node can refer to any object node in the target aggregation region structure graph.

[0084] For example, refer to Figure 7 For object node n in the aggregation region structure graph 700, with a depth of 2, the rooted subgraph of object node n at depth 0 includes rooted subgraph 701, the rooted subgraphs of object node n at depth 1 include rooted subgraphs 702 and 703, and the rooted subgraphs of object node n at depth 2 include rooted subgraphs 704 and 705. The adjacent nodes of object node n include object node n1 and object node n2. Object node n1 has a depth of 2, and the rooted subgraph of object node n1 at depth 0 includes rooted subgraph 706, the rooted subgraph of object node n1 at depth 1 includes rooted subgraph 707, and the rooted subgraph of object node n1 at depth 2 includes rooted subgraph 708. The depth of object node n2 is 2. The rooted subgraphs of object node n2 at depth 0 include rooted subgraph 709, the rooted subgraphs of object node n2 at depth 1 include rooted subgraph 710, and the rooted subgraphs of object node n2 at depth 2 include rooted subgraph 711. Therefore, the set of rooted subgraphs corresponding to object node n in the aggregation region structure graph 700 includes rooted subgraphs 701 to 711.

[0085] 2. Obtain the one-hot encoding of each aggregate region structure graph in the aggregate region structure graph set, and obtain the one-hot encoding of each rooted subgraph corresponding to each object node in each aggregate region structure graph.

[0086] Optionally, the dimension of one-hot encoding can be determined based on the number of aggregate region structure graphs in the aggregate region structure graph set. For example, the dimension of one-hot encoding can be the same as the number of aggregate region structure graphs in the aggregate region structure graph set.

[0087] 3. For the target aggregation region structure graph, combine the one-hot encoding of the target aggregation region structure graph and the one-hot encoding of any rooted subgraph corresponding to the target aggregation region structure graph into an input pair corresponding to the target aggregation region structure graph. Use the same method to obtain multiple input pairs corresponding to the target aggregation region structure graph.

[0088] 4. Obtain multiple input pairs corresponding to each aggregation region structure graph in the aggregation region structure graph set, and generate an input pair set.

[0089] Alternatively, the input layer of the graph embedding model can be used to obtain the one-hot encoding of the aggregate structure graph and the one-hot encoding of the rooted subgraph in the input pair based on the labels of the aggregate structure graph and the rooted subgraph in the input pair.

[0090] Step 202d: Using the graph embedding model, obtain the transition output vector set corresponding to the input pair set. The transition output vectors in the transition output vector set are used to transitionally represent the distributed representation of the rooted subgraph in the input pair.

[0091] Optionally, the specific process for obtaining the transition output vector can be as follows: For the target input pair in the input pair set, the one-hot encoding of the aggregation region structure graph in the target input pair is multiplied by the first weight matrix to obtain the first transition matrix. The first transition matrix is ​​used to transitionally represent the distributed representation of the aggregation region structure graph in the target input pair. The first transition matrix is ​​multiplied by the second weight matrix to obtain the transition output vector corresponding to the target input pair. Here, the target input pair can refer to any input pair in the input pair set.

[0092] For example, refer to Figure 6 The target input pair is embedded into the input graph of the input layer 601 into the model 600. The one-hot encoding x (dimension 1*V) of the aggregation region structure graph in the target input pair is combined with the first weight matrix W. V*N Multiply them to obtain the first transition matrix h (dimension 1*N), then combine the first transition matrix h with the second weight matrix W′. N*V Multiplying them yields the transition output vector (dimension 1*V) corresponding to the target input pair, as shown in the figure y. 1j To y Cj , where y Cj This is the transition output vector corresponding to the C-th input pair of the j-th aggregation region structure diagram.

[0093] Step 202e: Based on the one-hot encoding of the rooted subgraphs in the transition output vector set and the input pair set, obtain the objective function of the graph embedding model. The objective function is used to characterize the degree of association between the rooted subgraphs in the input pair and the aggregate region structure graph in the input pair.

[0094] Alternatively, the specific process of obtaining the objective function can be as follows:

[0095] 1. For a target input pair in the input pair set, a second sub-objective function is obtained based on the transition output vector corresponding to the target input pair and the one-hot encoding of the rooted subgraph in the target input pair. The second sub-objective function is used to characterize the degree of correlation between the transition output vector corresponding to the target input pair and the rooted subgraph in the target input pair.

[0096] For example, the second sub-objective function can be expressed as follows: Where y is the one-hot encoding of the rooted subgraph in the target input pair. The target input is the corresponding transition output vector.

[0097] In some other examples, y represents the true label of the rooted subgraph in the target input pair belonging to the aggregate region structure graph in the target input pair (e.g., 1). Given the likelihood probability corresponding to the target input (i.e., the transition output vector after logistic regression, ranging from 0 to 1), the second objective function described below is the third objective function described below. In other words, the third objective function can be directly determined as the objective function of the graph embedding model. By directly optimizing the third objective function, the optimized first weight matrix can be obtained.

[0098] 2. Summate the second sub-objective function sets corresponding to each input pair in the input pair set to obtain the second objective function of the graph embedding model.

[0099] For example, the second sub-objective function can be expressed as follows: Where N is the number of input pairs in the input pair set, y i For the one-hot encoding of the rooted subgraph in the i-th input pair, Let be the transition output vector corresponding to the i-th input pair.

[0100] 3. Perform logistic regression on the transition output vector corresponding to the target input pair to obtain the likelihood probability of the target input pair. The likelihood probability is used to characterize the probability that there is a rooted subgraph in the target input pair that belongs to the aggregate region structure graph in the target input pair.

[0101] For example, the likelihood probability can be expressed as follows:

[0102]

[0103] Wherein, Φ(G) is the distributed representation of the aggregation region structure graph in the target input pair (i.e., the first transition matrix mentioned above), Φ(sg) is the distributed representation of the rooted subgraph in the target input pair (i.e., the transition output vector), and Φ(C) is the set of distributed representations of all rooted subgraphs.

[0104] 4. For the target aggregation region structure graph, the third sub-objective function corresponding to the target aggregation region structure graph is obtained by multiplying the likelihood probabilities of the multiple input pairs corresponding to the target aggregation region structure graph.

[0105] For example, the third sub-objective function can be expressed as follows:

[0106] -logP(Φ(sg1),Φ(s g 2), ..., Φ(sg n )|Φ(G));

[0107] Where Φ(sgn) is the distributed representation of the nth rooted subgraph corresponding to the target aggregation region structure graph.

[0108] 5. Based on the sum of the third sub-objective functions corresponding to each aggregation region structure graph, the third objective function of the graph embedding model is obtained.

[0109] For example, the third objective function can be expressed as follows:

[0110]

[0111] Where V is the total number of aggregation region structure diagrams, Φ(G v ) is the distributed representation of the structure graph of the v-th aggregation region, Φ(sg) n ) is G v The distributed representation of the nth rooted subgraph.

[0112] 6. Based on the second and third objective functions, obtain the objective function of the graph embedding model.

[0113] Alternatively, the sum of the second and third objective functions can be used as the objective function of the graph embedding model, or the second and third objective functions can be weighted and summed to obtain the objective function of the graph embedding model.

[0114] Step 202f involves iteratively optimizing the objective function of the graph embedding model to obtain the optimized first weight matrix.

[0115] Optionally, the gradient descent method can be used to iteratively optimize the objective function to obtain the optimized first weight matrix and the optimized second weight matrix.

[0116] Step 202g: Based on each row element of the optimized first weight matrix, obtain the distributed representation corresponding to each aggregation region structure diagram.

[0117] Optionally, each row element of the optimized first weight matrix can be determined as a distributed representation of an aggregation region structure graph, thereby obtaining the distributed representations corresponding to each aggregation region structure graph.

[0118] In this embodiment, based on the multiple aggregation region structure graphs corresponding to the network to be detected obtained using BigCLAM, Graph2vec is used to obtain the distributed representation of the aggregation region structure graphs. Since there are a large number of overlapping rooted subgraphs among the multiple aggregation region structure graphs corresponding to the network to be detected, there are many common features among the distributed representations of different aggregation region structure graphs. This allows for accurate characterization of the correlation between the distributed representations of different aggregation region structure graphs, thereby improving the accuracy of determining the object group. Simultaneously, the presence of a large number of overlapping rooted subgraphs effectively improves the optimization efficiency of the first weight matrix, thus increasing the efficiency of obtaining the distributed representation.

[0119] Step 203: Match the distributed representation corresponding to the reference object group with the multiple distributed representations corresponding to the network to be detected, and determine the target aggregation region structure diagram from the multiple aggregation region structure diagrams corresponding to the network to be detected.

[0120] Among them, the target aggregation region structure map refers to the aggregation region structure map that has similar structural features to the aggregation region structure map corresponding to the reference object group among the multiple aggregation region structure maps corresponding to the network to be detected.

[0121] In one example, the process of obtaining the target aggregation region structure map can be as follows: obtain the distributed representation corresponding to the reference object group and the similarity between it and the multiple distributed representations corresponding to the network to be detected; determine the aggregation region structure map whose similarity among the multiple aggregation region structure maps corresponding to the network to be detected satisfies the first condition as the target aggregation region structure map.

[0122] In this embodiment, the first condition mentioned above may refer to a predetermined number of similarities belonging to a similarity sequence arranged in descending order. This predetermined number can be set according to actual usage requirements. For example, similarities are sorted in descending order of magnitude to obtain a similarity sequence; the predetermined number of similarities in the similarity sequence is determined as the target similarity; and the aggregation region structure diagram corresponding to the target similarity is determined as the target aggregation region structure diagram. Alternatively, multiple aggregation region structure diagrams corresponding to the network to be detected are sorted in descending order of magnitude of similarity to obtain an aggregation region structure diagram sequence; and the predetermined number of aggregation region structure diagrams in the aggregation region structure diagram sequence is determined as the target aggregation region structure diagram.

[0123] Optionally, the first condition mentioned above can also refer to a similarity greater than a set threshold. For example, the target similarity values ​​that are greater than the set threshold are obtained, and the aggregation region structure diagram corresponding to the target similarity value is determined as the target aggregation region structure diagram.

[0124] Alternatively, similarity can be calculated using algorithms including but not limited to: cosine similarity, Euclidean distance, Manhattan distance, Chebyshev distance, etc.

[0125] In another example, the process of obtaining the target aggregation region structure map can also be as follows: obtain the black label concentration corresponding to each of the multiple aggregation region structure maps corresponding to the network to be detected. The black label concentration is used to characterize the proportion of objects with black labels in the aggregation region structure map; remove the aggregation region structure maps with black label concentration less than the fourth threshold to obtain the remaining aggregation region structure maps corresponding to the network to be detected; match the distributed representation corresponding to the reference object group with the distributed representation corresponding to the remaining structure map to determine the target aggregation region structure map from the remaining structure map.

[0126] The fourth threshold can be set based on empirical values. The black label concentration can be determined by the ratio of the number of black-labeled object nodes to the total number of object nodes in the aggregation region structure diagram. Black labels are used to characterize illegal objects (such as those involved in illegal activities) or black market objects with misconduct (such as fraud, loan fraud, etc.).

[0127] This application embodiment performs preliminary screening of multiple aggregation region structure diagrams corresponding to the network to be detected, thereby reducing the amount of similarity calculation and the workload of matching object groups, and further improving the efficiency of determining object groups.

[0128] Step 204: Obtain the target object group corresponding to the target aggregation region structure diagram.

[0129] Optionally, the objects corresponding to each object node in the target aggregation region structure diagram are combined to obtain a target object group. This target object group refers to a group of objects that have similar structural features to the reference object group.

[0130] In summary, the technical solution provided in this application, by directly matching the distributed representation corresponding to the reference object group with multiple distributed representations corresponding to the network to be detected, achieves the determination of the object group at the aggregation region level, without needing to match each object individually, thus reducing the matching workload and improving the efficiency of object group determination. Furthermore, since the distributed representation of the aggregation region structure graph encompasses the structural features of the aggregation region structure graph, it considers the overall association of the object group, rather than being limited to a single object, thereby improving the accuracy of object group determination.

[0131] Furthermore, by using Graph2vec to obtain the distributed representation of the multiple aggregation region structure graphs corresponding to the network under test obtained based on BigCLAM, and considering the numerous overlapping rooted subgraphs among these graphs, a large number of common features exist between the distributed representations of different aggregation region structure graphs. This allows for a more accurate and realistic characterization of the correlations between the distributed representations of different aggregation region structure graphs, thereby improving the accuracy of object group determination. Simultaneously, the presence of numerous overlapping rooted subgraphs effectively improves the optimization efficiency of the first weight matrix, thus enhancing the efficiency of distributed representation acquisition and further improving the efficiency of object group determination.

[0132] Please refer to Figure 8 It illustrates a flowchart of a method for partitioning a relational structure graph according to an embodiment of this application. The execution entity of each step of the method can be... Figure 1 The terminal 10 or server 20 in the implementation environment of the scheme shown may include the following steps (801-804).

[0133] Step 801: Based on the object information of each object in the network to be detected, construct the object node set corresponding to the network to be detected.

[0134] Step 802: Based on the association information between various objects in the network to be detected, construct the relationship edges between various object nodes in the object node set, and generate the relationship structure graph corresponding to the network to be detected.

[0135] Steps 801 and 802 are the same as those described in the above embodiments. For any content not described in steps 801 and 802, please refer to the above embodiments. They will not be repeated here.

[0136] Step 803: Based on each object node in the relational structure graph, construct a fitting structure graph corresponding to the relational structure graph; wherein, the fitting structure graph includes an initial number of initial aggregation regions, and the set of relational edges corresponding to the fitting structure graph is determined based on the initial membership strength between each object node in the relational structure graph and each initial aggregation region, and the initial membership strength is used to initialize the probability that an object node belongs to an initial aggregation region.

[0137] In this embodiment, fitting a structure graph refers to fitting a relational structure graph to obtain a structure graph. This fitted structure graph uses object nodes in the relational structure graph as nodes and fits the relational edges in the relational structure graph using the initial membership strength between each object node and each initial aggregation region. The initial membership strength can be a large value, which gradually converges as the fitted structure graph is iteratively optimized. Initial aggregation regions are used to initialize the aggregation regions of the fitted relational structure graph. The initial membership strength is a non-negative random value. For a target object node and a target initial aggregation region, the greater the initial membership strength between the target object node and the target initial aggregation region, the greater the probability that the target object node belongs to the target initial aggregation region.

[0138] In one example, a fitted structure graph can be constructed based on the AGM (Community Affiliation Graph Model). The AGM model can be used to fit the relationship edges in the relationship structure graph, which can specifically include the following:

[0139] For a target initial aggregation region c, the probability that there is a relational edge between object node a and object node b is:

[0140]

[0141] Among them, F ac F represents the initial membership strength between object node a and the target initial aggregation region c. bc The initial membership strength between object node b and the target initial aggregation region c.

[0142] Let the initial number be C, then the probability that there is a relational edge between object node a and object node b (hereinafter referred to as the edge generation probability) is:

[0143]

[0144] Where, for i∈C, 1-P i (a, b) represents the probability that the initial aggregation region i cannot create a relational edge between object node b and the target initial aggregation region c. Let C represent the probability that none of the C initial aggregation regions can create a relational edge between object node b and the target initial aggregation region c. This represents the probability that there is a relation edge between object node a and object node b.

[0145] Using the same method as above, the edge generation probability between each object node in the fitted structure graph is obtained. Then, based on the edge generation probability between each object node in the fitted structure graph, the set of relational edges corresponding to the fitted structure graph is obtained. Finally, based on the set of object nodes and the set of relational edges corresponding to the fitted structure graph, the fitted structure graph is constructed.

[0146] Step 804: Based on the initial membership strength, obtain the first objective function corresponding to the fitted structure graph. The first objective function is used to characterize the likelihood between the set of relation edges corresponding to the fitted structure graph and the set of relation edges corresponding to the relation structure graph.

[0147] Optionally, the specific process of obtaining the first objective function can be as follows:

[0148] 1. For the first object node and the second object node in the fitted structure graph, based on the initial membership strength between the first object node and each initial aggregation region, and the initial membership strength between the second object node and each initial aggregation region, a first probability function is obtained between the first object node and the second object node. The first probability function is used to characterize the possibility that there is a relation edge between the first object node and the second object node.

[0149] Here, the first object node can refer to any object node in the fitted structure graph, and the second object node can refer to any object node in the fitted structure graph other than the first object node. This first probability function is the edge generation probability mentioned above.

[0150] 2. Integrate the first probability functions between each object node in the fitted structure graph to obtain the first sub-objective function.

[0151] For example, based on the above embodiments, the first sub-objective function can be expressed as follows:

[0152]

[0153] Where (u, v) refers to any pair of object nodes in the fitted structure graph, E is the set of relation edges corresponding to the fitted structure graph, and C is the initial number corresponding to the initial aggregation region.

[0154] 3. Based on the first probability function between each object node in the fitted structure graph, obtain the second probability function between each object node in the fitted structure graph. The second probability function is used to characterize the possibility that there is no relation edge between the first object node and the second object node.

[0155] For example, based on the above embodiments, the second probability function can be expressed as follows: 1-P uv .

[0156] 4. Integrate the second probability functions between each object node in the fitted structure graph to obtain the second sub-objective function;

[0157] For example, the second sub-objective function can be expressed as follows:

[0158]

[0159] 5. Integrate the first sub-objective function and the second sub-objective function to obtain the first objective function.

[0160] For example, the first objective function can be expressed as follows:

[0161]

[0162] Step 805: Iteratively optimize the first objective function to obtain the optimized membership strength between each object node and each initial aggregation region in the fitted structure diagram.

[0163] Optionally, gradient descent can be used to iteratively optimize the first objective function to obtain the optimized membership strength between each object node and each initial aggregation region. During the iterative optimization process, the first objective function can be simplified first, since log(e...) x If x = 1, then we can take the logarithm of the first objective function, and thus:

[0164]

[0165] Optionally, an L1 regularization term can be added to the first objective function, allowing the fitted structure graph to learn the number of true aggregation regions. For object node u, the initial membership strength between object node u and each initial aggregation region is adjusted towards maximizing the first objective function. This process can be represented as follows:

[0166]

[0167] Where L(F) is the first objective function, λ is the weight parameter, and F ui The initial membership strength between object node u and initial aggregation region i.

[0168] Step 806: Based on the optimized membership strength between each object node and each initial aggregation region in the fitted structure graph, the object nodes in the relationship structure graph are divided into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected.

[0169] Optionally, for each initial aggregation region, object nodes with an optimized membership strength greater than a first threshold are assigned to the target initial aggregation region; the number of object nodes corresponding to each initial aggregation region is obtained; initial aggregation regions with zero object nodes are removed to obtain multiple real aggregation regions corresponding to the network to be detected; and a structure diagram of multiple aggregation regions corresponding to the network to be detected is obtained based on the multiple real aggregation regions corresponding to the network to be detected.

[0170] Here, the target initial aggregation region can refer to any of the various initial aggregation regions. The first threshold can be set based on empirical values. Optionally, for an object node, it can be divided into multiple initial aggregation regions; that is, as long as the optimized membership strength between the object node and the initial aggregation region is greater than the first threshold, the object node can be determined to belong to that initial aggregation region. After obtaining multiple real aggregation regions, an aggregation region structure graph corresponding to the real aggregation region can be delineated from the relational structure graph based on the object nodes corresponding to the real aggregation regions. That is, the aggregation region structure graph is composed of the object nodes corresponding to the real aggregation regions and the relational edges of the object nodes corresponding to the real aggregation regions in the relational structure graph.

[0171] In summary, the technical solution provided in this application, by directly matching the distributed representation corresponding to the reference object group with multiple distributed representations corresponding to the network to be detected, achieves the determination of the object group at the aggregation region level, without needing to match each object individually, thus reducing the matching workload and improving the efficiency of object group determination. Furthermore, since the distributed representation of the aggregation region structure graph encompasses the structural features of the aggregation region structure graph, it considers the overall association of the object group, rather than being limited to a single object, thereby improving the accuracy of object group determination.

[0172] Furthermore, by using BigCLAM to divide the relationship structure graph into aggregation regions, BigCLAM can effectively capture the overlapping areas between aggregation region structure graphs, reflecting the correlation between multiple aggregation region structure graphs. Compared to obtaining each independent aggregation region structure graph, dividing the relationship structure graph into aggregation regions based on BigCLAM can obtain a more realistic and accurate aggregation region structure graph. Therefore, based on a more realistic and accurate aggregation region structure graph, the determination of object groups can be effectively improved, thus improving the accuracy of object group determination.

[0173] Please refer to Figure 9 The diagram illustrates a flowchart of a training method for a correlation scoring model provided in one embodiment of this application. The execution entity for each step of this method can be... Figure 1 The terminal 10 or server 20 in the implementation environment of the scheme shown may include the following steps (901-903).

[0174] Step 901: Obtain the training sample set, which includes the association information corresponding to the first type of object pairs, the association information corresponding to the second type of object pairs, and the association information corresponding to the third type of object pairs. The first type of object pair refers to the object pair in the reference object group whose association score is greater than the second threshold. The second type of object pair refers to the object pair in the network to be detected whose association score is greater than the second threshold. The third type of object pair refers to the object pair in the reference object group whose association score is less than the third threshold, and the third threshold is less than the second threshold.

[0175] Relevance scoring refers to the rating of the degree of relevance between objects. Relevance information can be used to reflect the relevance between objects, which can refer to social relationships, payment relationships, intimacy, etc. The aforementioned relevance information may include resource transfer information, such as the quantity of resource transfers (e.g., transfer amount, payment amount), the number of resource transfers (e.g., number of transfers, number of payments), the quantity of resource gifts (e.g., red envelope amount), the number of resource gifts (e.g., number of red envelopes), etc. It may also include social relationship information, such as the number of shared friends, the number of shared social groups, the duration of friendship relationships, etc. It may also include object usage information, such as document usage information, device usage information, etc., which are not limited in this embodiment of the application.

[0176] For example, taking a payment application as an example, the above-mentioned associated information may include at least: transfer amount, number of transfers, amount of resource donation, number of resource donations, amount of QR code payment, number of QR code payment, whether the same device is used (yes, marked as 1), whether the same ID is used (yes, marked as 1), number of the same friends, number of the same social groups, and duration of the friendship relationship (hereinafter simply marked as x1-x9).

[0177] The second and third thresholds can be set based on empirical values. The first type of object pair can refer to object pairs with high correlation (e.g., closeness) in the reference object group, the second type of object pair can refer to object pairs with high correlation (e.g., closeness) in the network to be detected, and the third type of object pair can refer to object pairs with low correlation (e.g., closeness) in the reference object group.

[0178] This application does not limit the proportion of the first type of object pairs, the second type of object pairs, and the third type of object pairs in its embodiments. For example, assuming the training sample set includes the association information of n pairs of object pairs, the association information corresponding to n / 4 pairs of the first type of object pairs, the association information corresponding to n / 4 pairs of the second type of object pairs, and the association information corresponding to n / 2 pairs of the third type of object pairs can be obtained.

[0179] Step 902: Obtain the correlation prediction score set corresponding to the training sample set through the correlation scoring model.

[0180] The correlation scoring model is used to score the correlation between pairs of objects. The network structure of the correlation scoring model can be a neural network, a deep neural network, a convolutional neural network, a logistic regression network, etc., and this application embodiment does not limit it.

[0181] The correlation prediction score set includes multiple correlation prediction scores. Each correlation prediction score corresponds to an object pair in the training sample set. The correlation prediction score is used to characterize the predicted correlation between the object pairs.

[0182] Optionally, the association information can be standardized before inputting it into the association scoring model. For example, the association information x1-x9 can be standardized to the range of 0-1 to obtain standardized association information x1-x9, and then the standardized association information x1-x9 can be output to the association scoring model to train the association scoring model.

[0183] Step 903: Based on the degree of difference between the correlation prediction score set and the correlation label score set corresponding to the training sample set, train the correlation score model to obtain the trained correlation score model.

[0184] The correlation labeling score set includes multiple correlation labeling scores, which are used to label the true correlation between object pairs and can be used as training labels for the model.

[0185] Alternatively, the process of obtaining the training loss of the correlation scoring model can be as follows:

[0186] For each target object pair in the training sample set, the sub-training loss for that target object pair is obtained based on its corresponding association label score and its corresponding association prediction score. The training loss of the association scoring model is then obtained based on the sum of the sub-training losses for each object pair in the training sample set.

[0187]

[0188] Where m is the number of object pairs in the training sample set, D is the training sample set, (u, v) are object pairs in D, y is the association label score corresponding to (u, v), and y′ is the association prediction score corresponding to (u, v).

[0189] The relevance scoring model is trained by minimizing its training loss, resulting in a fully trained model. Let the relevance information be x1-x9 as described above; then the relevance scoring model can be expressed as:

[0190] Among them, w i For x iThe corresponding weight parameters.

[0191] In summary, the technical solution provided in this application, by directly matching the distributed representation corresponding to the reference object group with multiple distributed representations corresponding to the network to be detected, achieves the determination of the object group at the aggregation region level, without needing to match each object individually, thus reducing the matching workload and improving the efficiency of object group determination. Furthermore, since the distributed representation of the aggregation region structure graph encompasses the structural features of the aggregation region structure graph, it considers the overall association of the object group, rather than being limited to a single object, thereby improving the accuracy of object group determination.

[0192] Furthermore, by comprehensively considering various social and payment relationships between objects, relational edges are constructed between object nodes. Compared to considering only social or payment relationships, the relational edges in this embodiment carry more information, thus more accurately reflecting the degree of association between objects. This improves the accuracy of constructing the relational structure graph and, consequently, the accuracy of identifying object groups.

[0193] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0194] refer to Figure 10 This diagram illustrates a block diagram of an apparatus for determining a group of objects according to an embodiment of this application. The apparatus has the functionality to implement the method example described above; this functionality can be implemented in hardware or by hardware executing corresponding software. The apparatus can be the computer device described above, or it can be located within a computer device. Figure 10 As shown, the device 1000 includes: a structure diagram acquisition module 1001, a distributed representation acquisition module 1002, a structure diagram matching module 1003, and a target group acquisition module 1004.

[0195] The structure graph acquisition module 1001 is used to acquire the aggregation region structure graph corresponding to the reference object group, and to acquire multiple aggregation region structure graphs corresponding to the network to be detected; wherein, the aggregation region structure graph is used to characterize the correlation between objects belonging to the same aggregation region.

[0196] The distributed representation acquisition module 1002 is used to acquire the distributed representation corresponding to the reference object group based on the aggregation region structure diagram corresponding to the reference object group, and to acquire multiple distributed representations corresponding to the network to be detected based on multiple aggregation region structure diagrams corresponding to the network to be detected; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram.

[0197] The structure graph matching module 1003 is used to match the distributed representation corresponding to the reference object group with the multiple distributed representations corresponding to the network to be detected, and to determine the target aggregation region structure graph from the multiple aggregation region structure graphs corresponding to the network to be detected.

[0198] The target group acquisition module 1004 is used to acquire the target object group corresponding to the target aggregation area structure diagram.

[0199] In one exemplary embodiment, such as Figure 11 As shown, the structure graph acquisition module 1001 includes: an object node construction submodule 1001a, a relationship edge construction submodule 1001b, and a structure graph acquisition submodule 1001c.

[0200] The object node construction submodule 1001a is used to construct the object node set corresponding to the network under test based on the object information of each object in the network under test.

[0201] The relation edge construction submodule 1001b is used to construct the relation edges between each object node in the object node set based on the association information between each object in the network to be detected, and generate the relation structure graph corresponding to the network to be detected.

[0202] The structure graph acquisition submodule 1001c is used to divide the relationship structure graph into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected. At least two of the multiple aggregation region structure graphs corresponding to the network to be detected have overlapping regions.

[0203] In one exemplary embodiment, the structural diagram acquisition submodule 1001c is used for:

[0204] Based on each object node in the relational structure graph, a fitting structure graph corresponding to the relational structure graph is constructed; wherein, the fitting structure graph includes an initial number of initial aggregation regions, and the set of relational edges corresponding to the fitting structure graph is determined based on the initial membership strength between each object node in the relational structure graph and each initial aggregation region, and the initial membership strength is used to initialize the probability that the object node belongs to the initial aggregation region;

[0205] Based on the initial membership strength, a first objective function corresponding to the fitted structure graph is obtained. The first objective function is used to characterize the likelihood between the set of relation edges corresponding to the fitted structure graph and the set of relation edges corresponding to the relation structure graph.

[0206] The first objective function is iteratively optimized to obtain the optimized membership strength between each object node and each initial aggregation region in the fitted structure graph.

[0207] Based on the optimized membership strength between each object node in the fitted structure graph and each initial aggregation region, the object nodes in the relationship structure graph are divided into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected.

[0208] In one exemplary embodiment, the structural diagram acquisition submodule 1001c is further configured to:

[0209] For the first object node and the second object node in the fitted structure graph, based on the initial membership strength between the first object node and each of the initial aggregation regions, and the initial membership strength between the second object node and each of the initial aggregation regions, a first probability function is obtained between the first object node and the second object node. The first probability function is used to characterize the possibility that there is a relation edge between the first object node and the second object node.

[0210] The first sub-objective function is obtained by multiplying the first probability functions between each object node in the fitted structure graph.

[0211] Based on the first probability function between each object node in the fitted structure graph, a second probability function between each object node in the fitted structure graph is obtained. The second probability function is used to characterize the possibility that there is no relation edge between the first object node and the second object node.

[0212] The second sub-objective function is obtained by multiplying the second probability functions between each object node in the fitted structure graph.

[0213] The first objective function is obtained by multiplying the first sub-objective function and the second sub-objective function.

[0214] In one exemplary embodiment, the structural diagram acquisition submodule 1001c is further configured to:

[0215] For each of the target initial aggregation regions, object nodes whose optimized membership strength with the target initial aggregation region is greater than a first threshold are assigned to the target initial aggregation region.

[0216] Obtain the number of object nodes corresponding to each of the initial aggregation regions;

[0217] The initial aggregation regions with zero object nodes are removed to obtain multiple real aggregation regions corresponding to the network to be detected.

[0218] Based on the multiple real aggregation regions corresponding to the network under test, a structural diagram of the multiple aggregation regions corresponding to the network under test is obtained.

[0219] In an exemplary embodiment, the relation edge construction submodule 1001b is used for:

[0220] Obtain the association information between various objects in the network to be detected within the target time period;

[0221] Based on the association information, an association score is determined between each object in the network to be detected, and the association score is used to characterize the degree of association between the objects;

[0222] For the third object node and the fourth object node in the object node set, the correlation score between the third object node and the fourth object node is determined as the weight value of the relationship edge between the third object node and the fourth object node;

[0223] Based on the weight values ​​of the relationship edges between each object node in the object node set, a relationship structure graph corresponding to the network to be detected is generated.

[0224] In an exemplary embodiment, the relation edge construction submodule 1001b is further configured to:

[0225] Obtain a training sample set, which includes association information corresponding to a first type of object pair, association information corresponding to a second type of object pair, and association information corresponding to a third type of object pair; wherein, the first type of object pair refers to the object pair in the reference object group whose association score is greater than a second threshold, the second type of object pair refers to the object pair in the network to be detected whose association score is greater than the second threshold, and the third type of object pair refers to the object pair in the reference object group whose association score is less than a third threshold, wherein the third threshold is less than the second threshold;

[0226] The correlation score model is used to obtain the correlation prediction score set corresponding to the training sample set.

[0227] Based on the degree of difference between the correlation prediction score set and the correlation label score set corresponding to the training sample set, the correlation scoring model is trained to obtain the trained correlation scoring model.

[0228] In one exemplary embodiment, the distributed representation acquisition module 1002 is configured to:

[0229] The aggregation region structure diagram corresponding to the reference object group and the multiple aggregation region structure diagrams corresponding to the network to be detected are combined to obtain an aggregation region structure diagram set;

[0230] A graph embedding model is constructed, which includes a first weight matrix and a second weight matrix. The number of rows in the first weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set, and the number of columns in the second weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set.

[0231] Obtain the input pair set of the graph embedding model; wherein, the input pair set includes multiple input pairs corresponding to each aggregation region structure graph, and for a target aggregation region structure graph in the aggregation region structure graph set, the input pair corresponding to the target aggregation region structure graph includes a set of one-hot codes of the target aggregation region structure graph and a one-hot code of a rooted subgraph corresponding to the target aggregation region structure graph;

[0232] The graph embedding model is used to obtain the transition output vector set corresponding to the input pair set. The transition output vectors in the transition output vector set are used to transitionally represent the distributed representation of the rooted subgraph in the input pair.

[0233] Based on the one-hot encoding of the rooted subgraphs in the transition output vector set and the input pair set, the objective function of the graph embedding model is obtained. The objective function is used to characterize the degree of association between the rooted subgraphs in the input pair and the aggregation region structure graph in the input pair.

[0234] The objective function of the graph embedding model is iteratively optimized to obtain the optimized first weight matrix;

[0235] Based on each row element of the optimized first weight matrix, the distributed representations corresponding to each aggregation region structure diagram are obtained.

[0236] In one exemplary embodiment, the distributed representation acquisition module 1002 is further configured to:

[0237] For a target input pair in the input pair set, a second sub-objective function is obtained based on the transition output vector corresponding to the target input pair and the one-hot encoding of the rooted subgraph in the target input pair. The second sub-objective function is used to characterize the degree of correlation between the transition output vector corresponding to the target input pair and the rooted subgraph in the target input pair.

[0238] The second objective function of the graph embedding model is obtained by summing the second sub-objective function sets corresponding to each input pair in the input pair set.

[0239] Logistic regression is performed on the transition output vector corresponding to the target input pair to obtain the likelihood probability of the target input pair. The likelihood probability is used to characterize the probability that there is a rooted subgraph in the target input pair that belongs to the aggregation region structure graph in the target input pair.

[0240] For the target aggregation region structure graph, the third sub-objective function corresponding to the target aggregation region structure graph is obtained by multiplying the likelihood probabilities of the multiple input pairs corresponding to the target aggregation region structure graph.

[0241] The third objective function of the graph embedding model is obtained based on the sum of the third sub-objective functions corresponding to each aggregation region structure graph.

[0242] Based on the second objective function and the third objective function, the objective function of the graph embedding model is obtained.

[0243] In one exemplary embodiment, the distributed representation acquisition module 1002 is further configured to:

[0244] For a target input pair in the input pair set, the one-hot encoding of the aggregation region structure graph in the target input pair is multiplied by the first weight matrix to obtain a first transition matrix. The first transition matrix is ​​used to transitionally characterize the distributed representation of the aggregation region structure graph in the target input pair.

[0245] Multiply the first transition matrix and the second weight matrix to obtain the transition output vector corresponding to the target input pair.

[0246] In one exemplary embodiment, the structure diagram matching module 1003 is configured to:

[0247] The similarity between the distributed representations corresponding to the reference object group and the multiple distributed representations corresponding to the network to be detected is obtained respectively.

[0248] The cluster structure diagram whose similarity satisfies the first condition among the multiple cluster structure diagrams corresponding to the network to be detected is determined as the target cluster structure diagram.

[0249] In an exemplary embodiment, the structure diagram matching module 1003 is further configured to:

[0250] Obtain the black label concentration corresponding to each of the multiple aggregation region structure diagrams of the network to be detected. The black label concentration is used to characterize the proportion of objects with black labels in the aggregation region structure diagram.

[0251] The aggregation region structure diagrams with black label concentrations less than the fourth threshold are removed to obtain the remaining aggregation region structure diagrams corresponding to the network to be detected.

[0252] The distributed representation corresponding to the reference object group is matched with the distributed representation corresponding to the remaining structure graph, and the target aggregation region structure graph is determined from the remaining structure graph.

[0253] In summary, the technical solution provided in this application, by directly matching the distributed representation corresponding to the reference object group with multiple distributed representations corresponding to the network to be detected, achieves the determination of the object group at the aggregation region level, without needing to match each object individually, thus reducing the matching workload and improving the efficiency of object group determination. Furthermore, since the distributed representation of the aggregation region structure graph encompasses the structural features of the aggregation region structure graph, it considers the overall association of the object group, rather than being limited to a single object, thereby improving the accuracy of object group determination.

[0254] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0255] Please refer to Figure 12 This diagram illustrates a structural block diagram of a computer device according to an embodiment of this application. This computer device can be used to implement the method for determining a group of objects provided in the above embodiments. Specifically, it may include the following:

[0256] The computer device 1200 includes a central processing unit (such as a CPU, GPU, or FPGA) 1201, a system memory 1204 including RAM (Random-Access Memory) 1202 and ROM (Read-Only Memory) 1203, and a system bus 1205 connecting the system memory 1204 and the central processing unit 1201. The computer device 1200 also includes a basic input / output system (I / O system) 1206 to facilitate information transfer between various devices within the server, and a mass storage device 1207 for storing the operating system 1213, application programs 1214, and other program modules 1215.

[0257] The basic input / output system 1206 includes a display 1208 for displaying information and an input device 1209 for user input, such as a mouse or keyboard. Both the display 1208 and the input device 1209 are connected to the central processing unit 1201 via an input / output controller 1210 connected to the system bus 1205. The basic input / output system 1206 may also include the input / output controller 1210 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1210 also provides output to a display screen, printer, or other types of output devices.

[0258] The mass storage device 1207 is connected to the central processing unit 1201 via a mass storage controller (not shown) connected to the system bus 1205. The mass storage device 1207 and its associated computer-readable media provide non-volatile storage for the computer device 1200. That is, the mass storage device 1207 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0259] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 1204 and mass storage device 1207 described above can be collectively referred to as memory.

[0260] According to an embodiment of this application, the computer device 1200 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1200 can be connected to the network 1212 via the network interface unit 1211 connected to the system bus 1205, or the network interface unit 1211 can be used to connect to other types of networks or remote computer systems (not shown).

[0261] The memory further includes at least one instruction, at least one program, code set, or instruction set, wherein at least one instruction, at least one program, code set, or instruction set is stored in the memory and configured to be executed by one or more processors to implement the method for determining the above-mentioned group of objects.

[0262] In one exemplary embodiment, a computer-readable storage medium is also provided, the storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set, when executed by a processor, implements the method for determining the above-described object group.

[0263] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0264] In one exemplary embodiment, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the method for determining the object group described above.

[0265] It should be noted that in the specific embodiments of this application, object information and other related data are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the object is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the acquisition of object operation information, object information, and relationship information between objects involved in this application are all carried out with the permission of the object and in compliance with relevant laws, regulations, and standards.

[0266] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0267] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for determining a group of objects, characterized in that, The method includes: Obtain the cluster structure diagram corresponding to the reference object group, and obtain multiple cluster structure diagrams corresponding to the network to be detected; wherein, the cluster structure diagram is used to characterize the correlation between objects belonging to the same cluster; Based on the aggregation region structure diagram corresponding to the reference object group, a distributed representation corresponding to the reference object group is obtained; and based on the multiple aggregation region structure diagrams corresponding to the network to be detected, multiple distributed representations corresponding to the network to be detected are obtained; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram. The similarity between the distributed representations corresponding to the reference object group and the multiple distributed representations corresponding to the network to be detected is obtained respectively. The cluster structure diagram whose similarity satisfies the first condition among the multiple cluster structure diagrams corresponding to the network to be detected is determined as the target cluster structure diagram; Obtain the target object group corresponding to the target aggregation region structure diagram.

2. The method according to claim 1, characterized in that, The step of obtaining the structure diagram of multiple aggregation regions corresponding to the network to be detected includes: Based on the object information of each object in the network to be detected, construct the object node set corresponding to the network to be detected; Based on the association information between various objects in the network to be detected, the relationship edges between various object nodes in the object node set are constructed to generate the relationship structure graph corresponding to the network to be detected. The relational structure graph is divided into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected. At least two of the multiple aggregation region structure graphs corresponding to the network to be detected have overlapping regions.

3. The method according to claim 2, characterized in that, The step of dividing the relational structure graph into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected includes: Based on each object node in the relational structure graph, a fitting structure graph corresponding to the relational structure graph is constructed; wherein, the fitting structure graph includes an initial number of initial aggregation regions, and the set of relational edges corresponding to the fitting structure graph is determined based on the initial membership strength between each object node in the relational structure graph and each initial aggregation region, and the initial membership strength is used to initialize the probability that the object node belongs to the initial aggregation region; Based on the initial membership strength, a first objective function corresponding to the fitted structure graph is obtained. The first objective function is used to characterize the likelihood between the set of relation edges corresponding to the fitted structure graph and the set of relation edges corresponding to the relation structure graph. The first objective function is iteratively optimized to obtain the optimized membership strength between each object node and each initial aggregation region in the fitted structure graph. Based on the optimized membership strength between each object node in the fitted structure graph and each initial aggregation region, the object nodes in the relationship structure graph are divided into aggregation regions to obtain multiple aggregation region structure graphs corresponding to the network to be detected.

4. The method according to claim 3, characterized in that, The step of obtaining the first objective function corresponding to the fitted structure graph based on the initial membership strength includes: For the first object node and the second object node in the fitted structure graph, based on the initial membership strength between the first object node and each of the initial aggregation regions, and the initial membership strength between the second object node and each of the initial aggregation regions, a first probability function is obtained between the first object node and the second object node. The first probability function is used to characterize the possibility that there is a relation edge between the first object node and the second object node. The first sub-objective function is obtained by multiplying the first probability functions between each object node in the fitted structure graph. Based on the first probability function between each object node in the fitted structure graph, a second probability function between each object node in the fitted structure graph is obtained. The second probability function is used to characterize the possibility that there is no relation edge between the first object node and the second object node. The second sub-objective function is obtained by multiplying the second probability functions between each object node in the fitted structure graph. The first objective function is obtained by multiplying the first sub-objective function and the second sub-objective function.

5. The method according to claim 3, characterized in that, The optimized membership strength between each object node in the fitted structure graph and each initial aggregation region is used to divide each object node in the relationship structure graph into aggregation regions, resulting in multiple aggregation region structure graphs corresponding to the network to be detected, including: For each of the target initial aggregation regions, object nodes whose optimized membership strength with the target initial aggregation region is greater than a first threshold are assigned to the target initial aggregation region. Obtain the number of object nodes corresponding to each of the initial aggregation regions; The initial aggregation regions with zero object nodes are removed to obtain multiple real aggregation regions corresponding to the network to be detected. Based on the multiple real aggregation regions corresponding to the network under test, a structural diagram of the multiple aggregation regions corresponding to the network under test is obtained.

6. The method according to claim 2, characterized in that, The step of constructing relationship edges between object nodes in the object node set based on the association information between objects in the network to be detected, and generating a relationship structure graph corresponding to the network to be detected, includes: Obtain the association information between various objects in the network to be detected within the target time period; Based on the association information, an association score is determined between each object in the network to be detected, and the association score is used to characterize the degree of association between the objects; For the third object node and the fourth object node in the object node set, the correlation score between the third object node and the fourth object node is determined as the weight value of the relationship edge between the third object node and the fourth object node; Based on the weight values ​​of the relationship edges between each object node in the object node set, a relationship structure graph corresponding to the network to be detected is generated.

7. The method according to claim 6, characterized in that, The correlation score is obtained through a correlation score model, and the training process of the correlation score model is as follows: Obtain a training sample set, which includes association information corresponding to a first type of object pair, association information corresponding to a second type of object pair, and association information corresponding to a third type of object pair; wherein, the first type of object pair refers to the object pair in the reference object group whose association score is greater than a second threshold, the second type of object pair refers to the object pair in the network to be detected whose association score is greater than the second threshold, and the third type of object pair refers to the object pair in the reference object group whose association score is less than a third threshold, wherein the third threshold is less than the second threshold; The correlation score model is used to obtain the correlation prediction score set corresponding to the training sample set. Based on the degree of difference between the correlation prediction score set and the correlation label score set corresponding to the training sample set, the correlation scoring model is trained to obtain the trained correlation scoring model.

8. The method according to claim 1, characterized in that, The step of obtaining the distributed representation corresponding to the reference object group based on the aggregation region structure diagram corresponding to the reference object group, and obtaining the multiple distributed representations corresponding to the network to be detected based on the multiple aggregation region structure diagrams corresponding to the network to be detected, includes: The aggregation region structure diagram corresponding to the reference object group and the multiple aggregation region structure diagrams corresponding to the network to be detected are combined to obtain an aggregation region structure diagram set; A graph embedding model is constructed, which includes a first weight matrix and a second weight matrix. The number of rows in the first weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set, and the number of columns in the second weight matrix is ​​the same as the number of aggregate region structure graphs in the aggregate region structure graph set. Obtain the input pair set of the graph embedding model; wherein, the input pair set includes multiple input pairs corresponding to each aggregation region structure graph, and for a target aggregation region structure graph in the aggregation region structure graph set, the input pair corresponding to the target aggregation region structure graph includes a set of one-hot codes of the target aggregation region structure graph and a one-hot code of a rooted subgraph corresponding to the target aggregation region structure graph; The graph embedding model is used to obtain the transition output vector set corresponding to the input pair set. The transition output vectors in the transition output vector set are used to transitionally represent the distributed representation of the rooted subgraph in the input pair. Based on the one-hot encoding of the rooted subgraphs in the transition output vector set and the input pair set, the objective function of the graph embedding model is obtained. The objective function is used to characterize the degree of association between the rooted subgraphs in the input pair and the aggregation region structure graph in the input pair. The objective function of the graph embedding model is iteratively optimized to obtain the optimized first weight matrix; Based on each row element of the optimized first weight matrix, the distributed representations corresponding to each aggregation region structure diagram are obtained.

9. The method according to claim 8, characterized in that, The step of obtaining the objective function of the graph embedding model based on the one-hot encoding of the rooted subgraphs in the transition output vector set and the input pair set includes: For a target input pair in the input pair set, a second sub-objective function is obtained based on the transition output vector corresponding to the target input pair and the one-hot encoding of the rooted subgraph in the target input pair. The second sub-objective function is used to characterize the degree of correlation between the transition output vector corresponding to the target input pair and the rooted subgraph in the target input pair. The second objective function of the graph embedding model is obtained by summing the second sub-objective function sets corresponding to each input pair in the input pair set. Logistic regression is performed on the transition output vector corresponding to the target input pair to obtain the likelihood probability of the target input pair. The likelihood probability is used to characterize the probability that there is a rooted subgraph in the target input pair that belongs to the aggregation region structure graph in the target input pair. For the target aggregation region structure graph, the third sub-objective function corresponding to the target aggregation region structure graph is obtained by multiplying the likelihood probabilities of the multiple input pairs corresponding to the target aggregation region structure graph. The third objective function of the graph embedding model is obtained based on the sum of the third sub-objective functions corresponding to each aggregation region structure graph. Based on the second objective function and the third objective function, the objective function of the graph embedding model is obtained.

10. The method according to claim 8, characterized in that, The step of obtaining the transition output vector set corresponding to the input pair set through the graph embedding model includes: For a target input pair in the input pair set, the one-hot encoding of the aggregation region structure graph in the target input pair is multiplied by the first weight matrix to obtain a first transition matrix. The first transition matrix is ​​used to transitionally characterize the distributed representation of the aggregation region structure graph in the target input pair. Multiply the first transition matrix and the second weight matrix to obtain the transition output vector corresponding to the target input pair.

11. The method according to any one of claims 1 to 10, characterized in that, The method further includes: Obtain the black label concentration corresponding to each of the multiple aggregation region structure diagrams of the network to be detected. The black label concentration is used to characterize the proportion of objects with black labels in the aggregation region structure diagram. The aggregation region structure diagrams with black label concentrations less than the fourth threshold are removed to obtain the remaining aggregation region structure diagrams corresponding to the network to be detected. The step of obtaining the similarity between the distributed representations corresponding to the reference object group and the multiple distributed representations corresponding to the network to be detected includes: The similarity between the distributed representation corresponding to the reference object group and the distributed representation corresponding to the remaining aggregation region structure diagram is obtained respectively.

12. A device for determining a group of objects, characterized in that, The device includes: The structure graph acquisition module is used to acquire the aggregation region structure graph corresponding to the reference object group, and to acquire multiple aggregation region structure graphs corresponding to the network to be detected; wherein, the aggregation region structure graph is used to characterize the correlation between objects belonging to the same aggregation region; The distributed representation acquisition module is used to acquire the distributed representation corresponding to the reference object group based on the aggregation region structure diagram corresponding to the reference object group, and to acquire multiple distributed representations corresponding to the network to be detected based on multiple aggregation region structure diagrams corresponding to the network to be detected; wherein, the distributed representation is used to characterize the structural features of the aggregation region structure diagram; The structure graph matching module is used to obtain the similarity between the distributed representation corresponding to the reference object group and the multiple distributed representations corresponding to the network to be detected; and to determine the aggregate region structure graph whose similarity satisfies the first condition among the multiple aggregate region structure graphs corresponding to the network to be detected as the target aggregate region structure graph. The target group acquisition module is used to acquire the target object group corresponding to the target aggregation region structure diagram.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the method for determining a group of objects as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement the method for determining a group of objects as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, The computer program product includes computer instructions that are executed by a processor to implement the method for determining a group of objects as described in any one of claims 1 to 11.

Images