Artificial intelligence-based classification method and apparatus, electronic device, and storage medium

Abstract

The application provides an artificial intelligence-based classification method and device, electronic equipment and computer readable storage medium; the method comprises the following steps: constructing a network structure of a social group based on the interaction behavior of each member in the social group; determining the corresponding member vector of each member in the social network, and combining the corresponding member vector of each member to determine the combination result as the group vector of the social group; performing cascaded aggregation processing on the network structure of the social group and the social group vector to obtain the group characteristics of the social group; and performing group classification processing based on the group characteristics of the social group to obtain the group category of the social group. Through the application, the social group can be hierarchically aggregated, and the accuracy of social group classification can be improved.

Description

Technical Field

[0001] This application relates to artificial intelligence technology, and more particularly to a classification method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence. Background Technology

[0002] Artificial Intelligence (AI) is a comprehensive technology within computer science that studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities. AI technology is a multidisciplinary field encompassing a wide range of areas, including natural language processing and machine learning / deep learning. With technological advancements, AI will be applied in more fields and play an increasingly important role.

[0003] Group classification is one of the important applications in the field of artificial intelligence. A group classification system refers to a system that uses computers to process, analyze, and understand social groups in order to determine the category of the social groups.

[0004] In related technologies, social group categories are determined by the presence of keywords related to group categories within the group. However, this method of group classification relies on keywords specific to group categories, which reduces the accuracy of social group classification. Summary of the Invention

[0005] This application provides an artificial intelligence-based classification method, apparatus, electronic device, and computer-readable storage medium that can perform hierarchical aggregation of social groups and improve the accuracy of social group classification.

[0006] The technical solution of this application embodiment is implemented as follows:

[0007] This application provides an artificial intelligence-based classification method, including:

[0008] Based on the interactive behavior of each member in the social group, construct the network structure of the social group;

[0009] Determine the member vector corresponding to each member in the social network, and combine the member vectors corresponding to each member to determine the group vector of the social group;

[0010] By combining the network structure of the social group and the social group vectors, a cascaded aggregation process is performed to obtain the group features of the social group;

[0011] Based on the group characteristics of the social groups, group classification processing is performed to obtain the group categories of the social groups.

[0012] In the above technical solution, the step of constructing the network structure of the social group based on the interactive behavior of each member in the social group includes:

[0013] Perform the following processing for any two members in the social group:

[0014] When the number of interactions between the two members in the social group exceeds an interaction threshold, an edge is established between the corresponding nodes of the two members, and

[0015] The number of interactions between the two members in the social group is determined as the weight of the edge.

[0016] The edge between the nodes corresponding to the two members and the weight corresponding to the edge are determined as the connection information between the two members;

[0017] Based on the connection information between any two members in the social group, construct the network structure of the social group.

[0018] The method in the above technical solution further includes:

[0019] By combining the network structure of the social group samples and the group vectors of the social group samples, a cascaded aggregation process is performed to obtain the group features of the social group samples.

[0020] Based on the group features of the social group samples, group classification processing is performed to obtain the predicted group category of the social group;

[0021] Based on the predicted group category of the social group and the group category label of the social group, a loss function for the social group prediction model is constructed;

[0022] Update the parameters of the social group prediction model until the loss function converges, and use the updated parameters of the social group prediction model when the loss function converges as the parameters of the trained social group prediction model.

[0023] This application provides an artificial intelligence-based classification device, comprising:

[0024] The building module is used to construct the network structure of the social group based on the interaction behavior of each member in the social group;

[0025] The determining module is used to determine the member vector corresponding to each member in the social network, and combine the member vectors corresponding to each member to determine the group vector of the social group.

[0026] The aggregation module is used to perform cascaded aggregation processing on the network structure of the social group and the social group vector to obtain the group features of the social group.

[0027] The classification module is used to perform group classification processing based on the group characteristics of the social group to obtain the group category of the social group.

[0028] In the above technical solution, the device further includes:

[0029] The processing module is used to construct the network structure of the social network based on the interaction behavior of each member in the social network;

[0030] Perform the following processing for any member in the social network:

[0031] Starting from the member, the network structure of the social network is sampled to obtain the sampled trajectory of the member;

[0032] The sampling trajectory of the member is processed by word vector encoding to obtain the node representation vector of the member;

[0033] Based on the node representation vector of the member, determine the member vector of the member in the social network.

[0034] In the above technical solution, the processing module is further used to encode the shared text of the member on the social network to obtain the member's property representation vector;

[0035] The target words are extracted from the text shared by the member on the social network to obtain the number of times the target words appear, and the number of times the target words appear is determined as the sensitive representation vector of the member;

[0036] The node representation vector, the property representation vector, and the sensitivity representation vector of the member are fused to obtain the member vector.

[0037] In the above technical solution, the processing module is further configured to perform the following processing on any two members of the social network:

[0038] When the interaction weight between the two members is greater than the interaction threshold, an edge is established between the corresponding nodes of the two members, and

[0039] The interaction weight between the two members is determined as the weight corresponding to the edge;

[0040] The edge between the nodes corresponding to the two members and the weight corresponding to the edge are used as the connection information between the two members;

[0041] The network structure of the social network is constructed based on the connection information between any two members in the social network.

[0042] In the above technical solution, the interactive behavior includes communication behavior and transaction behavior; the processing module is also used to determine the communication weight corresponding to the communication behavior of the two members in the social network;

[0043] Determine the transaction weights corresponding to the transaction behaviors of the two members in the social network;

[0044] The interaction weight between the two members is determined based on the communication weight and the transaction weight.

[0045] In the above technical solution, the processing module is further configured to determine the weight of the transaction behavior by multiplying the amount corresponding to the transaction behavior by the transaction weight;

[0046] The sum of the weight of the transaction behavior and the weight of the communication is determined as the interaction weight between the two members.

[0047] In the above technical solution, the processing module is further configured to randomly sample members adjacent to the member in the network structure of the social network to obtain the sampling information corresponding to the first sampled member;

[0048] Using the first sampled member as a new starting point, continue random sampling until a preset number of sampled information are obtained, and

[0049] The identification information of the member and the preset number of sampling information are concatenated to obtain the sampling trajectory of the member;

[0050] The sampling information corresponding to the i-th sampled member includes at least one of the following data: the identifier of the i-th sampled member; the interaction weight between the i-th sampled member and the (i-1)-th sampled member, where 0 < i ≤ N, i is a natural number, and N is the preset number.

[0051] In the above technical solution, the aggregation module is further used to aggregate the network structure of the social group and the social group vector through the first aggregation layer in a plurality of cascaded aggregation layers to obtain the aggregation result of the first aggregation layer;

[0052] The aggregation result of the first aggregation layer is output to subsequent cascaded aggregation layers, so that aggregation processing and output of aggregation results continue in the subsequent cascaded aggregation layers, until the output is reached in the last aggregation layer.

[0053] The output of the last aggregation layer is determined as the group characteristics of the social group.

[0054] In the above technical solution, the aggregation module is also used to perform a first aggregation process on the network structure of the social group to obtain a first aggregated network structure;

[0055] A second aggregation process is performed on the network structure of the social group and the social group vector to obtain the first aggregated social group vector;

[0056] The network structure of the first aggregation and the social group vector of the first aggregation are determined as the aggregation result of the first aggregation layer;

[0057] The following processing is performed through the j-th aggregation layer in the plurality of cascaded aggregation layers:

[0058] Perform the first aggregation process on the (j-1)th aggregation network structure to obtain the jth aggregation network structure;

[0059] A second aggregation process is performed on the network structure of the (j-1)th aggregation and the social group vector of the (j-1)th aggregation to obtain the social group vector of the j-th aggregation.

[0060] The network structure of the j-th aggregation and the social group vector of the j-th aggregation are determined as the aggregation result of the j-th aggregation layer;

[0061] The aggregation result of the j-th aggregation layer is output to the (j+1)-th aggregation layer;

[0062] Where 2≤j<J, j is a natural number, and J is the total number of the multiple cascaded aggregation layers.

[0063] In the above technical solution, the aggregation module is further used to determine the weight vector of the first aggregation layer by multiplying the network structure of the social group, the social group vector, and the aggregation weight matrix in the first aggregation layer.

[0064] A nonlinear transformation is performed on the weight vector of the first aggregation layer to obtain the aggregation matrix in the first aggregation layer;

[0065] The transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the product of the aggregation matrix are used to determine the network structure of the first aggregation.

[0066] In the above technical solution, the aggregation module is further used to determine the feature vector of the first aggregation layer by multiplying the network structure of the social group, the social group vector, and the latent vector transformation matrix in the first aggregation layer.

[0067] A nonlinear transformation is performed on the feature vector of the first aggregation layer to obtain the intermediate hidden vector of the first aggregation layer;

[0068] The product of the transpose of the aggregation matrix in the first aggregation layer and the intermediate latent vector of the first aggregation layer is determined as the social group vector of the first aggregation.

[0069] In the above technical solution, the aggregation module is further used to determine the feature vector of the last aggregation layer by multiplying the network structure output by the previous aggregation layer, the social group vector output by the previous aggregation layer, and the latent vector transformation matrix of the last aggregation layer.

[0070] A nonlinear transformation is performed on the feature vector of the last aggregation layer to obtain the intermediate hidden vector of the last aggregation layer.

[0071] The intermediate latent vector of the last aggregation layer is determined as the group feature of the social group.

[0072] In the above technical solution, the construction module is further configured to perform the following processing on any two members of the social group:

[0073] When the number of interactions between the two members in the social group exceeds an interaction threshold, an edge is established between the corresponding nodes of the two members, and

[0074] The number of interactions between the two members in the social group is determined as the weight of the edge.

[0075] The edge between the nodes corresponding to the two members and the weight corresponding to the edge are determined as the connection information between the two members;

[0076] Based on the connection information between any two members in the social group, construct the network structure of the social group.

[0077] In the above technical solution, the device further includes:

[0078] The training module is used to perform cascaded aggregation processing on the network structure of social group samples and the group vectors of the social group samples to obtain the group features of the social group samples.

[0079] Based on the group features of the social group samples, group classification processing is performed to obtain the predicted group category of the social group;

[0080] Based on the predicted group category of the social group and the group category label of the social group, a loss function for the social group prediction model is constructed;

[0081] Update the parameters of the social group prediction model until the loss function converges, and use the updated parameters of the social group prediction model when the loss function converges as the parameters of the trained social group prediction model.

[0082] This application provides an electronic device for classification, the electronic device comprising:

[0083] Memory, used to store executable instructions;

[0084] The processor, when executing executable instructions stored in the memory, implements the artificial intelligence-based classification method provided in the embodiments of this application.

[0085] This application provides a computer-readable storage medium storing executable instructions for inducing a processor to execute and implement the artificial intelligence-based classification method provided in this application.

[0086] This application provides a computer program that includes computer instructions stored in a computer-readable storage medium. These instructions are used to cause a processor to execute the artificial intelligence-based classification method provided in this application.

[0087] The embodiments of this application have the following beneficial effects:

[0088] By determining the member vector corresponding to each member in the social group in the social network, the connection between each member in the social group and other members in the social network is established, thereby enriching the group vector of the social group. Then, by combining the network structure of the social group and the social group vector, cascaded aggregation processing is performed to effectively aggregate the information of the social group, thereby accurately classifying the social group. Attached Figure Description

[0089] Figure 1 This is a schematic diagram illustrating an application scenario of the social group classification system 10 provided in this application embodiment;

[0090] Figure 2 This is a schematic diagram of the structure of the electronic device 500 for classification provided in an embodiment of this application;

[0091] Figures 3-6 This is a flowchart illustrating the artificial intelligence-based classification method provided in an embodiment of this application;

[0092] Figure 7 This is a schematic diagram of the network structure of a group in a social network provided in an embodiment of this application;

[0093] Figure 8 This is a schematic diagram of the trajectory sampling process provided in an embodiment of this application;

[0094] Figure 9 This is a schematic diagram of the structure of the social group prediction model provided in the embodiments of this application;

[0095] Figure 10 This is a schematic diagram of the interface of the instant messaging application provided in the embodiments of this application;

[0096] Figure 11 This is a schematic diagram of the interface of the security management application provided in the embodiments of this application;

[0097] Figure 12 This is a schematic diagram of hierarchical aggregation provided in an embodiment of this application. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0099] In the following description, the terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0100] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0101] In the implementation of this application, the collection and processing of relevant data should be strictly in accordance with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0102] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0103] 1) Graph Neural Network (GNN): A type of neural network that operates directly on graph structures, primarily processing data with non-Euclidean spatial structures (graph structures). It ignores the input order of nodes; during computation, the representation of a node is influenced by its neighboring nodes, while the graph's connections remain unchanged; the graph structure allows for graph-based reasoning. Typically, a GNN consists of two modules: a propagation module and an output module. The propagation module transmits information and updates the state between nodes in the graph, while the output module provides vector representations of the nodes and edges, defining the objective function according to different tasks. Types of GNNs include Graph Convolutional Networks (GCNs), Gated Graph Neural Networks (GGNNs), and Graph Attention Networks (GATs) based on attention mechanisms.

[0104] This application provides an artificial intelligence-based classification method, apparatus, electronic device, and computer-readable storage medium that can perform hierarchical aggregation of social groups and improve the accuracy of social group classification.

[0105] The AI-based classification method provided in this application can be implemented by the terminal / server alone; or it can be implemented collaboratively by the terminal and the server. For example, the terminal can independently undertake the AI-based classification method described below, or the terminal can send the identification information of a social group to the server, and the server can execute the AI-based classification method based on the received identification information of the social group.

[0106] The electronic device for group classification provided in this application can be various types of terminal devices or servers. 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. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0107] Taking servers as an example, such as server clusters deployed in the cloud, AI as a Service (AIaaS) is offered to users. The AIaaS platform breaks down several common AI services and provides them as independent or packaged services in the cloud. This service model is similar to an AI-themed marketplace, where all users can access and use one or more AI services provided by the AIaaS platform through application programming interfaces.

[0108] For example, one type of AI cloud service could be a group classification service, where a cloud server encapsulates the classification program provided in this application embodiment. Users invoke the classification service in the cloud service through a terminal (running a client, such as an instant messaging client), causing the cloud-deployed server to call the encapsulated classification program. This program combines the network structure of social groups with social group vectors for cascading aggregation, and performs group classification processing based on the obtained group characteristics to obtain the group category of the social group. For example, in an instant messaging application, group classification is performed on a social group to obtain its group category (e.g., dance interest group, singing interest group, etc.).

[0109] See Figure 1 , Figure 1 This is a schematic diagram of the application scenario of the classification system 10 provided in the embodiment of this application. The terminal 200 connects to the server 100 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0110] Terminal 200 (running a client, such as an instant messaging client) can be used to obtain a classification request for a social group, which includes relevant information about the social group (e.g., the identifier of the social group, the identifiers of the members in the social group, etc.).

[0111] In some embodiments, a classification plugin may be embedded in the client running on the terminal to implement an AI-based classification method locally on the client. For example, after the terminal 200 obtains a classification request for a social group (including the identifier of a social group to be classified), it calls the classification plugin to implement an AI-based classification method. This method combines the network structure of the social group with the social group vector for cascaded aggregation, and performs group classification processing based on the obtained group features to obtain the group category of the social group. For example, for instant messaging applications, group classification processing is performed based on the obtained group features to obtain the group category of the social group (e.g., fan group, celebrity group, gossip group, etc.).

[0112] In some embodiments, after the terminal 200 obtains a classification request for a social group (including the identifier of a social group to be classified), it calls the classification interface of the server 100 (which can be provided as a cloud service, i.e., a classification service). The server 100 performs cascaded aggregation based on the social group's network structure and social group vector according to the social group's classification request, and performs group classification processing based on the obtained group characteristics of the social group to obtain the group category of the social group.

[0113] The structure of the electronic device for classification provided in the embodiments of this application is described below. See also... Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device 500 for classification provided in an embodiment of this application. The example given is of an electronic device 500 that is a server. Figure 2 The illustrated electronic device 500 for image processing includes at least one processor 510, a memory 550, at least one network interface 520, and a user interface 530. The various components in the electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to implement communication between these components. In addition to a data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 540.

[0114] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0115] Memory 550 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 550 described in this application embodiment is intended to include any suitable type of memory. Memory 550 may optionally include one or more storage devices physically located away from processor 510.

[0116] In some embodiments, memory 550 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0117] Operating system 551 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0118] The network communication module 552 is used to reach other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0119] In some embodiments, the AI-based classification device provided in this application can be implemented in software, for example, as a social group classification plugin in the terminal mentioned above, or as a social group classification service in the server mentioned above. Of course, this is not limited to this; the AI-based classification device provided in this application can be provided in various software embodiments, including applications, software, software modules, scripts, or code.

[0120] Figure 2 An AI-based classification device 555 stored in memory 550 is shown. It can be software in the form of programs and plugins, such as image processing plugins, and includes a series of modules, including a construction module 5551, a determination module 5552, an aggregation module 5553, a classification module 5554, a processing module 5555, and a training module 5556. The construction module 5551, the determination module 5552, the aggregation module 5553, the classification module 5554, and the processing module 5555 are used to implement the social group classification function provided in the embodiments of this application, and the training module 5556 is used to implement the training function of the social group prediction model.

[0121] As mentioned above, the AI-based classification method provided in this application can be implemented by various types of electronic devices. See also Figure 3 , Figure 3 This is a flowchart illustrating the artificial intelligence-based classification method provided in the embodiments of this application, combined with... Figure 3 The steps shown are explained.

[0122] In step 101, the network structure of the social group is constructed based on the interactive behavior of each member in the social group.

[0123] As an example of obtaining a classification request for a social group, when a social group in an instant messaging client is active, such as when members interact within the group, the client automatically obtains a classification request for that social group (including relevant information about the group, such as shared content from all members). The client forwards this classification request to the server, which then executes an AI-based classification method. This method first constructs the social group's network structure based on the interactions of each member, allowing for subsequent cascading aggregation processing based on this network structure.

[0124] In some embodiments, a network structure of a social group is constructed based on the interaction behavior of each member in the social group, including: performing the following processing on any two members in the social group: when the number of interactions between the two members in the social group is greater than the interaction threshold, establishing an edge between the corresponding nodes of the two members, and determining the number of interactions between the two members in the social group as the weight of the edge; determining the edge between the corresponding nodes of the two members and the weight of the edge as the connection information between the two members; and constructing the network structure of the social group based on the connection information between any two members in the social group.

[0125] As an example, the number of times any two members appear in a social group within a certain time period is defined as the interaction count between those two members. When the interaction count between these two members exceeds an interaction threshold, it indicates that they are relatively active in the social group. These two members can then be identified as nodes, and an edge can be established between their corresponding nodes. The interaction count between the two members in the social group is then defined as the weight of the edge. This edge and its weight constitute the connection information between the two members. Once the connection information between all members in the social group is determined, the network structure of the social group is constructed based on the connection information between any two members. This network structure represents the connection relationships among the members in the social group, facilitating subsequent cascading aggregation based on the social group's network structure.

[0126] For example, such as Figure 7As shown, nodes with different numbers represent different members in the social group. The edge between two nodes indicates that the members corresponding to the two nodes interact frequently, that is, the number of interactions between the members corresponding to the two nodes in the social group is greater than the interaction threshold. When the interaction threshold is 1 and member i (corresponding to node 701) and member j (corresponding to node 702) appear 3 times in the social group within a preset time, it is determined that the number of interactions between member i and member j in the social group is 3, and an edge 703 is established between the nodes corresponding to member i and member j. The weight corresponding to edge 703 is determined to be 3. Then, the edge between the nodes corresponding to member i and member j and the weight (3) corresponding to the edge are determined as the connection information between member i and member j.

[0127] In step 102, the member vector corresponding to each member in the social network is determined, and the member vectors corresponding to each member are combined to determine the group vector of the social group.

[0128] Steps 101 and 102 do not have a clear sequential order. After the instant messaging client receives a classification request for a social group, it forwards the request to the server. The server then executes an AI-based classification method to determine the classification of the social group. First, it identifies each member in the social group. Then, it determines the member vector corresponding to each member in the social network. These member vectors are then combined, and the combined result is determined as the group vector of the social group. In other words, the group vector of the social group is composed of the member vectors of each member in the social group. Since each member in the social group is connected to other members in the social network, determining the member vector corresponding to each member in the social group enriches the group vector of the social group. Subsequent cascading aggregation processing can then be performed based on the group vector of the social group.

[0129] See Figure 4 , Figure 4 This is an optional flowchart illustrating an artificial intelligence-based classification method provided in this application embodiment. Before determining the member vector corresponding to each member in the social network, it is also necessary to determine the member vector of each member in the social network based on the network structure. Figure 4 Show Figure 3It also includes steps 105 to 108: In step 105, the network structure of the social network is constructed based on the interaction behavior of each member in the social network; In step 106, the following processing is performed for any member in the social network: starting from the member, the network structure of the social network is sampled to obtain the member's sampled trajectory; In step 107, the sampled trajectory of the member is processed by word vector encoding to obtain the member's node representation vector; In step 108, the member vector of the member in the social network is determined according to the member's node representation vector.

[0130] For example, firstly, based on the interaction behavior of each member in the social network, the network structure of the social network is constructed. After determining the network structure, for any member k (k is a natural number) in the social network, the trajectory of the constructed social network structure is sampled starting from member k to obtain the sampled trajectory of member k. Then, word vector encoding is performed on the sampled trajectory of member k to obtain the node representation vector of member k. Based on the node representation vector of member k, the member vector of member k is determined, thereby determining the member vector of all members in the social network. Subsequently, after determining all members in the social group, the corresponding member vector in the social network is determined based on the members in the social group, thereby determining the group vector of the social group.

[0131] In some embodiments, the network structure of a social network is constructed based on the interaction behavior of each member in the social network, including: performing the following processing on any two members in the social network: when the interaction weight between the two members is greater than the interaction threshold, establishing an edge between the corresponding nodes of the two members, and determining the interaction weight between the two members as the weight corresponding to the edge; using the edge between the corresponding nodes of the two members and the weight corresponding to the edge as the connection information between the two members; and constructing the network structure of the social network based on the connection information between any two members in the social network.

[0132] Continuing with the example above, for any two members (member n and member m) in a social network, if the interaction weight between members n and member m is greater than the interaction threshold, it indicates that members n and member m are relatively active in the social network. Member n and member m can be identified as nodes, and an edge can be established between the corresponding nodes of members n and member m. The interaction weight between members n and member m can be determined as the weight of the edge. Thus, the edge between the corresponding nodes of members n and member m and the weight of the edge are determined as the connection information between members n and member m. After determining the connection information between all members in the social network, the network structure of the social network can be constructed based on the connection information between any two members in the social network. The network structure of the social network represents the connection relationship between each member in the social network, so that the member vector of each member in the social network can be determined later based on the network structure of the social network.

[0133] In some embodiments, the interaction behavior includes communication behavior and transaction behavior; before establishing the edge between the corresponding nodes of the two members, the method further includes: determining the communication weights corresponding to the communication behaviors of the two members in the social network; determining the transaction weights corresponding to the transaction behaviors of the two members in the social network; and determining the interaction weights between the two members based on the communication weights and transaction weights.

[0134] Following the example above, to determine whether a social group is a dance interest group, when constructing the network structure of the social network, it is necessary to determine the communication weight corresponding to the communication behavior and the transaction weight corresponding to the transaction behavior of any two members in the social network. After determining the communication weight and the transaction weight, the communication weight and the transaction weight are weighted and summed to obtain the interaction weight between the two members, so that the connection information between the two members can be constructed subsequently based on the interaction weight between the two members.

[0135] For example, in order to determine the interaction weight between two members based on communication weight and transaction weight, the product of the amount corresponding to the transaction behavior and the transaction weight can be used to determine the weight of the transaction behavior, and the sum of the weight of the transaction behavior and the communication weight can be used to determine the interaction weight between the two members.

[0136] In some embodiments, before determining the member vector of a member in a social group, the method further includes: encoding the text shared by the member on the social network to obtain the member's property representation vector; extracting target words from the text shared by the member on the social network to obtain the number of times the target words appear, and determining the number of times the target words appear as the member's sensitivity representation vector; correspondingly, determining the member vector of a member in the social group based on the member's node representation vector includes: fusing the member's node representation vector, the member's property representation vector, and the member's sensitivity representation vector to obtain the member vector.

[0137] Continuing the example above, for member k in a social network, we determine the text (e.g., status updates) shared by member k on the social network, and perform natural language processing on the shared text to obtain the property representation vector of member k. Additionally, we can determine the number of times target words appear in the shared text of member k on the social network using keyword extraction techniques, and use the frequency of target word occurrences as the sensitivity representation vector of member k. Finally, we perform a weighted sum of member k's node representation vector, property representation vector, and sensitivity representation vector to obtain member k's member vector. By fusing member's node representation vector, property representation vector, and sensitivity representation vector, we can aggregate various scattered information about the member to obtain a richer member vector.

[0138] For example, to determine whether a social group is a dance interest group, keyword extraction technology is used to determine the number of times target words (including words such as "popular dance" and "street dance") appear in member k's shared text on social networks; to determine whether a social group is a gossip group, keyword extraction technology is used to determine the number of times target words (e.g., words such as "celebrity" and "popularity") appear in member k's shared text on social networks.

[0139] In some embodiments, trajectory sampling is performed on the network structure of a social network to obtain the sampling trajectory of a member, including: randomly sampling members adjacent to the member in the network structure of the social network to obtain the sampling information corresponding to the first sampled member; taking the first sampled member as a new starting point, continuing random sampling until a preset number of sampling information is sampled, and connecting the member's identification information and the preset number of sampling information to obtain the member's sampling trajectory.

[0140] For any member k in the social network, the sampling information corresponding to the i-th sampled member includes at least one of the following data: the identifier of the i-th sampled member; the interaction weight between the i-th sampled member and the (i-1)-th sampled member, where 0 < i ≤ N, i is a natural number, and N is a preset number. For example, if the sampling information corresponding to the i-th sampled member includes the identifier of the i-th sampled member and the interaction weight between the i-th sampled member and the (i-1)-th sampled member, then the sampling information corresponding to the 1-th sampled member includes the identifier of the 1-th sampled member and the interaction weight between the 1-th sampled member and the 0th sampled member (i.e., member k).

[0141] For example, for any member k in a social network, firstly, random sampling is performed on the members adjacent to member k to obtain the sampling information corresponding to the first sampled member. Then, the first sampled member is used as a new starting point to continue random sampling on the members adjacent to the first sampled member to obtain the sampling information corresponding to the second sampled member. Then, the second sampled member is used as a new starting point to continue random sampling until N sampling information is obtained, that is, the sampling information corresponding to the first to Nth sampled members. Finally, the identification information of member k and the N sampling information are connected to obtain the sampling trajectory of member k.

[0142] like Figure 8 As shown, nodes with different numbers represent different members in the social network. An edge between two nodes indicates that the members corresponding to those two nodes interact frequently in the social network, meaning that the interaction weight of the members corresponding to those two nodes is greater than the interaction threshold. After determining the network structure of the social network, for any member (node ​​801) in the social network, starting from node 801, random sampling is performed on the members adjacent to member k to obtain the sampling information corresponding to the first sampled member (node ​​802). This first sampled member is then used as a new starting point to continue random sampling on the members adjacent to the first sampled member to obtain the sampling information corresponding to the second sampled member (node ​​803). Finally, the identification information of member k, the sampling information corresponding to node 802, and the sampling information corresponding to node 803 are concatenated to obtain the sampling trajectory of member k.

[0143] In step 103, the network structure of the social group and the social group vector are combined and aggregated in a cascaded manner to obtain the group features of the social group.

[0144] After the server obtains the network structure and vectors of the social groups, it performs cascaded aggregation processing by combining the network structure and vectors of the social groups to obtain the group features of the social groups. Through hierarchical aggregation, it effectively extracts scattered information from the social groups so that accurate group classification can be performed based on accurate group features in the future.

[0145] See Figure 5 , Figure 5 This is an optional flowchart illustrating an artificial intelligence-based classification method provided in an embodiment of this application. Figure 5 Show Figure 3 Step 103 can be implemented through steps 1031 to 1033: In step 1031, the network structure and social group vector of the social group are aggregated through the first aggregation layer in a series of cascaded aggregation layers to obtain the aggregation result of the first aggregation layer; In step 1032, the aggregation result of the first aggregation layer is output to the subsequent cascaded aggregation layers to continue the aggregation processing and output of aggregation results in the subsequent cascaded aggregation layers until the last aggregation layer is output; In step 1033, the output of the last aggregation layer is determined as the group feature of the social group.

[0146] The aggregation network used for aggregation consists of multiple cascaded aggregation layers. After the server obtains the network structure and social group vectors of the social groups, the network structure and social group vectors are aggregated through the first aggregation layer in the aggregation network to obtain the aggregation result of the first aggregation layer. The aggregation result of the first aggregation layer is then output to subsequent cascaded aggregation layers to continue the aggregation processing and output of aggregation results in subsequent cascaded aggregation layers until the output is reached at the last aggregation layer. Finally, the output of the last aggregation layer is determined as the group feature of the social group.

[0147] As an example, see Figure 9 The social group prediction model for social group prediction obtains the network structure of social groups through the construction module 5551. The construction module 5551 outputs the network structure of social groups to the first aggregation layer in the aggregation module 5553. After obtaining the social group vector through the determination module 5552, the determination module 5552 outputs the social group vector to the first aggregation layer in the aggregation module 5553. The first aggregation layer in the aggregation module 5553 aggregates the network structure and social group vector of social groups to obtain the aggregation result of the first aggregation layer. The aggregation result of the first aggregation layer is then output to the second aggregation layer to continue the aggregation processing and output of the aggregation result until the output reaches the Jth aggregation layer. Finally, the output of the Jth aggregation layer is determined as the group feature of the social group, where J is the total number of cascaded aggregation layers.

[0148] In some embodiments, the network structure and social group vectors of social groups are aggregated to obtain the aggregation result of the first aggregation layer, including: performing a first aggregation process on the network structure of social groups to obtain a first aggregated network structure; performing a second aggregation process on the network structure and social group vectors of social groups to obtain the first aggregated social group vectors; determining the first aggregated network structure and the first aggregated social group vectors as the aggregation result of the first aggregation layer; correspondingly, aggregation processing and aggregation result output are continued in subsequent cascaded aggregation layers, including: through multiple The j-th aggregation layer in the cascaded aggregation layers performs the following processing: A first aggregation process is applied to the network structure of the (j-1)-th aggregation to obtain the network structure of the j-th aggregation; a second aggregation process is applied to the network structure of the (j-1)-th aggregation and the social group vector of the (j-1)-th aggregation to obtain the social group vector of the j-th aggregation; the network structure of the j-th aggregation and the social group vector of the j-th aggregation are determined as the aggregation result of the j-th aggregation layer; the aggregation result of the j-th aggregation layer is output to the (j+1)-th aggregation layer; where 2≤j<J, j is a natural number, and J is the total number of multiple cascaded aggregation layers.

[0149] Continuing with the example above, the following processing is performed in the first aggregation layer of multiple cascaded aggregation layers: A first aggregation is performed on the network structure of the social groups to obtain the network structure of the first aggregation; a second aggregation is performed on the network structure and social group vector of the social groups to obtain the social group vector of the first aggregation; the network structure and social group vector of the first aggregation are then determined as the aggregation result of the first aggregation layer, and the aggregation result of the first aggregation layer is output to the subsequent cascaded aggregation layers. The following processing is performed in the j-th aggregation layer (2≤j<J, where j is a natural number and J is the total number of cascaded aggregation layers): A first aggregation is performed on the network structure of the (j-1)-th aggregation to obtain the network structure of the j-th aggregation; a second aggregation is performed on the network structure and social group vector of the (j-1)-th aggregation to obtain the social group vector of the j-th aggregation; the network structure and social group vector of the j-th aggregation are then determined as the aggregation result of the j-th aggregation layer; finally, the aggregation result of the j-th aggregation layer is output to the (j+1)-th aggregation layer.

[0150] In some embodiments, before performing a first aggregation process on the network structure of the social group to obtain the first aggregated network structure, the method further includes: determining the weight vector of the first aggregation layer by multiplying the network structure of the social group, the social group vector, and the aggregation weight matrix in the first aggregation layer; performing a nonlinear transformation on the weight vector of the first aggregation layer to obtain the aggregation matrix in the first aggregation layer; correspondingly, performing a first aggregation process on the network structure of the social group to obtain the first aggregated network structure includes: determining the first aggregated network structure by multiplying the transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the aggregation matrix.

[0151] Continuing with the example above, before aggregation in the first aggregation layer, the aggregation matrix in the first aggregation layer is determined based on the network structure of the social groups, the social group vectors, and the aggregation weight matrix in the first aggregation layer. This matrix is ​​then used for subsequent aggregation processing. The formula for calculating the aggregation matrix in the first aggregation layer is as follows: ,in, The network structure representing social groups, Represents a social group vector. This represents the aggregation weight matrix in the first aggregation layer. This represents the aggregation matrix in the first aggregation layer. The aggregation matrix in the first aggregation layer of this application is not limited to... Other variations of the formula are also applicable to the embodiments of this application.

[0152] Correspondingly, after obtaining the aggregation matrix in the first aggregation layer, the server determines the network structure of the first aggregation by multiplying the transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the aggregation matrix. The calculation formula is as follows: ,in, This represents the aggregation matrix in the first aggregation layer. The network structure representing social groups, This represents the network structure of the first aggregation. The network structure of the first aggregation in this embodiment is not limited to... Other variations of the formula are also applicable to the embodiments of this application.

[0153] Correspondingly, the product of the social group's network structure, the social group vector, and the latent vector transformation matrix in the first aggregation layer is determined as the feature vector of the first aggregation layer. A nonlinear transformation is then performed on the feature vector of the first aggregation layer to obtain the intermediate latent vector of the first aggregation layer. The formula for calculating the intermediate latent vector of the first aggregation layer is as follows: ,in, This represents the intermediate hidden vector of the first aggregation layer. The network structure representing social groups, Represents a social group vector. This represents the implicit vector transformation matrix in the first aggregation layer. This represents a nonlinear transformation, i.e., an activation function (such as the ReLU function). This represents the feature vector of the first aggregation layer. Finally, the product of the transpose of the aggregation matrix in the first aggregation layer and the intermediate latent vector of the first aggregation layer is used to determine the social group vector of the first aggregation. The formula for calculating the social group vector of the first aggregation is as follows: The intermediate hidden vectors of the first aggregation layer in this application embodiment are not limited to... Other variations of the formula are also applicable to the embodiments of this application.

[0154] Correspondingly, the product of the network structure output from the previous aggregation layer, the social group vector output from the previous aggregation layer, and the latent vector transformation matrix of the last aggregation layer is determined as the feature vector of the last aggregation layer. A non-linear transformation is then performed on the feature vector of the last aggregation layer to obtain the intermediate latent vector of the last aggregation layer. This intermediate latent vector is then determined as the group feature of the social group. The formula for calculating the intermediate latent vector of the first aggregation layer is as follows: ,in, This represents the intermediate hidden vector of the last aggregation layer. This represents the network structure output by the (J-1)th aggregation layer. This represents the social group vector output by the (J-1)th aggregation layer. This represents the implicit vector transformation matrix in the last aggregation layer.

[0155] In step 104, group classification is performed based on the group characteristics of the social group to obtain the group category of the social group.

[0156] After the server obtains the group features of the social group, it uses a classifier to map the group features of the social group to a probability distribution of group categories, and determines the group category corresponding to the highest probability in the probability distribution as the group category of the social group.

[0157] As an example, see Figure 9 The aggregation module 5553 outputs the group features of the social group to the classification module 5554. The classification module 5554 performs group classification processing based on the group features of the social group to obtain the group category of the social group.

[0158] For example, such as Figure 10As shown, for unnamed social groups 1001 in instant messaging applications, group classification is performed based on the group characteristics of unnamed social groups to obtain the group category of the social group (e.g., fan group, celebrity group, gossip group, etc.); Figure 11 As shown, group classification is performed based on the group characteristics of the obtained social groups to obtain the group categories of the social groups, such as the idol-chasing group 1101.

[0159] See Figure 6 , Figure 6 This is an optional flowchart illustrating an AI-based classification method provided in this application embodiment. It requires training a social group prediction model to obtain the trained model for social group classification. Figure 6 Show Figure 3 The method also includes steps 109 to 112: In step 109, the network structure of the social group samples and the group vectors of the social group samples are combined and aggregated in a cascaded manner to obtain the group features of the social group samples; in step 110, group classification is performed based on the group features of the social group samples to obtain the predicted group category of the social group; in step 111, the loss function of the social group prediction model is constructed according to the predicted group category of the social group and the group category label of the social group; in step 112, the parameters of the social group prediction model are updated until the loss function converges, and the updated parameters of the social group prediction model when the loss function converges are used as the parameters of the trained social group prediction model.

[0160] Steps 109 to 112 do not have a clear sequential order with steps 101 to 104. For example, the following processing is performed through the social group prediction model: Based on the interaction behavior of each member in the social group sample, a network structure of the social group sample is constructed; the member vector corresponding to each member in the social network is determined, and the member vectors corresponding to each member are combined to determine the group vector of the social group sample; the network structure of the social group sample and the social group vector are combined and aggregated to obtain the group features of the social group sample; group classification is performed based on the group features of the social group to obtain the predicted group category of the social group; after determining the value of the loss function of the social group prediction model according to the predicted group category and the group category label of the social group, it can be determined whether the value of the loss function of the social group prediction model exceeds a preset threshold. When the value of the loss function of the social group prediction model exceeds the preset threshold, the error signal of the social group prediction model is determined based on the loss function of the social group prediction model, the error information is backpropagated in the social group prediction model, and the model parameters of each layer are updated during the propagation process.

[0161] Here, we explain backpropagation. Training sample data is input into the input layer of the neural network model, passes through the hidden layers, and finally reaches the output layer to output the result. This is the forward propagation process of the neural network model. Since there is an error between the output result and the actual result, the error between the output result and the actual value is calculated and propagated back from the output layer to the hidden layers until it reaches the input layer. During the backpropagation process, the values ​​of the model parameters are adjusted according to the error. This process is iterated until convergence. The social group prediction model is an example of a neural network model.

[0162] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0163] The embodiments of this application can be applied to various social group classification application scenarios, such as the application scenario of determining whether a social group is a target group, such as... Figure 1 As shown, terminal 200 connects to server 100 deployed in the cloud via network 300. After installing a security management application on terminal 200 and obtaining a social group classification request, it calls the social group classification interface of server 100. Server 100 classifies the social group according to the social group classification request to obtain the group category of the social group.

[0164] With the widespread use of social media, due to considerations of privacy and security, social media platforms cannot solely rely on filtering the content of messages sent by individuals to determine whether a group in a social network is a target group.

[0165] In related technologies, three techniques are used to identify whether a group is a target group: Solution A) Keyword filtering technology, which determines the theme and characteristics of the group by filtering keywords. If the group uses a large number of keywords, it indicates that the group is a target group; Solution B) Object manual verification technology, which performs manual verification based on screenshots or records uploaded by the informant after the object provides feedback on the group; Solution C) Readout graph convolutional neural network technology, which constructs a group network, transmits information within the group, and finally represents the group by reading out the representations of all nodes in the group through a readout function.

[0166] However, the methods in the related technologies have the following problems: In Solution A, keyword filtering has a significant impact, and more difficult-to-identify terms are designed to avoid keywords. Therefore, in practical applications, the accuracy of this method is low. Solution B mainly relies on manual feedback from the target audience. However, for internal groups, the probability of internal targets (members) reporting is low. Therefore, it is difficult to obtain organized group-building behavior through this method. Solution C requires reading the information of all nodes in the group through a read function to represent the group. When there are many nodes in the group, since the read function is a pooling function, it is difficult for the read function to take into account the detailed information of the objects in the group. A lot of information will be lost after pooling and reading.

[0167] To address the aforementioned issues, this application proposes a classification method based on hierarchical pooling graph neural networks (an artificial intelligence-based classification method). This method performs bottom-up hierarchical aggregation of nodes within a group during graph convolution propagation, enriching the information within the aggregated group and ultimately representing the group as a single point. The proposed method, through hierarchical aggregation, can more effectively extract fragmented information from the group structure, thereby enabling more accurate identification of target groups. Furthermore, by leveraging the relationships between members within the group within the social network, information from members outside the group can be effectively integrated.

[0168] Compared to Scheme A and Scheme B, the method proposed in this application uses an automatic detection method based on graph neural networks to automatically identify group functions. Since it is not affected by the language used by the objects or the reporting behavior of the objects, even if the objects that create the group use automatic detection avoidance technology, the method proposed in this application can still effectively detect the target group through factors such as social relationships between objects, attribute characteristics, and group member connections. Compared to Scheme C, the method proposed in this application, based on a hierarchical aggregation graph neural network, can gradually reduce the number of nodes in the group by using one point to represent multiple points through aggregation, thereby increasing the amount of information in each node in the group, thus effectively solving the problem of information loss in the pooling function.

[0169] The classification method based on hierarchical pooling graph neural networks proposed in this application is described in detail below:

[0170] The method proposed in this application includes the following: 1) a social network construction module; 2) an object representation learning module; 3) a swarm graph network construction module; 4) a hierarchical aggregation swarm representation module; and 5) a swarm prediction module.

[0171] 1) Social Network Building Module

[0172] First, by measuring the behavior between objects, all objects in the network are connected into a social network. Then, the object representation of all objects in the network is learned using the social network object representation method.

[0173] The edges between objects are determined by their one-to-one interactions. If there are many interactions between two objects, the edge between them has a higher weight; if there are few interactions, the edge has a lower weight; and if there are no interactions, there is no edge between them.

[0174] Interactions between objects consist of the following: communication behaviors (e.g., sending messages, liking and replying); and transaction behaviors (e.g., sending red envelopes and transferring money). Communication behaviors can be measured by the following factors: the number of interaction behaviors, the cumulative duration of interaction behaviors, the interaction discussion, the order of interaction frequency, and the number of interaction days in the most recent week. Transaction information can also be measured in the same way. When constructing a social network, it is necessary to weigh the weights of communication behaviors and transaction behaviors simultaneously. The transaction behavior can consider the size of the transaction amount. Therefore, the weight between two objects can be calculated using the following formula (1):

[0175] (1)

[0176] in, This represents the weight (interaction weight) between object i and object j. This represents the weight of the interaction behavior between object i and object j. This represents the weight of the transaction behavior between object i and object j. This represents the transaction amount between object i and object j. Multiplying the object transaction weight and the object transaction amount is to describe the behavior between the objects from both the frequency and amount perspectives.

[0177] 2) Object Representation Learning Module

[0178] The object representation learning module refers to learning through social network embeddings and combining this with descriptions of object characteristics based on their published statements. Below, we need to comprehensively describe the object from three aspects as attribute features.

[0179] 1. Social network embedding methods include DeepWalk and Node2Vec node embedding methods. Taking DeepWalk as an example, based on a given social network, starting from each node in the social network, multiple random walks are performed according to the network structure and object connection weights obtained from the social network construction module. Then, all the walks are used as a corpus and input into the word2vec word vector embedding algorithm. The word2vec algorithm is used for training, and finally, the node representation vector of each object is obtained. In the node representation vector, the representations of nearby nodes in the network are relatively similar.

[0180] 2. Information such as the object's avatar and status can be transformed into a vector (property representation vector) describing the object's properties through image processing technology, natural language processing (NLP) technology, object tagging technology, etc.

[0181] 3. Use keyword extraction technology to extract the frequency of occurrence of keywords in the shared content of the object, and determine the frequency of occurrence as the target word vector (sensitive representation vector) describing the object.

[0182] Ultimately, these three aspects are combined to describe the object as attribute characteristics.

[0183] 3) Group Graph Network Construction Module

[0184] For example, a group is a subgraph in a complete social network, and groups have a completely different network structure from social networks. This application's embodiment uses a context-order-based group network construction module. That is, if two people reply within a very short period, their relationship is relatively close. Therefore, by counting the number of times each member in the group co-occurs in a recent period (e.g., 3 days), the connection weight between objects can be obtained. The weight between objects is defined as follows: (2)

[0185] (2)

[0186] in, This indicates the number of times object i and object j appear together. This represents the weight (interaction weight) between object i and object j in group m. To avoid sparse connections within the social network, objects that co-occur once or less are defined as unconnected. This represents the adjacency matrix of the m-th group (the network structure of the m-th group).

[0187] The weights between nodes can be constructed using the method described above, resulting in the following: Figure 7 The network structure of the group shown is a network structure of a group in a social network, in which there are connections between nodes, for example, there is a connection between node 701 and node 702.

[0188] 4) Readout group representation module

[0189] Through the above group adjacency matrix And the object representation matrix X is used to learn the representation of each group. The readout group representation consists of hierarchical aggregation and information readout. First, it employs... The attribute features of the group are represented by the representation of the corresponding object in the group extracted from the object representation X of the social network. First, the standardized adjacency matrix is ​​calculated as shown in formula (3):

[0190] (3)

[0191] Where D is The Laplace matrix D has only diagonal elements that are non-zero. The diagonal elements can be calculated using the following formula (4):

[0192] (4)

[0193] Among them, can This can be simply referred to as the initial adjacency matrix. (The initial adjacency matrix of the m-th group).

[0194] Regarding K-layer graph convolutional neural networks, we first start with the original number of group members and then continuously aggregate group members using a soft aggregation method until the Kth layer is reached. Specifically, we employ... Let m represent the initial group members of the m-th group. This indicates the initial number of group members.

[0195] In this context, soft aggregation refers to a group member being able to be divided into several different categories, where the amount of information about that member in each category is determined by the soft aggregation matrix. Conversely, hard aggregation means that each member can only be assigned to one category and cannot be assigned to any other category. Both soft and hard aggregation are applicable to the embodiments of this application.

[0196] In order to soft aggregate group members, a soft aggregation matrix is ​​used to update two parts of variables: one is the update of the adjacency matrix A, and the other is the update of the member vector X.

[0197] Here we introduce a soft aggregation matrix S, where the dimension of S is... Given a soft aggregation matrix, the update of the adjacency matrix A can be shown in equation (5):

[0198] (5)

[0199] Correspondingly, the member vectors are updated as shown in formulas (6) and (7):

[0200] (6)

[0201] (7)

[0202] in, Indicates the first l Hidden variables of objects in the layer, Indicates the first l The latent variable transformation matrix of the layer.

[0203] Accordingly, the update of the soft aggregation matrix is ​​shown in Equation (8):

[0204] (8)

[0205] in, Indicates the first l The soft aggregation weight matrix of the layer.

[0206] This allows you to set the number of nodes in each layer of the hierarchical aggregation, for example... , It is determined by the structure of each group; different groups are different. It is set to 1. Therefore, after passing through K layers, it can be transmitted through a vector. To describe the group. For example... Figure 12 As shown, the convolutional neural network in the diagram includes three aggregation layers. The network structure of the m-th group includes 16 initial nodes (i.e., the initial group member representations). After aggregation processing by the first aggregation layer, the network structure of the m-th group yields a network structure with 5 nodes. The network structure with 5 nodes is then input into the second aggregation layer, resulting in a network structure with 2 nodes. The network structure with 2 nodes is then input into the third aggregation layer, resulting in a network structure with one node 1201. That is, the vector representation corresponding to node 1201 is a vector. .pass This represents the representation vector of the m-th group.

[0207] 5) Group Prediction Module

[0208] The group prediction module is used to train the model to predict whether a group is a legitimate group based on the input group's representation vector. Here, a multilayer perceptron is used as the classifier for the group prediction module (social group prediction model), and its calculation formula is shown in formula (9):

[0209] (9)

[0210] Where MLP stands for Multilayer Perceptron, and its input dimension is... The output dimension is 1-dimensional.

[0211] Furthermore, the model is trained using squared error. In this embodiment, all trainable parameters are: Until Its loss function is shown in equation (10):

[0212] (10)

[0213] in, Let m represent the true label of the m-th group. If the m-th group is the target group, then... If the m-th group is not the target group, then .

[0214] In practical group prediction, a group graph network is first constructed. Then, a hierarchical aggregation group representation module is used to propagate information and obtain the group representation. Finally, the predicted value is calculated using the prediction function (formula (9)) in the group prediction module to determine whether the group belongs to the target group.

[0215] This concludes the description of the AI-based classification method provided in this application, using the exemplary application and implementation of the server provided in the embodiments of this application. Figure 2 An AI-based classification device 555 stored in memory 550 is shown. It can be software in the form of programs and plugins, such as image processing plugins, and includes a series of modules, including a construction module 5551, a determination module 5552, an aggregation module 5553, a classification module 5554, a processing module 5555, and a training module 5556. The construction module 5551, the determination module 5552, the aggregation module 5553, the classification module 5554, and the processing module 5555 are used to implement the social group classification function provided in the embodiments of this application, and the training module 5556 is used to implement the training function of the social group prediction model.

[0216] The following describes the scheme for classifying social groups by coordinating the various modules in the AI-based classification device 555 provided in this application embodiment.

[0217] The construction module 5551 is used to construct the network structure of the social group based on the interaction behavior of each member in the social group; the determination module 5552 is used to determine the member vector corresponding to each member in the social network, and combine the member vectors corresponding to each member to determine the group vector of the social group; the aggregation module 5553 is used to perform cascaded aggregation processing on the network structure of the social group and the social group vector to obtain the group features of the social group; the classification module 5554 is used to perform group classification processing based on the group features of the social group to obtain the group category of the social group.

[0218] In some embodiments, the AI-based classification device 555 further includes: a processing module 5555, configured to construct the network structure of the social network based on the interaction behavior of each member in the social network; and to perform the following processing for any member in the social network: starting from the member, performing trajectory sampling on the network structure of the social network to obtain the member's sampling trajectory; performing word vector encoding processing on the member's sampling trajectory to obtain the member's node representation vector; and determining the member vector of the member in the social network based on the member's node representation vector.

[0219] In some embodiments, the processing module 5555 is further configured to encode the shared text of the member on the social network to obtain the member's property representation vector; extract target words from the shared text of the member on the social network to obtain the number of times the target words appear, and determine the number of times the target words appear as the member's sensitivity representation vector; and fuse the member's node representation vector, the member's property representation vector, and the member's sensitivity representation vector to obtain the member's member vector.

[0220] In some embodiments, the processing module 5555 is further configured to perform the following processing on any two members in the social network: when the interaction weight between the two members is greater than the interaction threshold, establish an edge between the corresponding nodes of the two members, and determine the interaction weight between the two members as the weight corresponding to the edge; use the edge between the corresponding nodes of the two members and the weight corresponding to the edge as the connection information between the two members; and construct the network structure of the social network based on the connection information between any two members in the social network.

[0221] In some embodiments, the interactive behavior includes communication behavior and transaction behavior; the processing module 5555 is further configured to determine the communication weight corresponding to the communication behavior of the two members in the social network; determine the transaction weight corresponding to the transaction behavior of the two members in the social network; and determine the interaction weight between the two members based on the communication weight and the transaction weight.

[0222] In some embodiments, the processing module 5555 is further configured to determine the weight of the transaction behavior by multiplying the amount corresponding to the transaction behavior by the transaction weight; and to determine the interaction weight between the two members by summing the weight of the transaction behavior and the communication weight.

[0223] In some embodiments, the processing module 5555 is further configured to, in the network structure of the social network, randomly sample members adjacent to the member to obtain sampling information corresponding to the first sampled member; take the first sampled member as a new starting point, continue to perform random sampling processing until a preset number of sampling information is sampled, and perform connection processing on the member's identification information and the preset number of sampling information to obtain the member's sampling trajectory; wherein, the sampling information corresponding to the i-th sampled member includes at least one of the following data: the identification of the i-th sampled member; the interaction weight between the i-th sampled member and the (i-1)-th sampled member, 0 < i ≤ N, where i is a natural number and N is the preset number.

[0224] In some embodiments, the aggregation module 5553 is further configured to perform aggregation processing on the network structure of the social group and the social group vector through the first aggregation layer in a plurality of cascaded aggregation layers to obtain the aggregation result of the first aggregation layer; output the aggregation result of the first aggregation layer to the subsequent cascaded aggregation layers to continue the aggregation processing and aggregation result output in the subsequent cascaded aggregation layers until the last aggregation layer is output, and determine the output of the last aggregation layer as the group feature of the social group.

[0225] In some embodiments, the aggregation module 5553 is further configured to perform a first aggregation process on the network structure of the social group to obtain a first aggregated network structure; perform a second aggregation process on the network structure of the social group and the social group vector to obtain a first aggregated social group vector; determine the first aggregated network structure and the first aggregated social group vector as the aggregation result of the first aggregation layer; and perform the following processing through the j-th aggregation layer in the plurality of cascaded aggregation layers: perform a first aggregation process on the (j-1)-th aggregated network structure to obtain a j-th aggregated network structure; perform a second aggregation process on the (j-1)-th aggregated network structure and the (j-1)-th aggregated social group vector to obtain a j-th aggregated social group vector; determine the j-th aggregated network structure and the j-th aggregated social group vector as the aggregation result of the j-th aggregation layer; and output the aggregation result of the j-th aggregation layer to the (j+1)-th aggregation layer; where 2≤j<J, j is a natural number, and J is the total number of the plurality of cascaded aggregation layers.

[0226] In some embodiments, the aggregation module 5553 is further configured to determine the weight vector of the first aggregation layer by multiplying the network structure of the social group, the social group vector, and the aggregation weight matrix in the first aggregation layer; to perform a nonlinear transformation on the weight vector of the first aggregation layer to obtain the aggregation matrix in the first aggregation layer; and to determine the network structure of the first aggregation by multiplying the transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the aggregation matrix.

[0227] In some embodiments, the aggregation module 5553 is further configured to determine the feature vector of the first aggregation layer by multiplying the network structure of the social group, the social group vector, and the latent vector transformation matrix in the first aggregation layer; to perform a nonlinear transformation on the feature vector of the first aggregation layer to obtain the intermediate latent vector of the first aggregation layer; and to determine the social group vector of the first aggregation by multiplying the transpose of the aggregation matrix in the first aggregation layer and the intermediate latent vector of the first aggregation layer.

[0228] In some embodiments, the aggregation module 5553 is further configured to determine the feature vector of the last aggregation layer by multiplying the network structure output by the previous aggregation layer, the social group vector output by the previous aggregation layer, and the latent vector transformation matrix of the last aggregation layer; perform a nonlinear transformation on the feature vector of the last aggregation layer to obtain the intermediate latent vector of the last aggregation layer, and determine the intermediate latent vector of the last aggregation layer as the group feature of the social group.

[0229] In some embodiments, the construction module 5551 is further configured to perform the following processing for any two members in the social group: when the number of interactions between the two members in the social group is greater than an interaction threshold, establish an edge between the corresponding nodes of the two members, and determine the number of interactions between the two members in the social group as the weight corresponding to the edge; determine the edge between the corresponding nodes of the two members and the weight corresponding to the edge as the connection information between the two members; and construct the network structure of the social group based on the connection information between any two members in the social group.

[0230] In some embodiments, the apparatus further includes: a training module 5556, configured to perform cascaded aggregation processing by combining the network structure of social group samples and the group vectors of the social group samples to obtain group features of the social group samples; perform group classification processing based on the group features of the social group samples to obtain the predicted group category of the social group; construct a loss function for the social group prediction model according to the predicted group category of the social group and the group category label of the social group; update the parameters of the social group prediction model until the loss function converges, and use the updated parameters of the social group prediction model when the loss function converges as the parameters of the trained social group prediction model.

[0231] This application provides a computer program product or computer program that includes 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 executes the computer instructions, causing the computer device to perform the artificial intelligence-based classification method described above in this application.

[0232] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the artificial intelligence-based classification method provided in this application. For example... Figure 3-6 The example shown is a classification method based on artificial intelligence.

[0233] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0234] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0235] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0236] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0237] It is understood that in the embodiments of this application, user data related to shared text and shared content are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is obtained, and the collection, use and processing of related data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0238] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A group classification method based on artificial intelligence, characterized in that, include: Based on the interactive behavior of each member in the social group, construct the network structure of the social group; Determine the member vector corresponding to each member in the social network, and combine the member vectors corresponding to each member to determine the group vector of the social group; The weight vector of the first aggregation layer is determined by multiplying the network structure of the social group, the group vector, and the aggregation weight matrix in the first aggregation layer through the first aggregation layer in a series of cascaded aggregation layers. A nonlinear transformation is performed on the weight vector of the first aggregation layer to obtain the aggregation matrix in the first aggregation layer; the transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the product of the aggregation matrix are determined as the network structure of the first aggregation. A second aggregation process is performed on the network structure of the social group and the group vector to obtain the first aggregated group vector; The network structure of the first aggregation and the group vector of the first aggregation are output as the aggregation result of the first aggregation layer to the subsequent cascaded aggregation layers, so as to continue the aggregation processing and output of the aggregation result in the subsequent cascaded aggregation layers until the last aggregation layer is output, and the output of the last aggregation layer is determined as the group feature of the social group. Based on the group characteristics of the social groups, group classification processing is performed to obtain the group categories of the social groups.

2. The method according to claim 1, characterized in that, Before determining the member vector corresponding to each member in the social network, the method further includes: The network structure of the social network is constructed based on the interactive behavior of each member in the social network. Perform the following processing for any member in the social network: Starting from the member, the network structure of the social network is sampled to obtain the sampled trajectory of the member; The sampling trajectory of the member is processed by word vector encoding to obtain the node representation vector of the member; Based on the node representation vector of the member, determine the member vector of the member in the social network.

3. The method according to claim 2, characterized in that, Before determining the member vector of the member in the social network, the method further includes: The shared text of the member on the social network is encoded to obtain the property representation vector of the member; The target words are extracted from the text shared by the member on the social network to obtain the number of times the target words appear, and the number of times the target words appear is determined as the sensitive representation vector of the member; Determining the member vector of a member in the social network based on the member's node representation vector includes: The node representation vector, the property representation vector, and the sensitivity representation vector of the member are fused to obtain the member vector.

4. The method according to claim 2, characterized in that, The process of constructing the network structure of the social network based on the interactive behavior of each member within the social network includes: Perform the following processing on any two members in the social network: When the interaction weight between the two members is greater than the interaction threshold, an edge is established between the corresponding nodes of the two members, and The interaction weight between the two members is determined as the weight corresponding to the edge; The edge between the nodes corresponding to the two members and the weight corresponding to the edge are used as the connection information between the two members; The network structure of the social network is constructed based on the connection information between any two members in the social network.

5. The method according to claim 4, characterized in that, The interactive behaviors include communication behaviors and transaction behaviors; Before establishing the edge between the corresponding nodes of the two members, the method further includes: Determine the communication weights corresponding to the communication behaviors of the two members in the social network; Determine the transaction weights corresponding to the transaction behaviors of the two members in the social network; The interaction weight between the two members is determined based on the communication weight and the transaction weight.

6. The method according to claim 5, characterized in that, Determining the interaction weight between the two members based on the communication weight and the transaction weight includes: The weight of the transaction is determined by multiplying the amount corresponding to the transaction behavior by the transaction weight. The sum of the weight of the transaction behavior and the weight of the communication is determined as the interaction weight between the two members.

7. The method according to claim 2, characterized in that, The step of sampling the network structure of the social network to obtain the sampled trajectory of the member includes: In the network structure of the social network, members adjacent to the member are randomly sampled to obtain the sampling information corresponding to the first sampled member; Using the first sampled member as a new starting point, continue random sampling until a preset number of sampled information are obtained, and The identification information of the member and the preset number of sampling information are concatenated to obtain the sampling trajectory of the member; The sampling information corresponding to the i-th sampled member includes at least one of the following data: the identifier of the i-th sampled member; the interaction weight between the i-th sampled member and the (i-1)-th sampled member, where 0 < i ≤ N, i is a natural number, and N is the preset number.

8. The method according to claim 1, characterized in that, The process of continuing aggregation processing and outputting aggregation results in the subsequent cascaded aggregation layers includes: The following processing is performed through the j-th aggregation layer in the plurality of cascaded aggregation layers: Perform the first aggregation process on the (j-1)th aggregation network structure to obtain the jth aggregation network structure; The network structure of the (j-1)th aggregation and the group vector of the (j-1)th aggregation are subjected to a second aggregation process to obtain the group vector of the j-th aggregation; The network structure of the j-th aggregation and the group vector of the j-th aggregation are determined as the aggregation result of the j-th aggregation layer; The aggregation result of the j-th aggregation layer is output to the (j+1)-th aggregation layer; Where 2≤j<J, j is a natural number, and J is the total number of the multiple cascaded aggregation layers.

9. The method according to claim 1, characterized in that, The second aggregation process, which involves combining the network structure of the social group and the group vector to obtain the first aggregated group vector, includes: The product of the network structure of the social group, the group vector, and the latent vector transformation matrix in the first aggregation layer is determined as the feature vector of the first aggregation layer. A nonlinear transformation is performed on the feature vector of the first aggregation layer to obtain the intermediate hidden vector of the first aggregation layer; The product of the transpose of the aggregation matrix in the first aggregation layer and the intermediate implicit vector of the first aggregation layer is determined as the group vector of the first aggregation.

10. The method according to claim 9, characterized in that, Determining the output of the last aggregation layer as the group feature of the social group includes: The product of the network structure output by the previous aggregation layer, the group vector output by the previous aggregation layer, and the latent vector transformation matrix of the last aggregation layer is determined as the feature vector of the last aggregation layer. A nonlinear transformation is performed on the feature vector of the last aggregation layer to obtain the intermediate hidden vector of the last aggregation layer. The intermediate latent vector of the last aggregation layer is determined as the group feature of the social group.

11. A classification device based on artificial intelligence, characterized in that, The device includes: The building module is used to construct the network structure of the social group based on the interaction behavior of each member in the social group; The determining module is used to determine the member vector corresponding to each member in the social network, and combine the member vectors corresponding to each member to determine the group vector of the social group. An aggregation module is used to determine the weight vector of the first aggregation layer by multiplying the network structure of the social group, the group vector, and the aggregation weight matrix in the first aggregation layer among multiple cascaded aggregation layers; perform a nonlinear transformation on the weight vector of the first aggregation layer to obtain the aggregation matrix in the first aggregation layer; determine the network structure of the first aggregation by multiplying the transpose of the aggregation matrix in the first aggregation layer, the network structure of the social group, and the aggregation matrix; perform a second aggregation process on the network structure of the social group and the group vector to obtain the group vector of the first aggregation; output the network structure of the first aggregation and the group vector of the first aggregation as the aggregation result of the first aggregation layer to subsequent cascaded aggregation layers, so as to continue the aggregation process and output the aggregation result in the subsequent cascaded aggregation layers until the output is reached to the last aggregation layer, and determine the group feature of the social group by the output of the last aggregation layer. The classification module is used to perform group classification processing based on the group characteristics of the social group to obtain the group category of the social group.

12. The apparatus according to claim 11, characterized in that, The AI-based classification device further includes a processing module for: The network structure of the social network is constructed based on the interactive behavior of each member in the social network. Perform the following processing for any member in the social network: Starting from the member, the network structure of the social network is sampled to obtain the sampled trajectory of the member; The sampling trajectory of the member is processed by word vector encoding to obtain the node representation vector of the member; Based on the node representation vector of the member, determine the member vector of the member in the social network.

13. The apparatus according to claim 11, characterized in that, The processing module is further configured to: The shared text of the member on the social network is encoded to obtain the property representation vector of the member; The target words are extracted from the text shared by the member on the social network to obtain the number of times the target words appear, and the number of times the target words appear is determined as the sensitive representation vector of the member; Determining the member vector of a member in the social network based on the member's node representation vector includes: The node representation vector, the property representation vector, and the sensitivity representation vector of the member are fused to obtain the member vector.

14. The apparatus according to claim 11, characterized in that, The processing module is further configured to: Perform the following processing on any two members in the social network: When the interaction weight between the two members is greater than the interaction threshold, an edge is established between the corresponding nodes of the two members, and The interaction weight between the two members is determined as the weight corresponding to the edge; The edge between the nodes corresponding to the two members and the weight corresponding to the edge are used as the connection information between the two members; The network structure of the social network is constructed based on the connection information between any two members in the social network.

15. The apparatus according to claim 11, characterized in that, The processing module is further configured to: Determine the communication weights corresponding to the communication behaviors of the two members in the social network; Determine the transaction weights corresponding to the transaction behaviors of the two members in the social network; The interaction weight between the two members is determined based on the communication weight and the transaction weight.

16. The apparatus according to claim 11, characterized in that, The processing module is further configured to: The weight of the transaction is determined by multiplying the amount corresponding to the transaction behavior by the transaction weight. The sum of the weight of the transaction behavior and the weight of the communication is determined as the interaction weight between the two members.

17. The apparatus according to claim 11, characterized in that, The processing module is further configured to: In the network structure of the social network, members adjacent to the member are randomly sampled to obtain the sampling information corresponding to the first sampled member; Using the first sampled member as a new starting point, continue random sampling until a preset number of sampled information are obtained, and The identification information of the member and the preset number of sampling information are concatenated to obtain the sampling trajectory of the member; The sampling information corresponding to the i-th sampled member includes at least one of the following data: the identifier of the i-th sampled member; the interaction weight between the i-th sampled member and the (i-1)-th sampled member, where 0 < i ≤ N, i is a natural number, and N is the preset number.

18. The apparatus according to claim 11, characterized in that, The aggregation module is also used for: The product of the network structure of the social group, the group vector, and the latent vector transformation matrix in the first aggregation layer is determined as the feature vector of the first aggregation layer. A nonlinear transformation is performed on the feature vector of the first aggregation layer to obtain the intermediate hidden vector of the first aggregation layer; The product of the transpose of the aggregation matrix in the first aggregation layer and the intermediate implicit vector of the first aggregation layer is determined as the group vector of the first aggregation.

19. The apparatus according to claim 11, characterized in that, The aggregation module is also used for: The product of the network structure output by the previous aggregation layer, the group vector output by the previous aggregation layer, and the latent vector transformation matrix of the last aggregation layer is determined as the feature vector of the last aggregation layer. A nonlinear transformation is performed on the feature vector of the last aggregation layer to obtain the intermediate hidden vector of the last aggregation layer. The intermediate latent vector of the last aggregation layer is determined as the group feature of the social group.

20. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the artificial intelligence-based classification method according to any one of claims 1 to 10.

21. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the artificial intelligence-based classification method according to any one of claims 1 to 10.

22. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the artificial intelligence-based classification method according to any one of claims 1 to 10.

Images