Modeling method of network community class group and user representation model
By employing a collaborative modeling approach at the community and user levels, and utilizing heterogeneous graphs and self-supervised learning, this approach addresses the issue of low efficiency in existing network community group modeling technologies, enabling efficient information representation and learning across various downstream tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2023-08-25
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies, when studying network community groups, neglect the complex relationships between different objects, resulting in low modeling efficiency and high dependence on data labels, making them difficult to apply effectively in various downstream tasks.
We adopt a collaborative modeling approach at the community and user levels. Through heterogeneous graph construction, structural encoding, and self-supervised learning, we train the model to interact with information between different levels and use self-supervised tasks and hierarchical iterative strategies for pre-training.
It achieves more efficient information representation in various downstream tasks, reduces dependence on data labels, and improves the model's learning ability at the community and user levels.
Smart Images
Figure CN117312489B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computers, and in particular to a graph pre-training method for unified modeling of network community groups and users. Background Technology
[0002] Online community groups play an increasingly important role in modern society. In order to study the group characteristics and representation information of different online community groups, researchers have built various models with the online community group as a whole or individual users as the research object, and designed corresponding downstream tasks to evaluate the models. They have all achieved excellent results in their respective fields.
[0003] Current research methods often model a single task object and discuss the model's performance in downstream tasks related to that object, which ignores the complex relationships that exist between different objects in a network community group. Summary of the Invention
[0004] The purpose of this specification is to provide a graph pre-training method for unified modeling of network community groups and users. The collaborative modeling of the community and user levels provided in this specification can offer a unified framework applicable to various downstream task objectives. Based on this collaborative modeling method, this specification provides a hierarchical iterative pre-training approach, introducing multiple self-supervised learning tasks at different stages. This enables the model to train a collaborative understanding of the user and community levels and to better learn the representations of both levels.
[0005] To achieve the above objectives, this specification provides a collaborative modeling method for community-level and user-level hierarchies. The method includes: priori classifying network community groups and extracting semantic information of different network community groups, as well as structural information between different network community groups; extracting semantic information of different users and structural information between different users within a network community group; extracting structural information between network community groups and users, and integrating all semantic and structural information to construct a heterogeneous social network that includes community-level and user-level hierarchies.
[0006] In one implementation, the steps of dividing and extracting information from different levels include: for each network community category, selecting text features within the community as the semantic information of the community; for different network community categories, selecting interaction methods between network community categories as structural information between communities; for each user, selecting text features posted by the user as the semantic information of the user; for different users, selecting social interaction methods between users as structural information between users; and for different network community categories and users, using the user's affiliation status with different network community categories as structural information between the community level and the user level.
[0007] This specification also provides a pre-training method for a model, comprising: a heterogeneous graph construction unit, which constructs a heterogeneous graph of a social network based on existing semantic and structural information in the data, serving as the input structure for representing the model pre-training; a structure encoding unit, which encodes and outputs the relationships between nodes in the graph through a node attribute encoder and an edge generation encoder; a self-supervised learning unit, which performs generative learning on occluded nodes and edges in the graph through two types of self-supervised tasks: node attribute reconstruction and edge reconstruction; and a hierarchical iteration unit, which performs generative learning on the community level and the user level respectively through iterative training on different node types.
[0008] In one embodiment, the method includes: a heterogeneous graph construction unit that constructs a social network heterogeneous graph containing two node types and three edge types using semantic and structural information at the community and user levels.
[0009] In one embodiment, the method includes: a structure encoding unit and a self-supervised learning unit, which respectively mask network community cluster nodes and user nodes on a heterogeneous graph, and use the generated output of the masked position through a node attribute encoder as the conceptual representation of the masked node; calculate the similarity between the generated encoding and the original masked features to guide the model to represent the semantic information on the nodes; mask various relationships formed by community nodes and user nodes on the heterogeneous graph, predict the probability that there is an edge between unconnected pairs of nodes through an edge generation encoder, and optimize the model to represent the structural information on the heterogeneous graph through negative contrastive estimation.
[0010] In one embodiment, the method includes: a hierarchical iterative unit, wherein at the user level, user nodes are sampled as target nodes, and a corresponding user-related self-supervised task is used for training in the self-supervised learning unit; at the community level, community nodes are sampled as target nodes, and a corresponding community-related self-supervised task is used for training in the self-supervised learning unit; after training the current stage for a certain number of rounds, the current model state is saved, and the target nodes of the next stage are used for training, and this operation is iterated throughout the training process; each iteration always starts from the user level and ends at the community level.
[0011] As can be seen from the technical solutions provided in the embodiments of this specification above, the purpose of these embodiments is to provide a graph pre-training method for unified modeling of network community groups and users. The method for collaborative modeling at the community and user levels provided in these embodiments can effectively address downstream tasks with various research objectives; the model pre-training method provided in these embodiments can promote information interaction between the community and user levels, while simultaneously training the model to better learn the representations of both levels. Experimental results also demonstrate that the model, training method, and apparatus provided in this specification can effectively learn various information representations of network community groups and users. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0013] Figure 1 This is an illustrative diagram illustrating a framework of different community modeling methods provided in this manual;
[0014] Figure 2 This is a schematic diagram of constructing a heterogeneous graph structure for a social network, provided in this specification.
[0015] Figure 3 This is a schematic diagram of a pre-training framework for unified modeling of network community groups and users provided in this specification;
[0016] Figure 4 This is a schematic diagram of an ablation experiment result provided in this instruction manual. Detailed Implementation
[0017] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments in this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.
[0018] This specification provides a representation model for online community groups and users, applicable to the collaborative modeling of online community groups and users. For illustrative purposes, please refer to [link to relevant documentation]. Figure 1 . Figure 1 The left side presents a widely adopted paradigm for modeling online community groups, which involves designing a specific model for the downstream task being applied, given community or user data. This approach performs well in individual downstream task tests, but it requires designing multiple models to handle different data and tasks, which is resource-intensive and inefficient, and also highly dependent on data labels.
[0019] In this embodiment, refer to Figure 1 On the right, community data and user data will be pre-integrated as composite data input, and a unified framework model of network community groups and users will be built on this basis, which can be applied to a variety of different downstream task scenarios.
[0020] See Figure 2 The method described in this embodiment may include the following steps:
[0021] Step S1.1: Divide the network community groups: Align the community data and user data, filter users belonging to the target network community group through the network community group index, and then truncate the tail communities according to the amount of user data corresponding to each network community group (for example, remove network community groups that do not meet the requirements), thereby realizing two-way filtering of network community groups.
[0022] Step S1.2: Construct node types; Nodes at the community level are represented by network community groups. In particular, different network community groups have different representation methods, but they can all be divided into multiple secondary network community groups or sub-communities according to the rules of this implementation method. Nodes at the user level are represented by individual users. An individual user is the sum of all community activities and data information under the same community account.
[0023] Step S1.3: Construct node relationships; For relationships between network community groups, hyperlinks between communities are used as the construction rule. When information published in community A includes or mentions community B, a hyperlink is created between community A and community B, and an edge relationship is constructed. For relationships between individual users, social interactions of users are used as the construction rule. In particular, social interactions have various different manifestations in different network community groups, but they can all be converted into comments by users on content published by other users according to the rules of this implementation method. When user A and user B are commented on by the same user, an edge relationship is constructed between user A and user B. For relationships between network community groups and users, user activity characteristics are used as the construction rule. In particular, user activity characteristics have various different manifestations in different network community groups, but they can all be converted into user affiliation relationships with different sub-communities according to the rules of this implementation method. When user A belongs to community B, an edge relationship is constructed between user A and community B.
[0024] Step S1.4: Assign semantic information to nodes; For community-level nodes, the semantic information is represented by extracting the text of the hyperlink edges. All original texts containing the hyperlink are statistically analyzed using dictionary information such as LIWC to serve as the macro-level semantic information of the network community group, and the mean is taken to one dimension as the initial embedding feature; For user-level nodes, the semantic information is represented by extracting the title text of the content published by the user in the community. After encoding the title text using the BERT model, the mean of multiple encodings from the same user is taken to one dimension as the initial embedding feature.
[0025] In this embodiment, graph G is defined as follows:
[0026] G = (V, E)
[0027] Where V represents the set of all nodes, i.e., all individual users and communities, and E represents the set of all edges. To intuitively describe the information of nodes and edges in the graph, for graph G, there are corresponding adjacency matrix A and feature matrix F:
[0028]
[0029]
[0030] The dimension of the adjacency matrix A is the number of nodes in the graph. The element A(u,v) in the u-th row and v-th column represents the edge relationship between nodes u and v. In the undirected graph designed by this method, A(u,v) = A(v,u) = 0, indicating that there is no edge between nodes u and v; conversely, A(u,v) = A(v,u) = 1, indicating that there is an edge between nodes u and v.
[0031] The feature matrix can be viewed as the concatenation of the node feature matrix (first term) and the edge feature matrix (second term). N represents the index of all nodes and edges with features, and L represents the dimension of the stored feature attribute vector. Therefore, F(u,∶) represents the feature information of the u-th node or edge. This information is usually encoded into a low-dimensional vector space using machine learning methods as the initial features of the graph. Generally, the features of nodes are only related to the text sequence and do not have semantic information with structural features.
[0032]
[0033] In heterogeneous social networks, the feature matrix F is composed of two parts of node features. The node features Fv can be further decomposed into user features Fu and community features Fc.
[0034] In contrast, existing social networks create homogeneous graphs between users or heterogeneous graphs with users as the main focus. The nodes may be users and tweets, but they are not constructed using users and communities.
[0035] Clearly, in this implementation, the structural characteristics of users and communities are explicitly presented through the relationships between edges between nodes in the graph, while semantic information is described as node features through the graph's feature matrix. By integrating all semantic and structural information, a heterogeneous social network comprising community and user levels is constructed.
[0036] The representation model provided in this implementation enables the alignment of community and user data from different sources, thereby achieving joint modeling of the community and user levels and providing a unified framework for information interaction between these two levels. Experiments show that the method provided in this specification can effectively model the community and user levels and learn more generalized community and user representations.
[0037] This specification also provides a pre-training method for a community user representation model, applied to self-supervised training of the representation model, the method comprising the following steps:
[0038] Step S2.1: Based on the current target node type, sample a subgraph on the input heterogeneous graph, and randomly mask the target node type and mask the edges related to the target node type.
[0039] Step S2.2: Generate the code of the occluded node by reconstructing the encoder through node attributes, generate the possible edges of the occluded node by reconstructing the encoder through edge, and calculate the similarity and confidence respectively, and perform comparative learning optimization.
[0040] Step S2.3: Change the current target node type, repeat steps S2.1 and S2.2, and iterate the training.
[0041] Please see Figure 3 In this implementation, by introducing a self-supervised pre-training task and a hierarchical iterative strategy, the model can generate encoded outputs of nodes and edges simultaneously during a single training run, enabling simultaneous representation learning at both the community and user levels.
[0042] In this embodiment, the self-supervised pre-training method includes: a node attribute reconstruction task, whereby the attribute X of the occluded node i is... i The mean attribute of all occluded nodes is used instead, and the encoded output h of the node is calculated by the node attribute encoder. i Attr Calculate the similarity between it and the original attribute that was covered. The similarity calculation and optimization objective is:
[0043]
[0044] Community nodes and user nodes will respectively mask the nodes during their respective training phases and use their respective mean attributes to replace the original node attributes.
[0045] In this embodiment, the self-supervised pre-training method includes: an edge reconstruction task, in which the edges related to the target node are occluded, and the node i is encoded by an edge reconstruction encoder to obtain the output h. i Edge The existence of an edge is predicted by the encoded output of a set of nodes using an edge decoder. Let j be the node associated with the occluded edge. + Let j be all nodes that are not connected to node i. For nodes i and j... + Calculate the edge confidence and use negative contrast estimation as the optimization objective:
[0046]
[0047] In this embodiment, the training objective of the representation model of network community groups and users during the training process is:
[0048]
[0049] In this embodiment, the pre-training method includes a hierarchical iterative training strategy. Network community groups have an inherent hierarchical structure, from text to users, and finally to communities. This is reflected in the relationship between community and user levels in the heterogeneous graph of a social network. Text serves as the initial embedding to represent the semantic information of both levels. Therefore, the specific steps of the hierarchical iterative training are as follows: In the user level stage, user nodes are sampled as target nodes, and corresponding user-related self-supervised tasks are used for training in self-supervised learning, namely, the attribute reconstruction task of user nodes, the edge reconstruction task between users, and the edge reconstruction task between communities and users; In the community level stage, community nodes are sampled as target nodes, and corresponding community-related self-supervised tasks are used for training in self-supervised learning, namely, the attribute reconstruction task of community nodes, the edge reconstruction task between communities, and the edge reconstruction task between communities and users; After training the current stage for a certain number of rounds, the current model state is saved, and training is then performed using the target nodes of the next level, iterating this operation throughout the entire training process; Each iteration always starts from the user level and ends at the community level.
[0050] This specification also provides a model training device applied to network community groups and user representation models, as described in the embodiments above. Since the principle of a model training device in solving a problem is similar to that of a model training method, the implementation of a model training device can refer to the implementation of a model training method, and repeated details will not be elaborated further. The terms "unit" or "module" used below can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated. Specifically, the device may include: a heterogeneous graph construction unit, which constructs a social network heterogeneous graph based on existing semantic and structural information in the data, serving as the input structure for pre-training the representation model; a structure encoding unit, which encodes and outputs the relationships between nodes in the graph through a node attribute encoder and an edge generation encoder; a self-supervised learning unit, which performs generative learning on occluded nodes and edges in the graph through two types of self-supervised tasks: node attribute reconstruction and edge reconstruction; and a hierarchical iteration unit, which performs generative learning on the community level and user level respectively through iterative training on different node types.
[0051] The implementation method of this specification also provides two downstream task evaluation methods, which include: a user community affiliation task with users as the inference object, and a community aggression detection task with communities as the inference object.
[0052] In one implementation, for a community affiliation task that uses users as the inference object, the method steps include: using the user's affiliated community as a label, a downstream task model predicts a list of communities to which the user belongs based on the input user coding features. This task can measure the pre-trained learning ability of user attributes and the edge relationship between communities and users.
[0053] In one implementation, for a community aggression detection task that uses communities as the inference object, the method steps include: using sentiment tendencies in hyperlinks as labels, a downstream task model predicts the aggression label classification of a community based on the input community encoding features. This task can measure the pre-training's ability to learn community attributes and the relationship between communities and community edges.
[0054] During the experiment, the dataset was sampled and divided into a pre-training set, a training set, a validation set, and a test set. Detailed data can be found in Table 1.
[0055] Table 1: Experimental dataset partitioning.
[0056]
[0057] In the experiment, for implementation details, the initial embedding features of community nodes and user nodes were aligned to the same dimension before training to enable information transfer between the two levels. During pre-training, the heterogeneous graph Transformer model was used as the base model encoder. The community level and the user level were trained for 10 rounds each in one iteration. The entire pre-training process went through 6 iterations, for a total of 120 rounds.
[0058] During the experiment, for downstream task testing, only 30% of the full dataset was used for testing. A prediction head was added to the model's output to map the output representation to the label space. All downstream tasks were trained for 50 epochs.
[0059] Table 2 reflects the overall performance of the experiment. For the pre-training phase, the overall optimization objective was a weighted sum of node attribute reconstruction and edge reconstruction tasks, reflecting the model's learning ability for heterogeneous input graphs. Therefore, a lower loss function value indicates that the model can better represent the original semantic information and structural features of the graph. During the pre-training phase, the implementation method (HIPT) in this specification achieved the lowest loss function value in the user target node stage (Stage-1).
[0060] In the downstream task testing phase, the method described in this specification outperformed other model methods in the user-community affiliation prediction task at the user target node stage. Specifically, the evaluation metrics NDCG and MRR used in this task are relatively low in absolute terms. This is due to the large number of sub-communities, resulting in long-tailed labeling in this downstream task; however, conclusions can still be drawn from relative comparisons between different model methods.
[0061] In the downstream task testing phase, specifically the community attack detection task, the method described in this specification outperformed other model methods in the community target node stage (Stage-2). This indicates that the community target node stage effectively aggregated relevant information from user nodes, guiding the updating of its own features.
[0062] Table 2: Overall performance of different models. Bold numbers indicate the best performance for each metric.
[0063]
[0064] In this scenario example, to explore the relationship between the number of iterations in hierarchical iteration and the number of word iteration training rounds, a hyperparameter experiment was conducted in conjunction with a user-community affiliation downstream task test. With a fixed total training round of 120, in one iteration, the two stages were trained for 1, 5, 10, 20, and 30 rounds respectively, corresponding to a total of 120, 24, 12, 6, and 4 iterations. As shown in Table 3, with each iteration training for 10 rounds and 12 iterations, the method described herein achieved the best performance in the user-community affiliation downstream task test.
[0065] Table 3: Hierarchical Iterative Setting Experiment. Bold numbers indicate the best performance for each metric.
[0066]
[0067] In this example scenario, to explore the relationship between community nodes and user nodes, an ablation experiment was conducted, coupled with a user-community affiliation downstream task test. In the experiment, the edge reconstruction task between communities and users was removed from the pre-training, and the model is denoted as HIPT cut. Please refer to [link to relevant documentation]. Figure 4In downstream task testing, the pre-trained model that removed the community-user edge to reconstruct the model performed worse than the optimal model in this embodiment and also worse than the model with randomly initialized downstream parameters. Experimental results indicate that the community-user edge plays a crucial role in the unified community and user modeling representation model of this embodiment, serving as a bridge between the structural features and information interaction of the community and user hierarchies. The HIPT cut model, lacking mutual information, overemphasized community or user hierarchical attributes during pre-training, thus being misled overall and resulting in poor test performance.
[0068] The devices or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware, or a module that implements the same function can be implemented by a combination of multiple sub-modules, etc. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed.
[0069] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.
[0070] This application can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.
[0071] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.
[0072] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. This application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.
[0073] Although this application has been described by way of examples, those skilled in the art will know that this application has many variations and modifications without departing from the spirit of this application, and it is intended that the appended claims cover such variations and modifications without departing from this application.
Claims
1. A modeling method for representational models of online community groups and users, characterized in that: The method for collaborative modeling of community and user levels includes: The online community categories are filtered by users belonging to those categories; Extract semantic information from each network community category and the semantic information of each individual user within a network community category; extract structural information between different network community categories, between different individual users, and between network community categories and users, including: Node types are constructed, where nodes at the community level are represented by network community groups, and nodes at the user level are represented by individual users; Node relationships are constructed, wherein hyperlinks between online community groups are used as the construction rule for relationships between online community groups, and social interactions of users are used as the construction rule for relationships between individuals; and user activity characteristics are used as the construction rule for relationships between online community groups and individual users. Nodes are endowed with semantic information. For nodes at the community level, the semantic information is represented by extracting the text of the hyperlink edges. All original texts containing the hyperlink are statistically analyzed using dictionary information such as LIWC to serve as macro-level semantic information for the network community group. For nodes at the user level, the semantic information is represented by extracting the title text of the content published by the user in the community. The title text is encoded using the BERT model, and the average of multiple encodings from the same user is calculated to one dimension as the initial embedding feature. It also extracts semantic information from different online community groups, as well as structural information between different online community groups; Semantic information of different users and structural information between different users are extracted from users within the online community category. Extracting structural information between online community groups and users; By integrating all semantic and structural information, a heterogeneous social network with community and user levels is constructed.
2. The method as described in claim 1, characterized in that: The steps for dividing and extracting information from different levels include: For each network community category, the textual features within the community are selected as the semantic information of the community; for different network community categories, the interaction methods between the network community categories are selected as the structural information between the communities. For each individual user, the text features published by the user are selected as the user's semantic information; for different users, the social interaction methods between users are selected as the structural information between users. For different online community groups and users, the user's affiliation status with different online community groups is used as the structural information between the community level and the user level.
3. A network community category and user representation model, characterized in that, It is constructed using the modeling method described in any one of claims 1 or 2.
4. A pre-training method for the network community group and user representation model as described in claim 3, characterized in that: The method, applied to self-supervised training of the representation model, includes: The network community cluster nodes and user nodes on the heterogeneous graph are masked respectively. The generated output of the masked position is used as the conceptual representation of the masked node by the node attribute encoder. The similarity between the generated code and the original masked features is calculated to guide the model to represent the semantic information on the node. The algorithm masks various relationships between community nodes and user nodes in a heterogeneous graph, predicts the probability of an edge between unconnected node pairs using an edge generation encoder, and optimizes the guiding model to represent the structural information on the heterogeneous graph using negative contrast estimation.
5. A training device for the network community group and user representation model as described in claim 4, characterized in that: The device, applied to network community category and user representation models, includes: Heterogeneous graph construction unit constructs a heterogeneous graph of the social network based on the existing semantic and structural information in the data, which serves as the input structure for pre-training the representation model; The structural encoding unit encodes and outputs the relationships between nodes in the graph through a node attribute encoder and an edge generation encoder. The self-supervised learning unit performs generative learning on occluded nodes and edges in the graph through two types of self-supervised tasks: node attribute reconstruction and edge reconstruction. The hierarchical iterative unit enables the model to perform generative learning at the community level and the user level through iterative training on different node types.
6. The model training apparatus as described in claim 5, characterized in that: The execution steps of a hierarchical iterative unit include: At the user level, sampling is performed with user nodes as target nodes, and user-related self-supervised tasks are used for training in the self-supervised learning unit. At the community level, sampling is performed with community nodes as target nodes, and community-related self-supervised tasks are used for training in the self-supervised learning unit. After training the current stage for a certain number of rounds, the current model state is saved, and the target node of the next stage is used for training. This operation is iterated throughout the training process. Each iteration always starts from the user level and ends at the community level.
7. A task evaluation method for a network community group and user representation model as described in claim 3, characterized in that, The method includes: a user community affiliation task with users as the inference object, and a community aggression detection task with communities as the inference object.
8. The task evaluation method as described in claim 7, characterized in that: The step "For the community affiliation task with users as the inference object" includes predicting a list of communities to which the user belongs based on the input user coding features.
9. The task evaluation method as described in claim 7, characterized in that: The step "For the community aggression detection task with the community as the inference object" includes predicting the aggression label classification of the community based on the input community coding features.