A Method and System for Detecting Abnormal Users in Chat Scenarios Based on Graph Neural Networks

Abstract

This invention relates to a method and system for detecting abnormal users in chat scenarios based on graph neural networks. The method includes: processing different group chat messages into semantic graphs and temporal graphs respectively; training a graph neural network model and a classifier model in labeled group chat messages, wherein the graph neural network model learns user node representations in the semantic graph and temporal graph; training a graph neural network model in unlabeled group chat messages using adversarial generation, and aligning it with the graph neural network model trained in labeled group chat messages; classifying user nodes using the trained classifier model based on the user node representations learned by the graph neural network model, thereby detecting abnormal users in all group chat messages. This invention can reflect the relationships between different users from both temporal and semantic perspectives, and can flexibly combine content and structural features to capture the anomalies of graph nodes, achieving more accurate identification of abnormal users.

Description

Technical Field

[0001] This invention belongs to the field of information technology, and in particular relates to a method and system for detecting abnormal users in chat scenarios based on graph neural networks. Background Technology

[0002] In recent years, the number and proportion of cybercrime cases have been increasing year by year. Chat tools have become the most frequently used criminal tools in cyber fraud crimes. More than half of the cyber fraud cases involve criminal elements committed through chat tools such as WeChat. Therefore, it is of great significance to discover and monitor abnormal users in a timely manner based on chat scenarios.

[0003] While considerable research has been conducted on anomaly user detection, most methods preprocess user information such as IP addresses and friend relationships as features, then use clustering or manually specified rules to determine whether a user is suspicious. However, current fraud cases involve criminals exploiting chat groups to collaboratively defraud victims, with multiple malicious users potentially existing within the group and frequently supporting each other. Furthermore, the pattern of victimization across different group chats may be similar, typically involving initial contact, suspicion, small gains, and significant losses. The fraudsters' tactics may be similar at each stage. Therefore, how to quickly identify victims and fraudsters by analyzing group chat logs is a crucial area of ​​focus for addressing this new type of crime.

[0004] Graph neural networks (GNNs) learn high-dimensional representations of nodes by simultaneously capturing content features and structural relationships. GNNs achieve this by encoding content and structural features into functions and optimizing these functions under supervised or unsupervised training signals. By analyzing chat logs in group chats and constructing fine-grained graphs in both temporal and semantic dimensions, GNNs can more accurately reflect the identities of different users within the group. Summary of the Invention

[0005] This invention proposes a method for detecting abnormal users in chat scenarios based on graph neural networks, which can more accurately identify abnormal users.

[0006] The principle of this invention is as follows: Chat logs from different group chats are modeled as semantic graphs for capturing user opinions and temporal graphs for capturing user interactions. Then, the abnormal user detection problem is reconstructed as a node classification problem. Guided by partially labeled group chat information, a graph neural network model is used to flexibly combine content and structural features to learn user representations and train a general classifier for classification. Furthermore, to ensure that the classifier trained on labeled group chat information can be applied to unlabeled scenarios, adversarial generative learning is used. The generator is responsible for generating user representations in unlabeled group chat information, while the discriminator is responsible for determining whether a sample comes from a labeled group chat. Through the game between the generator and the discriminator, the graph neural network model is aligned in high-dimensional space under different data conditions. Finally, the trained classifier is used to classify all group chat users.

[0007] The method for detecting abnormal users in chat scenarios based on graph neural networks of the present invention includes the following steps:

[0008] 1) Process the given different group chat information into semantic graphs and temporal graphs respectively;

[0009] 2) Train a graph neural network model and a classifier model in labeled group chat messages, where the graph neural network model learns the user node representations in the semantic graph and temporal graph of step 1);

[0010] 3) Train a graph neural network model in unlabeled group chat messages using adversarial generative methods, and align it with the graph neural network model trained in step 2).

[0011] 4) Based on the user node representation learned by the graph neural network model obtained in step 3), the user nodes are classified using the classifier model trained in step 2, thereby completing the detection of abnormal users in all group chat information.

[0012] Furthermore, the graph neural network model and classifier model are trained in the labeled group chat information. The relational graph convolutional neural network model RGCN is used to learn the node representations of semantic graphs and temporal graphs with heterogeneous edge types. The l-th layer of RGCN calculates the node representation H by aggregating messages from k types of edges. l :

[0013]

[0014] Among them, E i =A [：，：，i] The adjacency matrix represents the edges of type i. This represents the adjacency matrix with added self-loops, where I represents the adjacency matrix with respect to E. i Diagonal matrices of the same size; This represents the degree matrix composed of edges of type i. This represents the edge connection between node u and node v with respect to type i; W i l It is a trainable parameter matrix of type i, where l represents the l-th layer in the network, Agg represents the aggregation function, and H represents the initial hidden node. 0 Given the original node feature matrix X, after L layers of message propagation, the hidden representation HL of the last layer is used as the node representation H, where H[u] represents the representation of node u; for and Calculate the corresponding and For all user nodes, the node representations of the two graphs are concatenated to form the final node representation. in

[0015] Furthermore, the graph neural network model and classifier model are trained in the labeled group chat information, and a multilayer perceptron (MLP) model is selected as the classifier model to determine whether a node belongs to a fraudster.

[0016] Furthermore, the step of training a graph neural network model in unlabeled group chat information using adversarial generative methods and aligning it with a graph neural network model trained in labeled group chat information involves using the method of training an adversarial generative network (GAN) model to align RGCN models trained on different datasets.

[0017] Furthermore, the adversarial generative network includes a generator. and discriminator Responsible for providing samples of unlabeled group chat user node representations. Used to distinguish Does the given sample belong to a pre-trained, labeled group chat user? Finally, Unable to distinguish Does the given node belong to a different group chat? By aligning different group chat representations, MLP can ultimately determine the probability that an unlabeled user node represents a fraudster.

[0018] A system for detecting abnormal users in chat scenarios based on graph neural networks, employing the above method, includes:

[0019] Semantic graph and temporal graph construction modules are used to process different group chat information into semantic graphs and temporal graphs respectively;

[0020] The model training module is used to train a graph neural network model and a classifier model in labeled group chat messages, where the graph neural network model learns user node representations in semantic graphs and temporal graphs; and to train a graph neural network model in unlabeled group chat messages using adversarial generative methods, and align it with the graph neural network model trained in labeled group chat messages.

[0021] The abnormal user detection module is used to classify user nodes based on the user node representation learned by the graph neural network model, and thus complete the detection of abnormal users in all group chat information.

[0022] Compared with existing methods, the beneficial effects of the present invention are as follows:

[0023] 1) A novel method for modeling group chat information as a graph is proposed, which can reflect the relationships between different users in terms of temporal sequence and semantics;

[0024] 2) The problem of abnormal user detection is reconstructed into a node classification problem. The graph neural network model is used to flexibly combine content and structural features to capture the abnormality of graph nodes, so as to more accurately identify abnormal users.

[0025] 3) By utilizing adversarial generation, the model can classify all group chat users with minimal annotation, reducing the manual workload. Furthermore, it can be easily and quickly extended to new group chat information, resulting in high efficiency. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the training process of the overall model of this invention. Detailed Implementation

[0027] To make the objectives and technical solutions of this invention clearer, the implementation of this invention will be described in detail below with reference to the accompanying drawings.

[0028] This invention proposes a method for detecting abnormal users in chat scenarios based on graph neural networks. Figure 1 This describes the training process for the overall model.

[0029] Suppose there are N group chat messages, of which N train Each group has a label, meaning each user in the group has a specific identity (assuming only two categories: victims and fraudsters), while the remaining N... test =NN train Each group has no tags. Each group has its corresponding group chat information S. i Each group chat message S i This can be abstracted as a length of Len i sequence < id j textj timestamp j >, where 1≤i≤N and 1≤j≤Len i id is in the current group S i The text is a unique identifier for each user, such as their username, the text is the comment posted by that user, and the timestamp is the timestamp of the comment.

[0030] Next, according to S i Construct semantic graphs for each of the following (1≤i≤N) groups. A timing diagram

[0031] For semantic graphs Let n be the adjacency matrix, representing the number of nodes in the graph. i Each node. All user IDs in the current group chat history. j (1≤j≤Len i The set Id i As a user node, its size is (The same user may post multiple times). All posts in the group are displayed in text format. j (1≤j≤Len i Using existing viewpoint extraction tools, it is further subdivided into segments of length Len. i <id j , aspect j opinion j > Yes, where aspect j For this text j The subject being described, opinion j For ID j This aspect j Emotions (positive, negative).

[0032] Furthermore, we obtain a size of Aspect, a collection of subjects (where different users may express opinions on the same subject). i As the main node and satisfying (User nodes and principal nodes constitute Go) i (All nodes). Furthermore, if This indicates that there exists a type k between nodes u and v. o (k o An edge ∈ {0, 1}, where 0 represents a positive direction and 1 represents a negative direction, signifies a user node u (1 ≤ u ≤ n). user For the main node v(n) user <v≤n) have expressed their opinions, otherwise Let X be the node content feature matrix, where X o [u] indicates that the size of node u is d. o eigenvectors.

[0033] For timing diagrams This is an adjacency matrix, representing the number of nodes in the graph. Each node is composed of the user IDs of all users in the current group chat. j The size formed by (1≤j≤Leni) is set Id i composition. This indicates that there is a line of type between nodes u and v. The edge, meaning within a time period At that time, both user u and user v made comments. Let X be the node content feature matrix, where X t [u] indicates that the size of node u is d. t eigenvectors.

[0034] When processing node content features, this invention first converts messages of different lengths from different nodes into fixed-length features, while also addressing the issue of similar meanings for different keywords. Let u be a node in the graph, and text(u) contain the statements published by u. For each statement m∈text(u), m contains multiple keywords. To construct the content features of u, existing word vector tools are used to calculate the representation of each keyword, and all representations are averaged to obtain the final content feature vector of node u.

[0035] Furthermore, the Relational Graph Convolutional Neural Network (RGCN) model from the graph neural network model family is used to learn node representations with heterogeneous edge types for semantic graphs and temporal graphs. The RGCN at layer l (1≤l≤L) can compute the node representation H by aggregating messages from k types of edges. l :

[0036]

[0037] Among them, E i =A [：，：，i] , represents the adjacency matrix of edges of type i. This represents the adjacency matrix with added self-loops, where I represents the adjacency matrix with respect to E. i Diagonal matrices of the same size; This represents the degree matrix composed of edges of type i. This represents the edge connection between node u and node v with respect to type i; W i l Let be the trainable parameter matrix for type i, l represent the l-th layer in the network, and Agg represent the aggregation function, which is average pooling chosen here. The initial hidden node representation is H.0 Let X be the original node feature matrix. After L layers of message propagation, the hidden representation H of the last layer is used. L H is a node representation, where H[u] represents the representation of node u.

[0038] Furthermore, for and Calculate the corresponding and For all user nodes, the node representations of the two graphs are concatenated to form the final node representation. in

[0039] Furthermore, a multilayer perceptron (MLP) model is chosen as the classifier to determine whether a node belongs to a fraudster. The model can be modeled as P i =MLP(H i ),in For the final representation of user nodes, The model output represents the probability that a user node belongs to the fraudster, P. i [u] represents the probability that user u in group chat message i is a fraudster.

[0040] Furthermore, in the labeled group chat dataset (1≤i≤N) train ), train 2*N using stochastic gradient descent. train One RGCN (one for semantic graph and one for temporal graph) is used to learn the final representation of user nodes, and one MLP is used to classify user nodes, employing cross-entropy loss. c As the loss function, uy represents the labeled result of u (0 represents the victim, 1 represents the fraudster):

[0041]

[0042] Furthermore, in order to make the classifier MLP trained on labeled group chat datasets applicable to other unlabeled group chat datasets (N... train (<i≤N), this invention uses the method of training a Generative Adversarial Network (GAN) model to align RGCN models trained on different datasets.

[0043] Generative Adversarial Networks (GANs) consist of two neural network models: a generator and a microcontroller. and discriminator Used to generate samples, so that they are processed It was later determined to be a real sample; Used to estimate the probability that the current sample belongs to a real sample. The meaning of adversarial generation is... We need to try our best to make Will The generated sample was identified as real, while It is necessary to try to The generated samples were identified as fake. The training is supervised learning. The training is unsupervised learning. In this invention, Responsible for providing samples of unlabeled group chat user node representations. Used to distinguish Determine whether the sample belongs to a pre-trained, labeled group chat user. Finally, Unable to distinguish Does the given node belong to a different group chat? Alignment of different group chat representations was achieved. Therefore, the MLP can ultimately determine the probability that an unlabeled user node representation belongs to a fraudster.

[0044] For discriminator Its goal is to provide a node representation H f [u] Determine whether it belongs to a user in a tagged group chat or The generated unlabeled group chat user representations. An MLP model is also used as the... This represents the probability that user u has a label.

[0045] For generators Its goal is to generate, given the sequence graph and semantic graph in an unlabeled group chat scenario, a... It is impossible to distinguish whether the node representation is trained on labeled group chat data. RGCN is also used as... The representation matrix of user nodes.

[0046] GAN will use the discriminator The score is used as a rating for the generator Rewards The generator is trained using stochastic gradient ascent. To maximize its reward. When training converges, By inferring the user node representations of unlabeled group chat data and feeding them into an MLP model trained on a labeled dataset, it is possible to detect anomalous users based on chat logs.

[0047] The specific process of the abnormal user detection method in chat scenarios based on graph neural networks of the present invention includes the following steps:

[0048] 1) Process group chat information into semantic graphs and sequence graphs.

[0049] 2) Train a graph neural network model and a classifier model in labeled group chat messages.

[0050] 3) Train a graph neural network model in unlabeled group chat messages using adversarial generative programming, and align it with the graph neural network model trained in step 2). Alignment means that for the same identity type, the model trained in this step and the graph neural network model trained in step 2) can obtain similar representations in low-dimensional space, allowing the classifier trained in labeled group chat messages to be applied to unlabeled group chat messages.

[0051] 4) Use the classifier model trained in step 2) to classify the node representations learned by the graph neural network model obtained in step 3), thereby completing the detection of abnormal users in all group chat information.

[0052] The calculation process for the first stage, i.e., step 1), is as follows:

[0053] a) Given a group chat message set S of size N, which is divided into groups of size N train Group chat messages with tags S train And size N test =NN train Untagged group chat messages S test Number of time segments k t ;

[0054] b) For each S i (1≤i≤N), perform the following operations respectively:

[0055] c) Traverse S i Calculate the user set ID i The viewpoint triples P processed using existing viewpoint extraction methods i And the viewpoint collection Aspect i ;

[0056] d) Calculate the number of nodes

[0057] e) Initialize the semantic graph adjacency matrix and feature matrix

[0058] f) Enumerate viewpoint triples P i (id j apsect j opinion j Construct a semantic graph adjacency matrix;

[0059]

[0060] g) Process the node feature matrix using existing word vectors and text classification tools (e.g., fastText).

[0061] h) Initialize the adjacency matrix and characteristic matrix of the sequence graph.

[0062] i) S i After sorting by timestamp, it is divided into k t Section, obtained

[0063] j) Traversal Construct a temporal graph adjacency matrix. For any users u and v, if they both exit...

[0064] Now In the middle, then A t [u][v][j]=A t [v][u][j] = 1;

[0065] k) Utilize existing word vectors and text classification tools (e.g., fastText) to process the node feature matrix X. t ;

[0066] l) Repeat steps b)-k) until all group chat messages have been processed.

[0067] The calculation process for the second stage, i.e., step 2), is as follows:

[0068] a) Given a group chat message S with tags train Processed timing diagram semantic graph Number of training cycles e mlp ;

[0069] b) Initialize the classifier model MLP and the graph neural network model. and (1≤i≤N train );

[0070] c) Iterate from 1 to i up to N train ;

[0071] d) Calculation and spliced ​​together

[0072] e) Calculation Calculate loss c Backpropagation updates parameters;

[0073] f) Repeat steps c)-e) until the required number of training cycles e is completed. mlp ;

[0074] g) Calculation

[0075] The calculation process for the third stage, i.e., step 3), is as follows:

[0076] a) Given a group chat message S with tags train And untagged group chat messages S test Processed timing diagram Japanese

[0077] Yitu The trained node representation H label Number of training cycles e gan Training batch size;

[0078] b) Initialization and (N train <i≤N) as generator MLP as a discriminator

[0079] c) From H label The sampled batch of nodes represents H. t ;

[0080] d) From Generating the same number of nodes represents H. f ;

[0081] e) Fixed The parameters are used to calculate the reward during training.

[0082] f) From The generation of batch nodes is represented as H. f ;

[0083] g) Fixed The parameters are used to calculate the reward during training.

[0084] h) Repeat steps c)-g) until the required number of training cycles e is completed. gan .

[0085] The calculation process for the fourth stage, or step 4), is as follows:

[0086] a) Given a processed timing diagram semantic graph Trained graph neural network and (1≤i≤N), a well-trained classification model MLP, and a fraud threshold T. fVictim threshold T v (T v <T f );

[0087] b) Initialize the fraudster set F = {} and the victim set V = {};

[0088] c) Calculation and spliced ​​together

[0089] d) Calculation

[0090] e) For If P i [u]≥T f Then F = F + {u}, if P i [u]≤T v Then V = V + {u};

[0091] f) Repeat c) to d) until i has been traversed from 1 to N;

[0092] g) Return the set of detected fraudsters F and the set of victims V.

[0093] This invention is not limited to the methods described in the above embodiments, for example:

[0094] 1. When preprocessing node content features, other preprocessing methods can be selected, such as BERT.

[0095] 2. When learning node information, other heterogeneous graph neural network models are used, such as RGAT.

[0096] 3. Abnormal user identities are not limited to victims and fraudsters. The classifier can be changed from single-class to multi-class based on the tag information without affecting other model structures.

[0097] Based on the same inventive concept, another embodiment of the present invention provides a chat scenario abnormal user detection system based on graph neural networks using the method of the present invention, comprising:

[0098] Semantic graph and temporal graph construction modules are used to process different group chat information into semantic graphs and temporal graphs respectively;

[0099] The model training module is used to train a graph neural network model and a classifier model in labeled group chat messages, where the graph neural network model learns user node representations in semantic graphs and temporal graphs; and to train a graph neural network model in unlabeled group chat messages using adversarial generative methods, and align it with the graph neural network model trained in labeled group chat messages.

[0100] The abnormal user detection module is used to classify user nodes based on the user node representation learned by the graph neural network model, and thus complete the detection of abnormal users in all group chat information.

[0101] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smartphone, etc.) including a memory and a processor, wherein the memory stores a computer program configured to be executed by the processor, and the computer program includes instructions for performing the steps of the method of the present invention.

[0102] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk), wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, it implements the various steps of the method of the present invention.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the principles and scope of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. A method for detecting abnormal users in chat scenarios based on graph neural networks, characterized in that, Includes the following steps: The given different group chat information is processed into semantic graphs and temporal graphs respectively; Train a graph neural network model and a classifier model in labeled group chat messages, where the graph neural network model learns user node representations in semantic graphs and temporal graphs; We use adversarial generative programming to train a graph neural network model in unlabeled group chat messages and align it with a graph neural network model trained in labeled group chat messages. Based on the user node representation learned by the graph neural network model, the trained classifier model is used to classify the user nodes, thereby completing the detection of abnormal users in all group chat information; The step of processing different group chat information into semantic graphs and temporal graphs respectively includes: Suppose there are N group chat messages, of which N train Each group has a label, and the remaining N test =NN train Each group has no tags; each group has its corresponding group chat information S. i Each group chat message S i Abstracted to a length of Len i sequence <id j ,text j timestamp j > where 1≤i≤N and 1≤j≤Len i id is in the current group S i The text is a unique identifier for the user, the text is the comment posted by the user, and the timestamp is the timestamp of the comment being posted. For semantic graphs Let n be the adjacency matrix, representing the number of nodes in the graph. i Each node; all user IDs in the current group chat history. j set Id i As a user node, its size is All messages in the group (text) j Using an opinion extraction tool, it was subdivided into segments of length Len. i of <id j ,aspect j opinion j Yes, where aspect j For this text j The subject being described, opinion j For ID j This aspect j The emotion; further obtained in terms of size Aspect i As the main node and satisfying if This indicates that there exists a type k between nodes u and v. o An edge is defined as an edge where a user node u has expressed an opinion on the principal node v; otherwise... Let X be the node content feature matrix. o [u] indicates that the size of node u is d. o eigenvectors; For timing diagrams This is an adjacency matrix, representing the number of nodes in the graph. Each node is composed of the user IDs of all users in the current group chat. j The size of the structure is set Id i composition; This indicates that there is a line of type between nodes u and v. The edge, meaning within a time period Both user u and user v made comments at that time; Let X be the node content feature matrix. t [u] indicates that the size of node u is d. t eigenvectors.

2. The method according to claim 1, characterized in that, The process involves training a graph neural network model and a classifier model on labeled group chat messages. A relational graph convolutional neural network (RGCN) model is used to learn node representations with heterogeneous edge types in semantic and temporal graphs. The l-th layer of the RGCN calculates the node representation H by aggregating messages from k edge types. l : Among them, E i =A [:,:,i] The adjacency matrix represents the edges of type i. This represents the adjacency matrix with added self-loops, where I represents the adjacency matrix with respect to E. i Diagonal matrices of the same size; This represents the degree matrix composed of edges of type i. This represents the edge connection between node u and node v with respect to type i; W i l It is a trainable parameter matrix of type i, where l represents the l-th layer in the network, Agg represents the aggregation function, and H represents the initial hidden node. 0 Given the original node feature matrix X, after L layers of message propagation, the hidden representation H of the last layer is used. L As a node representation H, where H[u] represents the representation of node u; for and Calculate the corresponding and For all user nodes, the node representations of the two graphs are concatenated to form the final node representation. in 3. The method according to claim 2, characterized in that, The process involves training a graph neural network model and a classifier model in labeled group chat information, selecting a multilayer perceptron (MLP) model as the classifier model to determine whether a node belongs to a fraudster.

4. The method according to claim 3, characterized in that, In a labeled group chat dataset, train a 2*N dataset using stochastic gradient descent. train One RGCN is used to learn the final representation of user nodes, and one MLP is used to classify user nodes, employing cross-entropy loss. c As a loss function: Among them, P i [u] represents the probability that user u in group chat message i is a fraudster; uy represents the labeling result of u, where 0 represents a victim and 1 represents a fraudster.

5. The method according to claim 1, characterized in that, The method of training a graph neural network model in unlabeled group chat information using adversarial generative methods and aligning it with a graph neural network model trained in labeled group chat information is to align RGCN models trained on different datasets by using the method of training an adversarial generative network model (GAN).

6. The method according to claim 5, characterized in that, The adversarial generative network includes a generator. and discriminator Responsible for providing samples of unlabeled group chat user node representations. Used to distinguish Does the given sample belong to a pre-trained, labeled group chat user? Finally, Unable to distinguish Does the given node belong to a different group chat? By aligning different group chat representations, MLP can ultimately determine the probability that an unlabeled user node represents a fraudster.

7. A system for detecting abnormal users in chat scenarios based on graph neural networks, characterized in that, include: Semantic graph and temporal graph construction modules are used to process different group chat information into semantic graphs and temporal graphs respectively; The model training module is used to train a graph neural network model and a classifier model in labeled group chat messages, where the graph neural network model learns user node representations in semantic graphs and temporal graphs; and to train a graph neural network model in unlabeled group chat messages using adversarial generative methods, and align it with the graph neural network model trained in labeled group chat messages. The abnormal user detection module is used to classify user nodes based on the user node representation learned by the graph neural network model and the trained classifier model, thereby completing the detection of abnormal users in all group chat information. The step of processing different group chat information into semantic graphs and temporal graphs respectively includes: Suppose there are N group chat messages, of which N train Each group has a label, and the remaining N test =NN train Each group has no tags; each group has its corresponding group chat information S. i Each group chat message S i Abstracted to a length of Len i sequence <id j ,text j timestamp j > where 1≤i≤N and 1≤j≤Len i id is in the current group S i The text is a unique identifier for the user, the text is the comment posted by the user, and the timestamp is the timestamp of the comment being posted. For semantic graphs Let n be the adjacency matrix, representing the number of nodes in the graph. i Each node; all user IDs in the current group chat history. j set Id i As a user node, its size is All messages in the group (text) j Using an opinion extraction tool, it was subdivided into segments of length Len. i of <id j ,aspect j opinion j Yes, where aspect j For this text j The subject being described, opinion j For ID j This aspect j The emotion; further obtained in terms of size Aspect i As the main node and satisfying if This indicates that there exists a type k between nodes u and v. o An edge is defined as an edge where a user node u has expressed an opinion on the principal node v; otherwise... Let X be the node content feature matrix. o [u] indicates that the size of node u is d. o eigenvectors; For timing diagrams This is an adjacency matrix, representing the number of nodes in the graph. Each node is composed of the user IDs of all users in the current group chat. j The size of the structure is set Id i composition; This indicates that there is a line of type between nodes u and v. The edge, meaning within a time period Both user u and user v made comments at that time; Let X be the node content feature matrix. t [u] indicates that the size of node u is d. t eigenvectors.

8. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method described in any one of claims 1 to 6.

Images