A guided topic detection method based on dynamic group characteristics

Abstract

The application belongs to the field of social network analysis, and particularly relates to a guiding type topic detection method based on dynamic group characteristics, which comprises extracting relevant attributes from topic data, the relevant attributes including a topic propagation network, user basic attributes, user interaction degree, internal driving factors and external driving factors; a Louvain implicit community mining method based on user interaction degree and user attribute similarity is used to obtain a user group set of the topic at each moment; an IG2vec representation learning algorithm is used to obtain user feature representation at each moment from a topic outbreak stage; a sentiment mutual influence force model based on game theory is used to quantify the sentiment mutual influence force, and the sentiment of the user is corrected through the sentiment mutual influence force to obtain a sentiment feature representation of the topic group; a GRU network is used to capture the change of the topic group feature after fusion at each moment, and a full connection layer and a Softmax function are used to obtain a prediction result; the application can effectively identify guiding type topics.

Description

Technical Field

[0001] This invention belongs to the field of social network analysis, specifically relating to a guided topic detection method based on dynamic group characteristics. Background Technology

[0002] The detection of leading topics has become particularly important, as it involves how to accurately identify and track key topics with the potential to influence public opinion.

[0003] In recent years, a large number of scholars have conducted research in the field of social networks and achieved considerable success. Previous studies on rumor detection have focused on observing the textual characteristics of users during the spread of leading topics and the changing trends of these characteristics over time. However, they did not consider the impact of changes in group structure and sentiment characteristics within the topic on the detection of leading topics. Therefore, researching changes in group characteristics within the leading topic space can lead to more efficient detection of leading topics.

[0004] With the continuous expansion of social network data and the gradual maturation of deep learning technology, nonlinear models such as neural networks and deep learning have gradually become the mainstream choice for analyzing complex networks. Choi et al. (CHOI J,KO T,CHOI Y, et al. Dynamic graph convolutional networks with attention mechanism for rumor detection on social media[J]. PLOS ONE, 2021, 16(8):e0256039.) proposed a dynamic graph convolutional network model DyGCN based on attention mechanism for rumor detection, which effectively combines the content and propagation structure of rumors to extract features, and also captures the dynamic information of rumor propagation. This paper illustrates that the content features and propagation structure of topics can be used to effectively detect leading topics. Inspired by this paper and the group segmentation algorithm, this invention proposes a leading topic detection method based on group features, which combines the emotional features and structural features of the group, and uses a GRU network to capture the changes in group features during the topic evolution process to detect leading topics.

[0005] In summary, the existing technology has the following problems:

[0006] Question 1: The anonymity of guiding topic groups. Due to the uniqueness of users' interests and preferences, as well as the influence of various factors such as users' implicit preferences and emotions, guiding topic groups generally possess anonymity.

[0007] Question 2: The dynamic nature of guiding topic group characteristics. As the user group changes constantly during the evolution of the topic, the characteristics of guiding topic group are generally dynamic.

[0008] Question 3: Quantifying the Group Interest in Leading Topics. Understanding the emotional inclinations of the topic participants is crucial for detecting leading topics. Therefore, quantifying fine-grained user emotional inclinations to obtain the group's interest at different times is particularly important for using GRU networks to detect leading topics. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention provides a guided topic detection method based on dynamic group characteristics, specifically including the following steps:

[0010] S1: Utilize publicly available dataset websites or APIs provided by social networks to comprehensively collect topic data, including user comments, detailed user information, and interactions between users.

[0011] S2: Extract relevant attributes from topic data, including topic propagation network, basic user attributes, user interaction, internal driving factors, and external driving factors;

[0012] S3: The Louvain implicit community mining method based on user interaction degree and user attribute similarity is adopted to update the weighted adjacency matrix of the topic propagation network and obtain the user group set of the topic at each time point;

[0013] S4: Use the IG2vec representation learning algorithm to represent user features at each time step starting from the topic outbreak stage;

[0014] S5: Quantify the mutual influence of emotions based on the emotional mutual influence model obtained from game theory, correct users' emotions through the emotional mutual influence, mine users' true emotions, and then obtain the emotional characteristics representation of the topic group based on the results of group segmentation;

[0015] S6: The changes in the characteristics of the topic group after fusion at each time step are captured by the GRU network. After passing through a fully connected layer, the output is converted into the probability values ​​of two categories using the Softmax function. The category corresponding to the maximum probability value of the two categories is taken as the prediction result.

[0016] During periods of widespread discussion, the number of users online surges. Efficiently segmenting user groups to detect trending topics becomes a pressing issue. This invention offers at least the following advantages:

[0017] 1. This invention proposes a Louvain latent community mining method based on user interaction degree and user attribute similarity. By updating the weighted adjacency matrix, this method can more comprehensively evaluate the connection strength and attribute similarity between nodes, thereby achieving more accurate group segmentation and significantly improving the efficiency and accuracy of topic detection.

[0018] 2. During the propagation of topics, complex and ever-changing relationships exist between entities such as users, messages, and networks, posing challenges to feature extraction and simplification. To effectively address this challenge, this invention innovatively proposes an iterative topic representation method (IT2vec) based on predecessor topic representation. This method fully considers the connections and differences between the current topic and predecessor topics, and vectorizes the feature space of derived topics into low-rank vectors, greatly enriching the feature representation of topics and providing strong support for topic detection and analysis.

[0019] 3. Furthermore, to address the issue of quantifying group interests, this invention introduces game theory and a multiple linear regression algorithm to construct a group interest matrix, which is then integrated into topic features. This innovative method not only achieves accurate quantification of group interests but also enhances the depth and breadth of topic analysis. Simultaneously, considering the time-sensitive nature of topic dissemination cycles, this invention performs time-slicing processing on topic data and further proposes a guided topic model, providing strong technical support for real-time topic monitoring and early warning. Attached Figure Description

[0020] Figure 1 This is a flowchart of the social network-guided topic detection method of the present invention;

[0021] Figure 2 This is a schematic diagram of group division according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the iterative representation of topic features in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the mechanism for quantifying user emotional interaction in this invention;

[0024] Figure 5 This is a schematic diagram of the GRU neural network of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides a guided topic detection method based on dynamic group characteristics, specifically including the following steps:

[0027] S1: Utilize publicly available dataset websites or APIs provided by social networks to comprehensively collect topic data, including user comments, detailed user information, and interactions between users.

[0028] S2: Extract relevant attributes from topic data, including topic propagation network, basic user attributes, user interaction, internal driving factors, and external driving factors;

[0029] S3: The Louvain implicit community mining method based on user interaction degree and user attribute similarity is adopted to update the weighted adjacency matrix of the topic propagation network and obtain the user group set of the topic at each time point;

[0030] S4: Use the IG2vec representation learning algorithm to represent user features at each time step starting from the topic outbreak stage;

[0031] S5: Quantify the mutual influence of emotions based on the emotional mutual influence model obtained from game theory, correct users' emotions through the emotional mutual influence, mine users' true emotions, and then obtain the emotional characteristics representation of the topic group based on the results of group segmentation;

[0032] S6: The changes in the characteristics of the topic group after fusion at each time step are captured by the GRU network. After passing through a fully connected layer, the output is converted into the probability values ​​of two categories using the Softmax function. The category corresponding to the maximum probability value of the two categories is taken as the prediction result.

[0033] In this embodiment, the process involves three aspects: group segmentation, feature extraction, and detection model establishment, specifically including the following steps:

[0034] First, the topic is divided into groups at each time point. This involves obtaining topic data from public datasets, including basic information, network relationship information, and behavioral information (likes, follows, and reposts) of participating users. The adjacency matrix of the Louvain algorithm is then optimized based on user interaction and user attribute similarity to obtain the group set at each time point of the topic.

[0035] Secondly, group feature extraction is performed. That is, using the group set at each time step, the IG2vec representation learning algorithm is proposed to iteratively represent the topic data and vectorize the topic group feature space into a low-rank vector. Then, game theory and multiple linear regression algorithm are introduced to represent the group interest features.

[0036] Finally, the model building part obtains a new group feature matrix by incorporating the group interest features at each time step into the group feature matrix, and then inputs it into the GRU network to detect whether the current topic is a guiding topic.

[0037] This embodiment provides a specific implementation process, such as... Figure 1 The present invention provides a guided topic detection method based on dynamic group characteristics, which specifically includes the following steps:

[0038] S1: Utilize publicly available dataset websites or APIs provided by social networks to comprehensively collect topic data. This data includes user comments, detailed user information, and interactions between users.

[0039] Specifically, raw topic data can be obtained from the APIs provided by the official websites of public social media platforms. What needs to be obtained here are all comments, basic information and historical behavior information of commenters, for both guided and non-guided topics throughout their lifecycle. The basic information of commenters includes user age, user gender, number of followers, and number of fans. User historical behavior information includes the user's historical reputation and the posts the user has published in the past month.

[0040] S2: After collecting the original topic data, it is necessary to perform detailed attribute extraction. These attributes cover multiple dimensions such as topic propagation network, basic user attributes, user interaction, internal driving factors, and external driving factors. Based on these attributes, this invention constructs the original topic propagation network, providing an intuitive network structure reference for subsequent user group segmentation.

[0041] S3: The Louvain implicit community mining method, based on user interaction and user attribute similarity, is used to obtain the user group set of each topic at each time.

[0042] S4: The IG2vec representation learning algorithm is used to perform iterative representations at each time step starting from the topic outbreak stage, so as to obtain the low-rank vectorized feature representation of the topic group.

[0043] This invention employs the IG2vec representation learning method to process topic propagation networks. First, each user node in the network is mapped to a feature vector. Simultaneously, an update function is used to iteratively update the user feature matrix to obtain the user feature matrix. Then, user influence is used to weight and fuse user node features, so that the features of different nodes have different weights in the group feature representation, and finally, the group feature matrix is ​​obtained.

[0044] S5: Quantify the mutual influence of emotions based on the emotional mutual influence model obtained from game theory, correct users' emotions through the emotional mutual influence, mine users' true emotions, and then obtain the emotional characteristics representation of the topic group based on the results of group segmentation;

[0045] S6: The changes in the characteristics of the topic group after fusion at each time step are captured by the GRU network. After passing through a fully connected layer, the output is converted into the probability values ​​of two categories using the Softmax function. The category corresponding to the maximum probability value of the two categories is taken as the prediction result.

[0046] like Figure 2 As shown, a dynamically weighted adjacency matrix is ​​constructed based on user interaction degree and user attribute similarity. The dynamic Louvain algorithm is used to obtain the user group set under each topic at each time step. The specific process is as follows: first, an initial community division is performed, treating each node as an independent community. Then, iterative optimization is performed. In each iteration, the modularity gain brought by moving a node to an adjacent community is evaluated, and operations that increase modularity are executed until the network structure no longer changes. The specific steps include:

[0047] Treating each user as a community, we initialize the modularity as follows:

[0048]

[0049] Where Q represents modularity; A ij This represents the adjacency relationship between the i-th user node and the j-th user node in the adjacency matrix formed by users when a topic breaks out. In this embodiment, if there are interactive behaviors such as forwarding, commenting, and liking between the i-th user node and the j-th user node, then A... ij The value of A is 1, otherwise A ij The value is 0; d i d represents the degree of the i-th user node. j δ(c) represents the degree of the j-th user node; i ,c j δ(c) indicates whether the i-th node user and the j-th node user belong to the same community. If they belong to the same community, then δ(c) represents the number of nodes in the community. i ,c j ) = 1, otherwise δ(c) i ,c j ) = 0;

[0050] Perform iterative updates, moving a user to their neighboring community, updating the modularity after the move, and saving the strategy that increases the modularity until the modularity no longer increases. Output the community at this point. The updated modularity after the move is represented as:

[0051]

[0052] Among them, W T (u i ,u j ) represents the weight between user i and user j in the weighted adjacency matrix of the topic propagation network after the topic erupts at time T.

[0053] In dynamic social networks, the network structure changes at each time point. To capture these changes, this invention introduces a weighted adjacency matrix W. T (u i ,u j The calculation formula is as follows:

[0054]

[0055] Where T represents the time elapsed since the topic broke out, i.e., the current time is time T since the topic broke out; U T This represents the entire set of users who participated in the discussion after the topic broke out at time T.

[0056] The cosine similarity of the related attributes between user i and user j in the feature space is expressed as follows:

[0057]

[0058] Among them, UserProp(u j UserProp(u) represents the feature vector obtained by mapping the relevant attributes of user i to the feature space. j ) represents the feature vector obtained by mapping the relevant attributes of user j to the feature space; ∥·∥ represents taking the norm of the vector;

[0059] UserInteract(u i ,u j T) represents the degree of interaction between user i and user j during the evolution of the topic, and the calculation process is as follows:

[0060]

[0061] Among them, l i,jThis represents the relationship level between the i-th and j-th user nodes, i.e., the minimum number of edges required to connect the i-th and j-th user nodes. For example, if the i-th and j-th user nodes follow each other, the relationship level is 1, meaning they are connected by one edge. Alternatively, if the i-th and j-th user nodes require at least k nodes to connect, the relationship level is k+1, meaning there are at least k+1 edges between them. M represents the total number of original messages posted by users i and j in the topic up to time t. ma This represents the m-th message posted by the i-th user node and the j-th user node, and the time when the action a is performed. When a=1, it means the action of commenting; when a=2, it means the action of forwarding; and when a=3, it means the action of liking.

[0062] like Figure 3 The iterative representation of the node embedding vectors at each time step of the topic is combined with a function to fuse them into a user feature matrix at each time step. Specifically, this includes the following steps:

[0063] 101. Based on the topic propagation network at time t User set U t And the interaction between users E U Calculate the weighted adjacency matrix of the topic propagation network at the current time. In this embodiment, the topic propagation network... Represented as:

[0064] 102. Generate a node sequence by performing a random walk based on the weighted adjacency matrix, optimize the node sequence using the Skip-Gram model, and generate an embedding vector for each node; during the random walk, to more accurately describe the interactions between users, use the edge weight W used in the group partitioning. T (u i ,u j The user walks through the system, and this weighting considers not only the historical interactions between users but also the correlations between user attributes. The probability of walking from user i to user j is expressed as:

[0065]

[0066] Wherein, P(u i ||u j α represents the probability that the i-th user node will move to the j-th user node; p,q (u i ,u j W(u) represents the wandering bias term; i ,u j) represents the weight between user i and user j in the weighted adjacency matrix of the current topic propagation network; z represents the normalization coefficient; p represents the return parameter, controlling the probability of revisiting a previously visited node; q represents the intrinsic and extrinsic parameters, controlling whether the traversal direction is inward or outward search; d ui,uj This represents the distance between the i-th user node and the j-th user node.

[0067] 103. Generate a node sequence through random walk and optimize the node embedding representation using the Skip-Gram model. The Skip-Gram model generates an embedding vector for each node based on the input node sequence.

[0068] 104. At each time point t, construct the feature matrix for that time point using the embedding vectors. To capture the evolution of the topic across different time points, use the softmax function and the scoring function. To update the embedding vector at the current time step, that is:

[0069]

[0070] in, Let represent the embedding vector of the i-th user at time t+1; d is the dimension of the embedding vector.

[0071] This represents the embedding vector of the i-th user at time t+1. The scoring function.

[0072] The feature matrix at each time point t is composed of the updated embedding vectors at time point t. Each row of the feature matrix is ​​the embedding vector of a node, and the feature matrix includes all nodes in a community. Based on the group segmentation results, the feature representation of the topic group is obtained, i.e.:

[0073]

[0074] Where Marix(G,t) represents the feature representation of the topic group consisting of all user nodes in the topic propagation network G at time t, i.e., the group feature matrix; Weight(u i UserInf(u) represents the normalized value of the user influence of the i-th user node in the topic group composed of all user nodes in the topic propagation network G at time t; i W represents the user influence of the i-th user node in the topic group composed of all user nodes in the topic propagation network G at time t, where n is the number of user nodes in the topic propagation network G at time t; i Let W be the weight of the historical reputation score of the i-th user node in the topic group composed of all user nodes in the topic propagation network G at time t. i∈[0,1]; Let t be the number of followers of the i-th user node in the topic group composed of all user nodes in the topic propagation network G at time t; Let t be the number of users followed by the i-th user node in the topic group composed of all user nodes in the topic propagation network G at time t.

[0075] like Figure 4 User emotional influence function Emo(u i It consists of two aspects: internal driving factors and external influencing factors. First, the internal driving factors, Emo, are constructed from the textual features of user comments. in (e i ), constructing external influencing factors Emo from the emotional tendencies of users' friend groups. out (u i Internal driving factors Emo in (e i External influencing factors Emo out (u i They are represented as follows:

[0076]

[0077] in, This represents the internal driving factors, p, used to construct sentiment tendency from the textual features of user comments. This represents the external driving factor, p, which is constructed from the textual features of user friends' comments, indicating a sentiment tendency. t represents the time elapsed since the topic originated. emoSum(e i emoNum(e) represents the total number of sentiment-related words in all comments from the i-th user node during the time t from the topic's outbreak to the current moment. i emoSum(u,p) represents the number of comments with sentiment p for the i-th user node. i emoNum(u) represents the total number of sentiment-related words in all comments from the friends of the i-th user node. i ,p) represents the number of comments with sentiment tendency p among the friends of the i-th user node.

[0078] Considering both internal and external user factors, a sentiment influence function is constructed using a multiple linear regression algorithm, as follows:

[0079]

[0080] Among them, Inf pos (u i,t Inf represents the positive sentiment influence function of the i-th user node at time t; neg (u i,tInf represents the negative sentiment influence function of the i-th user node at time t; none (u i,t ) represents the unemotional influence function of the i-th user node at time t; ρ0, ρ1, and ρ2 are the partial regression coefficients of the multiple linear regression algorithm. The partial regression coefficients of the positive, negative, and unemotional influence functions are updated according to the multiple linear regression algorithm. The specific values ​​can be modified by those skilled in the art based on historical data, and will not be described in detail in this embodiment. p represents the internal driving factor of user i's sentiment tendency based on the text features of user comments at time t. When p=1, the sentiment tendency is positive; when p=2, the sentiment tendency is negative; and when p=3, the sentiment tendency is no sentiment. Let p be the external driving factor of user i's sentiment tendency based on the text features of user friends' comments at time t.

[0081] According to game theory, three game strategies are defined: "positive emotion," "negative emotion," and "no emotion bias." The payoff functions of these three strategies are respectively derived from Inf. pos (u i,t ),Inf pos (u i,t ) and Inf pos (u i,t The calculation process is as follows:

[0082]

[0083] Among them, Mut pos (u i,t Mut represents the influence of positive sentiment on the i-th user node at time t; neg (u i,t Mut represents the influence of negative sentiment on the i-th user node at time t; none (u i,t ) represents the influence of no emotion on the i-th user node at time t; ω1 and ω2 are two adjustable parameters used to represent the different degrees of influence of positive and negative emotions on no emotion tendency.

[0084] The sentiment feature matrices of users within the same community are aggregated together, that is, each user's sentiment feature vector is treated as a row vector, and the sentiment feature vectors of all users constitute the group sentiment feature matrix (i.e., ... Figure 4 The user interest matrix is ​​used to normalize the group emotion feature matrix, and then the group feature matrix is ​​concatenated with the group feature matrix to obtain the group feature matrix that integrates group emotions.

[0085] like Figure 5The group feature matrix, which incorporates group sentiment, is input into the GRU neural network to capture changes in group features at various time points. The Softmax function is then used to convert the GRU output into probability values ​​for two categories, predicting whether the current topic is a guiding topic. This can be expressed as:

[0086]

[0087] Where P(n,r|topic) represents the probability that the current topic is a rumor topic r or a non-rumor topic n; This is a group feature matrix that integrates group emotions; gru_model(·) represents the GRU neural network.

[0088] The category corresponding to the maximum probability value between the two categories is taken as the prediction result, that is:

[0089]

[0090] in, This indicates that the current topic has been determined to be a non-leading topic, while This indicates that the current topic has been identified as a leading topic.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A guided topic detection method based on dynamic group characteristics, characterized in that, Specifically, the following steps are included: S1: Utilize publicly available dataset websites or APIs provided by social networks to comprehensively collect topic data, including user comments, detailed user information, and interactions between users. S2: Extract relevant attributes from topic data, including topic propagation network, basic user attributes, user interaction, internal driving factors, and external driving factors; S3: The Louvain implicit community mining method based on user interaction degree and user attribute similarity is adopted to update the weighted adjacency matrix of the topic propagation network and obtain the user group set of the topic at each time point; S4: Use the IG2vec representation learning algorithm to represent user features at each time step starting from the topic outbreak stage; S5: Quantify the mutual influence of emotions based on the emotional mutual influence model obtained from game theory, correct users' emotions through the emotional mutual influence, explore users' true emotions, and then obtain the emotional characteristics representation of the topic group based on the results of group segmentation. S6: The changes in the characteristics of the topic group after fusion at each time step are captured by the GRU network. After passing through a fully connected layer, the output is converted into the probability values ​​of two categories using the Softmax function. The category corresponding to the maximum probability value of the two categories is taken as the prediction result.

2. The guided topic detection method based on dynamic group characteristics according to claim 1, characterized in that, The process of updating the weighted adjacency matrix of the topic propagation network includes the following steps: Among them, W T (u i ,u j ) represents the weight between user i and user j in the weighted adjacency matrix of the topic propagation network after the topic outbreak at time T; UserInteract(u) represents the cosine similarity of related attributes between user i and user j in the feature space. i ,u j (T) represents the degree of interaction between user i and user j during the evolution of the topic; T represents the time elapsed since the topic erupted; U T This represents the entire set of users in the current topic.

3. The guided topic detection method based on dynamic group characteristics according to claim 2, characterized in that, The process of obtaining the user group set for each topic at each time point using the Louvain implicit community mining method based on user interaction and user attribute similarity includes: Treating each user as a community, we initialize the modularity as follows: Perform iterative updates, moving a user to their neighboring community, updating the modularity after the move, and saving the strategy that increases the modularity until the modularity no longer increases. Output the community at this point. The updated modularity after the move is represented as: Where Q represents modularity; A ij This represents the adjacency relationship between the i-th user node and the j-th user node in the adjacency matrix formed by users when a topic breaks out; d i d represents the degree of the i-th user node. j The degree of the j-th user node is represented by δ(c); m represents the number of edges in the network; δ(c i ,c j δ(c) indicates whether the i-th node user and the j-th node user belong to the same community. If they belong to the same community, then δ(c) represents the number of nodes in the community. i ,c j ) = 1, otherwise δ(c) i ,c j ) = 0.

4. The guided topic detection method based on dynamic group characteristics according to claim 2, characterized in that, During the evolution of the topic, the degree of interaction between user i and user j at time t is UserInteract(u i ,u j ,t) is represented as: Among them, l i,j M represents the number of layers in the relationship between the i-th and j-th user nodes; M represents the total number of original messages posted by user i and user j in the topic up to time t; t ma This represents the m-th message posted by the i-th user node and the j-th user node, and the time when the action a is performed. When a=1, it means the action of commenting; when a=2, it means the action of forwarding; and when a=3, it means the action of liking.

5. The guided topic detection method based on dynamic group characteristics according to claim 1, characterized in that, The process of obtaining the low-rank vectorized feature representation of the topic group at time t includes the following steps: Based on the topic propagation network at time t User set U t And the interaction between users E U Calculate the weighted adjacency matrix of the topic propagation network at the current time. A node sequence is generated by performing a random walk based on the weighted adjacency matrix, and the node sequence is optimized using the Skip-Gram model to generate an embedding vector for each node. Based on the user's embedding vector at the current and previous time steps, update the current user's embedding vector using the scoring function and the softmax function; The embedding vectors of all user nodes in each community at each time point are obtained to form a group representation vector. The group representation vectors are combined to obtain the group feature matrix at that time point.

6. The guided topic detection method based on dynamic group characteristics according to claim 5, characterized in that, During a random walk, the probability of moving from user i to user j is expressed as: Wherein, P(u i ||u j ) represents the probability that the i-th user node will move to the j-th user node; a p,q (u i ,u j W(u) represents the wandering bias term; i ,u j ) represents the weight between user i and user j in the weighted adjacency matrix of the current topic propagation network; z represents the normalization coefficient; p represents the return parameter, which controls the probability of revisiting the previously visited node; q represents the intrinsic and extrinsic parameters, which control whether the traversal direction is inward or outward. This represents the distance between the i-th user node and the j-th user node.

7. The guided topic detection method based on dynamic group characteristics according to claim 5, characterized in that, The vector representation of each user node is updated based on the current user embedding vector and the user embedding vector at the previous time step, using a scoring function and a softmax function. The update process includes: in, Let represent the embedding vector of the i-th user at time t+1; d is the dimension of the embedding vector. This represents the embedding vector of the i-th user at time t+1. The scoring function.

8. The guided topic detection method based on dynamic group characteristics according to claim 1, characterized in that, User emotions are quantified using an emotional mutual influence model derived from game theory. The quantified positive emotional influence, negative emotional influence, and neutral emotional influence are respectively represented as follows: Among them, Mut pos (u i,t Mut represents the influence of positive sentiment on the i-th user node at time t; neg (u i,t Mut represents the influence of negative sentiment on the i-th user node at time t; none (u i,t Inf represents the influence of no emotion on the i-th user node at time t; pos (u i,t Inf represents the positive sentiment influence function of the i-th user node at time t; neg (u i,t Inf represents the negative sentiment influence function of the i-th user node at time t; none (u i,t ) represents the unemotional influence function of the i-th user node at time t; ω1 and ω2 are two adjustable parameters used to represent the different degrees of influence of positive and negative emotions on unemotional inclination.

9. The guided topic detection method based on dynamic group characteristics according to claim 8, characterized in that, The positive, negative, and unemotional influence functions of user i at time t are respectively expressed as: Where ρ0, ρ1, and ρ2 are the partial regression coefficients of the multiple linear regression algorithm, and the positive, negative, and no-emotional influence functions update the partial regression coefficients according to the multiple linear regression algorithm. p represents the internal driving factor of user i's sentiment tendency based on the text features of user comments at time t. When p=1, the sentiment tendency is positive; when p=2, the sentiment tendency is negative; and when p=3, the sentiment tendency is no sentiment. Let p be the external driving factor of user i's sentiment tendency based on the text features of user friends' comments at time t.

10. The guided topic detection method based on dynamic group characteristics according to claim 5, characterized in that, At time t, the internal and external driving factors of user i's sentiment tendency based on the text features of user comments are represented as follows: in, This represents the internal driving factors, p, used to construct sentiment tendency from the textual features of user comments. This represents the external driving factor, p, which is constructed from the textual features of user friends' comments, where t represents the time elapsed since the topic broke out. emoSum(e i emoNum(e) represents the total number of sentiment-related words in all comments from the i-th user node during the time t from the topic's outbreak to the current moment. i emoSum(u,p) represents the number of comments with sentiment p for the i-th user node. i emoNum(u) represents the total number of sentiment-related words in all comments from the friends of the i-th user node. i ,p) represents the number of comments with sentiment tendency p among the friends of the i-th user node.

Images