A group awareness discovery method for a guided topic

Abstract

The application belongs to the field of social network analysis, and particularly relates to a guiding type topic-oriented group consciousness discovery method; comprising: obtaining guiding type topic data and preprocessing the same to obtain user short texts and topic short texts; processing the two kinds of texts to obtain user text feature vectors and topic text feature vectors; fusing the two kinds of text feature vectors to obtain a topic text content feature matrix; obtaining a topic network and processing the same to obtain a topic topology structure feature matrix; fusing the topic text content feature matrix and the topic topology structure feature matrix to obtain a user node matrix; processing the guiding type topic data according to a cognitive game theory to obtain the influence of three kinds of forwarding behaviors of users, and optimizing a topic network adjacency matrix; calculating the aggregation features of user nodes according to the user node matrix and the topic network adjacency matrix; inputting the aggregation features into a classifier for processing to obtain the forwarding behavior prediction results of all users; the application has accurate prediction results and improves the group consciousness clarity.

Description

Technical Field

[0001] This invention belongs to the field of social network analysis, specifically relating to a method for discovering group consciousness in guiding topics. Background Technology

[0002] In today's information-saturated society, numerous trending topics emerge daily, becoming a crucial tool for those with ulterior motives to achieve their goals. The spread of these topics often relies on the guidance of key users or groups, frequently giving rise to guiding topics, some of which seriously impact social stability. Currently, guiding topics have become a new trend in topic dissemination, and their application in advertising, marketing, and expanding personal influence has become a research hotspot. Therefore, the dissemination and trends of guiding topics have become important issues in both society and academia. Consequently, predicting the spread of guiding topics plays a vital role in combating rumors, guiding public values, and monitoring media misinformation.

[0003] In recent years, the spread of trending topics has attracted widespread attention from scholars both domestically and internationally. Research on the spread of trending topics has yielded rich results, examining its dissemination patterns from multiple dimensions. The main research areas are dissemination models, identification and monitoring, and situational awareness. In establishing dissemination models, researchers primarily rely on the classic infectious disease model (SIR) to categorize social media users into different states and establish dynamic topic dissemination models. In identification and monitoring, features are extracted from the topic's dissemination network and content, and machine learning or deep learning methods are used to classify rumors and non-rumor topics. In situational awareness, researchers start with the characteristics and network structure of users participating in the topic, utilizing deep learning algorithms to mine the evolutionary mechanism of the topic and predict its dissemination trend.

[0004] However, predicting the potential behavior of user groups in leading topic networks still faces the following challenges:

[0005] 1. The Hidden Nature of Group Consciousness in Guided Topics. Under the guidance of key users in a topic, group consciousness can influence the behavioral tendencies of the group. Therefore, uncovering hidden group consciousness and identifying key users' insights into group consciousness is crucial.

[0006] 2. The problem of topic polarization caused by dynamic confrontation of group consciousness. The group consciousness surrounding guiding topics can develop into dynamic confrontation as the topic spreads, leading to polarization of user behavior. Therefore, accurately quantifying the dynamic confrontation of group consciousness is a pressing issue.

[0007] 3. The problem of blurred group consciousness in the early stages of a topic's evolutionary lifecycle. In the early stages of a trending topic's emergence, the topic space contains a large amount of data unrelated to the trending topic, resulting in sparse effective topic data. This can affect the accuracy of the entire topic propagation model. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a method for discovering group consciousness in the context of guiding topics. The method includes: acquiring guiding topic data and inputting it into a trained group consciousness discovery model for processing to obtain the group consciousness discovery model.

[0009] The process by which the group consciousness detection model processes guided topic data includes:

[0010] S1: Obtain the guiding topic data and preprocess it to obtain user short texts and topic short texts;

[0011] S2: Use a pre-trained BERT model to process user short texts and topic short texts to obtain user text feature vectors and topic text feature vectors;

[0012] S3: Merge user text feature vectors and topic text feature vectors to obtain the topic text content feature matrix;

[0013] S4: Obtain the topic network from the guided topic data, process the topic network using the TGC2vec algorithm, and obtain the topic topology feature matrix;

[0014] S5: Combine the topic text content feature matrix and the topic topology feature matrix to obtain the user node matrix;

[0015] S6: Process the guiding topic data according to cognitive game theory to obtain the influence of three types of user forwarding behavior, and optimize the topic network adjacency matrix based on the influence.

[0016] S7: Calculate the aggregate features of user nodes based on the user node matrix and the topic network adjacency matrix; input the aggregate features into the classifier for processing to obtain the prediction results of the forwarding behavior of all users.

[0017] Preferably, the preprocessing of the guiding topic data includes: removing emoticons and meaningless symbols from the guiding topic data, and performing word segmentation on the topic text and user text of the guiding topic data to obtain user short text and topic short text.

[0018] Preferably, the process of integrating user text feature vectors and topic text feature vectors includes: calculating the fit between users and topics based on user text feature vectors and topic text feature vectors; concatenating the fit with user text feature vectors to obtain user text representation features; and integrating the text representation features of all users to obtain the topic text content feature matrix.

[0019] Preferably, the process of obtaining the topic topology feature matrix includes:

[0020] Calculate the intimacy level between users based on guided topic data;

[0021] The edge weights of the topic network are calculated based on the intimacy between users;

[0022] The transition probability of user nodes is calculated based on the edge weights of the topic network.

[0023] Each user node performs a walk based on the transition probability, resulting in a walk sequence for each user node; the walk sequences of all user nodes are integrated to obtain the user walk matrix;

[0024] The user roaming matrix is ​​embedded to obtain the topic topology feature matrix.

[0025] Furthermore, the formula for calculating the intimacy between users is:

[0026]

[0027] in, Indicates user and Between The weight of each interaction method Indicates user and In the The total number of interactions under each type of interaction behavior and Representing users respectively , In the The total number of interactions with other users under each type of interactive behavior.

[0028] Preferably, the process of obtaining the interaction effect of the user's three forwarding behaviors includes:

[0029] Based on the guided topic data, calculate user basic attributes, user activity, user intimacy, neighborhood trust, and topic popularity;

[0030] Internal user factors are calculated based on user basic attributes and user activity; external topic factors are calculated based on user intimacy, neighborhood trust, and topic popularity.

[0031] Based on internal user factors and external topic factors of user forwarding behavior, a multiple linear regression algorithm is used to construct an influence function for three user positions;

[0032] Calculate the payoff functions for the three positions based on their influence functions;

[0033] Based on cognitive game theory, the influence of three user forwarding behaviors is calculated according to the payoff functions of three positions.

[0034] Furthermore, the formulas for calculating user basic attributes, user activity, user intimacy, neighborhood trust, and topic popularity are as follows:

[0035]

[0036]

[0037]

[0038]

[0039] in, Indicates user The basic attributes, Indicates user gender, Indicates user age, Indicates user The total number of fans Indicates user Number of friends Indicates user activity level This represents the average number of posts a user makes per day. This indicates the user's participation behavior in the topic. Indicates user Trust level in the field express, express, Indicates user Official certification level This indicates the introduction of a weighting factor. Indicate topic At any moment The topic's popularity at the time Indicates the first regulating factor. Indicates the second regulating factor. Indicate topic The number of reposts, Indicate participation in the topic Total number of users Indicates the topic's popularity decay factor. For the current moment, This marks the moment when the topic began to explode.

[0040] Preferably, in step S6, the process of optimizing the topic network adjacency matrix based on influence includes:

[0041] Calculate the influence matrix of the three stances based on the influence of the user's three forwarding behaviors;

[0042] The adjacency matrix is ​​optimized based on the influence matrices of the three positions, using the following formula:

[0043]

[0044] in, This represents the optimized topic network adjacency matrix. Represents a non-linear activation function. This represents the original adjacency matrix of the topic network. An influence matrix representing a positive stance. An influence matrix representing negative stances. An influence matrix representing a neutral stance. Indicates the weight of positive stance. Indicates the weight of negative positions. Indicates the weight of a neutral stance.

[0045] Preferably, in step S7, the process of calculating the aggregation features of nodes includes:

[0046] A multi-head attention mechanism is adopted to calculate multiple attention coefficients between user nodes based on the user node matrix;

[0047] User node features are aggregated based on attention coefficients and topic network adjacency matrices to obtain multiple attention aggregated features;

[0048] By concatenating multiple attention aggregation features, the aggregated features of the user node are obtained.

[0049] Preferably, during the training process of the group consciousness detection model, the original topic domain is used as the source domain and the guiding topic domain is used as the target domain; the text features of the source domain and the target domain are extracted, and the text features are input into the topic domain classifier and the behavior classifier for training. After training, the trained topic domain classifier is used to transfer the original topic domain data to the guiding topic.

[0050] The beneficial effects of this invention are as follows: The proposed method for discovering group consciousness in guiding topics effectively predicts the spread of guiding topics; by comprehensively considering the multi-dimensional cascading influence of user characteristics and topic characteristics on group consciousness, as well as the mutual influence of multiple positions in guiding topics, this invention integrates user and topic content representation, uncovers hidden key users, and improves the accuracy of the prediction results of the group consciousness discovery model; simultaneously, from the perspective of topic derivation, this invention uses domain adversarial neural networks to adapt features between the original topic domain and the guiding topic domain, realizing data migration between different topic domains and further enhancing the model's accuracy; by predicting the potential behavior of user groups in guiding topic networks, this invention helps to control the development trend of online rumors, more accurately discover different rumor groups, and more effectively guide relevant rumor subjects to correctly spread information, thus avoiding irreparable damage caused by large-scale harmful online rumors on the Internet. Attached Figure Description

[0051] Figure 1 This is a flowchart of the group consciousness discovery method for guided topics in this invention;

[0052] Figure 2 This is an example diagram illustrating the fusion of user topic content in this invention;

[0053] Figure 3 This is a schematic diagram illustrating the acquisition of the topic topology feature matrix using the TGC2vec algorithm in this invention;

[0054] Figure 4 This is a schematic diagram of the interaction mechanism of multiple positions in this invention;

[0055] Figure 5 This is a schematic diagram illustrating the adversarial migration of the topic domain in this invention. Detailed Implementation

[0056] 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.

[0057] This invention proposes a method for discovering group consciousness in guided topics, such as... Figure 1 As shown, the method includes the following:

[0058] The guided topic data is obtained and input into a trained group consciousness discovery model for processing, thus obtaining the group consciousness discovery model.

[0059] The process by which the group consciousness detection model processes guided topic data includes:

[0060] S1: Obtain the guiding topic data and preprocess it to obtain user short texts and topic short texts.

[0061] Guided topic data is obtained through publicly available online data sources and APIs provided by public social platforms. Guided topic data includes topic dissemination data and user behavior data. Among them, topic dissemination data includes topic text information and topic dissemination network information throughout the entire life cycle of the guided topic; user behavior data includes basic user attribute information, user likes, comments, reposts, follows, mentions, etc.

[0062] The preprocessing of the guiding topic data involves several steps. First, considering the presence of numerous meaningless emojis and other symbols in the topic text data, which can interfere with user topic data representation and model performance, emojis and meaningless symbols are removed, retaining only regular character data to obtain cleaned guiding topic data. Then, the topic text and user text of the cleaned guiding topic data are segmented into words, and stop words are removed to reduce noise, resulting in short user text. and topic short text , This represents the l-th short text segment for user u. The l-th short text representing topic t.

[0063] S2: Use a pre-trained BERT model to process user short texts and topic short texts to obtain user text feature vectors and topic text feature vectors.

[0064] like Figure 2 As shown, a pre-trained BERT model is used to process user short texts and topic short texts to obtain user text feature vectors. and topic text feature vectors .

[0065] S3: Combine user text feature vectors and topic text feature vectors to obtain the topic text content feature matrix.

[0066] The fit between a user and a topic is calculated based on the user's text feature vector and the topic's text feature vector. The calculation formula is as follows:

[0067]

[0068] in, Indicates user Relevance to the topic Indicates user User text feature vectors, Represents the topic text feature vector. Represents the magnitude of a vector.

[0069] splicing fit With user text feature vector We obtain the text representation features of users; by integrating the text representation features of all users, we obtain the topic text content feature matrix. .

[0070] S4: Obtain the topic network from the guided topic data, process the topic network using the TGC2vec algorithm, and obtain the topic topology feature matrix.

[0071] Obtaining topic networks from guided topic data, for Momentary Topics Network , This represents the set of all user nodes participating in the topic. This represents the network of relationships among users participating in the topic. For example... Figure 3 As shown, this invention designs a new TGC2vec algorithm walking strategy based on a random walk strategy combined with a user affinity factor between user nodes. Specifically:

[0072] Randomly select network nodes Starting from the node, the normalized affinity between nodes is used as a weighting factor to determine the next node to traverse; the objective function of the TGC2vec algorithm is defined as:

[0073]

[0074] in, For nodes The set of neighboring nodes, It is to make nodes A mapping function that maps to vectors. Represents a node The probability of the occurrence of a neighboring node.

[0075] Calculate user intimacy based on guided topic data:

[0076] This invention uses user intimacy to measure the degree of intimacy between users during the dissemination of a topic. The degree of intimacy between users is often reflected in various behaviors such as following, liking, forwarding, replying, and mentioning each other. Generally, the higher the frequency of interaction between users, the higher the intimacy. This invention considers that different behaviors represent different degrees of interaction; therefore, it assigns different weighting coefficients to these behaviors. To define user intimacy as:

[0077]

[0078] in, Indicates user and Between The weight of each interaction method is assigned, with more intimate behaviors receiving higher weight. For example, in user interactions, replying is considered more intimate than liking, therefore replying carries a higher weight. It should be higher than the act of liking; Indicates user and In the The total number of interactions under each type of interaction behavior and Representing users respectively , In the The total number of interactions with other users under each type of interactive behavior.

[0079] The edge weights of the topic network are calculated based on the intimacy between users:

[0080]

[0081] First, let L be the maximum length of the random walk, and then choose the starting point of the random walk as... The node sequence obtained after the random walk is as follows Considering the influence of the affinity between user nodes and the depth of the topic space, the transition probability of user nodes is calculated based on the edge weights of the topic network:

[0082]

[0083] in, Represents user node To the next user node The transition probability, Indicates the wandering bias term. Represents user node With the next user node Edge weights between them Represents the normalization constant. Represents the set of edges in the topic network. Represents user node With the next user node Intimacy level, Represents user node The set of neighboring nodes.

[0084] wandering bias term :

[0085]

[0086] in, To return the probability of the previous node, when When the time is right, it is more likely to wander to a wider area; otherwise, it is more likely to wander to a local area. Controlling the probability of reaching a new node during the walk, when When moving, the roaming tends to be more inclined to visit nearby nodes, and vice versa, it tends to visit distant nodes. Represents a node and The shortest distance between, with values ​​ranging from {0, 1, 2}.

[0087] Each user node performs a walk according to the transition probability, resulting in a walk sequence for each user node; the walk sequences of all user nodes are integrated to obtain the user walk matrix.

[0088] Embedding the user walk matrix, i.e., representing it as a low-rank eigenvector matrix, yields the topic topology feature matrix. Where N represents the number of user nodes in the guiding topic network space. This represents the dimension of the feature vector.

[0089] S5: Combine the topic text content feature matrix and the topic topology feature matrix to obtain the user node matrix.

[0090] topic text feature matrix Topological structure feature matrix of topics Merging is performed along the feature dimension, preserving all information from the original features and merging them into a larger feature space. Specifically:

[0091]

[0092] in, The number of user nodes in the topic's network space. This represents the dimension of the user feature vector.

[0093] S6: Process the guiding topic data according to cognitive game theory to obtain the influence of three types of user forwarding behavior, and optimize the topic network adjacency matrix based on the influence.

[0094] Based on guided topic data, the following metrics are calculated: user basic attributes, user activity, user intimacy, neighborhood trust, and topic popularity:

[0095] User basic attributes refer to factors such as gender, number of followers, age, and number of friends. These attributes have a certain influence on whether a user participates in the dissemination of guided topics. Therefore, this invention defines user basic attributes as follows:

[0096]

[0097] in , , and Representing users respectively The user's gender, age, number of followers, and number of friends.

[0098] User activity level represents a user's level of engagement with a topic. The more frequently a user likes, comments, and shares a topic, the higher their activity level within that topic. Highly active users not only represent high engagement, but their participation also influences the topic behavior of other users. Therefore, user activity level is defined as:

[0099]

[0100] in, This represents the average number of posts a user makes per day. This represents the sum of user engagement activities within a topic, including likes, comments, and shares.

[0101] Social networks often contain a large number of users, but they are also rife with malicious users such as fake followers and paid trolls. Therefore, this invention utilizes BERT to vectorize user historical behavior and topic information, and defines the degree of fit between users and topics by calculating their cosine similarity, thereby reducing the influence of irrelevant users on topic dissemination. Simultaneously, "opinion leaders" often possess high domain trust, and their posted topics or comments have a significant impact on other users. Therefore, this invention defines user domain trust as:

[0102]

[0103] in, On behalf of users Official certification level On behalf of users The total number of followers. Official accounts tend to have higher credibility compared to other accounts with a large number of followers, therefore a weighting factor is introduced. Reduce the impact of follower count on user trust levels.

[0104] Topics Indicates in Moment, topic The volume of traffic, or the popularity of a topic, is crucial. On one hand, the lifecycle of a topic depends on its popularity; user behavior differs significantly at different stages of a topic's evolution. On the other hand, the higher the topic's popularity, the easier it is to attract users to participate in discussions. Therefore, topic popularity is defined as follows:

[0105]

[0106] in, Indicate topic The number of reposts, Indicate participation in the topic Total number of users and These are adjustable hyperparameters, among which... This primarily serves to adjust the difference in the impact of the number of topic reposts and the total number of participants on topic popularity. Additionally, considering that topic popularity typically peaks briefly and then gradually declines, a time decay function is used. To simulate this process, in which For the current moment, The moment the topic began to explode This is the topic popularity decay factor, whose value is obtained by fitting real data.

[0107] This invention takes into account the user's perspective in guided discussions and comprehensively considers internal user factors. External factors related to the topic ; Calculate the internal user factors and external topic factors that influence user forwarding behavior:

[0108] Internal user factors mainly consist of basic user attributes and user activity, namely:

[0109]

[0110] External factors for a topic mainly consist of user intimacy, neighborhood trust, and topic popularity.

[0111]

[0112] If the current user is a user with a positive stance, then It can be represented as Similarly, the users of the other two positions... This can be represented as , .

[0113] like Figure 4 As shown, based on internal user factors and external topic factors of user forwarding behavior, a multiple linear regression algorithm is used to construct an influence function for three user stances:

[0114]

[0115] in, This indicates that the positive stance of the group towards users The influence This indicates the negative stance of the group towards the user. The influence This indicates that the neutral group is in favor of users. Its influence. , , These represent external factors influencing a topic, representing a positive, negative, and neutral stance, respectively. , , These are the partial regression coefficients obtained from training the multiple linear regression algorithm.

[0116] The payoff functions for the three positions are calculated based on their influence functions. Specifically, this invention defines three game strategies based on cognitive game theory: a "positive position," a "negative position," and a "neutral position." The payoff functions for the three positions are as follows:

[0117]

[0118] in, , , Indicates user The proportion of positive, negative, and neutral posts forwarded within a friend network, and .

[0119] Based on cognitive game theory, a model of mutual influence among multiple positions was established. The mutual influence of quantified guiding topic group consciousness is calculated based on the benefit functions of three positions to obtain the mutual influence of users' three forwarding behaviors:

[0120]

[0121] in, , and These represent the positive, negative, and neutral positions for the user after the game. The impact of forwarding.

[0122] For each node Then, fill in the influence value into the corresponding influence matrix. For example, the influence matrix for a positive stance:

[0123]

[0124] in, For nodes The neighboring nodes.

[0125] Similarly, , Ultimately, this yielded an influence matrix with three different stances: , , .

[0126] Finally, the edge weights of the topic network topology are optimized based on the influence matrices of the three positions, representing the influence relationships and interaction strengths of nodes on different messages propagated by neighboring nodes. Specifically, the original adjacency matrix is ​​combined with the game-related mutual influence matrix to obtain a new adjacency matrix.

[0127]

[0128] in, Represents a non-linear activation function. This is a hyperparameter for adjusting the weights of each influence matrix.

[0129] S7: Calculate the aggregate features of user nodes based on the user node matrix and the topic network adjacency matrix; input the aggregate features into the classifier for processing to obtain the prediction results of the forwarding behavior of all users.

[0130] Considering that traditional convolutional neural networks cannot handle non-Euclidean data such as social networks, while Graph Attention Networks (GATs) with attention mechanisms can assign different weights to node information in the topic network, focusing on more valuable nodes, this invention uses GAT to process guiding topic data and predict user forwarding behavior influenced by the topic group during the propagation of guiding topics. Based on the topic group being divided into three main groups, user behavior is also divided into three types: positive stance, negative stance, and neutral stance.

[0131] To capture the features of different nodes and improve the model's expressive power, this invention uses a multi-head attention mechanism to optimize the aggregated features of nodes. Specifically, the multi-head attention mechanism calculates multiple attention coefficients between user nodes based on the user node matrix. The calculation formula is as follows:

[0132]

[0133] in, This represents the attention coefficient between user node i and user node j. The linear transformation matrix representing the attention head. This represents the feature of user node i obtained from the user node matrix. This represents the feature of user node j obtained from the user node matrix. Represents a node The set of neighboring nodes, This represents a vector concatenation operation. Map the result of concatenating vectors to a real number. It is a non-linear activation function.

[0134] User node features are aggregated based on attention coefficients and topic network adjacency matrices to obtain multiple attention aggregated features. The formula for obtaining the attention aggregated features is as follows:

[0135]

[0136] in, This represents the attention aggregation feature of user node i. This represents the activation function. This represents the edge weight between node i and node j in the adjacency matrix of the topic network.

[0137] By concatenating multiple attention-aggregated features, the aggregated features of the user node are obtained:

[0138]

[0139] in, This represents the aggregation feature of user node i. Indicates the number of heads of attention. Indicates the first Attention coefficients calculated by each attention head. Indicates the first A linear transformation matrix for each attention head.

[0140] The aggregated features are input into a classifier for processing to obtain the forwarding behavior prediction results for all users; specifically, the classifier includes a pooling layer and a fully connected layer, and the model output is:

[0141]

[0142] in, The output obtained after modeling The group behavior of users who constantly guide topics. This indicates that the user group will forward messages with a positive stance. This indicates that the user group will forward messages with a neutral stance. This indicates that the user group will forward messages with negative stances; Indicates user The probability of forwarding messages with a positive stance. Indicates user The probability of forwarding a neutral message. Indicates user The probability of forwarding messages with negative stances; This represents the classifier output.

[0143] In the early stages of a trending topic's explosive growth, the topic space contains a large amount of data unrelated to the trending topic, resulting in sparse effective data. Discovering the collective consciousness of a trending topic often relies on training with a large amount of effective data; using sparse data from the trending topic domain makes it impossible to train an effective topic propagation model. However, trending topics typically originate from numerous trending topics, and because the original topic domain and the trending topic domain share significant similarities, migrating a large amount of effective data from the original topic domain to the trending topic domain can greatly alleviate the early data sparsity problem. Inspired by Domain-adversarial Neural Network (DANN), this invention performs feature adaptation between the original topic domain and the trending topic domain during the training process of the collective consciousness discovery model, achieving data migration between different topic domains. Figure 5 As shown, specifically:

[0144] Original topic domain As the source domain, the guiding topic domain As the target domain, and These represent the user feature vectors of the original topic domain and the guided topic domain, respectively. and These represent the tag data for the original topic domain and the guided topic domain, respectively.

[0145] Extract text features from the source and target domains, and input these features into the topic domain classifier and behavior classifier for pitfall training. Specifically:

[0146] The extracted text feature data are then input into the topic domain classifier. and behavior classifier During the training phase, accurate classification of the topic domain dataset is required, and the guided topic domain dataset and the native topic domain dataset need to be obfuscated to maximize the topic domain classification error. This invention defines the adversarial transfer loss function as:

[0147]

[0148] in, The loss function represents the behavior classifier. This represents the loss function of the topic domain classifier. It is a hyperparameter that controls the relationship between two losses.

[0149] The parameters of the behavior classifier are updated by minimizing the loss function, and the parameters of the topic domain classifier are updated by maximizing the loss function.

[0150]

[0151] Among them, Indicates the parameters of the feature extractor. This represents the parameters of the task classifier. This represents the parameters of the domain classifier. First, minimize the loss function. Meanwhile, regarding the parameters of the feature extractor Parameters of behavior classifier Optimization, that is, finding the method that minimizes the overall loss. At the same time, maximize the loss function. Parameters of the behavior classifier That is, to find the one that causes the greatest overall loss. .

[0152] This invention combines data from the native topic domain and the guided topic domain, and introduces DANN as the domain adaptation layer of the model. This effectively transfers the native topic domain data to the guided topic domain, enhances the early group consciousness representation effect of the guided topic, and improves the accuracy of the model.

[0153] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for discovering group consciousness in guiding topics, characterized in that, include: The process involves acquiring guiding topic data and inputting it into a pre-trained group consciousness detection model for processing. During the training of the group consciousness detection model, the original topic domain is used as the source domain, and the guiding topic domain is used as the target domain. Text features are extracted from the source and target domains and input into the topic domain classifier and behavior classifier for training. After training, the pre-trained topic domain classifier is used to transfer the original topic domain data to the guiding topic. The process by which the group consciousness detection model processes guided topic data includes: S1: Obtain the guiding topic data and preprocess it to obtain user short texts and topic short texts; S2: Use a pre-trained BERT model to process user short texts and topic short texts to obtain user text feature vectors and topic text feature vectors; S3: Merge user text feature vectors and topic text feature vectors to obtain the topic text content feature matrix; S4: Obtain the topic network from the guided topic data, process the topic network using the TGC2vec algorithm, and obtain the topic topology feature matrix; S5: Combine the topic text content feature matrix and the topic topology feature matrix to obtain the user node matrix; S6: Process the guiding topic data according to cognitive game theory to obtain the influence of three user forwarding behaviors, and optimize the topic network adjacency matrix based on the influence; the process of obtaining the mutual influence of the three user forwarding behaviors includes: Based on the guided topic data, calculate user basic attributes, user activity, user intimacy, neighborhood trust, and topic popularity; Internal user factors are calculated based on user basic attributes and user activity; external topic factors are calculated based on user intimacy, neighborhood trust, and topic popularity. Based on internal user factors and external topic factors of user forwarding behavior, a multiple linear regression algorithm is used to construct an influence function for three user positions; Calculate the payoff functions for the three positions based on their influence functions; Based on cognitive game theory, the influence of three user forwarding behaviors is calculated according to the payoff functions of three positions. The process of optimizing the topic network adjacency matrix based on influence includes: Calculate the influence matrix of the three stances based on the influence of the user's three forwarding behaviors; The adjacency matrix is ​​optimized based on the influence matrices of the three positions, using the following formula: ； in, This represents the optimized topic network adjacency matrix. Represents a non-linear activation function. This represents the original adjacency matrix of the topic network. An influence matrix representing a positive stance. An influence matrix representing negative stances. An influence matrix representing a neutral stance. Indicates the weight of positive positions. Indicates the weight of negative positions. Indicates the weight of a neutral stance; S7: Calculate the aggregate features of user nodes based on the user node matrix and the topic network adjacency matrix; input the aggregate features into the classifier for processing to obtain the prediction results of the forwarding behavior of all users.

2. The method for discovering group consciousness in a guided discussion according to claim 1, characterized in that, The preprocessing of guiding topic data includes: removing emoticons and meaningless symbols from the guiding topic data, and performing word segmentation on the topic text and user text of the guiding topic data to obtain user short text and topic short text.

3. The method for discovering group consciousness in a guided discussion according to claim 1, characterized in that, The process of fusing user text feature vectors and topic text feature vectors includes: calculating the fit between users and topics based on user text feature vectors and topic text feature vectors; concatenating the fit with user text feature vectors to obtain user text representation features; and integrating the text representation features of all users to obtain the topic text content feature matrix.

4. The method for discovering group consciousness in a guided discussion according to claim 1, characterized in that, The process of obtaining the topic topology feature matrix includes: Calculate the intimacy level between users based on guided topic data; The edge weights of the topic network are calculated based on the intimacy between users; The transition probability of user nodes is calculated based on the edge weights of the topic network. Each user node performs a walk based on the transition probability, resulting in a walk sequence for each user node; the walk sequences of all user nodes are integrated to obtain the user walk matrix; The user roaming matrix is ​​embedded to obtain the topic topology feature matrix.

5. A method for discovering group consciousness in a guided discussion according to claim 4, characterized in that, The formula for calculating the intimacy level between users is: ； in, Indicates user and The level of intimacy between them Indicates user and Between The weight of each interaction method Indicates user and In the The total number of interactions under each type of interaction behavior and Representing users respectively , In the The total number of interactions with other users under each type of interactive behavior.

6. The method for discovering group consciousness in a guided discussion according to claim 1, characterized in that, The formulas for calculating user basic attributes, user activity, user intimacy, neighborhood trust, and topic popularity are as follows: ； ； ； ； in, Indicates user The basic attributes, Indicates user gender, Indicates user age, Indicates user The total number of fans Indicates user Number of friends Indicates user activity level This represents the average number of posts a user makes per day. This indicates the user's participation behavior in the topic. Indicates user Trust level in the field express, express, Indicates user Official certification level This indicates the introduction of a weighting factor. Indicate topic At any moment The topic's popularity at the time Indicates the first regulating factor. Indicates the second regulating factor. Indicate topic The number of reposts, Indicate participation in the topic Total number of users Indicates the topic's popularity decay factor. For the current moment, This marks the moment when the topic began to explode.

7. The method for discovering group consciousness in a guided discussion according to claim 1, characterized in that, In step S7, the process of calculating the aggregation features of nodes includes: A multi-head attention mechanism is adopted to calculate multiple attention coefficients between user nodes based on the user node matrix; User node features are aggregated based on attention coefficients and topic network adjacency matrices to obtain multiple attention aggregated features; By concatenating multiple attention aggregation features, the aggregated features of the user node are obtained.

Images