Early guide type topic detection method based on emotion accumulation and dynamic graph representation

Abstract

The invention belongs to the field of social network analysis, and particularly relates to an early guide type topic detection method based on emotion accumulation and dynamic graph representation, which comprises the following steps: obtaining topic data generated before and after topic outbreak, including basic information, comment information and friend relationship of comment making users in topics; establishing a relation weight matrix between the users according to theme distribution similarity of historical comments made by the users and emotional influence of the users; converting the basic information of the user into a vector representation; and inputting the vector representation of each user under the topic and the relation weight matrix between the users into a graph attention network based on time coding, and judging whether the current topic belongs to a guide type topic or not. The guiding type topic can be effectively recognized, and the method is crucial for timely controlling further evolution of guiding type topics such as rumors and correctly guiding network public opinions.

Description

Technical Field

[0001] The present invention belongs to the field of social network analysis, and in particular relates to an early guided topic detection method based on sentiment accumulation and dynamic graph representation. Background Art

[0002] In today's era of information explosion, social networks have become the main platform for information dissemination. The rapid spread of topics can affect public opinion and social dynamics in a short period of time. The analysis of topic data not only helps us understand the mechanism of information dissemination, but also effectively monitor and guide public opinion. However, due to the openness and dynamism of social networks, topic data presents a high degree of complexity and dynamic changes, which puts higher requirements on data collection, processing and analysis.

[0003] In recent years, many scholars have achieved remarkable results in the field of social network research. In early studies, scholars detected guiding topics by analyzing the user text features and topic propagation features in the process of guiding topic propagation. Effectively identifying and curbing malicious guiding topics in the early stages is of vital importance to the purification of the network environment, social stability and national security. However, topics usually face the problem of sparse data in the initial stage. At the same time, users' behaviors are affected by the emotions of their previous participation in the topic, and the effective representation of users is also a major problem. Therefore, conducting research on data compensation, emotion accumulation and dynamic representation of topics for guiding topics can more effectively detect such topics.

[0004] With the continuous growth of social media data and the advancement of deep learning technology, nonlinear models such as neural networks and deep learning have become the preferred methods for processing complex network data. Recently, Wei et al. (Wei S, Wu B, Xiang A, et al. DGTR: Dynamic graph transformer for rumor detection [J]. Frontiers in Research Metrics and Analytics, 2023, 7: 1055348.) developed a Transformer-based dynamic graph model specifically for rumor detection. This study confirmed the practicality of using temporal attributes and propagation structures for topic detection by effectively capturing the long-term dependencies of structure and time series. Inspired by this study and the effect of emotional accumulation, this study designed a method for early identification of guided topics that combines emotional accumulation and dynamic graph technology. This method quantitatively analyzes the impact of emotional accumulation and uses the TGAT network to track the changes in user characteristics in real time during the development of topics to improve the detection efficiency of guided topics.

[0005] In summary, the existing technical problems are:

[0006] Problem 1: Data sparsity in the early stages of the outbreak of leading topics. Due to low user participation and sparse number of comments, there is insufficient data on early topics, which affects the model's ability to accurately capture topic features.

[0007] Problem 2: The problem of user emotion accumulation in guiding topics. The emotional inclination of users participating in the topic is the core of solving the problem of guiding topic detection. Before the target topic breaks out, the user's emotion on the previous topic will have an impact on the current topic. We propose an emotion accumulation model to quantify this influence;

[0008] Problem 3: Effective embedding of user nodes of guiding topics. Due to the continuous accumulation, increase, and change of the number of users, user relationships, and user comments during the evolution of topics, it is inaccurate to embed nodes of users at a certain moment.

[0009] To sum up, in the initial stage of a topic outbreak, the number of users on the network is relatively small, and it is very important to detect malicious leading topics and curb them immediately. How to effectively compensate for the data in the early stage of the topic and then detect leading topics has become a problem that needs to be solved urgently. Summary of the invention

[0010] In view of the problems of the prior art in the background technology, the present invention proposes an early guidance topic detection method based on emotion accumulation and dynamic graph representation, which specifically includes the following steps:

[0011] Obtain topic data generated before and after the topic outbreak, including basic information, comment information, and friend relationships of users who posted comments in the topic;

[0012] According to the similarity of the topic distribution of the current topic comments and the interactive relationship between users, a relationship weight matrix between users is established;

[0013] Obtain the user's emotional accumulation representation, convert the user's basic information into a vector representation, and concatenate the user's emotional accumulation representation with the vector representation as the user's final vector representation;

[0014] The final vector representation of each user under the topic and the relationship weight matrix between users are input into the time-coded graph attention network to determine whether the current topic is a guiding topic.

[0015] The present invention has at least the following beneficial effects:

[0016] 1. The present invention extracts the topic distribution of topics and finds other topics with similar semantics to the target topic based on the topic similarity of the topics. The information of these similar topics is used for data compensation, which significantly improves the efficiency and accuracy of topic detection.

[0017] 2. In view of the impact of the user's previous accumulated emotions on other topics on this topic, the present invention proposes an emotional accumulation influence model to construct user emotional influence and integrate it into subsequent user characteristics. This innovative method solves the problem of user emotional accumulation.

[0018] 3. In the process of topic propagation, entities such as users, messages, and networks are constantly changing, which brings challenges to the effective representation of user node vectors in topics. In order to effectively meet this challenge, the present invention proposes a TGAT dynamic graph representation method, which fully considers the impact of changes in topic content on user embedding, greatly enriches user feature expression, and provides strong support for topic detection and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the early guidance topic detection method based on emotion accumulation and dynamic graph representation of the present invention;

[0020] Figure 2 A schematic diagram of data compensation according to an embodiment of the present invention;

[0021] Figure 3 A schematic diagram of a cumulative emotional influence model according to an embodiment of the present invention;

[0022] Figure 4 A schematic diagram for constructing the explicit and implicit relationship matrix of users in the present invention;

[0023] Figure 5 Schematic diagram of the TGAT dynamic graph representation learning method in the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The present invention proposes an early guidance topic detection method based on emotion accumulation and dynamic graph representation, which specifically includes the following steps:

[0026] Obtain topic data generated before and after the topic outbreak, including basic information, comment information, and friend relationships of users who posted comments in the topic;

[0027] Establish a relationship weight matrix between users based on the similarity of topic distribution of users' historical comments and the emotional influence of users;

[0028] Convert the user's basic information into a vector representation;

[0029] The vector representation of each user under the topic and the relationship weight matrix between users are input into the time-coded graph attention network to determine whether the current topic is a guiding topic.

[0030] like Figure 1 This embodiment proposes a specific implementation process of an early guidance topic detection method based on emotion accumulation and dynamic graph representation, including the following steps:

[0031] S1: Use the API interface provided by public dataset websites or social networks to comprehensively collect topic data, including user comments, detailed user information data, and interactive relationships between users;

[0032] S2: After collecting topic data, we perform detailed attribute extraction. These attributes cover multiple dimensions such as topic propagation network, topic element popularity, user credibility, etc. Based on these attributes, we construct the original topic propagation network, which provides an intuitive network structure reference for subsequent data compensation and node representation.

[0033] S3: Adopt the data compensation method based on LDA topic model, compensate the original topic by derived topic, solve the problem of sparse data in the early stage of topic, and then obtain the compensated topic space;

[0034] S4: To address the problem of users’ accumulated emotions on other topics before the current topic breaks out, we established an emotional cumulative influence model to capture the propagation and cumulative effect of users’ emotions in multiple topics, thereby quantifying their influence on the current topic.

[0035] S5: Adopt TGAT dynamic graph representation learning algorithm, which continuously updates the user's node embedding vector through dynamic graph representation, simulates the influence between users in the process of topic evolution, and accurately represents the user's node vector;

[0036] S6: Further aggregate the user vectors in the graph through the attention mechanism to obtain the graph representation vector of the topic. After passing through the fully connected layer, the Softmax function is used to convert the output into probability values ​​of two categories, and the category corresponding to the maximum probability value of the two categories is taken as the prediction result.

[0037] like Figure 2 As shown in the figure, based on the topic distribution of the extracted topic and the topic similarity of the topic, other topics with similar semantics to the current topic are found. The information of these similar topics can be used for data compensation. The specific process is to first standardize the text data, convert the document into a word frequency vector or a vocabulary item-document matrix, and then train the LDA model to obtain the topic distribution of each document. Finally, similar topics are found by comparing the similarity of the topic distribution.

[0038] Suppose each topic Ti Contains user commentsD i ={d i1 ,d i2 ,…,d iM}, these documents are transformed into the topic distribution θ of each topic through the LDA model i , in order to find the s Topics with similar topic distribution can be calculated by calculating the topic distribution similarity between topics. In this embodiment, the cosine similarity between topic distributions is used as the topic distribution similarity. s and topic T r , whose topic distributions are θ s and θ r , then the cosine similarity calculation formula is:

[0039]

[0040] where θ s ·θ r is the dot product of two topic distribution vectors, ||θ s ||||θ r || are the products of the norms of the two topic distribution vectors. The value of cosine similarity ranges from 0 to 1. The closer the value is to 1, the more similar the topics of the two topics are.

[0041] Therefore, some hot topics will generate a large number of strongly related derivative topics in the early stage, so it is necessary to select topic elements with higher credibility to compensate for the original domain of the topic. First, the user's historical reputation Cre (u i ) and topic popularity Pop(t):

[0042] Cre(u i )=grade(u i )+α×fansNum(u i )

[0043]

[0044] Among them, Cre(u i ) represents user u i The historical reputation of grade(u i ) indicates the official account level of the user, fansNum(u i ) represents the number of followers of the user. t and n represent the current time and the time when the topic is published, respectively. disNum(t) represents the number of comments and reposts of the topic at time t, and disNum(t-1) represents the number of comments and reposts of the topic at time t-1. At the same time, the half-life function is introduced. Simulate the process of a topic's popularity rapidly decreasing after reaching its peak.

[0045] The formula for the final element credibility is:

[0046] Cre(e i )=Cre(u i )+β×Pop lev (e i )

[0047] Among them, the topic element e i By user u i Release; Pop lev (e i ) represents the topic element e i The heat level is converted from the heat level of the topic element at the current moment Pop(t); topics posted by users are often related to their own positions, so the impact of user characteristics with high credibility on this study is higher than that of topic elements with high heat. The influence of topic element heat on the credibility of topic elements is weakened by the attenuation factor β∈[0,1].

[0048] In this embodiment, topics whose topic element credibility is greater than a set threshold are screened, and topics whose pre-similarity to the current topic is greater than the set threshold are obtained, and user comment information in these topics is used as implicit relationships:

[0049]

[0050] Where Cor(u i,j ) represents the user comment set e i and e j The relevance of comments, where Vec(e i ) and Vec(e j ) represents the user comment set e i and e j Rel is converted into a vector form using the representation learning doc2vec algorithm. cov (u i,j ) represents user u i and u j , and β represents the boundary of text correlation.

[0051] Since the current topic data has been compensated before and the topic elements with higher credibility have been screened out, a topic explicit and implicit adjacency matrix is ​​established, in which the explicit relationship refers to the direct and observable interactive relationship between users. The quantitative formula is as follows:

[0052] Rel dom (u i,j )=Int com (e i,j )+Int fri (u i,j)

[0053] Among them, Rel dom (u i,j ) represents user u i and u j Explicit relations of Int com (e i,j ) represents the user comment set e i and e j Comment relation weight, Int fri (u i,j ) represents user u i and u j The friend relationship weight of the two elements is represented by 0 or 1.

[0054] Then, the explicit and implicit relationships are combined to construct the user relationship network of the early topic domain. At the same time, based on the explicit and implicit relationship weights, the relationship weights of the nodes in the element relationship network are quantified. The quantification formula is as follows:

[0055]

[0056] Among them, Rel weight (u i ,u j ) represents user u i and user u j The weight between cov (u i ,u j ) represents the user u obtained according to the topic model i and user u j The implicit relationship between dom (u i ,u j ) represents the user u obtained based on emotional influence i and user u j The explicit relationship between is the impact factor and

[0057] like Figure 5 , proposed a time-encoded graph attention network (TGAT) to obtain graph embedding vectors and detect topics through a fully connected layer. First, the topic needs to be sliced ​​to obtain the topic network set at each moment Indicates t i The relationship network of the topic at the moment;

[0058] In order to integrate time information, the attention coefficient depends not only on the characteristics of the nodes, but also on the time difference information between nodes. The attention coefficient is expressed as:

[0059] e ij =LeakyReLU(a T [Wh i ||Wh j ||W t Δt ij ])

[0060] Among them, e ij Represents user u i The final vector representation of user u j The final vector represents the attention coefficient between them; LeakyReLU(·) represents the LeakyReLU function; a is the parameter vector used to calculate the attention coefficient; W is the conversion matrix. In the present invention, the function of the conversion matrix is ​​to map the feature vectors to be processed into the same vector space for easy calculation; h i For user u i The final vector representation of h j For user u j The final vector representation of || represents the concatenation operation; W t is the conversion matrix of the time difference vector, which is used to convert the time difference encoding into the same dimension as the user's vector representation; Δt ij is the time difference vector, whose value is the absolute value of the time difference between the interactions between users. For example, when user A posts a dynamic at time a, and user B likes, comments, or reposts user A's dynamic at time b, the time difference between user A and user B is |ab|.

[0061] In this embodiment, the final vector representation of the user is obtained by vectorizing the basic information of the user and concatenating it with the accumulated emotion representation of the user. The basic information of the user may be the basic information filled in when the user applies for an account, including information such as gender and age. The process of obtaining the accumulated emotion representation of the user includes the following steps:

[0062] Use sentiment analysis tools to classify the sentiment of users' speeches or comments on different topics. Sentiment classification includes positive sentiment, negative sentiment, and neutral sentiment.

[0063] Get topic T k The sentiment polarity score of each comment of the user in is calculated by summarizing the polarity scores of all comments, calculating the sentiment score of the entire topic, and averaging the polarity scores of all comments to obtain the overall sentiment tendency of the topic. i In Topic T k The sentiment score in is expressed as:

[0064]

[0065] Among them, S(T k ,u i) represents user u i In Topic T k The sentiment score in is represented by Polarity(w); Polarity(w) represents the sentiment polarity value of word w in the text, which can be obtained from the dictionary or pre-trained model, usually in the range of [-1,1], where the polarity value of positive words is positive (such as "good", "amazing" has positive sentiment polarity), the polarity value of negative words is negative (such as "bad", "terrible" has negative sentiment polarity), and the sentiment polarity of neutral words is 0; Modifier(w) represents the influence factor of modifiers, including negative words and emphasis words. It modifies the polarity of sentiment words. Negative word corrections (such as "not", "never") will reduce or reverse the scores of sentiment words, and emphasis words (such as "very", "extremely") will amplify the scores of sentiment words; |T k | indicates topic T k The length of the review, that is, the number of sentiment words involved in the calculation, ensures that the polarity score is in the range of [-1,1];

[0066] By calculating the user's sentiment score in the previous topic and considering the time decay of sentiment, the user's sentiment accumulation value E in the current topic is calculated. acc (u i ,T c ), expressed as:

[0067]

[0068] Where E acc (u i ,T c ) represents user u i In the current topic T c The emotional accumulation value in S(T k ,u i ) represents user u i In the previous topic T k Sentiment score in t c ,t k Respectively represent the time of the current topic and the previous topic, λ represents the time decay factor, and controls the decrease of emotional influence over time. Control the influence of the previous topic sentiment on the current topic. The older the sentiment, the smaller the influence. n represents the number of topics the user has participated in before.

[0069] Obtain the user's historical emotional accumulation value E acc (u i ,T c ), and then combined with the user's sentiment score S(T c ,u i), calculate the user's emotional influence on the current topic I(u i ,T c ), the formula is as follows:

[0070] I(u i ,T c )=α·S(T c ,u i )+β·E acc (u i ,T c )

[0071] Among them, I(u i ,T c ) represents user u i About the current topic c Emotional influence; S(T c ,u i ) represents the sentiment score of the user in the current topic; E acc (u i ,T c ) represents the historical emotion accumulation value of the user; α and β are weight coefficients, which control the relative contribution of current emotion and historical emotion accumulation to the emotion influence;

[0072] Finally, in the process of information dissemination, users receive information about a topic from multiple sources, and this information may have different emotional tendencies (positive, negative, neutral), and the user's emotions on the topic may fluctuate. Especially when the information transmitted by different sources is inconsistent, the user's emotional cognition will become more uncertain. Due to the deviation of the information received by users during the topic dissemination process, the user's emotions fluctuate. Entropy can capture the impact of this fluctuation, thereby adjusting the user's final emotional cumulative influence. The final emotional cumulative influence formula is:

[0073]

[0074] I final (u i ,T c )=I(u i ,T c )·(1-H(u i ,T c ))

[0075] Among them, H(u i ,T c ) represents user u i In the current topic T c Emotional uncertainty in j represents the probability of each emotional state of the user in the previous topic; I(u i ,Tc ) represents user u i About the current topic c The raw emotional impact of final (u i ,T c ) represents the final emotional influence after correction. ij Perform Softmax operation to normalize and get the final attention coefficient α ij :

[0076]

[0077] in Represents user u i The neighbor set of user u i Other users with whom you have direct or indirect friendships;

[0078] The process of updating the vector representation of the user according to the attention weight includes:

[0079]

[0080] Among them, h′ i Indicates the updated user u j ; σ(·) represents a nonlinear activation function, and the ReLU activation function is selected in this embodiment; α ij Represents user u i The vector representation of user u j The vector representation of the attention weight between ; Represents user u i The neighbor set of , W is the transformation matrix.

[0081] Finally, the user's reputation value is used to calculate the weight, and the weight is used to weight the vector representation of each updated user, and the weighted cumulative value of all users is used as the representation of the topic, that is:

[0082]

[0083] in, is the representation of the topic, α i is the weight calculated according to the user's reputation value, and d is the number of users.

[0084] Finally, the Softmax function is used to map the detection result of whether the topic belongs to the guiding topic. That is, the Softmax function is used to obtain the probability that the topic belongs to the guiding topic and the non-guiding topic, and the larger value is used as the detection result.

[0085] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An early guided topic detection method based on sentiment accumulation and dynamic graph representation, characterized in that: The specific steps include: Obtain topic data generated before and after the topic outbreak, including basic information, comment information, and friend relationships of users who commented in the topic, and convert the user's basic information into a vector representation; Establish a relationship weight matrix between users based on the similarity of the topic distribution of the current topic comments and the comment information and friend relationships of the commenting users; The vector representation of the user's basic confidence under the topic, the user's accumulated emotion representation, and the relationship weight matrix between users are input into the time-coded graph attention network to determine whether the topic is a guiding topic.

2. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 1 is characterized in that: In the relationship weight matrix between users, user u i and user u j The weight between is expressed as: Among them, Rel weight (u i ,u j ) represents user u i and user u j The weight between cov (u i ,u j ) represents the user u obtained according to the topic model i and user u j The implicit relationship between dom (u i ,u j ) represents the user u obtained based on emotional influence i and user u j The explicit relationship between is the impact factor and 3. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 2 is characterized in that: According to the user u obtained by the topic model i and user u j The implicit relationship between cov (u i ,u j ) process includes: Among them, Cor(e i , e j ) represents user u i Comments collection i With user u j Comments collection j The cosine similarity of , β is the set threshold.

4. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 1, 2 or 3, characterized in that: When establishing the relationship weight matrix between users, users are screened, that is, users whose cosine similarity of the distribution of historical comments posted by two users in the topic is greater than the set threshold are selected.

5. The early guided topic detection method based on emotion accumulation and dynamic graph representation according to claim 2 is characterized in that: User u obtained based on emotional influence i and user u j Rel dom (u i ,u j ) include: Rel dom (u i ,u j )=Int com (e i ,e j )+Int fri (u i ,u j ) Among them, Int com (e i , e j ) represents user u i Comments collection i With user u j Comments collection j The comment relationship weight, if there is comment interaction between two users, then Int com (e i , e j )=1, otherwise Int com (e i , e j )=0;Int fri (u i ,u j ) represents user u i With user u j The weight of the friendship relationship, if there is a direct good relationship between two users, Int fri (u i ,u j )=1, otherwise Int fri (u i ,u j )=0.

6. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 1 is characterized in that: The process of graph attention network prediction based on temporal encoding includes:

101. Slice the topic to obtain the topic network under each time slice, set t=1, and initialize the topic network using the currently obtained user node; 102. Calculate the attention coefficient between two users according to the vector representation of the basic information of the users at time t, and calculate the attention weight according to the attention coefficient; 103. Update the vector representation of the user according to the attention weight to obtain the vector representation of the user at time t+1. If t is not greater than the set number of time slices, set t=t+1 and return to step 102.

104. Weight the vector representation obtained by updating the last time slice according to the user's reputation value, and use the accumulated value of the weighted vector representations of each user as the vector representation of the topic; 105、 Update them sequentially according to the order of time slices, and weight the vector representations of all users obtained from the last update based on the attention mechanism, and use the weighted accumulated vector representation as the vector representation of the topic; The softmax function is used to obtain the probability that the topic vector represents the leading topic and the non-leading topic, and the larger value is used as the final prediction result of the topic.

7. The early guided topic detection method based on emotion accumulation and dynamic graph representation according to claim 6 is characterized in that: The process of calculating the attention coefficient between two users based on the vector representation of the users include: e ij =LeakyReLU(a T [Wh i ||Wh j ||W t Δt ij ]) Among them, e ij Represents user u i The vector representation of user u j The vector represents the attention coefficient between them; LeakyReLU(·) represents the LeakyReLU function; a is the parameter vector used to calculate the attention coefficient; W is the transformation matrix, h i For user u i The vector representation of h j For user u j The vector representation of || represents the concatenation operation; W t is the conversion matrix of the time difference vector, which is used to convert the time difference encoding into the same dimension as the user's vector representation, Δt ij is the time difference vector, whose value is user u i With user u j The absolute value of the time difference between interactions.

8. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 6 is characterized in that: The process of updating the vector representation of the user according to the attention weight includes: Among them, h′ i Indicates the updated user u j The vector representation of ; σ(·) represents the nonlinear activation function; α ij Represents user u i The vector representation of user u j The vector representation of the attention weight between ; Represents user u i The neighbor set number of , W is the transformation matrix.

9. The early guidance topic detection method based on emotion accumulation and dynamic graph representation according to claim 1 is characterized in that: The process of obtaining the user's emotional accumulation representation include: I final (u i ,T c )=I(u i ,T c )·(1-H(u i ,T c )) Among them, I final (u i , T c ) represents user u i In the current topic T c The final emotional impact in i , T c ) represents user u i In the current topic T c The original emotional impact in H(u i , T c ) represents user u i In the current topic T c Emotional fluctuations in.

Images