A hotspot event controversial analysis method based on a heterogeneous symbolic attribute network
By using a method based on heterogeneous symbolic attribute networks, this study addresses the shortcomings of controversy analysis in hot topics on social networks. By utilizing self-supervised training and graph convolution operations, it achieves effective fusion of heterogeneous information and sentiment features, providing comprehensive controversy analysis results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2023-05-05
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies struggle to effectively handle heterogeneous information, attribute information, and sentiment characteristics when conducting controversy analysis on trending events in social networks, resulting in incomplete analysis results.
We employ a heterogeneous symbolic attribute network-based approach, which combines heterogeneous and symbolic networks for analysis through data collection, preprocessing, attribute fusion, community discovery, and controversy measurement, utilizing self-supervised training and graph convolution operations.
It enables multi-faceted and comprehensive analysis of the controversies surrounding trending events on social networks, accurately identifying the controversial aspects of events and providing detailed explanations.
Smart Images
Figure CN116777659B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data analysis technology, specifically a method for analyzing the controversy surrounding hot topics based on heterogeneous symbolic attribute networks. Background Technology
[0002] Every day, a large number of users participate in discussions of trending topics online, including public policies, emergencies, and social issues. Social networks shape communities of opposing viewpoints through these trending topics, and these communities, to some extent, influence the controversial nature of these events. Assessing the controversial nature of events within these communities is crucial for governments and businesses. Governments can adjust policies based on public attitudes, and businesses can optimize their products based on user opinions regarding their controversies. In social networks, trending topics involve diverse information, and the interaction of this information is often complex. A single user may post multiple comments and shares, and users differ in their level of influence; furthermore, the emotional (positive / negative) characteristics of users affect the community's outcome. Utilizing computer technology to quickly and accurately analyze massive amounts of information related to trending topics is a challenging task.
[0003] Current algorithms for controversial analysis mainly fall into three categories: content-based, network-based, and a combination of network and content. Content-based algorithms typically utilize textual features and probabilistic models to determine controversialness. This approach was favored by sociologists in early controversial analysis. Network-based models construct the topological structure between users through their interactions, primarily using node degree or boundary information to partition the graph. Generally, users with more interactions have higher node degrees and exhibit stronger internal connections within the community; Random-Walk Controversy is a representative example. The network-content combined model is a currently popular algorithm, combining the advantages of the previous two methods. It typically uses deep learning methods such as graph neural networks to determine controversialness, propagating textual features through the graph neural network along the topological structure, which better aligns with the understanding of social communication in reality.
[0004] Currently, methods for analyzing controversy still have some shortcomings. Content-based models analyze user-generated content, which is highly dependent on factors such as language and culture; textual feature representations may vary significantly across different linguistic contexts. Network-based models, on the other hand, are limited to homogeneous information between users, generally abstracting it into relationships between users, but cannot model multiple comments from a single user, inevitably leading to a lack of effective information. Models combining network and content consider both content and the network, and are currently a research hotspot. However, they do not handle crucial heterogeneous and complex information, including attribute information and sentiment characteristics, which can affect the results of controversy analysis. Modeling heterogeneous information, attribute information, and sentiment characteristics is key to successful controversy analysis. Furthermore, controversy analysis should consider methods for analyzing hot topics in communities based on heterogeneous symbolic attribute networks, trending keywords, and the community's main viewpoints; however, current research often fails to provide a comprehensive explanation of controversial events. Summary of the Invention
[0005] The purpose of this invention is to provide a method for hot topic controversy analysis based on heterogeneous symbol attribute networks, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for analyzing the controversy of trending events based on heterogeneous symbolic attribute networks is proposed. This method measures the degree of community difference from multiple perspectives and includes the following steps: 1) Data collection: Obtaining raw datasets from social networks using web crawling technology, which contain attribute information; 2) Data preprocessing and graph construction: Preprocessing by removing stop words and deleting data with low interaction, performing text summarization on all Weibo comments, obtaining event information based on heterogeneous symbolic attribute networks, and constructing a heterogeneous graph; 3) Attribute fusion: Integrating attribute information with textual and structural information; 4) Community discovery: Performing sentiment analysis based on user comment information, extracting the symbolic network between users, completing self-supervised training of users under both heterogeneous and symbolic networks, and obtaining community results; 5) Measurement of controversy: Measuring based on community differences, betweenness centrality, and user propagation representation.
[0008] Based on the above technical solutions, the present invention also provides the following optional technical solutions:
[0009] In one alternative approach: in step 2) data preprocessing and graph construction, the text summarization of all comments on Weibo is performed using a pre-trained Seq-Seq model, and the method for obtaining thematic entities in hot events is based on the Hidden Dirichlet Allocation Model.
[0010] In one alternative approach: In step 3) attribute fusion, the text information of microblogs, comments and summaries is represented by vectors using BERT, while the user representation is obtained through the Random Walk graph. Attribute levels are divided according to the attribute distribution, with higher levels for larger quantities, resulting in attribute codes. Vector concatenation technology is used to concatenate the attributes with the corresponding node vector representations, thereby completing attribute fusion.
[0011] In one alternative approach: In step 4), the heterogeneous relationships in the heterogeneous network are represented using different adjacency matrices, and graph convolution operations are performed layer by layer from the initial node representation. The convolution formula is:
[0012]
[0013] in, Represents the Laplace matrix under the τ relation. This represents the feature representation of the l-th layer. This represents the projection matrix of layer l, where σ is the activation function;
[0014] In symbolic networks, graph convolution operations are performed on networks under both balanced and unbalanced conditions to obtain vector representations.
[0015] The model converges based on the loss function of KL divergence, and a stable user node representation is obtained, so that the symbolic network and the heterogeneous network can achieve two-layer self-supervised training.
[0016] Community detection is achieved by performing Community Softmax operation on feature vectors from heterogeneous networks.
[0017] The KL loss function is expressed as:
[0018]
[0019] In one alternative approach: an attention mechanism is added in step 4, whereby the attention mechanism is used to capture the importance of different nodes, based on the attention scores of the current node and its neighboring nodes, and the formula is as follows:
[0020]
[0021] in, It is the attention vector, which is the node vector under the weighted sum of the normalized attention score and the feature representation; the sum of the attention vector and the node vector is used as the node representation in the heterogeneous network.
[0022] In one alternative approach: In step 5), the greater the difference between different communities in the dispute measurement, the higher the consensus and the lower the dispute. The centrality is measured based on structural characteristics. User propagation is represented by the closer the node distances of users in the same community, the lower the dispute. The average of the results of calculating community differences, centrality, and user propagation dispute is used as the final dispute judgment index. The closer the result is to 1, the greater the dispute, and the closer it is to 0, the less dispute.
[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0024] 1. The heterogeneous network in this invention contains various nodes and relationships, enabling message passing based on different heterogeneous relationships. Furthermore, to facilitate the analysis of the hot topic controversy analysis method based on heterogeneous symbolic attribute networks, the nodes of the hot topic controversy analysis method based on heterogeneous symbolic attribute networks are extracted, and the majority of opinions in the community are extracted as a summary. Attribute information is also of significant importance to the network model.
[0025] 2. This invention utilizes a self-supervised algorithm model to discover user communities and uses multiple measurement methods to determine the controversy surrounding the community results;
[0026] 3. The algorithm proposed in this invention solves the problem of heterogeneous symbolic attribute network community discovery, and conducts controversy analysis from different perspectives based on the community discovery results. Furthermore, it provides convenience for users by developing an online service platform to analyze hot events. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating the controversial analysis of this invention.
[0028] Figure 2 This is a schematic diagram of the heterogeneous network of the present invention.
[0029] Figure 3 This is a schematic diagram illustrating the community discovery process of the present invention.
[0030] Figure 4 This is a schematic diagram illustrating the controversial measurement method of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments listed herein are merely illustrative and not intended to limit the scope of the invention. Any obvious modifications or alterations made to this invention do not depart from the spirit and scope of the invention.
[0032] In one embodiment, such as Figure 1As shown, a method for hot topic controversy analysis based on heterogeneous symbol attribute networks is provided. The specific implementation of this method is as follows:
[0033] During the data collection phase, web crawling technology is used to obtain raw datasets from social networks, which also contain attribute information.
[0034] In data processing and graph construction, preprocessing operations are performed by removing stop words and deleting data with low interaction. Text-to-Text Transfer Transformer technology is used to perform text summarization processing on all comments of Weibo. The topic of the event is obtained according to Latent DirichletAllocation, and the heterogeneous graph is constructed.
[0035] Attribute fusion aims to integrate attribute information with textual and structural information. For textual information such as microblogs, comments, and summaries, BERT technology is used to obtain vector representations, while user representations are obtained using graph random walk technology. Furthermore, attribute levels are assigned based on attribute distribution, with higher levels indicating greater quantity, resulting in attribute encoding. Finally, vector concatenation techniques are used to combine the attributes with their corresponding node vector representations, thus completing attribute fusion.
[0036] The model structure for community discovery is as follows: Figure 3 As shown, sentiment analysis is performed based on user comments. Positive sentiments are connected by "+" edges, and negative sentiments by "-" edges, thus extracting the symbolic network between users. Next, self-supervised training is completed for users in both heterogeneous and symbolic networks. In the heterogeneous network, different adjacency matrices are used to represent heterogeneous relationships. Graph convolution operations are performed layer by layer from the initial node representation. The convolution formula is:
[0037]
[0038] in Represents the Laplace matrix under the τ relation. This represents the feature representation of the l-th layer. Let σ represent the projection matrix of layer l, and σ be the activation function.
[0039] The attention mechanism captures the importance of different nodes. It is based on the attention scores between the current node and its neighboring nodes, expressed by the formula:
[0040]
[0041] The attention vector that needs to be learned is obtained by weighting and summing the normalized attention score and the feature representation. The node vector under attention is then summed to obtain the node representation in the heterogeneous network.
[0042] In the symbolic network, graph convolution operations are performed on the networks under both balanced and imbalanced conditions to obtain vector representations. These vector representations are then fused using both concatenation and attention methods to obtain node representations within the symbolic network. To achieve two-layer self-supervised training for both networks, the model converges using a loss function based on KL divergence, resulting in stable user node representations. Finally, Community Softmax operations are performed on the feature vectors from the heterogeneous network to achieve community detection. The KL loss function can be expressed as:
[0043]
[0044] Controversy measurement involves judging the controversial nature of a community, and it is measured based on community differences, middle centrality, and user communication representation. Figure 4 These methods are illustrated in the document. The principle of community difference is that the greater the gap between different communities, the higher the consensus and the lower the controversy. Betweenness centrality is measured based on structural characteristics; user propagation represents the proximity of user nodes within the same community. The average of these three methods for calculating controversy is used as the final criterion for judging controversy; a result closer to 1 indicates greater controversy, while a result closer to 0 indicates less controversy. The analysis of event controversy can be approached from different angles, such as community sentiment, hot topic controversy analysis methods based on heterogeneous symbolic attribute networks, and summaries, thus providing specific explanations for the events.
[0045] This invention, utilizing the concepts of heterogeneous networks and symbolic networks, first extracts entities from collected data. Based on this, it extracts summaries and a hot topic controversy analysis method based on heterogeneous symbolic attribute networks to facilitate subsequent analysis. Based on the extracted nodes and relationships, it constructs a system as follows: Figure 2 The heterogeneous attribute network shown is hierarchical because of the differences in attributes between entities. This hierarchical processing is used to better highlight the importance of some nodes and to fuse them with text features.
[0046] A two-layer self-supervised training method using heterogeneous networks and symbolic networks was employed to obtain diverse community users. An attention mechanism was also incorporated to account for differences in relationships. Finally, the controversy surrounding the community results was measured, considering three methods: community differences, network structure, and user propagation representation. The analysis also covered aspects such as trending keywords, sentiment, and a method for analyzing the controversy surrounding trending events based on heterogeneous symbolic attribute networks.
[0047] The case study focuses on the hotly debated topic of "learning advanced mathematics." Based on the algorithm model of this application, it is determined that this is a controversial issue.
[0048] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.
Claims
1. A method for analyzing the controversy of hot topics based on heterogeneous symbolic attribute networks, characterized by measuring the degree of difference in communities from multiple different perspectives, including the following steps: 1) Data collection: Using web scraping technology to obtain raw datasets from social networks, including attribute information; 2) Data preprocessing and graph construction: Preprocessing is performed by removing stop words and deleting data with low interaction. Text summarization is performed on all comments on Weibo to obtain information on the controversy of hot events based on heterogeneous symbolic attribute networks, and the heterogeneous graph is constructed. 3) Attribute fusion, which integrates attribute information with textual and structural information; 4) Community discovery: Sentiment analysis is performed based on user comments to construct a symbolic network among users. Self-supervised training of users is completed under heterogeneous networks and symbolic networks to obtain community results. 5) Controversiality measures are taken based on community differences, intermediary centrality, and user communication representation.
2. The hot topic controversy analysis method based on heterogeneous symbolic attribute networks according to claim 1, characterized in that, In step 2) data preprocessing and graph construction, the text summarization of all comments on Weibo is performed using a Seq-Seq pre-trained model, and the method for obtaining topic entities in hot events is based on the Hidden Dirichlet Allocation Model.
3. The hot topic controversy analysis method based on heterogeneous symbolic attribute networks according to claim 1, characterized in that, In step 3), attribute fusion, the text information of microblogs, comments, and summaries is represented by vectors using BERT, while user representations are obtained using the Random Walk graph. Attribute levels are divided according to the attribute distribution, with higher levels for larger quantities, resulting in attribute codes. Vector concatenation technology is then used to concatenate the attributes with the corresponding node vector representations, thereby completing attribute fusion.
4. The hot topic controversy analysis method based on heterogeneous symbolic attribute networks according to claim 1, characterized in that, In step 4), the heterogeneous relationships in the heterogeneous network are represented using different adjacency matrices. Graph convolution operations are performed layer by layer from the initial node representation. The convolution formula is: in, Represents the Laplace matrix under the τ relation. This represents the feature representation of the l-th layer. This represents the projection matrix of layer l, where σ is the activation function; In symbolic networks, graph convolution operations are performed on networks under both balanced and unbalanced conditions to obtain vector representations. The model converges based on the loss function of KL divergence, and a stable user node representation is obtained, so that the symbolic network and the heterogeneous network can achieve two-layer self-supervised training. Community detection is achieved by performing Community Softmax operation on feature vectors from heterogeneous networks. The KL loss function is expressed as:
5. The hot topic controversy analysis method based on heterogeneous symbolic attribute networks according to claim 4, characterized in that, In step 4, an attention mechanism is added. This mechanism captures the importance of different nodes and is based on the attention scores between the current node and its neighboring nodes. The formula is as follows: in, It is the attention vector, which is the node vector under the weighted sum of the normalized attention score and the feature representation; the sum of the attention vector and the node vector is used as the node representation in the heterogeneous network.
6. In the hot topic controversy analysis method based on heterogeneous symbolic attribute networks according to claim 1, in step 5) controversy measurement, the greater the difference between different communities, the higher the consistency of opinions and the lower the controversy. The middle centrality is measured according to structural characteristics. User propagation means that in the same community, the closer the node representation of users is, the lower the controversy. In calculating the controversy of community difference, middle centrality and user propagation, the average of the results is used as the final controversy judgment index. When the result is closer to 1, it indicates greater controversy, and closer to 0, it indicates less controversy.