The proliferation of online media and social platforms has created a complex environment for public opinion formation and diffusion. Online public opinion events can rapidly escalate, posing challenges for management and governance. Key figures play a pivotal role in shaping and disseminating public opinion, with different roles influencing the event's trajectory. Traditional social network analysis methods often fall short due to their limitations in capturing the multidimensional aspects of online interactions. This study leverages the concept of a "supernetwork," a multi-layered network encompassing various interconnected relationships, to address this challenge. The research aims to design and evaluate an algorithm capable of simultaneously identifying different types of key figures involved in online public opinion dissemination, providing a deeper understanding of public opinion risk points and diffusion mechanisms. The study focuses on constructing a four-dimensional public opinion supernetwork, classifying key figures into distinct roles (opinion leaders, focus figures, and communication figures), and developing a classification and recognition algorithm that integrates multidimensional similarity and K-shell analysis.

Existing research on network public opinion dissemination highlights various influencing factors, including time, space, and social distance among participants, as well as attributes like emotions, opinions, and communication motivations. Social Network Analysis (SNA) methods are commonly used to identify key figures, with algorithms such as K-shell focusing on network location and centrality measures. However, limitations exist in the traditional single-layer SNA approach, which often overlooks implicit relationships and the multifaceted roles of key figures. Recent studies have explored multilayer and multidimensional networks, and the application of supernetwork theory to identify opinion leaders. Machine learning algorithms are also increasingly used to analyze user behavior and predict social influence. This study builds upon this existing research by incorporating both explicit and implicit relationships, classifying key figures into distinct roles beyond simple opinion leaders, and developing a novel algorithm that integrates both network location and attribute similarity.

The study constructs a "social-psychology-opinion-convergent" four-dimensional public opinion supernetwork model. This model integrates four layers: a social network representing forwarding relationships, an opinion network representing thematic affiliations, a psychological network representing emotional attributes, and a convergent network representing the roles of key figures. The algorithm uses a combination of multidimensional similarity and K-shell analysis to identify key figures. The multidimensional similarity index considers similarities across the four network layers, calculating similarities based on overlapping neighbour nodes in social, psychological, and opinion networks. This index helps capture the influence of individuals based on their shared characteristics and relationships. The K-shell decomposition algorithm is then used to identify nodes based on their position within the overall network structure. The combination of these two approaches provides a more comprehensive measure of influence. The influence index (Inf) is computed by combining the multidimensional similarity index and the K-shell index. Nodes are classified into four quadrants based on their Inf scores and K-shell values: opinion leaders (high similarity and high K-shell value), focus figures (low similarity and high K-shell value), communication figures (high similarity and low K-shell value), and ordinary figures (low similarity and low K-shell value). The algorithm is evaluated using AUC analysis, network destructive experiments, and fine-grained comparison with other methods, focusing on identifying opinion leaders, focus figures, and communication figures.

The study used data from the China Eastern Airlines plane crash event on Sina Weibo. Data preprocessing involved cleaning the text data, segmenting words, and removing stop words. Opinion extraction utilized LDA topic modeling combined with manual labeling to identify 82 core opinions. Sentiment analysis classified emotional intensity as positive, neutral, or negative. The algorithm effectively identified key figures across different phases of the event: opinion leaders (e.g., "China News Agency," "Mo Chen Mo Chen," "China Daily"), focus figures (e.g., "Modern Express," "Li Sweet Sauce," "China News Network"), and communication figures (e.g., "mind recorder," "big Qiqi is a big Qipa," "Yabo flavour Pop Rocks"). AUC analysis showed that the proposed algorithm outperformed baseline methods (CI, forwarding volume, degree centrality, K-shell, multidimensional similarity) in identifying key figures, with significantly higher AUC scores. Network destructive experiments demonstrated that the algorithm identified nodes whose removal caused greater network damage compared to nodes identified by baseline methods. Fine-grained analysis further revealed the algorithm's ability to effectively filter out less influential core-like nodes, identifying key figures with various roles in the network. Opinion leaders had extensive connections, focus figures demonstrated strong local influence, and communication figures served as bridging nodes between different communities.

The findings demonstrate the effectiveness of the proposed algorithm in identifying key figures with distinct roles in online public opinion events. The integration of multidimensional similarity and K-shell analysis provides a more nuanced understanding of influence compared to traditional methods. The algorithm's superiority in identifying focus figures and communication figures, often overlooked by simpler approaches, is a significant contribution. The case study highlights the algorithm's practical applicability in real-world scenarios. The algorithm’s ability to differentiate between global and local influencers adds valuable insights into public opinion dynamics. The results support the notion that understanding the diverse roles of key figures is crucial for effective public opinion management and crisis communication.

This study presents a novel algorithm for identifying key figures in online public opinion, incorporating multidimensional similarity and K-shell analysis within a supernetwork framework. The algorithm demonstrates superior performance compared to existing methods, effectively identifying opinion leaders, focus figures, and communication figures. Future research could focus on further refining the role classification, predicting the emergence of future key figures, and extending the algorithm to identify key topics within public opinion events. Exploring the integration of more advanced machine learning techniques within the supernetwork framework could further enhance the accuracy and efficiency of key figure identification.

While the algorithm demonstrates strong performance, several limitations should be noted. The study focuses on a single case study; further validation with diverse datasets is needed to ensure generalizability. The weighting scheme used in the multidimensional similarity index is relatively simple; more sophisticated weighting methods could improve performance. The algorithm relies on the availability of comprehensive data, which might not always be the case in real-world scenarios. Finally, the identification of 'potential' key figures, those who may become influential in the future, remains an area for future exploration.