
Protecting privacy in social graphs may require obscuring nodes’ membership in sensitive communities. However, doing so without significantly disrupting the underlying graph topology remains a key challenge. In this work, we address the community membership hiding problem, which involves strategically modifying the graph structure to conceal a target node’s affiliation with a community, regardless of the detection algorithm used. We reformulate the original discrete, counterfactual graph search objective as a differentiable constrained optimisation task. To this end, we introduce ∇-CMH, a new gradient-based method that operates within a feasible modification budget to minimise structural changes while effectively hiding a node’s community membership. Extensive experiments on multiple datasets and community detection methods demonstrate that our technique outperforms existing baselines, achieving the best balance between node hiding effectiveness and graph rewiring cost, while preserving computational efficiency.
The proliferation of online social networks (OSNs) like Twitter (X) has given rise to a concerning issue: the absence of a robust instance verification mechanism allows users with malicious intentions to disseminate an alarming volume of fake news, with potentially dire consequences. Consequently, the evaluation of user credibility, especially for those who propagate misleading information, has become a crucial research area, necessitating careful exploration across multiple dimensions and innovative methodologies. However, most existing methods for assessing user credibility predominantly rely on textual analysis of their posts, overlooking the rich multimodal data available in OSNs, such as multimedia content (images). Therefore, to fill this gap, in this paper, we introduce a novel user credibility assessment model, called CredUser, that leverages textual and multimedia data. To this end, we first propose the Tweet Credibility Model (TCM), which focuses on evaluating the credibility of tweets based on textual features. Then, as a significant contribution to this paper, we present Tweet-embedded Multimedia Data Credibility Model (TeMdCM), which not only incorporates textual features from TCM but also harnesses multimedia-based features to assess tweet credibility.To validate our models, we conducted experiments involving 563 users and 9356 tweets, accompanied by a total of 9641 images. Our experimental results reveal the substantial impact of multimedia data on the credibility assessment process, where TeMdCM improves accuracy by +6.53% for tweet credibility and by +7.10% for user credibility compared to TCM. This underscores the potential of multimedia data as a vital source for evaluating content and user credibility.
The development of social media has facilitated the emergence of opinion-based groups, online public opinion polarization attracting significant scholarly attention. Despite the proliferation of polarization metrics, little is known about their differences and consistency. Recent advancements in large language models (LLMs), such as GPT-4, offer new opportunities to measure polarization using large-scale social media text. This study reviews 14 commonly used polarization metrics and constructs a dataset of 603,591 Weibo posts covering 20 high-profile public events. Using GPT-4-mini and Baidu’s sentiment analysis API, we compute and compare 14 polarization metrics. Results show that metrics sharing similar distributional characteristics tend to align, with the proportion of negative sentiment demonstrating the strongest correlation with expert evaluations. Furthermore, the composite weighted metric constructed by Lasso regression outperforms all individual metrics in both numerical value and ranking consistency. Future research may consider utilizing this more comprehensive metric.
Since the rise of social media platforms, professional athletes have been outspoken about the impact of online abuse on their mental health, with some stepping away from platforms to avoid toxic commentary. Researchers have examined how athletes’ own social media activity relates to their performance. This study investigates whether fans’ social media comments are associated with athletic performance. We analyze 1,249,107 Instagram comments directed at male professional soccer players during the 2022/2023 English Premier League season. Using web scraping, we collected player comment histories, applied sentiment analysis and topic modeling techniques, and aggregated comment features prior to matches. We then tested whether these features predicted player performance across selected match-day metrics, employing both linear regression models and machine-learning approaches. Our findings suggest that overall sentiment valence does not predict immediate on-field performance outcomes. However, player fixed effects models reveal that the content framing of comments, particularly critical commentary, predicts disciplinary actions within players, even after absorbing all stable player characteristics. This distinction between sentiment and content framing suggests that the relationship between fan commentary and athletic performance is more selective than previously assumed.