Comparing structurally identical graphs is critical for evaluating network models, particularly when a forecasted network needs to be evaluated against the observed network, and compared to other forecasting models. For example, forecasting military conflict in a spatial grid. Existing metrics for comparing these models are limited, even moreso when the network contains probabilistic attributes and we want a proper scoring rule to take uncertainty into account. To address this challenge, we introduce a framework for quantifying differences between graphs whose nodes are associated with probability functions, as in conflict forecasting. Our approach uses a multi-component distance measure based on the dissimilarity between corresponding node-level distributions and the distributional shift in relationships across connected nodes. We use our method to compare and evaluate vertex-distribution networks in theoretical and real-world applications. The findings reveal that nodelevel and relational components provide distinct yet complementary patterns, supporting their practical use in comparing conflict forecasting models from public competitions.
Cyberbullying on social networking sites has become more prevalent. Most cyberbullying detection models often lack consideration of adversarial threads, leaving them vulnerable. This study evaluates the resilience of text-based cyberbullying detection models, constrained by limited available datasets, against word-level substitutions and character-level perturbations. We consider well-established ML techniques with real-world data and more recent LLM-based approaches to uncover model weaknesses. The results reveal that adversarial attacks can significantly reduce detection accuracy, e.g., most models are vulnerable to word- and character-level attacks with success rates up to 88% and 44%, respectively. We also find that LLM-based models such as CyberBERT are more resistant to both types of attack while maintaining strong detection performance. We show that model architecture and text vectorization choices significantly impact attack resistance and that adversarial training can help improve robustness, with tailored combinations of models and vectorizers showing the best results. These findings can guide the development of safer online platforms, as tailored strategies can make cyberbullying detection models more resilient and effective.
In this paper we consider what remains to create a unified theory of cognition, updating Newell's [1] report. We examine ACT-R as a new exemplar. After quickly summarizing what has been covered, we create a new schematic for ACT-R to add capabilities and mechanisms for a 5 to 10-year research plan. This includes more meso-level features of cognition and high-level aspects of cognition (that will inevitably increase the functionality while decreasing the usability of ACT-R unless features can be turned off). We suggest integrating more aspects of physiology and psychology as well as more social aspects. To accomplish this, we introduce six new functional modules. These additions enable the architecture to account for physiological and emotional conditions, stable personal traits, value-based evaluations, experiential learning, and gradual behavioral change. As a result, this extended ACT-R architecture offers a more comprehensive and flexible model of human cognition, bridging previously underrepresented areas of psychology, physiology, and behavior.
dSocial Identity Theory" is one of the most influential ideas in social science, but identity is notoriously difficult to operationalize and observe. In a previous project, we introduced a method for studying identity at massive scale by longitudinally observing a close proxy: social media bios. Here, we discuss a new variation on the method, which focuses on identity transitions: the amendment of an identity following a particular event. We illustrate the method through a test case taken from millions of Twitter bios from 2015-2023: patterns in the addition and deletion of the acronym "MAGA" ("Make America Great Again"). We show that certain categories of bio words at one time point can predict the addition of "MAGA" at a later point. We also examine the sorts of words that tended to replace "MAGA" when users amended their bios following the insurrection of January 6th, 2021. Finally, we discuss potential future applications of the method, focusing on the topic of "stigma exits."
Deception detection remains a critical challenge for security, human-centered computing, and behavioral research, yet most existing systems struggle to generalize across real-world contexts. This paper examines multimodal deception detection across four domainsacademics, biography, well-being, and crime. Using 400 videos from two datasets, including a new one we have released, we evaluate facial, affective, gaze, and gesture features in single- and cross-domain settings. Gaze and gesture consistently outperform facial and affective features, especially in high-risk or narrative contexts, despite being selected less often. Facial features, though common, often lack predictive value, underscoring the need to prioritize performance over frequency. We find that expressivity, not response length, improves detectability, and generalization is stronger when domains share risk or narrative structure. These insights support the design of adaptable, context-aware deception detection systems.
Large Vision Language Models (LVLMs) stand to catalyze transformative advances in social computing by enabling deep multi modal understanding, precise retrieval, and contextually informed generation. Existing LVLMs excel at understanding video content but incur prohibitive computational and memory costs due to dense, fixed-rate sampling. To capture fine-grained visual details, most LVLMs sample at 1-2 frames per second, yielding an unmanageable frame volume for long videos. Although various sampling optimizations have been proposed, finding the optimal balance between efficiency and accuracy remains elusive. We introduce EVRAG: Enhanced Video Retrieval-Augmented Generation, a training-free pipeline that leverages scene-change detection to select one representative frame per scene, thereby drastically reducing frame volume without degrading performance. To support robust information retrieval, EVRAG employs ChromaDB to construct an auxiliary index of supplementary frames, enabling sub-millisecond query response times and seamless interoperability with existing LVLM architectures. This optimized indexing yields a minimal resource footprint, permitting deployment on cost -constrained and low-energy platforms and thus broadening access to advanced video language AI. By lowering infrastructural barriers, EVRAG democratizes video-based social computing and enables organizations of all scales to perform real-time multimedia analysis. Experimental results demonstrate that EVRAG achieves performance on par with leading models while effecting multi -order -ofmagnitude reductions in both computational and memory overhead.
This study examines the persistence likelihood of 4,946,718 Twitter mention ties between U.S. Twitter users by linking a pre-2015 corpus with a COVID-era corpus (March 2020June 2022). Our analysis reveals how geographic distance, pre-existing tie strength, and local network structure jointly shape long-term tie survival. Key findings include: (1) systematic decay in survival rates with respect to geographic distance, from 16.4% at 05 miles to 9.9% beyond 2000 miles; (2) powerful protective effects of tie strength, with survival rising from under 1% in the weakest to 17% in the strongest quantile; (3) a null effect of reciprocity, measured by balance in mention frequency; and (4) substantial statelevel heterogeneity unexplained by geography or demographics. Urban environments amplify the protective effect of tie strength, while the lack of an independent imbalance effect challenges conventional reciprocitybased models of tie persistence. These results demonstrate how strong pre-existing bonds create resilient connections that can bridge geographic and temporal gaps, suggesting that long-surviving ties may function as bridges connecting increasingly divergent social worlds.
Recommendation systems, while optimized for personalization, can inadvertently create recommendation traps. In this study, we investigate such traps within YouTube's recommendation network using the Cheng Ho propaganda dataset, a case rich in sociopolitical relevance. We apply Focal Structure Analysis (FSA) to identify dense subgraphs and compare them against structural groupings derived from various centrality-based and community detection methods. To evaluate the behavioral pull of these structures, we introduce a metric named attraction, which captures how easily a subgraph is reached from outside. Random walk simulations demonstrate that focal structures exhibit stronger attraction than other group types. To assess semantic coherence, we apply a weighted topical uniformity framework using BERTopic, revealing that focal structures also tend to exhibit higher thematic consistency. Our findings suggest a strong alignment between structural cohesion and content homogeneity, offering a novel lens for identifying and analyzing recommendation traps in algorithmic systems.
Collecting social network data is often an imperfect process bearing missing connections that past research has shown can harm the accuracy of metrics calculated on the resulting network. This study examines the problem of constructing confidence intervals for the true value of a network metric when edges are known to be missing from a network. By comparing various methods for bootstrapping networks through simulated experiments, this research shows that different network metrics and different network types benefit from different methodologies for bootstrapping confidence intervals. Modified link prediction algorithms improved interval quality for average degree and betweenness centrality on small-world networks, whereas scale-free networks and triangle density called for the use of approaches based on random dot product graphs. Network scientists relying on data with missing links should therefore carefully consider their particular circumstances when seeking to quantify the uncertainty in their analyses.
The role of conversational context in the prediction of toxicity in the cross-platform remains underexplored across digital environments. This study investigates whether comment toxicity patterns from immediate parents (P), grandparents (GP), and great-grandparents (GGP) predict toxic responses differently in the COVID-19 Reddit health discourse (1.2 million comments, 23,000 threads) and the Russia-Ukraine Telegram political discourse (518 channels, 5.2 million posts). Using ensemble machine learning with 5-fold cross-validation, we achieved F1 scores of 68-77% across platforms, confirming the predictive validity between platforms. Statistical analysis revealed universal and contextdependent patterns with meaningful effect sizes: immediate parent influence showed consistent patterns across platforms for extreme toxicity (Cohen's d=0.24 d = 0.24), while moderate toxicity exhibited substantial context-dependent variations. In particular, the influence of the grandparents achieved a better performance in moderate political toxicity (74% F1) compared to the patterns dominant by the parents in health contexts (68% F1), representing a large effect size ( d=0.71 d = 0.71) and a 6-percentage-point performance difference. Cross-validation confirmed these patterns in 4 toxicity categories with 87% thread reconstruction accuracy for Telegram conversations. These findings establish a validated framework for context-dependent generational influence analysis, providing empirical evidence for discourse-specific behavioral patterns in online toxicity prediction.
This paper analyzes emotional expression patterns in over 653 million Twitter messages from the first six months of the COVID-19 pandemic (February-July 2020). We employ natural language processing and Fourier analysis to identify eight primary emotions plus loneliness in tweets containing pandemic-related keywords. Through Fourier analysis, the study reveals distinct weekly (7-day) and monthly (29-day) periodicities in emotional expressions, with loneliness uniquely operating outside typical weekly cycles. The study uncovers unexpected differences between human and automated accounts: contrary to expectations, bot messages contained fewer multiple co-occurring emotional cues than human posts, but for single emotion usage tended to mimic their respective human category - except for news-bots which amplified more fear messaging cues on average. These findings demonstrate how emotional expressions function as coordination mechanisms during crisis events and highlight the strategic role of bots in information dissemination across social media ecosystems during the pandemic.
Ransomware attacks have become a major threat to organizations worldwide, leveraging both technical vulnerabilities and social engineering to infiltrate target networks. While many security research teams provide extensive information about the tactics, techniques, and procedures (TTPs) used by threat groups and campaigns, there is still limited insight into the actual attack flows and transition patterns between techniques. In this study, we selected 19 MITRE ATT&CK groups and campaigns with confirmed histories of ransomware activity and systematically modeled their attack flows as network graphs. By integrating these into a unified ransomware attack flow network, we performed comprehensive analyses including hub centrality, betweenness centrality, and frequency analysis. Through this analysis, we identified nine recurring patterns within the integrated ransomware attack network. We also found techniques related to discovery and tool transfer exhibited high hub centrality and techniques associated with execution were found to have high betweenness centrality. Based on these findings, we of network-based analysis in identifying key vulnerabilities and improving organizational preparedness against evolving ransomware threats.
This study examines the impact of symbolic content in TikTok videos on trust formation during Taiwan's 2024 presidential election, addressing key questions about digital political communication and democratic discourse. We analyzed TikTok videos using a novel dualframe approach, comparing the first frames with key frames to determine the optimal methods for extracting symbolic information. Through advanced Large Language Model analysis, we systematically detected and categorized social, cultural, and political symbols, measuring their respective information gains and trust-building effectiveness. Our findings reveal that standout frames provide significantly higher information gain for non-symbolic content compared to symbolic content, where first frames prove more effective, particularly for social symbols. Trust analysis reveals that comprehensive symbolic representation yields the highest trust scores, with cultural and social symbols substantially outperforming political symbols in fostering democratic confidence. These results provide evidence-based insights for enhancing democratic communication strategies and combating disinformation through strategic symbolic messaging that strengthens rather than undermines public trust in electoral processes.
Trolling", defined as posting online with the goal of upsetting people, mostly for one's own amusement, is an activity linked with the personality trait of everyday sadism. Given the importance of online communities to the MAGA movement, support for Trump within known groups of trolls, and the challenges large pro-Trump communities have faced on mainstream social media platforms because of their frequent Terms of Service violations, we investigate the connection between trolling, online sadism, and voting behavior. Our results show that people who voted for Trump in 2024 scored significantly higher on trolling behavior and online sadism than Harris voters with average differences of over 30% between the two groups. These results have implications for content moderation, managing online harassment, and the online-to-offline violence pipeline.
This paper examines a new approach to traditional media literacy training: encouraging social media users to engage in social corrections or other countermeasures online when they encounter misinformation. Unlike most prior media literacy efforts that aim to improve accuracy discernment, this work focuses on increasing users' willingness to actively intervene. An experiment involving government analysts was conducted, where participants viewed a series of false or misleading social media posts and described how they would respond. Surveys were administered before and after an in-person, interactive training session. After the training, participants reported being more likely to intervene with greater effort when faced with misinformation, such as by commenting publicly or messaging the poster privately. However, this increase was mainly among those already engaging in low-effort actions, like reporting misinformation. Additionally, a qualitative analysis showed that factors such as the content of the post, the type of account posting the misinformation, and the platform all influence participants' willingness to intervene. Respondents reported being significantly more likely to counter closer contacts than less close ones. The promising findings from this case study can help inform future platform design to better promote individual interventions against misinformation.
Large Language Models (LLMs) are increasingly being used in sensitive domains such as healthcare and education, where safety is critical. While much LLM safety research focuses on deliberate attacks (e.g., jailbreaks, prompt injections), the impact of benign but imperfect user input, such as typos or paraphrasing remains underexplored. In this study, we investigate how such semantic-preserving perturbations affect the safety behavior of aligned LLMs. We find that these small perturbations can increase instability, e.g., causing unsafe responses to flip into safe refusals. This bidirectional instability reveals that current safety alignment mechanisms are fragile and context-dependent. Using a wide set of realistic perturbations in harmful queries, we evaluated several open-source models. Our results highlight the need for more contextaware and model-sensitive evaluation frameworks and training methods that ensure robust behaviors in the face of natural and noisy input. Github: https://github.com/Ztrimus/llm-sensitivity.
As Large Language Models (LLMs) become more sophisticated, there is a possibility to harness LLMs to power social media bots. This work investigates the realism of generating LLM-Powered social media bot networks. Through a combination of manual effort, network science and LLMs, we create synthetic bot agent personas, their tweets and their interactions, thereby simulating social media networks. We compare the generated networks against empirical bot/human data, observing that both network and linguistic properties of LLM-Powered Bots differ from Wild Bots/Humans. This has implications towards the detection and effectiveness of LLM-Powered Bots.
Social network motifs are recurring patterns of small subgraphs that indicate fundamental patterns of social communication. In this work, we study the simple star network motifs that recur on X during the COVID-19 discourse. We study the profile of the manifestation of the star network among bot and human users. There are six primary patterns of the star motif, differentiating by the bots and humans being either egos and alters. We describe the presentation of each of these six patterns in our data, demonstrating how the motif patterns can inform social media behavioral analysis.
Attempts to manipulate webgraphs can have many downstream impacts, but analysts lack shared quantitative metrics to characterize actions taken to manipulate information environments at this level. We demonstrate how the BEND framework can be used to characterize attempts to manipulate webgraph information environments, and propose quantitative metrics for BEND community maneuvers. We demonstrate the face validity of our proposed Webgraph BEND metrics by using them to characterize two small web-graphs containing SEO-boosted Kremlin-aligned websites. We demonstrate how our proposed metrics improve BEND scores in webgraph settings and demonstrate the usefulness of our metrics in characterizing webgraph information environments. These metrics offer analysts a systematic and standardized way to characterize attempts to manipulate webgraphs using common Search Engine Optimization tactics.
The rapid rise of Generative AI (GenAI), particularly LLMs, poses concerns for journalistic integrity and authorship. This study examines AI-generated content across over 40,000 news articles from major, local, and college news media, in various media formats. Using three advanced AI-text detectors (e.g., Binoculars, Fast-Detect GPT, and GPTZero), we find substantial increase of GenAI use in recent years, especially in local and college news. Sentence-level analysis reveals LLMs are often used in the introduction of news, while conclusions usually written manually. Linguistic analysis shows GenAI boosts word richness and readability but lowers formality, leading to more uniform writing styles, particularly in local media.
