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.
