Influencer-centered communities on social media are connected through overlapping audiences and shared topical attention, yet their directed sentiment dependencies remain difficult to observe directly. We reconstruct inter-community influence networks from sentiment time series on Weibo, a major Chinese microblogging platform. We track several prominent technology bloggers over six months, aggregating comment-level sentiment into collective sentiment trajectories for each blogger’s commenting audience. These audiences are treated as macro-agents, an aggregation justified by shared content exposure and homophily-driven sentiment alignment. Using opinion dynamics models as network inference tools, we estimate directed influence graphs capturing how collective sentiment in one community shapes that in others. The inferred networks exhibit pronounced sparsity and clear structural asymmetry. Some communities act as hubs whose sentiment radiates outward, while others are largely internally driven, reacting to their blogger’s posts and weakly coupled to peer communities. These patterns remain consistent across model variants, and the learned models demonstrate stable two-period-ahead predictive performance on held-out data. Our findings show that collective online sentiment is well described by structured inter-community averaging, yielding interpretable, empirically grounded influence graphs. The proposed framework provides an interpretable approach for mapping inter-community sentiment dependencies from social media time series.
Social media platforms serve as spaces for both open expression and the spread of harmful narratives, often targeting marginalized communities. In the context of Hindi, automated detection of homophobic and transphobic comments at span level is still not studied widely. This work introduces a span-annotated dataset of Hindi social media comments, where identity-based toxic comments are labeled at the span level using character-based indices and encoded in BIO (Begin, Inside, Outside) format to support token-level classification. To address the scarcity of annotated resources, we adopt a human-in-the-loop framework that combines expert-labeled seed data with systematically generated synthetic data using GPT-4o prompting, followed by human validation to ensure cultural and contextual accuracy. We proposed novel augmentation strategies for annotation homophobic/transphopic at the span level which incorporates context-level, entity-level, combined context/entity, and noise-level strategies, enabling richer and more diverse training material. The dataset was built in three stages: a seed version, an intermediate augmented version, and a final version. We benched-marked our dataset with multiple prompting strategies, including zero-shot, few-shot, and augmentation-based approaches, across both vanilla and cooperative multi-agent framework. Our proposed multi-agent framework decomposes the detection process into specialized roles such as self-annotation, span-related feature extraction, demonstration scoring, and final prediction, resulting in a structured and interpretable pipeline for span-level homophobic/transphobic speech detection. Experiments are conducted on both the test split and the combined extended test set, with evaluation using strict token-level, hybrid, and span-overlap metrics to capture both exact and tolerant boundary matches. Our proposed multi-agent framework using augmented V3 data on the combined extended test set, achieved a Macro F1 score of 0.90, outperforming the strongest single-agent baseline and demonstrating the effectiveness of combining augmentation, multi-agent reasoning, and refined evaluation. Disclaimer: This study includes examples of toxic language strictly for research and educational purposes. The content does not reflect the views of the authors. Reader discretion is advised. GitHub Link: https://github.com/Unit-for-Inclusive-AI/Multi-Agent-HT-Span-Hindi
Social media platforms have increasingly become conduits for the dissemination of hate speech targeting the LGBTQ+ community, significantly undermining their mental health and overall well-being. This phenomenon presents a critical challenge within the field of Natural Language Processing (NLP). While hate speech detection is a well-established area of research, counter-narratives (CNs)-defined as polite, empathetic responses designed to neutralize harmful content-offer a promising alternative to traditional moderation. However, the availability of CN datasets remains severely limited for low-resource Dravidian languages, such as Tamil. In this paper, we introduce a newly curated dataset comprising 5,428 Tamil HS-CN pairs and propose VIVATHAM, a novel multi-agent persona debate framework for generating high quality Tamil counter-narratives. This dataset was developed through a human-in-the-loop (HITL) process, in which domain experts reviewed and refined initial CNs to ensure linguistic accuracy, politeness, and cultural sensitivity. Unlike conventional methods, our approach employs a system of agents with distinct personas to debate and iteratively refine responses from diverse perspectives. This debate-based mechanism facilitates the production of nuanced, non-aggressive, and contextually appropriate counter-narratives. Our results demonstrate that the multi-agent debate framework significantly outperforms standard CN generation approaches. This research contributes the first large-scale Tamil CN dataset and highlights the potential of multi-agent systems to enhance the quality and inclusiveness of online discourse. GitHub Link:https://github.com/Unit-for-Inclusive-AI/VIVATHAM.
Mental health detection from social media has gained increasing attention due to the spontaneous expression of psychological states on online platforms. However, accurate interpretation remains challenging because social media language is highly contextual, ambiguous, and often includes sarcasm, metaphors, or non-clinical expressions, which can lead to unreliable predictions. Existing approaches based on machine learning, deep learning, transformers, and large language models (LLMs) primarily rely on text-driven patterns and often lack explicit ontology-guided grounding, structured knowledge validation, and interpretability. Although LLMs improve contextual reasoning, their predictions can be inconsistent and prone to hallucination when not supported by structured domain knowledge and a unified cross-component validation mechanism. To address these limitations, this paper proposes an ontology-guided framework that integrates a Knowledge Graph (KG), an Ontology-Informed Retrieval Classifier (ORC), and a Large Language Model for interpretable mental health detection from social media text. The framework leverages the Human Phenotype Ontology (HPO) to guide symptom extraction and Knowledge Graph construction, normalizing informal linguistic expressions into ontology-grounded representations. A key design contribution is an explicit KG–ORC cross-validation consistency gate, which requires both the Knowledge Graph and the ORC module to independently reach agreement before a prediction is committed reducing erroneous inferences from either component alone. The LLM is deliberately restricted to a post-validation role, generating human-readable explanations and supportive recommendations only from validated outputs, thereby reducing hallucination risks and preserving knowledge-consistency. The proposed framework is evaluated on the Dreaddit dataset, a large-scale, multi-domain Reddit corpus for stress and mental health analysis. Empirical evaluation on the HPO guided symptom extraction confirms differential detection rates of 56.04
Community detection is a central problem in network science, traditionally performed on graphs, often via modularity maximization. Yet, many real-world systems inherently involve higher-order interactions between more than two entities, naturally modeled as hypergraphs. In such contexts, pairwise projections obscure higher-order structure and may lead to misleading communities. Recent works thus addressed higher-order community detection, with hypermodularity as one promising candidate. The state-of-the-art algorithm h-louvain [Kaminski et al., J. ComNetw 2024] mixes greedy optimization techniques for both modularity and hypermodularity in a multilevel scheme. While being effective with default parameters, the algorithm lacks guidance on how to combine the two objectives and thus needs parameter tuning. This parameter tuning step uses probabilistic strategies such as Bayesian Optimization Technique (BOT), resulting in a very significant overhead in running time. In this paper, we propose three algorithmic variants of h-louvain to recover communities of high accuracy without the need for time-consuming searches of mixing parameters. Our variants consist of new ideas for creating starting solutions for greedy multilevel algorithms and additional post-processing steps inspired by the community detection algorithm leiden [Traag et al., SciRep 2019] for graphs. In extensive experiments with over 50 real-world and randomly generated graphs, we show that across nearly all test cases, a member of our algorithm suite matches or surpasses h-louvain w. r. t. Asymmetric RMI; hence, they provide a more faithful community representation than the state of the art.
The positions users express during political discussions on social media function as acts of social identity work. While the ideological alignment between online sources and their audiences is well documented, whether this alignment extends to the rhetorical style they use to express their positions, remains largely unexplored. We conduct a large-scale analysis of immigration-related YouTube discussions spanning 2020–2024. We release a corpus of user comments and video transcripts span-annotated with divisive rhetorical techniques and develop token-level multi-label span detection models for both domains, enabling source–audience comparison via a unified rhetorical intensity measure. Commenting audiences are segmented using the combination of the channels where users choose to post their comments and their stance towards immigration. YouTube channels exhibit distinct rhetorical profiles shaped jointly by political leaning and source type. Audiences display cluster-specific rhetorical signatures related to the ideological composition of each group, while sharing a common macro-category hierarchy consistent with divisive rhetoric functioning as a domain-level communicative toolkit. We provide statistical evidence ( ρ̅= 0.475 ) of the rhetorical alignment between matched source and audience groups. Interpreted through the lens of rhetorical ecology, our findings are consistent with users and channel owners co-constitutively enacting, normalising and reproducing the rhetorical means through which identity work is performed, though the correlational design cannot adjudicate whether this reflects audience selection, source-driven influence, or both.
Terrorist organisations increasingly form dynamic collaboration networks whose evolution amplifies operational capability and security risk, yet existing computational approaches struggle to capture their adaptive, strategic, and interpretable nature. This paper proposes an explainable multi-agent reinforcement learning framework for modelling the temporal evolution of terrorist collaboration networks across multiple geopolitical regions. Collaboration formation is framed as a sequential decision-making process in which organisational agents learn cooperative and non-cooperative strategies from evolving network states. To ensure transparency and analytical value, the framework integrates a multi-layer explainability pipeline combining feature attribution, sensitivity analysis, counterfactual reasoning, domain-informed structural assessment, and interaction-based synergy analysis. The approach is evaluated on regional networks spanning Southeast Asia, South Asia, the Middle East and North Africa, Latin America, North America, and Sub-Saharan Africa. Experimental results demonstrate stable learning and early convergence across heterogeneous regions, while explainability analyses reveal structurally meaningful drivers of collaboration, region-specific causal factors, and emergent coalition effects. The findings show that explainable multi-agent reinforcement learning can move beyond static prediction toward interpretable, decision-aware analysis of evolving terrorist networks, providing actionable insights for network science and security-oriented applications.
Representing researchers in large-scale scholarly environments is increasingly challenging due to the rapid growth and interdisciplinary nature of scientific publication. Most existing approaches rely on observable relationships such as co-authorship or citation links, which often reinforce established connections and overlook latent topical structure. In this work, we construct topic-based scholarly networks using publication content alone, representing researchers through topic distributions derived from both probabilistic and embedding-based topic models, including LDA and BERTopic, to compare different modeling paradigms. A central challenge is publication imbalance, where dominant research themes of prolific authors overshadow their secondary or emerging interests. When publication content is aggregated at the researcher level, these dominant themes disproportionately shape the resulting representation, distorting similarity estimates and downstream community structure. To address this issue, we introduce a cloning-based representation that partitions a researcher’s publications into topic-specific components and represents each component as an independent node in the network. This enables researchers to participate in multiple topical communities and naturally induces overlapping community structure. Empirical analysis on institutional publication data, together with validation against co-authorship patterns, suggests that accounting for publication imbalance improves the alignment of topic-based representations with observed collaborative relationships and reveals latent interdisciplinary structure that is often obscured under standard aggregation.
Fairness in both machine learning (ML) and human decision-making is crucial, yet both are inherently prone to distinct forms of bias: algorithmic or data-driven in ML models, and subjective or inconsistency-related in humans. This study investigates fairness within the context of university admissions using a real-world dataset of 1,121 applicant profiles, of which 870 correspond to students with local qualifications. We evaluate three ML models, namely the Extreme Gradient Boosting (XGB), Bidirectional Long Short-Term Memory (Bi-LSTM), and k-Nearest Neighbors (KNN), enhanced with BERT embeddings to capture textual features from application materials. To assess individual fairness, we employ a consistency-based metric that measures the agreement between predictions made by ML models and evaluations from human experts of diverse backgrounds. Results demonstrate that ML models exhibit superior consistency compared to human evaluators, with improvements of more than 14
Digital Twin Technology (DTT) represents a cutting-edge paradigm enabling real-time digital replication of physical entities. The DTT research domain has seen exponential growth, particularly in the last decade, reflecting its wide-ranging applications across manufacturing, healthcare, smart cities, and beyond. This study aims to systematically map and visualise global research trends on DTT by identifying prolific authors, contributing institutions and countries, influential publications, and evolving keyword networks. It also explores the collaborative structure of research in this domain using advanced scientometric indicators. A longitudinal scientometric analysis was conducted on 30,327 cleaned and validated bibliographic records from the Scopus database using Biblioshiny and VOSviewer tools. The PRISMA 2020 protocol guided the data selection and cleaning process. Scientometric indicators like Degree of Collaboration (DC), Collaboration Index (CI), and Collaboration Coefficient (CC) were computed to assess collaboration patterns. The findings reveal a sharp rise in research output post-2015. China has emerged as a leading country, contributing nearly 25
The spread of hate speech has become a serious issue in online communities, negatively influencing user interaction and platform integrity. Although major social media companies have devoted substantial effort to developing automated detection systems, reliably classifying hateful content continues to be difficult. One key reason is shortcut learning, where models rely on frequent or sensitive words instead of understanding sentence meaning and context. Many hate speech detection systems achieve high accuracy while relying on surface-level lexical cues rather than robust semantic understanding. This shortcut-based behavior can lead to unstable predictions, particularly for neutral texts containing identity-related terms. In this paper, we analyze the prediction stability of transformer-based hate speech models when classifying neutral content. We construct a bilingual dataset of English and Norwegian social media posts annotated for binary hate speech detection and evaluate several Norwegian-specific and multilingual transformer models under identical training conditions. To mitigate shortcut learning, we introduce a neutral only stability regularization objective that encourages consistent predictions under stochastic perturbations. In post-training, we assess residual instability by applying controlled masking-based input variations through random token masking and measuring prediction variability. Our results show that Norwegian-specific models, especially Nor-BERT _base , produce more stable predictions for neutral content while maintaining competitive classification performance. In contrast, the best-performing multilingual model, mm-BERT _base , achieves the highest overall accuracy of 86
Community detection in social networks aims to group users into clusters or modules. These clusters are defined in such a way that the internal connections within clusters are dense and strong, while the connections between clusters are weaker. In community detection, influential nodes (seeds) may exhibit a higher number of connections both within communities (clusters) and between communities. Therefore, identifying influential nodes that have the ability to establish more connections than regular users can lead to the formation of clusters with higher modularity. In other words, the presence of influential nodes, which have a greater impact on the structure of communities, creates the necessary conditions for forming clusters with high modularity characteristics. In this paper, a novel method for community detection based on influence maximization using collective intelligence, named SICDIM, is proposed. In this method, the Manta Ray Foraging Optimisation (MRFO) algorithm is first used to solve the influence maximization problem in order to identify influential and independent nodes whose neighbors do not overlap. Then, the Penguin Search Optimisation Algorithm (PeSOA) is utilized for community detection from the clusters of users connected to influential nodes. This method improves the performance of community detection on several well-known datasets, including Karate, Dolphins, PolBooks, and Football, with modularity (Q) values of 0.2570, 0.33, 0.4737, and 0.4692, and normalized mutual information (NMI) values of 0.5589, 0.398, 0.6228, and 0.3904, respectively. Moreover, the SICDIM method outperforms the PeSOA algorithm in various aspects such as modularity, NMI, precision, and recall, particularly in community detection.
In the digital era, the internet and social media have emerged as essential platforms for individuals facing mental health issues, often used for seeking information and community support. Despite the resources of informal advice available on social media, the complexity of these issues frequently exceeds non-expert knowledge. Specialized sites such as CounselChat and 7Cups offer professional guidance, yet many at-risk individuals still rely on unmoderated sources and general web search. We address this gap by investigating ranking strategies that match pre-existing expert advice to incoming mental-health questions. We introduce CounselingQA, a collection built from two specialized websites, pairing user questions with verified expert responses. We address the task as answer retrieval (AR): given a question, rank expert answers by relevance. We evaluate dense retrieval with SentenceBERT and MentalBERT, and propose a second stage that improves the initial ranking via transformer-based models and large language models (LLMs), used for filtering non-relevant candidates and for reordering. Beyond retrieval, we analyze linguistic style and affective attributes across topics, questions, and responses. Results show that dense retrieval provides strong candidates and that transformer/LLM-driven reranking further elevates relevant, on-topic advice to the top positions. We further conduct qualitative error analyses, including human evaluation to study the benefits and limitations of our approaches. Taken together, these findings indicate that retrieval-first pipelines can help scale access to professional guidance.
This study examines university students’ attitudes toward Instagram and transformations in privacy norms, and as well as with sociodemographic characteristics. In the study, a literature review was conducted, and the Instagram and “Privacy Transformation Scale” prepared by Kalaman was applied to volunteer participants with permission from the researcher. Statistical analyses were performed using the IBM SPSS Statistics 26 program. As a result of the study, the Kolmogorov–Smirnov test showed that the numerical variables were approximately normally distributed, so parametric tests were applied in the study. Descriptive statistics (frequencies, percentages, means, and standard deviations) were computed for participant demographic variables. Results indicated that male students exhibited higher levels of awareness regarding Instagram-related privacy transformations compared to female students. Participants aged 18–20 had lower mean privacy-transformation scores than older age groups. A high proportion of respondents reported sharing personal information (e.g., date of birth, place of residence, educational status, relationship status, occupation, and hometown) on Instagram.
As artificial intelligence (AI) reshapes educational and technological landscapes, understanding public online discussions may provide valuable insights into how AI and education are perceived, debated, and interpreted. This study investigated public discourse on AI and education by analyzing 49,641 posts and comments from Reddit and the Artificial Intelligence Stack Exchange, revealing how online communities shape perceptions of AI- and education-related sentiment and practices. Using a mixed-methods approach, combining computational grounded theory and qualitative constant comparison analysis, we identified five key themes: (1) AI Augments Teaching for Pedagogical Innovation, (2) Educators Advocating for Ethical AI Practice, (3) Students Develop Thinking Skills for Knowledge Construction, (4) Equipping Students with AI literacy for a Digitally Transformed Future, and (5) Human Connection and Collaboration for Human-AI Synergy in Learning. These findings highlighted social, ethical, and collaborative dimensions, with implications for policy and digital literacy development. We discussed implications for policy, curriculum design, and institutional strategy to ensure responsible and impactful adoption of AI technologies.
Link prediction aims to identify missing or future connections between entities of a complex system when the latter is modeled as a network. This research problem has attracted significant attention due to its relevance in numerous fields. In this work, we propose Network Centrality Link Prediction (NCLP), a novel approach that integrates various network centrality metrics in order to predict the likelihood of future connections between entities of a system. Specifically, we consider weighted betweenness, closeness, and Katz centralities, in addition to the Resource Allocation and Adamic–Adar indices. A genetic algorithm is employed to optimize the weights of these metrics, reflecting their contributions to link prediction. The proposed approach is evaluated on several real-world benchmark networks with varying structural characteristics. Experimental results demonstrate that our method achieves superior predictive performance compared to state-of-the-art link prediction techniques. These findings highlight the effectiveness and robustness of integrating centrality-based measures through an optimization-driven approach, emphasizing the potential of hybrid methodologies for improving link prediction accuracy across diverse network domains. Data and code related to the proposed approach can be found at the following GitHub repository: https://github.com/elissalichaaelkhoury/NCLP-Link-Prediction.
Fractal dimension can measure the complexity of a branching structure. The fractal dimension of branching in botanical trees, for example, can indicate unhealthy states when they are too sparse or too entangled. We hypothesize that the branching structures of online conversations can also use fractal dimension to measure sparse versus complex branching, with implications for conversational “health”. To test this we measured the fractal dimension of Reddit posts about AI. Posts about purely technical content (e.g. distinctions between different algorithms) had lower fractal dimension than those about social controversies (job loss, racial bias, etc.), suggesting that the controversial conversations had more complex branching structures. A sentiment analysis revealed that social posts had more negative sentiment, consistent with characterizing them as more controversial. We also found that even within each category (social vs. technical), higher fractal dimension was associated with more negative sentiment. The fractal model offers further insights when considering its analogous biological models. While it is common to use the metaphor of “conversation tree” we find that fractal metrics reveal a structure closer to Diffusion Limited Growth, found in bacteria colonies, fungi, and rhizomatic plant spread, where “sub-trees” can vary in fractal dimension from the parent. The fractal dimension of social controversy subtrees have a stronger regression slope sensitivity to that of the parent than do the technical subtrees, which has potential implications for the semantic process differences. Overall the application of fractal models to online conversations shows that it allows correlations between structural and semantic aspects, and offers a new way to illuminate the underlying characteristics.
Structured tax fraud schemes frequently rely on complex relational arrangements designed to obscure beneficial ownership and coordinated economic behavior. This study investigates whether graph-derived structural and weighted descriptors can support supervised tax fraud detection in ego-centered fiscal relationship networks. Graph-level features capturing compactness, local cohesion, path organization, and weighted relational influence were extracted from institutional investigation graphs and evaluated using multiple supervised classifiers. Additionally, graph convolutional networks (GCN) and GraphSAGE (graph sample and aggregate) were evaluated as graph-native benchmark models operating directly on the original relational structures. Exploratory analyses revealed a dominant structural axis separating compact and distance-driven configurations between fraudulent and non-fraudulent networks. Among the evaluated models, XGBoost achieved the highest cross-validated F1-score and strong ROC–AUC discrimination, while GraphSAGE achieved competitive performance among graph-native approaches. Results suggest that interpretable graph-derived descriptors capture substantial predictive relational information embedded in the investigated fiscal networks. These findings highlight the potential of graph-based descriptors as auditable tools for tax fraud detection in institutionally sensitive environments.
The dissemination of information at scale on social media has profound real-world consequences, necessitating advanced methods for cascade graph mining to identify and mitigate the spread of harmful content. Cascade classification, which involves categorizing information diffusion patterns based on their structural characteristics, is a fundamental task in this domain. While Contrastive Cascade Graph Learning (CCGL) has emerged as a promising self-supervised framework, its effectiveness and robustness for cascade classification remain under-explored, particularly in label-scarce scenarios. In this study, we provide a comprehensive evaluation of CCGL for classifying both synthetic cascades (generated by diverse network and diffusion models) and real-world cascades from multiple social media platforms. We further benchmark CCGL against representative baselines, including DeepHawkes and CasFlow. Our experimental results demonstrate that CCGL achieves strong performance, particularly on large-scale and structurally complex real-world cascades, where it outperforms representative baselines. However, the results also show that this advantage is conditional: sequence-based and cascade modeling baselines are competitive or superior in some settings, such as small citation-based cascades and synthetic network-model classification. Furthermore, we find that CCGL maintains high classification accuracy even with only 20
GRAFT-Rec (Graph-Augmented Fusion with Gradient-Boosted Recommendation) is a hybrid meta-ranking framework that combines heterogeneous collaborative, graph-based, content-derived, and popularity-aware signals for top-K recommendation. In response to the reviewers, the revised framework removes outdated language-model prompting and LightGCN-fusion descriptions, integrates RP3 β as an internal graph-based candidate and residual signal, and selects the final architecture exclusively by validation performance. The final MovieLens-1 M configuration uses a NoBiRank meta-feature variant with RP3 β residual blending and activity-stratified adaptive mixing. On MovieLens-1 M, GRAFT-Rec achieves P@10 = 0.0108, HR@10 = 0.1080, NDCG@10 = 0.0570, and MRR@10 = 0.0419 on 1,037 held-out users, outperforming RP3 β (NDCG@10 = 0.0396) by 43.95
