Source identification in propagation graphs is an inverse problem that aims to recover the initial infected or activated nodes from an observed propagation snapshot. This paper studies a static, structure-driven setting in which temporal traces, observer timestamps, and large labeled diffusion histories are unavailable. Voting and ranking-fusion methods are often motivated by the ensemble-learning intuition that several imperfect signals can be more useful after aggregation than any single signal alone. The study tests whether this intuition also helps multi-source localization on propagation graphs, where the individual signals are structural node heuristics rather than learned classifiers. The evaluation covers ten structural heuristics, four ensemble aggregation families, and four implemented source-localization baselines across 22 network-by-propagation cases and 198 saved propagation instances. The analysis separates fixed-method comparisons from diagnostic headroom, compares structural ensembles with non-structural source-localization baselines, and examines whether source counts can be estimated from score gaps or score thresholds rather than supplied as an oracle k. The results support a qualified conclusion. Structural ensembles expose a statistically reliable direction of improvement over selected single-heuristic references, but practical fixed-method gains are modest and depend on how member weights or configurations are chosen. Average-distance and topological centrality are the strongest recurring structural signals, suggesting that future methods should model proximity and graph-position evidence more explicitly rather than enumerate more voting combinations. Adaptive epsilon-gap and threshold rules partially control source-count error, but they do not remove the source-count problem under static snapshots. Overall, ensemble ranking fusion is useful as a calibrated diagnostic and sometimes as a practical aggregation layer, while richer temporal information, validation data, or more principled source-count estimators are still needed for robust deployment.
Information propagation in Online Social Networks (OSNs) is commonly modeled as a diffusion process based on fixed interaction rules. However, such models are limited in their ability to capture adaptive and context-dependent interactions. In this work, we propose an attention-based formulation of information propagation in which influence emerges from dynamic interactions between node representations. Rather than following predefined transmission rules, propagation is governed by compatibility between node states and evolves over time. Experiments on synthetic networks show that the proposed model exhibits non-monotonic activation dynamics, in contrast to classical cascade processes. A key result is the emergence of distinct propagation regimes controlled by a structural parameter, which governs a transition from globally distributed, high-entropy interactions to localized, low-entropy dynamics. These findings indicate that propagation can be understood as a process of dynamic attention redistribution rather than cumulative transmission. As a result, the proposed approach provides a unified and flexible approach to modeling information propagation, generalizing classical diffusion models and capturing a broader range of behaviors.
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.
