The increased use of large language models (LLMs) in sensitive domains leads to growing interest in how their confidence scores correspond to fairness and bias. This study examines the alignment between LLM-predicted confidence and human-annotated bias judgments. Focusing on gender bias, the research investigates probability confidence calibration in contexts involving gendered pronoun resolution. The goal is to evaluate if calibration metrics based on predicted confidence scores effectively capture fairness-related disparities in LLMs. The results show that, among the six state-of-the-art models, Gemma-2 demonstrates the worst calibration according to the gender bias benchmark. The primary contribution of this work is a fairness-aware evaluation of LLMs confidence calibration, offering guidance for ethical deployment. In addition, we introduce a new calibration metric, Gender-ECE, designed to measure gender disparities in resolution tasks.

Health misinformation poses a significant public threat by eroding trust in scientific expertise and diminishing adherence to health guidelines, which collectively weaken community resilience to preventable diseases. For these reasons, detecting health misinformation is crucial to protect public health. However, manual detection requires substantial human effort and expertise, making it impractical at scale, particularly in low-resource settings where technological and linguistic resources are limited. Developing automated techniques for identifying false or misleading claims is therefore essential to ensure timely intervention. Advancing these automated detection methods depends on the development of robust datasets, as they enable more accurate modeling and adaptation for specific languages and contexts. To the best of the authors’ knowledge, no misinformation detection techniques or datasets have yet been developed specifically for the Estonian language within the health domain. Addressing this gap, the primary objective of this study is to develop a reliable system for generating ground truth labels for health misinformation in Estonian, thereby contributing to misinformation detection in low-resource settings. Leveraging pre-labeled datasets in English, the proposed Cross-lingual Alignment and Confident Prediction Sampling (CAPS) approach employs a hybrid two-phase methodology involving semantic similarity measurements, manual annotation, classification, and confidence sampling. This methodology enables the efficient generation of misinformation labels with minimal reliance on manual annotation, contributing a valuable resource for advancing misinformation detection in underrepresented languages. The resulting dataset of 8,795 annotated news articles represents a significant advancement in health misinformation detection for the Estonian language.

The propagation of aggressive behavior in online social networks presents a growing threat to digital well-being and social harmony. While existing research focuses on modeling aggression diffusion or detecting aggressive content, forecasting individual user aggression remains an open challenge. This work fills this gap by introducing Temporal Social Graph Attention Network (TSGAN), a social-aware sequence-to-sequence architecture designed to forecast aggressive behavior in dynamic social networks. The core of TSGAN is an adaptive socio-temporal attention module that dynamically models social influence and temporal dynamics. To capture global social influence, TSGAN employs a graph contrastive learning approach to generate global network context embeddings. TSGAN utilizes an aggression intensity metric derived from a proposed hybrid aggression content detection model (92.87% F1), combining a fine-tuned transformer with a large language model to quantify user aggression over time. TSGAN uniquely addresses user inactivity, models dynamic follower relationship impacts, and accounts for temporal behavioral decay while scaling to large networks. Experiments on real-world datasets (X for aggression forecasting and Flickr for popularity prediction) demonstrate TSGAN’s versatility and effectiveness. TSGAN outperforms baselines in forecasting across hourly, daily, and weekly temporal intervals, showing up to 24.8% improvement in daily aggression predictions.

In this study, we investigate multimodal foundation models (MFMs) for emotion recognition from non-verbal sounds. We hypothesize that MFMs, with their joint pre-training across multiple modalities, will be more effective in non-verbal sounds emotion recognition (NVER) by better interpreting and differentiating subtle emotional cues that may be ambiguous in audio-only foundation models (AFMs). To validate our hypothesis, we extract representations from state-of-the-art (SOTA) MFMs and AFMs and evaluated them on benchmark NVER datasets. We also investigate the potential of combining selected foundation model (FM) representations to enhance NVER further inspired by research in speech recognition and audio deepfake detection. To achieve this, we propose a framework called MATA (Intra-Modality Alignment through Transport Attention). Through MATA coupled with the combination of MFMs: LanguageBind and ImageBind, we report the topmost performance with accuracies of 76.47%, 77.40%, 75.12% and F1-scores of 70.35%, 76.19%, 74.63% for ASVP-ESD, JNV, and VIVAE datasets against individual FMs and baseline fusion techniques and report SOTA on the benchmark datasets.

Rhymes and poems are a powerful medium for transmitting cultural norms and societal roles. However, the pervasive existence of gender stereotypes in these works perpetuates biased perceptions and limits the scope of individuals’ identities. Past works have shown that stereotyping and prejudice emerge in early childhood, and developmental research on causal mechanisms is critical for understanding and controlling stereotyping and prejudice. This work contributes by gathering a dataset of rhymes and poems to identify gender stereotypes and propose a model with 97% accuracy to identify gender bias. Gender stereotypes were rectified using a Large Language Model (LLM) and its effectiveness was evaluated in a comparative survey against human educator rectifications. To summarize, this work highlights the pervasive nature of gender stereotypes in literary works and reveal the potential of LLMs to rectify gender stereotypes. This study raises awareness and promotes inclusivity within artistic expressions, making a significant contribution to the discourse on gender equality.

Social media platforms, particularly Twitter, have emerged as vital media for organizing online protests worldwide. During protests, users on social media share different narratives, often coordinated to share collective opinions and obtain widespread reach. In this paper, we focus on the communities formed during a protest and the collective narratives they share, using the protest on the enactment of the Citizenship Amendment Act (#CAA) by the Indian Government as a case study. Since #CAA protest led to divergent discourse in the country, we first classify the users into opposing stances, i.e., protesters (who opposed the Act) and counter-protesters (who supported it) in an unsupervised manner. Next, we identify the coordinated communities in the opposing stances and examine the collective narratives shared by coordinated communities of opposing stances. We use content-based metrics to identify user coordination, including hashtags, mentions, and retweets. Our results suggest mention as the strongest metric for coordination across the opposing stances. Next, we decipher the collective narratives in the opposing stances using an unsupervised narrative detection framework and found call-to-action, on-ground activity, grievances sharing, questioning, and skepticism narratives in the protest tweets. We analyze the strength of the different coordinated communities using network measures, and perform inauthentic activity analysis on the most coordinated communities on both sides. Our findings also suggest that coordinated communities, which were highly inauthentic, showed the highest clustering coefficient towards a greater extent of coordination.

Misinformation has emerged as a major societal threat in the recent years in general; specifically in the context of the COVID-19 pandemic, it has wrecked havoc, for instance, by fueling vaccine hesitancy. Cost-effective, scalable solutions for combating misinformation are the need of the hour. This work explored how existing information obtained from social media and augmented with more curated fact checked data repositories can be harnessed to facilitate automated rebuttal of misinformation at scale. While the ideas herein can be generalized and reapplied in the broader context of misinformation mitigation using a multitude of information sources and catering to the spectrum of social media platforms, this work serves as a proof of concept, and as such, it is confined in its scope to only rebuttal of tweets, and in the specific context of misinformation regarding COVID-19. It leverages two publicly available datasets, viz. FaCov (fact-checked articles) Sharma et al., 2022 and misleading (social media Twitter) Sharma et al., 2024 data on COVID-19 vaccination.

In this work, we collect a moderate-sized representative corpus of tweets (over 200 000) pertaining to COVID-19 vaccination spanning for a period of seven months (September 2020–March 2021). Following a transfer learning approach, we utilize a pretrained transformer-based XLNet model to classify tweets as misleading or nonmisleading and manually validate the results with random subsets of samples. We leverage this to study and contrast the characteristics of tweets in the corpus that are misleading in nature against non-misleading ones. This exploratory analysis enables us to design features such as sentiments, hashtags, nouns, and pronouns which can, in turn, be exploited for classifying tweets as (non-)misleading using various machine learning (ML) models in an explainable manner. Specifically, several ML models are employed for prediction, with up to 90% accuracy, with the importance of each feature is explained using SHAP Explainable AI (XAI) tool. While the thrust of this work is principally exploratory in nature to obtain insight on the online discourse on COVID-19 vaccination, we conclude the article by outlining how these insights provide the foundations for a more actionable approach to mitigate misinformation. We have made the curated data as well as the accompanying code available so that the research community at large can reproduce, compare against, or build upon this work.

Fact-checking is a crucial task as it ensures the prevention of misinformation. However, manual fact-checking cannot keep up with the rate at which false information is generated and disseminated online. Automated fact-checking by machines is significantly quicker than by humans. But for better trust and transparency of these automated systems, explainability in the fact-checking process is necessary. Fact-checking often entails contrasting a factual assertion with a body of knowledge for such explanations. An effective way of representing knowledge is the Knowledge Graph (KG). There have been sufficient works proposed related to fact-checking with the usage of KG but not much focus is given to the application of reinforcement learning (RL) in such cases. To mitigate this gap, we propose an RL-based KG reasoning approach for explainable fact-checking. Extensive experiments on FB15K-277 and NELL-995 datasets reveal that reasoning over a KG is an effective way of producing human-readable explanations in the form of paths and classifications for fact claims. The RL reasoning agent computes a path that either proves or disproves a factual claim, but does not provide a verdict itself. A verdict is reached by a voting mechanism that utilizes paths produced by the agent. These paths can be presented to human readers so that they themselves can decide whether or not the provided evidence is convincing or not. This work will encourage works in this direction for incorporating RL for explainable fact-checking as it increases trustworthiness by providing a human-in-the-loop approach.

We demonstrate the Misinformation Concierge, a proof-of-concept that provides actionable intelligence on misinformation prevalent in social media. Specifically, it uses language processing and machine learning tools to identify subtopics of discourse and discerns non/misleading posts; presents statistical reports for policy-makers to understand the big picture of prevalent misinformation in a timely manner; and recommends rebuttal messages for specific pieces of misinformation, identified from within the corpus of data - providing means to intervene and counter misinformation promptly. The Misinformation Concierge proof-of-concept using a curated dataset is accessible at: https://demo-frontend-uy34.onrender.com/

Fake News on social media platforms has attracted a lot of attention in recent times, primarily for events related to politics (2016 US Presidential elections), and healthcare (infodemic during COVID-19), to name a few. Various methods have been proposed for detecting Fake News. The approaches span from exploiting techniques related to network analysis, Natural Language Processing (NLP), and the usage of Graph Neural Networks (GNNs). In this work, we propose DEAP-FAKED, a knowleDgE grAPh FAKe nEws Detection framework for identifying Fake News. Our approach combines natural language processing (NLP) and tensor decomposition model to encode news content and embed Knowledge Graph (KG) entities, respectively. A variety of these encodings provides a complementary advantage to our detector. We evaluate our framework using two publicly available datasets containing articles from domains such as politics, business, technology, and healthcare. As part of dataset pre-processing, we also remove the bias, such as the source of the articles, which could impact the performance of the models. DEAP-FAKED obtains an F1-score of 88% and 78% for the two datasets, which is an improvement of ~21 %, and ~3%, respectively, which shows the effectiveness of the approach.

Hate speech classification has become an important problem due to the spread of hate speech on social media platforms. For a given set of Hate Terms lists (HTs-lists) and Hate Speech data (HS-data), it is challenging to understand which hate term contributes the most for hate speech classification. This paper contributes two approaches to quantitatively measure and qualitatively visualise the relationship between co-occurring Hate Terms (HTs). Firstly, we propose an approach for the classification of hate-speech by producing a Severe Hate Terms list (Severe HTs-list) from existing HTs-lists. To achieve our goal, we proposed three metrics (Hatefulness, Relativeness, and Offensiveness) to measure the severity of HTs. These metrics assist to create an Inter-agreement HTs-list, which explains the contribution of an individual hate term toward hate speech classification. Then, we used the Offensiveness metric values of HTs above a proposed threshold minimum Offense (minOffense) to generate a new Severe HTs-list. To evaluate our approach, we used three hate speech datasets and six hate terms lists. Our approach shown an improvement from 0.845 to 0.923 (best) as compared to the baseline. Secondly, we also proposed Stable Hate Rule (SHR) mining to provide ordered co-occurrence of various HTs with minimum Stability (minStab). The SHR mining detects frequently co-occurring HTs to form Stable Hate Rules and Concepts. These rules and concepts are used to visualise the graphs of Transitivities and Lattices formed by HTs.

Online Social Media (OSM) platforms such as Twitter, Facebook are extensively exploited by the users of these platforms for spreading the (mis)information to a large audience effortlessly at a rapid pace. It has been observed that the misinformation can cause panic, fear, and financial loss to society. Thus, it is important to detect and control the misinformation in such platforms before it spreads to the masses. In this work, we focus on rumors, which is one type of misinformation (other types are fake news, hoaxes, etc). One way to control the spread of the rumors is by identifying users who are possibly the rumor spreaders, that is, users who are often involved in spreading the rumors. Due to the lack of availability of rumor spreaders labeled dataset (which is an expensive task), we use publicly available PHEME dataset, which contains rumor and non-rumor tweets information, and then apply a weak supervised learning approach to transform the PHEME dataset into rumor spreaders dataset. We utilize three types of features, that is, user, text, and ego-network features, before applying various supervised learning approaches. In particular, to exploit the inherent network property in this dataset (user-user reply graph), we explore Graph Convolutional Network (GCN), a type of Graph Neural Network (GNN) technique. We compare GCN results with the other approaches: SVM, RF, and LSTM. Extensive experiments performed on the rumor spreaders dataset, where we achieve up to 0.864 value for F1-Score and 0.720 value for AUC-ROC, shows the effectiveness of our methodology for identifying possible rumor spreaders using the GCN technique.

We contribute MIDAS as a novel sensing solution for characterizing everyday objects using thermal dissipation. MIDAS takes advantage of the fact that anytime a person touches an object, it results in heat transfer. By capturing and modeling the dissipation of the transferred heat, e.g., through the decrease in the captured thermal radiation, MIDAS can characterize the object and determine its material. We validate MIDAS through extensive empirical benchmarks and demonstrate that MIDAS offers an innovative sensing modality that can recognize a wide range of materials – with up to 83% accuracy – and generalize to variations in the people interacting with objects.

In the last decade, humanity has faced many different pandemics such as SARS, H1N1, and presently novel coronavirus (COVID-19). On one side, scientists are focusing on vaccinations, and on the other side, there is a need to propose models that can help in understanding the spread of these pandemics as it can help governmental and other concerned agencies to be well prepared, especially for pandemics, which spreads faster like COVID-19. The main reason for some epidemic turning into pandemics is the connectivity among different regions of the world, which makes it easier to affect a wider geographical area, often worldwide. Also, the population distribution and social coherence in the different regions of the world are non-uniform. Thus, once the epidemic enters a region, then the local population distribution plays an important role. Inspired by these ideas, we proposed a mobility-based SIR model for epidemics, which especially takes into account pandemic situations. To the best of our knowledge, this model is the first of its kind, which takes into account the population distribution and connectivity of different geographic locations across the globe. In addition to presenting the mathematical proof of our model, we have performed extensive simulations using synthetic data to demonstrate our model's generalizability. To demonstrate the wider scope of our model, we used our model to forecast the COVID-19 cases for Estonia.

Financial institutions such as banks regularly forecast the amount of finances an individual will have in his/her account in the near future. This can help banks in categorizing their customers so that banks can recommend financial products that matches the needs of their customers. In this work, we explored the historical financial transactions for predicting the amount a customer will receive through his/her transacting partners at a specific time. In particular, we use the Bitcoin transactional dataset, which has two main characteristics: i) network, and ii) temporal. This paper contributes by exploiting a specific kind of Graph Neural Network approach called Temporal-Graph Convolutional Network (T-GCN) for predicting the amount of Bitcoins received by a customer at a particular timestamp. The lower errors obtained using T-GCN approach compared to 11 baseline approaches (such as Support Vector Regression (SVR), Random Forest Regression (RFR), Vector Auto-Regressive (VAR), Long Short-Term Memory (LSTM), etc.) clearly demonstrate the effectiveness of T-GCN approach. In addition, our findings reveal that time is an important feature for such kind of predictive tasks.

Person to Person transactions have come a long way from trading goods, then to cash and now the possibility of online transactions. In the era of the internet, individuals often make payments through various financial platforms such as mobile payments, bank's online interfaces, etc. This paper presents an analysis of a large person-to-person (P2P) financial transactional network of three Baltic countries, namely Estonia, Latvia and, Lithuania. We collaborated with one of the largest financial institutions operating in these three countries to analyse an anonymous dataset of more than two million customers. We modeled these transactions from network science and explored the data with four different objectives. The first objective was to explore the network of transactions from a structural perspective and interpret their meanings. The next two objectives included the attributes of the nodes. In the second objective, we analysed the network for the similarity of interacting nodes in terms of income level. In the third objective, we analysed the spending pattern similarity among the interacting customers. The fourth objective aimed at exploring the relation between income and spending patterns. Our results indicate that the online payment network of Estonia is the most intact, among all the three countries, confirming the slogan of e-Estonia and Lithuanian network being the most fragmented. In addition, Latvia and Lithuania are more similar to each other in terms of their income vs. similarity patterns compared to Estonia. To the best of our knowledge, this is the first work that has studied such a large P2P financial network across three countries and the results of this study can provide insight into the financial behavior of the three Baltic countries.

The ability to detect critical spots in transportation networks is fundamental to improve traffic operations and road-network resilience in smart cities. Real-time monitoring of these networks, especially in very large metropolitan areas, is a compelling challenge due to the complexity of computing robustness metrics. This paper presents a framework for identifying vulnerabilities in very-large road networks. The framework adopts graph-based modeling of road networks and exploits big-data techniques and technologies for processing such large and complex graphs. First, we use the framework to prove the existence of a significant correlation between global efficiency and betweenness centrality. Then, we focus on an efficient algorithm, integrated in the framework, to rank the nodes according to this metric for finding potential vulnerabilities of a road network. To keep computation time under a "quasi" real-time threshold, a fast, requirement-driven, approximated strategy for computing betweenness centrality is adopted. The evaluation shows that the algorithm, integrated in the framework, exhibits a very good approximation for the most critical nodes, thus being well-suited for on-line operational monitoring.

Betweenness centrality is a metric widely used in several domains (social, biological, transportation, computer) to identify critical nodes of networks. Its exact computation is very demanding, with an O(nm) time complexity for unweighted graphs (where n is the number of nodes and m is the number of edges). Such complexity becomes an obstacle to the adoption of betweenness centrality for continuous monitoring of critical nodes in very large networks. Several solutions have been proposed to reduce computation time, mainly via parallelism, approximation or incremental recalculation. In this paper, we propose an algorithm for computing approximated values of betweenness that allows for tuning its performance on the basis of a tolerable error. The algorithm aims at reducing the number of single-source shortest-paths explorations via a pivot-based technique that exploits topological properties of graphs and clustering. It is evaluated by identifying the vulnerabilities (critical nodes) of a real-world, very-large road network. The evaluation shows that the approximation error does not significantly affect the most critical nodes, thus making the algorithm well-suited for on-line operational monitoring of road networks.

Mobile offloading is a promising technique to aid the constrained resources of a mobile device. By offloading a computational task, a device can save energy and increase the performance of the mobile applications. Unfortunately, in existing offloading systems, the opportunistic moments to offload a task are often sporadic and short-lived. We overcome this problem by proposing a social-aware hybrid offloading system (HyMobi), which increases the spectrum of offloading opportunities. As a mobile device is always co-located to at least one source of network infrastructure throughout of the day, by merging cloudlet, device-to-device and remote cloud offloading, we increase the availability of offloading support. Integrating these systems is not trivial. In order to keep such coupling, a strong social catalyst is required to foster user's participation and collaboration. Thus, we equip our system with an incentive mechanism based on credit and reputation, which exploits users’ social aspects to create offload communities. We evaluate our system under controlled and in-the-wild scenarios. With credit, it is possible for a device to create opportunistic moments based on user's present need. As a result, we extended the widely used opportunistic model with a long-term perspective that significantly improves the offloading process and encourages unsupervised offloading adoption in the wild.

A common problem in social network analysis is the presence of missing data. This problem has been extensively investigated in single layer networks, that is, considering one network at a time. However, in multilayer networks, in which a holistic view of multiple networks is taken, the problem has not been specifically studied, and results for single layer networks are reused with no adaptation. In this work, we take an exhaustive and systematic approach to understand the effect of missing data in multilayer networks. Differently from the single layer networks, depending on layer interdependencies, the common network properties can increase or decrease with respect to the properties of the complete network. Another important aspect we observed through our experiments on real datasets is that multilayer network properties like layer correlation and relevance can be used to understand the impact of missing data compared to measuring traditional network measures.

There is an inherent presence of communities in online social networks. These communities can be defined based on i) link structure or ii) the attributes of individuals. Attributes can indicate as interests in specific topics, like science-fiction books or romantic movies, or more in general their explicit affiliation to a group inside the network. In this paper, we analyze community structures as defined by how people are associated to third concepts like attributes. To understand the community patterns we analyze three large and one small social network datasets. Our analysis shows that, irrespective of the number of nodes for any particular interest in the network, at least 50% of the nodes are part of the same connected component in the graph induced by each interest. Another interesting result of our analysis is that the majority of sub-communities (50% or above) for any interest are separated by small hops (two to three) from each other.

In the last decade, various social platforms have been introduced on the web. Due to their specific orientation (friendship, professional connections, image sharing, etc.) users often join multiple networks. An important problem across these networks is the resolution of users profiles. That is, to identify if set of user profiles from different networks with different user ids or nicknames belong to the same user. The problem is more meaningful for resolving different profiles in digital forensic and criminal investigations. In this paper, we propose a method for profile resolution with the help of pictures being posted on different social platforms. We use the smartphone cameras which have become the source of instant image capturing and uploading process. In particular, we exploit the characteristic noise present in the images due to the manufacturing defects, to match user profiles across social platforms. To test our approach we select five different smartphones with two pairs of identical models, and three social platforms, namely Facebook, Google+ and WhatsApp. We evaluate our approach using real dataset of 1000 high-resolution pictures. The results indicate that even in the worst case our approach can provide profile matching upto 89.83%.

Modeling and understanding social network structure has interested researchers from many backgrounds including social science, computer science, theoretical physics and graph theory. Notable models include [1] and [2] achieving graphs with power-law degree distribution using preferential attachment and small-world characteristics using randomized rewiring of a regular ring lattice respectively. In contrast to a body of follow-up research which refine upon these seminal works to better capture the graph structure and characteristics (such as improving clustering coefficient by considering social triads along with preferential attachment [3]), this work aims additionally to model the geographic spread in social networks. With increased mobility in our society as well as enhanced communication opportunities social networks are increasingly spread all over the globe. Synthetic graphs imitating real-world social network characteristics are often used for driving simulations for planning and decision support. Incorporating geographic spread can facilitate better infrastructure provisioning in distributed systems supporting social and collaborative applications or model information of malware diffusion, word-of-mouth marketing, etc. The proposed model is tunable and modular. The model can be tuned to produce graphs with different geographic spread. The model is modular in the sense that existing geographic spread agnostic social network models can be plugged into our model to achieve desirable geographic spread in addition to other characteristics (such as degree distribution, clustering coefficient) that such a model would natively support.

Friend-to-friend networks, i.e. peer-to-peer networks where data are exchanged and stored solely through nodes owned by trusted users, can guarantee dependability, privacy and uncensorability by exploiting social trust. However, the limitation of storing data only on friends can come to the detriment of data availability: if no friends are online, then data stored in the system will not be accessible. In this work, we explore the tradeoffs between redundancy (i.e., how many copies of data are stored on friends), data placement (the choice of which friend nodes to store data on) and data availability (the probability of finding data online). We show that the problem of obtaining maximal availability while minimizing redundancy is NP-complete; in addition, we perform an exploratory study on data placement strategies, and we investigate their performance in terms of redundancy needed and availability obtained. By performing a trace-based evaluation, we show that nodes with as few as 10 friends can already obtain good availability levels.