Recent developments in language models have created new opportunities for time series forecasting by enabling models to capture complex temporal dependencies in data. Although Large Language Models (LLMs) have demonstrated their potential across various domains, their application in modeling structured time series data is still underexplored, particularly in low-resource settings. Our study investigates the effectiveness of Time Series Language Models (TSLMs) for macroeconomic forecasting across 27 European Union countries categorized by income level. We assess the performance of three open-source Small Language Models without fine-tuning: Chronos, Moirai, and TimesFM, based on their ability to forecast three economic variables: GDP percentage growth, inflation, trade balance, and one demographic variable: population growth, and compare them against traditional econometric models. Our study shows that the Moirai and TimesFM models consistently outperform econometric models in various contexts, particularly in middle and higher-income countries. Our findings highlight the potential of TSLMs as flexible time series forecasting tools.

Large language models (LLMs) have demonstrated potential to transform many industries, but their reliance on vast internet data raises concerns about inherent biases. This study investigates occupational bias in leading open-source LLMs, highlighting differences between technical and creative careers. Using occupation-specific prompts from the BOLD dataset and evaluating outputs with the regard metric, chosen for its ability to capture nuanced social perceptions such as respect and stereotyping, we quantify sentiment patterns underlying occupational bias. Our analysis reveals a consistent tendency for LLMs to express more positive sentiment toward technical occupations, with this bias intensifying as model size increases. We evaluate multiple mitigation strategies, including fairness-aware fine-tuning and fairness-oriented prompt engineering, finding that their combination most effectively reduces bias without compromising performance. These findings underscore the need for proactive bias assessment and integrated mitigation approaches to promote fairness and equity in AI-driven language generation systems.

The majority of financial transactions are now conducted virtually around the world. The widespread use of credit cards and online transactions encourages fraudulent activity. Thus, one of the most demanding real-world challenges is fraud detection. Unbalanced datasets, in which there are a disproportionately high number of non-fraud samples compared to incidents of fraud, are one of the key obstacles to effective fraud detection. A further factor complicating the learning process for cutting-edge machine learning classifiers is how quickly fraud behaviour changes. Thus, in this study, we suggest an efficient fraud detection methodology. We propose a unique nonlinear embedded clustering to resolve imbalances in the dataset, followed by a Deep Belief Network for detecting fraudulent transactions. The proposed model achieved an accuracy of 94% with a 70:30 ratio of training-validation dataset.

In this paper we demonstrate the feasibility of IoT deployment for noise level measurement to time-limited and high-intense, high-volume data, events. Through an iterative process, a prototype solution were designed and implemented in a real-time, privacy-compliant IoT sensor system under tight constraints concerning budget and development time. Our sensor system enables festival management to easily track, document and further, by applying real time big data analytics to the harvested information, have fact-full insights generated for decision making in terms of resolving noise disturbances. The whole approach was demonstrated by the use of lightweight Internet of Things architecture demonstrating how web technologies can be used throughout the technology stack in and IoT big data analytics case.

In recent years, there has been a growing research attention and practitioner interest in exploring the suitability of Blockchain technology for decentralised applications in multiple domains. This paper investigates the application of Blockchain technology to address some of the key challenges faced by the domain of Social Business (SB). SB is a business model for investments in social causes for the socio-economic development of under-privileged communities. We have modelled a small example of micro-credit use-case from microfinance activities of SB using a semi-formal modelling approach using Blockchain technology. We identified that the Blockchain technology provide solutions that enhance trust, transparency and auditability in SB activities. However, we have also identified challenges related to creating a native cryptocurrency for SB, and barriers to infrastructure and technology adoption by the different stakeholders in SB.

We present a big social media data study that comprises of 1 million individuals who interact with Facebook pages of the seven major political parties CDU, CSU, SPD, FDP, Greens, Die Linke and AfD during the 2017 German federal election. Our study uses the Social Set Analysis (SSA) approach, which is based on the sociology of associations, mathematics of set theory, and advanced visual analytics of event studies. We illustrate the capabilities of SSA through the most recent version of our Social Set Analysis (SoSeVi) tool, which enables us to deep dive into Facebook activity concerning the election. We explore a significant gender-based difference between female and male interactions with political party Facebook pages. Furthermore, we perform a multi-faceted analysis of social media interactions using gender detection, user segmentation and retention analysis, and visualize our findings. In conclusion, we discuss the analytical approach of social set analysis and conclude with a discussion of the benefits of set theoretical approaches based on the social philosophical approach of associational sociology.

This paper presents a novel approach that evaluates the right model for post engagement and predictions on Facebook. Moreover, paper provides insight into relevant indicators that lead to higher engagement with health care posts on Facebook. Both supervised and unsupervised learning techniques are used to achieve this goal. This research aims to contribute to strategy of health-care organizations to engage regular users and build preventive mechanisms in the long run through informative health-care content posted on Facebook.

This paper reports on a survey about the perceptions and practices of social media managers and experts in the area of public health. We have collected Facebook data from 153 public health care organizations and conducted a survey on them. 12% of organizations responded to the questionnaire. The survey results were combined with the findings from our previous work of applying clustering and supervised learning algorithms on big social data from the official Facebook walls of these organizations. In earlier research, we showed that the most successful strategy that leads to higher post engagement is visual content. In this paper, we investigated if organisations pursue this strategy or some other strategy that was successful and has not been uncovered by the machine learning algorithms. Performance of each organisation on Facebook is based on the number of posts (volume share) and the number of actions (value share). Calculation of performance with number of actions in the numerator and number of posts in the denominator reduces possible bias in the conclusions due to the varied size of organizations on social media. Moreover, our survey attempts to better understand the behaviour of organizations and to explain why almost half of the public health care content posted on Facebook is in the form of a short text message, where as the information can be communicated through seven other post types. Similar patterns and characteristics for different engagement clusters, also high and low performing companies suggests that a mixed-methods research approach consisting of machine learning techniques combined with expert knowledge using qualitative methods can offer important insights.

This paper tests whether accurate sales forecasts for Nike are possible from Facebook data and how events related to Nike affect the activity on Nike's Facebook pages. The paper draws from the AIDA sales framework (Awareness, Interest, Desire, and Action) from the domain of marketing and employs the method of social set analysis from the domain of computational social science to model sales from Big Social Data. The dataset consists of (a) selection of Nike's Facebook pages with the number of likes, comments, posts etc. that have been registered for each page per day and (b) business data in terms of quarterly global sales figures published in Nike's financial reports. An event study is also conducted using the Social Set Visualizer (SoSeVi). The findings suggest that Facebook data does have informational value. Some of the simple regression models have a high forecasting accuracy. The multiple regressions have a lower forecasting accuracy and cause analysis barriers due to data set characteristics such as perfect multicollinearity. The event study found abnormal activity around several Nike specific events but inferences about those activity spikes, whether they are purely event-related or coincidences, can only be determined after detailed case-by-case text analysis. Our findings help assess the informational value of Big Social Data for a company's marketing strategy, sales operations and supply chain.

This paper examines the changes in consumer behaviour and opinions due to the transition from a public to a commercial broadcaster in the context of broadcasting international media events. By analyzing TV viewer ratings, Facebook activity and its sentiment, we aim to provide answers to how the transition from airing Winter Olympic Games on NRK to TV2 in Norway affected consumer behaviour and opinion. We used text classification and visual analytics methods on the business and social datasets. Our main finding is a clear link between negative sentiment and commercials. Despite positive change in customer behaviour, there was a negative change in customer opinion. Based on media events and broadcaster theories, we identify generalisable findings for all such transitions.

This paper explores the relationship between TV viewership ratings for Scandinavian's most popular talk show, Skavlan and public opinions expressed on its Facebook page. The research aim is to examine whether the activity on social media affects the number of viewers per episode of Skavlan, how the viewers are affected by discussions on the Talk Show, and whether this creates debate on social media afterwards. By analyzing TV viewer ratings of Skavlan talk show, Facebook activity and text classification of Facebook posts and comments with respect to type of emotions and brand sentiment, this paper identifes patterns in the users' real-world and digital world behaviour.

Current state-of-the-art in big social data analytics is largely limited to graph theoretical approaches such as social network analysis (SNA) informed by the social philosophical approach of relational sociology. This paper proposes and illustrates an alternate holistic approach to big social data analytics, social set analysis (SSA), which is based on the sociology of associations, mathematics of set theory, and advanced visual analytics of event studies. We illustrate our new approach by applying it to relate real-world events with their reflections in terms of user interactions on social media platforms. We present and discuss a theoretical and conceptual model of social data followed by a formal description of our technique based on set theory and event studies with a real-world social data example from Facebook. We then illustrate our new approach by reporting on the design, development, and evaluation results of a state-of-the-art visual analytics dashboard, the Social Set Visualizer (SoSeVi). Using SoSeVi, we conducted a real-world case study that consists of approximately 90 million Facebook user interactions from 11 different companies that have been mentioned in the traditional media in relation to the garment factory accidents in Bangladesh, and analyze the results. The enterprise application domain for the dashboard is corporate social responsibility (CSR) and the targeted end-users are CSR researchers and practitioners. The design of the dashboard was based on the social set analysis approach to computational social science mentioned above. The development of the dashboard involved cutting-edge open source visual analytics libraries (D3.js) and creation of new visualizations such as of actor mobility across time and space, conversational comets, and more. Evaluation of the dashboard consisted of technical testing, usability testing, and domain-specific testing with CSR students and yielded positive results. In conclusion, we discuss the new analytical approach of social set analysis and conclude with a discussion of the benefits of set theoretical approaches based on the social philosophical approach of associational sociology.