DOI: 10.3897/jucs.168955

Authors: Davut Çulha

Abstract: Large Language Models (LLMs) have shown impressive capabilities in natural language understanding and reasoning; however, their internal conceptual organization remains largely opaque. This study introduces a novel prompt engineering approach to explore the latent conceptual spaces of LLMs by identifying maximally divergent conceptual regions. It draws on the conceptual spaces framework, in which knowledge is represented as geometric regions defined by quality dimensions. Within this framework, a proxy metric referred to as the conceptual divergence count is proposed. This metric represents the number of maximally divergent conceptual regions identified through structured prompts. Although the metric does not measure dimensionality in the geometric or architectural sense, it serves as an indicator of a model’s conceptual diversity. The method is applied to a range of LLMs, including models from the Gemini, GPT, and Claude families, as well as additional models such as Llama-3.3-70B-Instruct, Mistral-7B-Instruct-v0.3, Gemma-3-12B-IT, and Grok-3. The results show substantial variation in diversity, with gemini-2.5-flash-preview-05-20 achieving the highest value of 91 and claude-3-7-sonnet-20250219 recording the lowest value of 7. These findings suggest that diversity, as measured by this method, may provide insights into the internal organization of conceptual representations. While the metric is to be an indicator of a model’s factual knowledge structure, providing complementary insight into its internal organization, it does not directly measure knowledge accuracy. The proposed approach contributes a geometric perspective to the evaluation of LLM knowledge, complementing existing benchmarks and supporting interpretability.

HTML

PDF

DOI: 10.3897/jucs.114607

Authors: Marija Kuštelega, Renata Mekovec, Ahmed Shareef

Abstract: As technology advances and becomes more extensively used, digital twins are likely to play an increasingly crucial role in defining the future of industry, trade, and society. Despite the stated advantages and potential of digital twin technology, certain research and implementation gaps exist, which have hindered the adoption and advancement of digital twins. This study investigates how current research on the digital twin implementations has been positioned in front of practical challenges focused on privacy and security issues. The research method adopted was a systematic literature review, employing the PRISMA approach. A total of 47 publications were identified and analyzed. The results indicate that the privacy and security challenges for digital twin implementation are complicated and may be divided into six primary groups: (1) data privacy, (2) data security, (3) data management, (4) data infrastructure and standardization, (5) ethical and moral issues, (6) legal and social issues.

HTML

PDF

DOI: 10.3897/jucs.109080

Authors: F. Didem Alay, Nagehan İlhan, M. Tahir Güllüoğlu

Abstract: Photovoltaic (PV) energy systems are a leading type of renewable energy systems globally. Predicting PV energy production accurately is crucial for maintaining efficient energy grids, making informed decisions in the energy market, and reducing maintenance costs. To ensure high accuracy and optimal production, it is essential to monitor and analyze these variables regularly. Solar radiation and temperature are two meteorological variables that directly affect the quantity of PV energy generated in PV facilities. The Performance Ratio (PR) is a critical parameter for assessing PV plant performance. A comprehensive model was constructed in this study to forecast solar radiation and temperature using multiple machine learning methods, including Instance-Based K-Nearest Neighbor Algorithm (IBK), Linear Regression, Random Forests, Random Tree, Multilayer Perceptron (MLP), and MLP Regression. Moreover, we used time series approaches, such as Simple Exponential Smoothing (SES), Error-Trend-Seasonality (ETS), Autoregressive Integrated Moving Average (ARIMA) and Holt Winter's Seasonal Method (HWES) models for PV systems prediction. Initially, we conducted daily forecasts as well as 1-step ahead forecasts at 5-minute intervals for both solar radiation and temperature. It is crucial to subject both variables to the same methodology in order to construct precise models for forecasting PV. Secondly, we compared the predicted values of solar radiation and temperature with the actual energy yield of the power plant to calculate energy production. Subsequently, a relative analysis of data mining models and time series models have been performed depending on the statistical error criteria like RMSE, MAPE, MABE, MAE, MSE, and direction accuracy (DAC).

HTML

PDF

DOI: 10.3897/jucs.108445

Authors: Emre Sadıkoğlu, İrfan Kösesoy, Murat Gök

Abstract: With the increasing popularity and usage of artificial intelligence systems, it has become crucial to address their vulnerability to cyber-attacks. In this study, we propose a novel gradient descent-based method to generate fake data that can be accepted as positive by a targeted machine learning model. Our method is designed to generate a large number of positive samples with a minimal number of probes to the model, making it difficult to detect by security systems. Additionally, we develop an alternative model to the attacked model using a reverse engineering approach, trained on a dataset composed of the samples generated by our method. We evaluate the success of our proposed method and the alternative model through a series of experiments. We conducted experiments on six distinct datasets, each of which was trained using three separate machine-learning algorithms. This resulted in a total of eighteen unique models that were evaluated and compared in our analysis. In the evaluation of results, the most commonly used metrics in the literature, including effective attack rate (EAR), accuracy, precision, recall, and F1 score, were employed. Focusing particularly on EAR-oriented assessments, our method demonstrates its effectiveness with a notably high EAR of 97% in the combination of the kNN method and the Cancer dataset. According to the results of our experiments, the proposed method demonstrates high effectiveness as a data-driven attack method.

HTML

PDF

DOI: 10.3897/jucs.94514

Authors: Fernando Terroso-Saenz, Andres Muñoz

Abstract: Predicting human mobility is a key element in the development of intelligent transport systems. Current digital technologies enable capturing a wealth of data on mobility flows between geographic areas, which are then used to train machine learning models to predict these flows. However, most works have only considered a single data source for building these models or different sources but covering the same spatial area. In this paper we propose to augment a macro open-data mobility study based on cellular phones with data from a road traffic sensor located within a specific motorway of one of the mobility areas in the study. The results show that models trained with the fusion of both types of data, especially long short-term memory (LSTM) and Gated Recurrent Unit (GRU) neural networks, provide a more reliable prediction than models based only on the open data source. These results show that it is possible to predict the traffic entering a particular city in the next 30 minutes with an absolute error less than 10%. Thus, this work is a further step towards improving the prediction of human mobility in interurban areas by fusing open data with data from IoT systems.

HTML

XML

PDF

DOI: 10.3897/jucs.66375

Authors: Alessia D'Andrea, Maria Chiara Caschera, Fernando Ferri, Patrizia Grifoni

Abstract: The paper aims to provide a method to analyse and observe the characteristics that distinguish the individual communication style such as the voice intonation, the size and slant used in handwriting and the trait, pressure and dimension used for sketching. These features are referred to as Communication Extensional Features. Observing from the Communication Extensional Features, the user’s behavioural features, such as the communicative intention, the social style and personality traits can be extracted. These behavioural features are referred to as Communication Intentional Features. For the extraction of Communication Intentional Features, a method based on Hidden Markov Models is provided in the paper. The Communication Intentional Features have been extracted at the modal and multimodal level; this represents an important novelty provided by the paper. The accuracy of the method was tested both at modal and multimodal levels. The evaluation process results indicate an accuracy of 93.3% for the Modal layer (handwriting layer) and 95.3% for the Multimodal layer.

HTML

XML

PDF

DOI: 10.3897/jucs.2020.007

Authors: Anastasija Nikiforova, Janis Bicevskis, Zane Bicevska, Ivo Oditis

Abstract: The paper proposes a new data object-driven approach to data quality evaluation. It consists of three main components: (1) a data object, (2) data quality requirements, and (3) data quality evaluation process. As data quality is of relative nature, the data object and quality requirements are (a) use-case dependent and (b) defined by the user in accordance with his needs. All three components of the presented data quality model are described using graphical Domain Specific Languages (DSLs). In accordance with Model-Driven Architecture (MDA), the data quality model is built in two steps: (1) creating a platform-independent model (PIM), and (2) converting the created PIM into a platform-specific model (PSM). The PIM comprises informal specifications of data quality. The PSM describes the implementation of a data quality model, thus making it executable, enabling data object scanning and detecting data quality defects and anomalies. The proposed approach was applied to open data sets, analysing their quality. At least 3 advantages were highlighted: (1) a graphical data quality model allows the definition of data quality by non-IT and non-data quality experts as the presented diagrams are easy to read, create and modify, (2) the data quality model allows an analysis of "third-party" data without deeper knowledge on how the data were accrued and processed, (3) the quality of the data can be described at least at two levels of abstraction - informally using natural language or formally by including executable artefacts such as SQL statements.

HTML

XML

PDF

DOI: 10.3217/jucs-025-10-1323

Authors: Adi Darliansyah, M. Naeem, Farhaan Mirza, Russel Pears

Abstract: Personality distinctively characterises an individual and profoundly influences behaviours. Social media offer the virtual community an unprecedented opportunity to generate content and share aspects of their life which often reflect their personalities. The interest in using deep learning to infer traits from digital footprints has grown recently; however, very limited work has been presented which explores the sentiment information conveyed. The present study, therefore, used a computational approach to classify personality from social media by gauging public perceptions underlying factors encompassing traits. In the research reported in this paper, a Sentiment-based Personality Detection system was developed to infer trait from short texts based on the 'Big Five' personality dimensions. We exploited the spirit of Neural Network Language Model (NNLM) by using a uni ed model that combines a Recurrent Neural Network named Long Short-Term Memory (LSTM) with a Convolutional Neural Network (CNN). We performed sentiment classi cation by grouping short messages harvested online into three categories, namely positive, negative, and nonpartisan. This is followed by employing Global Vectors (GloVe) to build vectorial word representations. As such, this step aims to add external knowledge to short texts. Finally, we trained each variant of the models to compute prediction scores across the ve traits. Experimental study indicated the e ectiveness of our system. As part of our investigation, a case study was carried out to investigate the existing correlation of personality traits and opinion polarities which employed the proposed system. The results support the prior ndings of the tendency of persons with the same traits to express sentiments in similar ways.

HTML

XML

PDF

DOI: 10.3217/jucs-025-09-1199

Authors: François Bouchaud, Gilles Grimaud, Thomas Vantroys, Pierrick Buret

Abstract: The Internet of Things (IoT) is everywhere around us. Smart communicating objects are offering the digitalization of lives. They create new opportunities within criminal investigations. In recent years, the scientific community sought to develop a common digital framework and methodology adapted to IoT-based infrastructure. However, the difficulty in exploiting the IoT lies in the heterogeneous nature of the devices, the lack of standards and the complex architecture. Although digital forensics are considered and adopted in IoT investigations, this work only focuses on the collection. The identification phase is quite unexplored. It addresses the challenges of locating hidden devices and finding the best evidence to be collected. The matter of facts is the traditional method of digital forensics does not fully fit the IoT environment. Furthermore, the investigator can no longer consider a connected object as a single device, but as an interconnected whole one, anchored in a cross-disciplinary environment. This paper presents the methodology for identifying and classifying connected objects in search of the best evidence to be collected. It offers techniques for detecting and locating the appropriate equipment. Based on frequency mapping and interactions, it transfers the concept of "fingerprinting" into the field of crime scene. It focuses on the technical and data criteria to successfully select the relevant IoT devices. It gives a general classiffication as well as the limits of such an approach. It shows the collection of digital evidence by focusing on pertinent information from the Internet of Things.

HTML

XML

PDF

DOI: 10.3217/jucs-025-08-0925

Authors: Ashot Harutyunyan, Arnak Poghosyan, Naira Grigoryan, Narek Hovhannisyan, Nicholas Kushmerick

Abstract: We address several problems in intelligent log management of distributed cloud computing applications and their machine learning solutions. Those problems concern various tasks on characterizing data center states from logs, as well as from related or other quantitative metrics (time series), such as anomaly and change detection, identification of baseline models, impact quantification of abnormalities, and classification of incidents. These are highly required jobs to be performed by today's enterprise-grade cloud management solutions. We describe several approaches and algorithms that are validated to be effective in an automated log analytics combined with analytics from time series perspectives. The paper introduces novel concepts, approaches, and algorithms for feasible log-plus-metric-based management of data center applications in the context of integration of relevant technology products in the market.

HTML

XML

PDF

DOI: 10.3217/jucs-024-09-1153

Authors: Ahyoung Choi, Hangsik Shin

Abstract: Recently, many studies have been conducted on how to manage and analyze various types of health data such as clinical data, genomic data, and wirelessly collected multiple sensory data. In this paper, we propose a web-based healthcare data integration and management platform that collects heterogeneous types of health-related medical record as well as real-time lifelogging data. This platform provides flexible architecture to different types of data exchanges. The platform manages real-time data such as heart rate, blood pressure, and activity information extracted from various healthcare devices and provides functions to transmit them to the server. Then it analyses the risk based on a domain knowledge and individual differences by applying machine learning tools, then visualizes the result to the patient and doctor dynamically based on information simplification method. It also controls the data access authority concerning the level of expertise and role. For evaluation of integrated data analysis, we apply open database and evaluate the proposed risk analyser result. The proposed platform could be utilized for future healthcare service to share accumulated healthcare data in various situations.

HTML

XML

PDF

DOI: 10.3217/jucs-023-11-1057

Authors: Giangiacomo Bravo, Mike Farjam

Abstract: Computational social sciences (CSS) refer to computer-enabled investigations of human behaviour and social interaction. They include three main components - (i) computational modelling and social simulation, (ii) the analysis of digital traces of online interactions, (iii) virtual labs and online experiments - and allow researchers to perform studies that were even hard to imagine a few decades ago. Moreover, CSS favour a more systematic test of theories and increase the possibility of study replication, two factors holding the potential to help social sciences reach a higher scientific status. Despite the huge potential of CSS, we follow previous works in identifying several impediments to a larger adoption of computational methods in social sciences. Most of them are linked with the humanistic attitude and a lack of technical skills of many social scientist. Significant changes in the basic training of social scientist and in the relation patterns with other disciplines and departments are needed before the potential of CSS can be fully exploited.

HTML

XML

PDF

DOI: 10.3217/jucs-023-11-1038

Authors: Mikko Laitinen, Jonas Lundberg, Magnus Levin, Alexander Lakaw

Abstract: This paper presents the Nordic Tweet Stream, a cross-disciplinary digital humanities project that downloads Twitter messages from Denmark, Finland, Iceland, Norway and Sweden. The paper first introduces some of the technical aspects in creating a real-time monitor corpus that grows every day, and then two case studies illustrate how the corpus could be used as empirical evidence in studies focusing on the global spread of English. Our approach in the case studies is sociolinguistic, and we are interested in how widespread multilingualism which involves English is in the region, and what happens to ongoing grammatical change in digital environments. The results are based on 6.6 million tweets collected during the first four months of data streaming. They show that English was the most frequently used language, accounting for almost a third. This indicates that Nordic Twitter users choose English as a means of reaching wider audiences. The preference for English is the strongest in Denmark and the weakest in Finland. Tweeting mostly occurs late in the evening, and high-profile media events such as the Eurovision Song Contest produce considerable peaks in Twitter activity. The prevalent use of informal features such as univerbated verb forms (e.g., gotta for (HAVE) got to) supports previous findings of the speech-like nature of written Twitter data, but the results indicate that tweeters are pushing the limits even further.

HTML

XML

PDF

DOI: 10.3217/jucs-023-11-1035

Authors: Giangiacomo Bravo, Mikko Laitinen, Magnus Levin, Welf Löwe, Göran Petersson

Abstract: The ubiquity of sensor, computing, communication, and storage technologies provides us with access to previously unknown amounts of data - Big Data. Big Data has revolutionized research communities and their scientific methodologies. It has, for instance, innovated the approaches to knowledge and theory building, validation, and exploitation taken in the engineering sciences. The humanities and social sciences even face a paradigm shift away from data-scarce, static, coarse-grained and simple studies towards data-rich, dynamic, high resolution, and complex observations and simulations. The present focused topic presents investigations from different research fields in which the focus is either on utilizing Big Data or charting the benefits of using such evidence in basic research.

HTML

XML

PDF

DOI: 10.3217/jucs-021-01-0066

Authors: Sean Goggins, Wanli Xing, Xin Chen, Bodong Chen, Bob Wadholm

Abstract: This study proposes a process-oriented, automatic, formative assessment model for small group learning based on complex systems theory using a small dataset from a technology-mediated, synchronous mathematics learning environment. We first conceptualize small group learning as a complex system and explain how group dynamics and interaction can be modeled via theoretically grounded, simple rules. These rules are then operationalized to build temporally-embodied measures, where varying weights are assigned to the same measures according to their significance during different time stages based on the golden ratio concept. This theory-based measure construction method in combination with a correlation-based feature subset selection algorithm reduces data dimensionality, making a complex system more understandable for people. Further, because the discipline of education often generates small datasets, a Tree-Augmented Naïve Bayes classifier was coded to develop an assessment model, which achieves the highest accuracy (95.8%) as compared to baseline models. Finally, we describe a web-based tool that visualizes time-series activities, assesses small group learning automatically, and also offers actionable intelligence for teachers to provide real-time support and intervention to students. The fundamental contribution of this paper is that it makes complex, small group behavior visible to teachers in a learning context quickly. Theoretical and methodological implications for technology mediated small group learning and learning analytics as a whole are then discussed.

HTML

XML

PDF