|
|
 |
Hybrid Event
|
Artificial Intelligence Systems and Optimization
Section chair: Darko Huljenić
|
J. Mileski, M. Toshevska , S. Gievska (Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, Skopje, Macedonia)
Autonomous Control and Path Planning of UAV with Deep Reinforcement Learning 
Autonomous unmanned aerial vehicles (UAV) can be applied as a substitution for many manual processes, which results in solutions that are more cost-optimal and less prone to human error. In this paper, we consider a task that requires a quadrotor UAV to reach waypoints from an environment as fast as possible. This work presents various reinforcement learning experiments on autonomous control and path planning, exploring the potential of current state-of-the-art algorithms, including the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, which addresses the overestimation of value estimates and suboptimal policies commonly present in continuous control domain actor-critic models. The experiments also include different optimization techniques for finding the best set of hyperparameters. We evaluate the trained reinforcement learning agents and provide a detailed comparison and discussion of the results.
|
M. Schacht, J. Murach, A. Guldner, L. Begic Fazlic, L. Creutz, S. Naumann, K. Gollmer, G. Dartmann (Umwelt-Campus Birkenfeld, Birkenfeld, Germany)
Architecture of an AI on Device Board for Local, Automated Training and Inference
This paper presents the conceptualization and implementation of an AI on Device board consisting of a model-node for training of AI models and a sensing-node in form of a TinyML module for model inference and data collection, coupled with a shared storage unit for data exchange between the nodes.
With increasing complexity of artificial intelligence applications and the resulting cost of energy, data, and hardware requirements, assessments of resource- and energy-efficiency are often neglected. We plan to address these areas with innovative AI hardware solutions.
We demonstrate an AI on Device implementation without external transfer requirements, present the hardware and software concept and implement an use-case. We show how these methodologies achieve higher energy and resource efficiency while maintaining the desired performance. Furthermore the transfer of existing methodologies to an AI on Device board is demonstrated to show the adaptability and interoperability of different application areas of the AI sector.
|
G. Oparin, V. Bogdanova, A. Pashinin (Institute for System Dynamics and Control Theory of SB RAS, Irkutsk, Russian Federation)
Permutational Logical Networks
A special class of permutational logical networks is considered. Dynamic models of these networks have a phase space with only fixed points and cyclic trajectories. Within the framework of the Boolean constraint method, the available dynamic equations allow one to define the permutability property as a quantified Boolean formula or a system of Boolean equations. Examples of linear and nonlinear permutational logical networks are presented. A technology for studying the permutability property based on modern QSAT and SAT solvers is described.
|
M. Popovic (University of Novi Sad, Faculty of Technical Sciences, Novi Sad, Serbia), M. Popovic (RT-RK Institute for Computer Based Systems, Novi Sad, Serbia), M. Djukic, I. Basicevic (University of Novi Sad, Faculty of Technical Sciences, Novi Sad, Serbia)
Translating Federated Learning Algorithms in Python into CSP Processes Using ChatGPT
The Python Testbed for Federated Learning Algorithms is a simple Python FL framework that is easy to use by ML&AI developers who do not need to be professional programmers and is also amenable to LLMs. In the previous research, generic federated learning algorithms provided by this framework were manually translated into the CSP process and algorithms’ safety and liveness properties were automatically verified by the model checker PAT. In this paper, a simple translation process is introduced wherein the ChatGPT is used to automate the translation of the mentioned federated learning algorithms in Python into the corresponding CSP processes. Within the process, the minimality of the used context is estimated based on the feedback from ChatGPT. The proposed translation process was experimentally validated by successful translation (verified by the model checker PAT) of both generic centralized and decentralized federated learning algorithms.
|
J. Sabljo, R. Čorić, M. Đumić (School of Applied Mathematics and Informatics, Josip Juraj Strossmayer University of Osijek, Osijek, Croatia), M. Đurasević (Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia)
Evolving Priority Rules to Optimize Multi-Objective Criteria for the Resource-Constrained Project Scheduling Problem
The resource-constrained project scheduling problem (RCPSP) is a combinatorial optimization problem that entails allocating a set of tasks to limited resources, such as budget, workers, or equipment while aiming to optimize a defined objective function. Since this problem belongs to the class of NP-hard problems, heuristic approaches such as priority rules (PRs) are often used to solve them. In literature, genetic programming (GP) was used to evolve PRs for RCPSP while optimizing a single criterion, but optimizing only a single criterion is rare in real-world problems. Usually, real-world problems require the optimization of several criteria simultaneously. In this paper, new PRs are evolved using GP for multi-objective RCPSP. More precisely, sets of Pareto optimal solutions were obtained using non-dominated sorting genetic algorithm II (NSGA-II) and non-dominated sorting genetic algorithm III (NSGA-III). Through experiments, these two algorithms were compared using multi-objective performance measures. Additionally, a comparison between multi-objective PRs and single-objective PR was conducted.
|
S. Delalić (Faculty of Science, University of Sarajevo, Sarajevo, Bosnia and Herzegovina), R. Mutapčić, I. Fejzić (Info Studio d.o.o., Sarajevo, Bosnia and Herzegovina)
Clustering Approaches in Vehicle Routing Problems: A Comparative Study on Real-World Scenarios
The Vehicle Routing Problem (VRP) is among the most complex optimization challenges. Practical solutions require addressing real-world constraints such as time windows, vehicle capacities, delivery restrictions, driver working hours, and heterogeneous vehicle fleets. Solutions are often implemented in two stages: the first involves clustering customers, while the second focuses on incremental routing of these clusters to reduce complexity and improve solution control and explainability. However, the second stage heavily depends on the quality of the first, and clustering methods vary depending on client requirements. This paper explores various clustering methods and their impact on the final routing results, with a focus on real-world examples. The study includes diverse client scenarios, ranging from small-scale distribution systems with a limited number of customers to large-scale operations managing thousands of deliveries daily, covering both small and large orders. From fixed clustering and geographic partitioning to dynamic clustering algorithms and hybrid approaches, the advantages and limitations of each method are analyzed. The findings aim to provide actionable insights into selecting clustering methods that align with specific use cases, ensuring enhanced efficiency and adaptability in practical applications.
|
| Break |
Computer Vision
Section chair: Marina Ivašić-Kos |
D. Marušić (Independent Researcher, Zagreb, Croatia), S. Popović, Z. Kalafatić (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
Template Matching in Images Using Segmented Normalized Cross-Correlation
In this paper, a new variant of an algorithm for normalized cross-correlation (NCC) is proposed in the context of template matching in images. The proposed algorithm is based on the precomputation of a template image approximation, enabling more efficient calculation of approximate NCC with the source image than using the original template for exact NCC calculation. The approximate template is precomputed from the template image by a split-and-merge approach, resulting in a decomposition to axis-aligned rectangular segments, whose sizes depend on per-segment pixel intensity variance. In the approximate template, each segment is assigned the mean grayscale value of the corresponding pixels from the original template. The proposed algorithm achieves superior computational performance with negligible NCC approximation errors compared to the well-known Fast Fourier Transform (FFT)-based NCC algorithm, when applied on less visually complex and/or smaller template images. In other cases, the proposed algorithm can maintain either computational performance or NCC approximation error within the range of the FFT-based algorithm, but not both.
|
T. Krčmar, D. Šabanović, M. Habijan, I. Galić, I. Lukić (Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, Croatia)
MRI Image Reconstruction Using Adaptive Sampling and Diffusion Models
Efficient image sampling is essential for balancing acquisition speed, resolution, and computational cost in computer vision. Insufficiently sampled data often leads to quality degradation and reconstruction artifacts, posing challenges for downstream applications. To address this, we developed a diffusion-based model with a dynamic adaptive sampling strategy, combining randomly generated masks with denoising to enhance reconstruction quality. Comprehensive analyses validate the model’s effectiveness, highlighting its potential to optimize workflows in data-intensive image processing applications.
|
A. Antunović, R. Marković, M. Karajko (Faculty of Electrical Engineering, Computer Science and Information Technology, Osijek, Croatia), E. Fernández (University of Granada, Granada, Spain), G. Martinović, I. Aleksi, J. Balen (Faculty of Electrical Engineering, Computer Science and Information Technology, Osijek, Croatia)
Advancements in Fire Detection: A Review of Computer Vision and Deep Learning Approaches
This paper reviews existing research and effort on an important contemporary topic in the field of computer vision. Specifically, it examines current methodologies and algorithms for the detection of fire and smoke. As fires are becoming more frequent and spread rapidly, causing substantial material damage and potentially endangering human lives, it is necessary to develop technology that can detect fires promptly and notify the relevant authorities. This paper specifically emphasizes the work conducted with computer vision and deep learning based approaches. Additionally, the paper highlights the advantages of using unmanned aerial vehicles equipped with RGB and IR cameras for operations in hard-to-reach fire areas. The implementation of IoT technology enhances the fire detection system since it improves the accuracy of early and real-time fire detection, while also supplying valuable data for subsequent analysis and fire preventive efforts. Despite the many advantages offered by the current technology, due to the unpredictable fire nature and specific machine learning characteristics, there is still room for further progress in this research domain.
|
S. Zhang, A. Zakariya (SatakuntaUniversityof Applied Sciences, Pori, Finland), T. Aaltonen (SatakuntaUniversityof Applied Sciences, Ulvila, Finland)
Advancing Welding Automation: Molten Pool Segmentation and Real-Time Size Prediction
This study explores novel approaches for analyzing the welding molten pool and estimating its size in real-time to enhance welding quality and automation. A domain-adapted segmentation method is proposed, leveraging fine-tuned modern segmentation models to accommodate the unique characteristics of molten pool imagery. Building on segmentation results, a real-time molten pool size estimation technique is introduced, utilizing geometric references to guide adaptive adjustments to welding machine parameters, thereby reducing risks and optimizing the process.
Furthermore, the integration of long-range feature modeling with local feature refinement is investigated to address challenges in capturing both global and localized information in molten pool imagery. Experimental evaluations underscore the performance of the proposed methods and highlight pathways for developing robust solutions in welding automation. This work advances efforts to improve welding precision and reliability through innovative segmentation and size prediction techniques.
|
D. Sofiykov (Burgas Free University, Burgas, Bulgaria), P. Georgieva (BurgasFreeUniversity, Burgas, Bulgaria)
A Computer Vision System for Real-time Vehicular Traffic Management
This paper explores the integration of computer vision technology into a real-time traffic management system. It outlines the development and implementation of a computer vision model designed to optimize traffic flow, detailing technical aspects of the project. The system’s backend is developed in Python, utilizing widely used artificial intelligence libraries, enabling administrators to monitor and maintain the system effectively. A key feature of the system is an application that recognizes vehicles based on their license plates and color, cross-referencing this information with a database to detect traffic rule violations such as unpaid fees, fines, or missing vignettes. By combining advanced computer vision techniques with efficient programming frameworks, this project demonstrates a scalable, practical approach to modern traffic management challenges.
|
J. Maltar, L. Borozan (School of Applied Mathematics and Informatics, J. J. Strossmayer University of Osijek, Osijek, Croatia)
Siamese Neural Networks for Visual Place Recognition
Siamese neural networks are powerful constructs in the domain of deep learning and are currently the subject of intense research. Siamese neural networks describe the use of neural networks rather than the design of a network itself. The use of Siamese neural networks allows the possibility of adjusting the parameters of a learnable function either from scratch or with fine-tuning to better fit a particular task. In addition to adjustment, another useful side-effect according to the design of a learnable function, is dimensionality reduction. In this paper, we will utilize numerous fundamental and state-of-the-art concepts related to Siamese networks in order to solve different tasks. A special attention will be paid to the problem of visual place recognition.
|
B. Kotevski, S. Koceski, N. Koceska (Goce Delcev University, Stip, Macedonia)
Deep Learning-Based System for Detection and Classification of Household Entry Point States Using YOLOv11
Object state detection in home environments presents a crucial challenge for smart home systems and robotics applications. This paper presents deep learning approach for detecting and classifying the states of common household entryways - specifically doors and windows - as either open or closed. It uses the You Only Look Once (YOLO) v11 object detection model. Custom dataset comprising images of doors and windows in various lighting conditions and angles was constructed for this purpose.
Using the Roboflow platform, images were annotated with bounding boxes and preprocessed for standardization, with data augmentation applied to increase sample diversity. The resulting dataset was then divided into training, validation, and testing sets for model development and evaluation. The trained model was deployed through Roboflow's API, enabling seamless integration with Python through the dedicated library, which allowed for efficient real-time inference in our application. Experimental results indicate that the proposed approach offers a reliable solution for automated entryway state monitoring in home environments, with potential applications in smart home systems, security monitoring, and robotic navigation. The implementation demonstrates robust performance across varying lighting conditions and viewing angles, making it suitable for real-world applications. This research contributes to the growing field of automated home monitoring systems by providing a practical solution for entryway state detection using contemporary deep learning techniques.
|
Natural Language Processing - I
Section chair: Marko Horvat |
M. Rana, Z. Bhuiyan , M. Limon (Department of Computer Science and Engineering, Independent University, Bangladesh, Dhaka, Bangladesh), A. Rahman (Center of Excellence in Teaching and Learning, Green University, Dhaka, Bangladesh), M. Hasan, T. Habib (Department of Computer Science and Engineering, Independent University, Bangladesh, Dhaka, Bangladesh)
Review of Modern Chatbot Technologies: Performance, Capabilities, and Applications
The rapid evolution of conversational AI has introduced numerous chatbots with distinct features and applications tailored to specific industry needs. This paper surveys the performance of prominent models- ChatGPT, Gemini, DeepSeek, Blackbox, LLMA, Perplexity, and Claude evaluating capabilities such as natural language understanding, contextual reasoning, domain-specific expertise, and multimodal interactions. The study extends to advanced tasks like coding assistance, image generation, PDF interpretation, and graph generation, employing a robust mix of qualitative and quantitative metrics, including response accuracy, latency, and user satisfaction. This comprehensive analysis is crucial for guiding future innovations and enabling stakeholders to make informed decisions about integrating chatbots in various sectors. The findings emphasize that while models like ChatGPT and Gemini excel in contextual understanding and multimodal capabilities, others such as DeepSeek and Claude address more niche challenges, enhancing the AI's potential to improve human-computer interaction and setting the stage for significant advancements in conversational AI. The study also highlights the need to address emerging ethical challenges and suggests directions for future research to optimize these technologies further.
|
M. Toshevska, G. Mirceva, S. GIevska (Ss. Cyril and Methodius University, Faculty of Computer Science and Engineering, Skopje, Macedonia)
Exploring Large Language Models for Data Augmentation: A Case Study for Text Style Transfer
Text style transfer is the task that involves modifying a sentence to adapt to a desired target style while preserving its original meaning. It often requires high-quality parallel datasets that are not always available. This paper explores data augmentation techniques for text style transfer leveraging large language models (LLMs) to address the challenge of dataset scarcity. Our approach generates synthetic parallel data by prompting LLMs to paraphrase and/or rewrite sentences in diverse styles, enabling the creation of larger and more varied datasets. We demonstrate the applicability of this approach across three tasks: formality transfer with the GYAFC dataset, sentiment transfer with the Yelp dataset, and personal style transfer with the Shakespeare dataset. This work introduces an approach to enhance dataset availability, aiming to foster further research in the field and support broader application of LLMs.
|
I. Mikulić, M. Vlaić, G. Delač, M. Šilić, K. Vladimir (Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
Integrating External Knowledge with LLMs: A Systematic Review of RAG Approaches
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation, exhibiting exceptional in-context learning abilities without requiring extensive fine-tuning. However, these models often suffer from significant limitations, including hallucinations where fabricated or incorrect information is presented as fact, and the need for constant retraining in order to effectively integrate new knowledge. Retrieval-Augmented Generation (RAG) has emerged as a promising paradigm to address these challenges by combining LLMs with information retrieval (IR) techniques. By leveraging external knowledge bases, RAG pipelines retrieve and incorporate relevant information dynamically, enhancing the factual accuracy and adaptability of LLMs. This paper provides an overview of the state-of-the-art methods for implementing RAG in LLMs, examining key components of the RAG pipeline, including data preparation, retrieval, reranking, and post-retrieval techniques. We aim to highlight how these components collectively enable LLMs to achieve greater reliability and flexibility, paving the way for more robust AI applications
|
T. Sternak (UniZG Faculty of Electrical Engineering and Computing, Zagreb, Croatia), D. Runje (airt technologies d.o.o., Zagreb, Croatia)
Automating prompt leakage attacks on Large Language Models using agentic approach
This paper presents a novel approach to evaluating the security of large language models (LLMs) against prompt leakage—the exposure of system-level prompts or proprietary configurations. We define prompt leakage as a critical threat to secure LLM deployment and introduce a framework for testing the robustness of LLMs using agentic teams. Leveraging AG2 (AutoGen), we implement a multi-agent system where cooperative agents are tasked with probing and exploiting the target LLM to elicit its prompt.
Guided by traditional definitions of security in cryptography, we further define a prompt leakage-safe system as one in which an attacker cannot distinguish between two agents: one initialized with an original prompt and the other with a prompt stripped of all private information. In a safe system, the agents’ outputs will be indistinguishable to the attacker, ensuring that private or proprietary details remain secure. This cryptographically inspired framework provides a rigorous standard for evaluating and designing secure LLMs.
This work establishes a systematic methodology for adversarial testing of prompt leakage, bridging the gap between automated threat modeling and practical LLM security.
|
S. Ribić, R. Turčinhodžić Mulahasanović, K. Hodžić (Univerzitet u Sarajevu/ Elektrotehnički fakultet, Sarajevo, Bosnia and Herzegovina)
Torturing the Turing Test with Suggestive Anagrams in South Slavic Languages
The impressive results achieved by language recognition using a generative pre-trained transformer have led to divided opinion as to whether or not the Turing test has finally been passed. After understanding the working principles of the GPT program, it was observed that the tokenization concept that GPT uses to pass to the neural network that generates the text leads to the loss of word-to-letter association. Through about 20 specially prepared anagrams with a description of a term in a verse in the languages of the South Slavs, it was shown that ChatGPT is far more capable in understanding the semantic connection between words and allusions, than in the relatively simple task of searching for an adequate word from the offered letters.
|
V. Davidović (Polytechnic of Rijeka, Rijeka, Croatia), S. Martinčić-Ipšić (Faculty of Informatics and Digital Technology, Rijeka, Croatia)
Fine-Tuning BART for Croatian Abstractive Text Summarization Using a Novel News Dataset
The aim of abstractive text summarization is to generate a condensed version of a text while preserving its essential information. Recently, language models based on transformer architecture have shown promising results in generating summaries. However, these models are not explicitly tailored for the Croatian language. Moreover, they are trained in a supervised setup using text summarization datasets, but there is still no available dataset specifically designed for Croatian summarization. This paper addresses both of these challenges by adapting the BART (Bidirectional and Auto-Regressive Transformers) model for Croatian abstractive text summarization, using a novel Croatian news summarization dataset curated from online news portals. Specifically, we fine-tune the pre-trained BART model for Croatian by utilizing this new dataset. We also discuss the challenges and pitfalls involved in constructing the dataset, particularly when using large language models as annotation experts. The performance of the fine-tuned BART model is evaluated using ROUGE metrics. The results demonstrate improvement in the quality of Croatian text summarization.
|
| Break |
Natural Language Processing - II
Saction chair: Marko Horvat |
I. Dunđer, S. Seljan (Faculty of Humanities and Social Sciences, Zagreb, Croatia), M. Odak (Faculty of Humanities and Social Sciences, Mostar, Bosnia and Herzegovina)
Phishing Attacks in the Age of Embeddings: A Word-Vector Approach to Cyber Threats
The number of cyber threats such as phishing is steadily rising each year worldwide. Phishing attacks are a common type of cyber threats in which criminals pose as reliable organizations or trustworthy individuals in order to trick victims into revealing valuable information. Conventional detection techniques that rely purely on rule-based or heuristic methods struggle to keep up with the increasing number and complexity of phishing attacks. Developments in machine learning and natural language processing provide new ways to identify phishing attempts in order to overcome the difficulties in finding subtle patterns in the textual content of phishing messages. The aim of this paper is to examine whether by applying a machine learning technique known as word-vector representations to a phishing dataset it is possible to gain deeper comprehension about semantically similar words, and to increase understanding of potential phishing attempts. This is achieved by examining word embeddings in a continuous vector space model without the use of further resources, and only through an overview of the observed semantic similarities of the dataset’s textual features. Experimental results confirm the potential of the word-vector approach to phishing detection and the benefits of understanding the inherent characteristics of phishing datasets.
|
S. Delalić (Faculty of Science, University of Sarajevo & Info Studio d.o.o., Sarajevo, Bosnia and Herzegovina), Z. Kadrić, J. Jerkić, F. Mehmedović (Info Studio d.o.o., Sarajevo, Bosnia and Herzegovina)
Efficient CV Parsing: Lightweight Models and Heuristic Methods for Diverse and Non-Standard Formats
This paper addresses the challenge of analyzing CVs to parse their content into structured formats suitable for further processing and analysis. The proposed solution processes CVs provided as images or PDFs, handling diverse input formats, including free-form, multi-language, non-standardized layouts, and highly structured documents. Various heuristic approaches are employed for layout analysis, complemented by lightweight language models (LLMs) for extracting information. While multimodal models demonstrate strong performance, their cost and deployment complexity remain significant barriers. This study explores alternative methods optimized for computational efficiency, processing accuracy, and easier deployment. A comparative analysis of approaches is conducted on a standard dataset and a production dataset containing CVs from diverse clients and job roles, ranging from entry-level to specialized positions in various domanins. The findings highlight the potential of these tailored, efficient solutions for scalable and secure CV parsing.
|
A. Jankov, M. Toshevska, S. Gievska (Faculty of Computer Science and Engineering, Skopje, Macedonia)
Enhancing LLMs with LoRA Fine-Tuning using Medical Data and Knowledge Graph Enrichment for Improved Healthcare Outcomes
This research paper investigates the enhancement of large language models (LLMs) within the medical domain, focusing on members of the Llama family of LLMs. While LLMs have demonstrated remarkable success across various general-purpose natural language processing tasks, their application in specialized domains like medicine is often hindered by limited training on domain-specific data, resulting in suboptimal accuracy and contextual relevance. To address these limitations, this research employs low-rank adaptation (LoRA) to fine-tune LLMs on real-world patient-physician dialogues, effectively capturing the intricacies of medical discourse. Additionally, the knowledge of the LLM is enriched with the SPOKE knowledge graph, a structured repository of medical domain information, allowing the model to generate outputs that are both contextually and scientifically grounded. The experimental results underscore the transformative impact of this dual approach, demonstrating significant advancements in tasks such as automatic diagnosis generation and personalized drug recommendation. This work underscores the role of enriched LLMs in fostering innovation in medical informatics and shaping the future of healthcare delivery.
|
D. Bužić (College for Information Technologies, Zagreb, Croatia), J. Dobša (Faculty of Organization and Informatics, Varaždin, Croatia)
Emotions in Eurovision Song Contest Lyrics
The Eurovision Song Contest, started in 1956, is the longest-running annual TV music competition. So far, more than 1,500 songs have been sung at the competition, the lyrics of which contain numerous emotions. Emotion mining, one of the areas of sentiment analysis, deals with the automatic detection of emotions in text. In this paper, we present the results of emotion detection in the lyrics of songs from this international music show. Emotion detection was performed by NRC emotion lexicon. We found that the lyrics are dominated by emotions of joy, trust, and sadness, while disgust is the least represented. It is particularly interesting that over the decades, there has been a decrease in positive emotions (joy and trust) in songs, while negative emotions (anger, disgust, fear and sadness) have increased.
|
J. Meena, I. Jha, P. Girdhar (Delhi Technological University, New Delhi, India)
Customizable Chatbots for Real Estate - An LLM Driven Approach to Information Retrieval
One of the major challenges in real estate is the ability to retrieve information from legal documents. We suggest building a customized chatbot that is able to query the real estate related legal pdf documents in this paper. We then use some recent progressions in Large Language Models and retrieval-based question answering to introduce a system that uses this information. It uses a sentence transformer on top of the text to let it talk to a FAISS vector index, so you can do a semantic fast search. Utilizing data up through October 2023, the chatbot's responses are crafted with the power of Google Gemini Pro, combining the TF-IDF process that ascertains relevant content; this also means that answers are accurate and contextual. It enhances engagement through fast and accurate response for complex queries. This solution demonstrates the true power of utilities like LangChain in the process of extracting unstructured legal text into something understandable and digestible which in turn boosts client communication, making the operational aspects of the firm more efficient. In addition, the paper discusses ethical issues pointed out using language models, other data storage methodologies and outlines central limitations. This new research illustrates the scalable and customizable chatbot solutions available to best support the broader ecosystem of real estate.
|
J. Virtanen (Department of Mathematics and Statistics, Tampere University, Tampere, Finland), M. Toshevska (Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, Skopje, Macedonia)
A Comparison of GEC Tools for Grammatical Error Correction in English
Using the Building Educational Application (BEA) benchmark1, this study compares the capabilities of Google Gemini2, ChatGPT3, DeepSeek4, and the built-in grammar checkers in Google Docs and Microsoft Word for grammatical error correction (GEC). These tools correct a variety of errors, some of which overlap. Based on the BEA evaluation results, Google Gemini and the Google Docs grammar checker achieve the best scores of 60.2 and 65.86, respectively. Google Docs grammar checker is easy to use and, according to this evaluation, performs well, thus proving to be a viable option for GEC. However, standard grammar checkers are not typically designed for rewriting text to the same extent as GenAI tools; hence, it may be advisable especially for non-native speakers to combine traditional and GenAI grammar correction for the best possible results. However, it is necessary to check the grammatical corrections of LLMs, since generative AI tools suffer from hallucinations, which refers to the tendency of LLM chatbots to generate information that can be factually incorrect.
|
Artificial Intelligence Applications and Other Topics – I
Section chair: Alan Jović |
M. Vlaić, I. Mikulić, G. Delač, M. Šilić, K. Vladimir (Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
A Review of Time Series Dimensionality Reduction Methods |
D. Kopljar, V. Drvar (CROZ AI, Zagreb, Croatia), J. Babić, V. Podobnik (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia)
xAAD in Practice: Explainable Anomaly Detection in Cybersecurity
Anomaly detection systems are increasingly crucial across various industries, yet ensuring their outputs align with real-world needs remains challenging. Previously, we introduced xAAD—integrating Active Anomaly Discovery with Isolation Forest-based explainability—to enable expert-guided refinement of anomaly rankings and enhanced feature-level insights. This paper moves beyond theoretical foundations to showcase xAAD’s practical value in a high-stakes domain: cybersecurity.
In this scenario, xAAD helps analysts quickly identify and interpret abnormal network activities. By incorporating expert feedback, it not only improves detection accuracy but also provides clear explanations for why certain events are flagged as anomalies. This transparency empowers cybersecurity teams to respond more effectively, reducing investigation times and supporting informed decisions under pressure.
While xAAD’s adaptable framework can serve diverse stakeholders—developers, domain experts, regulators, and end-users—this paper emphasizes its tangible benefits in a complex, real-world setting. By aligning technical outputs with operational priorities, xAAD transforms explainable anomaly detection from a research concept into a practical asset. Our results bridge the gap between theoretical advances and industry demands, offering a roadmap for practitioners seeking to deploy explainable, feedback-driven anomaly detection systems across various domains.
|
D. Oreščanin (Legit Software d.o.o., Zagreb, Croatia), V. Kužina (Data Driven d.o.o., Zagreb, Croatia)
Engine for Discovery of Personal Data Based on Machine Learning
Although numerous privacy laws existed prior to its enactment, the GDPR has brought privacy issues to the forefront of regulatory attention. Organizations must ensure that their personal data management processes, particularly those involving the collection, storage, and processing of data within IT systems, comply with regulatory requirements. Personal data discovery in IT systems refers to the process of identifying and cataloguing personal data across an organization's digital infrastructure. It involves scanning databases, file systems, and applications to locate sensitive information such as names, addresses, or financial details. By automating this process, organizations can better understand their data landscape, manage risks, and ensure proper handling of Personal identifiable information (PII). Personal data discovery is a critical first step for any later activity related to the personal data in IT systems in organizations. This paper describes the design of the Personal Data Discovery Engine (PDDE), process flows and how main modules and their functionalities were implemented. A Pseudo algorithm based on Machine Learning (ML) for discovery and classification of personal data from heterogenous text and tabular datasets found in IT Systems is defined.
|
V. Vodilovska, N. Ackovska, I. Ivanoska (Ss Cyril and Methodius University, Skopje, Macedonia)
Leveraging Graph Attention Networks for Blood-Brain Barrier Permeation Prediction
The development of pharmaceuticals requires an extensive screening of molecules to evaluate their biochemical properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET). In this work, we explore the combination of Graph Neural Networks (GNNs) for the prediction of ADMET properties, with a specific focus on Blood-Brain Barrier (BBB) Permeation. Our GNN model performs consistently well across all metrics, evaluated on three different benchmark datasets. The main takeaway is the ability to maintain performance on unknown datasets that contain chemically diverse compounds, which we validated using the Tanimoto Similarity comparison. Our results demonstrate that GNNs can reliably predict BBB Permeation. Next steps for improvement include experimenting with different molecular features, model architectures, and exploring transfer learning strategies. Finally, the GNN model we developed for predicting BBB Permeation is designed to be general and can be extended for different small molecule prediction tasks, supporting both SMILES and InChi data formats. Our focus in future experiments is to extend this model to predict other ADMET properties.
|
L. Batistić, S. Ljubic (University of Rijeka, Faculty of Engineering, Rijeka, Croatia), K. Griparić, D. Sušanj (Juraj Dobrila University of Pula, Faculty of Engineering, Pula, Croatia)
Comparison of Image-based Deep Learning Classification Models for Motor Imagery EEG Signals
Motor impairments significantly impact individuals’ lives, prompting researchers to develop technologies like brain-computer interfaces (BCIs) that utilize electroencephalography (EEG) to restore or replace lost abilities. This research investigates the use of image representations derived from EEG signals for classifying motor imagery (MI) signals recorded via EEG-based BCI. MI is a key technique for BCI control, enabling users to imagine movements to interact with devices. Traditional MI classification often relies on analyzing EEG signals directly. This study explores a different approach by transforming EEG data into images and leveraging advanced neural network architectures, inspired by successes in image recognition. The research systematically evaluates different neural network architectures to identify the most effective method for accurate MI classification. Furthermore, the study will assess the generalizability of the proposed approach by applying it to two datasets.
|
| Break |
Artificial Intelligence Applications and Other Topics - II
Section chair: Alan Jović
|
L. Dobruna, A. Berisha (University of Prishtina, Prishtina, Kosovo)
A Data-Driven Approach for Predicting Solar Energy Production using Machine Learning Techniques
Machine learning techniques have a very wide application in various fields, including also renewable energy, which is now regarded as highly important topic due to its minimal negative impact on the environment. Solar energy is one of the most important renewable energy sources, and its installation capacity continues to grow. Forecasting solar energy production is a very high interests topic nowadays. In this paper, various machine learning algorithms are analysed, together with some hyperparameter optimization methods for prediction of solar energy production. Different methods for forecasting solar energy production based on selected performance metrics are reviewed and analyzed. The dataset is built based on real data collected from a solar park in Kosova and meteorological institute of Kosova and the results have been derived using the Python programming language. From the results, we concluded that the Extreme Gradient Boosting algorithm, along with Sequential Model- Based Optimization hyperparameter optimization method achieve the best result, in terms of forecasting the energy production from the solar panels, with an accuracy of 95.67%.
|
P. Medur (Faculty of Engineering, University of Rijeka, Rijeka, Croatia), M. Lubbers (Radboud University, Institute for Computing and Information Sciences, Nijmegen, Netherlands), G. Mauša (Faculty of Engineering, University of Rijeka, Rijeka, Croatia)
Optimizing Keyword Spotting Classifier based on Tiny Machine Learning for Low-Power Embedded Devices
Keyword spotting (KWS) is essential for enabling voice recognition in smart home systems and although deploying reliable and energy-efficient models on low-power embedded devices remains a challenge, tiny machine learning offers a promising solution. In this paper, raw audio data from Google’s Speech Commands Dataset v0.02 was used to develop and evaluate neural network architectures optimized for KWS based on Mel-frequency cepstral coefficients. The audio signals are formatted as compact 2D matrix of time-frequency features, making it ideal for convolutional neural networks. Ten networks were trained with 10-fold cross-validation, fine-tunned on speaker-specific data, and, by using post-training quantization, converted from TensorFlow (TF) to TensorFlow Lite (TFLite) format for deployment on microcontrollers. The trade-offs between model size, performance and energy efficiency are analyzed, and the results showed that the TFLite models considerably reduced the energy consumption, with energy savings ranging from 1-4% for smaller capacity models to 99.73% for larger ones, while maintaining performance similar to TF models. Furthermore, fine-tuning models of greater capacity slightly improved their accuracy (0.51%) and energy efficiency (1.75%), whereas smaller models were unaffected by fine-tuning. This study lays the foundation for the integration of KWS models based on tiny machine learning into microcontrollers for real-time applications.
|
M. Banjanin (Department of Computer Science and Systems, Faculty of Philosophy Pale, University of East Sarajevo,, Belgrade, Serbia), M. Stojčić, M. Vasiljević, A. Stjepanović, G. Jotanović, M. Husić, G. Jauševac (Faculty of Transport and Traffic Engineering Doboj, University of East Sarajevo, Republic of Srpska,, Doboj, Bosnia and Herzegovina)
Prediction of the Number of Parcels Sent Via Express Mail Service in a Time Series Using Statistical and Machine Learning Models
The prediction of the number of parcels sent via express mail during a working day from a single postal center to multiple destination locations is crucial for optimizing the logistics and operational processes of postal operators. Operators can identify seasonal and daily patterns in the time series of parcel shipments, enabling more efficient decision-making through analysis. Expected benefits include improved planning of human, transport, and storage resources and capacities, as well as enhanced customer experience. This study focuses on the Zenica Postal Center in Bosnia and Herzegovina (BiH), where the time series represents the total number of parcels sent hourly. The designed database encompasses a time series spanning 23 days and 5 hours, with each working day lasting 12 hours, resulting in a total of 281 observations. Traditional statistical models, such as Seasonal AutoRegressive Integrated Moving Average (SARIMA) and Prophet, were developed for time series analysis and prediction, alongside models based on machine learning techniques, including Support Vector Regression (SVR) and Artificial Neural Network (ANN). Each model was tested on 10% of the total dataset observations, and the evaluation and comparison of predictive performance were conducted using Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics.
|
B. Okreša Đurić (Fakultet organizacije i informatike, Varaždin, Croatia), C. Carrascosa (Departamento Sistemas Informáticos y Computación, Universitat Politècnica de València, Valencia, Spain)
Ontology-Driven Multiagent System Generation: A Framework
This paper presents a novel framework for developing multiagent systems using ontology-based modelling. Our method systematically maps concepts and relationships to agent roles, behaviours, and interactions by capturing domain knowledge in an ontology. The system automatically generates implementation templates for the modelled multiagent system from this structured representation. The approach reduces the need for extensive manual coding, ensuring that conceptual integrity is maintained throughout development.
A case study showcases how the framework can speed up the design cycle, reduce coding errors, and enhance system extensibility. Using ontology-driven generation ensures consistency and interoperability, as each agent artefact follows the established domain semantics. This approach is especially beneficial in complex, knowledge-intensive environments that require collaborative and distributed solutions. Ultimately, it leads to a flexible and scalable process that evolves from domain modelling to multiagent implementation.
Our framework provides a compelling strategy for bridging the gap between conceptual design and agent-based execution, streamlining multiagent system development and ensuring robust alignment with domain-specific ontologies.
|
D. Demeterfi, T. Kovačević, Z. Kostanjčar (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia), V. Bilas (Notitia Ltd, Zagreb, Croatia)
Bridging the Gap between Model Complexity and Data Availability for Forecasting Tourist Activity
Recent years have witnessed a surge in the application of complex machine learning (ML) models across industrial and scientific domains, driven by their ability to outperform traditional statistical methods in specific tasks. However, the efficacy of these models is closely linked to the availability and quality of data, which often limits the complexity and scope of feasible ML approaches. In this paper, we present a novel hybrid forecasting approach that combines ML with classical parametric curve fitting to mitigate data limitations. Specifically, our method first employs an XGBoost-based regression model to predict the residual annual tourist spending from early-season data. These predictions are then incorporated into a generalized logistic growth curve to model the cumulative spending dynamics, yielding interpretable parameters such as total expenditure, seasonal intensity, and peak activity timing. Focusing on the early prediction of annual tourist activity in the Adriatic Croatia region, our empirical evaluation demonstrates that the proposed hybrid model not only outperforms benchmark methods but also delivers actionable insights for real-world tourism planning.
|
|
Basic information:
Chairs:
Darko Huljenić (Croatia), Alan Jović (Croatia)
Steering Committee:
Andrea Budin (Croatia), Bojan Cukic (United States), Marko Đurasević (Croatia), Marina Ivašić-Kos (Croatia), Domagoj Jakobović (Croatia), Ruizhe Ma (United States), Neeta Nain (India), Stjepan Picek (Netherlands), Slobodan Ribarić (Croatia), Vitomir Štruc (Slovenia)
Program Committee:
Karla Brkić (Croatia), Marko Čupić (Croatia), Marko Đurasević (Croatia), Nikolina Frid (Croatia), Marko Horvat (Croatia), Tomislav Hrkać (Croatia), Ivo Ipšić (Croatia), Domagoj Jakobović (Croatia), Zoran Kalafatić (Croatia), Georgina Mirceva (North Macedonia), Josip Šarić (Croatia)
Registration / Fees:
|
REGISTRATION / FEES
|
Price in EUR
|
EARLY BIRD
Up to 23 May 2025 |
REGULAR
From 24 May 2025 |
| Members of MIPRO and IEEE |
270 |
297 |
| Students (undergraduate and graduate), primary and secondary school teachers |
150 |
165 |
| Others |
300 |
330 |
The student discount doesn't apply to PhD students.
NOTE FOR AUTHORS: In order to have your paper published, it is required that you pay at least one registration fee for each paper. Authors of 2 or more papers are entitled to a 10% discount.
Contact:
Darko Huljenić
Faculty of Electrical Engineering and Computing
Unska 3
HR-10000 Zagreb, Croatia
E-mail: huljenicdarko@gmail.com
Alan Jović
Faculty of Electrical Engineering and Computing
Unska 3
HR-10000 Zagreb, Croatia
Phone: +385 1 612 9548
E-mail: alan.jovic@fer.hr
The best papers will get a special award.
Accepted papers will be published in the ISSN registered conference proceedings. Papers in English presented at the conference will be submitted for inclusion in the IEEE Xplore Digital Library.
Location:
 Opatija is the leading seaside resort of the Eastern Adriatic and one of the most famous tourist destinations on the Mediterranean. With its aristocratic architecture and style, Opatija has been attracting artists, kings, politicians, scientists, sportsmen, as well as business people, bankers and managers for more than 180 years.
The tourist offer in Opatija includes a vast number of hotels, excellent restaurants, entertainment venues, art festivals, superb modern and classical music concerts, beaches and swimming pools – this city satisfies all wishes and demands.
Opatija, the Queen of the Adriatic, is also one of the most prominent congress cities in the Mediterranean, particularly important for its ICT conventions, one of which is MIPRO, which has been held in Opatija since 1979, and attracts more than a thousand participants from over forty countries. These conventions promote Opatija as one of the most desirable technological, business, educational and scientific centers in South-eastern Europe and the European Union in general.
For more details, please visit www.opatija.hr and visitopatija.com.
|
|
|
|
| Currently there are no news |
|
|
|
|
|