Abstracts for Session on Information Systems

• Karolis Trinkūnas, Jolanta Miliauskaitė, “AI-Based Evaluation of Business Requirement Changes Based on Stakeholder Feedback”

  Requirements engineering is essential for defining, documenting, and validating stakeholder needs in software development. However, business requirements written in natural language often contain quality defects such as ambiguity, inconsistency, incompleteness, subjectiveness, vagueness, non-verifiability, and negativity. These issues make it difficult to assess requirement quality and determine whether stakeholder feedback has been implemented correctly. This study investigates artificial intelligence-based methods for evaluating business requirement changes based on stakeholder feedback. The research includes an analysis of related work in requirements engineering, requirement quality assessment, and artificial intelligence applications for requirement processing. It also examines available datasets, AI training methods, and requirement quality indicators relevant to the identification of requirement defects. Particular attention is given to the detection of ambiguity, subjectiveness, vagueness, non-verifiability, and negativity using multi-label text classification models. To evaluate stakeholder feedback implementation, the proposed method combines artificial intelligence models with a hierarchical fuzzy inference system. In this framework, multi-label text classification is used to identify requirement defects, while the fuzzy inference component supports higher-level evaluation of requirement quality and the correctness of feedback-driven changes. This hybrid approach enables a more systematic assessment of business requirement modifications. The study concludes that artificial intelligence methods can support requirements engineering by improving the evaluation of business requirements and stakeholder feedback implementation. The proposed framework enables objective and repeatable comparison of different stakeholder feedback implementation methods and provides a structured approach for evaluating requirement changes.

• Anastasija Safonova, Diana Kalibatienė “Applying Fuzzy Inference and Machine Learning for Prediction of Floods: Lithuanian Case Study”

  This study presents the development of a flood prediction system based on the Adaptive Neuro-Fuzzy Inference System (ANFIS), with a focus on lowland rivers in Lithuania. The results demonstrate that the proposed ANFIS model achieves a high level of predictive accuracy and exhibits performance characteristics comparable to advanced machine learning approaches, including recurrent neural networks (RNN) and XGBoost. In flood event scenarios, the ANFIS model outperforms other machine learning methods in terms of accuracy. A key advantage of the proposed approach lies in its interpretability and transparent “white-box” structure, which enables detailed analysis of individual decision-making processes. These properties make the model a promising tool for application as an intelligent decision-support system in hydrometeorological services.

• Deborah Amenigy “The Evolution of Machine Learning Models in Financial Anomaly Detection” - winner in category "Best Visualisation"

  Machine learning now plays a key role in detecting financial anomalies, but research in this area has evolved unevenly. This systematic review analyses 46 peer-reviewed studies published from 2011 to 2026 using a PRISMA-guided process across Web of Science and Google Scholar. It tracks how learning methods have shifted from traditional classifiers to deep learning and ensemble techniques, with Random Forest and Artificial Neural Networks emerging as the most frequently used models. A notable finding is the focus on fraud: 44 of 46 papers target deliberate financial manipulation, while unintentional errors remain largely unexplored. The review also highlights a consistent gap between model performance on curated datasets and on real-world corporate data, raising concerns about the applicability of published results. These findings highlight the importance of conducting more extensive research, particularly focusing on unintentional errors, and developing more robust evaluation methods that accurately reflect the complexity of real-world financial data

• Davina Racine Jackson, Diana Kalibatienė “Development Project Quality Assessment Using Event-Based Methods”

  Open-Source Software serves as critical digital infrastructure yet lacks formal reliability guarantees. Traditional evaluation metrics often fail to capture complex socio-technical risks, such as the "Bus Factor" (i.e., centralization risk) or the operational trade-off between development speed and code scrutiny. This study proposes a novel, hybrid event-based analysis methodology to extract, visualize, and quantify these hidden reliability issues from repository data.

  The methodology follows a two-phase framework. In the first phase, a custom Python algorithm extracted event data from fifteen diverse GitHub repositories. Spearman correlation analysis established the statistical relationship between four metrics: Review Rigor, PR Merge Ratio, Issue Resolution Rate, and Contributor Diversity Index. The analysis quantified a "Speed vs. Rigor" trade-off (i.e., ρ=−0.39), informing an objective weighting system. In the second phase, a repeatable workflow was implemented, triangulating three paths: process mining visualization (i.e., ProM), statistical Multi-Criteria Decision Making (i.e., WASPAS), and deterministic AI validation (i.e., Gemini 2.5 Pro).

  Experimental results demonstrate that process models translate abstract metrics into visible structural behaviors, revealing bifurcated resource bottlenecks versus iterative quality-control loops. A critical finding emerged from the AI-driven validation demonstrating that while the mathematical WASPAS model effectively measures operational efficiency, deterministic AI acts as a strict auditor capable of uniquely detecting existential "Bus Factor" risks that linear math obscures. Additionally, the work includes an initial proof-of-concept for an automated cloud pipeline, demonstrating the potential of the proposed framework for real-time reliability monitoring. This holistic framework provides multi-dimensional insights, moving beyond simple health checks to ensure long-term repository resilience.

• Deividas Kalvelis, Diana Kalibatienė “Research on Machine Learning Validation and Verification System”

  Machine learning (ML) models are increasingly deployed in critical domains such as healthcare and finance, yet their „black box“ nature poses significant risks regarding reliability, fairness and robustness. Traditional software testing methods are often insufficient for ML systems due to their non-deterministic behavior and dependence on data quality. This paper addresses these challenges by proposing a novel and comprehensive approach for ML model Validation and Verification (V&V). The approach integrates data integrity checks, fairness assessment, adversarial robustness testing, and probability calibration into a unified lifecycle. To validate the proposed approach, a comparative case study was conducted using the SIIM-ISIC Melanoma Classification dataset. The results demonstrate that the proposed approach significantly outperforms a standard baseline pipeline, achieving a 460% improvement in Recall and a 202% improvement in F1-score, while successfully identifying demographic biases and adversarial vulnerabilities that standard metrics failed to detect. The findings underscore the necessity of a holistic V&V strategy to ensure the development of trustworthy and ethical AI systems.

• Dilki Sandunika Rathnayake, Asta Slotkienė “Evaluating Knowledge Graph Enhanced RAG for Automated Functional Requirements Extraction”

  Functional requirements (FRs) are essential for software design, development, and validation, yet manual extraction from diverse sources is labor-intensive, error-prone, and hard to scale. Large Language Models (LLMs) enable automation but often hallucinate and lack precision in regulated domains like healthcare, where compliance with standards such as HIPAA and IEEE 830 is critical. Retrieval-Augmented Generation (RAG) improves LLM reliability by grounding outputs in verified documents, though standard vector-based RAG overlooks complex semantic relationships needed for accurate FR identification.

  This study compares baseline Vector RAG against Knowledge Graph-Enhanced RAG (GraphRAG) using two knowledge bases (standard and refined) and three prompt strategies (Zero-Shot, Few-Shot, Chain-of-Thought) on a corpus of seven healthcare documents. Performance is assessed via RAGAS-adapted metrics: Faithfulness, Relevance, Technical Coverage, Compliance, and Overall Score.

  Results reveal knowledge base quality as the primary driver of extraction success. On the standard KB, GraphRAG extracted more FRs (32 vs. 20) but suffered relevance drops (0.42 vs. 0.53). The refined KB optimized precision, reducing extractions by 17% (43 total) while boosting scores, GraphRAG achieved the highest overall (0.83) and compliance (0.88). Few-Shot prompting with GraphRAG on refined KB yielded top gains, emphasizing curation's role in mitigating noise and false positives, vital for high-stakes healthcare applications.

  Key insight: GraphRAG excels conditionally with curated KBs, prioritizing precision over recall. Future work should refine graph construction and scale to larger corpora, advancing automated requirements engineering.

• Povilas Unčiuris, Dalius Mažeika “Architectural Refactoring from Monolithic Applications to Microservices”

  Modern information systems are continuously facing increasing technological and organizational challenges in ensuring rapid adaptation to changing business needs, scalability, and performance. Large systems that have been developed over a long period based on traditional architectures often become too slow and too complex to meet these requirements. This is especially evident when there is a need to quickly deploy new functionality.

  Organizations are increasingly considering migration to more modern architectures, such as microservices, because this allows them to manage growing technological demands more effectively. However, this transition is not simple. Refactoring a system’s architecture requires deep technical knowledge and tools that can help decompose systems in a way that meets new standards of flexibility, autonomy, and resilience. Current solutions and methods in this area are often not sufficiently comprehensive or accurate, and there is no de facto standard for architecture migration, which makes it difficult to choose from the many existing methods and strategies.

  To make this process easier and more adaptable to different situations and systems, it is necessary to develop new methods that would enable more effective identification of potential system parts that could be separated as independent components. Particularly important are automated solutions that would help analyze system semantics, identify logical boundaries, and propose an optimal architectural transformation.

• Eimantas Repšys “Development of an Artificial Intelligence (AI)-Based Computer Network Management Assistant”

  AI-powered Network Operations Assistant that combines Large Language Models (Claude) with the Model Context Protocol (MCP) to enable natural language interaction with live network devices.

  The system implements a dual-server MCP architecture separating read-only observation tools from write operations, enforcing a two-step propose-and-confirm workflow for all configuration changes. A FastAPI backend serves as the orchestration layer, routing natural language queries through Claude, which autonomously selects and executes from 30 MCP tools — including health checks, topology discovery, configuration management, and compliance validation — all operating over SSH connections to Cisco IOS devices via Netmiko.

  The assistant is deployed as an Electron desktop application spawning three concurrent Python processes — web interface, scheduler, and syslog server — providing a unified interface for monitoring, diagnostics, and controlled remediation. All write operations are audit-logged with user attribution, and pre-change impact analysis automatically flags high-risk modifications.

• Anastasija Grubinskienė, Andrius Katkevičius “Wearable Physiological Sensor Fusion for Pain Episode Detection”

  Wearable-based pain episode detection and forecasting is emerging as a promising way to complement subjective and irregular self-reported pain assessments. In everyday life, however, physiological changes are influenced not only by pain, but also by movement, emotional arousal, sleep, and environmental context. As a result, the main challenge is not simply to identify informative signals, but to design robust fusion strategies that remain reliable under real-world conditions.

  This presentation reviews recent literature on wearable physiological sensing for pain assessment and forecasting, focusing on signal fusion, signal quality management, and context-aware modeling. The reviewed studies cover acute clinical pain monitoring, long-term chronic pain assessment, and episodic pain forecasting, including migraine prediction. The most frequently used signal families include electrodermal activity, photoplethysmography, heart rate and heart rate variability, motion sensing, temperature, and, in more specialized settings, EEG and fNIRS.

  Two practical synthesis ideas are highlighted. The first is the 2+1 rule, which suggests that reliable wearable pain episode detection typically requires at least two physiological signal families and one context channel, such as motion or sleep segmentation. The second is the KKS principle (quality- and context-aware synthesis), which treats signal quality indicators and contextual information as active inputs that dynamically guide fusion and decision-making.

  The presentation argues that single-signal solutions are usually too fragile for real-world use, while multimodal, quality-controlled, and context-aware fusion provides a more robust direction for practical wearable pain monitoring. Future work should emphasize stronger validation across individuals and contexts, better protocol comparability, and clinically meaningful evaluation of generalizability.

• Dmytro Teplov “Investigation of the Efficiency of Artificial Neural Networks Trained with a Small Amount of Data”

  Artificial neural networks are widely used to solve complex problems in areas such as computer vision and natural language processing, but their performance typically depends on large volumes of labeled data. This reliance presents a major limitation in domains where data collection is costly, time-consuming, or ethically constrained, such as medical imaging and materials science. As a result, developing methods that enable effective learning from small datasets has become an important research focus.

  Several approaches have been proposed to address this challenge, including data augmentation, transfer learning, and synthetic data generation. These techniques aim to increase data diversity, leverage existing knowledge, and reduce overfitting, thereby improving model performance under limited data conditions. While promising, their comparative effectiveness and practical application remain areas of ongoing research.

  This thesis investigates the efficiency of such methods through systematic evaluation and experimental comparison. In particular, it focuses on training object detection neural networks with limited labeled data, aiming to provide practical insights into data-efficient model development.

• Priyankesh Shivam “Research on The Development of Sustainable Information Systems”

  This thesis investigates the relationship between SQL query complexity and energy consumption in relational database systems, contributing to the broader field of sustainable information systems. As modern data-driven applications increasingly rely on database management systems, their energy footprint has become a critical concern, particularly in large-scale data centers. Understanding how software-level factors such as query design and execution strategies influence power usage is essential for improving both performance and environmental sustainability. An experimental framework is developed to systematically analyze the energy behavior of SQL queries executed on a Linux-based system using a PostgreSQL database populated with the TPC-DS benchmark dataset. Energy consumption is measured at the hardware level using Intel’s Running Average Power Limit (RAPL) interface, enabling precise and fine-grained observation of CPU package energy usage. Queries are categorized into simple, medium, and complex classes based on structural characteristics such as joins, aggregations, and subqueries, and are evaluated under controlled conditions. The study first establishes baseline measurements in a single-user environment, identifying clear correlations between query complexity, execution time, and energy consumption. It then extends the analysis to concurrent workloads, examining how varying levels of parallel query execution impact overall energy efficiency, throughput, and latency. The results reveal that increasing concurrency does not always lead to proportional performance gains and can, beyond certain thresholds, result in higher energy costs with diminishing returns. Building on these findings, the thesis explores the concept of energy–performance trade-offs and identifies optimal operating regions where systems can maximize computational work while minimizing energy consumption.

• Rokas Povilionis “Design and Implementation of a Domain Specific Modelling Language”

  This presentation introduces the design and implementation of a domain-specific modeling language (DSML) aimed at improving rapid application development within the Ruby on Rails ecosystem. As modern software systems grow in complexity, traditional general-purpose modeling and programming approaches often fail to provide sufficient abstraction, automation, and domain alignment. This research addresses these limitations by proposing a DSML that integrates domain-driven design (DDD) principles with model-driven engineering (MDE) and AI-assisted code generation.

  The core contribution of the thesis is a novel methodology that enables developers to model system architectures visually using a constrained DSML, and automatically transform these models into functional application code. The approach leverages graph-based modeling tools, structured metadata, and a Model Context Protocol (MCP) to bridge the gap between high-level design and implementation. By combining diagram parsing, rule-based validation, and large language model (LLM) integration, the system ensures that generated code adheres to architectural best practices while reducing manual development effort.

  A prototype implementation demonstrates how domain models expressed through concepts such as bounded contexts, aggregates, entities, and services can be translated into Ruby on Rails code. The DSML enforces well-formed rules and semantic correctness, ensuring consistency and maintainability of both models and generated artifacts. Additionally, the use of AI enhances flexibility and adaptability, enabling context-aware code generation based on both diagrams and existing codebases.

  The results indicate that this approach enhances development efficiency, reduces boilerplate code, and strengthens the alignment between system design and implementation.

• Elvin Jabbarli “Research on Website Accessibility Compliance”

  This research focuses on web accessibility as an essential requirement for ensuring equal access to digital material for all users. The Web Content Accessibility Guidelines (WCAG) 2.1 by the World Wide Web Consortium (W3C) is an important framework for evaluating and improving website accessibility. However, most existing evaluation methods focus on one approach and thus limit assessment scope. Automated tools like WAVE, SiteImprove, TAW, and axe DevTools excel at pointing out basic structural issues but are insufficient for detecting semantic, contextual, and interactive barriers. Expert manual evaluation and user-based testing compensate for such deficits; however, they are almost never combined.

  In this thesis, we propose an evaluation methodology that incorporates automated techniques along with expert manual evaluation and user testing within an established framework. This methodology is based on the approach recommended by Wohlin et al. (2012) and was experimentally tested using five actual websites from various areas such as government institutions, e-commerce, news portals, education resources, and university sites. The sample consisted of ten accessibility experts and eleven users with different accessibility requirements who evaluated fifteen WCAG 2.1 AA success criteria.

  To fill the gap between raw evaluation data and practical development advice, a prototype using AI support was designed. This prototype uses a large language model (Gemini) for processing mixed evaluation data and generating JSON reports with per-criterion severity ratings, problem descriptions, and suggestions for resolving issues along with relevant code examples.

  It is found that the combination of methods provides more extensive identification of accessibility barriers compared to individual methods considered alone.

• Samral Feyziyev “Semantic Analysis”

  This presentation explores the growing impact of digital transformation on modern society, focusing on how emerging technologies are reshaping communication, education, and everyday life. In recent years, rapid advancements in artificial intelligence, data analytics, and digital platforms have significantly influenced the way individuals interact, learn, and make decisions. The purpose of this presentation is to examine both the opportunities and challenges associated with this transformation.

  The discussion begins by outlining key technological developments and their role in creating a more connected world. It highlights how digital tools have improved access to information, enabled remote collaboration, and increased efficiency across various sectors. Particular attention is given to the field of education, where online learning environments and digital resources have expanded opportunities for students worldwide.

  However, the presentation also addresses the potential drawbacks of digitalization. Issues such as data privacy, information overload, and digital inequality are critically analyzed. Not all individuals have equal access to technology, which can deepen existing social and economic disparities. Furthermore, the increasing reliance on digital systems raises concerns about cybersecurity and ethical use of data.

  The presentation concludes by emphasizing the importance of responsible innovation and digital literacy. It suggests that individuals, institutions, and policymakers must work together to ensure that technology is used in a way that benefits society as a whole. By understanding both the advantages and risks, we can better navigate the digital future and create a more inclusive and sustainable world.

• Eugenijus Jončas, Arūnas Ribikauskas “Subject Matter Language Recognition Using Training”

  An experiment is designed to develop an ASR system for a specific area or field in which it will be used. General use models may do well in casual conversational settings but are generally ineffective when dealing with industry-specific, technical terms, because there are no databases that include this type of information and therefore have no basis for understanding how the words sound when spoken. The purpose of the research is to create a working ASR model for Lithuanian speech that has been dictated by expert.

  The ASR Model was created using the open-source Kaldi toolset. In addition to utilizing the open-source Kaldi toolset as its base, the ASR Model also utilized the Mozilla Common Voice corpus and a private dataset of pre-recorded expert speech. For extracting features based on acoustic properties of the recorded audio, Mel-Frequency Cepstral Coefficients (MFCC) were employed. Additionally, in order to improve upon a basic acoustic model that would be produced from monophone units; the acoustic model was further developed into a context dependent triphone unit. Finally, the MFCC's were transformed using Linear Discriminant Analysis (LDA), Maximum Likelihood Linear Transformations (MLLT) and Feature Space Maximum Likelihood Linear Regression (fMLLR) in order to increase the overall robustness and ability of the ASR Model to adapt to different speakers.

  Using a baseline Unigram Language Model the ASR System achieved a total word error rate (WER) of 59.06%. Although the WER could be improved through the utilization of a more sophisticated language model and increased number of hours of training data; the ASR System successfully completed processing of each of the 5517 test utterance and produced output that was linguistically accurate. As such this experiment provided a viable starting point for developing additional ASR Systems for Lithuanian that can provide very high precision in professional environments.

• Džiugas Pumputis, Asta Slotkienė “Method of Generation of System Tests from Requirements”

  Presentation introduces a method for generating system test cases from software requirements using large language models (LLM). Usual test case creation is a process that requires significant human effort and deep understanding of system requirements. Recent advances in natural language processing (NLP) and generative AI make it possible to automate parts of this process by transforming textual requirements into structured testing scenarios.

  The proposed approach focuses on generating test scenarios from non-functional requirements using several LLMs. The method involves generation of structured test cases that include preconditions, steps, and expected results. Generated scenarios are evaluated using an automated “LLM-as-a-judge” framework that assesses quality using multiple metrics such as coverage, boundary value analysis, logical correctness, and verifiability.

  An experimental evaluation compares several modern LLMs in terms of generated test quality, processing time, and cost efficiency. The results show that LLM-based approaches can significantly assist in software testing by automating test creation and improving requirement coverage while also reducing manual effort.

• Karol Ravdo, Rokas Štrimaitis “Real-Time Data Classification Using Machine Learning Models”

  Modern data-driven systems often need to process live data and make predictions in real time. In such environments, classification becomes difficult because the data may change over time and true labels may arrive only after a delay. These challenges are known as concept drift and delayed supervision, and they can reduce the effectiveness of standard machine learning methods. This presentation introduces a study on real-time stream classification using incremental learning. The research focuses on how concept drift, delayed supervision, and different adaptation strategies affect classification performance. A real-time experimental environment was designed to simulate streaming conditions and evaluate model behaviour under changing data distributions. The experimental analysis was carried out on three datasets with different drift characteristics: SEA with abrupt drift, Rotating Hyperplane with gradual drift, and Elec2 as a real-world dataset with irregular and noisy drift. Several fixed adaptation strategies were compared, including reset-based and gradual adaptation methods. In addition, an adaptive strategy-selection approach called AutoBandit was evaluated. The results show that delayed labels reduce the efficiency of concept drift adaptation and lower overall classification accuracy. They also show that no single fixed adaptation strategy performs best in all scenarios. AutoBandit achieved the strongest results in uncertain and changing conditions, especially when supervision was delayed. These findings suggest that adaptive strategy selection is a promising approach for real-time stream classification.

• Andrej Šareiko, Dalius Mažeika “LLM-Driven Generative Pipelines for Robotic Design and Agentic Systems”

  Presentation introduces an agentic architecture for generative robotics, arguing that physically viable robot design from natural-language prompts exceeds the capabilities of a single LLM. It decomposes the pipeline into three specialized agents—a Designer for CAD generation, a Physicist for physical property and inertia inference, and a Validator for simulation-based assessment—coordinated through Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication. The work highlights two central challenges, the Hollow URDF problem and the Plateau of Iteration, and proposes four research directions: an RL-Sim feedback loop, a physical property inference agent, a human-in-the-loop co-design framework, and Robot-Gen-Bench. Together, these ideas outline a path toward physically grounded, auditable generative robotics systems.

• Dovilė Dikovičiūtė “Balancing Immediate and Future User Engagement in Personalized News Recommendation”

  Personalized news recommendation systems are widely used to help users find relevant content among large amounts of online news. However, many traditional recommendation methods mainly focus on immediate user actions, such as clicks or reading time, and do not consider whether the user will stay engaged later. Because of this, recommendations based only on short-term behavior may not always support long-term user engagement.

  This research explores the use of reinforcement learning to balance immediate and future user engagement in personalized news recommendations. The proposed approach is based on offline reinforcement learning using Conservative Q-Learning (CQL). The model is designed to consider two types of engagement: short-term engagement, reflected by immediate reading behavior, and future engagement, reflected by continued user activity. This allows the recommendation process to better reflect both what attracts the user now and what may keep the user active over time.

  The experimental study is carried out using the EB-NeRD dataset and compares the proposed method with several baseline recommendation approaches. The evaluation focuses not only on recommendation accuracy, but also on how well the method supports different aspects of user engagement. The results suggest that focusing only on immediate user interaction is not enough when the goal is to maintain user engagement over time. The proposed reinforcement learning approach shows potential for generating more balanced recommendations by taking both short-term and future user behavior into account.

• Ansu Mary Jacob, Jolanta Miliauskaitė “AI-Enabled Requirements Specification for Banking Chatbot Systems”

  Traditional Requirements Engineering Specification (RES) is often manual, time-consuming and prone to ambiguity especially in dynamic AI-driven systems. This research aims to improve the RES processes by applying AI-based development by proposing AI-enabled RES approach using Retrieval Augmented Generation (RAG) to automatically transform natural language user queries into structured Functional Requirements (FR). By combining semantic retrieval, Large Language Models (LLM), Knowledge Base (KB). By solving this problem is important because it can reduce manual effort, improve requirement quality and make software development more efficient and reliable in modern AI-based systems.

• ChuXian Chen, Diana Kalibatienė. “Application of Machine Learning Methods for Fairness-Aware Decision-Making in the Loan Approval Business Process”

  The study addresses the issue that traditional machine learning models in the field of loan approval decision-making are prone to inheriting and amplifying social biases present in historical data. This paper proposes and validates a loan approval decision system that incorporate fairness-aware strategies (threshold shifting techniques) into ML models. The effectiveness of this approach was ultimately demonstrated through model evaluation and statistical validation of the significance of the improvements in fairness.

• Marius Bitinas, Rokas Štrimaitis “Vilnius City Traffic Volumes Forecasting”

  Accurate prediction of traffic conditions in cities remains complex due to the dynamic nature of traffic flow and the influence of external factors such as weather and road conditions. Many existing approaches rely primarily on historical traffic data and do not fully incorporate these variables, which can limit prediction accuracy. This highlights the need for more comprehensive, data-driven forecasting methods. The aim of this research is to develop and evaluate a multivariate model for traffic speed prediction and congestion estimation in Vilnius using real-world data. Data from multiple sources were collected and integrated, including traffic flow, weather and road condition data obtained through external APIs. These heterogeneous sources were pre-processed and combined into a unified dataset representing real-world traffic dynamics. The study consisted of three stages. First, existing forecasting methods and influencing factors were analysed. Second, a prediction framework was developed, including preprocessing, feature engineering and model design. Third, experiments were conducted to evaluate different approaches. Both statistical and deep learning models were implemented, including SARIMA, LSTM, BiLSTM, CNN-LSTM and Transformer architectures, using recursive and direct forecasting strategies. Results showed that deep learning models outperformed traditional statistical methods in capturing temporal patterns. The recursive BiLSTM model achieved the most stable and accurate performance. Feature analysis indicated that weather and road data affected accuracy differently across locations, sometimes improving results and sometimes introducing noise. In conclusion, integrating multiple data sources and applying advanced deep learning models enabled effective traffic prediction. The findings highlighted the importance of careful feature.

• Supun Chandimal Thilakarathna Ranasinghe Mudiyanselage, Jolanta Miliauskaitė “Application of Artificial Intelligence Methods in The Requirements Engineering Management Process”

  Manual requirements engineering (RE) processes are increasingly inadequate for large-scale software projects, suffering from inconsistency, ambiguity, and poor traceability. This paper presents an AI-driven three-stage pipeline for automated requirements engineering management, combining large language models (LLMs), transformer-based classification, and graph database technology within a unified human-in-the-loop architecture. The proposed system addresses three core RE challenges: requirement completeness detection and structuring (Stage 1), functional versus non-functional requirement classification with subtype identification (Stage 2), and automated dependency and conflict detection through knowledge graph construction (Stage 3). A multi-model ensemble combining a domain-fine-tuned NoRBERT classifier, a Qwen2.5-7B-Instruct LLM, and a rule-based classifier with consensus voting is introduced to mitigate LLM hallucination while preserving classification explainability. Experimental evaluation on a structured banking domain dataset demonstrated 100% completeness detection accuracy and 83.3% end-to-end automation, with automatic identification of a critical semantic conflict undetectable by keyword-based methods. Evaluation on the PROMISE NFR benchmark (n=125) yielded 96.8% binary classification accuracy for the fine-tuned model (F1=0.976) and 100% NFR precision for the ensemble with zero false positives across all test cases. Processing time of 310.71 seconds represents a 20–40× improvement over equivalent manual analysis. These results demonstrate that multi-model consensus voting effectively suppresses hallucinated classifications while maintaining high recall, and that graph-based traceability enables automated conflict detection beyond the capability of document-based approaches. The system contributes a reproducible, domain-agnostic framework for scalable, explainable requirements engineering automation.

• Roberta Gailiūtė “Comparative Analysis of CNN Architectures for Automated Textile Sorting and Recycling”

  This study investigates the application of deep learning models for automated textile waste classification in recycling processes. The increasing amount of textile waste and the limitations of traditional sorting methods highlight the need for intelligent classification systems capable of identifying textile materials efficiently. Convolutional neural networks (CNN) were selected as the primary approach due to their strong performance in image-based classification tasks.

  Several CNN architectures, including AlexNet, ResNet18, MobileNetV2, and EfficientNet, were evaluated using textile image datasets. The datasets were preprocessed using resizing, normalization, and data augmentation techniques to improve model generalization. Model performance was assessed using accuracy, F1 score, loss, and computational efficiency.

  Experimental results showed that transfer learning significantly improved classification performance. ResNet18 achieved stable accuracy, while MobileNetV2 provided the best balance between accuracy and computational cost. The experiments also demonstrated that dataset complexity influences model performance, with TextileNet presenting more challenging classification conditions than simpler clothing datasets.

  The findings indicate that deep learning models are suitable for automated textile classification and can support intelligent recycling systems. Lightweight architectures are particularly promising for real-time industrial applications where computational efficiency is critical.

• Mahesh Kumar Pratihar “Textual Data Clusterization Based on Self-Organizing Map and Word Embeddings”

  This presentation focuses on improving text clustering using a combination of modern word embeddings and Self-Organizing Maps (SOMs). Clustering textual data is difficult because text is often high-dimensional, sparse, and semantically complex. Traditional methods like TF-IDF and K-means often fail to capture the true meaning of text, leading to poor clustering results. To solve this problem, the proposed approach uses transformer-based models to generate meaningful sentence embeddings that capture context and semantics. These embeddings are then clustered using SOMs, which organize data into a low-dimensional map while preserving relationships between similar texts. This makes the clusters easier to interpret and visualize. The method is tested on several real-world datasets, including tweets, news articles, and reviews. Its performance is evaluated using standard clustering metrics such as Silhouette Score, Adjusted Rand Index, and Purity. The results show that this combined approach produces more accurate and meaningful clusters compared to traditional techniques.

  In summary, this work demonstrates that combining contextual embeddings with SOMs improves both the quality and interpretability of text clustering. The approach can be useful in applications like document organization, sentiment analysis, and information retrieval.

• Mirmatlab Karimli, Diana Kalibatienė “Research on Application of Machine Learning Methods for Drug Sales Prediction”

  This thesis focuses on improving drug sales prediction in pharmacy stores by applying machine learning methods. Accurate forecasting is very important in the pharmaceutical industry because it helps pharmacies maintain the right stock levels, reduce costs, and avoid situations where medicines are either overstocked or unavailable. Traditional forecasting approaches, such as ARIMA, are often not able to fully capture the complex and changing patterns in pharmaceutical sales, especially when seasonality and external factors are involved.

  In this research, three main models are analysed: XGBoost, ARIMA, and Long Short-Term Memory (LSTM). The study includes a comparative analysis of these methods using real pharmacy sales datasets. A complete forecasting framework is also proposed, including data preprocessing, feature engineering, model training, and evaluation. Special attention is given to seasonal behaviour and time-based dependencies, as these strongly influence drug demand.

  The results show that no single model performs best in all cases. ARIMA works well with more stable and structured data, while LSTM provides better results when the data is more irregular and complex. XGBoost is effective in capturing non-linear relationships but depends heavily on the quality of the engineered features. Therefore, the study highlights the importance of choosing models based on the specific characteristics of the dataset.

  Overall, the proposed approach improves forecasting accuracy and supports better inventory management in pharmacy stores, helping to reduce waste and ensure the availability of medicines.

• Islam Salamzade “Mining Method Application for Software System Event Log Analysis”

  Modern software systems generate large volumes of heterogeneous and unstructured event logs, which contain valuable information about system behavior, performance, and failures. However, these logs are not directly suitable for process mining because they often lack reliable correlation identifiers (case IDs), making it difficult to reconstruct meaningful execution traces. This limitation reduces the applicability of process mining techniques in real-world system monitoring scenarios.

  This study addresses the problem by proposing a Smart Trace Construction (STC) method for reconstructing process traces from raw software logs in correlation-weak environments. The method transforms unstructured log data into structured event logs by parsing log messages, extracting relevant attributes, and grouping events into traces based on temporal proximity and attribute similarity. A prototype implementation was developed to generate event logs in CSV/XES formats compatible with standard process mining tools.

  The proposed method was evaluated on real-world datasets, including the HDFS and BGL system logs. The evaluation focused on trace reconstruction quality using metrics such as purity and mixed-trace ratio, as well as downstream process mining performance using token-based replay metrics (fitness and precision). Experimental results demonstrate that the STC method achieves high reconstruction quality and produces more accurate process models compared to baseline approaches. Additionally, sensitivity analysis shows that the inactivity threshold parameter significantly affects reconstruction quality, highlighting a trade-off between trace fragmentation and merging.

  Overall, the results confirm that the proposed method is effective and robust across different types of system logs, even in the absence of reliable case identifiers, thereby improving the applicability of process mining in practical software system analysis.

• Tahmina Bayramova “Research on Decision Support System for Designing Cloud Infrastructure”

  The topic of research is the application of rule-based decision modelling and Multi- Criteria Decision Analysis (MCDA) techniques to develop a Decision Support System (DSS) for designing cloud infrastructure. Cloud computing services are becoming increasingly diverse and undergo constant modifications regarding their price, performance, scalability, and Service Level Agreements (SLAs). This creates additional challenges for software architects when making a choice among several alternative options. To solve the above-mentioned problem, the research proposes applying Decision Model and Notation (DMN) decision modeling for qualitative rule-based filtering and employing various MCDA techniques, including AHP and TOPSIS methods, to quantitatively evaluate and rank feasible alternatives. First of all, the DMN decision tables will be utilized to remove architecturally infeasible cloud configurations based on the analysis of functional and non-functional requirements. Afterward, the remaining options will be evaluated according to certain criteria (such as price, performance, availability, scalability, and security) and sorted accordingly. An experimental validation of the proposed approach will be performed in a local development environment using Visual Studio Code. It is expected that the integration of rule-based decision modelling with MCDA methods will contribute to increasing the transparency, consistency, and accuracy of cloud infrastructure design decisions.

• Fidan Babayeva “Integration of Ontologies: FrameNet, WordNet, SUMO, DBpedia”

  Ontology is a formal and explicit representation of concepts, entities, and the relationships between them within a specific domain. An ontology represents knowledge in a formal way for a particular domain, giving the types of entities-things that exist-and the relationships between them. It provides a common vocabulary to represent concepts and the relationships among them, which enables data sharing, integration, and inference processes in machines.

  The integration of multiple ontologies - FrameNet , SUMO , WordNet , and DBpedia, faces significant challenges. Indeed, these independently developed ontologies are used for distinct purposes: lexical relations with WordNet, frame semantics with FrameNet, an ontology formally developed with SUMO at a higher level, and extraction of structured information from Wikipedia to support DBpedia. The latter element increases the chances of semantic conflicts and inconsistencies by presenting a diversity in terminology, structure, and conceptual representation. Different ontologies use different structures, scopes, and levels of abstraction. Integrating these ontologies allows us to combine their individual strengths and create a more complete, context-aware understanding of language.

  To address this challenge, the report proposed and implemented an initial prototype that operationalizes the suggested approach with emphasis on Word Sense Disambiguation as a key enabling process for integration. A reproducible code in Python (Google Colab) has been developed using spaCy for linguistic preprocessing and a BERT model (via KerasNLP) for contextual embeddings, followed by an embedding-based WSD component that predicts WordNet synsets for ambiguous words.

  The experimentation carried out on a sample dataset of sentences demonstrated that the prototype has the capability of producing well-structured and reliable outputs suitable for further semantic annotation.

• Lukas Makaris, Asta Slotkienė “Research into the Quality of Unit Tests Generated by Large Language Models”

  This thesis examines the use of large language models for automated unit test generation in software engineering. As software systems become more complex, the need for efficient and reliable testing continues to grow. Large language models offer new possibilities for reducing manual effort and accelerating test creation, but the quality of generated tests remains inconsistent. This raises important questions about their reliability, maintainability, and practical value in real development environments. The research focuses on the factors that influence the quality of automatically generated unit tests and explores strategies that may improve generation outcomes. Special attention is given to prompt design, contextual information, and model behavior, as these elements strongly affect the structure and usefulness of generated test code. The study treats unit test generation not only as a code generation task, but also as a software quality challenge in which correctness, readability, robustness, and ease of integration are essential. To address this problem, the thesis combines theoretical analysis with empirical investigation. Relevant scientific literature is reviewed to identify current challenges, common limitations, and promising directions in automated test generation with large language models. On this basis, a framework is developed for comparing different generation approaches and assessing their impact on test quality according to practical evaluation criteria. The results contribute to a broader understanding of how large language models can support software testing processes. The thesis also highlights both the opportunities and the limitations of applying generative artificial intelligence in quality assurance. Overall, the work aims to support the development of more structured and effective approaches for integrating large language models into real-world test generation practices.

• Mahir Mustafayev “Research on Program Code Analysis Techniques”

  The current software systems written in TypeScript are growing larger and more complicated, and are composed of a great number of modules linked to each other. Dependencies among code components emerge as a key factor affecting the maintainability of software, architectural quality and technical debt as projects increase in size. Nevertheless, available TypeScript analysis tools primarily concentrate on basic analysis or simple dependency analysis of imports and do not usually have the capability to do more architecture-oriented analysis.

  This study examines the techniques of program code analysis, specifically dependency analysis of TypeScript projects. The study aim is to design and test a method which allows systematic extraction and presentation of dependencies in a form that can be used to analyse software systems architecturally.

  The thesis aims to overcome these shortcomings by proposing a dependency analysis algorithm combining the techniques of a static code analysis with the graph-based software structure representations. The presented methodology derives dependency relations on TypeScript projects with CodeQL queries and converts the resulting data to structured datasets. With this knowledge, a Design Structure Matrix (DSM) is created to indicate dependencies among software components in a small and analyzable architectural format.

  An experimental version of the suggested method was created to test the feasibility of the approach. The prototype pipeline is an automated dependency extraction, dataset generation, and DSM construction based on a mixture of CodeQL analysis and Python-based processing. Experimental data show that the technique is effective in modeling dependency relationships and can also aid developers to study the patterns of coupling and structural relationship in the system architecture.

• Mohini Naga Venkata Poojaswi Kankipati, Algirdas Laukaitis “AI-Powered Quiz Generation and Automated Evaluation System for MCQs and Written Responses System”

  This research presents the development of an AI-powered online quiz system that supports both automatic quiz generation and evaluation of multiple-choice and subjective written responses. Traditional quiz creation and assessment methods are time-consuming, require significant manual effort, and lack scalability, particularly in modern digital learning environments.

  The proposed system integrates AI-driven quiz generation with automated assessment capabilities. It leverages Natural Language Processing (NLP) models, including GPT and BERT, to generate questions based on user-defined topics and to evaluate written responses. Multiple-choice questions are assessed using rule-based logic to ensure accuracy and efficiency. The system is implemented using Google Colab, selected for its cloud-based environment, accessibility to computational resources, and seamless integration with AI libraries and APIs, enabling rapid development and deployment without local infrastructure dependencies.

  The system processes user inputs to generate quizzes, evaluate responses, and provide feedback. The expected outcome is a scalable and efficient solution that reduces manual workload while maintaining consistency and accuracy in both quiz creation and assessment.

  Future work includes improving model accuracy, enabling multilingual support, enhancing feedback quality, and extending the system to support evaluation of video-based responses in interview scenarios.

• Leonardas Vekrikas, Jelena Mamčenko ”Fraud Detection in Financial Data with The Help of AI Methods: Analysis and Investigation” - winner in category "Best Talk"

  This study evaluates the hybrid LightGBM-LSTM framework proposed by Yousefimehr and Ghatee (2025) for fraud detection in financial transactions. The model combines LightGBM for structured data analysis with LSTM for sequential dependency modelling, incorporating distribution-preserving balancing (HybridOS/HybridUS) and voting-based feature selection (F2Vote).

  The research consisted of three phases. First, the model was fully implemented and replicated on the European Credit Card Fraud Dataset (284,807 transactions), confirming the reproducibility of the original results. Second, testing on the PaySim synthetic dataset (1,048,575 transactions) revealed significant performance degradation, exposing the model's sensitivity to dataset characteristics.

  Third, a comparative experiment evaluated feature-level fusion (concatenating LightGBM leaf embeddings and LSTM hidden states via an MLP meta-classifier) against the original decision-level fusion (sum rule with threshold). Feature-level fusion achieved comparable AUC (0.967) but substantially lower F1-score (0.040 vs ~0.80), primarily due to class imbalance effects and overlapping discriminative signals. The findings validate decision-level fusion as the more practical approach, offering equivalent discriminative power with greater interpretability.

• Pavithra Madushanka Midimulla Kankanamlage, Jūratė Janutėnaitė – Bogdanienė “AI Based Emotion Recognition for Human Decision Making Under Stressful Conditions: A Systematic Review and Research Framework”

  Human decision-making in critical environments is highly vulnerable to emotional and cognitive stress, particularly in cybersecurity contexts where real-time responses are essential. Despite well-established technical defense mechanisms, emotional factors influencing human operators remain insufficiently addressed. This study presents a systematic review and conceptual framework for integrating artificial intelligence (AI)–based emotion recognition into real-time cybersecurity decision-making.

  A structured literature review was conducted using the Web of Science database (2016–2026), initially identifying 2,470 records, of which 100 studies were ultimately included following rigorous screening and eligibility assessment. The analysis reveals that existing research predominantly addresses three domains independently: emotion detection (e.g., facial, speech, and physiological signals), real-time cyberattack response, and human decision-making under stress. However, the intersection of these domains remains significantly underexplored.

  Findings indicate that unimodal emotion recognition models perform well in controlled environments but lack robustness in dynamic, real-world conditions. In contrast, multimodal approaches improved performance by capturing complex interdependencies across modalities.

  To address these gaps, this research concentrates on proposing a novel multimodal AI-based research framework that integrates real-time emotion recognition with cyberattack detection and decision-support mechanisms. The framework emphasizes context-aware modeling and adaptive multimodal fusion to enhance trust in high-stakes environments. This work contributes a unified perspective by systematically bridging emotion recognition, cybersecurity, and human decision-making, enabling the development of next-generation, human-centered cyber defense systems capable of supporting effective decision-making under stress.

• Paulius Bundza, Justas Trinkūnas “MCADS: Out-of-Distribution Generalization of Multi-Label Chest X-Ray Classification Models Across Seven Architectures” - winner in category "Best Scientist"

  Deep learning models for multi-label chest X-ray classification have shown strong in-distribution performance, yet their reliability degrades under domain shift — a critical barrier to safe clinical deployment. This thesis systematically evaluates the out-of-distribution (OOD) generalization, calibration, and fairness of seven pre-trained architectures (six DenseNet-121 variants and one ResNet-50 (Cohen et al., 2022)) by training on NIH ChestX-ray14 and testing on the BRAX dataset. The study proposes a comprehensive evaluation framework and an integrated web application (MCADS), providing actionable insights into how these models behave across demographic subgroups and unseen clinical environments.

• Rakshitha Shyamanur Revanaradhya, Urtė Radvilaitė “Research on Improving the Quality of Online Customer Support Using Artificial Intelligence”

  The rapid growth of digital services has significantly increased the demand for efficient and high-quality online customer support. However, traditional customer support systems often face critical challenges, including delayed response times, lack of personalization, limited scalability, and inconsistent service quality. These limitations negatively impact customer satisfaction, loyalty, and overall user experience. This research investigates how Artificial Intelligence (AI) can be leveraged to address these challenges and improve the quality of online customer support systems.

  The study proposes a hybrid AI-driven framework that integrates Natural Language Processing (NLP), Machine Learning (ML), and large language models to enhance customer interaction processes. The system is designed to understand customer queries through intent detection and sentiment analysis, enabling more accurate and context-aware responses. Additionally, a contextual knowledge retrieval mechanism based on semantic similarity is implemented to provide relevant and precise information. The framework also incorporates sentiment-aware response generation and escalation strategies to ensure that complex or sensitive cases are handled appropriately, including human intervention when necessary.

  An end-to-end prototype of the proposed system was developed and evaluated within an e-commerce domain. The experimental results demonstrate significant improvements in response accuracy, response time, and customer satisfaction compared to traditional rule-based systems. The findings highlight the effectiveness of combining rule-based logic with advanced AI techniques to achieve both reliability and adaptability.

• Robertas Pavlovskis “Method and Study of Automating Requirements Verification Using the RAG Model”

  Prezentacijoje bus nagrinėjamas reikalavimų verifikacijos automatizavimo metodas, taikant PPG modelį, bei pateikiami atlikto tyrimo rezultatai. Darbo tikslas buvo išanalizuoti, kaip skirtingos PPG strategijos, kalbiniai modeliai ir jų parametrų kombinacijos veikia automatizuotos reikalavimų verifikacijos tikslumą, stabilumą ir praktinį pritaikomumą. Tyrime buvo vertinami keli metodai: paprastas PPG, spekuliatyvus PPG ir susiliejančio PPG metodas. Šie metodai buvo testuojami naudojant skirtingus didžiuosius kalbos modelius, tokius kaip ChatGPT ir Gemini, bei įvairius jų veikimo parametrus. Tyrimo metu buvo siekiama ne tik palyginti atskirų metodų našumą, bet ir nustatyti, kokia modelio, paieškos strategijos ir parametrų kombinacija leidžia pasiekti geriausius rezultatus reikalavimų verifikacijoje.

• Ervin Barkovski, Nikolaj Goranin “Prototype of an Intrusion Detection System for Internet of Things Networks”

  Recent research has shown that IoT networks are characterized by specific security vulnerabilities related to limited device resources and rapidly evolving and increasingly sophisticated threats. The authors of the previous studies emphasize that traditional signature-based IDSs are unable to detect new attacks, even though they require fewer computational and memory resources. Anomaly detection and hybrid methods based on machine learning are considered more promising solutions, but their resource requirements depend on the chosen model training method. When comparing ML and DL methods, it is noted that ML is more suitable for intrusion detection in IoT networks, as it achieves accuracy comparable to that of DL methods while consuming fewer technical resources.

  An analysis of existing solutions revealed a clear distinction between commercial and open-source tools. Commercial offerings such as “Armis,” “Nozomi Networks,” “Claroty,” or “Darktrace” are tailored for IoT networks, accurate, and feature advanced analytics capabilities, but require significant financial investment. Unlike commercial solutions, open-source IDS tools such as Suricata, Wazuh, Zeek, and Snort are not tailored for IoT networks. However, this is not necessarily a drawback—since they are open-source, they can be configured for specific IoT needs, optimize resource usage, and achieve high intrusion detection accuracy by integrating an ML-based anomaly detection model.

  Therefore, this research focuses on analysing currently available algorithms for ML-based learning, presents the findings, and suggests which algorithm is most suited for implementation into the IDS tools.

• Danielius Krajeckas “Vehicle Route Optimization Adding Fuel Price Forecast“

  Magistro darbe nagrinėjamas maršruto planavimas atsižvelgiant į kuro kainų prognozę. Kuro kaina bei išlaidų kelionėje optimizavimas yra svarbu ne tik kasdieniam vartotojui, bet ir įmonėms, kurių veikla yra glaudžiai susijusi su transportu. Pagrindinis tyrimo tikslas – išanalizuoti ir visapusiškai įvertinti kiek galima sutaupyti planuojant maršrutą su kuro kainų prognozavimu. Tyrimas apima kuro kainų prognozavimą bei maršruto planavimą įtraukiant degalines, kurių kuro kaina pagal pasirinktą degalų rūšį yra mažiausia. Didesnis dėmesys skiriamas kuro kainų prognozei atsižvelgiant į prieš tai buvusias kainas, bei pasaulinę situaciją.