Selecting a Multi-Criteria Decision-Making (MCDM) method is critical for developing robust Decision Support Systems (DSS), yet limited attention has been given to assessing their stability under structural changes in decision problems. This study proposes a simulation-based framework for evaluating the robustness of MCDM methods when the least important criteria are iteratively removed. Four selected methods, namely Additive Ratio ASsessment (ARAS), COmplex PRoportional ASsessment (COPRAS), Measurement Alternatives and Ranking according to COmpromise Solution (MARCOS), and MultiAttributive Ideal-Real Comparative Analysis (MAIRCA) were tested across thousands of randomized scenarios, with performance assessed through mean ranking correlation, frequency of ranking alterations, and distribution of similarity values. The findings reveal consistent stability trends across methods while identifying differences in sensitivity to criteria reduction. Notably, MAIRCA and COPRAS exhibited more concise performance distributions, suggesting stronger resilience to problem changes. This work addresses a critical gap in understanding method robustness, supporting more informed selection of MCDM techniques for uncertain decision environments and enhancing the reliability of decision-making processes.
The widespread adoption of smart mobile devices (SMDs) with advanced computing capabilities presents a valuable resource for mobile crowd computing (MCC). Efficient task scheduling in MCC relies on selecting the right SMDs, which poses a complex multi-criteria decision-making challenge due to the diverse hardware specifications of the devices and the presence of non-compensatory parameters. Traditional multi-criteria decision analysis (MCDA) methods, such as the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), typically assume full compensability between criteria. However, this assumption may conflict with strong sustainability principles. To tackle this issue, the authors introduce the Strong Sustainability Paradigm based Technique for Order Preference by Similarity to Ideal Solution (SSP-TOPSIS) method, an extended version of TOPSIS that incorporates linear compensation reduction. This enhancement allows for a more accurate reflection of sustainability requirements in the decision-making process. The SSP-TOPSIS method demonstrates improved analytical capabilities compared to classical TOPSIS and provides a framework that supports sustainability-driven decisions.
Chatbots powered by large language models (LLMs) are increasingly prevalent in various domains. Nonetheless, they face challenges such as hallucinations and losing context during extended conversations. This study tackles these issues by proposing a multi-agent strategy for chat architecture where multiple LLMs focus on distinct tasks to enhance the quality of their output. The suggested solution involves a supervisor agent working in conjunction with a document search and review module. We assess the performance of information systems with chatbots designed to respond to sustainability questions in English and handle technical documentation for plant equipment in Polish. A comprehensive analysis of commercial and open-source models revealed that Qwen2.5 v14b’s performance is comparable to that of the Gemini family models.
This study examines the reliability of automatic evaluation metrics in assessing responses generated by large language models (LLMs) in the context of university recruitment. A total of 113 domain-specific questions were used to prompt five prominent LLMs, each in three configurations: basic, document-context, and internet-context. The generated responses were evaluated using three categories of metrics: lexical, semantic, and LLM-asa-Judge. These metric-based assessments were subsequently compared with expert evaluations conducted using a 5-point Likert scale. The findings indicate that although automatic metrics offer considerable efficiency, their consistency with expert judgments varies substantially. Moreover, the results suggest that both the model configuration and its underlying architecture significantly affect evaluation outcomes. Among the metric categories, LLM-as-a-Judge appears to yield the highest alignment with expert assessments, suggesting greater reliability in this approach.
This article presents an innovative approach to monitoring river water quality in real time by generating estimates of difficult-to-measure signals such as biochemical oxygen demand. Laboratory tests take too long for real-time monitoring. Therefore, an adaptive PDALM algorithm (Proportional Differential Algorithm with a Latch Mechanism) was developed, integrating mathematical modelling with measurement data to enable instantaneous estimation of water quality signals using a special latch mechanism. The forced eigenvalue distribution guarantees system dynamics and ensures stability and robustness to disturbances. In the proposed RTMS system, the PDALM algorithm functions as an adaptive soft sensor generating high-quality training data. This data is then used by a generative neural network for anomaly detection and forecasting of atypical scenarios in dynamic environmental systems. The system can function as an intelligent environmental monitoring module capable of learning, predicting, and responding to changing environmental conditions.
LoRaWAN networks, which are extremely popular today, are based on the LoRa protocol and offer very long communication ranges, but they also come with significant limitations. These limitations stem primarily from two factors: duty cycle and maximum message size. During image transmission over LoRa-based networks, packet loss is a common problem resulting from limited bandwidth and transmission interference. During image transmission, it leads to missing data in the received content, most often visible as vertical or horizontal lines. We present a method to repair such corrupted images using a fully convolutional neural network inspired by the U-Net architecture. The experiments carried out show that the proposed approach effectively reconstructs missing parts of the image, achieving high structural similarity (SSIM) to the original. The proposed method can be applied to image transmission and reconstruction on low-power devices, typical of IoT systems that use LoRa for communication.
Changing economic conditions and technological challenges are shaping the future of the insurance industry. Traditionally associated with a conservative approach to risk, the industry is faced with the need to make bold decisions that will allow it not only to manage new risks, but also to take advantage of the opportunities presented by the digitalisation of processes. Thus, the insurance market needs to adapt quickly to technological change and changing consumer expectations. The aim of this article is to analyse the areas and tools related to the digital transformation of the insurance sector. The research question posed is: how does digital transformation affect the development of the insurance market globally? The research presented in the article can contribute to the theoretical knowledge of digital transformation in the insurance sector. The research is based on data available in international reports published by Swiss Re, GlobalData, among others.
Data lineage is the set of techniques for tracking the flow of data throughout its lifecycle. These techniques are crucial for data management, governance, and compliance with regulations. Lineage links are maintained between data and database objects, but they are often broken by temporary objects and user defined functions. To the best of our knowledge, discovering broken lineage links has not been addressed yet in research. In this paper, we present a method for detecting broken lineage links between database objects. To this end we apply machine learning techniques on available metadata. We extract feature vectors and employ a classification approach to determine whether one database object is a source for another. Initial experiments on large database schemas show that the discovery of broken lineage links is possible at an acceptably high probability.
Companies all over the world benefit from Industry 4.0, to fully harness its potential they need a skilled workforce. Hence, the goal of the paper is to recognize whether students, who represent the future workforce, are ready for the new technological conditions and challenges that these bring. Education at the academic level should follow and, in some cases, precede business conditions to ensure that students have the skills required in the labor market after graduating, both in hard skills and soft skills areas.
Today, deep learning methods are being strongly developed and are used for many different tasks. This paper addresses the task of vehicle detection in parking lots. The focus of this study is to evaluate the performance of several versions of the YOLO (You Only Look Once) object detection algorithm on a self-created dataset, using the models in their default configurations with pre-trained weights from the COCO dataset. The dataset contains various lightning and weather conditions such as sunshine, cloudiness, and the presence of snow. Each YOLO version is evaluated using a range of metrics such as precision, recall, F1 score, and FPS. Methods for optimizing the use of models are then proposed and tested. The results demonstrate the trade-offs between detection accuracy and computational efficiency.
This paper presents a descriptor-based method for labeling point clouds using a two-stage transformer architecture. The first stage consists of an encoder that extracts descriptors from point cloud fragments. The second stage, a decoder, assigns labels to these fragments based on both the descriptor of the current fragment and an earlier predefined pattern descriptor. This approach functions as an interactive labeling tool similar to a brush, with the ability to reinforce or weaken the pattern through direct manipulation of its descriptor.
This research investigates the perceived naturalness of synthesized speech in the context of Polish medical terminology, a critical factor for applications such as voice-enabled medical dialogue systems. We conducted a comparative analysis of three speech synthesis models: SpeechGen, ElevenLabs, and a version of ToucanTTS fine-tuned on a specialized corpus of Polish medical recordings. The evaluation employed objective measures, the NISQA metric, and subjective assessments through Mean Opinion Score (MOS) surveys. Our findings indicate that SpeechGen and ElevenLabs produce synthesized speech that closely rivals the naturalness of human speech, as evidenced by both NISQA scores and MOS ratings. In contrast, despite improvements, the fine-tuned ToucanTTS model did not achieve comparable levels of perceived naturalness. Notably, participants occasionally rated the advanced synthesized speech as more natural than human speech recorded in non-studio environments, underscoring the potential of these technologies in real-world applications. This study emphasizes the significance of naturalness in enhancing user experience, particularly in specialized linguistic domains. It provides insights into speech synthesis's current capabilities and limitations for less-resourced languages like Polish.
This paper presents a privacy-preserving framework for distributed neural network modeling across heterogeneous data sources, where local datasets differ in both objects and attributes. To enable collaborative learning without sharing raw data or model parameters, each local decision table is independently transformed into a unified feature space using multiple dimensionality reduction techniques – Principal Component Analysis (PCA), Singular Value Decomposition (SVD), and Uniform Manifold Approximation and Projection (UMAP). Various types of neural networks – Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Simple Recurrent Network (SIMPLE), Multilayer Perceptron (MLP) and the Radial Basis Function Network (RBF) – are trained locally, and their outputs are aggregated using soft voting (simple average) to generate final predictions. Experimental results on benchmark datasets confirm the approach’s effectiveness, scalability, and robustness in decentralized learning settings.
Injuries pose a significant challenge in professional football, affecting player availability, team performance, and club finances. Accurate prediction of injury risk is crucial for implementing effective prevention strategies. This study develops a deep learning model to predict the likelihood of injury in professional football players using data collected through Catapult Sports tracking devices. The research is carried out in collaboration with KKS Lech Poznań, a Polish professional football club. The proposed model architecture combines bidirectional long- and short-term memory (BiLSTM) networks with an attention mechanism to learn from the time series data and predict player injury risk. The model is trained on sequences of Catapult data spanning 14 days before each recorded injury or non-injury event. To address class imbalance, a custom loss function was implemented that balances focal loss and the F_beta score. The model's performance is evaluated on an independent test set, achieving a specificity of 0.90, an accuracy of 0.90, and a recall of 0.40.
Feature selection plays a significant role in the development of categories of information systems related to decision support, such as diagnostic or recommendation systems. Such systems should ensure the possibility of identifying the most important features as well as analysing data from different locations, taking into account the specificity and characteristics of the local data sources. In the process of data analysis, the stage of data preparation, including the transformation of the attribute domain from continuous form to intervals, plays an important role, as the outcome of this process influences the subsequent stages of the analysis. In the paper, an approach to creating a global feature ranking that takes into account the specifics and characteristics of different discretisation algorithms was proposed. A new weight for the estimation of attribute importance was defined and compared with a measure that is implemented in the Python programming language library. Both types of weights were used to create a hierarchical structure of the global ranking of features. The experiments were carried out on datasets from the stylometry domain dedicated to the task of authorship attribution.
This work presents an intelligent support system for a novel, non-destructive (NDT), 2D method to identify parameters of reinforced concrete (RC) structures. Using association rule analysis (ARA), it detects relationships between signal changes and structure parameter modifications, identifying signal parameters influenced by a single structural parameter. Multitask learning is used to identify concrete cover thickness, reinforcing bar diameter, and steel class. Features are extracted from the three spatial components of magnetic induction via ACO decomposition, which is suited for creating complex databases. Genetic algorithms improve noise resilience in function approximation. Results are shown as Fuzzy Rough Sets. Three vertically placed sensors, combined with AI, enable precise identification of parameters, with changes in one not affecting others.
Recommendation systems are essential for delivering personalized content across e-commerce and streaming services. However, traditional methods often fail in cold-start scenarios where new items lack prior interactions. Recent advances in large language models (LLMs) offer a promising alternative. In this paper, we adopt the retrieve-and-recommend framework and propose to fine-tune the LLM jointly on warm- and cold-start next-item recommendation tasks, thus, mitigating the need for separate models for both item types. We computationally compare zero-shot prompting, in-context learning, and fine-tuning using the same LLM backbone, and benchmark them against strong PLM-based baselines. Our findings provide practical insights into the trade-offs between accuracy and computational cost of these methods for next-item recommendation. To enhance reproducibility, we release the source code under https://github.com/HayaHalimeh/LLMs-For-Next-Item-Recommendation.git.
Network intrusion detection systems face high-dimensional traffic, which degrades accuracy and raises computational costs. We evaluate Principal Component Analysis (PCA), Recursive Feature Elimination (RFE), and a deep autoencoder on the UNSW-NB15 dataset using eight classifiers. RFE delivers peak accuracy (92.5%) with minimal variability. PCA restores near-baseline accuracy while preserving 95% variance with minor tuning. The autoencoder yields nonlinear embeddings but demands extensive training and trails classical methods. These findings guide the selection of reduction strategies under accuracy requirements and resource constraints.
This paper presents a Design Science Research (DSR) approach to addressing visualization bottlenecks in environmental Information Systems Development (ISD). By developing a CUDA-based atmospheric effects framework utilizing the Material Point Method (MPM) and Marching Cubes algorithms, we demonstrate how GPU acceleration transforms ISD methodologies for data-intensive decision support systems (DSS). This research contributes to digital transformation of environmental monitoring platforms by enabling real-time processing of complex simulation data that traditionally require significant computational resources. The prototype demonstrates scalable performance handling up to 6.5 million particles while enabling configuration-driven customization that allows information systems developers to integrate sophisticated environmental visualization without specialized graphics expertise. This approach democratizes atmospheric data visualization for environmental monitoring systems. Empirical results demonstrate real-time visualization capabilities suitable for operational deployment.
This study empirically evaluates the performance of Chronos, a recent foundation model pre-trained on a large corpus of time series data, for the task of daily stock index forecasting. Using a rolling window framework on historical Nasdaq-100 and S&P 500 data from 1995 to early 2025, we compare zero-shot and fine-tuned Chronos variants against a diverse set of established forecasting methods, including statistical benchmarks (AutoARIMA, ETS), standard deep learning models (DeepAR, DLinear, SimpleFeedForward), other Transformer-based architectures (PatchTST), and ensemble approaches. Our results, based on standard forecasting metrics and simulated trading performance, indicate that zero-shot Chronos provides competitive forecasting accuracy. It is statistically comparable to the best traditional methods, but its derived trading performance lags top benchmarks. The fine-tuned Chronos variant statistically underperformed the zero-shot version in forecast accuracy. These findings highlight the potential of foundation models and underlines the significant challenges in effective fine-tuning.
