This study introduces a novel concept lattice reduction model that integrates heuristic and machine learning optimization approaches to balance expressiveness with computational efficiency. The resulting set of concepts is considered the kernel of the original context. The proposed kernel induction method applies an optimized genetic algorithm with machine learning support for kernel selection. In the efficiency comparison tests, we also analyzed the simulated annealing method, the particle swarm optimization method, and a derivative-free, ranking-based optimizer method. Experimental evaluations on synthetic and real-world datasets reveal that the proposed genetic algorithm variant outperforms the other benchmark methods in both computational efficiency and effectiveness while maintaining scalability. The applicability of the proposed method is demonstrated through a linguistic-domain case study on selecting an optimal kernel vocabulary.

Fault detection in industrial processes is challenging due to complex interactions between sensor measurements. This paper proposes a graph learning framework where process variables are modeled as nodes and edges encode their relationships. A Graph Convolutional Network learns normal behavior from sliding time windows, and prediction errors are used for anomaly detection. Node-level contributions enable fault localization and support operator decisions. The approach is evaluated on three benchmarks: Tennessee Eastman Process (TEP), Secure Water Treatment (SWaT), and Water Distribution (WADI). It is compared with LSTM, PCA, and autoencoder baselines. On SWaT, all methods achieve ROC-AUC above 0.91. On TEP, the graph model outperforms LSTM with ROC-AUC of 0.788 compared to 0.749. WADI remains difficult for all methods, with ROC-AUC below 0.20. The framework provides interpretable sensor-level insights for reliable monitoring and control.

Multilayer perceptron (MLP) is a powerful tool used in both regression and classification problems. However, training an MLP is inherently stochastic for several reasons, including stochastic optimization algorithms and weight initialization methods. Additionally, changing the architectural parameters of the MLP strongly influences the model's performance. This study examines how random the final accuracy and loss values of an MLP are when the network structure and input data are fixed. We executed extensive repeated-run experiments to examine performance variance rather than the single best results. We also investigated the influence of hyperparameters, including the number of hidden layers and the number of neurons, by assessing the sensitivity of each parameter on model performance and training time. Analyzing randomness in MLP training provides a broader and more reliable way to study and compare hyperparameter optimization methods.

The radial basis function and multilayer perceptron architectures diverge significantly as theoretical and practical approaches to neural networks. While each exhibits unique strengths in classification performance, the potential for integrating these tools remains a compelling, yet understudied, area of research. In this paper, we introduce two different approaches for the integrated architecture. The first approach applies the mixture of MLP dense layers and RBF layers. The proposed model applies an optimized kernel initialization mechanism to provide a fast convergence. In the second architecture model, a novel custom network layer architecture is presented which can balance between the MLP and RBF layer mode. Tested across various classification tasks, our proposed neural network model consistently demonstrated superior performance.

This study presents the design, implementation, and evaluation of an intelligent tutoring system tailored for dynamic learning domains such as computer science. The objective was to create a system that adapts not only to rapidly changing curricula, but also to diverse learner trajectories and self-regulated learning behaviors. The proposed system is built on the Evolving Knowledge Space Graph model, a graph-based knowledge representation framework that supports dynamic curriculum structure. The model was combined with a relational database architecture to facilitate real-time learning tracking, adaptive content delivery, and data-driven decision making. Furthermore, a generative AI-based assistant was used to automatically generate domain-specific content, including knowledge units, prerequisite relations, and quiz questions, thus significantly reducing the author’s time. The system incorporates multiple adaptive learning algorithms. The model uses intuitionistic fuzzy logic to represent learner knowledge states and forgetting over time. A classroom study with 45 participants was conducted to evaluate structural clarity, usability, learning outcomes, and algorithm performance. The results indicate that the integrated approach, combining graph-based modeling, database-driven tracking, and generative content creation, successfully supports the navigation of the learners and the comprehension of the domain. The Bayesian knowledge propagation algorithm demonstrated the highest knowledge gains, while log-data analysis revealed significant differences in self-regulated learning patterns between learners. These findings highlight the potential of hybrid ITS architectures that leverage structured domain models, relational data infrastructure, and generative AI to support personalized and scalable learning in dynamic educational settings.

Most knowledge representation models in tutoring systems treat knowledge as a static resource. In real world scenarios, knowledge evolves over time as circumstances change. Consequently, the information embedded in a tutoring model may become outdated or require revision, especially in rapidly changing domains such as software development. The Evolving Knowledge Space Graph model addresses this issue by introducing abstract time to represent knowledge dynamics. This paper analyzes the temporal dependencies within the Evolving Knowledge Space Graph model, introducing “before” and “after” relations, and extending the model to capture knowledge aging. Additionally, it proposes metrics for assessing knowledge differences and complexity. Finally, the feasibility of the approach is demonstrated through a case study that transforms the publicly available OpenJDK JMC system’s knowledge into an Evolving Knowledge Space Graph.

This paper proposes a novel approach to integrate tutoring functionality into AI systems to counteract the potential decline of human intelligence caused by AI-driven over-automation. Existing explainable AI methods primarily emphasize transparency while lacking inherent educational functionality. Consequently, users are essentially left as passive recipients of AI-driven decisions without any structured learning mechanism in place. To address this, this paper introduces the knowledge-sharing-bridge (KSB), a component designed to transform AI into an active tutor. Unlike traditional intelligent tutoring systems (ITS), which operate separately from AI decision-making processes, the KSB is embedded within AI frameworks, ensuring continuous and context-aware learning opportunities. The proposed framework uses structured knowledge representation tools, such as category maps and word-clouds, to improve the user’s understanding of the decisions made by the AI systems. Prototype implementation demonstrates how these elements work together to provide real-time, interactive learning experiences. The results indicate that integrating KSB into AI enhances both explainability and user learning. This approach promotes a more in-depth interaction with AI insights and enables AI systems to become lifelong learning companions, closing the gap between automation and education.

One of the most actively researched areas in the field of process mining is process discovery, which aims to construct a schema that aligns with existing event trace sequences. Current standard industrial workflow schema induction methods impose certain limitations on the system being examined. To address the shortcomings, this article proposes a novel solution that employs graph neural networks and convolutional neural networks to perform schema discovery. In the first phase of schema generation, we perform equivalence prediction, implemented as an edge prediction task. From the obtained equivalence network, we identify the target schema nodes, which correspond to the maximal quasi-cliques of this network. The results of the performed efficiency tests demonstrate that the proposed method can manage such complex cases that are not covered by standard process discovery methods, and it provides more compact and more precise schema graphs.

The accurate prediction of kinematic configurations, including joint limit violations, collisions, and trajectory anomalies, is essential for ensuring the safety and efficiency of 6-DOF robotic arms in industrial applications. In this study, we develop a novel inverse kinematics (IK) solver based on a data-driven approach utilizing Graph Neural Networks (GNNs). Our model effectively captures the complex spatial relationships governing kinematic behavior by representing the robotic system as a graph—where nodes correspond to joints and edges represent physical linkages. Trained on a dataset of direct kinematics, our GNN-based model infers joint angles given an end-effector position with high accuracy and efficiency. The proposed approach achieves an accuracy of 92% with a 100.0% success rate and a computational runtime of 2.4 ms, outperforming conventional numerical and optimization-based IK solvers. These results highlight the potential of GNNs in real-time inverse kinematics prediction, enabling improved motion planning, reduced computational costs, and enhanced adaptability in dynamic environments. This research paves the way for more efficient and scalable solutions in industrial automation, human-robot collaboration, and autonomous robotic systems.

In this study, we explore the novel application of clustering techniques in the context of regression analysis. Regression is a key method in data science aimed at predicting the value of real-valued functions for arbitrary input variables. The central hypothesis of our research is that clustering can be leveraged to identify subregions of the input domain where the target function exhibits greater regularity and stability, thus enabling more precise and reliable regression. Specifically, we propose a modified K-means clustering algorithm that optimizes clusters based not only on proximity in feature space but also on the homogeneity of the function values within each cluster, measured via reduced standard deviation. In this paper, we present a series of experimental evaluations analyzing the efficiency of the proposed method. These evaluations also include comparisons with standard k-means clustering and nearest neighbourhood regression methods. The results demonstrate that the proposed method not only reduces the standard deviation within clusters but also improves the regression accuracy. Received: 05 July 2025 / Accepted: 30 August 2025 / Published: 25 September 2025

The radial basis function architecture and the multilayer perceptron architecture are very different approaches to neural networks in theory and practice. Considering their classification efficiency, both have different strengths; thus, the integration of these tools is an interesting but understudied problem domain. This paper presents a novel initialization method based on a distance-weighted homogeneity measure to construct a radial basis function network with fast convergence. The proposed radial basis function network is utilized in the development of an integrated RBF-MLP architecture. The proposed neural network model was tested in various classification tasks and the test results show superiority of the proposed architecture. The RBF-MLP model achieved nearly 40 percent better accuracy in the tests than the baseline MLP or RBF neural network architectures.

Process-mining techniques have emerged as powerful tools for analyzing event data to gain insights into business processes. In this paper, we present a comprehensive analysis of road traffic fine management processes using the pm4py library in Python. We start by importing an event log dataset and explore its characteristics, including the distribution of activities and process variants. Through filtering and statistical analysis, we uncover key patterns and variations in the process executions. Subsequently, we apply various process-mining algorithms, including the Alpha Miner, Inductive Miner, and Heuristic Miner, to discover process models from the event log data. We visualize the discovered models to understand the workflow structures and dependencies within the process. Additionally, we discuss the strengths and limitations of each mining approach in capturing the underlying process dynamics. Our findings shed light on the efficiency and effectiveness of road traffic fine management processes, providing valuable insights for process optimization and decision-making. This study demonstrates the utility of pm4py in facilitating process mining tasks and its potential for analyzing real-world business processes.

In FCA, we often deal with a formal context K = ( G, M, I ) that is only partially known, i.e. only the attributes that belong to an observable set N C M are known. There must also exist a part H of the object set G - called a training set - that consists of elements with all attributes known. The concepts of K have to be determined using the subcontexts corresponding to the training object set H and to the observable attribute set N . In our paper, this problem is examined within the extended framework of the semiconcepts of the original context, which are generalizations of its concepts. Each semiconcept of the original context induces a semiconcept in both subcontexts. In this way, each semiconcept of the context is represented by an induced pair of semiconcepts, which can also be considered its approximations - as in the case of rough sets. We describe the properties of the mapping defined by this representation and prove that the poset formed by these semiconcept pairs is a union of two complete lattices. We show that these induced semiconcept pairs can be generated by using a simplified representation of them. As the number of semiconcepts grows exponentially with the size of the training set and the observable attribute set, an algorithm that selects the representation pairs for which their support and relevance reach a certain threshold is also presented.

The development of transformer-based language models brings a paradigm shift in the world of smart applications. The ChatGPT model opened new horizons in the field of natural language understanding and generation. This paper presents a survey on the history of transformer models, on the basic architecture and application areas. The last section is devoted to two use cases experiments on the application of ChatGPT. The first domain relates to Human-Level Programming and the second focuses on the semantic functional parsing of text sentences. The performed analysis demonstrates the big potential in the transformer language models.

This research explores the potential of Large Language Models in the context of healthcare solutions, with a specific focus on Sexual, Reproductive, and Maternal Health Rights (SRMHR) Question Answering (QA) in the low-resource language, Amharic (sic). To construct the dataset, we first collected data from medical textbooks and guidelines authored by reputable medical institutions and organizations. Utilizing automatic question-and-answer generation techniques, we then generated pairs for the dataset. Subsequently, the dataset underwent annotation, translation, and evaluation processes, resulting in a refined collection of 2.8k Amharic datasets. We use the dataset to fine-tune the LLaMA-2-Amharic model, with test results assessed using BLEU scores and human-level evaluations, demonstrating promising outcomes. The curated Amharic SRMHRQA dataset serves as a foundational resource for future research. However, further enhancements are necessary to optimize its efficacy, particularly within the realm of SRMHR for low-resource languages like Amharic. Future research could involve scaling up the dataset in terms of size, quality, and domain coverage.

This study focuses on the fundamental process of parsing sentences to create semantic graphs from textual documents. It introduces novel techniques for parsing phrases within semantic graph-based induction, employing both ChatGPT-based and Hybrid parser-based approaches. Through a thorough analysis, the study evaluates the performance of these methods in generating semantic networks from text, particularly in capturing detailed event descriptions and relationships. Results indicate a slight advantage in accuracy for the Hybrid parser-based approach (87%) compared to ChatGPT (85%) in sentence parsing tasks. Furthermore, efficiency analysis reveals that ChatGPT’s response quality varies with prompt sizes, while the Hybrid parser-based method consistently maintains excellent response quality.

One of the hot topics in machine learning is the field of GNN. The complexity of graph data has imposed significant challenges on existing machine learning algorithms. Recently, many studies on extending deep learning approaches for graph data have emerged. This paper represents a survey, providing a comprehensive overview of Graph Neural Networks (GNNs). We discuss the applications of graph neural networks across various domains. Finally, we present an advanced field in GNNs: graph generation.

The Vehicle Routing Problem is about optimizing the routes of vehicles to meet the needs of customers at specific locations. The route graph consists of depots on several levels and customer positions. Several optimization methods have been developed over the years, most of which are based on some type of classic heuristic: genetic algorithm, simulated annealing, tabu search, ant colony optimization, firefly algorithm. Recent developments in machine learning provide a new toolset, the rich family of neural networks, for tackling complex problems. The main area of application of neural networks is the area of classification and regression. Route optimization can be viewed as a new challenge for neural networks. The article first presents an analysis of the applicability of neural network tools, then a novel graphical neural network model is presented in detail. The efficiency analysis based on test experiments shows the applicability of the proposed NN architecture.

A key problem domain inside Robotic Process Automation is the automatic discovery of workflow process schemes. Considering current process mining technologies, graph-based approaches dominate the industry. On the other hand, the conventional methods suffer from low time efficiency and varying accuracy. Machine learning-based methods can provide better efficiency, but they have significant limitations considering schema flexibility. The paper presents a novel neural network-based schema induction model for the discovery of event patterns containing parallel and optional sequences of different actors. This model can process more complex event graphs and situations than the conventional methods. The performed analysis and test results show the unique power of this approach in process schema mining.

AbstractAn intelligent tutoring system is a computer-based educational tool designed to provide adaptive learning environment to learners, mimicking the role of a human tutor. Its most typical areas of application are language learning, mathematics education, programming courses and medical training. Intelligent Tutoring Systems are based on the knowledge-module that is holding the system's knowledge in a well-structured format. Considering the current state of the art knowledge-module representations, a model that can represent evolving information is lacking. Representing evolving information is needed for those tutoring systems that are working with dynamically changing domains, e.g., software science. In this paper a new combined model is presented that is based on the ontology model and the fundamentals of knowledge space theory. The proposed model introduces the term of abstract time to be able to formulate an evolving knowledge graph. This paper introduces the term of evoking-hooks that makes it possible to realize connections between external domain elements and the nodes of the proposed model.