Goran Keković

International Journal of Human–Computer Interaction

Exploring Learner Engagement in E-Learning Environments: A Predictive Analytics Perspective

Mikić, Vladimir; Keković, Goran; Mangaroska, Katerina; Ilić, Miloš; Kopanja, Lazar; Vesin, Boban

https://www.tandfonline.com/doi/full/10.1080/10447318.2025.2535530#abstract

During the last decade, embracing learner engagement in the development of educational technologies has contributed to the integration of favorable pedagogical practices and advanced learning tools. This research investigates relationships between student behavior interaction data in an e-learning environment and self-reported engagement data, and proposes a model for predicting engagement levels. Statistical analysis was conducted on data from 45 undergraduate students at the University of South-Eastern Norway in a one-semester programming course. Artificial neural networks were used to develop a prediction model for classifying engagement levels. The findings emphasize the importance of coding exercises, topic-based assessments, and explanatory hints, and demonstrate the feasibility of predicting engagement using learner activity, interaction time, and learning outcomes.

e-learning environment; learner engagement; behavior data; questionnaire

M21

RADIOENGINEERING

Artificial Bias Induction in Fourth-Order Cumulants Based Automatic Modulation Classification Algorithm in AWGN and Multipath Propagation Channel

Božović, Rade; Orlić, Vladimir; Keković, Goran

https://www.radioeng.cz/fulltexts/2025/25_02_0224_0233.pdf

Automatic modulation classification (AMC) is a widely used technique for recognizing modulation formats of signals considered a priori unknown. Due to low algorithm and hardware complexity, AMC algorithms based on fourth-order cumulants remain popular. This paper proposes a new AMC approach focused on manipulating theoretical expected cumulant values of real signal constellations through artificial bias induction, assuming improved performance. The method was evaluated through Monte Carlo simulations in AWGN and multipath propagation channels (with known and unknown impulse response). Results confirmed the superiority of the proposed approach in classification of real and complex signal constellations, especially at lower SNR values, with performance improvements up to 25%.

AMC; AWGN; bias; Binary Phase Shift Keying (BPSK); channel impulse response; cumulants; multipath; Quadrature Amplitude Modulation (QAM)

M23

Proceedings of International Scientific and Professional Conference “ALFATECH” Smart Cities and Modern Technologies

Optimization of AI methods for air pollution prediction

Keković, Goran; Božović, Rade

https://conference.alfatech.rs/

Air pollution is one of the major problems of large urban areas, and artificial intelligence (AI) methods can be used to predict pollution levels based on a wide range of parameters. This paper analyzes the use of artificial neural networks (ANN) based on the Levenberg–Marquardt algorithm with Bayesian regularization (LMBR). The results show high prediction accuracy, competitive with radial basis neural networks commonly used for regression tasks. It is also demonstrated that, in addition to tuned ANN parameters, selecting an optimal sample size can help balance the desired accuracy and deviation between simulated and real data. Relative error was used as a measure of deviation. The study shows that sample size alone is not always decisive for method efficiency, and that the full structure of input data must be considered.

air pollution; artificial neural networks; Bayesian regularization; Levenberg–Marquardt algorithm

M33

IEEE Transactions on Learning Technologies

Predicting Student Performance in a Programming Tutoring System Using AI and Filtering Techniques

Ilić, M.; Keković, G.; Mikić, V.; Mangaroska, K.; Kopanja, L.; Vesin, B.

https://ieeexplore.ieee.org/document/10608149

This study examines the use of machine learning (ML) and artificial neural networks (ANN) for predicting student grades in a programming tutoring system. It addresses the growing use of AI methods in e-learning prediction without sufficient analysis of the most appropriate techniques and input parameters. The experiment used interaction data from university students and explored correlations among predictors and between predictors and the target variable. By applying a filtering technique based on the minimum redundancy–maximum relevance (mrMR) criterion, the study showed that data structure and correlations play a major role in defining an appropriate prediction model. ANN (Levenberg–Marquardt algorithm with Bayesian regularization) outperformed ML methods and achieved the highest prediction accuracy.

artificial intelligence; accuracy; electronic learning; predictive models; artificial neural networks; statistical analysis; correlation

M21a

Proceedings of International Scientific and Professional Conference “ALFATECH” Smart Cities and Modern Technologies

A MODIFIED VERSION OF GRADIENT DESCENT ALGORITHM AS A SOLUTION OF LOCAL MINIMUM PROBLEM IN ARTIFICIAL NEURAL NETWORK

Keković, Goran; Božović, Rade; Stamenković, Negovan

https://conference.alfatech.rs/

This paper proposes software based on modified gradient descent and backpropagation algorithms for overcoming the local minimum problem in artificial neural networks. During network training, at the end of each epoch, the existence of the global minimum was checked through successive loss function values and by determining the percentage of successfully classified samples from training and test sets. The software was written in the C# programming language in an object-oriented and modular manner, with its own mathematical library, enabling further upgrades with additional artificial neural network algorithms.

artificial neural networks; neuron weights; global minimum; gradient descent; backpropagation algorithm; C# programming language

M33

Proceedings of International Scientific and Professional Conference “ALFATECH” Smart Cities and Modern Technologies

AN OPTIMAL THRESHOLD ADAPTATION OF ENERGY DETECTOR IN COGNITIVE RADIO FOR SMART CITIES APPLICATIONS

Božović, Rade; Keković, Goran

https://conference.alfatech.rs/

Smart cities represent a modern technological concept of interconnection among different devices for collecting and post-processing various types of data. Based on such information, the efficiency of city services (security, energy consumption, safety) can be optimized and improved, significantly enhancing citizens’ quality of life. Cognitive radio technology enables more efficient use of radio spectrum, while spectrum sensing is crucial for minimizing interference. Due to its low implementation complexity, a non-coherent energy detector is a suitable choice for spectrum sensing. This paper presents a closed-form solution for energy detector threshold adaptation in cognitive radio.

advanced smart communications; cognitive radio; energy detector; spectrum sensing; threshold adaptation

M33

ARTIFICIAL INTELLIGENCE CONFERENCE – Book of Abstracts (Serbian Academy of Sciences and Arts, SASA)

Stability of Machine Learning Methods

Keković, Goran

https://www.mi.sanu.ac.rs/~ai_conf/previous_editions/2024/AI_Conference_Book_of_Abstracts.pdf

This work investigates the stability (generalization property) of machine learning methods for data classification using four different publicly available databases. The analyzed methods were support vector machine (SVM), ensemble of bagged trees (bagged DT), and k-nearest neighbors (KNN). The input parameter space included mean coefficient of variation (CV), mean mutual entropy, and sample size. The input set was divided into equal subsets, where the first subset was used for training and validation, while the others served as unseen sets. The procedure was repeated iteratively twenty times, and mean total accuracy and mean relative error were determined. The results showed that the bagged DT method achieved particularly strong performance, especially in cases of low or negligible statistical dependence between input variables, and was recommended as the first choice under such conditions.

support vector machines; ensemble of bagged tree; k-nearest neighbor; coefficient of variation; mutual entropy

M34