Evaluation and selection of explanations in explainable artificial intelligence (XAI) is essential as AI systems grow, increasing transparency needs. Existing methods rarely built on theoretical grounding and failed to distinguish perceived and actual user understanding. This paper proposes a comprehensive user evaluation method for XAI explanations in a structured questionnaire format, enabling efficient data collection and comparison. Core explanation dimensions recognized in literature were integrated into our method: trust, satisfaction, mental models, and mental effort. Reliable instruments analyzed with exploratory factor analysis were used to assess trust and satisfaction. Mental models were assessed through reflection, prediction, and glitch detection tasks. Ten XAI techniques were evaluated: image-based (LIME positive and negative masking, feature occlusion, saliency maps, SHAP, LIME overlay mask) and numeric-based (SHAP bar chart, partial dependence plots, decision tree, counterfactual explanations, ANCHOR). Our method effectively differentiates explanations across multiple dimensions, aiding the selection of the most suitable technique for end-users.

The vast majority of machine learning research focuses on improving the correctness of the outcomes (i.e., accuracy, error-rate, and other metrics). However, the negative impact of machine learning outcomes can be substantial if the consequences marginalize certain groups of data, especially if certain groups of people are the ones being discriminated against. Thus, recent papers try to tackle the unfair treatment of certain groups of data (humans), but mostly focus on only one sensitive feature with binary values. In this paper, we propose an ensemble boosting FairBoost that takes into consideration fairness as well as accuracy to mitigate unfairness in classification tasks during the model training process. This method tries to close the gap between proposed approaches and real-world applications, where there is often more than one sensitive feature that contains multiple categories. The proposed approach checks the bias and corrects it through the iteration of building the boosted ensemble. The proposed FairBoost is tested within the experimental setting and compared to similar existing algorithms. The results on different datasets and settings show no significant changes in the overall quality of classification, while the fairness of the outcomes is vastly improved.

Digitalna knjižnica Univerze v Mariboru - institucionalni repozitorij Univerze v Mariboru: diplomska, magistrska in doktorska dela; publikacije, raziskovalni podatki, drugi raziskovalni rezultati; izdaje univerzitetne založbe.

Autoencoder, an hourly glass-shaped deep neural network capable of learning data representation in a lower dimension, has performed well in various applications.However, developing a high-quality AE system for a specific task heavily relies on human expertise, limiting its widespread application.On the other hand, there has been a gradual increase in automated machine learning for developing deep learning systems without human intervention.However, there is a shortage of automatically designing particular deep neural networks such as AE.This study presents the NiaNet method and corresponding software framework for designing AE topology and hyper-parameter settings.Our findings show that it is possible to discover the optimal AE architecture for a specific dataset without the requirement for human expert assistance.The future potential of the proposed method is also discussed in this paper.

This paper presents results of classification on imbalanced data with ensemble allocation method. The result of the allocation method were compared to traditional techniques for dealing with imbalaced datasets – the sampling methods. The allocation method is a two level ensemble that combines unsupervised and supervised learning. In this research the first level of allocation the unsupervised anomaly detection is used as an allocator which is combined with several traditional classification method on second level of ensemble. The allocation method is tested on imbalanced datasets and the results are compared to two well used sampling methods – under-sampling of majority instances, and over-sampling with SMOTE which introduces new artificial instances of minority class to the dataset. Results of all of the methods were compared on overall accuracy and average F-score metrics. The results show that allocation method produces the best classification model, which is also supported by statistical analysis.

Multi-population genetic algorithms have been used with success for several multi-objective optimization problems. In this paper, we present a new general multi-population genetic algorithm for evolving decision trees. It was designed to improve the possibility of evolving balanced decision trees, simultaneously optimized for the predictions of each class. Single-population genetic algorithms namely tend to construct decision trees with great variance in single class accuracies. The proposed approach is tested over 10 UCI datasets, and it is compared with a single-population genetic algorithm as well as with traditional decision-tree induction algorithms. Results show that the designed multi-population approach provides classification results comparable to C4.5 and CART in terms of accuracy and tree size, while outperforming them regarding balanced solutions (in terms of average class accuracy and range of single-class accuracies).

In this article we highlight the main challenges in ontology visualization and propose a solution. First we present most common ontology visualization challenges, discuss the functionalities that users usually miss in ontology representation tools and propose the solutions to these problems. We present an implementation that addresses detected challenges with three different graphical representations of ontologies, forming a multiple view. The visualizations follow well-established concepts from graphic design by which we achieved less clutter and simpler comprehension of ontologies' structures. Furthermore, the implemented system includes functionalities from which users will in our opinion benefit most (all-clickable elements enable easy exploration and in-depth research, filtering, reveal/hide hierarchy levels, advanced search, etc).