Although explainability is critical for trustworthy AI, Group Recommender Systems (GRS) lack dedicated, flexible software tools. To address this, we present GREX, a novel, open-source Python library designed to facilitate the development and evaluation of explanations in group settings. GREX is built on a modular and extensible architecture, providing implementations of three distinct explanation paradigms: counterfactual (Sliding-Window-Weighted), rule-based (EXPGRS), and local model-agnostic (LORE4GROUPS). We conducted a comprehensive description of GREX stages, overall including data preparation, training, recommendation, and explanation. GREX is focused on empowering researchers and practitioners to systematically develop and compare Explainable AI (XAI) methods, fostering progress in transparent and user-centric GRS.
The lack of proper formalization in non-symbolic explainable artificial intelligence (XAI) has resulted in a growing number of provable flaws in widely adopted methods of explainability. This paper reviews some of the most visible misconceptions of non-symbolic XAI, and shows how logic-based XAI has been applied to uncover and correct those misconceptions. The paper also summarizes ongoing research to develop a sound, scalable and human-understandable framework for rigorously defined XAI.
The growing deployment of large language models (LLMs) in policy-sensitive and distributed environments raises concerns about hallucinations, semantic opacity, and lack of traceability, especially in regulated domains such as healthcare and finance. Retrieval-augmented generation (RAG) helps address these issues by grounding outputs in external sources, while data spaces (DSs) offer a governance-driven infrastructure for federated and sovereign data exchange. However, their integration remains underdeveloped, limiting the design of AI systems that are both explainable and compliant with regulatory demands. This paper analyzes three RAG–DSs integration models from the literature and introduces two new architectures, Guided RAG and Federated RAG, that directly address their structural limitations by combining coordination, data sovereignty, and trust in generative systems. Guided RAG and Federated RAG are formally specified using Business Process Model and Notation (BPMN), and all five models, together with a non-retrieval baseline, are analyzed across operational dimensions including transparency, traceability, and control delegation. This work establishes an architectural foundation for trustworthy generative AI operating under distributed governance constraints, with explicit guarantees of data sovereignty. The proposed framework supports systematic evaluation in federated settings and serves as a foundation for methodological refinement and regulatory alignment.
This research examines the utilization of Inverse Reinforcement Learning and Deep Q-Networks in traffic signal management to enhance urban mobility through the automation of reward function design. The study illustrates an 81.35
The increasing reliance on artificial intelligence (AI) in education has highlighted the necessity for transparent and trustworthy recommender systems. This paper presents a framework that combines collaborative filtering, specifically Singular Value Decomposition (SVD), with post-hoc, model-agnostic explanation techniques to generate interpretable recommendations in e-learning environments. To achieve this, we implement both crisp and fuzzy association rule mining approaches for explaining recommendations based on learners’ historical behaviors. The fuzzy model leverages membership degrees to account for interaction intensity, while the crisp model uses binarized user–item transactions. We generate explanations by mapping recommended items to association rules, and explanation quality is assessed using the Average Fidelity metric. Experimental evaluations on a real-world MOOC dataset demonstrate the effectiveness of the framework in delivering recommendations with interpretable justifications. The results validate the feasibility of integrating fuzzy-based explainability into recommendation pipelines, as the proposed approach enhances the crisp-based explanation approach, taking into account fidelity-based measures.
The automatic identification of crops in images plays a crucial role in the digitization of agriculture, especially related to monitoring and preserving the agroecosystem. This study presents an analysis and comparison of traditional and modern methods for identifying olive trees in RGB images. Among these approaches, we examine classical computer vision algorithms using OpenCV as well as advanced deep learning models, such as U-Net, which leverage convolutional neural networks. Identifying crops in olive groves using only RGB images relies exclusively on color information, which can lead to ambiguities when distinguishing vegetation from surrounding elements in the ecosystem. Factors such as varying lighting conditions, seasonal changes in foliage and different types of terrain further complicate the identification process. This study aims to evaluate the strengths and limitations of each approach by testing RGB images under different brightness, resolution and distance conditions and to examine its potential impact on optimizing agricultural processes. The correct selection of these segmentation techniques could represent a significant advance in agricultural automation, allowing more efficient and reliable management of olive groves through real-time integration of artificial vision tools.
The paper addresses the classification of Exchange-Traded Funds (ETFs) to propose a set of profitable ETFs for portfolio selection. An intelligent classification tool has been applied to infer the weights of the decision criteria from holistic preference information from the decision-maker. It classifies the ETFs in three risk categories. The resulting classification is not sensitive to minor modifications of the weight values due to the exploration of values of decision variables that match the decision makers’ preferences.
The olive sector is crucial in the province of Jaén (Spain), as it is the main source of income. Within this sector, the identification of olive varieties is an essential task that is hampered by the morphological variability of the leaves and by the fact that there are numerous varieties that directly affect the quality of the oil. This study uses deep learning to classify olive tree varieties, notably the Picual variety, using RGB images. A balanced dataset of leaf images was created through acquisition and segmentation. Six pre-trained Convolutional Neural Networks (VGG16, ResNet50, InceptionV3, Xception, MobileNetV2 and DenseNet121) were evaluated and adjusted using transfer learning. Among them, DenseNet121 achieved 83.2
This work presents an initial approach to studying the explainability of deep learning models at the layer level using fuzzy logic techniques. The aim is to link neural activations to input text concepts, analyzing their coherence and relationships. For this, a dataset of movie reviews is used, where each text is classified into one of seven emotions via a sentiment analysis model. The resulting activation vectors are clustered, allowing each vector to belong to multiple clusters, thus reflecting the ambiguity of textual data. The resulting clusters are analyzed through coherence measures to assess their internal consistency and uncertainty. These analyses help reveal how similar concepts are grouped in the neural activations. Preliminary results are promising, motivating further exploration of this approach. As an additional contribution, this work proposes the application of the same methodology to a domain with more objective concepts, such as wine recommendation. In this case, fuzzy profiles are defined based on organoleptic attributes of wine, and affinity scores between wines and user profiles are computed using TSK FIS models. These affinity values are used as neural network training targets, allowing subsequent analysis of how the network organizes these clearer and more structured concepts in its hidden activations.
Reducing uncertainty and imprecision in medical data is essential to improve the quality of diagnoses and treatments. This paper presents a statistical—fuzzy hybrid framework that integrates information–theoretic metrics and classical statistical indicators with fuzzy logic techniques to address these challenges. The framework combines geometric mean filtering—as a computational operator—with generalized compensation operators and evaluates uncertainty reduction using mutual information, clairvoyance value (VoC), and traditional error indicators to enhance data clarity and minimize uncertainty. Validation was performed using synthetic and real–world datasets, including electrocardiographic (ECG) signals and computed tomography (CT) scans. The results showed a significant decrease in the VoC metric, indicating improved data quality and reduced uncertainty. In the ECG recordings, the VoC reduction was particularly notable, improving signal quality for subsequent analysis. Comparison with previous studies demonstrates that the proposed framework offers substantial improvements in uncertainty reduction. Practical implications include more accurate diagnoses and the potential for personalized treatments. The limitations of the study include the need for greater validation in diverse clinical datasets. In conclusion, the hybrid framework effectively improves the quality of medical data and contributes to emerging frontiers in medical data analysis through the integration of statistical methods, fuzzy logic, and uncertainty metrics to support optimized clinical decision-making.
This paper explores the use of machine learning and natural language processing (NLP) techniques for the automatic classification of text-based appeals in the field of housing and communal inspection (HCI). The proposed approach includes a data processing module and a set of classification models. An experimental study was conducted comparing traditional classification algorithms such as Naive Bayes, Support Vector Machines, Decision Trees, K-Nearest Neighbors (KNN), FastText, and transformer-based models using BERT. The approach—combining preprocessed data with sampling based on class distribution in the training set and leveraging top predictions from multiple models—demonstrated effective performance based on evaluation metrics including accuracy, top-k accuracy, F1-score, and ROC-AUC. This method significantly reduces the workload on human operators and improves service quality in the HCI sector.
Rheumatoid Arthritis (RA) is a complex autoimmune disease requiring sophisticated modelling approaches for accurate progression prediction and treatment optimisation. This study introduces a novel application of Physics-Informed Neural Networks (PINNs) that integrates established pathophysiological laws with data-driven learning to predict three critical RA biomarkers: C-reactive protein (CRP), Disease Activity Score 28 (DAS28), and lymphocyte count. Our approach embeds ordinary differential equations representing inflammatory dynamics, disease activity, and immunological response directly into the neural network's loss function. Validation on a synthetic dataset of 400 patients over 12 months demonstrates superior performance with a mean R-2 of 0.7053, representing a 34.5% improvement over linear regression baseline. The model successfully learned 12 interpretable physical parameters, revealing unexpected medical insights including moderate CRP-DAS28 coupling (gamma = 0.1075) and significant endogenous self-resolution capacity (gamma = 0.0939). Clinical correlation analysis confirmed preservation of known medical relationships while respecting physiological ranges. This work establishes PINNs as a powerful tool for chronic disease modelling, offering enhanced interpretability, improved generalisation, and the potential for discovering novel pathophysiological insights with direct clinical applications.
Recommender systems have transformed how users access information, particularly in the news domain. While these systems enhance personalization, they also introduce fairness concerns, potentially limiting users’ exposure to diverse perspectives. This experimental study aims to evaluate fairness across various recommendation algorithms by comparing collaborative and neural network-based approaches. To assess performance, both classical evaluation metrics, together with bias and fairness-aware metrics, such as Bias Disparity, Ranking-based Equal Opportunity, and Average Recommendation Popularity, are considered. The Adressa dataset is used to analyze recommendation behavior and their impact when using various user and item attributes. The results highlight important variations in fairness distribution across different algorithms, underscoring the necessity of incorporating fairness considerations into recommender system design.
The development of fair and robust remote photoplethysmography (rPPG) systems is fundamentally constrained by the lack of diversity in public datasets, especially a scarcity of data from individuals with the darkest skin phototypes (Fitzpatrick VI). To overcome this, this paper introduces two key contributions. First, the EquiNet-DB, a new large-scale video dataset that significantly extends previous benchmarks by adding a large cohort of new participants with a specific focus on Fitzpatrick skin types V and VI. Second, EquiNet, a novel neural architecture featuring a Joint Spatio-Temporal Attention (JSTA) mechanism designed to excel in low-SNR conditions. Trained and evaluated on the EquiNet-DB, the EquiNet model sets a new state-of-the-art, achieving unprecedented accuracy and demonstrating a dramatic reduction in performance disparity across skin tones. This work provides a foundational contribution towards reliable and equitable remote health monitoring.
Real-world recommender systems often suffer from extreme imbalance in user feedback, as most users provide only positive opinions. This paper demonstrates that such imbalance has substantial negative effects on the system’s ability to generalize, particularly in predicting negative preferences. To address this issue, the study investigates whether intra-domain transfer learning, leveraging users with more complete profiles (those who provide both positive and negative feedback) can help mitigate these limitations. A synthetic evaluation framework is introduced based on false-positive and false-negative user profiles, allowing for a detailed assessment of model behavior under incomplete user data. Experiments on image-based datasets reveal consistent improvements in true negative rates for underrepresented user groups. Although the mitigation is partial, the results suggest that selecting appropriate user subsets as a source domain may offer a promising direction, and that identifying which users generalize best could be key to future progress.
Endometriosis is a chronic and painful disorder that significantly affects many aspects of a woman’s life. Its complex symptomatology makes early diagnosis and effective treatment particularly challenging. Although deep Convolutional Neural Networks (CNNs) have shown promise in medical image classification, few studies have explored their application to endometriosis detection. This study evaluates and compares the performance of three advanced CNN models (DesNet121, InceptionV3, and Xception) using laparoscopic images to identify endometriotic tissue. A custom dataset was compiled by merging images from the ENDI and GLENDA databases, with preprocessing steps including normalization and data augmentation. The models were trained and validated using stratified splits, and assessed based on standard metrics (accuracy, precision, recall, AUC, and confusion matrices). The results revealed accuracy scores of 89
The uncontrolled growth of global tourism requires a new generation of recommendation systems that guide users toward more sustainable choices. Existing systems, however, often prioritize popularity, failing to account for the complex sustainability impacts of tourism. Addressing this gap, this paper introduces a novel, comprehensive hierarchical framework for evaluating Points of Interest based on both user-centric and sustainability criteria. We further propose a detailed, multi-faceted data acquisition and estimation strategy designed to overcome the critical challenge of data scarcity and populate this framework with robust values. This hybrid approach integrates primary data from platforms and open repositories with advanced inference methods. In particular, we propose to employ proxy variables and leverage Artificial Intelligence methods including Large Language Models to estimate hard-to-measure criteria. Building upon this approach, we present a systematic guide for implementing each criterion, specifying its relevant data source, acquisition method and evaluation scale. This work provides a foundational methodology that spans from conceptual structure to data implementation, enabling the development of next-generation recommendation systems that guide tourists toward genuinely sustainable choices.
Visual SLAM systems often suffer from degraded performance in dynamic environments due to the inclusion of feature points on moving objects, which violate the static world assumption. In this paper, we propose a dynamic object filtering framework designed to enhance the performance of monocular ORB-SLAM3 by excluding features associated with dynamic regions. Our method integrates deep learning-based object detection, multi-object tracking, and dense optical flow analysis to generate per-frame binary masks that identify and suppress dynamic content. A motion similarity metric, which combines directional- and magnitude-based flow differences with spatial weighting, is used to assess object motion relative to the background. Temporal consistency is enforced through a probabilistic state filter to smooth the classification over time. We evaluated our approach on the KITTI Odometry dataset, and it shows that our solution significantly reduced pose estimation errors in highly dynamic sequences, while maintaining performance in static scenes. The results demonstrate the effectiveness of the proposed filtering strategy as a lightweight and general-purpose enhancement to existing SLAM systems when operating in real-world environments.
Due to their ever-increasing number of applications, multi-label classification algorithms are facing a major challenge: learning from evolving data streams with distribution changes over time, under limited computational and memory resources. In this paper, we first revisit existing works through a meta-model of low-complexity multi-scale algorithms combining short-term memory for fast drift adaptation and long-term memory to integrate distribution changes over time. Then, we develop a very strong baseline from this family, called A2ML (Adaptive Memories for Multi-Label stream classification), specifically designed for non-stationary streams. Its long-term memory is managed via adaptive label clustering and biased reservoir sampling, ensuring linear-time model updates. A2ML is compared to 7 state-of-the-art algorithms on 15 stationary and 4 non-stationary streams with over 100,000 examples, generated to test various data changes and concept drifts. Results show A2ML performs well in both settings and has lower computation times than competitors.
IoT systems generate large volumes of time-series data, but sensor malfunctions often lead to missing values that reduce the effectiveness of machine learning models. We propose a novel hybrid architecture that interleaves Long Short-Term Memory (LSTM) layers with a multihead attention mechanism, where the first LSTM layer captures local temporal dependencies, the attention layer highlights long-range relationships, and the second LSTM layer integrates these features into a coherent sequence. This structured design, unlike conventional orderings, enhances robustness against irregular missingness. Evaluated on six months of soil surface temperature data with simulated missing rates from 10 R^2 score ( R^2 ) and root mean squared error (RMSE). Performance is also compared to a statistical technique k-Nearest Neighbour (KNN) and a deep learning technique Bidirectional Recurrent Imputation for Time Series (BRITS) baselines. Importantly, training with simulated missingness further improved generalization, underscoring the novelty and practical relevance of the proposed approach for real-world IoT scenarios.
