Multi-Criteria Recommender Systems (MCRS) provide personalized recommendations by incorporating multiple evaluation criteria, enabling nuanced preference modeling beyond single-criterion approaches. However, traditional MCRS methods face challenges in capturing complex non-linear interactions among criteria and handling data sparsity, while deep learning approaches lack interpretability in explaining criteria contributions. This paper proposes MC-JEPA-Hybrid, a novel framework integrating traditional multi-criteria aggregation operators—Ordered Weighted Averaging (OWA) and Choquet integral—with the Joint-Embedding Predictive Architecture (JEPA). The framework employs a two-stage architecture where JEPA learns high-quality user-item embeddings through self-supervised learning, followed by criteria aggregation using learnable fuzzy measures with enforced monotonicity constraints. Shapley values provide theoretically grounded explanations for criteria contributions. Experiments on the Yahoo! Movies multi-criteria dataset demonstrate statistically significant improvements over state-of-the-art methods ( p < 0.001 ): 19.4 ϕ = 0.32 , r = 0.905 ), validating the interpretability of the learned aggregation rules.

Motion detection from video streams remains a foundational yet challenging task in computer vision, especially in real-world scenarios where dynamic backgrounds, subtle movements, and hardware constraints hinder conventional methods. While deep learning models achieve impressive accuracy, they require large annotated datasets and high computational resources, rendering them impractical for real-time and embedded systems. This paper introduces ProMEM, a novel unsupervised motion detection framework that synergizes Single Pixel Imaging (SPI) with a statistical Energy Distance (EoD) measure to efficiently capture temporal changes without any need for training or background modeling. By projecting frame blocks onto structured illumination patterns, ProMEM extracts robust motion cues even in noisy or low-contrast conditions. We propose two algorithmic variants: a lightweight optimized version for real-time applications and an enhanced version with morphological refinements for increased robustness. Comprehensive evaluations on CDNet2014 and AAU Trimodal datasets demonstrate that ProMEM achieves competitive performance compared to both classical and deep learning-based methods while preserving computational efficiency suitable for deployment on edge devices. Our results highlight ProMEM as a promising alternative for resource-constrained, real-world surveillance and monitoring applications.

Recommendation systems increasingly influence user-facing and high-stakes decisions, yet often exhibit biases correlated with sensitive attributes such as gender or age. Existing fairness metrics, including demographic parity and Pearson correlation, mainly capture linear relationships and may therefore miss nonlinear forms of discrimination. This paper proposes EBFair, a fairness-aware recommendation framework that connects an energy-inspired dual-objective view with distance correlation from Energy Statistics. The method jointly optimizes recommendation accuracy and demographic fairness, where the fairness term penalizes statistical dependence between sensitive attributes and predicted ratings. Because distance correlation equals zero if and only if two variables are statistically independent, EBFair can detect both linear and nonlinear biases. Experiments on MovieLens-1M with binary gender as the sensitive attribute show that EBFair reduces demographic parity gap by 62

A collaborative filtering recommendation system typically involves recommending objects and items that meet the user's needs and expectations. Recommendation systems are of paramount importance because it guides the users to search their items, in which these items are right for themself. However, most of the recommendation systems are limited to using the measures between the users or between the items based on the pairwise ratings value, without considering relationships between one user’s rating value and all other user’s ratings values. In this article, we propose a collaborative filtering recommendation system, this system integrated with the t-Test distance correlation to measure the dependence between the random vectors. The energy distance measures are based on certain Euclidean distances between sample elements. After explaining the proposed system, the evaluation is performed on the Jester5k database with two experimental methods, when data is partitioned in two approaches: Split and Bootstrap. The results show that the Precision-recall value of the t-TestCF is always higher than the two compared collaborative filtering recommendation systems (these two recommendation systems are available in recommenderlab).

Detecting anomalous objects in aquaculture ponds is an important but challenging problem due to the lack of labeled data. This study proposes an unsupervised learning method combining Panoptic Segmentation, Energy Distance, and Point of Interest (PoI) to automatically detect anomalous objects from pond images. The system extracts a 12-dimensional feature vector specialized for aquatic environments and calculates an Energy Distance Map measuring feature distribution differences between image regions. A two-step segmentation strategy is applied: segmenting objects with low-to-medium energy, then creating instances for high-energy regions corresponding to potential anomalous objects. Classification is based on dynamic thresholds from the 75th percentile, allowing adaptation to different image conditions. Experiments on three diverse scenarios with varying density and lighting showed promising localization results, demonstrating feasibility for smart aquaculture applications.

The profound advancements witnessed in artificial intelligence algorithms, particularly within the domain of machine learning, have led to considerable progress. A hallmark of machine learning algorithms is their capacity to process immense quantities of data and accurately forecast outcomes for tasks such as image recognition and classification. The maritime industry, a vital contributor to the expansion of global trade activities, contends with the formidable challenge of detecting and controlling the multitude of vessels traversing the oceans. Despite satellite images offering an extensive reservoir of oceanic data, a comprehensive appraisal of machine learning algorithm performance in tackling the critical issue of ship identification and detection remains largely absent. This investigation leverages machine learning algorithms from the scikit-learn library to discern and locate ships within the publicly available Kaggle dataset. The results show that using SVM gives the highest result with 80.41

Skin lesion classification is crucial for early detection of skin cancer, yet traditional methods are often error-prone and current deep learning models face challenges with imbalanced datasets and nonlinear dependencies between features and labels. To overcome these limitations, this study introduces an Energy-Based Model (EBM) with a Partial Energy Correlation Loss function (LPEC), designed to capture nonlinear relationships and enhance robustness, particularly for underrepresented lesion categories. The model was trained and evaluated on the ISIC 2019 and ISIC 2020 benchmark datasets under both multi-class and binary classification scenarios. Experimental results demonstrate consistently high performance, achieving strong accuracy, sensitivity, and specificity in detecting malignant lesions such as Melanoma and Basal Cell Carcinoma, while also highlighting challenges with rare categories. These findings confirm the effectiveness of integrating Partial Energy Correlation into EBMs, providing a reliable framework for automated skin lesion classification. The proposed approach shows strong potential for integration into computer-aided diagnostic (CAD) systems in dermatology and paves the way for further research into improving robustness and generalization across diverse clinical datasets.

Context-aware recommender systems enhance personalization by incorporating situational information beyond traditional user-item interactions. However, they often face challenges due to sparse and heterogeneous contextual signals. We propose EBC-CARS, an energy-based framework that formulates user-item-context interactions as an energy minimization problem, enabling contextual conditions to reshape the compatibility structure in a principled manner. Building upon this formulation, we introduce ED-EBC-CARS, which integrates Energy Distance as a statistically grounded regularizer to mitigate distributional discrepancies under heterogeneous contextual settings. Experimental evaluations on MovieLens-25M, Amazon Reviews, and Yelp demonstrate consistent improvements over conventional collaborative filtering and representative context-aware baselines in terms of RMSE and Precision@10, indicating the effectiveness of distribution-aware energy modeling for stable and reliable preference estimation.

Context-aware recommender systems (CARS) have emerged as essential tools for personalized recommendations, where contextual signals such as time, location, and user situation are incorporated to capture the dynamic nature of user preferences. Among existing approaches, contextual modeling is considered the most principled paradigm as context is directly integrated into the prediction function. However, current methods are limited by several challenges: data sparsity is exacerbated by tensor factorization, only linear interactions are captured by factorization machines, and theoretical interpretability is lacking in deep learning approaches. To address these limitations, EBM-CARS (Energy-Based Modeling for Context-Aware Recommender Systems) is proposed in this paper. Two theoretical foundations are unified in this framework: Energy-Based Models (EBMs), where user–item–context compatibility is measured through energy functions with lower energy indicating stronger preference, and Székely-Rizzo energy statistics, by which scale invariance and statistical consistency are guaranteed. The energy landscape is shaped through contrastive learning, whereby the intractable partition function computation is bypassed. Extensive experiments were conducted on three benchmark datasets: MovieLens-25M, Amazon Product Reviews, and Yelp. Improvements of 10–18% in recommendation accuracy over state-of-the-art methods (FM, DeepFM, NFM, xDeepFM) were achieved, and the effectiveness of energy-based contextual modeling is validated through comprehensive ablation studies

In this paper, a novel approach for color-based image segmentation is proposed, using a combination of sigmoid and softmax functions for probability estimation. This method is designed to address challenges in color segmentation such as unclear boundaries and image noise. An energy function is constructed comprising two main components: a data term based on color distance and a smoothness term to ensure continuity in segmentation. Through the application of the sigmoid function, color distances are transformed into class membership probabilities, thereby enabling the generation of soft segmentations. Concurrently, the softmax function is applied for multi-label classification, creating probability distributions for various object classes. The model’s performance is evaluated on the VOC2012 dataset, demonstrating its effectiveness in improving accuracy and quality of object segmentation. Experimental results indicate that the proposed method achieves superior performance compared to traditional threshold-based and clustering approaches, particularly for objects with distinct color characteristics. The integration of morphological operations for post-processing further enhances boundary delineation and noise reduction. This probabilistic approach offers a robust framework for color-based image segmentation across diverse imaging conditions.

Personalized recommender systems have become fundamental in various domains such as e-commerce, streaming platforms, and social media. Despite the success of traditional collaborative filtering techniques, their effectiveness is often limited by ignoring the contextual factors that influence user preferences. In this work, we propose ECBM-CF, an interpretable framework for context-aware collaborative filtering that integrates Concept Bottleneck Architectures with Energy-Based Models. Our approach captures intricate user-item-context interactions while providing semantically meaningful concept-level explanations that enhance transparency and trustworthiness. We evaluate ECBM-CF on large-scale benchmark datasets including MovieLens 25M, Amazon Reviews, and Yelp. Experimental results demonstrate that our method significantly improves recommendation accuracy and interpretability compared to conventional baselines. This study highlights the potential of combining energy-based modeling with concept bottlenecks to advance explainable and context-sensitive recommendation systems.

Motion pattern analysis plays a crucial role in both computer vision and clinical rehabilitation, where assessing motor synchrony is essential for understanding coordination and monitoring recovery. In this paper, we propose a novel approach based on Energy Distance Correlation (EDC) to evaluate synchrony in human motion patterns. Unlike traditional methods that primarily focus on isolated kinematic parameters, EDC is capable of capturing both linear and nonlinear dependencies between motion variables, thereby providing a comprehensive view of the motor control system. Three analytical scenarios are developed: (1) global synchrony across different subject groups, (2) inter-joint synchrony between specific joint pairs, and (3) chain-level synchrony and bilateral symmetry. Experimental evaluation on the CMU Motion Capture dataset demonstrates the effectiveness of the proposed method in distinguishing healthy subjects from elderly, neurological, and orthopedic groups under walking, running, and jumping tasks. Results indicate that walking is highly sensitive to neurological disorders, running highlights age-related decline, and jumping reveals musculoskeletal limitations. These findings suggest that EDC offers an objective and robust tool for motion assessment, opening new directions for clinical diagnosis, targeted intervention, and progress tracking in rehabilitation.

Context: The Segment Anything Model (SAM) has been recognized as a revolutionary foundation model for image segmentation with remarkable zero-shot generalization capabilities across diverse visual domains. Problem: However, the requirement for user-provided prompts is considered a significant limitation for fully automated industrial applications where human intervention is impractical. Solution: In this paper, a novel framework called E-POI-SAM (Energy-based Point of Interest for Segment Anything Model) is proposed, which integrates Energy-Based Models (EBM) with Point of Interest (POI) detection to enable automatic prompt generation. Three energy terms are formulated: local energy E_local for feature distinctiveness based on gradient analysis, context energy E_context based on energy distance for boundary detection, and saliency energy E_saliency for visual attention modeling. An Energy-Guided Loss function is designed to combine SAM’s segmentation loss with energy-based constraints for end-to-end training. Results: Extensive experiments on COCO, PASCAL VOC, and Cityscapes datasets demonstrate that E-POI-SAM achieves 67.8

Early detection and classification of skin lesions present a significant challenge in dermatology and healthcare, particularly in the early identification of malignant skin lesions. Traditional methods often struggle to capture the complex relationships among features in dermatological images and face difficulties in handling imbalanced datasets. We propose a novel approach that integrates Energy-Based Models (EBMs) with the Energy Correlation (EC) methodology. Experiments conducted on the ISIC 2020 dataset demonstrated the model's capability in detecting and classifying skin lesions, particularly malignant ones, achieving an accuracy of $\mathbf{9 8, 2 3 \%}$. These results introduce a new direction for applying EBMs with Energy Correlation to other complex medical image detection and classification tasks, while also offering an effective diagnostic support tool for dermatology specialists.

Skin lesion classification is a significant challenge in the medical field, particularly in the early detection of malignant skin lesions. Traditional methods require large amounts of data and often struggle to handle heterogeneous datasets, especially in medical datasets where there is an uneven distribution of disease types. A novel approach has been proposed that integrates Energy-Based Models (EBM) with an Energy Distance (ED) approach. This method aims to incorporate ED into the training process of the EBM model. The experimental results on the ISIC_2019 dataset achieved classification accuracy of 74.19

The goal of the collaborative filtering problem is to find accurate and efficient mappings from previously rated data at items of the users. Improving item-based collaborative filtering (IBCF) and user-based collaborative filtering (UBCF) involves understanding the mathematics of distance measures and finding the right balance between calculating similarity and providing recommendations very accurate. However, the popular distance measures for recommendation models only focus on measuring pairwise rating values between one user and another, or between one item and another. In this article, the authors have proposed a new recommendation model, which consists of building a collaborative filtering model with the bias-corrected distance correlation statistic. The correlation method focuses on measuring the rating values of one object with all ratings of the other object; the Bias-Corrected Distance Correlation (BCDCOR) provides an improved estimate of the distance correlation; it corrects the bias present in the original distance correlation. Experimental results are developed on the Jester5k dataset, with two popular evaluation methods for the recommendation models, namely precision and recall values. The experimental results show that with the Bias-corrected-based recommendation model between users and users, the Precision and Recall values of the proposed model are higher than those of the compared collaborative filtering recommendation systems.