Abstract
Hyperspectral image (HSI) classification remains challenging in real-world scenarios due to severe class imbalance and ambiguous boundaries caused by mixed pixels. Existing methods typically address class imbalance at the data or loss level, while overlooking its intrinsic coupling with spectral–spatial representation learning. This limitation often leads to the degeneration of discriminative structures and the dominance of majority classes in feature space. To tackle this issue, this paper proposes a multi-scale spectral–spatial selective network (M3S-Net) that jointly enhances feature representation and enables distribution-aware feature learning and optimization. Specifically, a high-frequency residual spatial attention (HRSA) module is designed to learn a smooth background-response suppression mask inspired by high-pass residual modeling, thereby mitigating the inherent over-smoothing effect of convolutional operations and enhancing boundary-aware representations. Furthermore, a dynamic selective fusion module (DSFM) is developed to adaptively aggregate shallow spatial details and deep semantic features, effectively bridging the semantic gap across different scales. To address long-tail distribution and improve generalization, a distribution-balanced regularization strategy is proposed by integrating inverse class weighting and label smoothing, thereby jointly alleviating distribution bias and suppressing overfitting. In addition, a test-time augmentation (TTA) scheme is employed to enhance inference robustness without introducing additional training cost. Extensive experiments on four widely used benchmark datasets demonstrate that the proposed method achieves top or highly competitive performance under limited-label and imbalanced evaluation protocols. For instance, M3S-Net obtains 85.69 ± 1.49% overall accuracy using only 1% of the training samples on Indian Pines, exceeding CLSJE by 0.31 percentage points and maintaining balanced average accuracy. © 2026 Elsevier B.V.

Abstract
This article investigates the ability of large language models (LLMs) to evaluate semantic relations between word pairs by examining their alignment with human-generated semantic ratings. Semantic relations represent the degree of connection (e.g., relatedness or similarity) between linguistic elements and are traditionally validated against human-annotated datasets. Due to the challenges of building such datasets and recent progress in LLMs' capacity to model humanlike understanding, we explore whether LLMs can serve as reliable substitutes for traditional human ratings. We conducted experiments using multiple LLMs from OpenAI, Google, Mistral, and Anthropic, evaluating their performance across diverse English and Portuguese semantic relations datasets. We included in the analysis PAP900, a recently published dataset of semantic relations in Portuguese, to examine the influence of prior exposure to the dataset on LLM training. The results show that the LLM predictions correlate strongly with human ratings. The findings reveal the potential of LLMs to supplement or replace traditional semantic measure algorithms and crowd-sourced human annotations in semantic tasks.

Abstract

Abstract

This work examines how key modelling decisions in Deep Reinforcement Learning (DRL) shape the quality of learned scheduling policies. Although DRL has been increasingly applied to the Flexible Job-shop Scheduling Problem (FJSP), existing studies provide limited understanding of how distinct action and reward definitions respond to specific problem settings. We address this gap by integrating Constraint Programming (CP) into the design and training processes of a multi-agent DRL framework. Using CP-derived optimality bounds, we analyse the impact of job and machine heterogeneity in the suitability of distinct action and reward formulations, considering standard FJSP benchmarks. Building on these insights, we (i) design a set of reward functions that outperform existing ones, reducing the optimality gap by roughly 10% in high-heterogeneity scenarios, and (ii) introduce a hybrid DRL method (DQL-D) that incorporates expert demonstrations from CP solutions into a multi-agent Deep Q-Learning architecture. This hybrid strategy accelerates convergence and improves stability, yielding a 5% reduction in optimality gap relative to a standard DQL baseline. Overall, our findings demonstrate that optimality-guided DRL modelling enables lightweight, fast, and reliable scheduling agents, meeting the real-time and decentralized decision support demanded by modern manufacturing systems.

Abstract
This paper focuses on the comprehensive assessment of regional performance in attaining the 2030 Strategic Framework for Education and Training (ET2030) established by the European Union. To this end, we propose a composite indicator framework based on robust Benefit-of-the-doubt models empirically validated through an extensive analysis of data spanning 32 countries and 101 NUTS-I level regions for 2019. We integrate contextual variables into a robust conditional model to ensure an equitable evaluation among regions grappling with distinct circumstances. Specifically, the unemployment rate and the percentage of the population holding national citizenship are considered. Moreover, the research identifies best practices from high-performing regions that can serve as benchmarks for underperforming areas. Analyzing regional-level data is crucial for understanding disparities between European regions and within countries.

Abstract
Our team previously introduced an innovative concept for an "invisible" Electrocardiography (ECG) system, incorporating electrodes and sensors into a toilet seat design to enable signal acquisition from the thighs. Building upon that work, we now present a novel dataset featuring real-world, single-lead ECG signals captured at the thighs, offering a valuable resource for advancing research on thigh-based ECG for cardiovascular disease assessment. To our knowledge, this is the first dataset of its kind. The tOLIet dataset comprises 149 ECG recordings collected from 86 individuals (50 females, 36 males) with an average age of 31.73 ± 13.11 years, a mean weight of 66.89 ± 10.70 kg, and an average height of 166.82 ± 6.07 cm. Participants were recruited through direct contact with the Principal Investigator at Centro Hospitalar Universitario de Lisboa Central (CHULC) and via clinical consultations conducted at the same institution. Each recording includes four differential signals acquired from electrode pairs embedded in the toilet seat, with reference signals obtained from a standard 12-lead hospital ECG system.