This paper examines the relationship between learning environments and changes in Information and Communication Technologies (ICT)-related career aspirations across education systems in the context of ongoing digital transformation. The analysis uses country-level data from the Programme for International Student Assessment (PISA) 2018 and 2022, combining indicators of student autonomy, digital skills and teacher support. Digital readiness is operationalized as a latent country-level construct combining digital skills, student autonomy and perceived teacher support. The study uses regression analysis as the primary associational approach, complemented by clustering and latent-representation techniques used for exploratory profiling and visualization. Unlike prior research that treats learning loss, digital skills and career expectations separately, our analysis integrates them within a comparative longitudinal framework. It shifts the focus from short-term post-pandemic effects toward the broader capacity of education systems to support digital preparedness and future-oriented career expectations. Digital skills were consistently associated with ICT aspiration growth across education systems, while teacher support played a complementary contextual role. Autonomy shows weaker, context-dependent effects. The findings support interpreting digital readiness as a multidimensional construct reflecting combinations of digital skills, autonomy and teacher support.

This study investigates the structural relationships among reading, science and mathematics performance, socio-economic dispersion (SES_Variance) and gender gaps using cross-country PISA 2022 data. The primary objective is to assess whether reading functions as an upstream determinant of academic achievement and to evaluate the causal roles of SES and gender disparities within an integrated probabilistic framework. We employ Bayesian network (BN) modelling with both constraint-based (Peter-Clark) and score-based (hill-climbing with BDeu score) algorithms under multiple discretization schemes, complemented by ANOVA analyses. Policy-relevant interventions are simulated using do-operator logic to examine SES and gender gap equalization scenarios. The BN also supports probabilistic queries that define performance archetypes. Robustness is assessed through alternative binning strategies and bootstrap stability analysis. The BN reveals a stable performance chain (Reading -> Science -> Math) and a direct influence of SES_Variance on all subjects. Contrary to econometric associations, gender gaps exhibit no consistent direct causal effect on performance, except for a discretization-sensitive link from GenderGap_Math to Mean_Science_2022 in 4-bin models. SES equalization simulations produce counter-intuitive reductions in high-performance probabilities.

Purpose The complexity of energy communities (EC) and the growing body of scientific literature on energy sharing (ES), value sharing (VS) and local electricity market (LEM) mechanisms create a significant gap between theoretical model development in scientific papers and practical implementation. To address this gap, the authors' paper aims to propose a Paper-to-Code agentic framework for EC that translates peer-reviewed scientific literature into executable models for data-driven decision support.Design/methodology/approach Paper-to-Code integrates large language models, vector-based knowledge retrieval and expert supervision to formalize the ES/VS/LEM described in publications into validated specifications, pseudocode and modular code implementations.Findings The framework enables EC to benchmark alternative operational and market strategies under identical data conditions, bridging the gap between research and digital solutions for EC decision-making. Across 200 scientific papers, 3 EC test cases, 100 profiling conversations and 30 expert-evaluated simulations, the framework achieved: near-perfect grounding (approximate to 0.97-1.00), high algorithm preservation (>= 0.96), >= 95% profiling and recommendation consistency, expert validation (approximate to 9/10 usefulness score).Originality/value These models are applied to EC data sets through an agent-based simulation environment, where community-specific objectives such as cost minimization, self-consumption maximization and grid dependence reduction are evaluated. By coupling automated knowledge extraction with expert-in-the-loop validation and operational simulation, Paper-to-Code ensures traceability and reproducibility.

This paper investigates how generative-artificial intelligence (AI) is influencing job requirements, skill compositions and sectoral dynamics across global labor markets. It examines the evolving frequency and framing of AI-related competencies in job postings, exploring whether generative-AI functions primarily as an augmentative or substitutive component in the workplace. A large-scale, multi-source corpus of over 150,000 English-language job postings (2018–2025) is compiled from twelve open-access datasets and one public API. The analytical framework integrates lexical skill extraction, semantic framing, topic modeling and time-series forecasting. Skill mentions are categorized into five dimensions: AI_Data, Routine, Soft_Meta, Domain_Specific and Leadership, while cross-sectoral analyses and correlation matrices quantify interdependencies between competencies. Sentence-transformer embeddings and cosine similarity are used to compute a Framing Index, distinguishing augmentation- vs. automation-oriented discourse. Results reveal a statistically significant increase in AI-related skill mentions after 2021 (t = −4.70, p = 0.016), alongside a decline in routine-task skills such as data entry and manual coding. Forecasts indicate potential continued growth in AI_Data and Soft_Meta skills through 2025, signaling a structural convergence toward hybrid human-AI expertise as a new foundation of employability. However, the results do not imply direct evidence of workforce transformation beyond what job-posting data can support. Investigating job postings, our research contributes a replicable, data-driven methodology for mapping the diffusion of AI-related skills across industries and time.

The regulatory framework governing EU electricity markets is highly complex, fragmented across multiple normative acts and sensitive to citation accuracy and contextual completeness. While Large Language Models (LLMs) offer promising capabilities for regulatory question answering (QA), their tendency to hallucinate legal references and omit critical conditions makes them unreliable for compliance-sensitive domains. This paper presents the design of a domain-specific Retrieval-Augmented Generation (RAG) system for EU electricity market regulations, explicitly engineered to deliver source-grounded, traceable and low-hallucination answers. The answering component is based on Google's gemini-2.5-flash model. The Open AI's gpt-4o-mini model is responsible for both relevant document selection before building the RAG prompt and playing the judge LLM role for Retrieval Augmented Generation Assessment (RAGAS) evaluation. We build a legal corpus comprising multiple core EU regulatory acts related to REMIT and market operation and propose a regulatory QA architecture that integrates: (i) three chunking strategies (article-based, structure-aware, sliding window), (ii) two embedding models and (iii) a novel LLM-based document selection agent that restricts retrieval to the most relevant normative acts before vector search, improving contextual focus and retrieval precision. Using a fixed benchmark of regulatory questions and a reproducible evaluation protocol, we quantitatively assess system performance with RAGAS metrics and classical information-retrieval measures. While all configurations achieve strong faithfulness (up to 0.96), answer relevancy varies substantially with embedding and chunking choices. The findings confirm that retrieval engineering, particularly embedding selection, chunking strategy and pre-retrieval document filtering, has a high impact for building reliable regulatory AI systems. The sliding window strategy combined with bge-small-en-v1.5 delivered the strongest rank-sensitive retrieval performance, achieving the highest Precision@10 and NDCG@10. In contrast, article-level chunking with the same model yielded a modest improvement in Recall@10, indicating a clear trade-off between recall and precision-oriented ranking quality in legal corpora.

In this research, we propose a methodology for explaining AI models and determining the critical features for retaining customers to prevent churn. By aggregating feature importance values across multiple explainers, our research harmonizes different explainers into a unified importance score, reducing bias and variability. It focuses on impactful features by filtering rows with positive recall drops and comparing results with established techniques like SHAP and LIME. Customers on month-to-month contracts and those with high monthly charges are at the highest churn risk, particularly when tenure is below 20 months. The results also indicate strong alignment between feature importance rankings derived from SHAP and the proposed multiple Perturbation-Based Explainers (PBE). Spearman Rank Correlation, Kendall Tau and Jaccard Similarity, demonstrate high agreement between the methods considering several classifiers. Furthermore, Kolmogorov-Smirnov and Wasserstein distance tests confirm minimal differences in their feature importance distributions. A comparative evaluation of SHAP, LIME, PBE based on fidelity, consistency and runtime shows that SHAP offers the highest fidelity but is computationally expensive (289s). LIME achieves moderate fidelity with faster runtime (35s), while PBE is the most efficient (18s), making it ideal for real-time or resource-constrained scenarios. Runtime estimates for SHAP reveal significant scalability challenges. A sensitivity analysis using alternative objectives (e.g., F1-score drop) confirms the stability of feature rankings, demonstrating the robustness of the proposed metric-agnostic framework. Additionally, scalability and deployment considerations are discussed, highlighting the suitability of the approach for integration into real-time CRM systems and drift-aware operational environments.

Our research investigates the dynamic relationship between economic growth and environmental sustainability in Serbia, a country situated at the intersection of Western environmental standards and Eastern development models. Specifically, it examines whether economic growth, as measured by GDP per capita, conflicts with environmental sustainability. It also considers the roles of renewable energy consumption (RENC), urbanization (URB) and trade openness (TO) as mediating variables. The analysis is based on annual time series data for Serbia from 1995 onward. A Vector Error Correction Model (VECM) framework is employed to assess both short-run and long-run relationships among the variables. Our research addresses the underexplored question of how countries like Serbia can reconcile growth with sustainability in a policy space shaped by contrasting regional norms. The results indicate a short-run trade-off between GDP and environmental sustainability, as lagged CO₂ emissions and RENC negatively affect GDP growth. However, in the long run, growth is positively associated with TO and URB, while RENC is strongly driven by URB. CO₂ emissions appear to evolve relatively independently of TO and URB. The adjustment coefficients confirm that GDP, CO₂ and RENC significantly respond to deviations from long-run equilibrium, with URB playing a central role in stabilizing the system.

This research presents a regime-aware hybrid forecasting framework for the Bitcoin market’s nonlinear, nonstationary and regime-switching behavior. The architecture integrates econometric models, neural forecasting and meta-learning, unified under a regime-detection mechanism using probabilistic inference. Central to the approach is a Hidden Markov Model (HMM) trained on log returns, which infers latent market regimes, bull, bear and sideways, based on statistical characteristics rather than arbitrary thresholds. Each detected regime triggers a specialized forecasting model: ARIMAX for volatile bear markets, SARIMAX for cyclical sideways periods and NeuralProphet for nonlinear bullish dynamics. These models leverage historical returns (Jan. 2012-Jun. 2025) and external signals, including technical indicators (RSI, MACD, Bollinger bands) and volatility metrics. A meta-learning layer, implemented via XGBoost, dynamically selects the optimal model at each time step based on the regime. This enables real-time adaptation to evolving market conditions. Predictions are made on log returns and translated into price forecasts through exponentiation. The framework’s performance is evaluated using R2, Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The regime-aware model outperforms the no-regime model significantly across all metrics, especially in error reduction (MAE cut by 56

Research interest in Explainable Artificial Intelligence (XAI) has surged since 2018, accelerating after 2020. Our research explores the role of XAI in e-commerce, emphasizing its ability to enhance transparency, interpretability and trust in AI-driven systems. To address the limited research on XAI in e-commerce, this paper conducts a bibliometric and semantic analysis of 8,845 publications from the Web of Science. It employs four advanced topic modeling techniques: Latent Dirichlet Allocation (LDA), BERT-LDA, BERT-Clustering and BERTopic. The research identifies XAI tools, popular XAI applications in e-commerce and emerging trends. The analysis also examines market type evolution and measures research momentum through the Compound Annual Growth Rate (CAGR). Keywords such as “Natural Language Processing”, “Neural Networks”, “Transformer Models” and “SHAP” exhibit the highest CAGR values, exceeding 80%. Using Hugging Face NER pipeline, we obtain that SHAP is the most frequently used tool with 2,182 mentions, followed by ICE, ALE and CEM. Extracting “market” from abstracts, the analysis reveals a significant shift in XAI research from traditional financial markets to emerging areas like energy, sustainability and environmental sciences. In e-commerce, fraud detection is the most prominent XAI application with nearly 40 mentions, followed by inventory management, price prediction, churn and demand forecasting with 15-20 mentions each. Topic 2 by BERTopic is closely related to e-commerce and focuses on the theme of recommender systems.

Energy communities (ECs) members have a high degree of heterogeneity. Agent-based modelling (ABM) allows for the modelling of each entity as an agent with distinct characteristics and decision rules. This is particularly useful when the heterogeneity of agents affects outcomes. Market-based models handle competition and supply-demand interactions and often rely on aggregate or average behaviours, which might overlook important nuances due to individual differences. In this paper, we combine ABM with reinforcement learning (RL) and market models to trade the surplus and demand at the local electricity markets (LEMs) level embedding PyMarket and Mesa packages. It explores integrating electric vehicles, heating and flexibility as vectors to expand ECs using an RL agent to optimally schedule these devices for efficient bidding. Two strategies are proposed: S1-RL agent predicts only the bidding price and S2-RL agent predicts both price and quantity. The RL agent in S2 optimizes load to increase demand during local generation intervals, utilizing local surplus more effectively. Financially, S2 outperforms S1, offering more flexibility and optimization in LEMs trading. In summer, the financial savings (FST) increase from 11.58% in S1 to 17.88% in S2, while, in winter, they increase from 4.21% to 5.38%. The efficiency of trading is better, and the traded quantity doubles in summer in S2 compared to S1.

This paper proposes an AI-based trading framework that integrates supervised price forecasting with reinforcement learning (RL)-based decision-making. The objective is to enhance both profitability and risk management in cryptocurrency trading by equipping RL agents with forward-looking market information and risk-aware incentives. The proposed methodology follows a two-stage design. First, a univariate long short-term memory (LSTM) model generates 72 bitcoin price forecasts. These predictions are used to compute future technical indicators, which are combined with current market indicators to construct an enriched, forward-looking state representation. Second, an RL agent is trained in this environment using a novel long-term reward function that incorporates transaction costs, drawdown penalties, volatility penalties, and delayed rewards to promote stable and sustainable trading behavior. Four state-of-the-art RL algorithms (PPO, SAC, TD3, and A2C) are systematically evaluated over randomized 180-day episodes using hourly bitcoin data. The results demonstrate that the proposed agent consistently outperforms conventional buy-and-hold and moving average crossover strategies, achieving an average profit ratio of 32% and a Sharpe ratio of 1.34. These findings highlight the novelty and effectiveness of combining mid-term price forecasts, enriched technical states, and risk-aware RL training for robust cryptocurrency trading.

This paper presents a metadata analysis framework to investigate research trends and predict citations in the field of recommender systems. We apply Natural Language Processing (NLP) techniques, using spaCy and regex, to extract and analyze frequencies of different system types and technologies. To explore relevant researchtopics, we implement both BERTopic and Non-negative Matrix Factorization (NMF). We also introduce a zeroshot classification pipeline with the facebook/bart-large-mnli model to categorize articles into recommender system categories without requiring supervision. For citation prediction, we employ the distilbert-base-uncased model from Hugging Face to tokenize and vectorize text features, engineering 13 additional features to enrich the dataset. Random Forest, XGBoost and LightGBM models are used to evaluate predictive performance using RMSE and R² under both baseline and feature-enhanced scenarios. The proposed feature engineering and dimensionality reduction approach significantly improve citation prediction performance compared to the baseline. Specifically, RMSE is reduced from approximately 19–21 to around 6.8–10.6, while the coefficient of determination (R²) increases from 0.38–0.45 to up to 0.93 across the evaluated models. The most notable improvement is observed for the Random Forest and LightGBM models, where RMSE decreases by over 65% and R². Context-aware systems, with approximately 300 mentions and knowledge-based systems, with fewer than 100 mentions, receive comparatively less attention in the literature. BERTopic reveals a diverse and evolving landscape in recommender system research, spanning deep learning models, privacy and federated learning, domain-specific applications in tourism, education, healthcare and fake news. Explainable and cross-domain recommendations appear more prominently in recent years.

Typically, after the market clearing, some participants remain unmatched due to price discrepancies. As a result, they are left with the option to buy electricity at the higher Time-of-Use (ToU) rates or sell at the lower Feed-in-Tariff (FiT) rates. In this paper, we explore and compare 5 basic pricing mechanisms and 5 more complex mechanisms, including a modified Vickrey-Clarke-Groves, Nash bargaining solution, cooperative game theory methods like the Shapley value, constrained optimization using Lagrange multipliers and dynamic mechanisms with Reinforcement Learning (RL), to facilitate continued trading on Local Electricity Markets (LEM) and integrate more renewables. First, we propose an Iterative Proportional Adjusting (IPA) method based on bid-ask spreads. Second, to develop an adaptive pricing post-auction mechanism that leverages both simple and complex methods, we propose a framework that dynamically selects the most appropriate pricing strategy based on market conditions, participant behavior and desired outcomes.

Evidence-based clinical decision making requires specialists to identify, evaluate and synthesize relevant scientific literature. However, PubMed searches for complex clinical cases often return hundreds of publications that cannot be reviewed manually under time constraints. This study proposes SCEPTER (Single-Case Evidence-driven PubMed-To-rEcommendation Reasoner), a framework for transforming clinical case descriptions into evidence-based recommendations. SCEPTER combines PubMed retrieval, PubMedBERT semantic ranking, large language model (LLM)-based claim extraction, evidence-level weighting, contradiction detection, consensus analysis and multi-objective Pareto claim selection. The framework generates structured evidence syntheses and grounded actionable recommendations. A Paper Q&A module further enables interactive exploration of selected publications. The proposed framework introduces multi-objective reasoning model that integrates literature support, contradiction analysis and interactive literature interrogation into a unified clinical decision-support pipeline. Evaluation on 150 case studies demonstrated that the framework reduced an average search space of 576 papers to 53 retained papers, 7 Pareto-optimal claims and 3 final recommendations, corresponding to an overall compression ratio of 192:1. Despite this reduction, the retained evidence maintained high diversity (entropy=0.901). The ablation study showed that Pareto-based selection increased evidence diversity and recommendation utility compared with conventional ranking approaches.

This paper investigates the academic discourse surrounding TikTok, the globally influential short-video social media platform, with a particular focus on its relationship with politics (accounting for 24.7% of the total papers). It aims to uncover trends, entities and links within TikTok-related research, addressing broader topics on the platform’s impact and narrower themes involving its political implications. Using bibliometric analysis of 2,313 publications indexed in the Web of Science (2019–2024), our research employs a dual framework combining sentiment analysis and topic modeling. To ensure robustness and capture complementary perspectives, we apply three sentiment analysis approaches alongside three topic modeling approaches (LDA, BERT-LDA, BERT-Clustering). This multi-method strategy enables us to cross-validate findings, highlight convergences and expose different nuances. The three approaches share overlapping themes but vary in scope and depth. The analyses reveal rising negativity and declining positivity over time, with peaks in negativity in the recent years: 2023 and 2024. Positive sentiment was highest in 2019 but steadily declined, while all methods highlight increasing critical sentiment and less emotional polarization in recent years. Key findings reveal significant trends in TikTok’s societal, cultural and political influence, highlighting its role in shaping public opinion, amplifying misinformation and influencing democratic processes.

This paper investigates how generative-artificial intelligence AI is reshaping job requirements, skill compositions and sectoral dynamics across global labor markets. It examines the evolving frequency and framing of AI-related competencies in job postings, exploring whether generative-AI functions primarily as an augmentative or substitutive force in the workplace. A large-scale, multi-source corpus of over 150,000 English-language job postings 2018-2025 is compiled from twelve open-access datasets and one public API. The analytical framework integrates lexical skill extraction, semantic framing, topic modeling, BERTopic, LDA, KMeans, and time-series forecasting ARIMA. Skill mentions are categorized into five dimensions: AI_Data, Routine, Soft_Meta, Domain_Specific and Leadership, while cross sectoral analyses and correlation matrices quantify interdependencies between competencies. Sentence-transformer embeddings and cosine similarity are used to compute a Framing Index, distinguishing augmentation- versus automation-oriented discourse. Investigating job postings, our research contributes a replicable, data driven methodology for mapping the diffusion of AI related skills across industries and time. Results reveal a sharp post-2021 increase in AI-related skill mentions: prompt engineering, fine-tuning and model validation, accompanied by a decline in routine tasks: data entry and manual coding. Forecasts suggest sustained growth in AI_Data and Soft_Meta skills through 2025, signaling a structural convergence toward hybrid human-AI expertise as a new foundation of employability.

The growth of distributed renewables and residential storage has turned consumers into active prosumers, creating new energy management challenges that require advanced optimization and transparent support. This study proposes an agentic AI-driven decision-support framework for load optimization and behavioral guidance that integrates models for photovoltaic (PV) generation and load profiling, short-term operational optimization, long-term system planning and a comprehensive performance evaluation layer based on energy, economic and battery health KPIs. These components are orchestrated with a four-agent architecture, comprising a Profiler, a Short-Term Advisor, a Long-Term Advisor, and a Q&A Agent that combine large language models (LLM) with explicit rule-based policies to produce reliable, interpretable and economic recommendations for energy management. Short-term optimization significantly reduces grid imports by aligning flexible loads with PV production and stabilizing battery behavior. Long-term optimization achieves near-autonomous system configurations, increasing self-sufficiency from 69.6% to 97.55% for Prosumer P1 and from 37.2% to 72.4% for Prosumer P2, while cutting grid dependence by more than half. The advisory agents successfully prevent unnecessary system oversizing, enforce battery protection and generate economic recommendations. The Q&A agent achieves high performance, with intent accuracy above 96%, numerical correctness above 93% and high user-rated usefulness, confirming the framework's accessibility for nontechnical users.

The shift to 15 min market time units (MTUs) in single-day-ahead coupling (SDAC) increases temporal granularity, but complicates the interpretation of intra-hour electricity price spikes and rapid ramps. This paper examines whether architectural decomposition improves the reliability of large language model (LLM)-based diagnostics in price-only settings, rather than causal market analytics, under severe information constraints. We compare a proposed agentic workflow featuring structured context extraction, spike/ramp detection, hypothesis generation, consistency checks, and explicit uncertainty calibration against non-agentic baselines. The paper contributes: (i) a reproducible benchmark for 15 min diagnostic question answering in day-ahead markets, (ii) an agentic architecture tailored to structured time-series reasoning with explicit uncertainty handling, and (iii) empirical evidence that decomposition and verification improve evidence grounding and trustworthiness in market analytics. The evaluation includes 360 price-only cases sampled across autumn 2025, winter 2025-2026, and early spring 2026, balanced by bidding zone, temporal period, event type, and impact tier, comprising 180 spike and 180 ramp cases from six Central and Eastern European bidding zones (Bulgaria, Czechia, Hungary, Poland, Romania, and Slovakia). Using identical inputs, we assess automatic reliability metrics and human ratings. The agentic workflow improves reliability (triangle = +0.067, 95% CI [+0.049, +0.085]) and significantly increases calibrated price-only disclaimers (triangle = +0.500) relative to the monolithic LLM baseline. Human evaluation confirms higher overall quality (+0.74), helpfulness (+1.06), and correctness (+0.94), with a 65.5% pairwise win rate. Overall, the results support a narrower conclusion: structured decomposition and verification improve calibration and perceived explanation quality relative to a simple monolithic LLM baseline, but their advantages are not uniform across stronger non-agentic baselines and remain limited by the absence of exogenous market data.

Large Language Model (LLM)-based code analysis tools are adopted to automate software documentation tasks. However, the scalability of these approaches to real codebases, where Intermediate Representations (IR) exceed LLM context limits, remains underexplored. This paper introduces an agentic architecture with context engineering for automated UML diagram generation from source code repositories. It employs a hierarchy of five specialized agents: PlannerAgent, AnalyzerAgent, DiagramAgent, CorrectorAgent and DependencyAnalyzerAgent, built on the Claude Agent SDK, each addressing a distinct cognitive subtask. A deterministic, importance-weighted IR compaction layer transforms full project IRs into diagram-specific views guaranteed to fit within token constraints, requiring no LLM calls and completing in milliseconds. Thus, we evaluate the system across 12 open-source repositories in 4 programming languages (Java, JavaScript, PHP, Python) and 7 UML diagram types, producing 84 observations assessed on 5 automated metrics. Results demonstrate high syntactic validity (mean: 91.5

Learning human preferences in language models remains fundamentally challenging, as reward modeling relies on subtle, subjective comparisons or shades of gray rather than clear-cut labels. This study investigates the limits of current approaches and proposes a feature-augmented framework to better capture the multidimensional nature of human judgment. Using the Anthropic HHRLHF dataset, we evaluate ten diverse large language models LLMs under a standard pairwise preference setting, where baseline performance remains below 0.74 ROC AUC, highlighting the difficulty of the task. To address this, we enrich textual representations with interpretable signals: response length, refusal indicators, toxicity scores and prompt response semantic similarity, enabling models to explicitly capture key aspects of helpfulness, safety and relevance. The proposed hybrid approach yields consistent improvements across all models, achieving up to 0.84 ROC AUC and significantly higher pairwise accuracy, with DeBERTav3Large demonstrating the best performance. Beyond accuracy, we integrate SHAP and LIME to provide fine-grained interpretability, revealing that model decisions depend on contextualized safety and supportive framing rather than isolated keywords. We further analyze bias amplification, showing that while individual features have weak marginal effects, their interactions influence preference learning.