Early academic risk prediction aims to identify at-risk students from limited early-course behavioral observations and to provide evidence for timely educational support. Existing methods usually overlook the discrepancy between a student's current state and their historical trajectory, resulting in insufficient capture of early signals and limited interpretability. To address this issue, the authors propose EduRiskNet, an interpretable framework for early academic risk prediction. EduRiskNet integrates multiperspective behavioral encoding, semantic state modeling, and state deviation quantification to capture early risk formation signals in learning trajectories. A dual-scale dynamic module further combines long-term trajectory patterns with short-term anomalies for early risk identification. Experiments on two public datasets show that EduRiskNet outperforms the tested baselines. The model also organizes risk scores, state-deviation evidence, and dominant behavioral factors into structured information for decision-support priorities.
With the rise of location-based services applications, research on point-of-interest (POI) recommendation has gained significant attention. Conventional federated POI recommendation studies suffer from insufficient prediction accuracy. This study introduces a novel POI federated recommendation method: Spatio-Temporal self-Attention Recommendation for POI (STAR-POI). STAR-POI employs a spatiotemporal self-attention network to learn the spatial and temporal relationships between discontinuous check-ins, a projection enhancement gradient similarity index to cluster clients with their model updates. Model accuracy comparisons showed that STAR-POI outperforms three baseline models across all metrics, as it can capture more nuanced user behaviors. Further, variation analysis of model accuracy revealed the method’s saturation points concerning the number of clients, evaluation metrics, data distribution in the study area, and the projection enhancement factor λ. These insights can guide the selection of optimal parameters for model training and POI recommendation.
Online learning platforms accumulate student-question records for learning analytics, early warning, and institutional decision support. Mining these records requires accurate next-response prediction and stable modeling of temporal-relational patterns among items and knowledge components. Existing knowledge tracing methods use static structures or separate temporal and relational modeling, limiting their ability to capture evolving dependencies and causing relational drift. To address this issue, this paper proposes the Cross-Temporal Relational Alignment Network (CTRA-Net), a framework for joint sequential and dynamic structural modeling. CTRA-Net uses time-conditioned graph attention to reconstruct knowledge relations under the current learner state and a memory bank to preserve historical prototypes for cross-stage alignment. Student-level five-fold cross-validation on six public datasets shows that CTRA-Net outperforms sequential, graph-based, spatiotemporal, memory-based, and semantic-enhanced baselines. Drift, cold-start, and efficiency analyses show better stability with deployable cost.
Phishing attacks remain one of the primary cybersecurity threats because attackers exploit users through deceptive URLs focusing primarily on online banking. The research undertakes a comparative assessment between deep learning and machine learning detection algorithms applied to find phishing URLs focused on banking clients. The research implemented Convolution Neural Network (CNN), Long Short-Term Memory (LSTM) and alongside Logistic Regression (LR) and Random Forest (RF) and Support Vector Machine (SVM) and XGBoost as independent training models for evaluation. The CNN model proved its superiority in identifying complicated patterns of malicious URLs because it reached 99% accuracy during testing and training. Public performance standards enabled the assessment of robust model behavior through an extensive evaluation process. Upcoming research will investigate instant threat identification techniques as well as extra feature integration strategies to boost system performance for protective measures in changing security threats.
Trajectory-user linking (TUL) plays a vital role in multisource geospatial data analytics for behavioral pattern recognition and user identification. Addressing the limitations of conventional TUL approaches in computational efficiency and predictive accuracy, this study proposes a novel prediction model for TUL based on graph attention network (PMTULGAN) that harnesses graph attention networks to significantly enhance predictive performance of TUL tasks. PMTULGAN's key innovation lies in its dynamic attention mechanism, which adaptively allocates weights to nodes to facilitate more accurate extraction and interpretation of salient features within complex trajectory data. Extensive experimental evaluations reveal a substantial performance enhancement for the proposed method over conventional methods. These empirical results underscore the robustness and reliability of PMTULGAN in various data scenarios and substantiate its practical utility in real-world applications.
To address adaptability limitations in conventional model predictive control, this paper proposes a data-driven dual-channel (D-Channel) adaptive framework for wheeled mobile robots. By extending the traditional single-channel architecture into a parallel dual-network configuration and integrating an offline reinforcement learning channel, the proposed method improves control precision and robustness. The D-Channel structure refines predictive outputs, while the reinforcement learning module enhances adaptability to dynamic disturbances. Comprehensive simulations and hardware-in-the-loop experiments show that the D-Channel radial basis function neural networks-model predictive control outperforms single-channel counterparts. The results demonstrate improved tracking accuracy, faster convergence, and reduced steady-state error, confirming the effectiveness of combining data-driven learning with predictive optimization for complex control tasks.
This paper proposes a Weak Information Graph Neural Network (WIGNN) for weak information modeling. Firstly, WIGNN introduces a multi-scale feature propagation mechanism that iteratively aggregates both local and global information, thereby enhancing node representations. Secondly, the authors design a pseudo label optimization and semantic alignment module that leverages limited labeled data alongside generated pseudo labels to construct class prototypes, reinforcing semantic consistency of nodes across different views. Finally, WIGNN incorporates a supervised contrastive learning module that aggregates representations for the similar node in the original and augmented graphs while pushing apart representations of nodes from different classes, which effectively mitigates weak feature issues. Extensive experiments on several public benchmark datasets demonstrate that WIGNN significantly outperforms leading baselines under extreme conditions of weak structure, weak features, and weak labels, exhibiting superior generalization and robustness.
As ideological and political education strengthens in institutions of higher education, there is a growing need for automated methods that identify emotions and ideological stances in student comments. Existing approaches have focused on sentiment polarity and have struggled with multidimensional emotions, sentiment–ideology interactions, and indirect or metaphorical expressions, resulting in limited granularity and interpretability. To address these challenges, this study proposes an Aspect-level Coupled Reflective Network (ACR-Net) for the joint modeling of sentiment and ideology. ACR-Net extracts aspect semantics using large language models, builds fused representations through a module coupling ideology with emotion, and incorporates a sentiment weighted network and reflective calibration mechanism to capture local cues and handle implicit expressions. Experiments on real ideological and political student comments demonstrated that ACR-Net surpasses existing methods in accuracy, interpretability, and generalization. Our code is available at https://github.com/jialekang/Aspect-level-Coupled-Reflective-Network.git.
Affective computing aims to enable machines to recognize and simulate human emotions, forming a critical component of intuitive human-computer interaction. This study presents a systematic bibliometric review of 225 high-impact publications (2014–2024) from the Web of Science Core Collection. Utilizing tools like Bibliometrix and CiteSpace, this analysis maps the field’s evolution, identifying a paradigm shift from basic emotion recognition to deep learning-based multi-modal fusion, with generative models and large language models for affective synthesis emerging as a new frontier. Persistent challenges include integrating multi-modal context for personalization, fulfilling real-time processing requirements, and addressing ethical issues like bias. To bridge the gap between emotion recognition and the development of genuinely adaptive, context-aware systems, the study highlights the urgent need for generative affective frameworks and neuroscience-informed lightweight models. This review synthesizes the developmental trajectory of affective computing in human-computer interaction to guide future research.
Accurate demand forecasting enables efficient supply chain management. However, two major challenges persist: (a) insufficient modeling of external conditional variables and poor capture of complex multimodal demand distributions and (b) conditional information that is fused only once at the system input, causing information decay and reduced responsiveness to event-driven shocks. To this end, the authors propose a dual-phase conditional diffusion model (DP-CDM) in which a reverse sliding diffusion along the temporal axis exploits temporal continuity to build an autoregressive mechanism, enhancing sequence modeling and avoiding structural misalignment. In addition, noise-degradation diffusion enriches multimodal probabilistic representations and improves robustness to external disturbances. A conditional embedding module aligns multimodal features by aggregating local histories, global trends, and SHapley Additive exPlanations (SHAP)-quantified external factors, which are injected throughout the denoising processes. A 3.7% improvement in fitting performance showed the effectiveness of this model in capturing event-driven demand dynamics.
Big data settings are currently evolving from classical systems that focus on supporting advanced decision-support processes-as applied to many real-life scenarios, which are typically populated by distributed and heterogeneous data sources, such as conventional distributed data warehousing environments-to cooperative information systems. Different data formats contribute to define challenging big data systems, in which the main issue consists in supporting modern big data analytics involving massive amounts of data. As a consequence, a relevant research challenge is how to efficiently integrate, process, and mine such distributed knowledge, which composes the foundations of final big data analytics processes. Starting from these considerations, in this paper the authors propose an online analytical mining-based framework for supporting big data analytics, along with a formal model underlying this framework, called Multi-Resolution Ensemble-Based Model for Advanced Knowledge Discovery in Big Data Warehouses.
Big data settings are currently evolving from classical systems that focus on supporting advanced decision-support processes—as applied to many real-life scenarios, which are typically populated by distributed and heterogeneous data sources, such as conventional distributed data warehousing environments—to cooperative information systems. Different data formats contribute to define challenging big data systems, in which the main issue consists in supporting modern big data analytics involving massive amounts of data. As a consequence, a relevant research challenge is how to efficiently integrate, process, and mine such distributed knowledge, which composes the foundations of final big data analytics processes. Starting from these considerations, in this paper the authors propose an online analytical mining-based framework for supporting big data analytics, along with a formal model underlying this framework, called Multi-Resolution Ensemble-Based Model for Advanced Knowledge Discovery in Big Data Warehouses.
This paper introduces interactive-aware multi-objective style transfer network, an innovative framework designed to enhance digital artistic workflows by balancing computational efficiency, creative autonomy, and ethical transparency. By integrating a dual-path network for content preservation and style evolution, meta-learning for rapid style adaptation, and a hybrid evaluation system, interactive-aware multi-objective style transfer network achieves 85.7% style retention across diverse domains while reducing convergence iterations by 19.2%. The framework also employs gradient-weighted class activation mapping to align artificial intelligence, decisions with designer intent, achieving 78% congruence. These advancements address key limitations in opacity, latency, and domain generalization, providing a robust solution for intelligent creative tools. This work is significant for academic researchers and information technology professionals focused on advanced data processing and human-centered design.
Phishing through mobiles is becoming advanced, attacking the users through malware applications, SMS, and social media. Dynamic threats better the conventional detection techniques, thereby hybrid approaches integrating machine learning, deep learning, and heuristic rules are the essentials. Here the work is on mobile security utilizing AI in interaction with 5G and edge computing for detection in real time. This survey discusses ensemble learning, federated learning, blockchain, and privacy-preserving techniques for defending against adversarial attacks and limited resources. It discusses elastic defences for mobiles and explores Explainable AI and quantum machine learning for enhanced performance and explainability. The results are from peer-reviewed journals and sources (2018-2024) like IEEE, Springer, and ScienceDirect, showing a modern overview of hybrid phishing detection.
To improve photovoltaic (PV) power forecasting under variable meteorological conditions, this paper proposes a hybrid model combining signal decomposition, clustering, and deep learning. An improved complete ensemble empirical mode decomposition with adaptive noise method is used for multi-scale decomposition of meteorological inputs such as temperature, solar radiation, and wind direction. Sample entropy-guided K-means clustering segments signals into high, medium, and low-frequency components, with high-frequency parts further denoised using variational mode decomposition. A convolutional neural network-bidirectional long short-term memory network is then optimized by the crown porcupine optimization algorithm to fine-tune key hyperparameters. Experiments on real PV data show a 20% root mean squared error reduction (to 7.30 kW), demonstrating strong adaptability and robustness for intelligent PV scheduling.
This article introduces an innovative text summarization mechanism designed to tackle the inherent challenges of condensing lengthy and unstructured legal documents in the context of India. The authors' primary aim is to create a system proficient in extracting crucial information from these documents, producing concise summaries akin to those crafted by humans. The proposed methodology frames summarization as a binary classification problem, employing an extractive summarization technique rooted in statistical features and word vectors. The system strategically identifies summary statements from the comprehensive input text section. To automate the summarization process, they leverage various classifiers, including logistic regression, gradient boosting, and neural networks. Through this multifaceted approach, they endeavor to enhance the efficiency and accuracy of legal document summarization, addressing a critical need in the field.
Contrastive learning enables models to learn graph structural information through self-supervised learning in the absence of labels. However, real-world networks often contain both graph structural information and incomplete node attribute information. Based on this, this paper proposes a heterogeneous network text attribute fusion method based on multi-layer semantic relation contrastive learning. Firstly, the heterogeneous network components are reconstructed using semantic and thematic attribute acquisition methods at different levels, obtaining semantic representations of text attributes at various levels of abstraction. Then, the contrastive learning component of the heterogeneous network is employed to maximize the correlation between different views of the heterogeneous network, allowing the two heterogeneous networks to align in this space. This alignment helps to uncover the latent connections between text attribute features across different views, thereby achieving the fusion of information between views.
The rise of online education demands improved learning assessment and personalization. Current knowledge tracing methods struggle with feature extraction, limited information interaction within learning data, and insufficient utilization of structured relationships between knowledge points. To address these challenges, this article proposes a knowledge graph-enhanced interleaved multi-head attention knowledge tracing model. The model integrates bidirectional long short-term memory networks, an interleaved multi-head attention mechanism, and graph convolutional networks into a deep learning framework. The interleaved multi-head attention mechanism enhances the model's ability to capture long-distance dependencies, while the knowledge graph encoding module utilizes graph convolutional networks to mine structured relationships between knowledge points. This architecture considers both the dynamic learning process and integrates structured information from the knowledge system. Experiments on multiple public datasets validate the model's effectiveness.
The widespread use of electromagnetic space utilization technology across various fields—including maritime, terrestrial, aeronautical, orbital, electrical, and telecommunications—has generated vast amounts of electromagnetic environment data. To manage challenges, such as the storage of large raw datasets, data discrepancies, isolated data across multiple pathways, and low data value density, a big data-based electromagnetic environment data warehouse is proposed. This warehouse standardizes data from diverse sources, integrates and reconstructs it according to business themes, and uses a mix of relational and non-relational databases for storage. It meets the needs for high data reliability, fast access, and massive storage capacity, offering a solution to data overload while supporting data mining and knowledge discovery in the electromagnetic field.
This paper uses advanced analytics to explore emerging trends in sustainable human resource management (HRM). Analyzing publications from 1991 to 2023 via the Scopus database identifies key trends and contributions in sustainable HRM research, including sustainability, green HRM, and corporate social responsibility. The findings highlight an increasing research volume, significant global contributions, and a growing academic interest in integrating sustainability into HR practices. Despite its reliance on the Scopus database, this review provides valuable insights for HR practitioners and organizations aiming to enhance sustainability through HR strategies. It also maps future research directions in sustainable HRM.
