In the era of information overload, various types of information interconnect to form complex networks. To better manage diffusion paths within networks, we propose predicting information transmissibility—the probability of information being transmitted under the influence of other information in the network. Accurate transmissibility prediction has practical applications in recommendation systems and misinformation control, enabling relevant information to reach appropriate audiences while curbing the spread of less useful content. Given the characteristics of information networks, text-attributed graphs provide a natural representation that captures both network structure and content semantics. However, existing text-attributed graph representation methods fail to capture diffusion dynamics and incur high computational costs. Therefore, we propose a novel efficient textual-graph model, Language Temporal Variation Graph Network(LTVGN), to predict transmissibility by capturing time-varying features, structural information and textual information. Our proposed model is evaluated on the citation dataset HEP-TH. The results demonstrate that our model outperforms state-of-the-art models, achieving a low estimation error.

Wearable brain-computer interfaces (BCIs) have made it feasible to monitor brain activity for emotion recognition in real-world settings. While deep learning models achieve high classification accuracy on electroencephalography (EEG) data, they often lack interpretability, limiting their neuroscientific relevance. In this study, we present an interpretable framework for EEG-based emotion analysis rooted in energy landscape analysis. EEG signals from the DEAP dataset were standardized and binarized prior to quantification of neural state transitions. We found significant subject-specific correlations between the number of state transitions and emotional ratings of valence and arousal. Further analysis revealed that certain binary brain states, particularly complementary pairs, were among the most frequently observed and showed emotion-dependent frequency differences. Transitions between these state pairs varied across subjects, suggesting their role as local minima in the brain’s dynamic landscape. Our findings demonstrate that energy landscape analysis provides an interpretable alternative to black-box models, offering insights into how brain dynamics relate to emotional experiences. This approach contributes toward building explainable affective computing systems and supports the use of neural state modeling in emotion-aware BCIs.

Non-Independent and Identically Distributed (Non-IID) data pose a fundamental challenge in Federated Learning (FL). It usually causes a severe client drift issue (various client model update directions) and thus, degrades the global model performance. Existing methods typically address this by assigning appropriate weights to client models or optimizing model update directions. However, these methods overlook client model update trends. They focus solely on the final client models to be aggregated at the server at each communication round, ignoring model optimization trajectories, which may contain richer information to aid model convergence. To address this issue, we propose FedA ⁴, a novel FL framework with Anti-bias Aggregation and trAjectory-based Adaptation, which leverages clients’ optimization trajectories, rather than only their final model snapshots. For anti-bias aggregation, by observing a phenomenon termed model collapse, where biased clients tend to predict any input data as the dominant classes in their own datasets, we quantify the class dominance and analyze the level of client drift for each client. We evaluate a prediction entropy, namely concentration, so as to assign an optimal weight to each client at each training round. To further mitigate the negative effect of clients with high levels of client drift (biased clients), we then develop a gradient adaptation mechanism termed trajectory-based adaptation, which analyzes clients’ trajectories to correct each client’s contribution to the aggregated global model. Extensive experiments on CIFAR-10, CIFAR-100, STL-10, and Fashion-MNIST demonstrate that FedA ⁴ significantly outperforms state-of-the-art baselines, particularly in scenarios with extreme data heterogeneity (high level of Non-IID).

In this survey, we present a focused analysis of recent advances in Artificial Intelligence (AI) for music education, structured around four key pedagogical domains: learning and practicing, assessment, creation, and teaching. First, for learning and practicing, we examine AI-driven tools for personalized skill acquisition, including intelligent instruments and adaptive systems that provide real-time feedback. Second, for assessment, we review progress in the automated assessment of musical performance, moving beyond pitch and rhythm to more nuanced expressive qualities. Third, for creation, we investigate innovations in AI-augmented composition, where generative models act as collaborative partners and creative catalysts for students. Finally, for teaching, we explore systems for teacher empowerment, such as automated resource generation and learning analytics dashboards. The analysis highlights the transformative role of Deep Learning (DL) and Generative AI (GAI) in each domain while critically discussing persistent technical limitations and emerging ethical concerns. By synthesizing interdisciplinary developments, this survey aims to chart the current frontier and inform future research at the intersection of AI technology and music pedagogy.

Transcranial direct current stimulation (tDCS) has emerged as a versatile non-invasive neuromodulation approach that can alter cortical excitability and affect network plasticity. Recent advances in machine learning (ML) offer an opportunity to transform tDCS from largely heuristic practice into a quantitatively informed, adaptive intervention paradigm. Here, we synthesize developments from 2020 to 2025 at the intersection of tDCS and ML. Search results from structured PubMed and Google Scholar queries were screened for eligibility based on predefined inclusion criteria, retaining peer-reviewed studies that applied ML techniques to tDCS related studies. Eligible studies were evaluated for data integrity, and ML model validation methodology. Sixteen studies met inclusion criteria. Across these studies, ML was applied to heterogeneous datasets, including electroencephalography, neuroimaging, and clinico–demographic features, to predict stimulation outcomes, characterize neural responses, and identify biomarkers of tDCS sensitivity. Support vector machines and random forests remain prevalent, reflecting the modest scale and exploratory nature of current datasets; most studies rely on early-stage clinical or preclinical cohorts, resulting in promising yet fragmented evidence. Nevertheless, emerging results illustrate how ML can reveal latent physiological structure, guide dose–response optimization, and support the translation of tDCS toward precision neuromodulation. Drawing on this integrated analysis, we highlight key directions for the field: multimodal integration that unifies electrophysiological, structural, and behavioral signatures; incorporation of biophysically grounded forward models and pretrained deep-learning architectures; and development of adaptive, closed-loop control strategies capable of personalizing stimulation in real time. Together, these advances chart a pathway toward ML-guided tDCS systems that are mechanistically informed, clinically actionable, and scalable for widespread application.

The T-cell receptor (TCR) is a fundamental component of the adaptive immune system, playing a crucial role in the development and progression of autoimmune diseases through its remarkable diversity and antigen specificity. Advances in high-throughput sequencing technologies and multi-omics data integration have revolutionized the ability to characterize TCR repertoires at unprecedented resolution. Coupled with emerging machine learning methodologies, these advances have opened new avenues for unraveling the complex immunopathology underlying autoimmune disorders. This review comprehensively summarizes current knowledge on the dynamic regulation of TCR repertoires in autoimmune diseases, highlighting key processes such as central tolerance failure, clonal expansion of autoreactive T cells, and regulatory T cell dysfunction, as well as the influences of genetic predisposition and immunosenescence on shaping TCR diversity. This review also provides a 3 that demonstrates how to analyze publicly available TCR repertoire datasets. We compare V and J gene usage profiles and CDR3 summary features across clinical labels to characterize between-group variation and to inform feature engineering for downstream machine learning models. Furthermore, we detail various machine learning-based diagnostic models that utilize gene usage patterns and CDR3 sequence features to accurately classify autoimmune disease status, alongside recent breakthroughs in predicting TCR-epitope binding specificity. These computational approaches not only enhance diagnostic precision but also provide mechanistic insights into immune recognition and autoreactivity. By integrating immunological principles with data-driven techniques, this work aims to offer a robust theoretical framework and practical guidance for future research in immunology and precision medicine. Ultimately, the convergence of TCR repertoire profiling and machine learning promises to drive innovative strategies for early diagnosis, personalized therapy, and improved clinical management of autoimmune diseases, enabling the transition to antigen-specific tolerogenic therapies.

Reinforcement learning (RL) has shown promise in optimizing treatment strategies for sepsis, a life-threatening condition responsible for significant mortality in intensive care units. However, deploying RL policies in clinical settings requires not only optimizing patient outcomes but also ensuring adherence to established medical guidelines. In this paper, we propose a two-stage safety framework for offline RL-based sepsis treatment. The first stage employs Constraint-Penalized Q-learning combined with Implicit Q-Learning (CPQ-IQL), which incorporates clinical constraints through Lagrangian optimization during policy learning. The second stage applies a runtime safety filter that dynamically validates actions against clinical guidelines before execution. We evaluate our framework on the ICU-Sepsis benchmark with four clinically-motivated constraints derived from the Surviving Sepsis Campaign 2021 guidelines. Experimental results over 5 random seeds demonstrate that CPQ-IQL achieves the lowest constraint violation rate (22.88 ± 0.94%) among all baselines while maintaining competitive survival rates (78.4 ± 1.8%). When combined with the Safe Actions filtering mechanism, constraint violations are reduced by 97.2% (from 22.88% to 0.41%), demonstrating the effectiveness of our two-stage safety framework. Our analysis reveals that the Safe Actions filter modifies approximately 21% of policy decisions, highlighting the importance of runtime safety mechanisms for clinical deployment. These findings suggest that combining constraint-aware offline learning with runtime safety filtering provides a practical pathway toward safe and effective RL-based clinical decision support systems.

Federated continual learning (FCL) is a distributed training framework that allows for learning from sequences of tasks on different centers under privacy-preserving. Although FCL has been extensively studied in fields such as image recognition and image segmentation, it remains unexplored in multi-center corneal diseases diagnosis, where data is inherently distributed and asynchronous while data privacy, reliability, and interpretability are urgently required. Therefore, this paper proposes Powderless for multi-center corneal diseases diagnosis, which can effectively transfer corneal diseases knowledge through prompt aggregation and prompt selection across various sequentially learned tasks from different centers under privacy. To further enhance diagnosis performance, ensure detection reliability, and improve interpretability, we design three key components: a multi-modal ensemble mechanism, an energy-based uncertainty estimation module, and a decision explanation module grounded in causal intervention. Comprehensive experimental results on the keratitis dataset demonstrate that our method achieves significant improvements compared to the base model, single-modality version, and local training in terms of both accuracy and the alignment between accuracy and uncertainty.

With the rapid proliferation of multimodal social media platforms, fake news has been increasingly disseminated through multiple modalities such as text, images, and videos, posing serious threats to social stability, public cognition, and these platforms’ ecosystem. Existing unimodal fake news detection methods face much challenge in multimodal scenarios, as they can not capture fully cross-modal semantic correlations and inconsistencies. Multimodal fake news detection, which integrates heterogeneous information from text, visual, and audio modalities to explore inter-modal consistency and complementarity, has therefore become a major research focus in recent years. A comprehensive survey of recent advances in multimodal fake news detection is presented in the paper, which consists of systematically reviews of the fundamental concepts, detection tasks, and underlying technical principles in this field. Major benchmark datasets and commonly used evaluation metrics are also introduced, followed by a structured taxonomy of representative detection methods and a summary of their experimental results. Furthermore, the key challenges faced by current research are discussed and promising future research directions are outlined. Compared with existing surveys, this work presents a more comprehensive method categorization that emphasizes the evolution of detection techniques, offers clearer comparisons of datasets and experimental analyses, and provides more practical insights for researchers and practitioners in multimodal fake news detection.

This paper reviews recent advancements in autonomous driving safety, focusing on the evolution of autonomous driving systems from modular pipelines to end-to-end (E2E) frameworks and emerging vision-language-action (VLA) models. For modular systems, this paper analyzes how to mitigate error propagation between decoupled modules using multi-sensor redundancy and formal verification. For end- to-end systems, this paper delves into learning-based motion planning. It emphasizes safety innovations to address the lack of transparency in deep learning, such as interpretable cost maps and world-model-based simulations. For VLA models, this paper investigates integrating vision language models (VLMs) to enhance high-level semantic reasoning and understanding of long-tail driving scenarios. It discusses safety guardrail technologies, such as chain of thought (CoT) reasoning, to ensure that logic aligns with driving regulations. Finally, this paper summarizes current challenges and outlines future research directions, providing a systematic reference for building safe and reliable autonomous driving systems.