This study presents an explainable stock-forecasting framework that combines a Transformer with gradient-weighted class activation mapping (Grad-CAM) and compares autoregressive-only with selected multivariate inputs. Global-local aggregation (GLA) aggregates min–max-normalized Grad-CAM maps from multiple Transformer layers. Learned layer-weight GLA uses unequal weights selected on validation data, whereas Uniform GLA assigns equal weights. The contribution therefore lies in the aggregation and validation framework rather than in arithmetic averaging as a new operator. We also compare GLA with Shapley additive explanations (SHAPs) and integrated gradients (IGs), finding agreement on highly attributed features at substantially lower computational cost. On the small Jakarta Islamic Index (JII) dataset, autoregressive information dominates short-horizon forecasting. However, a selected three-input model reduces stable-window mean absolute percentage error (MAPE) from $\mathbf{0.745\%}$ to $\mathbf{0.610\%}$. Selected multivariate inputs also reduce NIFTY50 MAPE from $\mathbf{4.202\%}$ to $\mathbf{3.761\%}$, while remaining statistically comparable to univariate input on S&P500. Ten-seed paired $\boldsymbol{t}$-tests and exact Wilcoxon tests support the stable-window JII and NIFTY50 gains. Learned layer-weight GLA improves full-ranking perturbation faithfulness on S&P500 and NIFTY50 while remaining statistically tied with Uniform GLA on JII. These results show that both multivariate inputs and unequal layer weighting provide scenario-dependent benefits. The framework therefore offers competitive forecasts and transparent model-attribution diagnostics rather than causal explanations.
In this article, we experimentally and theoretically analyze practical consensus stimulated by a small number of committed players who stubbornly promote an alternative in a binary-choice quiz with graded confidence. Our study bridges the gap between theory and experiment on practical consensus, since prior work on practical consensus is limited to theoretical analyses. In practical consensus, socially chosen opinions converge to a bounded interval, while admitting some deviations from exact consensus. We formulate a novel heterogeneous population dynamics model that represents the diversity of boundedly rational decision-makers to capture practical consensus in an on-site multiround group experiment. From a theoretical analysis, we prove that practical consensus is shaped by the heterogeneity of boundedly rational decision-makers. Our findings contribute to a better understanding of practical consensus and may provide insights into the spread of new alternatives in binary-choice settings with graded confidence.
Link prediction is a foundational task in social network analysis that focuses on predicting potential or absent links between entities based on the existing network structure. Despite numerous algorithmic advances, the impact of graph density and nonedge (negative sample) selection strategies on the predictive performance of machine learning (ML) models remains underexplored. This study presents a systematic empirical investigation into these factors by constructing past graphs of varying densities; sparse, moderate, and dense, from present graphs through strategic edge removal. Negative samples are generated using three methods: Random sampling, small shortest path length (Small-SPL), and large shortest path length (Large-SPL). Ten similarity-based features are extracted, and seven ML models are evaluated using Facebook 1K (FB1K) and 4K datasets. Results demonstrate that both graph density and nonedge selection critically influence prediction outcomes. The decision tree (DT) classifier achieved the highest accuracy of 99.92% for dense graphs using the Large-SPL nonedge selection strategy. It also consistently outperformed other models across moderate and sparse graphs, achieving accuracies of 98.56% and 87.41%, respectively. These findings demonstrate the effectiveness of dense historical graphs and appropriate nonedge sampling strategies for improving link prediction accuracy. Large-SPL selection improves class separability and reduces bias, while sparse graphs lead to lower recall due to limited structural information. Ensemble models and DTs consistently perform well across conditions, whereas Naïve Bayes (NB) and SVM are more sensitive to sparsity and feature overlap. The findings highlight the importance of thoughtful dataset construction and validate the scalability of the proposed framework for robust link forecast.
Takeover requests (TORs) in autonomous vehicles (AVs) can impose cognitive load on drivers and affect the quality of control transfer. Detecting this cognitive load is difficult in real-world settings because synchronized multimodal data are costly and risky to collect. We developed a virtual reality (VR) driving platform for controlled TOR experiments and multimodal data collection. Based on the collected driving behavior, eye-tracking, and physiological signals, we propose momentum prototype network with attentive embedding aggregation for multimodal cognitive load detection network (MPA-MCDNet), a cognitive-load detection model that combines multimodal fusion with prototype-based episodic learning. The fusion module integrates heterogeneous time-series features, and the prototype-based module is used for the limited-trial setting in this study. Experiments on the collected dataset show that the platform supports TOR-related cognitive-load studies and that MPA-MCDNet performs well on five-class classification in our evaluation.
The development of financial technology has made significant breakthroughs with the rapid advancement of artificial intelligence (AI) and deep learning models. Fingerprint recognition has long been a commonly used biometric technology and can also be applied to identity verification in the financial sector. However, despite the progress of AI, many security issues still remain. Therefore, it is necessary to ensure the security of financial technology systems throughout their development. This study combines the concept of multiobjective optimization and employs the Nondominated Sorting Genetic Algorithm IV (NSGA-IV) to compute the perturbations required for generating adversarial attacks. Experimental results demonstrate that the proposed method can effectively attack deep learning models, and under the current experimental settings, the attack success rate is highly promising. In addition, the average structural similarity index (SSIM) between the generated samples and the original samples reaches approximately 0.9999. Furthermore, this study conducts ablation experiments on the objective functions to analyze their respective contributions and effectiveness. The experimental results of this method also indicate that current financial technology systems still face many threats from malicious attacks, and thus require effective defense mechanisms to ensure their security and robustness.
Disinformation propagation on online platforms poses significant threats to public trust and societal stability, underscoring the urgency of effective governance strategies. Although traditional rule-based modeling has provided valuable insights, it remains difficult to capture complex human behavior and cognition using rigid behavioral assumptions. To overcome these limitations, we introduce a multiagent disinformation propagation model driven by a large language model (LLM), referred to as LLM-driven multiagent disinformation propagation (LMDP). The LLM equips agents with reasoning capabilities and enables decisions based on information content. With the proposed model, this study examines disinformation dynamics shaped by cognitive heterogeneity, network topology, message framing, official interventions, and selection of official agents. Extensive simulation experiments are conducted, and the results show that disinformation is more likely to spread when radical agents are more prevalent or when messages are framed with greater elaboration. Conversely, effective suppression relies on timely and targeted interventions, together with the strategic deployment of official agents at vital network positions.
Recommender systems are vulnerable to poisoning attacks due to their open feedback nature. Existing black-box poisoning attacks typically rely on surrogate models to generate fake users, implicitly assuming that all users contribute equally during model training. However, recommender models exhibit nonuniform user-level fitting behavior, where users with higher fitting consistency are more closely aligned with the model’s mainstream preference representations and exert greater influence on parameter updates. Exploiting such users can yield stronger attack signals, but relying solely on high-consistency signals learned from a surrogate model risks overfitting and degrades cross-model transferability. To address the tradeoff between attack effectiveness and transferability, we propose a loss-aware dynamic weighting attack framework (LAW) with a two-stage design. LAW first leverages users with relatively high consistency to obtain stable attack directions and then progressively incorporates broader user signals to expand attack coverage and improve generalization. Extensive experiments across multiple datasets and recommender models demonstrate that LAW achieves more stable cross-model transferability while consistently outperforming existing attack methods.
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder for which early risk identification is essential for timely clinical assessment, care planning, and long-term disease management. However, computational modeling of AD remains challenging because clinically meaningful diagnosis cannot be inferred from numerical separability alone; it requires alignment with clinical assessment and biomarker validation. From a computational social systems perspective, a major challenge is the heterogeneity and fragmentation of patient-related information, including wearable actigraphy signals, clinical records, and demographic contexts. To address this challenge, we propose AlzDetect, an enhanced TabNet-based deep learning (DL) framework for screening-oriented AD risk modeling across clinically grounded and wearable-derived data domains. AlzDetect is not presented as a subject-level multimodal diagnostic system; rather, it is evaluated as a heterogeneous data risk stratification framework under domain shift, missingness, and cross-domain robustness settings. The wearable-derived (Google Pixel Watch 2) dataset is used only for feasibility and robustness analysis of behavioral signal modeling, not as biomarker-confirmed evidence for AD diagnosis or clinical staging. AlzDetect provides transparent decision support through TabNet’s sparse attention masks, which identify influential predictors at feature and modality levels. The extensive evaluation shows that AlzDetect achieves 96% predictive accuracy and outperforms representative baselines within the evaluated dataset specific settings, demonstrating strong predictive separability and robustness across heterogeneous AD relevant data sources.
In social networks, encrypted communication has become increasingly prevalent to protect user privacy. However, this also provides attackers with opportunities to launch covert attacks via encrypted traffic. Traditional traffic detection methods relying on rules and payload analysis fail due to content encryption. As a result, encrypted traffic detection methods based on statistical features and sequence features have gradually emerged as research hotspots. Nevertheless, statistical feature-based methods require access to all packets in a flow, making real-time detection challenging. Therefore, this article exploits sequential features and combines the attention mechanism with sequence prediction to model contextual dependencies among packets. This design allows the model to capture effective flow representations using only a small number of packets, thus helping the classifier achieve accurate classification. In addition, an attention optimization mechanism is introduced to improve the reliability of attention vectors, which contributes to the interpretability of model decisions. Experimental results show that the proposed method achieves competitive classification performance on five datasets using no more than 20 packets, which reduces latency for real-time threat detection. Notably, when using only the first five packets, the average waiting time is only 0.12 s, accounting for merely 0.05% to 3.58% of that required by comparison methods.
Graph neural networks (GNNs) have been widely adopted in collaborative filtering to model higher order relationships between users and items. However, existing methods mainly focus on interaction-based structures and neglect the explicit modeling of semantic relations among users and among items. Within these graph structures, deep graph message propagation tends to cause over-smoothing, where node embeddings converge to similar representations, hindering effective personalization. To address these challenges, we propose a semantic-augmented multiview graph-based recommendation framework (SAGRec). Specifically, SAGRec constructs semantic similarity graphs based on embeddings generated by large language models to capture semantic relations between users and between items. Next, we propose a frequency-aware propagation mechanism. It leverages representation differences between nodes and incorporates gated message passing to retain node distinctions during propagation. To further mitigate representation drift induced by multilayer propagation, we design an adaptive inter-layer contrastive learning strategy that enforces consistency while maintaining discriminability across layers. Comprehensive empirical evaluations conducted on three real-world datasets show that SAGRec achieves consistently superior performance over baseline methods with respect to normalized discounted cumulative gain and recall.
As artificial intelligence (AI) technologies become increasingly integrated into critical social systems ranging from healthcare diagnostics to smart city infrastructure, their security and computational efficiency are paramount. Among emerging threats, sponge poisoning attacks exploit deep learning models to silently increase training energy and latency without impacting accuracy, posing serious availability risks in real-world deployments. This article introduces MetaSpongeDetect, a meta-learning-based detection framework designed for online identification of such stealthy adversarial behaviors. The approach leverages a recurrent neural network (RNN) within a model-agnostic meta-learning (MAML) paradigm to capture temporal patterns in the training of model dynamics. By learning from a distribution of tasks constructed from model statistics across multiple architectures and datasets, MetaSpongeDetect generalizes effectively to previously unseen attack scenarios. Experimental results demonstrate that the proposed method achieves superior performance in few-shot settings, significantly outperforming baseline anomaly detection techniques. The framework offers a robust and scalable solution for online detection of training-time adversarial behaviors. The proposed approach enhances AI system resilience and safety, making it suitable for deployment in computational social systems, where robustness, energy efficiency, and public trust are critical. MetaSpongeDetect achieves 95%–99% detection accuracy across diverse models and datasets, demonstrating robust generalization to unseen sponge poisoning attacks. This scalable, adaptive framework enables online detection of training-time adversarial behaviors, enhancing the resilience and reliability of AI systems in computational social infrastructures.
Video domain adaptation (VDA) learns domain-adaptive and transferable knowledge across different videos, which improves the transferability and robustness of the corresponding model to different video tasks. Since the high-dimensional videos are generally embedded in nonlinear manifolds with low dimension, VDA on the Grassmann manifold has attracted widespread attention. However, existing Grassmann representation-based VDA methods ignores the implicit temporal information in representing videos as points on Grassmann manifold. To this end, we propose a novel VDA method with Grassmann manifold trajectory representation (VDA-GMTR), which extracts trajectory feature from video to preserve the temporal and spatial information. Further, to sufficiently alleviate domain discrepancy, the proposed VDA-GMTR aligns the features through collaborative video-level and clip-level alignments (CAs), capturing both coarse-grained and fine-grained information. The CA is embedded in trajectory representation learning, while the video-level alignment (VA) integrates the pseudolabel information to achieve semantic interaction and alignment. The extensive experimental results conducted on various tasks demonstrate that the developed VDA-GMTR achieves the superior classification performance and strong robustness. The scalability of VDA-GMTR for multimodal video task is further discussed.
Spatial public goods games involve localized interactions and high-dimensional state spaces, making stable cooperation difficult. Traditional approaches often struggle to effectively capture neighborhood-level strategic interactions and dynamically align individual incentives with collective welfare. To resolve this issue, this article introduces a novel intelligent decision-making framework called local mean field proximal policy optimization with neighbor-dependent cooperative density punishment (LMFPPO-NCDP). The conventional mean field concept is reformulated as a localized statistical representation embedded into policy learning, enabling agents to adapt to neighborhood level cooperation dynamics. Additionally, a NCDP mechanism penalizes defectors according to local cooperation density without direct costs to cooperators. Experimental results demonstrate that LMFPPO-NCDP promotes rapid and stable cooperation under low enhancement factors. Compared with proximal policy optimization (PPO), Q-learning, and the Fermi update rule, the proposed framework achieves lower cooperation thresholds and more robust cooperative outcomes. Statistical analyses further confirm its robustness and effectiveness across different enhancement factors. The proposed framework provides a computational framework for analyzing adaptive cooperation in spatially structured multiagent systems.
Voting is a widely accepted method for decision-making and for assigning positions to members of committees. For instance, each year, most IEEE Societies or Councils conduct elections to renew its Executive Committee (ExCom) or Administrative Committee (AdCom) in accordance with the Society’s or council’s bylaws. However, traditional voting methods, though easy to follow and broadly accepted, may not always yield the most effective outcomes. This article proposes a new, computation-based approach to committee member selection, grounded in the environments-classes, agents, roles, groups, and objects (E-CARGO) model and the role-based collaboration (RBC) methodology. Developed and refined over the past two decades, E-CARGO/RBC provides a solid foundation for modeling and solving complex collaborative and decision-making problems, particularly those involving team-based dynamics. Through comparative analysis and a case study, this work demonstrates how the proposed computational approaches surpass the traditional way in the selection of ExCom/AdCom members, and achieve more effective position assignments than traditional voting-based approaches.
Emotional distress and declining psychological well-being are increasingly prevalent among student populations in India, emerging as a significant concern relative to other societal segments. In the context of computational social systems, especially within human–device interaction (HDI) and social media–mediated communication, addressing these mental health (MH) challenges demands a nuanced understanding of affective states through both verbal expression and subtle nonverbal behavior. This creates technical and methodological obstacles for nonintrusive HDI systems, including the need to simulate naturalistic interaction settings, capture various social signals, decode authentic emotional responses, and ensure high fidelity in emotion recognition without relying on obtrusive sensing technologies. This study aimed to assess the emotional well-being of young adults in premier Indian institutions through various technological interventions during machine interactions. However, the lack of standardized noncontact databases has hindered the development and evaluation of emotional health assessment tools. To address this gap, we developed the Contactless Affect Recognition Database (CARD), built from data gathered across 53 participants, which is designed to support the advancement of robust and scalable models for emotion and mental disorder detection, particularly in online and social media–enabled environments such as video conferencing, live streaming, and virtual classrooms. The database includes RGB and infrared facial images, peak facial clips, breath, heart waveforms, and other physiological parameters recorded during machine interactions, enabling cross-platform applications from controlled laboratory settings to real-world social media contexts.
Personalized recommendation systems often suffer from various biases in data-driven paradigms, which can significantly degrade their fairness and generalization. Existing debiasing approaches typically rely on strong assumptions about bias types or purely statistical corrections, limiting their adaptability to unknown and changeable bias patterns, which fail to capture the complexities of human behavior and decision-making. To address this challenge, we propose a novel SVD-based dual adaptive debiasing learning (SDADL) framework. In SDADL, we distinguish between target samples (TS), which are observed positive interactions in training, and flaw samples (FS), which are highly ranked but potentially spurious positives. During training, node embeddings are decomposed via singular value decomposition (SVD) and further refined through dual adaptive learning pathways dedicated to TS and FS, respectively. Notably, the model is not explicitly constrained by the loss on the original embeddings; instead, it adaptively balances dual path learning signals under weak supervision, thereby discovering and mitigating hidden biases in a self-guided manner. Experiments demonstrate that SDADL consistently enhances base GCN models in a plug-and-play form and is effective in both supervised and self-supervised learning scenarios.
Steganography-based malware poses a growing concern across modern digital and social media ecosystems, further amplified by high-volume content generation in interconnected cyber–physical platforms. Traditional stego detection and filtering mechanisms often rely on ground-truth cover images, lack interpretability, and struggle to adapt to evolving embedding strategies under real-time constraints. Motivated by the broader shift toward intelligent multiagent systems in security-critical environments, this work introduces an adaptive framework that brings together several autonomous components to sanitize stego impurities without requiring access to their original covers. This architecture employs a reinforcement learning-driven controller that coordinates pretrained steganalysis and sterilization agents through a soft actor critic (SAC) policy guided by gradient-weighted class activation mapping (Grad-CAM) attention maps for explainable spatial attention. The off-policy SAC formulation enables efficient training and low-latency inference, suitable for deployment in high-throughput media streams. Experimental evaluations across multiple benchmark datasets demonstrate that the proposed method consistently removes around 95% of the embedded content with minimal impact on visual quality. By combining explainable spatial reasoning with adaptive learning, this work provides one of the first cover independent solutions for operational stego sterilization and offers a promising direction for extending intelligent multiagent concepts to broader cybersecurity challenges.
Detecting hate memes on social media presents a formidable challenge due to their multimodal and often subtle nature. The hateful intent typically emerges from a nuanced interplay between visual and textual elements, which existing methods often fail to capture by inadequately modeling these cross-modal correlations. To address this gap, we propose semantic context alignment and latent enhancement network (SCALE), a novel framework designed to achieve a deep understanding of hateful content. Our approach operates in two core stages. First, for Semantic Context Alignment, we introduce an attention-enhanced multimodal alignment (AEMA) module. It dynamically aligns image and text features into a unified semantic space, enabling a fine-grained grasp of the visual-textual context. Second, to achieve Latent Enhancement, we employ a lightweight convolutional semantic enhancement network (CSEN). This network refines the aligned and fused representations, distilling high-level, abstract semantics that are crucial for deciphering implicit hateful meanings. Extensive experiments on three benchmark datasets validate the superiority of SCALE. Notably, it achieves a new state-of-the-art area under the receiver operating characteristic curve (AUROC) of 92.95% on the HarMeme dataset, significantly outperforming prior methods and demonstrating its effectiveness in tackling the complexities of hate meme detection.
