Applied Intelligence最新文献

Predicting stock returns is difficult due to numerous dynamic and interrelated factors such as interest rates, exchange rates, economic conditions, corporate policies, investor behavior, and political developments. This study proposes a hybrid trading strategy that integrates the Decision Tree (DT) algorithm with the Markowitz Mean-Variance (MV) portfolio optimization model to improve prediction accuracy and optimize investment returns. Using data from 94 companies continuously listed on the NASDAQ 100 (NDX100) index between 2015 and 2024, the model uses technical analysis indicators as input for the DT to predict short-term stock returns and then applies MV optimization to allocate weights within dynamically generated portfolios. The empirical analysis includes 64 portfolio variations derived from different parameter settings and is evaluated using the Sharpe ratio, Sortino ratio, maximum drawdown, total return, and turnover ratios. The results show that portfolios created with higher trading target ratios and moderate MV optimization significantly outperform the NDX100 index, providing superior risk-adjusted returns and better stability. These findings highlight the effectiveness of combining machine learning-based regressor with classical optimization for dynamic portfolio creation and underscore its practical potential for improving trading decisions in volatile markets.

The systematic analysis of user-generated social media content, especially when enriched with geospatial context, plays a vital role in domains such as disaster management and public opinion monitoring. Although multimodal approaches have made significant progress, most existing models remain fragmented, processing each modality separately rather than integrating them into a unified end-to-end model. To address this, we propose an unsupervised, multimodal graph-based methodology that jointly embeds semantic and geographic information into a shared representation space. The proposed methodology comprises two architectural paradigms: a mono graph (MonoGrah) model that jointly encodes both modalities, and a multi graph (MultiGraph) model that separately models semantic and geographic relationships and subsequently integrates them through multi-head attention mechanisms. A composite loss, combining contrastive, coherence, and alignment objectives, guides the learning process to produce semantically coherent and spatially compact clusters. Experiments on four real-world disaster datasets demonstrate that our models consistently outperform existing baselines in topic quality, spatial coherence, and interpretability. Inherently domain-independent, the framework can be readily extended to diverse forms of multimodal data and a wide range of downstream analysis tasks.

This paper presents a deep learning-based numerical algorithm named Deep Rayleigh Quotient Iteration (DRQI), which extends the classical, high order convergent scheme in matrix eigenvalue problems, Rayleigh quotient iteration (RQI), to large scale discretization and further continuous linear differential operator through neural network parametrization. By embedding the RQI structure into a semi-implicit relaxation loss, which couples successive Rayleigh updates, preserves the convergence behavior of RQI in the extended functional setting. And the mean squared residual of the governing equation serves as a stopping criterion to avoid overfitting. The performance of DRQI is systematically analyzed from both theoretical and experimental perspective. Theoretical analysis proves that the DRQI inherits the higher-order convergence property of RQI and provides the guidelines for network design and standards for sampling density, establishing the consistency between discrete RQI and its continuous operator counterpart. While the benchmark experiments across the Laplace and Fokker-Plank operators demonstrate that DRQI is more robust in lower dimensional cases, outperforming the deep Ritz method and the inverse power method neural network in accuracy of eigenpairs estimation. The results also indicate that the relaxation factor enables control over the early training stages, enhancing adaptability in practical applications. Finally, DRQI is applied to the 1D neutron transport equation, achieving errors below 10^–5 within 1500 training epochs when estimating the effective multiplication factor. These results highlight DRQI as a general, higher-order convergent, and practically robust eigenvalue problem solver with strong potential for engineering applications.

Purpose: A decision support pathway for general practitioners (GPs) was explored through automated referral letter analysis, with large language models’ (LLMs) diagnostic roles comprehensively evaluated. Methods: The in-context learning performance of ChatGPT and GPT-4 for diagnostic decision support was evaluated using referral letters. Synthetic referral letters generated by ChatGPT addressed data scarcity, with distributional congruence quantified via Kullback-Leibler divergence. Two fine-tuning frameworks were comparatively assessed: encoder-based pre-trained language models (PLMs) for diagnostic classification, and decoder-based LLMs adapted to multiple-choice question-answering paradigms. Results: GPT-4 showed suboptimal few-shot accuracy (0.544). Synthetic letters demonstrated high fidelity (KL-divergence<0.05). Encoder-based PLMs consistently outperformed decoder-based LLMs when fine-tuned with augmented data, with BERT achieving 0.977 accuracy in mixed-train-collect-test protocols. Complementary F1 (0.9707) confirmed negligible diagnostic bias. Conclusion: LLMs exhibited insufficient diagnostic accuracy through both direct implementation (GPT-4 few-shot: 0.544) and fine-tuning approaches (accuracy 0.723), establishing fundamental limitations in clinical deployment. Crucially, their text-generation capability was leveraged for structured data augmentation, producing synthetic referral letters with high distributional fidelity (KL-divergence<0.05). This validated methodology enabled superior diagnostic performance through encoder-based PLM fine-tuning, where BERT achieved near-clinical-utility accuracy (0.977) - demonstrating 25.4% relative improvement over best-performing LLMs. Implementation pathways consequently prioritize this hybrid framework: LLM-mediated data augmentation followed by resource-efficient PLM classifiers, currently undergoing neurologist-piloted validation before multicenter expansion.

Multi-view clustering, which integrates complementary information from heterogeneous data sources, has witnessed substantial progress in graph-based methodologies. However, prevalent graph-based methods often suffer from sparse initial graph constructions and high computational demands for large-scale data. Furthermore, effectively learning consistent underlying structures while mitigating noise and view discrepancies remains challenging, frequently leading to fusion distortion. In this paper, we present a novel Anchor-based Tensor Fusion method via High-Order Relations (ATFMH) to address these limitations, ATFMH leverages anchor graphs to build high-order graphs, systematically mitigating initial sparsity and capturing deeper structural relationships. Furthermore, ATFMH employs a low-rank tensor nuclear norm constraint on the jointly learned similarity graphs. This constraint promotes a consistent shared subspace across views, enhancing robustness against view-specific variations and noise. Extensive experiments demonstrate that ATFMH achieves an optimal balance between efficiency and accuracy. ATFMH significantly reduces computational and spatial costs while maintaining or exceeding state-of-the-art clustering performance across multiple benchmarks. This work provides an efficient and robust solution for large-scale multi-view data analysis.

Multi-view clustering has been extensively studied to group large amounts of multi view data using multiple information sources. Traditional multi-view clustering algorithms often exhibit quadratic or even cubic complexity, posing challenges for large-scale datasets. Recently, several algorithms have employed anchor graphs to reduce expensive time complexity. However, these algorithms fail to respect the intrinsic geometric structure within individual views. To address this issue, this paper proposes a novel multi-view clustering algorithm, called fast multi-view clustering with geometric structures (FMCGS). FMCGS constructs an affinity graph to model the geometric structure within each view separately. Moreover, it can adaptively assign different weights to different views, expecting views that play a more important role in clustering to have greater weights. We also present an optimization scheme based on iterative updating of three factor matrices to solve the proposed model. Extensive experiments on nine real-world datasets validate the superiority of the proposed approach over state-of-the-art methods.

Effective medical image segmentation results are essential for the subsequent diagnosis. Fully convolutional networks and their variants based on the encoder-decoder structure have achieved excellent performance in medical image segmentation. Despite recent progress, segmenting targets with large-scale variations and complex backgrounds remains difficult. In this paper, we propose a Mixed-Scale Context Fusion (MSCF) network for medical image segmentation. First, we introduce a context fusion module between the encoder and decoder, which can fuse multi-level features from the encoder and extract contextual information useful for the segmentation task. Second, we introduce a mixed-scale attention module, it can extract features at different scales by two branches, and learn scale information adapted to the target size using an attention mechanism to enhance the capability of multi-scale feature extraction. We conduct extensive experiments on the LIDC and MSD datasets, where MSCFNet achieves Dice scores of 85.83% and 86.09%, improving over FCN by 8.36% and 6.57%, respectively.

Objective

Convolutional Neural Networks (CNNs) have become essential tools for classroom student behavior recognition but lack the capability of global information capturing. In recent years, Vision Transformer (ViT) has demonstrated strong global modeling capabilities, which can be employed to strengthen the multi-level spatial information representation ability of classroom student behavior recognition models.

Methods

First, Cascaded Residual Vision Transformer (CR-ViT) model was proposed. The outputs of residual convolutional layers were integrated into multiple ViT modules to learn both shallow and deep feature representations for multi-level spatial information extraction, followed with the LSTM network for further capturing the global dependency between the sequences of ViT modules. Second, Cascaded Residual Vision Transformer with Morlet wavelet (MCR-ViT) was proposed. Based on CR-ViT, morlet wavelet transform activation layers were employed to improve the sensibility for variations of edges in feature maps.

Results

The proposed methods were validated on our collected dataset named Student Behavior in Classroom (SBIC), as well as the publicly available dataset of Student Classroom Behavior (SCB). The CR-ViT model and the MCR-ViT model improved the accuracy on SBIC by 6.10% and 7.32%, and on SCB by 0.92% and 1.14%. The MCR-ViT with second-order derivative of Morlet wavelet achieved the highest accuracy improvement compared to its variants based on other wavelet.

Conclusion and significance

both CR-ViT and MCR-ViT exhibit superior performance, which can be leveraged to build high-performance student behavior recognition systems in classrooms.

In data mining research, streaming data analysis has emerged as a critical computational paradigm requiring real-time processing capabilities to ensure temporal relevance and practical applicability. However, dynamic data stream environments commonly exhibit complex challenges including multi-class imbalance phenomena with time-varying distribution ratios, and concept drift characteristics that collectively undermine the stability and accuracy of conventional classification frameworks. To address these interdisciplinary challenges, this study presents SCWE+, an innovative data stream classification algorithm specifically designed for multi-class imbalance scenarios with concept drift adaptation. The methodology comprises three core components: (1) A sample classification difficulty weighting mechanism integrating margin-based sample evaluation with supervised negative margin rescue loss, which strategically enhances model focus on both easily misclassified instances and minority class distributions through adaptive attention allocation; (2) A dynamic ensemble selection framework incorporating adaptive plasticity reward (APR), which implements hierarchical sliding window analysis across sample, class-specific, and difficulty-stratified dimensions to generate context-aware classifier performance evaluations and optimize prediction ensembles through weighted aggregation; (3) An expert validation architecture that constructs class-specialized expert groups by selecting high-performance classifiers from the classifier pool for specific class domains, enabling post-ensemble prediction verification to mitigate misclassification risks in low-confidence predictions. Comprehensive empirical evaluations were conducted across diverse synthetic and real-world data stream benchmarks, demonstrating statistically significant performance improvements compared to nine state-of-the-art data stream classification algorithms under varying imbalance ratios and concept drift scenarios.

Physics-informed neural networks (PINNs) have proven to be a powerful tool for scattering simulation, leveraging physics principles to inform the learning process. When solving the acoustic scattering problem induced by the scatterer, conventional PINNs that rely on multilayer perceptrons (MLP) neglect the inherent acoustic field dependencies, thus failing to propagate globally the scattering characteristics and accurately capture nonlinear features near the scatterer boundary. Researchers often define explicit architectures or increase training parameter counts to capture these dependencies, yet this inevitably introduces either structural constraints or heightened computational costs. In this paper, we propose a novel framework based on the Kolmogorov–Arnold Network (KAN), termed PIKAN, to address this limitation. Leveraging KAN-driven plane wave expansion and the Sine activation function’s efficient plane wave learning capacity, PIKAN effectively captures both global dependencies and fine-grained scattering features in the scattered acoustic field. We evaluate the effectiveness and generalization performance of the PIKAN model across three typical scattering scenarios: high-frequency case, irregular scatterer, and multiple scattering, to validate its practicality. Experiments demonstrate that PIKAN achieves a superior balance between both accuracy and computational efficiency, when compared against state-of-the-art baselines and alternative activation functions. Moreover, integrating with the RAD method enhances the model’s flexibility and further improves its performance.