Applied Intelligence最新文献

In recent years, dynamic multi-objective optimization problems (DMOPs) have received increasing attention, and numerous strategies have emerged to solve such problems. However, in most existing algorithms, the specific roles of each decision variable are often overlooked. This article proposes an adaptive dynamic multi-objective optimization algorithm utilizing two-stage correlation-based decision variable analysis named ADMOEA-TSCDVA. In the proposed decision variable analysis strategy, the decision variables are initially divided into convergence-related and diversity-related variables based on correlation analysis in the first stage. Subsequently, in the second stage, diversity-related variables are further classified into simple diversity-related and complex diversity-related variables. The results of the decision variable analysis will be used in a dual-model initialization strategy to guide the algorithm in generating a high-quality initial population when changes occur, and in an adaptive strategy to select a suitable static optimization algorithm when the environment remains unchanged. Comprehensive comparative experiments were conducted on two DMOP benchmark test suites against five state-of-the-art algorithms. The results show that the proposed ADMOEA-TSCDVA achieved the best MIGD and MHV results in 63% and 64% of the 57 test cases, demonstrating its excellent performance in dynamic multi-objective optimization scenarios.

Traffic flow prediction is a core task in intelligent transportation systems (ITS), challenged by complex spatiotemporal dynamics. Nodes in traffic networks exhibit diverse temporal patterns, including abrupt fluctuations during peak hours or sudden weather changes, while maintaining spatial correlations that evolve over time. Accurate prediction thus requires effectively capturing high-dimensional spatiotemporal dependencies. In this paper, we propose the Spatiotemporal Decoupling-Gated Transformer (STDGformer), whose key principle is decoupling for coupling. Instead of directly modeling entangled spatiotemporal features, STDGformer first decouples temporal and spatial representations using a Gated Temporal Transformer and a Gated Spatial Transformer. This allows the model to capture high- and low-frequency temporal patterns, as well as dynamic and static spatial dependencies, more effectively. Subsequently, a Cross Attention Transformer integrates the decoupled features, reconstructing their joint interactions and enabling precise modeling of multi-dimensional spatiotemporal dependencies. Extensive experiments on four real-world traffic datasets demonstrate that STDGformer consistently outperforms representative baselines, achieving superior accuracy and robustness while providing interpretable insights into temporal and spatial feature contributions. Specifically, STDGformer reduces MAE by 1.1% on PEMS04 and decreases MAPE by 1.2% on PEMS03, respectively.

Multi-view learning has been capturing widespread attention from a variety of areas, while very limited labeling information poses a great challenge for multi-view semi-supervised classification. Especially, rare labels often require multi-hop label propagation, which leads to over-smoothing and inaccurate prediction. In contrast, reinforcement learning can capture long-term policy series by a reward function. In this paper, we address the semi-supervised multi-view classification problem from a reinforcement learning perspective. The proposed approach frames the classification problem as a process where an agent, guided by reinforcement learning, starts from the state representing the unlabeled sample, transitions to the state corresponding to the label, and completes the label assignment. The reward matrix’s core is defined as the affinity matrix between samples, complemented by one-hot encoding in the label space and an identity matrix. The Q-table is then derived from the rewards in Q-learning, enabling the use of limited label information to effectively mine multi-view data. In the label assignment phase, the agent utilizes the table to eventually reach the state corresponding to the label and assign that label to the sample represented by the initial state. Our method’s effectiveness is confirmed through extensive experiments on multi-view semi-supervised classification tasks compared with several advanced approaches.

With the rapid development of intelligent tutoring systems (ITSs) in the past decade, tracing students’ knowledge state has become increasingly important for providing individualized learning guidance; this is the main idea of knowledge tracing (KT), which models students’ mastery of knowledge concepts (KCs, skills needed to solve a question) based on their past interactions on platforms. Many KT models have been proposed and have recently shown remarkable performance. However, the majority of these models use concepts to index questions, which implies that the predefined skill tags for each question are required in advance to indicate the specific KCs needed for answering the question correctly; this makes it difficult to apply to large-scale online education platforms where questions are often not well organized by skill tags. In this paper, we propose Q-matrix-based attentive knowledge tracing (QAKT), an end-to-end KT model that uses the attentive approach in situations where predefined skill tags are not available. With a novel hybrid embedding method based on the Q-matrix and Rasch model, the QAKT model is capable of modeling problems hierarchically and learning the Q-matrix efficiently based on students’ sequences. Moreover, the architecture of the QAKT ensures that it is friendly to questions associated with multiple skills and has outstanding interpretability. After conducting experiments on a variety of open datasets, we empirically verified that even without predefined skill tags, our model performs similarly to or even better than the state-of-the-art KT methods, by up to 2% in AUC in some cases. Moreover, our model outperforms existing models that do not require skill tags (by up to 7% in AUC) in predicting future learner responses. The results of further experiments suggest that the Q-matrix learned by the QAKT is highly model-agnostic and more information-sufficient than the one labeled by human experts, which could help with the data mining tasks in existing ITSs.

The aim of this paper is to propose a novel group decision-making (GDM) method that addresses expert weight uncertainty and efficiency interval estimation. Based on interval multiplicative probabilistic linguistic preference relations, this study combines entropy weight method, multiplicative data envelopment analysis (DEA) cross-efficiency and Bootstrap analysis to provide a more accurate and reliable GDM framework. First, logarithmic function is introduced into the entropy method to measure the fuzziness and hesitancy of preference matrix sufficiently and quantify the weight of expert information. Then, a multiplicative DEA cross-efficiency model is adopted for efficiency evaluation of the decision-making units. Logarithmic function is designed in this model to intuitively display efficiency differences. We further develop Bootstrap-DEA for efficiency correction, leveraging its statistical advantages to enhance the robustness and reliability of our GDM method. Finally, a numerical example and comparative analysis are provided to highlight the rationality and effectiveness of the proposed GDM method.

In the field of neuroscience, it has been confirmed that the hippocampal CA1 plays a crucial role in spatial navigation. Establishing an appropriate deep learning architecture to replicate this process can not only more accurately predict behavior but also better understand the working mechanisms of CA1. This study utilizes an open-source dataset to predict rat immediate behavior by constructing a hybrid network capable of processing spatial-temporal information and dynamically activating over input. Specifically, the KA-AttLSTMnet architecture is proposed to replicate the CA1’s navigation function. First, this study employs spike-related methods to encode CA1 activity and represents a total of eight open-field behavioral states under allocentric/egocentric strategies. Then, spike encoding and behavioral states are used as inputs and outputs, respectively, to establish eight different immediate behavior prediction models. Finally, in terms of model structure, a self-attention mechanism architecture based on recurrent neural network (RNN) is built, and Kolmogorov-Arnold network (KAN) is employed to further dynamically adjust the architecture, which already possesses spatial-temporal processing capabilities, to enhance the extraction of CA1 neural activity information and improve the prediction of the rat’s immediate behavior. By comparing the prediction results of the eight behavioral states, only two behaviors, direction of turn ({{{varvec{b}}}_{{varvec{t}}}}^{{varvec{i}}}) and speed ({{b}_{s}}^{i}), achieved relatively stable predictions, indicating that the CA1 internal circuit is more inclined to fully reflect an egocentric strategy rather than an allocentric one. Ablation experiments demonstrated that the unrolled-LSTM network is more effective in processing spike encoding over time and organizing it over space. Additionally, the mean squared error (MSE) for predicting ({{{varvec{b}}}_{{varvec{t}}}}^{{varvec{i}}}) and ({{b}_{s}}^{i}) decreased from 0.4203/0.0435 to 0.3687/0.0150, and eventually to 0.3255/0.0110. This reduction highlights the positive impact of the multi-head self-attention mechanism in RNN (AttLSTMnet) for extracting contextual information, as well as the dynamic regulation capability of neurons in the KAN (KA-AttLSTMnet), which differs from the traditional weight-activation function mechanism, both of which contribute significantly to improving the immediate behavior prediction model.

Monitoring railway fasteners is essential for track safety. However, the development of intelligent inspection systems is hindered by the scarcity of defective fastener images and the pronounced imbalance between defective and non-defective classes. Conventional generative approaches, such as GAN- and diffusion-based models, can increase data volume but often incur high computational cost, unstable convergence, or produce samples lacking structural fidelity, limiting their suitability for engineering deployment. This work proposes the Domain-Adaptive NeuroGenerative Framework (DA-NGF), a lightweight few-shot generative model tailored to rail fastener inspection. The framework leverages compact latent modeling, orientation-invariant self-supervision, and defect-aware attention to synthesize structurally consistent fastener images from extremely limited real samples. Synthetic samples are evaluated using a separate downstream classifier (FastenerNet + +), enabling quality validation without coupling generation and classification pipelines. Experimental results show that augmenting real data with DA-NGF synthetic samples improves downstream defect recognition accuracy by 21.13% and precision by 22.53%, with corresponding gains in recall and F1-score. The model executes efficiently on CPU hardware, underscoring its suitability for low-resource deployment.

Social media and digital platforms allow us to freely and easily express our opinions to a wide audience. In this study, we investigate gender-based differences in online communication, specifically on Twitter (now X), in the context of soccer, by analyzing patterns of participation, sentiment, and engagement across male and female users. As one of the most popular sports, soccer engages a diverse audience on social media, regardless of expertise. We collected 9.5 million tweets related to soccer in English and Portuguese during three months (March - June 2022). Only 18.38% tweets were identified as written by women, highlighting a possible gender gap already in the number of people who participated actively in this topic. We analyze tweets in Portuguese and English, as these languages represent different perspectives on soccer. English serves as a proxy for global discussions, while Portuguese reflects more localized but deeply engaged communities. We observe that women and men communicate more between each other in Portuguese than in English, exhibiting lower homophily within their social networks. However, this difference in homophily does not appear to influence how women and men express emotions and sentiments, suggesting that these aspects may be shaped by other factors such as societal expectations, gender role socialization, online community dynamics, visibility constraints, or other gender-related norms and characteristics. Women express their emotions more intensely in response to events than men, regardless of the differences in the number of tweets, and men tend to be more negative in their tweets than women. Our study reveals persistent gender gaps across both weeks and hours through qualitative and quantitative analyses, including detailed text-level and network-level examinations. These findings underscore the importance of identifying and reporting gender disparities in online communication to foster more inclusive spaces where individuals can freely express their opinions.

In recent years, multi-modal salient object detection methods (e.g., RGB-D or RGB-T) have made significant progress in complex scene applications. However, existing methods often neglect edge feature enhancement and detail preservation, which are crucial for the clarity and accuracy of foreground-background separation. Additionally, effectively highlighting spatially related features, balancing local and global information, and suppressing background noise remain major challenges. To address these issues, we propose a new enhanced squeezed directional large kernel multi-scale multi-modal fusion network, named SDLK-Net, to optimize the separation of foreground and background and enhance detail capture ability. Specifically, we introduce a Squeezed Large Kernel Edge-Enhanced Fusion (SLKEF) method, which combines the squeeze operation with bidirectional large kernel convolutions to effectively enhance the separation between foreground and background, achieving the regional integrity and edge clarity of salient object segmentation results. In addition, we develop an Adaptive Extraction Multi-path Matching (AEMM) method, which enhances the fine-grained feature representation of the ROI region through path decoupling and fusion to help better capture the detailed information. Lastly, we design a Multi-Scale Filter Gate Integration (MSFGI) model, which flexibly balances local and global contextual information through dynamic filtering and gating mechanisms, ensuring the highlighting of spatially related features while effectively suppressing background noise. Experimental results demonstrate that SDLK-Net significantly improves detection accuracy in RGB-D and RGB-T salient object detection tasks, outperforming existing methods and achieving state-of-the-art performance across multiple datasets.

Imbalanced multi-label learning (IMLL) aims to learn a multi-label classifier from instances with imbalanced label distribution. Most existing IMLL models either use resampling techniques for preprocessing or transform multi-label learning problems into single-label problems, and traditional imbalanced learning methods are applied. The resampling methods may distort data distribution, whereas the latter overlook dependency and correlation between labels. To address those challenges, this paper proposes an approach to imbalanced multi-label learning, nominated as MWLDAIML. In the proposed model, a label enhancement matrix is designed according to the imbalance rates of positive and negative instances to enlarge the influence of instances in minority classes. A linear mapping is learned to function as a classifier from the feature space to the enhanced label space, and simultaneously acts as a projection from a high-dimensional space to a low-dimensional space, ensuring a large divergence between inter-class instances and a tight distribution of intra-class instances via introducing a weighted linear discriminant analysis (WLDA). Moreover, the metric learning is embedded into WLDA to discern intra-class instances and inter-class instances to capture a more complex nonlinear relationship between instances and their multiple labels. Additionally, the graph Laplacian regularization is imposed to ensure that predicted labels inherit the topological structure of instances in the feature space. An efficient algorithm for implementing MWLDAIML is developed to implement the proposed model, and extensive experiments on real-world benchmark datasets demonstrate that the proposed model outperforms existing methods for imbalanced multi-label classification.