Cross-lingual cross-target stance detection addresses the critical challenge of identifying textual stances (support/oppose) toward specific targets when linguistic variations and target distribution discrepancies exist across multilingual contexts. Performance degrades due to cross-lingual feature misalignment and target representation inconsistency. Furthermore, existing methods typically model text and target representations independently, overlooking that stance depends on the joint representation of the text and target. To overcome these limitations, we propose the Dual-Expert Collaborative Model (DECM), a unified framework that synergizes text-target interaction modeling with cross-lingual adaptation and target distribution alignment through three designed components: Cross-Lingual Expert Module leverages prompt-enhanced fine-tuning on mBERT with cross-lingual consistency constraints, aligning stance semantics across languages; Cross-Target Expert Module addresses target heterogeneity via joint text-target encoding, contrastive representation clustering, and feature space optimization to disentangle stance-relevant target characteristics; Collaborative Adaptation Module integrates cross-expert knowledge through pseudolabel propagation on unlabeled target-language data, reinforced by confidence-weighted prediction aggregation and unsupervised contrastive representation refinement. Extensive experiments demonstrate that DECM significantly outperforms baseline methods across multiple cross-lingual and cross-target stance detection benchmarks. It achieves 77.05
Skeleton-based gait emotion recognition has attracted significant attention due to its wide-ranging application potential. Existing approaches, however, still face several limitations: spatial graph topologies often follow physical joint connections and therefore miss semantically related long-distance coordination, temporal modules based on local convolutions have limited receptive fields, and visually similar emotions can produce ambiguous skeleton patterns. To address these issues, we propose a multi-semantic hierarchical graph-transformer network (MSH-GT) for gait emotion recognition. Central to our model is the hierarchical graph-transformer module, which integrates hierarchical graph convolutions with Transformer mechanisms. This design enhances spatial relationship modeling by adopting a centroid diffusion strategy to construct multi-scale semantic connections and leverages global temporal context to capture long-range dynamics. Additionally, a motion-pose spatio-temporal fusion module is proposed for dual-stream information fusion, and a prototype-based refinement contrastive module improves feature discriminability for ambiguous samples. Comprehensive experiments under the adopted Emotion-Gait and ELMD benchmark protocols show that MSH-GT achieves competitive state-of-the-art accuracy while using fewer parameters and FLOPs than several strong graph-based baselines. The source code is released on https://github.com/lebronble/MSH-GT .
Fuzzy β covering rough set model represents a significant expansion of covering rough set model, which can effectively handle the uncertainty and fuzziness of data. Swarm intelligence algorithms, which simulate the collaboration and interaction between individuals in biological groups, find extensive application across machine learning domains to solve complex optimization problems. In attribute reduction, swarm intelligence algorithms facilitate the efficient search for optimal attribute subsets within a vast attribute combination space. This paper studies fuzzy β covering-driven attribute reduction for hybrid data via fuzzy conditional information entropy using matrix operation and cuckoo search algorithm. Initially, a distance function is utilized to ascertain the difference or similarity between objects in hybrid data, further establishing the corresponding fuzzy β covering decision information system. Subsequently, fuzzy conditional information entropy is defined within this system to measure the classification ability of attribute subsets, and then fuzzy conditional information entropy can be viewed as an evaluation function for attribute reduction. After that, a method for computing fuzzy conditional information entropy using matrix operation is proposed. Next, an attribute reduction algorithm based on fuzzy conditional information entropy and the greedy search strategy is designed. Additionally, an attribute reduction algorithm that incorporates fuzzy conditional information entropy and the cuckoo search strategy is also designed to search for more optimal attribute subsets. Finally, experiments are conducted comparing the two designed algorithms with nine existing advanced attribute reduction algorithms. With classification accuracy and F1 score as evaluation metrics, the experimental results show that the proposed algorithms are effective and achieve competitive classification performance on the evaluated datasets and classifiers.
In recent years, multi-scale feature processing has been regarded as key to achieving high-precision pose estimation. Among these approaches, High-Resolution Networks (HRNets) have demonstrated leading performance due to their unique multi-scale parallel processing and fusion mechanisms. However, their high computational complexity and massive parameter count severely hinder real-time applications on resource-constrained devices such as mobile platforms. To address this challenge, this paper proposes a lightweight high-resolution pose estimation network, UG-HRNet. The proposed design centers on two major innovations: First, a novel lightweight deformable convolution block (LDC-Block) is introduced. This module reduces redundancy through half-channel computation and channel mixing strategies. To preserve feature expressiveness despite lightweighting, a context-guided deformable convolution variant, termed UG-DConv, is integrated into its computational branch. UG-DConv preserves the standard deformable convolution formulation while using a decomposed large-kernel guidance branch to generate more context-aware offsets and modulation masks with limited additional cost, effectively compensating for the expressiveness loss caused by lightweighting. Second, at the architectural level, HRNet’s expensive fully parallel fusion is replaced by an iterative H/L alternating fusion mechanism. This mechanism decomposes multi-scale interactions into efficient top-down (H-Only) and bottom-up (L-Only) information flows, enabling progressive feature refinement. Finally, experimental evaluations on COCO and MPII datasets demonstrate that UG-HRNet achieves a remarkable balance between efficiency and accuracy. Based on the comprehensive analysis of model complexity, computational cost, and peak memory efficiency, the results suggest that UG-HRNet possesses significant potential for deployment on resource-constrained devices. It significantly reduces model complexity while maintaining competitive performance, serving as a practical alternative to existing popular lightweight networks.
Aiming at information fuzziness, uncertain decision-makers’ preferences and the lack of social network trust relationships in complex decisions, existing picture fuzzy information-based group decision-making methods have defects including a complex consistency check process, lack of objective weight allocation basis and difficulty in quantifying decision result reliability. This paper proposes a novel picture fuzzy information group decision-making method integrating the BrowseRank algorithm and judgment matrix consistency theory. Firstly, a breakthrough is made in judgment matrix consistency theory: transformation functions are used to convert picture fuzzy preference relations into fuzzy complementary judgment matrices to simplify data and reduce computational complexity. A derived matrix and an additive consistency determination theorem are proposed, solving cumbersome traditional verification processes. Secondly, innovations are made in social networks and weight allocation: a picture fuzzy trust propagation operator complements missing trust relationships; the BrowseRank algorithm is introduced to calculate decision-makers’ objective weights by integrating multiple parameters, breaking the limitation of the ignoring social attributes. An aggregation strategy combining the weighted geometric average operator and Hadamard product ensures the consistency of the aggregated matrix, filling relevant research gaps. Thirdly, a picture fuzzy set similarity measure defines decision-maker reliability, realizing a quantitative evaluation of decision result reliability and improving the evaluation system. Finally, the method is applied to new energy vehicle battery selection; its advantages in consistency check efficiency, weight rationality and result reliability are verified via simulation and comparative analysis.
Discriminative dictionary learning remains an effective approach for small-sample image classification, but existing methods often rely on iterative l_1 -based coding or heuristic updates of atom–class associations. To address these issues, this paper proposes Dictionary Category-Adaptive Fisher Discriminant Dictionary Learning (DCAFDDL), a supervised dictionary learning model that integrates Fisher-discriminative coefficient regularization with a category-adaptive atom-weight matrix. Compared with Fisher Discriminant Dictionary Learning (FDDL), the proposed model adopts l_2 regularization in the main coding stage, which enables an analytic coefficient update. Compared with Label-Driven Dictionary Learning (LDL), the category weights are optimized through a Karush–Kuhn–Tucker (KKT)-based solution rather than iterative projection. These design choices improve online coding efficiency and refine atom–class associations, while introducing higher offline training cost. Experiments on Extended Yale B, AR, Oxford Flowers-17, and Caltech101 show that DCAFDDL achieves competitive performance against representative discriminative dictionary learning baselines, including recent sparse-subset-selection and dictionary-pair-learning-related baselines. Additional analyses on Extended Yale B, including ablation, runtime, convergence, random-split stability, noise robustness, and class-imbalance evaluation, provide further insight into the behavior of the proposed method.
Grape leaf diseases are major risks to agricultural productivity around the world, and there is an urgent need for rapid and accurate detection methods of these diseases to prevent crop losses. Manual inspection, the old fashioned way is generally slow, expensive and error-prone. To overcome this limitation, in the present study, we introduce a deep learning model designed to automatically detect grape leaf diseases based on images. Our approach integrates transfer learning with attention mechanism, using a pretrained ResNet50 which is trained on ImageNet as feature extractor. To allow the model to attend on salient pathogenic regions in the leaf images, we augmented backbone network with a Convolutional Block Attention Module (CBAM) to boost its discriminative capacity. Moreover, a controlled two-phase training approach is applied. In the first phase, the ResNet50 backbone is frozen while the CBAM module and custom classification head are trained for task-specific feature extraction. In the second phase, the entire network is fine-tuned using a reduced learning rate to stabilize convergence and improve generalization performance. The model has been trained and tested on two different datasets (i.e. the public PlantVillage dataset (controlled) and the more challenging NGLD dataset (in-the-wild). The former reached 99.33
Maintaining apple plant health and ensuring agricultural productivity depend heavily on accurate identification of leaf diseases. In real-world field environments, variations in illumination, background clutter, contrast, and leaf texture introduce significant complexity, making it challenging for deep learning models to generalise effectively. To address these challenges, this study proposes a deep learning model for apple leaf disease classification. We begin by benchmarking three modern feature extractors, namely Swin Transformer-Tiny, MobileNetV3-Small, and ConvNeXt-Tiny, to establish a strong and diverse architectural foundation. Building upon these backbones, we integrate an enhanced Intuitionistic Fuzzy Fusion (IFF) module that explicitly models membership, non-membership, and hesitation degrees, thereby improving the model’s ability to image feature calibration in ambiguous and visually similar disease patterns. With the incorporation of the IFF mechanism, the proposed framework achieves a classification accuracy of 0.9918 and a macro-F1 score of 0.9905. Learning curves, confusion matrices, and ROC–AUC analysis further validate the improved discriminative capability of the model and its reduced misclassification rate across closely related disease classes. The proposed IFF fusion framework consistently maintains the performance of individual backbones as well as conventional fusion strategies, providing a generalizable feature calibration suitable for practical agricultural deployment.
To address feature selection in high-dimensional datasets with heterogeneity and asymmetry, this paper proposes L_1 -norm Nonparallel Hyperplane Support Vector Quantile Regression ( L_1 -NHSVQR). The proposed model incorporates a quantile parameter to capture heterogeneity across the data distribution and simultaneously constructs two nonparallel lower and upper regression functions to reflect asymmetry at each quantile level. Unlike sparse twin support vector regression, L_1 -NHSVQR maintains consistency in its training and testing procedures. In addition, the L_1 -norm regularization term endows the model with an inherent feature selection capability. Experimental results demonstrate that L_1 -NHSVQR not only identifies relevant features but also comprehensively captures both heterogeneity and asymmetry within the data.
Financial time series are characterized by high volatility, strong nonlinearity, and a pronounced dependence on exogenous drivers. Accurate forecasting of prices and risks is therefore crucial for investment decision-making and risk management. However, most existing methods still rely primarily on the historical behavior of a single target asset, while largely overlooking the latent coupling between the target and diverse exogenous variables such as macroeconomic indicators, sector indices, and cross-asset signals. To simultaneously characterize cross-asset interactions, multi-scale temporal dynamics, and the complex loss landscape, we propose FTSformer, a novel financial-initiated multi-scale exogenous-fusion framework for time series forecasting. For heterogeneous information embedding, we introduce a Dual-Domain Fusion Layer (DDF) that jointly extracts and fuses exogenous information in both the time and frequency domains, while explicitly emphasizing its intrinsic heterogeneity. We then design a Multi-Scale Temporal Partition (MSTP) mechanism together with a Channel-Modulation Preprocessor (CMP), which dynamically adapts the weights of each input channel to balance varying signal strengths and enhance the training stability and convergence of the core temporal modeling module. In addition, a BA-LSTM component is employed to capture both short-term fluctuations and long-term dependencies in financial markets. Extensive experiments conducted on multiple public benchmark datasets, as well as a real-world Canadian traffic dataset, demonstrate that FTSformer substantially outperforms existing baselines in terms of forecasting accuracy, training stability, and robustness to exogenous perturbations, reducing forecasting errors by up to 13.4
To enhance the intelligent collaborative execution capability of tasks by Coaxial-Rotor-Aircraft (CRA) cluster, the problems of rapid response and stability maintenance need to be addressed. This study investigates a decentralized adaptive tracking control method based on a leader-follower coordination strategy, capable of achieving both the rapid formation tracking response and flight stability of each CRA simultaneously. Specifically, the mathematical model of CRA is established and extended to the CRA cluster system. Then, the intelligent formation tracking control scheme is proposed, guided by the virtual leader strategy, and the control system is divided into a position subsystem and an attitude subsystem. The position subsystem maintains formation by adjusting the relative position of each CRA, and the attitude subsystem is responsible for changing the attitude of each CRA to the desired stable state. The position controller is developed using the back-stepping control strategy, while the attitude tracking controller is created by an adaptive radial basis function neural network (RBFNN) based on the position model. The stability of these controllers is mathematically proven with Lyapunov theory. In the verification section of this work, numerical simulations and software-in-the-loop (SITL) simulations of intelligent tracking are conducted. The obtained trajectories of CRAs are smooth and continuous, demonstrating excellent position response and minimal amplitude vibration.
Content-Based Image Retrieval (CBIR) has benefited significantly from deep learning approaches, particularly Convolutional Neural Networks (CNNs), which are effective in capturing local visual patterns. However, CNN-based representations may lack sufficient global contextual information. Vision Transformers (ViTs), on the other hand, are capable of modeling long-range dependencies and global relationships among image regions. Motivated by the complementary properties of these two architectures, this work proposes an efficient hybrid framework, termed HCV-MLO (Hybrid CNN-ViT with Multi-Loss Optimization), which integrates CNN and ViT branches to jointly exploit local and global semantic features for image retrieval. To further improve the discriminative capability of the learned embeddings, a multi-loss optimization strategy combining Triplet loss, Contrastive loss, and Cross-Entropy loss is adopted. This paper extends our previous conference version accepted at MAPR 2025 by providing a more comprehensive analysis of the role of each loss component through three single-loss variants: CNN-ViT-CE (Cross-Entropy only), CNN-ViT-Contrastive (Contrastive only), and CNN-ViT-Triplet (Triplet only). Extensive experiments conducted on CIFAR-100, CUB-200-2011, and Stanford Dogs datasets demonstrate that the proposed hybrid framework consistently outperforms single-network baselines and single-loss variants across multiple retrieval metrics. In particular, the full HCV-MLO model achieves an mAP@100 of 81.99 on CIFAR-100, showing a clear improvement over recent competitive methods. These results indicate that combining hybrid CNN-ViT representations with multi-loss learning provides a practical and effective solution for enhancing embedding quality and retrieval performance in CBIR tasks.
In complex network analysis, accurately identifying critical nodes is pivotal for understanding information propagation and network dynamics. Traditional methods like degree centrality and betweenness centrality often neglect spatial distributions, leading to clustered node selections that reduce propagation effectiveness. This study introduces the KSvoterank algorithm, which integrates k-shell decomposition with an enhanced voting mechanism to effectively disperse key nodes and mitigate the “rich-club” effect. By stratifying nodes via k-shell and employing a dynamic voting process that considers both direct and indirect neighborhood contributions, KSvoterank maintains global structural information while minimizing influence overlap. Comprehensive evaluations on multiple real-world networks demonstrate KSvoterank’s superior performance in propagation scope and robustness, validating its utility for influence maximization and network robustness analysis.
The interpretability issue in traditional collaborative filtering methods can be effectively addressed by Knowledge Graphs (KG), which offer comprehensive structured data. However, deep Graph Neural Network (GNN) models often face over-smoothing challenges as they grow in network depth, causing convergence of node representations and loss of personalized features. To tackle this, issue we introduce a novel model, the Multi-level Contrast Learning Knowledge Perception Recommendation System with a Residual Enhancement Mechanism (MCLRE). This mechanism incorporates the node representation from the preceding layer into the graph convolution of each layer, mitigating information loss and excessive smoothing. Moreover, our model employs a multi-level contrast learning approach that not only contrasts within the map but also interacts across maps. This strategy enhances the fusion of collaborative filtering signals and knowledge graph data, thereby enhancing recommendation accuracy and robustness. Experimental validation on various public datasets demonstrates that MCLRE outperforms existing state-of-the-art methods in terms of AUC, F1, and Recall@K metrics. Ablation studies further confirm the pivotal role of the residual enhancement mechanism and multi-level contrast learning strategy in enhancing model performance.
Collaborative filtering (CF) learns user—item preference patterns from interaction data and is widely used in recommender systems. However, traditional CF models may overlook individual-level bias, causing some users to receive poorer recommendations than others. Such disparity can reduce user satisfaction and engagement. Although prior studies have examined individual-user unfairness using bandits, unfairness metrics, and explanation scores, its effective mitigation remains relatively underexplored. To address this issue, this paper proposes a dual-stage counterfactual imputation framework for mitigating individual user unfairness in CF. The framework first identifies probable candidate users based on limited interaction histories and low training-side utility. It then temporarily imputes new interactions for these users and evaluates their effect on training-side utility. Probable candidate users whose temporary imputation improves either their individual training-side utility or the average training-side utility across all users are retained as final candidates. Based on the metric used for this evaluation, two Counterfactual Fairness Mitigation (CFM) variants are proposed: CFM-N using NDCG and CFM-F using F1. Experimental results on three datasets show that the best proposed variant improves overall-user F1@10 over the closest competing fairness method by 122.8
A critical bottleneck in self-supervised skeleton-based action recognition is the limited semantic distinctiveness of learned features. While conventional approaches predominantly utilize global data augmentation to facilitate contrastive learning across different views, they often fail to establish sharp semantic boundaries in the latent space. To mitigate this issue, we present a novel self-supervised architecture centered on graph prototype learning. Our framework integrates two pivotal modules: a Graph Prototype Learning Network, which employs learnable prototype vectors to reconstruct masked inputs thereby coercing the encoder to distill shared action semantics rather than fitting redundant noise and a Discriminative Class Contrastive Learning mechanism. The latter utilizes a memory bank to maintain dynamic class centroids, optimizing the alignment between samples and their respective prototypes via contrastive loss. Extensive evaluations on the NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD benchmarks reveal superior performance, particularly in terms of class discrimination and robustness to occlusion. Notably, our method sets a new state-of-the-art on the ARMM dataset with an accuracy of 95.70
In this paper, we propose a new fuzzy approximate reasoning framework covering Fuzzy Modus Ponens/Fuzzy Modus Tollens in a fuzzy system, which is based on an extended distance defined via the Least Common Multiple (LCM). It has no information loss and offers stronger reductive property and improved fuzzy controllability. Its basic idea is to use the LCM to align the dimensions of the antecedent and the consequent with different dimensions, and then drive the inference by distance measure, while designing two sign-vector forms, P(+ 1,0,−1) and P(+ 1,−1), in order to modulate a conclusion. Theoretical analyses show that the proposed method outperforms existing methods such as Compositional Rule of Inference, Triple Implication Principle, Quintuple Implication Principle, and Approximate Analogical Reasoning Scheme in terms of reductive property, computational time and complexity, and information loss. Application of a novel rotary inverted pendulum control is presented, which is based on the proposed method of the LCM and KSI principle. Experimental results based on real-word data have verified its effectiveness and the characteristics that conform to human thinking.
Early detection of plant diseases is essential for improving crop productivity and ensuring sustainable agriculture. However, many existing tomato leaf disease detection systems rely on either deep learning or handcrafted features alone, limiting robustness and interpretability in practical agricultural environments. To address these limitations, this study proposes a hybrid multi-modal feature fusion framework that integrates deep convolutional representations with handcrafted texture descriptors for tomato leaf disease classification. The proposed method combines features extracted from a modified MobileNetV3 network with Gray-Level Co-occurrence Matrix (GLCM) texture descriptors including contrast, homogeneity, entropy, and correlation. Additionally, multi-channel representations derived from RGB, HSV, Lab color spaces and edge-detection filters are incorporated to enhance spatial and spectral feature diversity. Particle Swarm Optimization (PSO) is applied for deep feature selection, while Principal Component Analysis (PCA) reduces dimensional redundancy in the fused feature space. Classification is performed using an Extreme Learning Machine (ELM), whose hyperparameters are optimized using PSO. Experimental evaluation on the PlantVillage tomato leaf disease dataset demonstrates that the proposed framework achieves an accuracy of 98.60
Utilizing rich video-level contextual information to perceive targets effectively is crucial for object tracking. However, existing methods typically capture limited-length context through simple multiple video or adjacent frames, resulting in underutilization of contextual information. To address this issue, we propose the DFCTrack tracking framework. It captures target behavior changes across the entire temporal span of the complete video sequence rather than a few images. Specifically, we propose a long-term context modeling mechanism based on collaborative architecture between Mamba and backbone networks. It utilizes Mamba’s hidden state iteration to enhance the backbone’s cross-frame spatiotemporal information aggregation capability. This integrates target evolution cues across full temporal dimensions and captures long-range spatiotemporal information. On the other hand, we propose a Short-term Target Redundancy Removal module. This module precisely selects high-value tokens relevant to the current target by jointly optimizing attention entropy and classification score. Provides recent high-quality target representations for the tracker. Finally, we design a Long-Short Term Context Fusion module, which leverages a self-attention mechanism to dynamically weight and fuse long-term context with recent high-quality token representations. This introduces global temporal information into the model, significantly enhancing its spatiotemporal awareness. Experiments demonstrate that DFCTrack achieves excellent tracking performance on multiple benchmarks such as GOT-10K, TrackingNet, and LaSOT.
