Applied Intelligence最新文献

Deep neural networks suffer from overfitting when training samples contain inaccurate annotations (noisy labels), leading to suboptimal performance. In addressing this challenge, current methods for learning with noisy labels employ specific criteria, such as small loss, historical prediction, etc., to distinguish clean and noisy instances. Subsequently, semi-supervised learning techniques are introduced to boost performance. Most of them are one-stage frameworks that aim to achieve optimal sample partitioning and robust SSL training within a single iteration, thereby increasing training difficulty and complexity. To address this limitation, we propose a novel two-stage noisy label learning framework called UCRT, which consists of uniform consistency selection and robust training. In the first stage, the emphasis lies on creating a more uniform and accurate clean set, while the second stage uniformly extends this clean set to improve model performance by introducing SSL techniques. Comprehensive experiments conducted on both synthetic and real-world noisy datasets demonstrate the stability of UCRT across various noise types, showcasing superior performance compared with state-of-the-art methods. The code will be available at: https://github.com/LanXiaoPang613/UCRT.

Graph-regularized representation methods have demonstrated promising performance in image clustering. However, with the exponential growth of image data scales, traditional graph-regularized methods are no longer efficient in handling large-scale datasets due to their high computational and spatial complexity. Due to both natural and non-natural factors, real-world application data often contain outliers. To address these issues, this paper proposes a Bipartite graph-regularized robust Low-rank Matrix Factorization (BLMF) method for semi-supervised image clustering. The bipartite graph structure reduces the computational complexity to (varvec{O(ndt)}), representing a significant improvement compared to traditional graph-regularized methods with computational complexity of (varvec{O(n}^{varvec{2}}varvec{dt)}). Furthermore, to mitigate the negative effects of outliers, the Maximum Correntropy Criterion (MCC) is introduced as a fidelity measure in constructing the optimization model. Additionally, we propose an anchor point selection strategy to reduce the influence of anchor outliers. An iterative algorithm based on Fenchel Conjugate (FC) and Block Coordinate Update (BCU) techniques is developed to solve our model effectively. The convergence properties of the proposed algorithm are rigorously analyzed, demonstrating that it satisfies both objective convergence and iterative sequential convergence. Experiments are conducted on 11 real image datasets, comparing the proposed BLMF method with 12 state-of-the-art algorithms. The results demonstrate that the proposed method outperforms competing methods in most cases across small, medium, and large datasets.

To achieve advanced driving functions, intelligent vehicle relies on lane mark captured by an optical sensor in traffic environment. The emerging dynamic vision sensor (DVS) exhibits an impressive dynamic range and time resolution, showing potential as the next generation of on-board sensor to overcome complex lighting conditions like glare, blur, or darkness. Different from the traditional RGB image, the event image produced by the DVS is sparse and lacks color. Although existing methods achieve competitive performance for RGB-based lane mark segmentation, they often result in discontinuous lane mark segmentation when applied to the sparse event image. In this paper, we propose the Event-based Large Kernel Network (EvLKNet) specifically for segmenting sparse event lane marks. The EvLKNet is a lightweight architecture that employs multi-scale large kernel convolution which enhances the ability to capture long-range relationships among sparse event pixels and mitigates the degradation of segmentation accuracy caused by image sparsity. Moreover, intermediate event feature map distillation is used to improve prediction accuracy without increasing inference cost. The performance of EvLKNet surpasses existing advanced methods on two event-based lane mark datasets, DET and Carla-DVS. The lightweight architecture allows EvLKNet to have only 3.83M parameters and a computational overhead of 11.92 GFLOPS.

The risk of corporate bankruptcy is increasing with the complex and changing global economic environment, making accurate and reliable bankruptcy prediction models essential. Traditional methods, such as machine learning-based models, often lack semantic clarity, making their results difficult to interpret. Understanding causal relationships is crucial for uncovering the causes and mechanisms of corporate bankruptcy. Causal inference can reveal these relationships, enhancing our understanding of bankruptcy phenomena. Additionally, financial data and market information provided by companies often contain ambiguity and uncertainty. Rough fuzzy set theory, combining rough set and fuzzy set advantages, can handle such data more accurately. This paper proposes a novel multi-attribute prediction decision-making method, CIRF-TOPSIS, which integrates causal inference and rough fuzzy set theory. Unlike existing bankruptcy prediction approaches, CIRF-TOPSIS simultaneously reveals causal mechanisms behind financial indicators and handles fuzziness and uncertainty in data. Empirical testing confirms that CIRF-TOPSIS significantly outperforms traditional MADM methods in both accuracy and interpretability, offering a transparent and reliable decision-support tool for risk management.

Three-dimensional reconstruction based on hyperspectral data can be applied in numerous fields. In recent years, the 3D Gaussian Splatting (3DGS) method has garnered widespread interest in RGB images due to its fast training and rendering speeds. Directly extending 3DGS to hyperspectral images faces challenges such as reduced training and rendering speeds and increased demand for computational resources, due to numerous channels in hyperspectral images. This paper proposes a faster, smaller, and noise-robust hyperspectral 3DGS method based on feature dimension compression. The method leverages the advantages of decoupling of spatial and spectral features using the point cloud representation in 3DGS. By employing a two-stage training approach and incorporating a high-frequency regions refinement, this method significantly reduces the training time on hyperspectral data, improves rendering speed, reduces the required storage space and achieves higher robustness to noise, while achieving comparable rendering results compared with the original 3DGS method. The PSNR of rendered image is 33.9 and training time is only 0.55 hour. This method can better perform tasks such as three-dimensional reconstruction and novel view synthesis of hyperspectral data in situations with limited computational resources.

In the field of medical image segmentation, techniques based on convolutional neural networks and Transformers have been extensively developed. However, the current mainstream methods for medical image segmentation are mainly based on standard convolutions or depthwise separable convolutions, which essentially process on fixed and ordered feature channels. This inherent constraint limits the model’s ability to learn cross-channel contextual relationships. To address these challenges, we introduce a novel architecture strategy with channel shuffle as the core operation and the Multi-Scale Dynamic Channel Shuffle Attention (MS-DCSA) mechanism. Our proposed mechanism actively disrupts the fixed order of channels to achieve more effective cross-channel information exchange, promoting the model to learn richer global-local feature representations. Experiments on the Synapse, ACDC, CVC-ClinicDB, and ISIC-2017 datasets have shown that MS-DCSNet achieves excellent segmentation accuracy in various types of medical image segmentation with average Dice scores of 84.22%, 92.48%, 88.70%, and 96.45%. These results not only significantly surpass most existing methods, but also demonstrate the powerful segmentation ability and excellent generalization performance of MS-DCSNet.

This paper introduces the Constrained Monte Carlo Tree Search (CMCTS) framework to enhance the mathematical reasoning capabilities of Large Language Models (LLM). By incorporating a constrained action space, Process Reward Model (PRM), and partial-order rules, CMCTS effectively addresses the limitations of existing MCTS methods in terms of state space diversity and action selection rationality. Specifically, during the expansion phase, CMCTS restricts action sampling to a predefined constrained action set to increase candidate state diversity. In the simulation phase, it introduces partial-order rules and PRM to optimize action selection and prevent unreasonable state transitions. Experimental results show that CMCTS performs outstandingly across multiple mathematical reasoning benchmarks. Under a zero-shot setting, a 7B model achieves 83.4% average accuracy, surpassing the 72B baseline model by 4.8%. Ablation studies demonstrate that each component of the framework is crucial for performance improvement, and their combined use fully leverages their respective strengths. Overall, the CMCTS framework provides an effective approach to enhancing LLM mathematical reasoning capabilities, supported by theoretical analysis, and offers novel insights for future reasoning tasks.

Intelligent and autonomous networks require precise and fast mechanisms to minimize errors and ensure efficient operation. Modern methods are increasingly based on artificial intelligence, in particular on machine learning, to reliably process large amounts of data. While high-quality datasets are essential to train machine learning models, assessing the quality of datasets can be challenging and is often overlooked or underestimated. This paper proposes a novel method of permutation testing to assess one relevant dataset quality dimension in the context of binary or multiclass classification problems: the strength of the association between data and labels. The method described is called Permutation for Quality of Dataset Assessment (PerQoDA). In this paper we introduce the method, the statistics and visualisations necessary for proper interpretation of the results, and lastly, the theoretical justification of limits of performance. According to our experiments carried out on both simulated as well as real network datasets, the PerQoDA method can correctly estimate the strength of relationships in labelled datasets across a range of scenarios.

Ensuring the reliability and safety of railway operational equipment is crucial for efficient train operations. Existing fault diagnosis methods lack integration with structured domain knowledge, limiting their scalability, interpretability, and accuracy. We propose a model, namely BB-ROEFKG, which combines unstructured textual fault with structured knowledge graphs to enhance diagnostic reasoning and semantic understanding. This integrated framework extracts contextual semantics from fault descriptions and fuses them with structured representations that the domain-specific knowledge graph of railway equipment faults provides. The sequential modeling network processes the fused information and employs a mechanism to highlight key semantic features relevant to fault categorization. We evaluate the proposed model on a real-world dataset that we collected from a Chinese railway bureau. Experimental results show that BB-ROEFKG significantly outperforms traditional text-based models, achieving a 16.44% point improvement in Macro F1 score. In addition, ablation studies confirm the contribution of each component in the model. To support practical deployment, we have integrated BB-ROEFKG into a web-based diagnostic system, and operational staff are currently conducting trial use. The system enables railway personnel to input fault descriptions in natural language and receive timely and accurate diagnostic results.

The goal of recommendation is to provide users with personalized item suggestions. However, data-driven recommendations suffer from popularity bias caused by imbalanced user–item interactions; this problem becomes more severe in multi-behavior settings because popularity bias is both complex across behaviors and transmissible along behavior cascades, making it infeasible to simply apply single-behavior debiasing methods directly. To this end, we propose a novel popularity debiasing framework for multi-behavior recommendation via Causal Inference and Data Augmentation (CIDA), which consists of a Popularity-aware Noise Weighting (PNW) module and a Simulated Unpopular item Augmentation (SUA) module. To cope with the complexity of popularity bias, PNW adjusts item representations by applying controllable noise weights based on item popularity. Meanwhile, SUA introduces simulated nodes into the causal graph to block the direct causal effect between items and recommendation outcomes. Inspired by imitation learning, we design a minimum distance constraint in feature space between simulated and real unpopular items to maintain representational similarity. A feature fusion strategy is then used to generate unbiased item representations. To integrate the two modules, we design a noise augmentation contrastive learning task, enabling the unbiased item features to propagate effectively across multi-behavior chain and suppress popularity bias transmission. Extensive experiments on three real-world datasets validate the effectiveness of our framework and the rationality of its design.