International Journal of Machine Learning and Cybernetics最新文献

Label distribution learning is an effective learning approach for addressing label polysemy in the field of machine learning. In contrast to multi-label learning, label distribution learning can accurately represent the relative importance of labels and has richer semantic information about labels. Presently label distribution learning algorithms frequently integrate label correlation into their models to narrow down the assumption space of the model. However, existing label distribution learning works on label correlation use one-to-one or many-to-one correlation which has limitations in representing more complex correlation relationships. To address this issue, we attempt to extend the existing correlation relationships to many-to-many relationships. Specifically, we first construct a many-to-many correlation mining framework based on self-representation. Then by using the learned many-to-many correlation, a label distribution learning algorithm is designed. Our algorithm achieved the best performance in (78.21%) of cases across all datasets and all performance metrics with the algorithm having the best average ranking. It also demonstrated statistical superiority compared to the comparison algorithms in pairwise two-tailed t-tests. This paper introduces a novel approach to representing and applying label correlations in label distribution learning. The exploitation of this new many-to-many correlation can enhance the representational capabilities of label distribution learning models.

To improve the speeds of the traditional nuclear norm minimization methods, a fast tri-factorization method (FTF) was recently proposed for matrix completion, and it received widespread attention in the fields of machine learning, image processing and signal processing. However, its low convergence accuracy became increasingly obvious, limiting its further application. To enhance the accuracy of FTF, a generalized tri-factorization method (GTF) is proposed in this paper. In GTF, the nuclear norm minimization model of FTF is improved to a novel ({{varvec{L}}}_{1,{varvec{p}}})(0 < p < 2) norm minimization model that can be optimized very efficiently by using QR decomposition. Since the ({{varvec{L}}}_{1,{varvec{p}}}) norm is a tighter relaxation of the rank function than the nuclear norm, the GTF method is much more accurate than the traditional methods. The experimental results demonstrate that GTF is more accurate and faster than the state-of-the-art methods.

Accurate short-term load forecasting (STLF) is crucial for the power system. Traditional methods generally used signal decomposition techniques for feature extraction. However, these methods are limited in extrapolation performance, and the parameter of decomposition modes needs to be preset. To end this, this paper develops a novel STLF algorithm based on multi-scale perspective decomposition. The proposed algorithm adopts the multi-scale deep neural network (MscaleDNN) to decompose load series into low- and high-frequency components. Considering outliers of load series, this paper introduces the adaptive rescaled lncosh (ARlncosh) loss to fit the distribution of load data and improve the robustness. Furthermore, the attention mechanism (ATTN) extracts the correlations between different moments. In two power load data sets from Portugal and Australia, the proposed model generates competitive forecasting results.

Heterogeneous graph neural networks have attracted considerable attention for their proficiency in handling intricate heterogeneous structures. However, most existing methods model semantic relationships in heterogeneous graphs by manually defining meta-paths, inadvertently overlooking the inherent incompleteness of such graphs. To address this issue, we propose a multi-graph aggregated graph neural network (MGAGNN) for heterogeneous graph representation learning, which simultaneously leverages attribute similarity and high-order semantic information between nodes. Specifically, MGAGNN first employs the feature graph generator to generate a feature graph for completing the original graph structure. A semantic graph is then generated using a semantic graph generator, capturing higher-order semantic information through automatic meta-path learning. Finally, we aggregate the two candidate graphs to reconstruct a new heterogeneous graph and learn node embedding by graph convolutional networks. Extensive experiments on real-world datasets demonstrate the superior performance of the proposed method over state-of-the-art approaches.

The path planning for unmanned mobile robots has always been a crucial issue, especially in unknown environments. Reinforcement learning widely used in path planning due to its ability to learn from unknown environments. But, in unknown environments, deep reinforcement learning algorithms have problems such as long training time and instability. In this article, improvements have been made to the deep deterministic policy gradient algorithm (DDPG) to address the aforementioned issues. Firstly, the experience pool is divided into different experience pools based on the difference between adjacent states; Secondly, experience is collected from various experience pools in different proportions for training, enabling the robot to achieve good obstacle avoidance ability; Finally, by designing a guided reward function, the convergence speed of the algorithm has been improved, and the robot can find the target point faster. The algorithm has been tested in practice and simulation, and the results show that it can enable robots to complete path planning tasks in complex unknown environments.

Unmanned aerial vehicles (UAVs) are extensively applied in military, rescue operations, and traffic detection fields, resulting from their flexibility, low cost, and autonomous flight capabilities. However, due to the drone’s flight height and shooting angle, the objects in aerial images are smaller, denser, and more complex than those in general images, triggering an unsatisfactory target detection effect. In this paper, we propose a model for UAV detection called DoubleM-Net, which contains multi-scale spatial pyramid pooling-fast (MS-SPPF) and Multi-Path Adaptive Feature Pyramid Network (MPA-FPN). DoubleM-Net utilizes the MS-SPPF module to extract feature maps of multiple receptive field sizes. Then, the MPA-FPN module first fuses features from every two adjacent scales, followed by a level-by-level interactive fusion of features. First, using the backbone network as the feature extractor, multiple feature maps of different scale ranges are extracted from the input image. Second, the MS-SPPF uses different pooled kernels to repeat multiple pooled operations at various scales to achieve rich multi-perceptive field features. Finally, the MPA-FPN module first incorporates semantic information between each adjacent two-scale layer. The top-level features are then passed back to the bottom level-by-level, and the underlying features are enhanced, enabling interaction and integration of features at different scales. The experimental results show that the mAP50-95 ratio of DoubleM-Net on the VisDrone dataset is 27.5%, and that of Doublem-Net on the DroneVehicle dataset in RGB and Infrared mode is 55.0% and 60.4%, respectively. Our model demonstrates excellent performance in air-to-ground image detection tasks, with exceptional results in detecting small objects.

Skin lesion segmentation is a fundamental task in the field of medical image analysis. Deep learning approaches have become essential tools for segmenting medical images, as their accuracy in effectively analyzing abnormalities plays a critical role in determining the ultimate diagnostic results. Because of the inherent difficulties presented by medical images, including variations in shapes and sizes, along with the indistinct boundaries between lesions and the surrounding backgrounds, certain conventional algorithms face difficulties in fulfilling the growing requirements for elevated accuracy in processing medical images. To enhance the performance in capturing edge features and fine details of lesion processing, this paper presents the Boundary-Prior-Guided Multi-Scale Aggregation Network for skin lesion segmentation (BGMAN). The proposed BGMAN follows a basic Encoder–Decoder structure, wherein the encoder network employs prevalent CNN-based architectures to capture semantic information. We propose the Transformer Bridge Block (TBB) and employ it to enhance multi-scale features captured by the encoder. The TBB strengthens the intensity of weak feature information, establishing long-distance relationships between feature information. In order to augment BGMAN’s capability to identify boundaries, a boundary-guided decoder is designed, utilizing the Boundary Aware Block (BAB) and Cross Scale Fusion Block (CSFB) to guide the decoding learning process. BAB can acquire features embedded with explicit boundary information under the supervision of a boundary mask, while CSFB aggregates boundary features from different scales using learnable embeddings. The proposed method has been validated on the ISIC2016, ISIC2017, and (PH^2) datasets. It outperforms current mainstream networks with the following results: F1 92.99 and IoU 87.71 on ISIC2016, F1 86.42 and IoU 78.34 on ISIC2017, and F1 94.83 and IoU 90.26 on (PH^2).

This paper aims to provide a novel design of a multiscale framelet convolution for spectral graph neural networks (GNNs). While current spectral methods excel in various graph learning tasks, they often lack the flexibility to adapt to noisy, incomplete, or perturbed graph signals, making them fragile in such conditions. Our newly proposed framelet convolution addresses these limitations by decomposing graph data into low-pass and high-pass spectra through a finely-tuned multiscale approach. Our approach directly designs filtering functions within the spectral domain, allowing for precise control over the spectral components. The proposed design excels in filtering out unwanted spectral information and significantly reduces the adverse effects of noisy graph signals. Our approach not only enhances the robustness of GNNs but also preserves crucial graph features and structures. Through extensive experiments on diverse, real-world graph datasets, we demonstrate that our framelet convolution achieves superior performance in node classification tasks. It exhibits remarkable resilience to noisy data and adversarial attacks, highlighting its potential as a robust solution for real-world graph applications. This advancement opens new avenues for more adaptive and reliable spectral GNN architectures.

The dual-mode 24/7 monitoring systems continuously obtain visible and infrared images in a real scene. However, differences such as color and texture between these cross-modality images pose challenges for visible-infrared person re-identification (ReID). Currently, the general method is modality-shared feature learning or modal-specific information compensation based on style transfer, but the modality differences often result in the inevitable loss of valuable feature information in the training process. To address this issue, A complementary feature fusion and identity consistency learning (CFF-ICL) method is proposed. On the one hand, the multiple feature fusion mechanism based on cross attention is used to promote the features extracted by the two groups of networks in the same modality image to show a more obvious complementary relationship to improve the comprehensiveness of feature information. On the other hand, the designed collaborative adversarial mechanism between dual discriminators and feature extraction network is designed to remove the modality differences, and then construct the identity consistency between visible and infrared images. Experimental results by testing on SYSU-MM01 and RegDB datasets verify the method’s effectiveness and superiority.

Measuring the semantic similarity between two texts is a fundamental aspect of text semantic matching. Each word in the texts holds a weighted meaning, and it is essential for the model to effectively capture the most crucial knowledge. However, current text matching methods based on BERT have limitations in acquiring professional domain knowledge. BERT requires extensive domain-specific training data to perform well in specialized fields such as medicine, where obtaining labeled data is challenging. In addition, current text matching models that inject domain knowledge often rely on creating new training tasks to fine-tune the model, which is time-consuming. Although existing works have directly injected domain knowledge into BERT through similarity matrices, they struggle to handle the challenge of small sample sizes in professional fields. Contrastive learning trains a representation learning model by generating instances that exhibit either similarity or dissimilarity, so that a more general representation can be learned with a small number of samples. In this paper, we propose to directly integrate the word similarity matrix into BERT’s multi-head attention mechanism under a contrastive learning framework to align similar words during training. Furthermore, in the context of Chinese medical applications, we propose an entity MASK approach to enhance the understanding of medical terms by pre-trained models. The proposed method helps BERT acquire domain knowledge to better learn text representations in professional fields. Extensive experimental results have shown that the algorithm significantly improves the performance of the text matching model, especially when training data is limited.