Selective deletion of data from deep models, known as unlearning, has become crucial for enforcing the right to be forgotten, while also mitigating the negative impact of flawed training data. Retraining deep models is often impractical due to data access restrictions and computational overhead. Existing retraining-free methods are typically based on the Fisher Information Matrix (FIM), which quantifies the importance of model parameters with respect to forgetting classes, applying equal dampening to these parameters. This approach implicitly assumes a semantically uniform representation space, where all retained classes are equidistant from the forgetting classes. However, this assumption often fails in real-world cross-modal retrieval scenarios characterized by multi-label and non-orthogonal semantics. To overcome this limitation, we propose Prior-Prototype guided Partitioned dampening (PPP), an effective strategy for selective forgetting in cross-modal retrieval. First, PPP defines prior-prototypes, which are semantic centers derived from well-trained models, to identify neighbor classes semantically close to the forgetting set. Then, PPP uses Fisher information to identify parameters sensitive to forgetting and partitions them into buffer and core regions based on their relative importance to the neighbor and retained sets. Finally, PPP applies a hierarchical dampening strategy, where core parameters receive stronger suppression guided by prototype-based semantic disparities. Comprehensive evaluations on four large-scale benchmarks show that PPP performs competitively with retraining-based baselines, highlighting its effectiveness and generalizability in selective unlearning for cross-modal retrieval.

Existing deep learning-based spatiotemporal fusion (STF) methods for remote sensing images primarily denoise low-frequency features. They neglect the impact of noise in high frequency features, often resulting in fused images with poor visual quality and excessive smoothness. To address this, a two-stage STF network based on wavelet denoising and conditional diffusion enhancement is proposed. Firstly, a dualsparse attention denoising network is designed to process high and low frequency image features via high and low pass filters, respectively, enhancing detail representation. Subsequently, a conditional enhancement network is introduced, using the preliminary fusion results as conditional input to enhance features and iteratively denoise during diffusion stage, thereby improving noise robustness. Finally, Fourier-based frequency domain modulation module is introduced to balance the frequency distributions of the backbone and skip branches within the denoising network, achieving further noise reduction. Extensive experiments on CIA and SW benchmark datasets demonstrate that WDCDiff outperforms various state-of-the-art (SOTA) methods across multiple metrics.

Modern deep hashing methods rely on data augmentation to fully realize their potential in the face of large-scale retrieval galleries and over-parameterized visual models. However, this work observes that mainstream label-preserving augmentation methods are unreliable for information retrieval because they lead to incomplete alignment between data and labels. This misalignment impairs metric losses in distinguishing original/augmented data during same-class clustering, compromising nearest-neighbor search efficacy. To address these issues, we propose an innovative plug-and-play data augmentation framework tailored for retrieval tasks, based on the concept ofGaining robust features by randomlyGiving up parts of the image (GG). Inspired by the ease with which visual changes induced by discard transformations can be estimated, we design two intuitive augmentation methods along with corresponding semantic shift estimators to measure the semantic changes introduced by each operation. Additionally, we optimize the metric loss based on the semantic retention scores, guiding the metric objective to properly allocate gradients for generated samples. This adjustment mitigates the adverse effects caused by incomplete alignment, optimizing the intra-class distance of both original and augmented data in the Hamming space, while ensuring the relevance and accuracy of the retrieval results. Extensive experiments conducted on six datasets demonstrate the effectiveness and robustness of our proposed framework. Code is available athttps://github.com/wuhulahu/GG.

Accurate glioma MRI segmentation is pivotal for diagnosis, tumor characterization, grade detection, molecular subtyping, and treatment response assessment. However, existing models are often hindered by high computational costs and insufficient segmentation performance, limiting their clinical utility. To address these challenges, we propose a lightweight Mixed Vision Transformer (MViT) with an adaptive residual-enhanced network, named MViT-ARNet, for 3D glioma MRI segmentation. MViT-ARNet employs an encoder-decoder architecture. The encoder leverages a lightweight framework incorporating MViT modules, which effectively capture both local details and longrange dependencies. Crucially, it fuses features extracted by convolutional layers with those learned through spatial reduction attention in the Transformer, significantly reducing model parameters and enhancing segmentation accuracy. Within the decoder, we introduce an adaptive residual-enhanced upsampling module, which progressively reconstructs high-resolution segmentation maps by adaptively integrating skip features from the encoder. Evaluated on a mixed multimodal MRI dataset, the model demonstrates notable efficiency, requiring only 4.41 M parameters and 22.27 GFLOPs. It achieves robust segmentation performance, with Dice scores of 0.7632 for the enhancing tumor, 0.8229 for the tumor core, and 0.9022 for the whole tumor on the test set. These results demonstrate the potential of MViT-ARNet for accurate glioma delineation and its viability for real-world clinical applications.

Accurate prediction of drug side effect frequencies is critical for drug safety evaluation and clinical decision-making. Current methods primarily emphasize the associations between drugs and side effects, yet they often neglect the underlying structural and semantic features of both, which limits further advancements in prediction accuracy. In this study, we propose a novel multi-view fusion framework, MVFDSP, which integrates pre-trained molecular representation of 1D and 2D views with graph-based side effect information for side effect frequency prediction. Firstly, we obtain both 1D and 2D molecular representations from the pretrained molecular language model, and combine them using an adaptive fusion strategy. Subsequently, we construct a similarity network based on the side effect frequency matrix using K-Nearest Neighbors (KNN), and incorporate semantic embeddings derived from the terminology system of MedDRA to construct a side effect information graph. A multi-head graph attention network is then employed to capture the multi-dimensional information within this graph, allowing the model to attend to diverse aspects of the semantic and structural relationships among side effects. The final frequency prediction matrix is derived from the inner product between the learned drug and side effect embeddings. Experimental results on the SIDER 4.1 dataset demonstrate that MVFDSP outperforms existing methods, highlighting its effectiveness in capturing complex relationships of drugs and side effects. The code and data are available at https://github.com/Sonder-Echo/MVFDSP.

Deep hashing methods typically rely on high-quality feature embeddings to generate compact hash codes that preserve the original semantic information. However, in supervised learning, the feature extractor can only capture the prior distribution of the training set, which results in significant performance degradation when confronted with unknown categories. To address this challenge, we propose a Prior-Prototypes based feature learning framework (PP framework) for Few-Shot hashing. Specifically, we introduce a novel Prior-Prototypes encoder (PP encoder) that generates PP embeddings by computing the differences between new samples and Prior Prototypes, thereby avoiding direct reliance on feature representations from pre-trained extractors. Furthermore, to enhance the hashing function’s ability to capture subtle features, we design a Self-Consistency Diffusion Module (SCDM) that imposes self-consistency constraints during the hashing learning, and improved information exchange among samples. Extensive experiments on three benchmarks and four hashing objectives show that our method outperforms existing approaches.

The goal of the Image Text Multimodal Named Entity Recognition (MNER) task is to identify and classify entities such as individuals, organisations, and places across text and image modalities. Most of the current work focuses on improving the cross-modal alignment model, which has certain limitations: 1. In Chinese social media domain datasets, the mix of Chinese and English language leads to the difficulty of achieving accurate segmentation in the current named entity recognition research. 2. Only image features are regarded as auxiliary information, thus ignoring the impact of image feature processing on the MNER task. To address these issues, we propose a hierarchical subnetwork (HSN) based approach in this paper. Specifically, the visually enhanced MNER model is implemented by constructing subnetworks with flexible different levels. Textual features are extracted through a linguistic model, while visual features are derived from the VGG19 model embedded in SMFE. Features from both text and visual models are fused through two subnetworks dynamically constructed based on the data features to enhance the consistency of the representations. Finally, we use Conditional Random Fields (CRF) to sequence tag the output of the multi-modal interaction module. The proposed method achieves F1 scores of 88.05% and 80.24% on the CMNER dataset and the Wukong-Cmner dataset, respectively, demonstrating the competitiveness of our method.

State Space models (SSMs) like PointMamba provide efficient feature extraction for point cloud self-supervised learning with linear complexity, surpassing Transformers in computational efficiency. However, existing PointMamba-based methods rely on complex token ordering and random masking, disrupting spatial continuity and local semantic correlations. We propose \textbf{ZigzagPointMamba} to address these challenges. The key to our approach is a simple zigzag scan path that globally sequences point cloud tokens, enhancing spatial continuity by preserving the proximity of spatially adjacent point tokens. Yet, random masking impairs local semantic modeling in self-supervised learning. To overcome this, we introduce a Semantic-Siamese Masking Strategy (SMS), which masks semantically similar tokens to facilitate reconstruction by integrating local features of original and similar tokens, thus overcoming dependence on isolated local features and enabling robust global semantic modeling. Our pre-training ZigzagPointMamba weights significantly boost downstream tasks, achieving a 1.59\% mIoU gain on ShapeNetPart for part segmentation, a 0.4\% higher accuracy on ModelNet40 for classification, and 0.19\%, 1.22\%, and 0.72\% higher accuracies respectively for the classification tasks on the OBJ-BG, OBJ-ONLY, and PB-T50-RS subsets of ScanObjectNN. Code is available at https://github.com/Rabbitttttt218/ZigzagPointMamba.

Advanced persistent threats (APTs) have become a major network threat due to their persistence, complexity, and multi-stage nature. Existing APT detection methods based on provenance graphs have been proven to be effective. However, most studies based on provenance graphs are limited to single event detection, and rely on rule design and prior knowledge, making it difficult to effectively extract contextual information hidden in the graph. To address these issues, this paper innovatively proposes an APT detection model based on self-supervised contrastive learning, APT-SSC. First, the model uses an improved GraphSage(Graph Sample and Aggregate) encoder combined with a multi-head attention mechanism to learn the embedding representation of nodes; then, through context-based (local-global) contrastive learning, the robustness and context-awareness of node embeddings are effectively enhanced; finally, the optimized node embeddings are used as the input of the classification detection model to achieve node-level anomaly detection and tactics detection. Experiments are carried out on two public datasets, CICAPT-IIoT2024 (containing a variety of complex tactics) dataset and DAPRA TC-E3 dataset. The experimental results show that APT-SSC improves F1-Scores by 5% and 7% in anomaly detection and tactic detection experiments, respectively. APT-SSC has been proven to be significantly superior to existing methods in APT detection and can also effectively identify APT tactics.

Accurate and efficient diagnosis of breast cancer is an important research topic of computer-aided diagnosis. However, due to the lack of breast cancer histopathological image data and the fact that the histopathological images themselves have problems such as cell overlap and uneven color distribution, accurate classification and efficient feature extraction of the breast cancer histopathological images remains still a challenge. To tackle this challenge, this paper proposes a multi-stage transfer learning method (MT) which conducts two sequential transfer learning stages using the general images and the breast cancer medical images, respectively, and then the low-level features of the general images and the high-level features of the breast cancer medical images are fused in the final stage to improve the classification performance. The method was tested on the BreakHis dataset with the images divided into four subclasses (for the benign and the malignant separately) and eight subclasses (the benign and the malignant all together). A comparison of the MT method with some other state-of-the-art methods showed that for the 4 benign subclasses, the MT method achieved accuracy of 99.20%, 98.68%, 96.01%, and 98.24% in 40×, 100×, 200×, and 400× magnification factors, respectively. For the 4 malignant subclasses, the MT method achieved accuracy of 99.37%, 99.12%, 98.74%, and 98.14% in 40×, 100×, 200×, and 400× magnification factors, respectively. Finally, For all the eight subclasses, the MT method achieved accuracy of 97.49%, 96.04%, 96.21%, and 95.05% in 40×, 100×, 200×, and 400× magnification factors, respectively. The experimental results demonstrated the MT method can provide an effective means in the breast cancer classification.

The deployment of Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surfaces (STAR-RIS) has proven to be an effective means to extend coverage and improve wireless signal quality. STAR-RIS in wireless networks for aided Unmanned Aerial Vehicle (UAV) communications enables a significant boost in network capacity and the provision of virtual line-of-sight links to efficiently meet the quality-of-service (QoS) requirements of user equipment (UE). Accordingly, this paper proposes a novel STAR-RIS-aided multi-UAV communication framework to exploit energy efficiency and total throughput maximally. We formulate the long-term optimization problem as a decentralized, partially observed Markov decision process (DEC-POMDP). Then, we formulate the discrete association scheduling problem as a non-cooperative theoretical game and propose the UA-CFG algorithm to realize the UE association scheme that converges to a Nash equilibrium (NE). Then, a multi-agent reinforcement learning (MARL) method with well-established robustness is devised to continuously optimize the trajectories and energetic consumption of UAVs through centralized training and distributed implementation. Experimental results reveal that the performance of the proposed algorithm is considerable compared to other traditional schemes.

The accelerated non-agriculturalization of cropland has increasingly highlighted the importance of remote sensing (RS) change detection (CD) for monitoring land-use transitions. However, variations in RS imaging conditions and irregular cropland changes often result in noisy or inaccurate change maps. To address these challenges, we propose a novel deep learning framework named change-aware and fourier feature exchange Network (CAFENet). The method introduces a dedicated change-aware (CA) branch to extract discriminative change cues from pseudo-video sequences and integrates them into the backbone network. A fourier feature exchange module (FFEM) is designed to reduce brightness, color, and style discrepancies between bitemporal images, thereby enhancing robustness under varying acquisition conditions. Fused features are further refined using an efficient multi-scale attention mechanism (EMSA) to capture rich spatial details. In the decoding stage, a dynamic content-aware upsampling module (DCAU), together with skip connections, progressively recovers spatial resolution while preserving structural information. Experimental results on three datasets—CLCD, SW-CLCD, and LuojiaSET-CLCD—demonstrate that CAFENet achieves superior performance over state-of-the-art methods in terms of both accuracy and robustness, particularly in complex agricultural landscapes.

In recent years, methods leveraging global context modeling for continuous road extraction from remote sensing images have garnered significant attention. These methods can effectively address the issues of road discontinuity and missed detection caused by complex background interference, but they generally suffer from high computational complexity. Therefore, we propose CGCNet, a simple yet effective road extraction network based on global context modeling. CGCNet adopts an encoder–decoder architecture with a compact global context-aware block (CGCB) embedded in the center part to capture long-range dependencies among road segments. This block effectively enhances the model’s global modeling capability and significantly reduces computational complexity using a compact representation of the embedded Gaussian nonlocal block (NLB). Furthermore, we introduce SW-XJU, the first remote sensing road dataset constructed explicitly for the unique landscapes of Western China. Extensive experiments show that CGCNet achieves a superior balance between efficiency and accuracy compared with the state-of-the-art methods, and the extracted roads exhibit stronger connectivity. The source code and dataset are available athttps://github.com/LPeng625/CGCNet.

Exploring gene-drug associations is a key step in identifying new drug candidates, but traditional experimental methods are often expensive and time-consuming. While Graph Neural Network (GNN)-based models have demonstrated effectiveness in association prediction tasks, they face challenges in the information aggregation process. Existing GNN models either treat all nodes uniformly or rely on simple attention mechanisms to assign weights to neighboring nodes, limiting their capacity to capture complex relationships and improve performance. To address these limitations, we propose a novel adaptive graph diffusion network, ADiffGDA, for gene-drug association prediction. The model begins by randomly initializing embeddings for genes and drugs, which are then updated through neighborhood information aggregation. A key feature of ADiffGDA is the incorporation of a heat kernel, enabling each node to dynamically adjust aggregation weights based on its local structure. This approach allows the model to better capture variations in node types and local neighborhood patterns. Through extensive comparative experiments, we show that ADiffGDA outperforms existing state-of-the-art methods. Furthermore, case studies validate its effectiveness as a predictive tool, offering valuable insights for future biological experiments. The code and datasets for ADiffGDA are freely available at https://github.com/one-melon/ADiffGDA.

Exploring the intricate connections between non-coding RNAs (ncRNAs) and drug resistance is crucial for understanding the molecular mechanisms behind drug resistance, identifying novel drug development targets, and uncovering key biomarkers to optimize therapeutic strategies. Traditional biological assays face significant challenges, including high costs and lengthy timelines, prompting the need for advanced computational methods to predict ncRNA-drug resistance associations. In this study, we introduce HGTRDA, a novel computational framework designed to predict potential associations between ncRNAs and drug resistance. HGTRDA leverages LightGCN to generate node representations that capture topological information from the surrounding node neighborhood. These representations are then dynamically optimized using a global hypergraph transformer to model the relationships between ncRNAs and drug resistance. To enhance the quality of the learned embeddings, HGTRDA employs self-supervised learning to fine-tune topology-aware embeddings, reducing the impact of noise and improving representation quality. The final association scores between ncRNAs and drugs are computed using an inner product method. Empirical evaluations on the ncRNADrug database demonstrate that HGTRDA outperforms six contemporary state-of-the-art methods in predicting ncRNA-drug resistance associations. Furthermore, case studies illustrate the practical utility of HGTRDA as a predictive tool in real-world scenarios. The code and dataset for HGTRDA are freely available at https://github.com/one-melon/HGTRDA.

Over the past few years, there have been significant advancements in deep learning technology, leading to remarkable progress in the field of image analysis. However, when it comes to handling complex remote sensing images, current semantic segmentation methods still face challenges and do not perform as well as desired. How to obtain both spatial detail information and semantic information at the same time is an urgent problem to be solved. This letter proposes a context fusion network based on boundary guidance (BGFNet), which incorporates the patch attention module (PAM), the feature maps are enriched with contextual information, improving their ability to capture spatial dependencies. In order to alleviate boundary ambiguity, a boundary guidance module (BGM) is used to weight features with rich semantic boundary information. Furthermore, the compatible fusion module (CFM) is employed to merge high-order and low-order features, creating novel features. Channel attention is then applied to the obtained features allows us to select the desired features by filtering out irrelevant information. We validate our model on the Vaihingen and Potsdam datasets reached 81.65% and 86.94% mean intersection over union (mIoU), respectively, indicating the superiority of the proposed model.

Coordinating multiple agents with diverse tasks and changing goals without interference is a challenge. Multi-Agent Reinforcement Learning (MARL) aims to develop effective communication and joint policies using group learning. Some of the previous approaches required each agent to maintain a set of networks independently, resulting in no consideration of interactions. Joint communication work causes agents receiving information unrelated to their own tasks. Currently, agents with different task divisions are often grouped by action tendency, but this can lead to poor dynamic grouping. This paper presents a two-phase solution for multiple agents, addressing these issues. The first phase develops heterogeneous agent communication joint policies using a Group Communication MARL framework (GCMA). The framework employs a periodic grouping strategy, reducing exploration and communication redundancy by dynamically assigning agent group hidden features through hyper-network and graph communication. The scheme efficiently utilizes resources for adapting to multiple similar tasks. In the second phase, each agent's policy network is distilled into a generalized simple network, adapting to similar tasks with varying quantities and sizes. GCMA is tested in complex environments like StarCraft II and UAV take-off, showing its well-performing for large-scale, coordinated tasks. It shows GCMA's effectiveness for solid generalization in multi-task tests with simulated pedestrians.

The outbreak of COVID-19 has caused a serious impact on the traditional logistics industry. Considering that the truck-drone collaborative delivery system can both reduce the risk of COVID-19 propagation and deliver supplies in a cost effective and timely manner, this paper introduces the Multiple visits Travelling Salesman Problem with Multiple Heterogeneous Drones (MTSP-MHD). The model allows a truck to carry a fleet of heterogeneous multi-visit drones for cooperative deliveries, where the drones are capable of delivering to multiple customers on a single route and the flight is restricted by energy consumption and payload constraints. To solve MTSP-MHD, we develop an approach that combines K-Means$++$clustering, Nearest neighbor search and Greedy strategies (KNG) to construct feasible solutions. Meanwhile, an Improved Artificial Bee Colony algorithm combining Metropolis acceptance criterion of Simulated Annealing, Tabu list of Tabu Search, and Elite selection strategies (IABC-MTE) is proposed to enhance the quality of solutions. Particularly, three problem-specific neighborhood operators are adopted to search for new solutions. The massive experimental results indicate that IABC-MTE achieves significant improvements over other competitors, with average objective value reductions ranging from 1.81% to 29.16% and standard deviations reduced by 0.04 to 26.44. Finally, the influencing factors of the drone fleet, the performance of different drone fleets and delivery modes are evaluated in detail.

Unmanned Aerial Vehicles (UAVs) can be employed as short-term aerial base stations or as access points for User Equipments (UEs) to communicate with other UEs effectively. However, communication links may be obstructed by buildings, leading to poor data transfer performance and significant energy consumption. Deploying Reconfigurable Intelligent Surfaces (RIS) as part of the UAV-assisted communication system proves to be an effective means to avoid building obstructions and enhance wireless information quality. However, the complexity of communication relationships in multi-UAV systems with RIS-aided communication poses a significant challenge in energy reduction. Therefore, this study investigates a new RIS-aided multi-UAV communication framework for edge computing systems. The system aims to meet the quality-of-service (QoS) for UEs while minimizing the total energy consumption. To optimize the total energy consumption of RIS-aided multi-UAV communication, the impact of communication between multiple UAVs and differences between UE clusters on that system’s performance is also considered. We introduce a Stackelberg game to deal with the communication relationship between multiple UAVs and design a K-means-based clustering algorithm to segment UEs periodically. A model-free deep reinforcement learning algorithm grounded in maximum entropy is proposed to jointly optimize UAV trajectory design, phase shift control, and power allocation to reduce energy consumption further. Experimental results indicate that the system proposed performs favorably concerning both energy consumption and throughput.

Achieving optimal protein solubility is pivotal for efficient high-throughput purification, especially in industrial settings. However, conventional experimental techniques for assessing protein solubility in such contexts are not only costly but also time-intensive. Currently, numerous methods are available for predicting protein solubility, yet their effectiveness remains limited. Most of these approaches are predominantly sequence-based, failing to harness the invaluable structural insights inherent in proteins. Addressing these limitations, we introduce PPSol, an innovative protein solubility prediction methodology. Operating on protein sequences, PPSol employs ESM2 to predict protein contact maps, forming the basis for constructing protein graphs. Subsequently, well-established techniques are employed to predict protein feature representations as node features, including the utilization of the Position-Specific Scoring Matrix (PSSM). The resulting graph is fed into a graph convolutional neural network (GCN), enabling the acquisition of spatial structural information from proteins. Concurrently, ESM2-generated features undergo dimensional reduction via fully connected layers, integrating into every layer of the GCN for precise protein solubility prediction. Our approach excels through the fusion of pre-trained protein language models and GCNs, surpassing existing methodologies. Notably, PPSol attains state-of-the-art performance, showcasing a remarkable 2.8% enhancement in AUROC performance com-pared to prior strategies.

Discovering associations between circular RNAs (circRNAs) and cellular drug sensitivity is essential for understanding drug efficacy and therapeutic resistance. Traditional experimental methods to verify such associations are costly and time-consuming. Thus, the development of efficient computational methods for predicting circRNA-drug associations is crucial. In this study, we introduce a novel computational predictor called HETACDA, aimed at predicting potential circRNA-drug sensitivity associations. HETACDA constructs a heterogeneous graph network, incorporating the characteristic structure graph of drugs and circRNAs, along with the circRNA-drug sensitivity topology graph. By employing a graph convolutional network, the drug embedding vector is computed from the molecular structure of the drug. Through a graph attention mechanism, HETACDA assigns distinct attention weights to nodes in order to emphasize the contribution of various neighborhood nodes to the central node. Subsequently, an association score between circRNA-drug sensitivity is predicted using a three-layer fully connected neural network. Extensive experimental comparisons against several state-of-the-art methods highlight the effectiveness of our proposed framework. The availability of our source code and datasets on GitHub (https://github.com/xiaoxiaojun131421/HETACDA) facilitates replication and further research in this area.

tRNAs play a pivotal role in protein synthesis by transporting amino acids to the ribosome according to mRNA instructions. These molecules are essential regulators in various biological processes, and their dysregulation is closely linked to human diseases. Predicting associations between tRNAs and diseases is valuable for uncovering biomarkers that aid in disease prevention, detection, prognosis, diagnosis, and treatment. However, experimental validation of such associations is resource-intensive, necessitating the development of robust computational methods. In this study, we propose PTDA-SWGCL, a novel model for predicting potential tRNA-disease associations. PTDA-SWGCL integrates tRNA and disease similarity information derived from Gaussian kernel similarity, sequence similarity, and semantic similarity. It initializes tRNA and disease embeddings using this similarity information and refines them through supplementarily weight and graph comparison learning training on the tRNA-disease association graph. The final association pair prediction is obtained by the inner product of the tRNA and disease embeddings. Experimental results demonstrate that PTDA-SWGCL outperforms state-of-the-art methods. Case studies confirm its effectiveness in predicting tRNA-disease associations. The code and data are available at https://github.com/ZYPssss/PTDA-SWGCL.

Most object detection methods have achieved impressive performance on several public benchmarks, instead, facing underwater detection tasks, it is challenging to detect marine targets because of the inherent illumination inhomogeneity in underwater images. Moreover, the imbalanced foreground-background proposals further aggravate the situation of capturing marine organisms. To address the problems, we analyze the deficiency of existing feature pyramid structures and propose a multi-depth and multi-breadth pyramid architecture named Refined Pyramid Grid (RPG). A Harmonizing Focal Loss (HFL) is then proposed to generalize the discrete labels in focal loss to the continuous version to improve the optimization. Experimental results on the real-world datasets have demonstrated the efficiency and reliability of the proposed framework regarding underwater object detection tasks.

Abstract Infrared target detection has strong anti‐interference ability, long working distance and can work day and night. So it is widely used in military security and transportation fields, and infrared road object detection is critical in traffic checkpoints and autonomous driving. However, the target scale in infrared images changes greatly, small targets are difficult to detect, the poor image quality and low signal‐to‐noise ratio are still huge challenges in infrared target detection. This paper proposes an improved infrared target detection model I‐CenterNet based on the anchor‐free model CenterNet. The EfficientNetV2 with the channel attention mechanism is used instead of the traditional structure as the backbone network to enhance feature extraction. In order to reduce the noise of the input infrared image, Dilated‐Residual U‐net (DRUNet) is used. Meanwhile, feature pyramid and Sub‐Pixel are combined for multi‐scale feature fusion. Data enhancement is implemented to improve model performance. The experimental results show that the average detection accuracy of this model on the Flir infrared data set is 87.9%, and the average detection speed reaches 14.2 frames/s.

A large number of experimental studies have shown that circRNAs can act as molecular sponges of microRNAs, interacting with miRNAs to regulate gene expression levels, thereby affecting the development of human diseases. Exploring the potential associations between circRNAs and miRNAs can help understand complex disease mechanisms. Considering that biological experiments are time-consuming and labor-intensive, this study proposes a computational model using a graph neural network and singular value decomposition (CMASG) for circRNA-miRNA association prediction. Specifically, graph neural networks are used to learn nonlinear feature representations of nodes, followed by matrix factorization algorithms to learn linear feature representations of nodes, and then combined feature representations learned from different perspectives. Finally, the lightGBM algorithm was used for circRNA-miRNA association prediction. The proposed CMASG model achieved an AUC value of 0.8804. The experimental results demonstrate the superiority and effectiveness of the CMASG model in predicting circRNA-miRNA association tasks.

Single image super-resolution (SISR) is a challenging ill-posed problem due to multiple high-resolution (HR) images can degenerate into the same low-resolution (LR) image. However, existing deep learning-based super-resolution (SR) methods always have blurred edge structures in the restored images. In addition, they mainly build more profound and more complex convolutional neural networks (CNN), which leads to substantial computational overhead. To address these issues, we propose the feature fusion super-resolution network (FFSRN) that uses the gradient map of the image to guide the restoration. In FFSRN, we propose the split and shuffle concat block (SSCB), which can extract rich features while controlling the model size and computational effort. We also introduce gradient branching to provide additional structural priors for the reconstruction process to restore high-resolution gradient mapping. Experimental results show that this method has a better peak signal-to-noise ratio, computational overhead and visual quality than the existing super-resolution algorithms. Code is available at https://github.com/Qyzs/FFSRN.

BACKGROUND: Circular RNAs (circRNAs) play essential roles in cancer development and therapy resistance. Many studies have shown that circRNA is closely related to human health. The expression of circRNAs also affects the sensitivity of cells to drugs, thereby significantly affecting the efficacy of drugs. However, traditional biological experiments are time-consuming and expensive to validate drug-related circRNAs. Therefore, it is an important and urgent task to develop an effective computational method for predicting unknown circRNA-drug associations. RESULTS: In this work, we propose a computational framework (GATECDA) based on graph attention auto-encoder to predict circRNA-drug sensitivity associations. In GATECDA, we leverage multiple databases, containing the sequences of host genes of circRNAs, the structure of drugs, and circRNA-drug sensitivity associations. Based on the data, GATECDA employs Graph attention auto-encoder (GATE) to extract the low-dimensional representation of circRNA/drug, effectively retaining critical information in sparse high-dimensional features and realizing the effective fusion of nodes' neighborhood information. Experimental results indicate that GATECDA achieves an average AUC of 89.18% under 10-fold cross-validation. Case studies further show the excellent performance of GATECDA. CONCLUSIONS: Many experimental results and case studies show that our proposed GATECDA method can effectively predict the circRNA-drug sensitivity associations.

In the era of social networks, the rapid growth of data mining in information retrieval and natural language processing makes automatic text summarization necessary. Currently, pretrained word embedding and sequence to sequence models can be effectively adapted in social network summarization to extract significant information with strong encoding capability. However, how to tackle the long text dependence and utilize the latent topic mapping has become an increasingly crucial challenge for these models. In this article, we propose a topic-aware extractive and abstractive summarization model named T-BERTSum, based on Bidirectional Encoder Representations from Transformers (BERTs). This is an improvement over previous models, in which the proposed approach can simultaneously infer topics and generate summarization from social texts. First, the encoded latent topic representation, through the neural topic model (NTM), is matched with the embedded representation of BERT, to guide the generation with the topic. Second, the long-term dependencies are learned through the transformer network to jointly explore topic inference and text summarization in an end-to-end manner. Third, the long short-term memory (LSTM) network layers are stacked on the extractive model to capture sequence timing information, and the effective information is further filtered on the abstractive model through a gated network. In addition, a two-stage extractive–abstractive model is constructed to share the information. Compared with the previous work, the proposed model T-BERTSum focuses on pretrained external knowledge and topic mining to capture more accurate contextual representations. Experimental results on the CNN/Daily mail and XSum datasets demonstrate that our proposed model achieves new state-of-the-art results while generating consistent topics compared with the most advanced method.

Piwi-interacting RNAs (piRNAs) are a novel class of small non-coding RNAs that interact with the PIWI protein family and are associated with various diseases. Identification of piRNA-disease associations provides candidate piRNA targets for disease treatment and provides promising candidate molecular targets to promote the drug design. However, detection of piRNA-disease associations by biological experiments is often high cost and time-consuming. Therefore, there is an urgent need for reliable computational methods for piRNA-disease associations identification. In this study, we proposed a new computational predictor named iPiDA-GBNN to predict potential piRNA-disease associations. The iPiDA-GBNN presented the piRNA-disease pairs by combining disease similarity information and piRNA similarity information. Disease similarity information includes disease GIP kernel similarity, disease Jaccard similarity, disease semantic similarity, and piRNA similarity includes sequence similarity and GIP kernel similarity. The Stacked auto-encoder(SAE) was then performed on piRNA features to extract the key features. The training datasets consisted of experiments that confirmed positive associations and the same quantity negative associations from the unknown pairs. Finally, the Gradient Boosting Neural Networks(GrownNet) [20] trained with the known piRNA-disease associations and the negative associations to predict new piRNA-disease associations. The experimental results showed that iPiDA-GBNN achieved superior predictive ability compared with other state-of-art predictors. The source code and datasets explored in this work are available at https://github.com/Tracyhuahua666/iPiDA-GBNN.

A large amount of evidence shows that circular RNAs(circRNAs) participate in transcription and translation regulation and function as “micro RNA(miRNA)-sponges”. Recognizing circRNA-miRNA interaction is helpful to understand the function of circRNAs, especially its role in complex diseases. Obtaining interactive information based on traditional biological experiments is usually small-scale, time-consuming, and laborious. Considering that there are few calculation methods, it is urgent to develop efficient and accurate methods to extract the interaction between circRNA and miRNA. In this work, we proposed a computational framework called CMIVGSD, which uses singular value decomposition and graph variational auto-encoders to predict circRNA-miRNA interaction. To our best knowledge, CMIVGSD is the first calculation framework to predict circRNA-miRNA interaction. CMIVGSD uses the singular value decomposition (SVD) algorithm to obtain linear features from the circRNA-miRNA interaction matrix. We have constructed the similarity networks of circRNA and miRNA, respectively. The graph variational auto-encoder (VGAE) is employed to mine the non-linear features of circRNA-miRNA in similarity networks. Finally, we combine linear and non-linear features and use LightGBM to predict interaction scores. We performed five-fold cross-validation experiments. Experimental results show that our proposed method is better than other methods. The case study further proves the effectiveness of CMIVGSD in predicting circRNA-miRNA interaction.

Many studies have shown that circRNAs play essential roles in various biological processes. With the development of technology, the associations between circRNA and diseases have been discovered, and these associations will help diagnose and treat diseases. However, it is time-consuming and costly to detect the associations between circRNAs and diseases with the experimental methods. Therefore, it is necessary to develop a feasible and effective computational method for predicting circRNA-disease associations. In this paper, we propose a new computational framework called VLCDA to identify the potential circRNA-disease associations. Initially, we construct features by fusing circRNA expression profile features and circRNA protein-coding ability features, disease semantic features, circRNA and disease GIP Kernel features, and use VGAE to mine its deep latent features. Finally, we use the fusion features to train the LightGBM classifier and the trained LightGBM to identify the circRNA-disease associations. The main contribution of VLCDA is that we firstly add circRNA protein-coding ability feature to the circRNA-disease association prediction model. In addition, VLCDA uses variational graph auto-encoders to extract the latent features of circRNA-disease associations to improve the prediction model’s accuracy further. VLCDA obtained the area under the ROC curve (AUC) scores of 0.9783 in 5-fold cross-validation. In addition, in the case studies, 16 of the top 20 circRNA-disease associations predicted by VLCDA have been confirmed by relevant literature.

The fact that the existing traffic sign images are easily affected by external factors, and the traffic signs are generally small targets on the images at different scales, has made it difficult in feature extraction when doing traffic sign detection. To achieve better detection results, a multi-target traffic sign detection method with channel attention and feature fusion network (CAFFNet in short) is proposed. This method effectively learns the correlation between feature channels through a lightweight channel attention network, realizes local cross-channel interaction without dimensionality reduction, and enhances the representation ability of the network. The feature pyramid network is used to achieve feature fusion and generate high-resolution multiscale semantic information. The dilated convolution is utilized to capture the multiscale context information to narrow the difference between features and improve the detection effect of the model. The experimental results show that the proposed method on the two datasets GTSDB and CTSD has achieved superior performance in the evaluation criteria compared with the existing detection algorithms.

Aerial image-based target detection has problems such as low accuracy in multiscale target detection situations, slow detection speed, missed targets and falsely detected targets. To solve this problem, this paper proposes a detection algorithm based on the improved You Only Look Once (YOLO)v3 network architecture from the perspective of model efficiency and applies it to multiscale image-based target detection. First, the K-means clustering algorithm is used to cluster an aerial dataset and optimize the anchor frame parameters of the network to improve the effectiveness of target detection. Second, the feature extraction method of the algorithm is improved, and a feature fusion method is used to establish a multiscale (large-, medium-, and small-scale) prediction layer, which mitigates the problem of small target information loss in deep networks and improves the detection accuracy of the algorithm. Finally, label regularization processing is performed on the predicted value, the generalized intersection over union (GIoU) is used as the bounding box regression loss function, and the focal loss function is integrated into the bounding box confidence loss function, which not only improves the target detection accuracy but also effectively reduces the false detection rate and missed target rate of the algorithm. An experimental comparison on the RSOD and NWPU VHR-10 aerial datasets shows that the detection effect of high-efficiency YOLO (HE-YOLO) is significantly improved compared with that of YOLOv3, and the average detection accuracies are increased by 8.92% and 7.79% on the two datasets, respectively. The algorithm not only shows better detection performance for multiscale targets but also reduces the missed target rate and false detection rate and has good robustness and generalizability.

Summary With the rapid increase in e‐commerce data, recommender systems (RSs) have become the most prevalent methods for providing recommended services in various commercial platforms. Deep learning–based recommender methods improve recommendation results by learning latent representations; however, most cannot capture the correlations between items and ignore additional information such as time information, which leads to suboptimal suggestions. To improve recommendation accuracy, we propose a novel recommender algorithm based on graph embedding and diffusion sampling (graph2vec). Our improved model constructs a graph based on users' behavior histories and embeds the graph to a low‐dimensional vector space with a deep learning approach. To obtain more accurate embedding results, we use a revised sampling method based on information diffusion theory to capture both the depth and breadth information of a graph. Then, we recommend the top‐N items to the target user depending on the final representation vectors. Experiments are carried out with real‐world datasets to demonstrate the superior performance of graph2vec. The results show that browse‐based graph construction and diffuse‐based graph embedding help improve the recommender accuracy of the new model compared with that of the selected state‐of‐the‐art models.

To the problem that random placement in distributed file system leads to the low utilization of servers, this article models according to the characteristics of remote sensing image data blocks, clusters data by setting storage centre with access frequency so that the adjacent remote sensing image data blocks in spatial position are near each other in physical storage as well. It promotes the response speed of the system, places data blocks according to data block groups, reshuffle the non-grouped data blocks at the system low load and turns off dispensable datanodes to achieve energy saving. The experiments suggest that in the inquiry of remote sensing image data, it is more efficient to place data blocks according to their own characteristics than random placement. Compared with common dynamic data placement strategies, this strategy performs better in energy saving when the system is at moderate load.