With the widespread adoption of microservice architectures, software systems have become increasingly complex and dynamic, leading to more diverse types of anomalies. Traditional single-modal anomaly detection methods are no longer sufficient for accurately identifying system anomalies and locating their root causes. To address these challenges, this paper proposes GRAFlow, an anomaly detection and root cause localization method based on cross-modal feature fusion and multi-scale graph attention networks. GRAFlow integrates three key modalities-logs, performance metrics, and traces-and employs a cross-modal attention mechanism to dynamically model the semantic interactions among these heterogeneous data sources, thereby enhancing detection accuracy and robustness. Additionally, a multi-scale graph attention network is introduced to capture both local dependencies and long-range propagation paths among microservices, enabling more comprehensive system state modeling. The model jointly optimizes both anomaly detection and root cause localization tasks. Experimental results on two real-world microservice datasets, TrainTicket and SocialNetwork, demonstrate that GRAFlow significantly outperforms existing state-of-the-art methods across multiple evaluation metrics, including accuracy, F1 score, HR@K, and NDCG@K, confirming its effectiveness and robustness in complex system environments.
Skin cancer is a prevalent and potentially life-threatening condition, where early and accurate detection is crucial for effective treatment. Traditional diagnosis, based on visual assessment, faces challenges due to inter-class variability and the similarity between benign and malignant lesions. Although automated diagnostic systems are gaining attention, few offer comprehensive evaluations. This study addresses this gap by systematically assessing such systems and reporting extensive performance metrics.This study contributes to both artificial intelligence and dermatological image analysis by introducing an automated system for multi-class skin lesion classification. The system was evaluated under various configurations, including original and artifact-inpainted images. Three approaches were employed: deep learning, radiomics-based and hybrid. Additionally, to enhance the interpretability, the system generates segmentation masks and attribution maps.Models were evaluated using five-fold cross-validation on the International Skin Imaging Collaboration (ISIC) 2018 Challenge Task 3 training dataset, with final performance assessed on the official test set. The best-performing model achieved a micro accuracy of 79.76 +/- 0.73, macro F1-score of 73.62 +/- 1.38, precision of 74.80 +/- 0.84, and recall of 73.93 +/- 2.19, surpassing comparable methods reported in the literature.The contribution to the research community is the release of a novel set of 57630 binary artifact labels for the ISIC-2018 Task 3 dataset, enabling research on the impact of visual artifacts such as hair, rulers, and interface fluids. This contribution is supported by statistical analysis and attribution map assessment, revealing significant differences in artifact occurrence across lesion classes and highlighting the utility of these labels in identifying and mitigating bias in skin lesion classification.
The integration of intelligent transportation systems and autonomous driving is reshaping modern mobility by mitigating the longstanding trade-off between traffic efficiency and road safety. Enabled by vehicle-to-everything (V2X) communications, connected and autonomous vehicles (CAVs) are increasingly integrated into vehicle-road-cloud collaborative networks, resulting in measurable improvements in traffic capacity, energy efficiency, and collision avoidance. However, the coexistence of CAVs, human-driven vehicles, and vulnerable road users (VRUs) introduces complex challenges in collaborative control, real-time risk perception, and data security. To address these issues, this review synthesizes recent advances through a tripartite lens: collaborative driving, risk-aware perception, and energy-efficient operation. Our analysis identifies three recurring scientific questions: (i) how to effectively couple dynamic risk perception with cooperative control under uncertainty; (ii) how to enable secure, privacy-preserving collaboration across heterogeneous agents; and (iii) how to guarantee VRU safety in mixed-autonomy traffic. Representative approaches fall into three categories: (1) multimodal collaborative decision-making frameworks combining hierarchical deep reinforcement learning and model predictive control; (2) federated learning architectures that preserve data privacy while enabling cross-vehicle knowledge sharing; and (3) human-centric safety mechanisms leveraging ultra-wideband sensing and heterogeneous graph neural networks for VRU detection and intent prediction. Collectively, these findings demonstrate that risk-aware, energy-efficient cooperative driving is technically feasible, yet its large-scale deployment hinges on interdisciplinary innovation, standardized communication protocols, and regulatory alignment.
Anomaly prediction is vital for ensuring the safe operation of pump station facilities. However, traditional models are often hampered by spurious correlation stemming from unobservable confounding factors, thereby degrading detection accuracy. To address this, this paper proposes the Causal Intervention with Front-door Adjustment Model (CIFAM) to eliminate spurious correlation in complex features. Initially, CIFAM integrates video and signal data, utilizing graph attention networks and clustering to construct a multi-level structural attribute graph that represents dynamic operational dependencies. The model incorporates causal dilated convolution to capture long-term memory while preventing information leakage, alongside a random edge dropping strategy to simulate causal intervention. By applying the front-door adjustment criterion, CIFAM extracts robust intermediate representations to block interference paths. Furthermore, contrastive learning and Adaptive Instance Normalization (AdaIN) are employed to decouple features, bolstering model robustness in complex scenarios. Experimental results demonstrate that CIFAM achieves state-of-the-art performance across multiple industrial benchmarks, specifically reaching 91.1% accuracy and a 90.6% Micro-F1 score on the OURS dataset. The model exhibits exceptional robustness and sample utilization efficiency. Furthermore, with minor modifications, this algorithm can be generalized to various monitoring scenarios, such as power generator set monitoring, semiconductor production lines, and UAV-assisted monitoring for oil pipelines or high-voltage power line maintenance.
This paper investigates future purchase prediction in e-commerce using sequences of user visits and behavioural micro-interactions. The study is based on anonymised production data from three e-commerce domains, comprising approximately 690,000 users and 2.61 million visits. User histories are modelled as temporally ordered sequences of visits, with each visit represented by 21 features. Six models are evaluated: three tabular approaches (SVM, XGBoost, CatBoost) and three sequential architectures (RNN, LSTM, GRU), each tested in two variants addressing class imbalance. The experimental setup employs a temporal train-test split, five random seeds, and bootstrap-based significance testing. The top-performing sequential models achieve an AUC-PR of approximately 0.379, significantly outperforming the strongest tabular baseline, XGBoost (AUC-PR approximate to 0.364). This advantage remains consistent across domain-specific evaluations as well as in a leave-one-shop-out setting. Further analysis shows that mouse movement and scrolling-related features provide the strongest predictive signal, whereas extending user histories beyond 40 visits yields only marginal performance gains.
Unsupervised defect detection is crucial for industrial inspection, but teacher-student (T-S) frameworks tend to overfit a single teacher's feature manifold, leading to poor generalization on subtle anomalies. We introduce TAD++, a dual-path distillation framework that combines heterogeneous Teacher-Assistant-Student (T-A-S) guidance with a pseudo-defect inverse-distillation branch. A compact assistant, structurally distinct from the teacher, is trained to co-distill the student, thereby mitigating single-teacher bias. In parallel, the inverse-distillation path tasks the student with reconstructing normal appearances from defect-injected inputs, serving as a regularization term to prevent anomaly leakage. A dynamic attention weighting module adaptively fuses these heterogeneous guidance signals. Crucially, the assistant, inverse branch, and weight modules are strictly training-only. This design ensures that while TAD++ benefits from a rigorous multi-phase optimization, it maintains zero additional inference latency and memory overhead compared to standard T-S baselines. On MVTec AD, BTAD, and VisA, TAD++ achieves consistent improvements in both image-level detection and pixel-level localization, with extensive ablations confirming the efficacy of the heterogeneous dual-path design.
Since the proposal of the blockchain, its application scenarios have been continuously expanded. However, the anonymity feature of the blockchain has hindered market regulation, leading to numerous illegal activities such as phishing fraud, which has now become a serious type of crime. Currently, most phishing fraud detection technologies on blockchain platforms use transaction data to construct basic raw transaction graphs and then use neural network methods to mine key information. This study proposes a graph gated recurrent neural network (GGRNN) model that fully integrates temporal and spatial information, effectively utilizing time-related information in the transaction graph. It first takes an account as the center node to obtain its second-order transaction data and then constructs a dynamic transaction graph (DTG). Subsequently, the DTG is fed to the GGRNN to process the temporal features in a gated recurrent unit (GRU) framework and introduce graph convolutional network (GCN) operations to fully use the node neigh-bourhood topology features, obtain the embedded representation of the graph, and then perform graph classification for phishing node detection. To verify the effectiveness of the proposed model, it was applied to real-world Ethereum transaction datasets. Numerical results show that the proposed GGRNN model significantly outperforms state-of-the-art methods.
Few-shot learning (FSL) aims to transfer knowledge from known to unknown categories using limited samples. However, the opaque nature of neural networks makes it challenging to discern the knowledge learned by the model, and existing methods often lack explainability, limiting their reliable application in high-stakes fields such as medical diagnosis and autonomous driving. To address this, we propose a visually explainable dynamic similarity network (VEDSNet), which achieves a balance of performance, explainability, and efficiency through a lightweight architecture (approximately 6.8M parameters, built on a ViT-Tiny backbone). The Feature Decomposition Module (FDM) generates fine-grained, semantically meaningful representations via parallel feature learning, providing intuitive visual insights into the model's decisions. The Dynamic Metric Module (DMM) employs a sample-adaptive dual-metric strategy to enhance discrimination with limited data, switching to a single metric for efficiency when data is sufficient. Experiments on standard datasets demonstrate that VEDSNet achieves high classification accuracy while providing clear visual explanations of its decision-making process, making it suitable for efficient deployment in resource-constrained scenarios.
This paper proposes CDH-YOLO, an efficient, real-time pedestrian detection model for nighttime RGB images. Built on YOLOv5, CDH-YOLO incorporates structural reparameterization to optimize the backbone network and integrates convolutional block attention module to enhance feature representation. Transposed convolution replaces nearest neighbor interpolation for upsampling to preserve semantic information. A lightweight decoupled head addresses spatial misalignment between classification and regression tasks, while SIoU loss improves training convergence and localization accuracy. Experiments on the KAIST dataset demonstrate that CDH-YOLO achieves superior accuracy with real-time performance, significantly outperforming existing methods in nighttime pedestrian detection.
In human society, opinion evolution and guidance for opinion evolution are useful for maintaining social stability, business development, and so on. To tackle these issues, we propose an opinion evolution and guidance model based on social networks and information networks (namded EGSDCN) for the first time. Firstly, we develop an opinion evolution model based on the information networks and social networks (ISOE). Specifically, we first update the individual's opinion by judging the quality of the information obtained by individual from the information network. Then, we filter the trusted neighbor set for individuals by quantifying individuals' attributes and update individual's opinion after weighting analysis of the trusted neighbor set. Finally, we conduct information exchange between the social and information networks. For guiding opinion evolution, we develop a group opinion guidance strategy based on individual stubbornness differences (termed PDGM). Specifically, we first divide the guided individuals into stubborn and non-stubborn groups. Then, for the non-stubborn group, a linear function model is used to intervene individual stubbornness. For the stubborn group, we propose the interest and opinion change functions to dynamically adjust individuals' opinion. Extensive simulation experiments have been conducted and proved that our proposed model is effective.
Multimodal image matching remains a challenging yet essential task in the field of computer vision. In recent years, detector-free methods have emerged as promising approaches, achieving high matching accuracy by leveraging global modeling capabilities. While transformer-based methods are effective, they often suffer from significant computational overhead, limiting their efficiency.To address this, we propose MambaSC, a novel framework that integrates Mamba with self-attention and cross-attention mechanisms to balance accuracy and efficiency. Specifically, MambaSC introduces the M2Backbone for efficient feature extraction and the MSC Module to enhance feature interaction and alignment.Extensive experiments across multiple multimodal image datasets demonstrate that MambaSC consistently outperforms state-of-the-art methods while maintaining computational efficiency, making it a compelling solution for complex multimodal image matching scenarios. Code is available at: https://github.com/LiaoYun0x0/MambaSC.
The integration of artificial intelligence (AI) into the music domain has catalyzed a transformative shift in how music is composed, performed, and taught. This paper introduces and frames the concept of music intelligence and employs bibliometric and systematic review methodologies to comprehensively analyze music intelligence. Music intelligence encompasses the development and application of intelligent systems that not only automate or enhance traditional musical tasks but also foster new modes of creativity, interaction, and pedagogy. Tracing the evolution from early rule-based systems to modern deep learning and multimodal models, we examine how AI is increasingly embedded in musical workflows. We highlight applications ranging from generative composition and expressive performance interpretation to real-time accompaniment and personalized education. By positioning AI as an active collaborator rather than a mere tool, this study underscores the need for collaborative efforts among computer scientists, musicians, educators, and cognitive scientists to fully realize the potential of intelligent music systems. Our biblio-metric analysis indicates an annual growth rate of 14.92%, with China, the US, and the UK contributing 52.9% of global research output. The findings reveal a rapidly expanding interdisciplinary field characterized by increasing international collaboration, methodological diversification, and a growing focus on human-AI co-creativity. However, persistent gaps remain in cultural inclusivity, interpretability, and ethical governance.
The goal of log anomaly detection is to accurately detect system anomalies from logs. Traditional methods often suffer from insufficient generalization and delayed anomaly detection when dealing with semantically diverse and loosely structured log data. As the complexity of the system increases, the size of the logs is getting larger and larger, and it has become impractical to analyze them manually. To this end, this paper proposes FRLog, a log anomaly detection framework based on large language model, which realizes contextualized semantic embeddings of log sequences by fusing BERT and LLaMA models, thereby enabling more accurate log anomaly detection. Meanwhile, the parameter fine-tuning strategy ReFT is introduced, and the semantic bootstrapping, representation alignment and global tuning process are optimized by a three-phase collaborative training mechanism. Experimental results on three typical log datasets, BGL, HDFS and Thunderbird, show that FRLog outperforms the existing mainstream methods in terms of F1, Precision and Recall, especially in complex scenarios, demonstrating stronger anomaly discrimination and sample efficiency, which verifies its superiority and robustness in the log anomaly detection task.
Dynamic community detection often suffers from the instability of results, making consistent community identification across network snapshots critically important. However, the cut off between snapshots might lead to the loss of some higher-order structures, such as closed triangle motifs. In view of this, we examine the relationship between the missing higher-order structures and the instability, and find a positive correlation between higher-order loss ratio (HOLR) and temporal smoothing normalized mutual information (TSNMI). Based on this finding, we propose a new-brand higher-order edge enhancement (HOEE) algorithm, aiming to effectively reconstruct higher-order interactions to overcome the instability issue. The HOEE algorithm employs the higher-order activity potential (HAP) of nodes between consecutive snapshots to recover the loss of higher-order information by the transformation of the triangle motif, thus ensuring the temporal stability of dynamic communities. Experimental evaluation on synthetic and real-world dynamic networks demonstrates that HOEE outperforms state-of-the-art methods in community detection accuracy and significantly reduces community instability. Theoretical analysis confirms stability guarantees and characterizes graph property changes induced by HOEE. The HOEE algorithm effectively enhances temporal community stability through higher-order interaction reconstruction, providing a robust solution for dynamic network analysis.
This paper tackles the challenge of achieving Input-to-state stable (ISS) synchronization in actuator-saturated sampled-data control (SDC) networks for Markovian jump (MJ) Lur'e networks. We investigate the impacts of actuator saturation on system performance and stability, proposing a control strategy that ensures synchronization in the presence of external disturbances. Our analysis employs a Wirtinger-based integral inequality alongside a modified free matrix-based integral inequality (MFMBII), providing a framework for examining Lur'e networks. Initially, we create an MFMBII that combines the dynamics of MJ Lur'e networks and takes into consideration time-varying delays. Second, we formulate two sufficient conditions for the SDC design that ensure mean-square ISS error of specification for the hybrid closed-loop system. We do this by combining the MFMBII method with the Lyapunov-Krasovskii functional (LKF). Through a systematic methodology, we demonstrate that the proposed method maintains bounded state responses and converges to a common trajectory at an exponential rate. The results highlight the effectiveness of integrating ISS with SDC in managing complex dynamical networks. Finally, the proposed ISS method is validated through a numerical example, confirming its efficacy.
Chest X-ray analysis is vital for clinical screening, diagnosis, and treatment planning. The increasing workload on radiologists calls for robust automated solutions to generate accurate and standardized reports. Conventional report generation models often struggle to detect rare and anomalous diseases, particularly when faced with imbalanced datasets, which can compromise diagnostic knowledge accuracy. To address these limitations, we propose ChestXGen, a novel multimodal framework for automated radiology report generation. Our model is based on a fully Transformer-based encoder-decoder architecture that integrates Memory Augmented Transformer (MAT) blocks with a Context-Aware Bi-Gate (CABG) mechanism. These enable the model to capture long-range dependencies, effectively fuse visual and textual features, and better handle underrepresented conditions. Visual features are extracted using a ResNet-101-V2 backbone and refined through a shared memory module that continuously reinforces cross-modal associations. This integrated approach facilitates the generation of comprehensive, accurate, and contextually coherent reports. Extensive evaluation on the large-scale MIMIC-CXR dataset, comprising 377,110 images and corresponding free-text reports demonstrate that ChestXGen outperforms previous models on BLEU-1, BLEU-2, BLEU-3, and METEOR metrics. The results demonstrate the efficacy of Transformer-based models in substantially reducing radiologists' reporting burden while concurrently enhancing the precision and reliability of diagnostic interpretations.
This paper introduces a pioneering approach to robot environment representation by integrating a novel hypergraph-based method for modeling fuzzy signatures with a quadtree-like structure for obstacle detection. By structuring fuzzy signatures through hypergraphs, we establish a robust framework that not only streamlines information representation but also simplifies the aggregation-based decision-making process. This synergy is applied to the domain of mobile robotics, where accurate and efficient environment representation is essential. Utilizing a quadtree-like structure for data organization, our technique systematically evaluates feature points against a set of fuzzy operations, determining the significance of obstacles and reconstructing the environmental model through the traversal of the quadtree-like structure. Furthermore, the hypergraph-based formalism sets the stage for a future transition to a tensor-based representation of fuzzy signatures, as envisioned in future work.
Differential Evolution (DE) has been found to be inefficient and inaccurate for high-dimensional complex problems. Quantum-inspired Differential Evolution (QDE) possesses quantum computational properties, enabling effective handling of high-dimensional problems. However, QDE is plagued by issues of excessive mutation and poor convergence. Therefore, a hybrid enhanced Quantum-inspired Differential Evolution algorithm, termed QAHQDE, is proposed. Within QAHQDE, an improved chaotic strategy is designed. Non-repeating distributed quantum positions are generated, enhancing the diversity of initialized individuals. A quantum-adaptive mutation strategy is adopted to address the over-mutation problem inherent in QDE. The mutation degree is adaptively reduced, and convergence performance is thereby improved. A novel hybrid mutation strategy is constructed. Weighted mutation operators are combined with standard differential evolution. Local and global search capabilities are balanced, and convergence accuracy is enhanced. The performance of QAHQDE was evaluated against 38 algorithms using 48 benchmark functions from CEC2005, CEC2010, and CEC2013, across dimensions D=100, 500, 1000, and 3000. Experimental results demonstrate that QAHQDE outperforms QDE by at least three orders of magnitude. Superior convergence performance, higher convergence accuracy, and excellent stability are exhibited by QAHQDE on most functions.
Single-image generation models generate high-quality and diverse images by learning the internal distribution of patches within a single image, addressing the issue of data scarcity and attracting increasing attention. However, existing methods are unsatisfactory when dealing with images with global structures, such as animal images. To address this issue, we propose Semantic fusion and Structure-guided global generation from a Single image with Diffusion models (S3Diff). Specifically, during training, we employ a semantic extractor to extract high-level semantic features from training images and use the proposed semantic fusion block to fuse semantic features with image features, enhancing the model's understanding of image semantics and improving the quality of the generated images. During sampling, we apply manifold constrained gradient based on image structure to enforce the generation path to regress to the manifold of the original image, preserving reasonable global structures. Extensive experiments on public datasets demonstrate the thorough exploration of hyperparameters and the rationality of key designs, with quantitative and qualitative comparisons against baseline methods and validating that our proposed method preserves reasonable semantic and structural relationships, can generate high-quality and diverse images, significantly improving the model's global generation capabilities.
