NetWork
Classification Method of Medical Insurance Fraud Suspects Based on Data Augmentation
MA Zongchen;DING Weilong;CAI Ruihao;SHAO Jingcheng;HE Haoyang;ZHAO Zhuofeng;To address the low accuracy in classifying fraud suspects due to imbalanced samples and difficulty in obtaining global information, a data augmentation method for classifying medical insurance fraud suspects was proposed. Firstly, a table of script vocabulary was constructed by extracting WeChat chat records, and the retrieval-augmented generation technique was employed to expand the minority categories.Key information semantically similar to suspect instances was generated and converted into a graph structure for storage in a graph database. Secondly, a graph attention network model was introduced in the graph classification stage. Combined with a mask matrix, the training process was optimized to enhance the model′s ability to extract key features, thereby achieving accurate node classification. Experimental results on real-world datasets demonstrated that compared to the graphSAGE and ChatGLM4 methods, the macro-average F1 scores of the proposed method increased by 3. 3 and 12. 9 percentage points respectively, significantly improving the classification accuracy.
Intelligent Recognition for Petroleum Drilling Conditions Based on Association Rule Mining and Dynamics Graph
SHAN Xin;YANG Zhongguo;ZHAO Zhuofeng;A drilling condition recognition method based on association rule mining and system dynamics graph construction was proposed to address the problems that multisource sensor time-series data in petroleum drilling were easily interfered by noise, that the existing deep learning methods had poor interpretability, and that traditional association rule mining methods were difficult to capture parameter mutations and dynamic associations. Firstly, the adaptive optimal smoothing window algorithm was introduced to denoise the original high-dimensional time-series data. Secondly, a correlation mining algorithm based on symbolization and mutation perception was proposed, and dynamic rules were extracted from the symbol sequences. Finally, a system dynamics graph was constructed based on the mined results to reveal the logical relationship chains of drilling parameters under different conditions. The experimental results showed that the proposed method could effectively identify the association patterns of key parameters in the drilling process, generate interpretable rule sets, and provide reliable support for intelligent monitoring and analysis in petroleum engineering.
Adaptive Multiple Undersampling for Imbalanced Data Ensemble Classification
ZHANG Zhen;GUO Yujie;TIAN Hongpeng;WANG Wenjuan;An adaptive multiple undersampling ensemble classification method was proposed to address the limitations of underfitting and information loss caused by fixed sampling ratios and random sample deletion in traditional undersampling techniques when dealing with imbalanced data. Firstly, an adaptive dynamic undersampling strategy was designed in the sampling stage, which automatically determined the optimal imbalance ratio through an iterative process and generated multiple diverse training subsets, thereby overcoming the limitation of a fixed sampling ratio while retaining key sample information. Secondly, in the classification stage, an evidence fusion strategy based on conflict discrimination and dynamic weighting was proposed to collaboratively train and integrate the base classifier ensemble. Both global and local weights were comprehensively evaluated to achieve accurate sample classification. Comparative experiments conducted on multiple real-world datasets demonstrated that the proposed method achieved significant improvements in key metrics such as accuracy, precision and Matthews correlation coefficient, thereby confirming its effectiveness and superiority.
Weak Termination Analysis of Microservice Systems Based on Model Checking
YANG Li;ZHANG Chu;LIU Guoxi;LI Lecheng;DAI Fei;To address the automatic detection of minimum requirements for reliable interaction in microservice systems, a weak termination analysis method for microservice systems was proposed based on model checking. Firstly, microservices were modeled as labelled transition systems, and under asynchronous communication( mailbox communication) based on first-in-first-out buffers, the microservice system was formed by asynchronous composition of microservices. Secondly, in order to describe both narrow termination and generalized termination simultaneously, the traditional terminal state set was extended to a weak terminal state set. From the perspectives of deadlock-freedom and well-formedness, weak termination was then defined to formally characterize the minimum requirements for reliable interaction in microservice systems. Model checking was used to automatically check whether the microservice system satisfied weak termination on the classical microservice system case dataset. Experimental results showed that the proposed method could effectively check weak termination in microservice systems and thus ensure reliable interactions among microservice systems.
Federated Prediction of Remaining Useful Life for Critical Components of Autonomous Vehicles in Multiple Ports
MA Chao;ZHANG Chengcheng;GUAN Zhibo;ZHANG Kaiqi;HUANG Hai;To address the challenge of remaining useful life prediction caused by data silos and privacy concerns among ports, a federated learning method based on long short-term memory( LSTM) was proposed. Firstly, autoencoders were locally deployed at each port to extract key features and construct a federated LSTM model. Secondly, a feature similarity-driven adaptive client grouping and weighted aggregation strategy was developed. In this strategy, clients with similar degradation patterns were grouped, and model aggregation was performed within each group based on validation performance to enhance global performance under non-independent and identically distributed conditions. Experiments conducted on a real-world port battery dataset demonstrated that the proposed method effectively mitigated model drift and significantly improved remaining useful life prediction accuracy.
Transformer Retrieval Framework Based on Few-shot Adversarial Learning
HU Guangcheng;HU Xinrong;A Transformer retrieval framework based on few-shot adversarial learning was proposed to address the performance degradation of dense retrieval methods when dealing with query variants. Firstly, few-shot learning was introduced to improve the generalization ability of the model under limited data, and adversarial training was employed to enhance robustness against query variants. Secondly, a proxy attention mechanism was designed, in which proxy tokens interacted with key features in queries and documents, so that implicit semantic levels were captured while global context was preserved, thereby improving semantic mapping performance. The experimental results on the TREC-DL-2019 and TREC-DL-2020 datasets showed that the proposed framework had excellent retrieval accuracy and robustness, providing a new perspective for information retrieval under complex query scenarios.
Research on Metaheuristic Algorithm Based on Dual Closed-loop PID
YU Chunming;ZHU Xiaodong;CHEN Ke;REN Chunxiao;A global optimization metaheuristic algorithm utilizing a dual closed-loop PID( ProportionalIntegral-Derivative) control mechanism(DCLPID) was proposed to improve global exploration capability and local exploitation efficiency for solving high-dimensional complex problems. Inspired by control theory, the algorithm incorporates a position-velocity dual closed-loop structure: the outer position loop enhances strengthen global search, while the inner velocity loop improves enhance local refinement. PID parameters were dynamically tuned to achieve an adaptive balance during optimization. Experiments using the CEC 2017 benchmark suite were conducted, showing that DCLPID surpassed eight state-of-the-art meta-heuristic algorithms in both convergence precision and rate; rank-sum tests confirmed significant superiority with significance. These findings confirm the algorithm′s superior optimization performance and broad applicability, offering an effective approach for complex optimization tasks.
Solving Graph Coloring Problems Based on Improved Deep Q-network
SONG Jiahuan;WANG Xiaofeng;DING Hongsheng;HU Simin;SUO Xiaona;YAN Dong;The graph coloring problem( GCP), a canonical NP-hard combinatorial optimization challenge, has been recognized as playing a critical role in diverse application domains such as wireless spectrum allocation, parallel task scheduling, and resource optimization. To address the computational bottlenecks posed by large-scale and structurally complex graphs, a reinforcement learning framework based on an enhanced dueling deep Q-network(DDQN) was proposed. The graph coloring process was formulated as a Markov decision process, with carefully designed state representations, action definitions, and reward functions that were used to guide the agent in minimizing color conflicts and reducing the total number of colors during training. The adopted DDQN architecture explicitly decoupled the state-value function from the action-advantage function, thereby improving policy evaluation accuracy and enhancing training stability. Extensive experiments were conducted on several standard benchmark graph datasets, and it was demonstrated that the proposed method significantly outperformed traditional heuristic algorithms in terms of solution quality, color utilization efficiency, and convergence speed. The research not only provided a scalable and generalizable intelligent optimization paradigm for the GCP but also offered a novel modeling and solution pathway for tackling complex combinatorial optimization problems via reinforcement learning.
Research on Pneumonia Classification and Localization Methods Based on X-ray Imaging
WEN Yaxue;LI Yuqin;JIANG Zhengang;SHI Weili;DAI Bofan;Pneumonia is a disease characterized by high mortality, and research on its treatment and early screening tools has garnered significant attention. However, the complexity of chest images, annotation difficulty, and the inter-class similarities and intra-class variations of pneumonia all pose challenges to pneumonia classification and localization tasks based on X-ray images. Firstly, a self-supervised learning approach utilizing a siamese network was applied to extract self-supervised visual feature representation of chest X-ray images, and a multi-layer perceptron(MLP) detection head was used to perform pneumonia classification task. Secondly, a spatial-channel attention module was designed in the backbone network of YOLOv10. By utilizing the spatial information across channels to enhance pneumonia features, the ability of the YOLOv10 network to identify pneumonia lesions was enhanced. Finally, the proposed algorithm was evaluated using the COVID-19 Radiography Database and RSNA Pneumonia Detection dataset. The results demonstrated its effectiveness of the proposed algorithm in pneumonia classification and localization tasks based on X-ray images.
Community Detection Algorithm Based on Graph Attention Mechanism and Multi-distance Analysis
ZHANG Zhen;WU Guohao;ZHANG Hongxia;ZHOU Qi;To address the low accuracy of community detection in large-scale social networks, which was caused by complex node relationships and hidden community structures, a community detection algorithm based on graph attention mechanisms and multi-distance analysis was proposed. Firstly, a multilayer perceptron(MLP) module was introduced during the pre-training phase so that the accuracy of sample recognition could be improved. Secondly, a graph attention-based readout module was proposed, by which the contribution weights of nodes in graph representation learning were adaptively learned. The weighted aggregation was enabled, the discriminative power of graph-level representations was enhanced, and the accuracy and robustness of community detection were improved. Finally, a multi-distance analysis method was designed to strengthen the detection of small communities and to further increase overall detection accuracy. It was demonstrated by experimental results that superior performance was delivered by the improved model across multiple datasets, On the Amazon dataset, the F1 score of 90. 22% was achieved, which represented improvements of 11. 33 and 5. 86 percentage points over the CLARE and ProCom models, respectively. For Jaccard similarity, a score of 83. 46% was obtained, corresponding to improvements of 14. 96 and 7. 62 percentage points over CLARE and ProCom, respectively.