To address the issues of high computational overhead and the inability to capture hierarchical structures when DBSCAN processes data with uneven density distributions, this paper proposes a density-driven clustering framework, namely MSGB based on granular balls (GBs). This method uses the GB as the basic operational unit. First, it automatically determines the initial number of partitions using mean drift and initializes cluster centers based on local density peaks to avoid the sensitivity to initial values inherent in traditional k-means. Second, it designs a radius calculation strategy based on the coefficient of variation-weighted median distance, allowing the GB scale to dynamically adjust according to local density variations, thereby preserving structural information while suppressing noise. Furthermore, by constructing a granular ball topological similarity graph, it transforms the density reachability determination into a search for connected components among granular balls, effectively handling scenarios with multi-density clusters and weak connections. To validate the effectiveness of the proposed method, this paper conducts comparative experiments on 10 synthetic datasets and 10 real-world datasets using ACC, NMI, and ARI as evaluation metrics. Specifically, on the synthetic datasets, MSGB improves ACC by 3.81%, NMI by 2.82%, and ARI by 4.49% compared to DBSCAN. On real-world datasets, it also achieves consistent improvements. The results show that MSGB outperforms several mainstream comparison algorithms on average across all datasets, demonstrating outstanding robustness and consistent advantages, particularly in noisy and non-uniform density scenarios, thereby validating the framework’s strong adaptability to data with complex density structures. The source code of the MSGB algorithm is available via the following network link https://github.com/Z-Lucky-M/MSGB.

Large Language Models (LLMs) are increasingly explored for mental health assessment and psychiatric decision support, but autoregressive generation remains vulnerable to hallucination, semantic drift, and weak traceability. These risks carry particular consequences in psychiatric settings because decisions depend on structured diagnostic criteria, longitudinal symptom relations, exclusion rules, and clinically interpretable evidence. Retrieval-Augmented Generation (RAG) can improve factual grounding, yet pipelines centered on dense semantic retrieval do not naturally represent the discrete topology, hierarchical organization, and multi-hop dependencies of clinical ontologies and diagnostic rules. We therefore treat neuro-symbolic RAG as an information-fusion problem for explainable mental-health applications. We organize the literature around three fusion loci: pre-retrieval fusion, intra-retrieval alignment and re-ranking, and post-retrieval decoding and decision-level fusion. These loci are compared through source heterogeneity, fusion granularity, constraint strength, and evaluative consequences. The central questions are where neural and symbolic signals meet in the RAG pipeline, which mechanism families recur at each fusion locus, and how those choices shape strengths, risks, and evaluation burdens in mental health applications. To answer them, we draw selectively on adjacent clinical, biomedical, information retrieval, and general RAG literatures. The evaluation lens moves beyond task accuracy to include faithfulness, logical consistency, uncertainty and abstention behavior, and auditability; these dimensions structure a stage-aware evaluation framework. We close by outlining challenges in temporal reasoning, patient-specific graph construction, evaluation standardization, and clinically deployable human-in-the-loop neuro-symbolic systems.

Due to its NP-hard nature, the electric vehicle charging scheduling problem requires a robust optimization algorithm capable of producing accurate solutions within a reasonable computational cost. Existing algorithms reviewed in the literature often suffer from either slow convergence speed or an inability to escape local optima. Therefore, this study proposes a novel evolutionary algorithm, termed FL-SHADE, designed to address these limitations by achieving a better balance between exploration and exploitation. This algorithm combines the FGO algorithm with the adaptive L-SHADE algorithm to present a new variant, namely FL-SHADE. The FGO algorithm has robust exploration that helps escape from local optima but suffers from a poor exploitation operator, leading to slow convergence. In contrast, the AL-SHADE algorithm is known for its effective exploitation but limited exploration. By leveraging the complementary strengths of both algorithms, FL-SHADE features strong abilities to avoid stagnation in local optima and accelerate convergence toward high-quality solutions. FL-SHADE is initially assessed using the CEC2017 benchmark and compared with several competing algorithms based on several performance indicators to evaluate its stability and effectiveness. According to the experimental results, FL-SHADE can outperform all algorithms on 15 of 29 test functions, be competitive on 12, and perform worse on only 2, demonstrating that it is a robust alternative for addressing continuous optimization challenges. Subsequently, FL-SHADE is evaluated on 12 charge scheduling problems under four different penetration levels and three scenarios to assess performance at small, medium, and large scales. In addition, it is compared with several high-performing and recently proposed optimization algorithms to validate its effectiveness and stability. The experimental results indicate that FL-SHADE outperforms competing algorithms in eight test cases, whereas AL-SHADE performs better in the remaining cases, suggesting that both algorithms are strong candidates for solving electric vehicle charge scheduling problems.

Cross-domain few-shot hyperspectral image classification has attracted significant attention due to its practical relevance in scenarios where labeled samples are scarce and domain distributions differ. Existing methods mainly focus on explicit feature alignment and distribution transfer between source and target domains, while often overlooking the spectral ambiguity caused by mixed pixels and the lack of hierarchical structure in Euclidean feature space. These limitations reduce intra-class consistency and weaken inter-class separability. To address these issues, we propose a fuzzy hyperbolic-aware fusion network for cross-domain few-shot hyperspectral image classification. Specifically, a fuzzy attention feature (FAF) embedding module is designed to adaptively fuse base convolutional features with spectral–spatial structural features through fuzzy attention, thereby alleviating spectral ambiguity and enhancing discriminative representation learning. On this basis, we further introduce a hyperbolic cross-sample relational attention mechanism, where support and query features are first attentively weighted in Euclidean space to capture cross-sample correlations and then projected into Lorentz hyperbolic space. Benefiting from the negative-curvature geometry, the proposed hyperbolic embedding explicitly models hierarchical relationships among samples, leading to significantly tighter intra-class clusters and more distinct inter-class boundaries. By jointly leveraging fuzzy attention and hyperbolic representation learning, the proposed method effectively enhances cross-domain discriminability and generalization capability under limited labeled samples. Extensive experiments on multiple hyperspectral datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches in both classification accuracy and stability.

Glioma segmentation in multi-modal MRI is critical for clinical decision-making, but it remains challenging due to tumor heterogeneity and the complexity of multi-modal feature fusion. Although recent deep learning methods have achieved notable progress, they often struggle to effectively extract discriminative feature information, and achieve effective cross-modal fusion. To address these challenges, we propose an Anatomical Symmetry-guided Multi-modal Fusion Network (ASMFNet) for precise glioma segmentation, which explicitly leverages anatomical symmetry. First, a hierarchical multi-encoder-decoder architecture is proposed to capture multi-modal complementary information. Second, an anatomical symmetry guidance module is proposed to guide the network to exploit tumor-induced asymmetries. Third, an intra-modality purification module is proposed to suppress irrelevant features while enhancing glioma-related features. Furthermore, a cross-modal fusion module is proposed to enable adaptive and context-aware multi-modal fusion. Finally, an anatomically-informed segmentation loss function is designed based on the characteristics of glioma to boost model convergence and improve segmentation performance. Extensive ablation and comparative experiments conducted on the BraTS datasets demonstrate that ASMFNet achieves statistically significant improvements over state-of-the-art methods, demonstrating its potential for reliable and clinically applicable brain tumor segmentation.

Medical insurance is designed to reimburse employees for monetary losses brought on by illness risks. However, the increasing costs of medical insurance make people tend to commit fraud. Medical insurance fraud (MIF) leads to losses for individuals, businesses, and governments, while also threatening the global sustainability of medical insurance systems. For the sake of classifying the severity of MIF, this study investigates a unique sorting method by integrating cloud model with technique for order preference by similarity to ideal solution (TOPSIS), i.e., Cloud-TOPSIS-Sort, to tackle multi-criteria sorting (MCS) problem. Firstly, given uncertainty in MCS, this study will adopt probabilistic interval-valued hesitant fuzzy set (PIVHFS) for handling uncertainty and propose a technique for objectively ascertaining the probability information in PIVHFS. Secondly, a method of transforming PIVHFS into cloud model is proposed for aggregating probabilistic interval-valued hesitant fuzzy elements (PIVHFEs). Thirdly, considering the important role of divergence in uncertainty theory, f-divergence is extended for cloud model to realize the effective processing of uncertain information. Then, a Cloud-TOPSIS-Sort method is raised, and a case study concerning MIF is used for illustrating the application of the suggested technique.

Circular intuitionistic fuzzy sets (C-IFSs) offer an expressive mathematical structure for representing uncertainty, extending traditional intuitionistic fuzzy sets by introducing a geometric interpretation through circular regions. Knowledge measures, which quantify information from the perspective of certainty, serve as a critical tool for capturing informational content and clarity. This study systematically extends the notion of knowledge measures to the framework of C-IFSs by introducing a suite of general construction techniques. These methods employ foundational mathematical constructs, including t-norms, t-conorms, automorphisms, and aggregation operators. The article ensures theoretical soundness by formally deriving expressions and proving mathematical results. Importantly, the proposed framework demonstrates that knowledge measures can be constructed not only from individual t-norms and t-conorms but also through their integration with automorphisms and aggregation strategies. The formulation of arguments within these operators plays a pivotal role, and the inherent non-uniqueness of such structures invites further scholarly exploration. Building on these insights, this study develops a novel score function that enables a refined ranking mechanism balancing both information content and clarity. Experts can adjust decision weights dynamically through tunable parameters. Additionally, the article proposes a new weight generation method based on knowledge-theoretic principles. To illustrate the practical efficacy of the proposed approach, this study introduces a comprehensive multi-attribute group decision-making framework and applies it to a real-world supplier selection problem in the manufacturing sector. The results demonstrate strong robustness and reliability of the proposed model, with an average Spearman rank correlation coefficient of 0.90 compared to existing approaches.

Multivariate time series forecasting plays a vital role in finance, meteorology, energy scheduling, transportation, and environmental monitoring, yet traditional statistical and machine learning methods often failto capture the complex nonlinearities, high-dimensional redundancies, and spatio-temporal dependencies inherent in real-world data. To overcome these limitations, we propose IFCM-DSACNN, a novel deep learning framework that synergistically integrates an improved fuzzy $C$-means clustering algorithm (IFCM) with a dual self-attention mechanism and a one-dimensional convolutional neural network (1D-CNN). The IFCM module first clusters features exhibiting similar temporal dynamics, thereby reducing feature redundancy and mitigating noise interference through a cross-entropy-based suppression strategy, while the alpha evolution algorithm autonomously determines the optimal number of clusters without manual tuning. Within each coherent feature subset, a dual self-attention mechanism is applied in parallel along both the sample and feature dimensions to capture complementary global dependencies, and the resulting relation-aware representations are fused and subsequently refined by a 1DCNN to extract local structural patterns for final prediction. Extensive experiments on six publicly available datasets, comparing against nine state-of-the-art forecasting models, demonstrate that IFCM-DSACNN consistently achieves the lowest RMSE and MAE across all datasets, along with near-zero MBE values, confirming itssuperior accuracy, robustness, and bias control. Ablation studies validate the complementary contributions ofeach module, and Friedman with Nemenyi statistical tests confirm the)significance of our improvements. These results establish IFCM-DSACNN as a powerful and reliable solution for complex multi-feature time series forecasting tasks.

The rapid advancement of medical imaging technology has produced massive volumes of data, making efficient retrieval from large-scale datasets a critical challenge. Hashing-based retrieval algorithms provide a promising solution for enabling efficient data retrieval. However, most existing approaches primarily exploit global image information while often neglecting small, localized pathological regions that are of greater diagnostic significance in medical imaging. To address this issue, a novel Dual-Channel Transformer Hashing with Co-Attention Fusion (DCTH-CF) algorithm is proposed, designed to capture both the overall context and fine-grained local features simultaneously. In the global channel, multi-scale learning is integrated into a Scale-Enhanced Transformer Encoder, combined with a Power Mean Transform layer, to improve nonlinear feature representation. For the local channel, a Local Primary Model is first constructed from individual image regions, and these features are further refined through the same Scale-Enhanced Transformer Encoder and Power Mean Transform layer to extract highly discriminative local representations. Finally, global and local features are integrated via a co-attention-based fusion module that leverages Global-Local and Local-Global Attention mechanisms, enabling the generation of effective hash codes for retrieval. Experimental results on large-scale medical image datasets show that DCTH-CF achieves mean Average Precision (mAP) of 0.368, 0.382, and 0.360 at 12, 24, and 36 bits on ChestX-ray14, and 0.714, 0.736, and 0.723 on ISIC2018. These results significantly outperform state-of-the-art methods, demonstrating the effectiveness of DCTH-CF in medical image retrieval.

The remarkable capabilities of large language models (LLMs) are fueled by massive Internet-scale corpora. However, scraped data owners often do not consent to its use for training, raising significant legal and ethical concerns over copyright and privacy. Data auditing techniques seek to verify whether a protected dataset was used in training a target LLM, typically framing the task as membership inference: estimating binary sample-level membership and aggregating to a dataset-level decision. In this article, we identify a fundamental limitation of this crisp binary paradigm: in realistic training pipelines, datasets are rarely used in full. Instead, models are trained on mixtures of partial subsets drawn from multiple sources. Existing auditing techniques, built upon an all-or-none assumption-declaring a dataset either entirely present or absent from training-collapse in partial dataset usage scenarios. Their predictions fluctuate unpredictably with the member ratio, causing unstable performance and high false-negative rates. Inspired by fuzzy set theory, we relax the crisp notion of binary membership to a continuous fuzzy membership in [0,1], quantifying each sample's degree of inclusion in the model's training set. We establish a theoretical bridge between sample-level fuzzy memberships and the dataset-level usage ratio, facilitating inference of the proportion of a protected dataset used during training. A neural network fuzzifier first estimates sample-level fuzzy memberships from binary labels in a reference set, then refines them using dataset-level member ratios as higher order supervision. Finally, a defuzzification stage aggregates calibrated memberships to determine partial usage. Across LLMs of varying scales and multiple auditing datasets, our fuzzy auditor substantially outperforms state-of-the-art crisp binary techniques in detecting partial usage, estimating member proportions, and identifying individual member samples.

Bipartite graphs are a fundamental resource for inferring user preferences and understanding consumer decision-making behavior, driving the rapid development of graph-based recommendation systems. The challenge of feedback scarcity (i.e., sparse interactions) in graph-based recommendation systems has prompted extensive research on exploration-exploitation strategies to trade off exploration-exploitation in interactions. However, the tight coupling between exploration-exploitation behavior and bipartite graph learning poses a major obstacle to achieving the balance needed for long-term recommendation gains (i.e., enhancing diversity while maintaining accuracy). To better understand this issue, we conduct preliminary theoretical analysis and find that uneven user-item interactions heighten the risk of exploration-exploitation imbalance. Even popular graph Transformer architectures, though effective in exploration-exploitation, may still exhibit uncertain and unintended behavior, which can result in the imbalance issues. To address this problem, we propose a pane-aware graph Transformer architecture for personalized recommendation from a dynamic resource allocation perspective. Our approach builds upon graph Transformers and introduces two key modules: 1) pane partitioning and 2) a two-stage learning strategy, to explicitly intervene in the exploration-exploitation process of graph Transformers, rather than relying on the models to perform potentially uncertain and unintended balancing behavior. Specifically, the first module uses clustering and large language model reasoning to accurately constrain exploration-exploitation regions, guiding the balancing behavior of downstream graph Transformers. The second module separates training focuses to improve information allocation and ensure marginal gains. Extensive experiments on real-world datasets demonstrate that the proposed method effectively balances exploration-exploitation in user-item interactions and achieves long-term recommendation gains.

Continuous Time Dynamic Graphs serve as abstractions of real-world dynamic networks. The internal logic of intricate interaction relies on the underlying correlation of spatio-temporal features among nodes in the process of continuous evolution. Recent studies attempt to mine this logic and solve the temporal correlation of graph topologies through feature timing coding. However, the existing explicit coding theory does not fully explore the continuous spatio-temporal state evolution, thus hindering the capture of historical state factors and limiting the accuracy of future state prediction. In this paper, we propose a Continuous Spatio-Temporal Graph Disentanglement with Selective State Spaces Model (CSTGD-Mamba) based on selective state space models. Specifically, an Anonymous Spatio-Temporal Aware feature extraction method is proposed to enhance the induction of current state spatio-temporal feature capture. Additionally, the spatio-temporal kernel operator disentangles the potential correlation of current state attribute features, spatial features and temporal features. Selected state space models capture long-range dependencies in continuous spatio-temporal state evolution, unify spatial and temporal messaging, and allow continuous time graph propagation and information aggregation to characterize stable and accurate future spatio-temporal dynamics. Our approach outperforms state-of-the-art methods in both transductive and inductive settings on four real-world datasets for spatio-temporal state prediction.

To address the coupled challenges of fault detection and isolation in plant-wide industrial processes, this article proposes a structure-collaborative distributed dictionary learning (SCDDL) framework. The core idea is to exploit the conditional structure associated with shared variables in a sparse Gaussian graphical model (GGM) as a common basis to characterize cross-block relationships, expose fault-induced structural deviations, and support coupling-aware detection and decoupled isolation in large-scale industrial systems. Guided by an overlapping-block sparse GGM (OSGGM), process variables are partitioned into independent, conditionally independent, and joint-overlap blocks. Based on this partitioning, two complementary dictionary learning paths are integrated: 1) an adaptive manifold-guided path that preserves local geometry via atom-wise regularization and 2) a graph-guided path that uses a GGM-derived conditional-mean projection as a structural prior to align observed dependencies and graph-induced dependencies. These paths are used to suppress local normal variations, whereas faults that violate local geometric consistency or graph-induced dependencies are amplified as residuals and coefficient perturbations, thereby improving detection sensitivity. For fault isolation, a hierarchical block-wise reconstruction strategy that involves constrained normal-pattern recovery, block-data substitution, and joint sparse decomposition is developed. This design explicitly removes cross-block interference and enables decoupled isolation. Numerical studies and industrial applications demonstrate the effectiveness and robustness of the proposed framework.

High-dimensional datasets often contain numerous irrelevant or redundant features, which may lead to overfitting and increased computational complexity. In most cases, feature selection can improve classification accuracy and reduce feature dimensionality, and thus can be regarded as a multi-objective optimization problem. In recent years, multi-objective evolutionary algorithms (MOEAs) have been demonstrated to perform well in multi-objective feature selection tasks. However, their performance often degrades when confronted with high-dimensional noisy data. To address these challenges, this paper proposes an evolutionary algorithm utilizing fuzzy autoencoders, termed FAE-MOEA, for multi-objective feature selection. The proposed method combines fuzzy theory with an autoencoder model to effectively suppress noise in the data and further enhance the discriminative capability of feature importance. Specifically, we employ mutual information in conjunction with k-means to classify feature correlations, then utilize intuitionistic fuzzy sets to handle uncertainty, and finally extract feature importance through an autoencoder equipped with an Evaluation layer. Moreover, to accelerate the optimization process and extract superior feature subsets, FAE-MOEA initializes a Guidance Matrix based on Feature Importance and Individual Importance, and uses this matrix to guide the evolution of the population. Accordingly, two key operators and an update strategy are designed to effectively direct the evolutionary search process during each iteration. Experimental results on twelve high-dimensional datasets show that the proposed algorithm achieves significant improvements over several existing algorithms with respect to the dimensionality of the retained subsets and the resulting classification performance.

Accurately extracting lesion regions from histopathological images is crucial for diagnosing cervical squamous cell carcinoma, and determining the locations and shapes of such lesions is very important for evaluating tumor size and metastasis trends. The existing deep learning-based methods for extracting lesion regions in histopathological images primarily employ attention mechanisms to focus on regions of interest or integrate multiscale features to enhance the attained segmentation performance. However, their accuracy remains suboptimal when histopathological images with indistinct or blurred edge features are being processed. Therefore, we propose a dense prediction framework for histopathological images that integrates visual texture pattern enhancement and region-aware topological semantics via a feature interaction fusion strategy, enabling precise lesion segmentation. Specifically, our approach addresses the challenge brought by indistinct edges in pathological images through a multiscale texture enhancement branch, which sharpens discriminative morphological patterns. To optimize the feature fusion process, we introduce a positive definite-constrained attention mechanism that facilitates multilevel interactions between texture-enhanced and topological semantic features. Additionally, a united possibility state-space-model module is further designed to extract robust region-level topological semantics, thereby enhancing the ability of TexSem-Net to comprehend complex structural relationships. Extensive experiments demonstrate that our method outperforms the state-of-the-art segmentation networks in terms of both accuracy and delineation precision.

Forecasting the future state of a single target sensor from multi-sensor observations is a critical task in intelligent transportation systems, supporting congestion warning, signal control, and fine-grained traffic management at key locations. Existing methods under this setting often face three limitations: they characterize feature contributions only at the point or window level, rely on reduction-oriented feature processing that may discard complementary information, and rarely transform extracted structural information into relation-aware cues for downstream prediction. To address these issues, this paper proposes TVFI-GST, a spatiotemporal forecasting framework driven by time-varying structural cues. TVFI-GST first constructs target-semantic phase partitions and models phase-level temporal feature importance. It then employs fuzzy information granules (FIG) as carriers of trend evidence and reliability information to enable reliability-aware granular soft fusion, suppressing redundancy while preserving and reconstructing task-relevant information. Finally, the learned fusion structure is converted into dynamic relational cues and lag-aware context for fusion-aware prediction, forming a closed loop between representation learning and downstream forecasting. In this way, FIGs support phase-level importance modeling, reliability characterization, and information-preserving fusion, elevating temporal feature importance from a local explanatory signal to a stable structural interface for subsequent modeling. Experiments on six real-world traffic datasets demonstrate that TVFI-GST achieves strong overall forecasting performance against competitive baselines, while also offering interpretable structural cues under complex traffic conditions.

Brain network dysfunction is an important feature of the disease. Acupuncture’s therapeutic effects are achieved through dynamic interregional brain interactions and the fusion of multisource neural information. In this study, we enrolled 177 patients with Functional Dyspepsia (FD) who received 20 sessions of acupuncture treatment and underwent functional magnetic resonance imaging before and after treatment. Subsequently, we constructed the Hierarchical Efficacy Network (HEffNet) to predict the improvement value of the Nepean Dyspepsia Symptom Index by iteratively identifying the core trunk (efficacy region) from the Brain Tree to form a hierarchical information integration path. The model uses a Kalman filter and an improved Gated Graph Neural Network (GGNN) to capture the topological connectivity of brain regions. Experimental results showed that the features of hierarchical efficacy regions were significantly related to therapeutic efficacy. Patients who experienced significant symptom improvement formed a multi-level stable network across the frontal lobe, limbic lobe, insula, and subcortical nuclei, confirming that acupuncture works by optimizing the information fusion between brain regions. The HEffNet model outperformed baseline models in the prediction task. Ablation experiments verified the effectiveness of the GGNN and residual modules. The therapeutic efficacy of acupuncture for FD stems from the hierarchical information fusion of multiple brain regions. HEffNet quantifies this mechanism to achieve precise efficacy prediction and provides a new paradigm for studying acupuncture’s central mechanisms.

Feature selection is a critical preprocessing technique in machine learning that helps reduce data redundancy and noise while improving model performance. In high-dimensional datasets with large search spaces, effective feature selection is essential for identifying valuable information. Most existing methods rely heavily on hyperparameter tuning and exhibit poor generalization, limiting their applicability across diverse datasets. To address these issues, a Differential Evolution Framework with multi-mutation strategies fusion and self-adaptive configuration (DEF-SAC) is proposed for high-dimensional feature selection problems. The main innovations are: 1) We design a self-configuring framework for multi-mutation strategies to enhance diversity and search efficiency; 2) A novel dual Hamming distance-based population initialization method is developed. Using randomly binary reference individuals as the reference core and leveraging Hamming distance to quantify feature correlations for achieving population initialization and diversity; 3) We additionally introduce a feature-weighted evaluation model that uses entropy values to guide population updating and accelerate convergence. By comparing DEF-SAC with seven representative meta-heuristic algorithms and five single-mutation variants of DEF-SAC, we validated its performance across 18 high-dimensional datasets: DEF-SAC achieved higher classification accuracy and smaller feature subset sizes than other methods on 16 datasets, demonstrating exceptional optimization capabilities.

Existing sliding window-based data stream clustering methods often struggle with fixed-granularity representations and delayed responses to abrupt pattern changes, leading to reduced adaptability and clustering accuracy. To address these limitations, we propose an Adaptive Fuzzy Granular-Ball framework for Stream clustering with sliding windows (AFGBStream), a novel sliding window-based data stream clustering framework that integrates fuzzy set theory with granular-ball computing to support adaptive and incremental updates. Specifically, a window-guided data granulation strategy is developed using a Fuzzy C-Means (FCM)-based granular-ball generation approach, which enhances representation robustness and reduces sensitivity to micro-cluster parameters. In addition, a window-aware temporal decay mechanism is employed to improve the model's responsiveness to evolving data by dynamically filtering outdated information. Our comprehensive experimental evaluation on 17 benchmark datasets demonstrates that AFGBStream achieves superior performance over five competing methods.