Existing ECG classification methods have difficulty jointly representing time- and frequency-domain information and are limited in their ability to fuse multiple views, model long-range dependencies, and handle class imbalance. To address these limitations, this study proposes a Time-Frequency Knowledge Attention Cross-Modal Fusion Network (TFKACF-Net). First, raw ECG signals and their DWT sequences are used to construct a dual-branch input for extracting complementary time–frequency features. Then, residual convolution, frequency-domain enhancement, and channel attention are employed to strengthen the representation of key waveform patterns. Finally, Transformer-based self-attention dynamically fuses multi-view features and captures long-range dependencies. Focal loss and AdEMAMIX are further combined to improve minority-class recognition. On the MIT-BIH five-class task, TFKACF-Net achieves an accuracy of 99.38
Incomplete Multi-view Clustering (IMVC) aims to cluster partially missing multi-view data. Existing methods impute missing data via semantic information. However, these methods ignore the evaluation of imputation quality, resulting in degraded clustering performance. Also, IMVC based on contrastive learning focuses on inter-view alignment while neglecting intra-view discriminability. Thus, we propose a method named Incomplete Multi-view Clustering based on Imputation-Quality Aware Adaptive Weighting and Spectral Constraint (IQASC). Unlike existing IMVC methods that equally weigh all imputed samples, our core novelty is two quality metrics for imputation reliability and dynamic sample-level confidence weights. First, to impute missing data, we design a Missing Imputation Network (MIN). This module adopts generative imputation based on self-attention mechanism. Second, to evaluate the quality of imputed data, we design a Sample Weighting Network (SWN). This auxiliary module evaluates imputation quality by integrating inter-iteration fluctuation and spatial neighborhood consistency, yielding dynamic confidence weights. Third, to optimize the intra-view clustering structure, we introduce a spectral loss into the MIN, which is rarely considered in contrastive IMVC frameworks. This loss function enforces compact same-cluster distributions and distinct different-cluster margins. Finally, we design an alternating optimization process for both MIN and SWN. Experiments on various datasets validate our method, with around 5
Accurate brain tumor segmentation from magnetic resonance imaging (MRI) remains challenging due to heterogeneous tumor morphology, intensity variations, and performance variability across imaging cohorts. This paper proposes an ACO-configured lightweight framework, ACO–DLV3+RFF, that integrates Ant Colony Optimization (ACO) with a modified DeepLabV3+ architecture enhanced by a Residual Feature Fusion (RFF) module. The ACO mechanism selects key hyperparameters—learning rate, dropout, dilation, fusion weight, batch size, and filter count—to identify an effective configuration within the predefined search space and experimental budget. The RFF module integrates high-level semantic features from Atrous Spatial Pyramid Pooling (ASPP) with fine-grained low-level features through ACO-selected weighted fusion, supporting contextual representation and structural-detail preservation. Comprehensive experiments across three benchmark datasets—FBTS, BraTS 2021, and BraTS 2018—evaluate the framework across independently processed MRI cohorts rather than establishing direct cross-domain transfer or clinical readiness. Under five-fold cross-validation, the proposed model achieves mean Dice Similarity Coefficient (DSC) and Jaccard Index (JI) values of 0.9755 and 0.9528, respectively, on FBTS. It also achieves DSC and JI values of 0.9708 and 0.9438 on BraTS 2021 and 0.9652 and 0.9345 on BraTS 2018, respectively. The controlled ablation study shows that the complete ACO–DLV3+RFF configuration achieves the highest DSC and JI among the same-protocol variants, whereas previously published methods are used only for contextual comparison. Nevertheless, the findings represent retrospective computational evidence, and independent external, multi-center, and prospective validation remains necessary before clinical translation can be considered.
Neighborhood rough-feature selection, referring to harnessing the neighborhood rough set approach in identifying informative features, has found widespread applications in data mining, knowledge discovery, pattern recognition, etc. Despite its broad usage, determining the neighborhood size is always a tricky problem, which plagued by several limitations, e.g., shaky adaptability to perceive the distributions of neighbors and poor capability to distinguish the qualities of features. To overcome these weaknesses, we introduce a generic method dubbed as FERN (Fuzzy Entropy Rate Neighboring) that facilitates the adaptive implementation of neighborhood rough-feature selectors. FERN defines a measure of fuzzy entropy rate which is leveraged to quantify the uncertainty in neighborhood. Through minimizing the uncertainty with a greedy strategy, FERN tailors the neighborhood size for each sample automatically. Serving as a plug-and-play framework, FERN can freely instantiate the existing neighborhood rough-feature selection algorithms. Extensive experiments on various existing neighborhood rough-feature selection algorithms demonstrate FERN’s effectiveness.
Time series clustering across distributed IoT devices enables meaningful pattern discovery, yet conventional centralized approaches require direct access to raw sensor data, leading to privacy risks and substantial communication overhead. In this study, we propose FedAE-KMeans, a federated clustering framework that integrates local autoencoder-based representation learning with centralized KMeans clustering in a shared latent space. We allow each device to train an autoencoder locally to learn compact feature representations, ensuring that sensitive raw data remain on-device while only compressed latent vectors are transmitted to the server. We conducted extensive experiments on MNIST and CIFAR-10 datasets under non-IID data distributions to evaluate clustering effectiveness and robustness. On the MNIST dataset, FedAE-KMeans achieved a silhouette score of 0.677 and a clustering purity of 89.3
