With the widespread adoption of multi-view data in numerous fields, multi-view unsupervised feature selection (MUFS) has made notable strides in both feature pruning and missing-view completion. Nonetheless, existing MUFS methods typically rely on centralized servers, which cannot meet real-world demands for privacy preservation and distributed learning, and they often suffer from suboptimal solution and weak convergence guarantees. To address these challenges, IMUFFS, an incomplete multi-view unsupervised federated feature selection via cooperative particle swarm optimization (CPSO) and tensor-aligned learning (TAL) is proposed. Specifically, each client executes CPSO-TAL at two stages: (i) an external optimization phase that involves a CPSO, inspired by the co-evolutionary mechanism of hybrid breeding optimization algorithm, performing a global search in the feature space, and (ii) an internal optimization phase that leverages TAL with imputation and CP decomposition, where CP decomposition reduces dimensionality by decomposing the original tensor into a sum of core components, to learn low-dimensional embeddings, while simultaneously updating anchor graphs and view preference weights, thereby harmonizing imputation and representation learning. On the server side, a federated aggregation strategy using adaptive normalized mutual information (NMI) weighting combines the locally optimized feature selection (FS) weights and NMI scores from clients, ensuring privacy while improving the quality of FS and convergence. Extensive experiments on multiple datasets demonstrate that IMUFFS consistently outperforms state-of-the-art methods, yielding more effective and robust FS and enhancing better missing-view completion.

Band selection (BS) is an effective way for reducing redundancy and enhancing processing efficiency by obtaining a subset with highly discriminative bands. However, existing methods focus on the correlations between neighboring bands, overlooking the nonlinear dependencies present in global contexts. Moreover, these approaches typically treat each band as an independent entity, disregarding the spatial structure and potential spectral-spatial information embedded within band combinations. To address these limitations, we propose S2HGC, an end-to-end spectral-spatial hypergraph convolutional network designed for unsupervised hyperspectral BS. Our method eliminates error accumulation from stepwise processing and enables the holistic optimization of band subset. Specifically, a controllable autoencoder (CAE) is designed to reflect the low-dimensional representation of full bands and partition them into multiple combinations. Furthermore, spatial topology is produced from these band combinations using superpixels, and a spectral-spatial information fusion (S2IF) module is designed to extract spatial features and integrate them with spectral features. Finally, an adaptive hypergraph convolution is employed to capture the nonlinear correlations between bands and enhance the information interaction across long-distance bands. The critical bands are then selected to reconstruct the original hyperspectral imagery, which is iteratively optimized to obtain the optimal band subset. Extensive experiments on three publicly available datasets demonstrate the superior performance of S2HGC compared to other state-of-the-art BS methods.

Band selection (BS) is one of the effective ways to reduce data dimensionality while retaining important information in hyperspectral imagery (HSI). The ultimate goal is to represent the original HSI using as few informative and discriminative bands as possible. However, existing BS methods usually divide all bands into independent spaces while neglecting the relationships between these spaces. Moreover, the evaluation of band importance through its neighbors requires more time to capture global information from the spectral perspective. Therefore, a dynamic-interactive subgraph Transformer (DISGT) is proposed for unsupervised hyperspectral BS. Specifically, a dynamic distance norm is designed to measure the correlation between bands and characterize several subspaces with better separability. Bands with uncertain subspace membership are shared with other subspaces through information interaction to enhance the diversity of band combinations, compensate for the effect of subspaces on global information processing, and construct subspaces into subgraphs. Furthermore, the subgraph Transformer (SGT) captures local and global spectral information using a sparse multihead attention (SparseMHA) mechanism to accurately evaluate band importance. Finally, the original HSI is reconstructed using the crucial bands to continuously adjust and obtain the optimal band subset. The superior performance of the DISGT is demonstrated through extensive comparisons with other state-of-the-art methods on three public datasets.

In the realm of high‐dimensional data analysis, numerous fields stand to benefit from its applications, including the biological and medical sectors that are crucial for computer‐aided disease diagnosis and prediction systems. However, the presence of a significant number of redundant or irrelevant features can adversely affect system accuracy and real‐time diagnosis efficiency. To mitigate this issue, this paper proposes two innovative wrapper feature selection (FS) methods that integrate the ant colony optimization (ACO) algorithm and hybrid rice optimization (HRO). HRO is a recently developed metaheuristic that mimics the breeding process of the three‐line hybrid rice, which is yet to be thoroughly explored in the context of solving high‐dimensional FS problems. In the first hybridization, ACO is embedded as an evolutionary operator within HRO and updated alternately with it. In the second form of hybridization, two subpopulations evolve independently, sharing the local search results to assist individual updating. In the initial stage preceding hybridization, a problem‐oriented heuristic factor assignment strategy based on the importance of the knee point feature is introduced to enhance the global search capability of ACO in identifying the smallest and most representative features. The performance of the proposed algorithms is evaluated on fourteen high‐dimensional biomedical datasets and compared with other recently advanced FS methods. Experimental results suggest that the proposed methods are efficient and computationally robust, exhibiting superior performance compared to the other algorithms involved in this study.

Ship classification using Synthetic Aperture Radar induces effective marine applications but suffers from imbalanced datasets. Common ship types (majority classes) contain many more instances than rare ship types (minority classes), resulting in performance loss since minority instances tend to be ignored. Meanwhile, it is difficult and time-consuming to produce reliable instances for rare ships. The synthetic minority oversampling techniques have shown great potential to balance the distribution of classes by synthesizing minority instances. However, the newly synthesized instances may overlap with the majority instances, reducing the separability among classes, or being far away from the classification boundary, which is meaningless. This paper proposes a clustering-based size-adaptive safer oversampling technique to address the imbalanced classification problem. Proven schemes, including cluster minority instances, adaptively allocate oversampling sizes, and assign weights are adopted for selecting minority instances. Then, new instances are synthesized according to the safe metric of selected instances, instead of randomly inserted. Furthermore, multiple handcrafted features are tested to provide a clearer classification boundary. The proposed comprehensively considers the usefulness and safety of synthesizing instances. Experiments on the OpenSARShip and the FUSAR-Ship datasets demonstrate that the proposed technique achieves significantly better results.

Terrestrial laser scanning (TLS) is an important technique for tree stem detection. In this paper, a point-based method for stem detection is proposed using single-scan TLS data. One of the main concerns is the point density, which decreases rapidly with the increasing distance to the scanner position. In the proposed method, the search radius is generated adaptively, based on the relationship between the distance and point density, to make sure that the neighborhood maintains a similar scale to the corresponding point density. The belonging of each point is recognized with cuckoo search-based support vector machine, and the points labeled as stem are then clustered and filtered for further verification. The threshold for the small cluster filtering is also adaptive to deal with the problem of the cluster point number decreasing as a function of distance. The stem position is calculated with the lowest cylinder from the cluster segmentation and modeling for the stem mapping. Experiments were carried out on two plots with radii of more than 130 m. The overall detection rate was 76.1%, and 75% of the stems outside 80 m were detected with the adaptive radius, despite the point density being less than 5 cm.