Search Results (153)

Satellite array receivers must suppress interference whose spatial, spectral, and temporal characteristics vary across operating conditions, often with limited known symbols for adaptation. This paper studies when learned subspace augmentation is useful under finite pilot support. The receiver retains the instantaneous complex-linear receive subspace and adds a low-dimensional orthogonal complement generated from temporal and spectral features, with all basis coefficients estimated jointly from known symbols. A finite-sample analysis separates oracle representation gain from the additional pilot-to-payload estimation error caused by basis enlargement. Across 720 seen-family conditions, the proposed receiver achieves mean uncoded bit error rate (BER) 0.06521, versus 0.06572 for the receive-span ablation and 0.06719 for minimum-variance distortionless response; the 0.78% reduction relative to receive-span is modest and concentrated in spatially loaded narrowband regimes. A same-frame trained-versus-untrained comparison shows an average benefit from parameter optimization in the reference realization. Independent retraining from three initializations yields closely clustered held-out performance, while the selected model order varies between $L=1$ and $L=2$. For three unseen interference families, mean net augmentation gain is negative and aggregate MSE does not improve over receive-span. The results therefore identify a condition-dependent regime in which subspace augmentation is useful when representation gain exceeds finite-pilot estimation cost, rather than a broadly transferable advantage across interference types. Full article

Multi-view clustering seeks a consensus partition from heterogeneous feature views while retaining view-specific information. We propose SGLog-${\gamma }^{\ast }$-MSC, a nonconvex multi-view subspace clustering framework that combines graph total variation, an ${\mathsf{\ell }}_{2,\log }$ penalty, and a tensor ${\gamma }^{\ast }$ spectral penalty. These terms promote locally consistent self-representations, model sample-wise corruption, and capture shared low-rank structure across views, respectively. We derive an alternating augmented-Lagrangian algorithm with candidate-selection rules that return global minimizers for both nonconvex proximal subproblems. We also state explicit conditions under which accumulation points satisfy the KKT system and establish conditional whole-sequence convergence through the Kurdyka–Łojasiewicz framework. Hyperparameters are selected by a label-free protocol specified before the audited rerun; ground-truth labels are never used during selection. Across 20 recorded k-means++ initializations of each fixed embedding, the method attains NMI $0.8206\pm 0.0155$ on Yale and $0.8962\pm 0.0211$ on Scene-15. Relative to nine literature-reported baselines, these are the highest reported NMI values on the two datasets. On UCI digits, BBCSport, and ORL, the method reaches NMI $0.9837$, $0.9634$, and $0.9892$, respectively, ranking second only to HLR-M²VS in the descriptive cross-paper comparison. Component-wise ablation identifies the ${\mathsf{\ell }}_{2,\log }$ term as the largest and most consistent contributor, while the effects of the ${\gamma }^{\ast }$ surrogate and sparse-gradient term depend on the dataset. Compared with the tensor nuclear norm, the ${\gamma }^{\ast }$ surrogate improves performance on BBCSport, ORL, and UCI digits; sensitivity analysis shows that an interior $\gamma$ also improves performance on Scene-15. Together, these results support combining robust error modeling with local and shared structural regularization while emphasizing the dataset dependence of individual components. Full article

Generalized category discovery (GCD) aims to classify known categories while discovering novel ones in unlabeled data, yet existing methods lack mechanisms to correct boundary-ambiguous samples that receive noisy pseudo-labels, as they primarily rely on visual feature learning without external semantic guidance. Vision-language models (VLMs) offer a natural source of cross-modal semantic correction. However, applying VLM-guided contrastive signals directly within the GCD training loop proves counterproductive because the locally-oriented InfoNCE loss conflicts geometrically with the globally oriented K-means objective in the shared backbone space. We identify the root cause as a dual resource allocation problem: the VLM-derived signal must be allocated to the correct feature subspace to avoid geometric conflict with K-means clustering (space allocation), and the limited VLM inference budget must be allocated to the correct samples to maximize discriminative return (budget allocation). These two decisions are coupled; failure on either renders the other ineffective. To resolve this, we propose DualGLEAN, a framework that addresses the dual allocation challenge through two coupled mechanisms: decoupled contrastive alignment (DCA), which routes the VLM-guided neighbor contrastive loss to a dedicated projector space while preserving the backbone space for global clustering, and compound uncertainty querying (CUQ), a three-stage filtering metric that jointly evaluates predictive entropy, boundary proximity, and local label inconsistency to direct VLM queries exclusively to truly boundary-critical samples. Extensive experiments on the AID and RSSDIVCS datasets demonstrate that DualGLEAN achieves strong performance, improves four diverse GCD baselines as a plug-in module, generalizes across seven VLM backbones, introduces zero additional trainable parameters to the base GCD network, and incurs a total VLM API cost of only CNY 2.45 per full training run on the AID dataset under the default search-scope configuration, with the cost scaling linearly with the query budget. Full article

Supervised learning in fields such as genomics and medical imaging is often hindered by the high cost of expert data annotation. Active learning addresses this bottleneck by iteratively selecting the most informative unlabeled samples for labeling. However, in high-dimensional environments, traditional diversity-based query strategies lose their effectiveness due to the degradation of global distance metrics. To address these challenges, this paper proposes a novel framework, Active Learning via Subspace Ensembles and Similarity (ALSES). Instead of relying on global distances, ALSES constructs a similarity matrix by sampling an ensemble of random feature subspaces. The subspaces are filtered based on their discriminative power, and pairwise sample similarities are aggregated using cluster co-occurrence. This structural representation is integrated into a hybrid batch selection strategy that balances model uncertainty and data representativeness. Extensive evaluations on simulated datasets and real-world high-dimensional cancer cohorts demonstrate that ALSES consistently outperforms standard active learning baselines. The framework effectively isolates informative variables and achieves superior classification accuracy with significantly fewer labeled instances, demonstrating its robustness in complex, noisy applications. Full article

In the domain of rotating machinery fault diagnosis, challenges such as multi-operating condition distribution heterogeneity and the difficulty of distinguishing fault features within multi-scale temporal signals persist. To address these issues, this paper introduces the Time-Gated Multi-Expert Generative Adversarial Network (TGME-GAN), a fault diagnosis approach that integrates a multi-expert gated conditional generative adversarial network with a clustering structure-aware feature enhancement. This method combines unsupervised K-means clustering with supervised discriminative learning. The optimal number of clusters is selected adaptively using the silhouette coefficient, and the distance vector from each sample to the cluster centers serves as a topological prior feature. A spatial–temporal joint representation matrix is then formed by concatenating PCA principal components, differential features, cumulative statistical features, and standardized change rates, which together capture both abrupt mutations and progressive degradation in fault signals. In the model, the discriminator incorporates a multi-expert gated network. Each expert learns a feature subspace corresponding to a distinct operating condition, and the gated network dynamically assigns fusion weights, allowing the discriminator to capture heterogeneous distributions across industrial conditions. The generator extracts multi-scale local patterns with a three-layer one-dimensional convolutional network and models sequential dependencies with a two-layer LSTM, producing high-quality fault samples that preserve intrinsic consistency. At the engineering level, TGME-GAN is deployed for gearbox fault diagnosis in uneven, small-sample industrial settings. In two gearbox fault experiments, this method substantially outperforms current mainstream models. Full article

Time-fractional diffusion equations describe subdiffusive transport in heterogeneous media, but their numerical treatment is complicated by the nonlocal Caputo derivative and by the weak singularity that the solution develops at the initial time. We study a global-in-time discretization that combines spectral collocation in time—on the fractional power basis ${\left\{t^{\mathsf{\ell }\alpha }\right\}}_{\mathsf{\ell }=0}^N$, evaluated at Chebyshev–Gauss–Lobatto nodes, which reproduces the leading terms of the singular expansion of the solution—with a second-order conservative finite-difference stencil in space that uses harmonic averaging of the diffusivity at the cell faces and therefore remains accurate across discontinuous media. The resulting fully discrete problem is a large, nonsymmetric, dense-in-time linear system whose two-norm condition number grows like the inverse square of the spatial mesh size, so that Krylov subspace iteration without preconditioning stalls under refinement. Exploiting the Kronecker sum structure of the discrete operator, we build a preconditioner by fast diagonalization of the spatial factor through the discrete sine transform. For constant diffusivity the preconditioner reproduces the operator exactly and yields a direct solver; for variable diffusivity it is spectrally equivalent to the operator, and we prove that the eigenvalues of the preconditioned system cluster in a disk centered at one whose radius depends only on the coefficient contrast, and not on the mesh, the number of temporal degrees of freedom, or the fractional order. Numerical experiments in one and two space dimensions confirm second-order spatial accuracy and a preconditioned iteration count that stays flat—twelve iterations from $M=32$ up to $M=1024$ in one dimension and eleven up to $M=256$ per direction in two—while the unpreconditioned count grows by more than two orders of magnitude. In time, the accuracy is spectral until round-off in the ill-conditioned Vandermonde matrix of the power basis takes over: the barrier is reached at $N=9,10,13$ for $\alpha =0.3,0.5,0.7$, where the attainable error is about ${10}^{-6}$. A benchmark against the L1 scheme on uniform and graded meshes, the Alikhanov L2-$1_{\sigma }$ scheme and Grünwald–Letnikov convolution quadrature quantifies when the global approach pays: on forced problems and on modes with $\kappa \lambda T^{\alpha }\lesssim 2$ it reaches a prescribed accuracy one to two orders of magnitude faster and with several times less memory, while for strongly damped modes the fractional power basis converges only algebraically and graded time marching is preferable below a relative error of ${10}^{-2}$. Full article

Open wireless propagation makes unmanned aerial vehicle (UAV) links vulnerable to eavesdroppers distributed over roads, building clusters, or other continuous regions. This paper proposes a low-rank threat-subspace method for cooperative secure beamforming from distributed ground transmitters to a legitimate UAV. Steering vectors sampled over one or multiple azimuth–elevation threat regions are concatenated into a training matrix, whose dominant left singular vectors compactly represent regional exposure. The legitimate steering vector is projected onto the orthogonal complement of this subspace and power-normalized. We prove global optimality of the normalized projection for every feasible retained rank, derive a leakage bound from the first discarded singular value, and introduce uncertainty padding for independently mismatched region boundaries. Simulations evaluate disconnected and volumetric regions, deterministic geometries, Rician scattering, channel and phase errors, non-colluding and colluding eavesdroppers, and covariance-reconstruction and sampled peak-leakage baselines. In the default setting, six modes retain 98% of the sector energy, and the proposed method reduces average leakage to −26.25 dB, compared with −19.41 dB for pointwise nulling and −9.89 dB for maximum-ratio transmission. Independent boundary-error tests show that interval padding stabilizes leakage at the cost of desired gain, while multi-region tests quantify the progressive increase in effective rank. The results establish both the applicability limits and the low-overhead advantages of spatial-structure-based secure beamforming. Full article

Traumatic brain injury (TBI) is associated with heterogeneous neurophysiological and behavioral outcomes that are not always captured by conventional clinical classifications. This study aimed to characterize electrophysiological patterns in individuals with TBI and examine their alignment with clinical groupings. EEG, event-related potentials (ERPs), spectral features, and behavioral performance were analyzed in 121 participants from a larger clinical cohort. Data were preprocessed using artifact subspace reconstruction (ASR), adaptive mixture independent component analysis (AMICA), and standardized quality control procedures. Spectral power, ERP components, and behavioral metrics were extracted and compared across clinical severity and OSU TBI-ID classifications, and unsupervised k-means clustering was applied to identify data-driven electrophysiological subgroups. Results demonstrated consistent ERP signal quality across participants, with all datasets meeting P300 global field power thresholds. While group-level comparisons revealed distributions of electrophysiological markers across clinical categories, clustering identified two electrophysiological subgroups characterized by differing spectral, connectivity, and ERP profiles that were not significantly associated with OSU TBI-ID classifications. Electrophysiological subgroup assignments were not significantly associated with OSU TBI-ID classifications. These findings suggest that data-driven electrophysiological subgrouping captures variability not reflected in traditional clinical groupings, highlighting the potential utility of electrophysiological measures for characterizing heterogeneity within chronic TBI populations. Full article

We propose a three-functional-layer decomposition framework for underwater acoustic (UWA) integrated sensing and communication (ISAC) preambles, instantiated as P5. Two spectrally separated chirp layers—hyperbolic frequency modulation (HFM) for wideband Doppler invariance and linear frequency modulation (LFM) for sub-meter ranging—are carried under a common constant-amplitude zero-autocorrelation (CAZAC) envelope that supplies cell identification and despreading against a root-blind attacker. Closed-form screening conditions constrain the layers to a near-orthogonal subspace, and direct cross-ambiguity measurement confirms the realized separation. In matched-filter Monte Carlo simulation, P5 meets the continuous-sensing target (range root mean square error ${\sigma }_R\le 1$ m at 10 dB signal-to-noise ratio) and has the smallest normalized matched-filter peak loss across twelve modeled UWA environments among four tested waveforms. Against four classical structure-aware attackers it stays below the strict $P_d\le 0.1$ low-probability-of-intercept target at 0 dB attacker-input SNR. A 10-seed, 11.2-million-parameter spectrogram ResNet-18 reaches $P_d=0.5$ against P5 at mean $+21.24$ dB total-energy SNR (95% CI $[+21.06,+21.42]$ dB) and $P_d=0.1$ at $+18.72$ dB ($[+18.45,+19.00]$ dB); these crossings are lower bounds on adversary capability, not a security guarantee. The integration also has explicit costs: composite peak-sidelobe level ($-7.60$ dB default, $-11.29$ dB optimized) remains inferior to equal-aperture single-waveform baselines, and sixteen-cell identification falls to ≤$0.07$ under a $+2$ dB near–far interferer. All-60-sounding WATERMARK replay further gives adverse P5^def–B5 losses of $-0.816$ dB on NOF1 (sounding-cluster 95% CI $[-1.021,-0.621]$ dB) and $-0.950$ dB on NCS1 ($[-0.977,-0.923]$ dB) after all waveforms are scaled into the same measured 8-kHz band. The evidence is therefore simulation dominant and supplemented by measured-channel replay of band-scaled variants; native-band transducer, pool, and sea-trial validation remain future work. Full article

Self-supervised graph collaborative filtering suffers from two geometric pathologies. In Dimensional Collapse, embeddings quietly collapse into a low-rank subspace, a well-documented but poorly solved problem. Semantic Collapse is more complex: stiff $L^2$-squared orthogonalization penalties that are supposed to push the embeddings apart end up ripping through the heavy-tailed community overlaps that contain the collaborative signal. Earlier studies attempted to mitigate sparsity by injecting static noise or structural perturbations, but such interventions did not pinpoint the root cause, i.e., the distortions in the global covariance geometry itself. In this paper, we propose a Huber-Contrastive Graph Convolutional Network (HCGCN) that combines a spatial message-passing backbone with an $O(1)$ contrastive augmentation overhead and a Relation-Aware Dual-View Gated Contrastive Network. The main novelty is a Huber Covariance Whitening module that imposes a geometry-aware threshold on the cross-correlation matrix of augmented views—below the threshold, the gradients follow an $L^2$ penalty (enforcing uniformity); above it, the penalty flattens to $L^1$ (protecting genuine semantic clusters from gradient explosion). This theoretically motivated dual-regime penalty actively preserves the macro-semantic topology of the graph while aggressively stamping out spurious noise correlations. The HCGCN is evaluated on the Yelp2018, Amazon-Book, and MovieLens datasets and performs significantly better than state-of-the-art baselines like LightGCN, SGL, SimGCL, and NESCL, especially under severe cold-start settings where covariance regulation proves most critical. Full article

Single-cell RNA sequencing data are characterized by high dimensionality, sparsity, and strong nonlinearity, hindering conventional single-view clustering methods from capturing linear and nonlinear feature subspaces simultaneously. Features from distinct dimensionality reduction approaches are inherently complementary: PCA (Principal Component Analysis) preserves global linear structures, UMAP (Uniform Manifold Approximation and Projection) maintains topology and local neighborhoods, and PHATE (Potential of Heat-diffusion for Affinity-based Trajectory Embedding) depicts gradual transitions in cell differentiation. To fuse these complementary sources, we adopt an inter-view correlation maximization paradigm. Canonical Correlation Analysis (CCA) integrates two views by maximizing projection correlation but is limited to pairwise scenarios. We extend it to Generalized Canonical Correlation Analysis (GCCA) for multi-view alignment and introduce a deep autoencoder to construct the DeepGCCA (Deep Generalized Canonical Correlation Analysis) framework. This method generates three views via PCA, UMAP, and PHATE, extracts nonlinear latent features with the autoencoder, projects multi-view representations into a unified subspace under weighted GCCA constraints, and performs K-means clustering. Experiments on the two simulated and three real single-cell datasets evaluated in this study show that DeepGCCA demonstrates competitive performance against all single-view baselines and performs favorably compared to several widely adopted methods. Moreover, downstream marker gene analysis supports the biological interpretability of the resulting clusters within these datasets. Within the scope of this benchmark, DeepGCCA provides a valuable reference for high-precision clustering of single-cell transcriptomic data, offering practical insights into multi-view integration and biological interpretability. Full article

Efficient multi-hop prediction over large-scale spatio-temporal knowledge graphs of aircraft taxiing trajectories remains challenging, as existing methods focus either on static multi-hop relations or on accuracy improvement for spatio-temporal single-hop predictions, leading to computational inefficiency. This paper proposes a vector-index-supported multi-hop prediction method. First, a knowledge graph embedding technique that integrates spatio-temporal features maps the trajectory graph into a low-dimensional complex vector space. Then, a hierarchical query acceleration structure based on IndexIVFFlat is constructed. A clustering strategy guided by the distribution of trajectory data partitions the vector space into subspaces, and approximate nearest neighbor search within those subspaces rapidly prunes the candidate set to accelerate multi-hop retrieval. Experiments on real aircraft taxiing trajectory datasets and general benchmarks show that the proposed method substantially improves prediction efficiency while maintaining competitive accuracy. The results demonstrate that the vector index mechanism effectively balances accuracy and efficiency, and the efficiency has been improved by at least 56.65%. This work provides a key technical foundation for real-time analysis and intelligent prediction of large-scale aircraft taxiing trajectories. Full article

The exponential growth of social media has revolutionized global communication, enabling instant idea exchange and transforming information sharing into a worldwide phenomenon while simultaneously accelerating the spread of abusive and hateful content that threatens online harmony and poses a serious risk to online community integrity and public trust. Although supervised deep learning approaches achieve impressive accuracy for hate speech detection, they remain fundamentally reliant on extensive annotated corpora, and their lack of interpretability makes them insufficient for transparent and scalable real-world hate speech detection. This study presents a category-oriented unsupervised architecture for English hate-speech detection and classification that substantially reduces reliance on large labeled datasets by requiring only minimal supervision (10% of labels for post hoc cluster interpretation), ensuring transparency and a high degree of semantic interpretability. We introduce an unsupervised Subspace Weighting Co-Clustering framework that uses HateBERT-driven contextual embeddings, enabling simultaneous interpretable feature weighting and semantic understanding for robust hate-speech detection. The obtained embeddings are further structured using the Subspace Weighting Co-Clustering approach, which enables the unsupervised discovery of latent subspaces and the organization of tweets into semantically coherent hate categories. The comprehensive evaluation shows that the framework achieves superior accuracy over existing methods, providing a more robust and effective mechanism for digital platforms to identify and mitigate hate speech and promote safer online interactions. Full article

Anchor-based bipartite graph methods provide scalable solutions for multi-view clustering, but most of them construct graphs in the original feature space, where high dimensionality distorts the proximity between samples and anchors and degrades graph quality. In addition, the K-means step commonly used to discretize spectral embeddings may produce different cluster assignments across random seeds. To address these limitations, this paper proposes projection-enhanced bipartite graph learning (PEBGL), which first projects each view onto a compact PCA subspace and then jointly performs bipartite graph construction, consensus graph fusion with adaptive view weighting, spectral embedding, and discrete label assignment within an alternating optimization framework. Most subproblems admit closed-form or efficient projection-based updates, and the final labels are obtained by connected-component detection on the learned consensus graph, reducing the dependence on K-means post-processing. Experiments on six benchmark datasets demonstrate that PEBGL achieves competitive clustering performance against recent graph-based and bipartite graph-based methods. These results validate the effectiveness of the proposed framework. Full article

Existing multi-view clustering approaches based on matrix factorization often fail to jointly capture global high-order correlations and local view-specific characteristics, and they typically suffer from instability in generating final clustering labels. To overcome these limitations, this paper presents a multi-view subspace clustering method termed dual-tensor constrained multi-view subspace clustering (DTCMVSC). Specifically, for each view, we learn an independent latent representation matrix, a projection matrix, and a basis matrix. The latent representations and projection matrices are stacked into third-order tensors, upon which tensor nuclear norm regularization is imposed to simultaneously exploit consensus structures and complementary information across views. Additionally, a consensus regularization term and adaptive view weights are introduced to align the latent representations of different views toward a unified consensus subspace. The resulting optimization problem is efficiently solved under the ADMM framework, after which a similarity matrix is constructed from the consensus representation and spectral clustering is performed to obtain the final labels. Experimental evaluations on six benchmark datasets demonstrate the superiority of DTCMVSC. Specifically, it achieves an ACC of 86.10% on CMU and an NMI of 94.17% on ORL, surpassing even the lowest-performing state-of-the-art baselines by 63.08 and 18.53 percentage points, respectively. Full article