Unsupervised feature selection aims to extract discriminative feature subsets from high-dimensional data, which typically exhibit intricate, multifaceted structures encompassing both local geometry and global semantics. Existing spectral-embedded methods often leverage multiple objectives to capture these distinct perspectives. However, using a single representation for manifold preservation and linear regression can lead to competing requirements on nonlinear or heterogeneous data. To address this issue, we propose a decouple-then-align framework, termed DTA-UFS, that seeks consensus between spectral embedding and pseudo-labels to coordinate their complementary structural information. The framework separates the graph-smoothed spectral representation from the relaxed cluster indicator used as the sparse-regression target. An explicit semi-orthogonal alignment parameterization records the correspondence between their latent bases, through which manifold information guides feature selection via the learned pseudo-labels. Theoretically, we derive an upper bound on the consensus offset between the learned pseudo-label representation and an ideal cluster space, linking the alignment error and manifold regularization term in the objective to the quality of the learned representation under graph perturbation. Experiments on synthetic and real-world datasets demonstrate competitive clustering performance against representative unsupervised feature-selection baselines.
Medical time series, such as electrocardiograms (ECG) and electroencephalograms (EEG), exhibit transient waveform changes and drifting spectral patterns, requiring models to integrate heterogeneous cues for reliable analysis. Existing studies typically analyze time series in either the time or frequency domain, each capturing only a subset of the underlying physiology. Although time–frequency fusion has been explored, existing methods typically integrate domains without fine-grained alignment across time, frequency and leads, limiting their ability to represent the evolving and non-stationary dynamics of long multi-lead recordings. To address these limitations, we propose DEFT (Dual-domain multi-rEsolution Fusion Transformer), a novel dual-domain framework that explicitly couples raw waveform dynamics with multi-resolution spectral evidence. DEFT decomposes each recording into slowly varying components reflecting baseline rhythms and band-limited details preserving transient morphology. These components drive two coordinated streams: a time-domain stream for beat-level waveform dynamics and a frequency-domain stream capturing scale-specific spectral evolution and cross-lead interactions. An interaction patch embedding integrates both streams with strict alignment, enabling unified interpretation of morphology within its surrounding spectral and spatial context. Experiments on large-scale ECG and EEG datasets demonstrate that DEFT surpasses strong uni-domain and fusion baselines, underscoring its robustness and generality for non-stationary clinical time series analysis.
