Sparse Learning Research Articles

Order spectrum-assisted sparse learning classification approach for wind turbine drivetrain fault diagnostics under variable operating conditions

CQSA-KT: Research on personalized knowledge tracing based on quantum-constructivism in sparse learning environments

To flexibly handle high-dimensional data, this paper develops a new sparsity-constrained optimization approach. By leveraging group l 0 -norm constrained neural networks, the proposed approach aims to simultaneously extract crucial features and estimate the underlying model function with statistically guaranteed accuracy. Under some mild conditions, we establish statistical theories of the proposed method for a general class of nonparametric regression-type loss functions. Moreover, two iterative greedy selection algorithms, which iterate between a standard gradient descent step and a hard thresholding step with or without debiasing, are presented to implement the computation. Convergence guarantees and sparsity recovery capabilities of these algorithms are rigorously examined. Empirical validation through a series of comprehensive experiments conducted on both real-world and synthetic datasets underscores the superior performance of our proposed estimator. In various scenarios, the proposed approach surpasses conventional Lasso-based sparse learning methods in terms of variable selection accuracy and prediction performance, thereby highlighting its efficacy in practical applications.

Suppressing complex mixed noise in seismic data poses a significant challenge for conventional methods, which often cause signal damage or leave residual noise. While sparse basis learning is a promising approach for this task, traditional data-driven learning methods are often insensitive to the physical properties of seismic signals, leading to incomplete noise removal and compromised signal fidelity. To address this limitation, we propose a physics-constrained sparse basis learning method for mixed noise suppression. Our method integrates local dip attributes—estimated and iteratively refined by a plane-wave destructor filter—as a physical constraint within the dictionary learning framework. This constraint guides the learning process to achieve high-fidelity signal reconstruction while effectively suppressing multiple noise types. Tests on complex synthetic and real data demonstrate that the proposed method outperforms conventional techniques and industry-standard workflows in attenuating mixed noise, including strong anomalous amplitudes, ground roll, and random and coherent components, thereby significantly enhancing the signal-to-noise ratio and imaging quality.

Multimodal emotion recognition leverages multiple modalities to capture emotional cues more comprehensively, thereby improving the accuracy and robustness of emotion recognition. From the perspective of multimodal data and feature learning, reducing information redundancy in multimodal data and enhancing the discriminability of deep feature co-learning can effectively boost recognition performance. Based on this, this paper proposes a multimodal emotion recognition method based on an Adaptive High-order Transformer Network (AHOT). This method constructs Adaptive Selection Transformer block (AST) and Cross-modal Feature Fusion block (CMFF) for each modality branch, aiming to fully capture non-redundant feature representations from each modality and the interactions between modalities. In addition, a sparse high-order feature learning module is designed to enable the learning of highly discriminative high-order features across modalities. Experimental results on two multimodal emotion recognition datasets (IEMOCAP and CMU-MOSEI) demonstrate that, compared with several related methods, the proposed AHOT effectively improves emotion recognition accuracy. Moreover, ablation studies and parameter analyses further validate the effectiveness of AHOT.

Multimodal emotion recognition leverages multiple modalities to capture emotional cues more comprehensively, thereby improving the accuracy and robustness of emotion recognition. From the perspective of multimodal data and feature learning, reducing information redundancy in multimodal data and enhancing the discriminability of deep feature co-learning can effectively boost recognition performance. Based on this, this paper proposes a multimodal emotion recognition method based on an Adaptive High-order Transformer Network (AHOT). This method constructs Adaptive Selection Transformer block (AST) and Cross-modal Feature Fusion block (CMFF) for each modality branch, aiming to fully capture non-redundant feature representations from each modality and the interactions between modalities. In addition, a sparse high-order feature learning module is designed to enable the learning of highly discriminative high-order features across modalities. Experimental results on two multimodal emotion recognition datasets (IEMOCAP and CMU-MOSEI) demonstrate that, compared with several related methods, the proposed AHOT effectively improves emotion recognition accuracy. Moreover, ablation studies and parameter analyses further validate the effectiveness of AHOT.

This study proposes a novel two-step sparse learning framework that combines Sparse Principal Component Regression (SPCR) with regularization methods, Lasso, Elastic Net, Ridge, and Smoothly Clipped Absolute Deviation (SCAD), to improve prediction and interpretability in high-dimensional settings. Simulation experiments were conducted under varying sample sizes, dimensionality levels, sparsity conditions, and predictor correlations to evaluate the performance of the hybrid estimators in comparison to traditional penalization approaches. Results show that SPCR-Lasso and SPCR-Enet consistently deliver superior accuracy and stability in high-dimensional, multicollinear contexts, with SPCR-Enet performing particularly well in extreme dimensionality. SPCR-SCAD demonstrated advantages in sparse, low-correlation scenarios, while Ridge regression contributed modest improvements. These findings underscore that estimator performance is strongly data-dependent and highlight the value of SPCR hybridization for mitigating multicollinearity while enhancing interpretability. The study offers practical guidance for applied researchers in fields such as genomics, finance, and climate science, and contributes methodologically by demonstrating the robustness of SPCR-based regularization in handling complex high-dimensional data structures.

Summary In the process of oil and gas exploration, lithology classification of well log data is crucial for accurately describing subsurface formations and subsequent oil and gas reserve evaluations. Traditional lithology classification methods rely on manual feature extraction and linear models, which have achieved good results in some cases. However, these methods are often constrained by issues, such as data quality, noise interference, and incomplete feature selection, when dealing with complex and variable well log data. To address this challenge, we propose a novel deep learning framework—causal intervention and sparse shift network or CISSNet—aimed at enhancing the accuracy and robustness of lithology classification by introducing causal intervention mechanisms and sparse transfer learning strategies. The core innovation of CISSNet lies in its combination of causal intervention networks and sparse transfer learning networks. First, CISSNet uses a causal intervention framework to capture the causal relationships in well log data, identifying potential causal connections between different lithology categories. By incorporating causal reasoning mechanisms into the model, CISSNet effectively filters out noise factors and highlights key factors influencing lithology classification. Additionally, CISSNet introduces a sparse transfer learning strategy, which shares knowledge and features between source and target domains to overcome the heterogeneity issue of well log data across different well locations, thereby enhancing the model’s crosswell adaptability. In the experimental section, several real well log data sets are used for validation. The results show that CISSNet outperforms traditional classification methods and existing deep learning models in both lithology classification accuracy and robustness. Notably, even in the presence of missing or incomplete data, CISSNet can still maintain high classification precision. Overall, CISSNet provides a new solution for lithology classification by combining causal intervention and sparse transfer learning, not only improving classification accuracy but also enhancing adaptability and stability in complex and incomplete data. Future research will focus on further optimizing the model, improving its generalization ability for different types of well log data, and exploring its potential applications in other geological exploration tasks.

Skeleton2Mask: Skeleton-supervised airway segmentation.

Dynamic load identification based on dual sparse dictionary learning

Multi-scale temporal representation with sparse dynamic graph learning for district heat load forecasting

Advancing the Biological Plausibility and Efficacy of Hebbian Convolutional Neural Networks.

Regularized multi-task learning with individual-feature-based task correlations for Alzheimer's cognitive score prediction.

Kernel Bayesian tensor ring decomposition for multiway data recovery.

Fast identification of maize varieties with small samples using near-infrared spectral feature selection and improved stacked sparse autoencoder deep learning

TransModel-BSDL: Bayesian Sparse Dictionary Learning for Development of Site-Specific Transformation Models Using Limited Data and a Database of Existing Models

A data-driven sparse learning approach to reduce chemical reaction mechanisms

Accurate calling of parental-child SNPs and Indels in family trios is very helpful for understanding genetic traits and diseases. Indel calling is even more important than SNP calling, as Indels may have led to substantial changes in protein structures that affect more of the traits of the organism. However, the best Indel calling methods have recall rates below 85%, precision below 92%, and F1 below 88% on 60times ONT Q20 data, much lower than their SNP calling’s recall performance of 99.87%, precision of 99.86%, and F1 of 99.86%. Difficulties in Indels calling include how to distinguish sequencing errors from genuine Indels and how to optimize the Mendelian genetic model. This work proposes sparse attention learning for high-performance calling of Indels from family-trios’ ONT long-read sequencing data, while still maintaining exceptional performance on SNP calling. Key steps include a sparsely connected attention network to convert fully aligned data cubes into essential features, and a deep learning on these features via ResNet and 3D convolutional blocks to enable accurate detection of family-trio variants. This attention network is in fact a dual attention network to aggregate both channel and spatial information, capable of selecting sub-cubes of critical channels and base locations that are resistant to the confounding effects of sequencing errors. Comparing with the current best-performing trio-variant detection method, our F1 is 5.6%–14.19% higher, recall is 7.07%–18.67% higher, and precision is 3.85%–7.87% higher on ONT Q20 datasets. Case studies of indel-dense regions in chromosome 20, including the centromere and disease-associated genes, demonstrate the significant impact of indel variations on disease pathogenesis, providing novel perspectives for future personalized and targeted therapies.

Underwater visual object tracking is crucial for marine resource exploration and military security. However, due to the effect of insufficient light and turbid background in underwater scenes, efficient and accurate target tracking cannot be realized on underwater edge devices with limited computing resources. To address this problem, we design an underwater object tracking network, namely DBSF, for edge computing devices based on sparse confidence feature learning guided by differential boundary attention. Specifically, we propose a differential boundary attention distribution model to compute the object edge distribution state to enhance the accurate perception of the underwater object edge structure. Then, the differential boundary attention-guided object tracking network learns to perceive the highly discriminative sparse features on the object structure, and computes the object sparse confidence matrix, which reduces the constraints of the edge devices with limited computational resources and ensures the tracking performance. Extensive experiments demonstrate that the DBSF network achieves accurate underwater target recognition and outperforms related advanced methods.

Identification of microbes with large impacts on their microbial communities, known as keystone microbes, is a topic of long-standing interest in microbiome research. However, many approaches to identify keystone microbes are limited by the inherent nonlinearity and state-dependence of microbial dynamics. Machine learning approaches have been applied to address these shortcomings but often require more data than is available for a given microbial system. We propose a keystone identification approach called KeySDL which reduces the amount of data required by incorporating assumptions about the type of microbial dynamics present in the experimental system. The data are modeled as originating from a Generalized Lotka-Volterra (GLV) model, an architecture commonly used to simulate microbial systems. The parameters of this model are then estimated using Sparse Dictionary Learning (SDL) Compared to existing methods, this approach allows accurate prediction of keystone microbes from small numbers of samples and provides an output interpretable as reconstructed system dynamics. We also propose a self-consistency score to help evaluate whether the assumption of GLV dynamics is reasonable for a given dataset, either through the application of KeySDL or other analysis tools validated using GLV simulation.