Dimensionality Reduction Research Articles

Intelligent Fruit Detection System Using Optimized Hybrid Deep Learning Models

Accurate and efficient detection of dragon fruit ripeness is crucial for optimizing harvesting schedules, reducing post-harvest losses, and ensuring fruit quality. This research investigates applying optimized hybrid deep learning (DL) models for intelligent dragon fruit ripeness classification using a dataset of 2,563 images. The feature extraction using pre-trained CNNs, specifically DenseNet-50 and ResNet-50, followed by dimensionality reduction using Principal Component Analysis (PCA). The reduced feature sets are then fed into various classifiers, including Support Vector Machines (SVM) with linear and RBF kernels, a Voting ensemble of SVMs, and a Multi-Layer Perceptron (MLP). The performance of models is evaluated using key metrics such as accuracy, AUC, etc. The experimental findings indicate that the DenseNet-50 features combined with PCA and an SVM Voting ensemble achieve the highest classification accuracy of 97.71%, along with a balanced recall, precision, and F1-score of 0.96. The ResNet-50 features coupled with an MLP also exhibit competitive performance.

A Large Language Model-Powered Map of Metabolomics Research.

We present a comprehensive map of the metabolomics research landscape, synthesizing insights from over 80,000 publications. Using PubMedBERT, we transformed abstracts into 768-dimensional embeddings that capture the nuanced thematic structure of the field. Dimensionality reduction with t-SNE revealed distinct clusters corresponding to key domains, such as analytical chemistry, plant biology, pharmacology, and clinical diagnostics. In addition, a neural topic modeling pipeline refined with GPT-4o mini reclassified the corpus into 20 distinct topics─ranging from "Plant Stress Response Mechanisms" and "NMR Spectroscopy Innovations" to "COVID-19 Metabolomic and Immune Responses." Temporal analyses further highlight trends including the rise of deep learning methods post-2015 and a continued focus on biomarker discovery. Integration of metadata such as publication statistics and sample sizes provides additional context to these evolving research dynamics. An interactive web application (https://metascape.streamlit.app/) enables the dynamic exploration of these insights. Overall, this study offers a robust framework for literature synthesis that empowers researchers, clinicians, and policymakers to identify emerging research trajectories and address critical challenges in metabolomics while also sharing our perspectives on key trends shaping the field.

MPN-RRT*: A New Method in 3D Urban Path Planning for UAV Integrating Deep Learning and Sampling Optimization

The increasing deployment of unmanned aerial vehicles (UAVs) in complex urban environments necessitates efficient and reliable path planning algorithms. While traditional sampling-based methods such as Rapidly exploring Random Tree Star (RRT*) are widely adopted, their computational inefficiency and suboptimal path quality in intricate 3D spaces remain significant challenges. This study proposes a novel framework (MPN-RRT*) that integrates Motion Planning Networks (MPNet) with RRT* to enhance UAV navigation in 3D urban maps. A key innovation lies in reducing computational complexity through dimensionality reduction, where 3D urban terrains are sliced into 2D maze representations while preserving critical obstacle information. Transfer learning is applied to adapt a pre-trained MPNet model to the simplified maps, enabling intelligent sampling that guides RRT* toward promising regions and reduces redundant exploration. Extensive MATLAB simulations validate the framework’s efficacy across two distinct 3D environments: a sparse 200 × 200 × 200 map and a dense 800 × 800 × 200 map with no-fly zones. Compared to conventional RRT*, the MPN-RRT* achieves a 47.8% reduction in planning time (from 89.58 s to 46.77 s) and a 19.8% shorter path length (from 476.23 m to 381.76 m) in simpler environments, alongside smoother trajectories quantified by a 91.2% reduction in average acceleration (from 14.67 m/s² to 1.29 m/s²). In complex scenarios, the hybrid method maintains superior performance, reducing flight time by 14.2% and path length by 13.9% compared to RRT*. These results demonstrate that the integration of deep learning with sampling-based planning significantly enhances computational efficiency, path optimality, and smoothness, addressing critical limitations in UAV navigation for urban applications. The study underscores the potential of data-driven approaches to augment classical algorithms, providing a scalable solution for real-time autonomous systems operating in high-dimensional dynamic environments.

A Data-driven Method for IGBT Open-Circuit Fault Diagnosis of NPC Inverters in Three-Phase Photovoltaic Grid-Connected Systems

Abstract Data-driven methods have exhibited excellent performance and promising prospects in the fault diagnosis of power electronics systems. This article proposed an open-circuit fault diagnosis method based on locally linear embedding (LLE) and regularized extreme learning machine (RELM) for insulated gate bipolar transistor (IGBT) single-tube open-circuit faults of three-level neutral point clamped (NPC) inverters in three-phase photovoltaic grid-connected power generation systems. By building a simulation model of the photovoltaic power generation system, the A-phase output current sample data of IGBT single-tube open-circuit faults under different light intensities are obtained. Subsequently, the sample data is processed by dimensionality reduction and used for the training of the fault diagnosis model. Through optimizing the dimensionality of LLE reduction and the number of hidden layers in the RELM, the model attains optimal diagnostic speed and accuracy. Finally, comparative experiments with other machine learning algorithms demonstrated that the method proposed in this article effectively improved the performance of the IGBT single-tube open-circuit fault diagnosis system for three-level photovoltaic NPC inverters.

Analysis of different IDS-based machine learning models for secure data transmission in IoT networks

Abstract The Internet of Things (IoT) encompasses a network of interconnected devices that collect, analyze, and exchange vast amounts of data. However, this connectivity creates opportunities for various types of cyberattacks, making IoT systems vulnerable and potentially leading to the compromise of sensitive information. Therefore, developing effective intrusion detection system (IDS) is one of the key challenges in IoT network security. The aim of this study is to develop a machine learning (ML) model for network traffic classification and attack detection in IoT environments. Through a comparative analysis of different algorithms, the study seeks to identify the model with the best performance, which could serve as a foundation for efficient IDS solutions tailored to the specific characteristics of IoT networks. The RT-IoT2022 dataset was used for experimental analysis, providing realistic framework for testing ML models, including k-nearest neighbors, Random Forest, XGBoost, multilayer perceptron, and various 1D convolutional neural network architectures. The study examines preprocessing techniques, focusing on dimensionality reduction (principal component analysis, variance inflation factor, Pearson’s test), outlier detection (interquartile range, Z-score, Isolation Forest), and transformation methods (Box–Cox, RobustScaler, Winsorization). Based on the results of the experiment, the most effective model and preprocessing technique were proposed.

Fast 3-D forward modelling of gravity field vector and its gradient tensor using FFT-based spectral method

Abstract In gravity anomalies forward modeling, traditional numerical approaches necessitate solving a complex linear system characterized by sparse equations through matrix inversion. Meanwhile, the calculated three-component gravity field and five-component gravity gradient, utilizing finite-difference or finite-element methods, may experience accuracy loss due to inherent numerical differentials in the gravitational potential. To overcome this issue, we present a fast Fourier transform (FFT)-based spectral approach for solving the 3-D gravitational potential Poisson equation associated with homogeneous Dirichlet boundary conditions. Initially, by applying the FFT technique along the x-, y-, and z-axes for the 3-D Poisson equation, the corresponding partial differential equations in spatial domain are converted to algebraic equations in the wavenumber domain, allowing us to easily obtain the 3-D wavenumber-domain gravitational potential. Furthermore, the precise computation of the three-component gravity field responses and the five-component gravity gradient responses in the spatial domain, leveraging the differential attributes of the Fourier transform for spatial derivatives, which significantly enhances the model's accuracy. Despite requiring an additional inverse FFT during the reduction of the research problem's dimensionality, the proposed method can avoid the challenges associated with large-scale matrix inversion. Finally, to validate the computational accuracy and performance of the proposed algorithm, we utilize two density models and a real-world terrain model. Additionally, in comparison to the hybrid wavenumber domain finite-element method in terms of computational efficiency, the FFT-based spectral method not only maintains high accuracy, but also reduces computational time, clearly demonstrating its superiority over conventional numerical methods.

Relevant, Hidden, and Frustrated Information in High-Dimensional Analyses of Complex Dynamical Systems with Internal Noise.

Extracting from trajectory data meaningful information to understand complex molecular systems might be nontrivial. High-dimensional analyses are typically assumed to be desirable, if not required, to prevent losing important information. But to what extent such high-dimensionality is really needed/beneficial often remains unclear. Here we challenge such a fundamental general problem. As a representative case of a system with internal dynamical complexity, we study atomistic molecular dynamics trajectories of liquid water and ice coexisting in dynamical equilibrium at the solid/liquid transition temperature. To attain an intrinsically high-dimensional analysis, we use as an example an abstract high-dimensional descriptor of local molecular environments (e.g., Smooth Overlap of Atomic Positions, SOAP), obtaining a large dataset containing 2.56 × 106 576-dimensional SOAP spectra that we analyze in various ways. Our results demonstrate how the time-series data contained in one single SOAP dimension accounting only <0.001% of the total dataset's variance (neglected and discarded in typical variance-based dimensionality reduction approaches) allows resolving a remarkable amount of information, classifying/discriminating the bulk of water and ice phases, as well as two solid-interface and liquid-interface layers as four statistically distinct dynamical molecular environments. Adding more dimensions to this one is found not only ineffective but even detrimental to the analysis due to recurrent negligible-information/non-negligible-noise additions and "frustrated information" phenomena leading to information loss. Such effects are proven general and are observed also in completely different systems and descriptors' combinations. This shows how high-dimensional analyses are not necessarily better than low-dimensional ones to elucidate the internal complexity of physical/chemical systems, especially when these are characterized by non-negligible internal noise.

Principal component Gaussian optimization for enhancing metaheuristic algorithms in high-dimensional problems

This paper introduces Principal Component Gaussian Optimization (PCGO), a novel hybrid framework that combines Principal Component Analysis (PCA) for dimensionality reduction with an adaptive Gaussian distribution mechanism to enhance search performance. The key innovation of PCGO lies in the dynamic interaction between these two components: PCA-based dimensionality screening identifies dominant search directions by analyzing population distribution, allowing particles to project onto principal components when fitness improvements stall, thereby reducing the search space while preserving meaningful exploration. Simultaneously, the Gaussian mechanism dynamically adjusts mean and variance parameters to optimize particle distribution and guide them along more promising search paths. This synergistic approach accelerates convergence and improves solution accuracy. We integrate PCGO with Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO), creating enhanced variants (PCGO-PSO and PCGO-GWO). The framework is rigorously evaluated on 13 standard benchmark functions (unimodal and multimodal) across 50, 100, and 500 dimensional spaces, as well as real-world engineering design problems. PCGO-enhanced PSO and GWO algorithms significantly improve performance on high-dimensional benchmarks and engineering problems, achieving up to 28.4% higher precision, 19.7% faster convergence, and 34.6% lower variability compared to their base versions. They also outperform state-of-the-art metaheuristics by 12–41% in accuracy, with statistically significant results, demonstrating PCGO's effectiveness in complex optimization tasks.

A unified generalization of the inverse regression methods via column selection

Abstract A bottleneck of sufficient dimension reduction (SDR) in the modern era is that, among numerous methods, only sliced inverse regression (SIR) is generally applicable in high-dimensional settings. The higher-order inverse regression methods, which form a major family of SDR methods superior to SIR at the population level, suffer from the dimensionality of their intermediate matrix-valued parameters which have excessive columns. In this paper, we propose to use a small subset of columns of the matrix-valued parameter for SDR estimation, which breaks the convention of using the ambient matrix in the higher-order inverse regression methods. With a quick column selection procedure, we then generalize these methods and their ensembles in high-dimensional sparse settings, in a uniform manner that resembles sparse SIR without additional assumptions. This is the first promising attempt in the literature to free the higher-order inverse regression methods from their dimensionality, thereby facilitating the application of SDR. Some numerical illustrations, including both simulation studies and a real data example, are provided at the end.

On the breakdown of dimensional reduction and supersymmetry in random-field models

We discuss the breakdown of the Parisi-Sourlas supersymmetry (SUSY) and of the dimensional-reduction (DR) property in the random field Ising and O(NN) models as a function of space dimension dd and/or number of components NN. The functional renormalization group (FRG) predicts that this takes place below a critical line d_\mathrm{DR}(N)dDR(N). We revisit the perturbative FRG results for the RFO(NN)M in d=4+\epsilond=4+ϵ and carry out a more comprehensive investigation of the nonperturbative FRG approximation for the RFIM. In light of this FRG description, we discuss the perturbative results in \epsilon=6-dϵ=6−d recently derived for the RFIM by Kaviraj, Rychkov, and Trevisani. We stress in particular that the disappearance of the SUSY/DR fixed point below d_\mathrm{DR}dDR arises as a consequence of the nonlinearity of the FRG equations and cannot be found via the perturbative expansion in \epsilon=6-dϵ=6−d (nor in 1/N1/N). We also provide an error bar on the value of the critical dimension d_\mathrm{DR}dDR for the RFIM, which we find around 5.11±0.095.11±0.09, by studying several successive orders of the nonperturbative FRG approximation scheme.

Manifold embeddings achieve comparable performance with multispectral imagery for time-series based land disturbance detection

ABSTRACT Dense long-term time series multispectral imagery is crucial for monitoring Earth's surface and detecting disturbances in near-real-time. However, the massive storage requirements of such data pose significant challenges. Dimensionality reduction techniques have been widely applied in remote sensing to address the curse of dimensionality, yet achieving lossless recovery of multispectral data across global locations and varying surface conditions remains difficult. Additionally, it is unclear whether reduced features retain temporal continuity and can integrate effectively with time series algorithms for disturbance detection. This study leverages Uniform Manifold Approximation and Projection (UMAP) for multispectral dimensionality reduction, trained on Harmonized Landsat Sentinel-2 (HLS) imagery. The resulting manifold embeddings are applied to the Continuous Change Detection and Classification (CCDC) algorithm for land disturbance detection. Two key findings emerge: (1) We developed a general UMAP-based dimensionality reduction model that works across global seasons, with manifold embeddings preserving time series coherence and exhibiting stable value ranges. (2) The embeddings achieved comparable performance to full-spectrum multispectral data in image prediction and disturbance detection with CCDC. This research highlights the potential of manifold learning to efficiently store and process dense satellite imagery while maintaining the ability to detect diverse land disturbances.

Comparative analysis of the effects of different dimensionality reduction algorithms on hyperspectral estimation of total nitrogen content in wheat soils

Comparative analysis of the effects of different dimensionality reduction algorithms on hyperspectral estimation of total nitrogen content in wheat soils

Active Gaze Labeling: Visualization for Trust Building.

Areas of interest (AOIs) are well-established means of providing semantic information for visualizing, analyzing, and classifying gaze data. However, the usual manual annotation of AOIs is time-consuming and further impaired by ambiguities in label assignments. To address these issues, we present an interactive labeling approach that combines visualization, machine learning, and user-centered explainable annotation. Our system provides uncertainty-aware visualization to build trust in classification with an increasing number of annotated examples. It combines specifically designed EyeFlower glyphs, dimensionality reduction, and selection and exploration techniques in an integrated workflow. The approach is versatile and hardware-agnostic, supporting video stimuli from stationary and unconstrained mobile eye tracking alike. We conducted an expert review to assess labeling strategies and trust building.

Selection and Placement of Sensors for Electric Motors: A Review and Preliminary Investigation

This review explores sensor selection and placement strategies for electric motor monitoring in industrial settings. A wide range of sensor types including temperature, vibration, current, and position sensors—are evaluated in terms of their technical features and application constraints. Preliminary experimental data on vibration sensors highlight how signal amplitude varies with sensor placement, reinforcing the importance of correct positioning. However, this study stops short of applying AI/ML techniques to optimize placement. Accordingly, this paper serves as a foundational step toward developing intelligent sensor deployment frameworks. Future work will build on this review by integrating supervised learning, dimensionality reduction, and reinforcement learning techniques to automate sensor placement and improve condition monitoring in electric motors.

Data-driven reduced-order modeling of hydrogen-fueled supersonic combustion

Efficient modeling and simulation of supersonic combustion processes are crucial in aerospace applications, requiring rapid prediction of complex multi-physics interactions in irregular computational domains. In this paper, we present a novel residual variational autoencoder-transformer (ResVAE-Trans) model, which is a data-driven method for dimensionality reduction and prediction of multi-physics fields in hydrogen-fueled supersonic combustion. The ResVAE projects high-dimensional dynamic systems onto a low-dimensional latent space, while the transformer constructs a reduced-order model within this space. Before applying the ResVAE-Trans model for dimensionality reduction and prediction, the proposed framework maps multi-physics data from irregular domains onto a structured grid and normalizes it. The framework is demonstrated through hydrogen-fueled supersonic combustion simulations of scramjet engines at the German Aerospace Center (DLR). This approach offers a solution for reduced-order modeling of multi-physics fields in irregular computational domains. Results show that the method successfully achieves dimensionality reduction and prediction of multi-physics fields. It enhances computational efficiency while maintaining prediction accuracy.

Statistical Shape Analysis of Human Bodies

ABSTRACTMorphological analysis of the human body is crucial for various applications in ergonomics and product design, with significant economic and commercial implications. This paper presents a novel exploration of statistical methods for the analysis of human body shapes based on 3D landmark data. It focuses on the application of well‐established statistical methods within the framework of Kendall's 3D shape space. It is well known that Kendall's 3D shape space has non‐constant curvature at all points. A key contribution of this study is the detailed examination of the curvature of Kendall's space at points corresponding to human body shapes, highlighting its implications for posterior statistical analysis. A study of the distances in the dataset is also carried out. From this point, we also compare the performance of intrinsic (Riemannian) methods and Euclidean approximations for mean shape estimation, group difference identification, and dimensionality reduction, providing a comprehensive assessment of their respective strengths and limitations in these contexts. These findings aim to improve the statistical understanding of body morphology and provide valuable guidance for applications in fields such as product design and ergonomics.

Efficient forward and inverse uncertainty quantification for dynamical systems based on dimension reduction and Kriging surrogate modeling in functional space

Efficient forward and inverse uncertainty quantification for dynamical systems based on dimension reduction and Kriging surrogate modeling in functional space

Low-Rank Modeling for Functional Linear Regression with Two-Way Functional Predictor

In this paper, we study a flexible functional linear regression model where the dependency of a scalar response on a functional predictor is function-valued process rather than conventional one-way processes. Additionally, we provide an intuitively appealing estimation approach to estimate the bivariate functional regression coefficient. We first represent the bivariate functional coefficient by using the data-driven bases function to achieve the dimension reduction, and then introduce an iterative least square procedure to estimate the coefficients after dimension reduction in the framework of low-rank structure. Theoretically, we investigate the convergence rate of bivariate functional coefficient estimator under mild conditions. Simulation studies indicate that the proposed methods perform well in finite samples and an empirical example is presented to illustrate its usefulness.

Machine-learning certification of multipartite entanglement for noisy quantum hardware

Abstract Entanglement is a fundamental aspect of quantum physics, both conceptually and for its many applications. Classifying an arbitrary multipartite state as entangled or separable---a task referred to as the separability problem---poses a significant challenge, since a state can be entangled with respect to many different of its partitions. We develop a certification pipeline that feeds the statistics of random local measurements into a non-linear dimensionality reduction algorithm, to determine with respect to which partitions a given quantum state is entangled. After training a model on randomly generated quantum states, entangled in different partitions and of varying purity, we verify the accuracy of its predictions on simulated test data, and finally apply it to states prepared on IBM quantum computing hardware.

StDGAC: A novel identifying spatial domains method via graph attention contrastive network for spatial transcriptomics data.

stDGAC: A novel identifying spatial domains method via graph attention contrastive network for spatial transcriptomics data.