Reduced Feature Space Research Articles

Joint feature selection and classification of low-resolution satellite images using the SAT-6 dataset

The modern industries of today demand the classification of satellite images, and to use the information obtained from it for their advantage and growth. The extracted information also plays a crucial role in national security and the mapping of geographical locations. The conventional methods often fail to handle the complexities of this process. so, an effective method is required with high accuracy and stability. In this paper, a new methodology named RankEnsembleFS is proposed that addresses the crucial issues of stability and feature aggregation in the context of the SAT-6 dataset. RankEnsembleFS makes use of a two-step process that consists of ranking the features and then selecting the optimal feature subset from the top-ranked features. RankEnsembleFS achieved comparable accuracy results to state-of-the-art models for the SAT-6 dataset while significantly reducing the feature space. This reduction in feature space is important because it reduces computational complexity and enhances the interpretability of the model. Moreover, the proposed method demonstrated good stability in handling changes in data characteristics, which is critical for reliable performance over time and surpasses existing ML ensemble methods in terms of stability, threshold setting, and feature aggregation. In summary, this paper provides compelling evidence that this RankEnsembleFS methodology presents excellent performance and overcomes key issues in feature selection and image classification for the SAT-6 dataset.

Use of machine learning methods in diagnosis of carpal tunnel syndrome

Carpal tunnel syndrome (CTS) is a common condition diagnosed using physical exams and electromyography (EMG) data. This study aimed to classify CTS severity using machine learning techniques. EMG data from 154 patients, including measurements of motor and sensory latency, velocity, and amplitude, were used to form a six-dimensional feature space. Classifiers such as DT, LDA, NB, SVM, k-NN, and ANN were applied, and the feature space was reduced using ANOVA, MRMR, Relieff, and PCA. The DT classifier with ANOVA feature selection showed the best performance for both full and reduced feature spaces.

Exhaustive Exploitation of Nature-Inspired Computation for Cancer Screening in an Ensemble Manner.

Accurate screening of cancer types is crucial for effective cancer detection and precise treatment selection. However, the association between gene expression profiles and tumors is often limited to a small number of biomarker genes. While computational methods using nature-inspired algorithms have shown promise in selecting predictive genes, existing techniques are limited by inefficient search and poor generalization across diverse datasets. This study presents a framework termed Evolutionary Optimized Diverse Ensemble Learning (EODE) to improve ensemble learning for cancer classification from gene expression data. The EODE methodology combines an intelligent grey wolf optimization algorithm for selective feature space reduction, guided random injection modeling for ensemble diversity enhancement, and subset model optimization for synergistic classifier combinations. Extensive experiments were conducted across 35 gene expression benchmark datasets encompassing varied cancer types. Results demonstrated that EODE obtained significantly improved screening accuracy over individual and conventionally aggregated models. The integrated optimization of advanced feature selection, directed specialized modeling, and cooperative classifier ensembles helps address key challenges in current nature-inspired approaches. This provides an effective framework for robust and generalized ensemble learning with gene expression biomarkers.

A novel voting ensemble model empowered by metaheuristic feature selection for accurate flash flood susceptibility mapping

This study addresses the challenges of flash flood susceptibility mapping in Yemen’s Qaa’Jahran Basin, characterized by complex terrain and limited hydro-meteorological data. To enhance predictive accuracy, we integrate metaheuristic feature selection with ensemble learning models. Initially, fifteen flash flood variables were retrieved using Geographic Information System (GIS) based remote sensing, setting the stage for a novel feature selection algorithm. Then, the Memo Search Algorithm (MSA), a metaheuristic approach is proposed to efficiently reduce feature space. Through comprehensive comparisons with established algorithms such as the Artificial Bee Colony (ABC) and Gray Wolf Optimizer (GWO), MSA refined the selection, identifying 'elevation’ and 'distance to streams’ as optimal factors. Statistical validations using the Friedman and Wilcoxon signed-rank tests confirmed the significant superiority of MSA over competing algorithms. Ensemble classifiers (bagging, boosting, stacking) were then applied to the reduced feature space. Comprehensive evaluation revealed the boosting ensemble with MSA outperformed traditional techniques reaching 98.75% accuracy, 0.9896 Area Under the Curve (AUC), and 98.95% the harmonic mean of the precision and recall (F1-score). Precision in identifying high-risk flash flood zones was underlined via spatial prediction, confirming the integrated framework’s ability to significantly improve forecast accuracy. The findings aid disaster management with powerful geographic mapping in data-poor regions. The proposed framework is adaptable globally for flash flood-prone areas with similar constraints. As climate change is expected to increase extreme rainfall events, communities globally will need robust data-driven methodologies for flash flood susceptibility mapping. The Key recommendations of the current study include investigating hybrid feature selection methods to better enhance predictive inputs and analyzing transferability across hydro-climatic zones.

A simple theory for training response of deep neural networks

Deep neural networks give us a powerful method to model the training dataset’s relationship between input and output. We can regard that as a complex adaptive system consisting of many artificial neurons that work as an adaptive memory as a whole. The network’s behavior is training dynamics with a feedback loop from the evaluation of the loss function. We already know the training response can be constant or shows power law-like aging in some ideal situations. However, we still have gaps between those findings and other complex phenomena, like network fragility. To fill the gap, we introduce a very simple network and analyze it. We show the training response consists of some different factors based on training stages, activation functions, or training methods. In addition, we show feature space reduction as an effect of stochastic training dynamics, which can result in network fragility. Finally, we discuss some complex phenomena of deep networks.

Botnet Attack Detection in IoT Devices using Ensemble Classifiers with Reduced Feature Space

The Internet of Things (IoT) is an advancing important technology offers multiple perks, such as webcams, baby monitors, room temperature controllers, smart security cameras and intelligent home automations resulting in the creation of intelligent settings that greatly simplify daily living. However, there are cybersecurity dangers associated with IoT devices due to their lack of protection. For example, Internet of Things botnets have become a major risk. IoT has been a boon for attackers to perform malicious attacks like information theft, DDoS, sending junk data to disrupt networks. IoT devices face serious security issues, from having default weak and common passwords, and a lack of security, rarely and poorly monitored, to having open access to management systems, always connected to the internet. In this paper, we used the N-BaIoT dataset which includes datasets of 9 IoT devices infected with 2 Bot viruses Mirai and Bashlite, where each botnet has 5 sub-attacks and the benign datasets of 9 devices. An analysis with the N-BaIoT dataset which initially had 115 features were reduced to 35 features by using manual reduction and further reduced to single feature in 5-time instances equivalent to 5 features using heat map. We then classified the sub-attacks of 2 botnets and benign of 9 IoT devices by using 7 Machine Learning based classifiers in the Weka tool and Python and compared our results with the manually reduced 35 Features and Heat map based 5 features. Performance metrics like correctly classified, incorrectly classified instances and time taken to build the model were evaluated to verify the proposed work. We found out that over 3 ensemble machine learning classifiers performed extremely well with 99 % accuracies for all devices. In order to verify the logic of our work we tried implementing our proposed model in a different dataset with 3 ensemble classifiers and were able to achieve high detection rates.

Unsteady cylinder wakes from arbitrary bodies with differentiable physics-assisted neural network.

This work describes a hybrid predictive framework configured as a coarse-grained surrogate for reconstructing unsteady fluid flows around multiple cylinders of diverse configurations. The presence of cylinders of arbitrary nature causes abrupt changes in the local flow profile while globally exhibiting a wide spectrum of dynamical wakes fluctuating in either a periodic or chaotic manner. Consequently, the focal point of the present study is to establish predictive frameworks that accurately reconstruct the overall fluid velocity flowfield such that the local boundary layer profile, as well as the wake dynamics, are both preserved for long time horizons. The hybrid framework is realized using a base differentiable flow solver combined with a neural network, yielding a differentiable physics-assisted neural network (DPNN). The framework is trained using bodies with arbitrary shapes, and then it is tested and further assessed on out-of-distribution samples. Our results indicate that the neural network acts as a forcing function to correct the local boundary layer profile while also remarkably improving the dissipative nature of the flowfields. It is found that the DPNN framework clearly outperforms the supervised learning approach while respecting the reduced feature space dynamics. The model predictions for arbitrary bodies indicate that the Strouhal number distribution with respect to spacing ratio exhibits similar patterns with existing literature. In addition, our model predictions also enable us to discover similar wake categories for flow past arbitrary bodies. For the chaotic wakes, the present approach predicts the chaotic switch in gap flows up to the mid-time range.

Deep Learning PCA Correlation Analysis of Teacher’s Educational Technology Quality and Teacher’s Professional Competence

In contemporary education, the integration of educational technology has become crucial for enhancing teaching quality and fostering student learning outcomes. This study employs a novel approach by combining deep learning techniques with Principal Component Analysis (PCA) to explore the relationship between teachers' educational technology proficiency and their professional competence. The research leverages a dataset comprising various metrics related to teachers' usage of educational technology tools and their performance in professional development programs. Initially, the study preprocesses the data and applies deep learning models to extract high-level features from the complex and heterogeneous dataset. Subsequently, PCA is employed to reduce the dimensionality of the feature space while preserving the underlying structure and variability of the data. Through this process, latent factors representing different aspects of educational technology proficiency and professional competence are identified. The correlation analysis conducted on the reduced feature space reveals intricate relationships between teachers' competency in educational technology and their overall professional capabilities. The findings indicate that certain dimensions of educational technology quality, such as adaptability to new tools and innovative pedagogical approaches, positively correlate with measures of professional competence, including classroom management skills, student engagement, and instructional effectiveness.

Inverse design of pore wall chemistry and topology through active learning of surface group interactions.

Design of next-generation membranes requires a nanoscopic understanding of the effect of biologically inspired heterogeneous surface chemistries and topologies (roughness) on local water and solute behavior. In particular, the rejection of small, neutral solutes, such as boric acid, poses a heretofore unsolved challenge. In prior work, a computational inverse design technique using an evolutionary optimization successfully uncovered new surface design strategies for optimized transport of water over solutes in smooth, model pores consisting of two surface chemistries. However, extending such an approach to more complex (and realistic) scenarios involving many surface chemistries as well as surface roughness is challenging due to the expanded design space. In this work, we develop a new approach that uses active learning to optimize in a reduced feature space of surface group interactions, finding parameters that lead to their assembly into ordered, optimal patterns. This approach rapidly identifies novel surface functionalizations that maximize the difference in water and boric acid transport through the nanopore. Moreover, we find that the roughness of the nanopore wall, independent of its chemistry, can be leveraged to enhance transport selectivity: oscillations in the pore wall diameter optimally inhibit boric acid transport by creating energetic wells from which the solute must escape to transport down the pore. This proof-of-concept demonstrates the potential for active learning strategies, in concert with molecular simulations, to rapidly navigate complex design spaces of aqueous interfaces and is promising as a tool for engineering water-mediated surface interactions for a broad range of applications.

Mitigating the curse of dimensionality using feature projection techniques on electroencephalography datasets: an empirical review

Electroencephalography (EEG) is commonly employed to diagnose and monitor brain disorders, however, manual analysis is time-consuming. Hence, researchers nowadays are increasingly leveraging artificial intelligence (AI) techniques for automatic analysis of EEG, involving task-specific feature extraction and classification. Extracting multiple non-linear features from multiple EEG channels enhances the performance of classification, but it also results in high-dimensional features. However, the presence of the "curse of dimensionality" poses challenges for classifiers in AI-based EEG analyzers, leading to overfitting and complexity in classification. Therefore, in this empirical review, the effectiveness of the existing dimensionality reduction techniques to mitigate the curse of dimensionality in EEG feature sets is analyzed. We begin this study, by overviewing the extracted high-dimensional features from EEG signals related to disorders such as schizophrenia, alcoholism, focal seizures, focal seizures with deep features, and depression. Subsequently, 23 reduction techniques were reviewed, which transform the high-dimensional EEG features into a new reduced feature space to improve the classification. The impact of the reduced features was evaluated using traditional AI-based classifiers (support vector machine (SVM) and k-nearest neighbor (k-NN)). 10-fold cross-validation was performed for training and testing, and the performance was evaluated using accuracy, selected features, and dimensionality reduction rate (DRR) metrics. Comprehensive analysis of projection techniques on diverse EEG datasets offers valuable insights to assist researchers in selecting the most suitable technique. The hybrid projection technique termed principal component analysis-based t-distributed stochastic neighbor embedding (PCA+t-SNE) achieved an impressive average accuracy of 93.36%, surpassing the k-NN classifier without reduction techniques.

Providing a Feature Selection Method for Lung Cancer Prediction Using Neural Network

Introduction: The remarkable growth of lung cancer and its associated impacts and consequences, along with the substantial costs it imposes on society, has driven the medical community to pursue programs aimed at further examination, prevention, early detection, and diagnosis. In medicine science, timely discovery and diagnosis of diseases can prevent many life-threatening conditions and save people's lives.Material and Methods: This study aims to predict lung cancer using a novel feature selection method integrated with a classifier. Our approach entails a comprehensive four-stage method. Initially, we calculate feature similarities within a lung cancer dataset using the absolute value of the Pearson correlation coefficient, followed by the clustering of initial features using the community detection algorithm called Louvain. Next, we employ techniques to determine the optimal subset of features using the concept of node centrality. Ultimately, lung cancer diagnosis is executed using the selected features, leveraging a classifier.Results: Comparative analysis reveals that our proposed method outperforms existing techniques in terms of reduced execution time and improved prediction accuracy. When compared with established methods, our approach demonstrates superior outcomes in terms of the number of selected features and classification accuracy. Our method reduced 12600 features to 118 features and its accuracy was 95.28, 95.49, 95.23 and 95.32 for Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB) and K-Nearest Neighbor (KNN) classifier. The comparison of runtime shows that the proposed method is significantly improved with a runtime of 2.146 seconds compared to other methods.Conclusion: The proposed feature selection method successfully reduced the initial feature set and significantly decreased computational time. Moreover, the achieved prediction accuracies underscore the reliability of our approach. This significant reduction in feature space while maintaining consistently high prediction accuracies serves as a strong validation of the potency and practical applicability of our methodology in the domain of lung cancer prediction. These compelling results strongly advocate for the potential real-world impact of our approach.

In defense of local descriptor-based few-shot object detection.

State-of-the-art image object detection computational models require an intensive parameter fine-tuning stage (using deep convolution network, etc). with tens or hundreds of training examples. In contrast, human intelligence can robustly learn a new concept from just a few instances (i.e., few-shot detection). The distinctive perception mechanisms between these two families of systems enlighten us to revisit classical handcraft local descriptors (e.g., SIFT, HOG, etc.) as well as non-parametric visual models, which innately require no learning/training phase. Herein, we claim that the inferior performance of these local descriptors mainly results from a lack of global structure sense. To address this issue, we refine local descriptors with spatial contextual attention of neighbor affinities and then embed the local descriptors into discriminative subspace guided by Kernel-InfoNCE loss. Differing from conventional quantization of local descriptors in high-dimensional feature space or isometric dimension reduction, we actually seek a brain-inspired few-shot feature representation for the object manifold, which combines data-independent primitive representation and semantic context learning and thus helps with generalization. The obtained embeddings as pattern vectors/tensors permit us an accelerated but non-parametric visual similarity computation as the decision rule for final detection. Our approach to few-shot object detection is nearly learning-free, and experiments on remote sensing imageries (approximate 2-D affine space) confirm the efficacy of our model.

LBO-MPAM: Ladybug Beetle Optimization-based multilayer perceptron attention module for segmenting the skin lesion and automatic localization

ABSTRACT In recent years, skin cancer has been the most dangerous disease noticed among people worldwide. Skin cancer should be identified earlier to reduce the rate of mortality. Employing dermoscopic images can identify and categorise skin cancer effectively. But, the visual evaluation is a complex procedure to be done in the dermoscopic image. However, Deep learning (DL) is an efficient method for skin cancer detection; however, segmenting the skin lesion and automatic localisation in an earlier stage is complicated. In this paper, a novel Ladybug Beetle Optimization-Double Attention Based Multilevel 1-D CNN (LBO-DAM 1-D CNN) technique is proposed to detect and classify skin cancer. To improve skin lesion type discriminability, the two types of attention modules are introduced. The Ultra-Lightweight Subspace Attention Module (ULSAM) is utilised for classifying the feature maps into different stages to validate the frequency from different image samples. However, the multilayer perceptron attention module (MLPAM) is determined to provide information regarding skin cancer classification and diminish the noise and unwanted data. To minimise data loss, it is then combined with hierarchical complementarity during classification. Second, a modified MLPAM is used to extract significant feature spaces for network learning, select the most important information, and reduce feature space redundancy. The Ladybug Beetle Optimization (LBO) algorithm provides the optimal classification solution by minimising the loss rate of DAM 1-D CNN architecture. The experimentation is conducted on three different datasets such as ISIC2020, HAM10000, and the melanoma detection dataset. The experimental results revealed that the proposed method is compared with different existing methods such as IMFO-KELM, Mask RCNN, M-SVM, DCNN-9, and TL-CNN with different datasets. These methods attained 94.56, 92.65, 90.56, 88.65, and 95.5 for the ISIC2020 dataset but the proposed method enhanced the classification performance by attaining 97.02. Also, the validation is based on metrics, namely, accuracy, precision, sensitivity, and F1-score of 97.03%, 97.05%, 97.58%, and 97.27% for a total of 500 epochs.

A Data Fusion Approach for Accurate Diagnosis of Parkinson's Disease

Diagnosing Parkinson's Disease (PD) can be quite challenging as it presents with symptoms and lacks biomarkers. Nevertheless, the use of data fusion, which combines types of data using machine learning techniques holds promise, for the timely detection of the disease. In this study, we explore the application of data fusion by employing Principal Component Analysis (PCA) as a step to reduce complexity. We then utilize the K Nearest Neighbors (KNN) classification to improve the accuracy of PD diagnosis. By analyzing nonlinear features associated with PD from a dataset PCA helps us extract attributes while maintaining important variations in the data. Subsequently, KNN is employed to identify patterns in this reduced feature space and effectively distinguish between individuals with PD and those who are healthy. Our results show improvements when using the KNN classifier compared to state-of-the-art approaches. This demonstrates its effectiveness in detecting PD leading to promising outcomes and providing a framework for precise PD diagnosis.

A study on the risk stratification for patients within 24 hours of admission for risk of hospital-acquired urinary tract infection using Bayesian network models.

Early identification of patients at risk of hospital-acquired urinary tract infections (HA-UTI) enables the initiation of timely targeted preventive and therapeutic strategies. Machine learning (ML) models have shown great potential for this purpose. However, existing ML models in infection control have demonstrated poor ability to support explainability, which challenges the interpretation of the result in clinical practice, limiting the adaption of the ML models into a daily clinical routine. In this study, we developed Bayesian Network (BN) models to enable explainable assessment within 24h of admission for risk of HA-UTI. Our dataset contained 138,250 unique hospital admissions. We included data on admission details, demographics, lifestyle factors, comorbidities, vital parameters, laboratory results, and urinary catheter. Models developed from a reduced set of five features were characterized by transparency compared to models developed from a full set of 50 features. The expert-based clinical BN model over the reduced feature space showed the highest performance (area under the curve = 0.746) compared to the naïve- and tree-augmented-naïve BN models. Moreover, models developed from expert-based knowledge were characterized by enhanced explainability.

Machine learning-based clinical decision support for infection risk prediction.

Healthcare-associated infection (HAI) remains a significant risk for hospitalized patients and a challenging burden for the healthcare system. This study presents a clinical decision support tool that can be used in clinical workflows to proactively engage secondary assessments of pre-symptomatic and at-risk infection patients, thereby enabling earlier diagnosis and treatment. This study applies machine learning, specifically ensemble-based boosted decision trees, on large retrospective hospital datasets to develop an infection risk score that predicts infection before obvious symptoms present. We extracted a stratified machine learning dataset of 36,782 healthcare-associated infection patients. The model leveraged vital signs, laboratory measurements and demographics to predict HAI before clinical suspicion, defined as the order of a microbiology test or administration of antibiotics. Our best performing infection risk model achieves a cross-validated AUC of 0.88 at 1 h before clinical suspicion and maintains an AUC >0.85 for 48 h before suspicion by aggregating information across demographics and a set of 163 vital signs and laboratory measurements. A second model trained on a reduced feature space comprising demographics and the 36 most frequently measured vital signs and laboratory measurements can still achieve an AUC of 0.86 at 1 h before clinical suspicion. These results compare favorably against using temperature alone and clinical rules such as the quick sequential organ failure assessment (qSOFA) score. Along with the performance results, we also provide an analysis of model interpretability via feature importance rankings. The predictive model aggregates information from multiple physiological parameters such as vital signs and laboratory measurements to provide a continuous risk score of infection that can be deployed in hospitals to provide advance warning of patient deterioration.

Multimodal fusion hierarchical self-attention network for dynamic hand gesture recognition

Recent improvements in dynamic hand gesture recognition have seen a shift from traditional convolutional architectures to attention-based networks. These attention networks have been proven to outclass CNN + LSTM architectures, showing higher accuracy as well as reduced model parameters. Especially, skeleton-based attention networks have been shown to outperform visual-based networks due to the rich information from skeleton-based hand features. However, there is an opportunity to introduce complementary features from other modalities like RGB, depth, and optical flow images to enhance the recognition capability of skeleton-based networks. This paper aims to explore the addition of a multimodal fusion network to a skeleton-based Hierarchical Self-Attention Network (MF-HAN) and test for increased model effectiveness. Unlike traditional fusion techniques, this fusion network uses features derived from other sources of multimodal data in a reduced feature space using a cross-attention layer. The model outperforms its root model and other state-of-the-art models on the SHREC’17 track dataset, especially in the 28 gestures setting by more than 1 % in gesture classification accuracy. The experimentation was tested on the DHG dataset as well.

An Intelligent Diagnostic System to Analyze Early-Stage Chronic Kidney Disease for Clinical Application

Chronic kidney disease (CKD) is a progressive condition characterized by the gradual deterioration of kidney functions, potentially leading to kidney failure if not promptly diagnosed and treated. Machine learning (ML) algorithms have shown significant promise in disease diagnosis, but in healthcare, clinical data pose challenges: missing values, noisy inputs, and redundant features, affecting early-stage CKD prediction. Thus, this study presents a novel, fully automated machine learning approach to tackle these complexities by incorporating feature selection (FS) and feature space reduction (FSR) techniques, leading to a substantial enhancement of the model’s performance. A data balancing technique is also employed during preprocessing to address data imbalance issue that is commonly encountered in clinical contexts. Finally, for reliable CKD classification, an ensemble characteristics-based classifier is encouraged. The effectiveness of our approach is rigorously validated and assessed on multiple datasets, and the clinical relevancy of the strategy is evaluated on the real-world therapeutic data collected from Bangladeshi patients. The study establishes the dominance of adaptive boosting, logistic regression, and passive aggressive ML classifiers with 96.48% accuracy in forecasting unseen therapeutic CKD data, particularly in early-stage cases. Furthermore, the effectiveness of the FSR technique in reducing the prediction time significantly is revealed. The outstanding performance of the proposed model demonstrates its effectiveness in addressing the complexity of healthcare CKD data by incorporating the FS and FSR techniques. This highlights its potential as a promising computer-aided diagnosis tool for doctors, enabling early interventions and improving patient outcomes.

Data Mining Technique for Breast Cancer Prediction using Fuzzy Theory

In order to find a reliable approach of breast cancer prediction, Data mining methods are used in the studies provided in this article. This study compares multiple patient clinical data in order to find a reliable model that can predict the occurrence of breast cancer. In this article, the support vector machine (SVM), artificial neural network (ANN), naive bayes classifier, and AdaBoost tree are used as four data mining methods. Furthermore, since it has such a significant impact on the efficacy and efficiency of the learning process, feature space is extensively examined in this work. Combining PCA with other data mining algorithms that use a PCA-like technique to compress the feature space is recommended. This hybrid is intended to assess the effect of feature space reduction. Wisconsin Breast Cancer Database (1991) and Wisconsin Diagnostic Breast Cancer (1995) are two frequently used test data sets that are used to assess the effectiveness of these algorithms. To calculate each model's test error, the method of 10-fold cross-validation is used. The findings of this research show a thorough trade-off between these tactics and also provide a thorough assessment of the models. In practical applications, it is anticipated that feature identification results would help to avoid breast cancer for both doctors and patients.

Improving Ultrasonic Testing by Using Machine Learning Framework Based on Model Interpretation Strategy

Ultrasonic testing (UT) is increasingly combined with machine learning (ML) techniques for intelligently identifying damage. Extracting significant features from UT data is essential for efficient defect characterization. Moreover, the hidden physics behind ML is unexplained, reducing the generalization capability and versatility of ML methods in UT. In this paper, a generally applicable ML framework based on the model interpretation strategy is proposed to improve the detection accuracy and computational efficiency of UT. Firstly, multi-domain features are extracted from the UT signals with signal processing techniques to construct an initial feature space. Subsequently, a feature selection method based on model interpretable strategy (FS-MIS) is innovatively developed by integrating Shapley additive explanation (SHAP), filter method, embedded method and wrapper method. The most effective ML model and the optimal feature subset with better correlation to the target defects are determined self-adaptively. The proposed framework is validated by identifying and locating side-drilled holes (SDHs) with 0.5λ central distance and different depths. An ultrasonic array probe is adopted to acquire FMC datasets from several aluminum alloy specimens containing two SDHs by experiments. The optimal feature subset selected by FS-MIS is set as the input of the chosen ML model to train and predict the times of arrival (ToAs) of the scattered waves emitted by adjacent SDHs. The experimental results demonstrate that the relative errors of the predicted ToAs are all below 3.67% with an average error of 0.25%, significantly improving the time resolution of UT signals. On this basis, the predicted ToAs are assigned to the corresponding original signals for decoupling overlapped pulse-echoes and reconstructing high-resolution FMC datasets. The imaging resolution is enhanced to 0.5λ by implementing the total focusing method (TFM). The relative errors of hole depths and central distance are no more than 0.51% and 3.57%, respectively. Finally, the superior performance of the proposed FS-MIS is validated by comparing it with initial feature space and conventional dimensionality reduction techniques.