Feature Dimensionality Reduction Research Articles

Intelligent Lung Cancer Detection: A Genetic Algorithm Machine Learning Fusion

Early detection of lung cancer is crucial for improving patient outcomes. While Deep Learning (DL) models have shown high accuracy in lung cancer diagnosis, they often require substantial computational resources. This study proposes a novel approach that leverages Genetic Algorithm (GA) to optimize feature selection and dimensionality reduction from lung cancer images. By integrating GA with conventional Machine Learning (ML) models, we demonstrate improved classification accuracy while minimizing computational requirements. Our experimental results show that combining GA with a feed-forward neural network classifier yields exceptional performance, achieving a classification accuracy of 99.70%. This approach offers a promising alternative to DL models for lung cancer detection, particularly in resource-constrained settings.

Modified Kernel Global-Local Marginal Fisher Analysis for Rolling Bearing Feature Extraction

Abstract The vibration signals of rolling bearings usually exhibit high-dimensional, nonlinear, and non-Gaussian distribution characteristics due to long-term operation under complex working conditions. Therefore, we proposed a novel algorithm named Modified Kernel Global-Local Marginal Fisher Analysis (MKGLMFA) for bearing feature extraction and dimensionality reduction. The proposed MKGLMFA algorithm introduces the kernel function to map data into a high-dimensional space to represent data nonlinearly first. It enhances the within-class compactness and between-class dispersibility by considering spatial relationships and label information when constructing adjacency graphs and simultaneously exploits the local and global geometry of data. Furthermore, a bearing fault diagnosis approach is presented based on MKGLMFA. It first processes the original vibration signals through MKGLMFA to obtain low-dimensional manifold features. Then these characteristics were input into the K-nearest neighbor (KNN) classifier to achieve fault pattern recognition. The superiority of the proposed MKGLMFA algorithm in feature extraction is verified in comparison with some existing state-of-the-art machine learning methods on three rolling bearings datasets. And the subsequent classification diagnosis experiments indicate the effectiveness and high efficiency of the newly raised MKGLMFA algorithm. In comparison with the representative diagnosis methods, the proposed method can extract more sensitive discriminant features, and the classification accuracy of diagnosis is significantly improved in consequence.

MTGWNN: A Multi‐Template Graph Wavelet Neural Network Identification Model for Autism Spectrum Disorder

ABSTRACTFunctional magnetic resonance imaging (fMRI) has been widely applied in studying various brain disorders. However, current studies typically model regions of interest (ROIs) in brains with a single template. This approach generally examines only the connectivity between ROIs to identify autism spectrum disorder (ASD), ignoring the structural features of the brain. This study proposes a multi‐template graph wavelet neural network (GWNN) identification model for ASD called MTGWNN. First, the brain is segmented with multiple templates and the BOLD time series are extracted from fMRI data to construct brain networks. Next, a graph attention network (GAT) is applied to automatically learn interactions between nodes, capturing local information in the node features. These features are then further processed by a convolutional neural network (CNN) to learn global connectivity representations and achieve feature dimensionality reduction. Finally, the features and phenotypic data from each subject are integrated by GWNN to identify ASD at the optimal scale. Experimental results indicate that MTGWNN outperforms the comparative models. Testing on the public dataset ABIDE‐I achieved an accuracy (ACC) of 87.25% and an area under the curve (AUC) of 92.49%. MTGWNN effectively integrates brain network features from multiple templates, providing a more comprehensive characterization of brain abnormalities in patients with ASD. It incorporates population information from phenotypic data, which helps to compensate for the limited sample size of individual patients and improves the robustness and generalization of ASD identification.

A Fault Diagnosis Method for Turnout Switch Machines Based on Sound Signals

The turnout switch machine, a vital outdoor component of railway signaling, controls train steering amidst complex operations and high frequencies. Its malfunction significantly disrupts train operations, potentially causing derailments. This paper proposes a sound-based fault diagnosis method, termed ERS (a method combining EMD, ReliefF, and SVM), for effective monitoring and detection of turnout switch machines. The method employs Eigenmode Decomposition (EMD) to smooth the sound signal, reduce noise, and extract key statistical parameters of both the time and frequency domains. To address redundant information in high-dimensional features, the ReliefF algorithm is utilized for feature selection, dimension reduction, and fault classification based on weighted parameters. Subsequently, the selected feature parameters are used to train the Support Vector Machine (SVM). A comparison with results obtained without ReliefF feature selection demonstrates the necessity of this step. The results show that the fault diagnosis accuracy reaches 98% in the positioning work mode and 95.67% in the reversing work mode, verifying the method’s effectiveness and feasibility.

Prediction of surgical outcomes in ovarian cancer based on CT imaging

This study aims to investigate the commonly used prediction models for the outcomes of ovarian cancer surgery. The study will select 71 cases from the TCGA-OV dataset and divide them into a training set (n=51) and a validation set (n=20) at a ratio of 7:3. The project is based on PyRadiomics Python3.0.1 and extracts 107 basic image features from the image group data, including first-order statistics, shape, gray-level co-occurrence matrix, gray-level size zone matrix, GLRLM, etc. On this basis, the LASSO algorithm is used for feature dimensionality reduction. In building the model, a fuzzy matrix is used to analyze multiple evaluation indicators, including the area under the receiver operating characteristic (ROC) curve (AUC). This study hopes to provide more precise diagnostic and treatment plans for patients, with the goal of achieving better treatment outcomes.

Feature selection and response prediction on a suspension bridge due to wind effect by machine learning

This research introduces an efficient machine learning approach for predicting the structural health of long-span suspension bridges, focusing on their responses to dynamic environmental load like earthquakes and wind. Unlike traditional analytical methods that rely on simplifications, this study utilizes real-world data from anemometers and accelerometers to enhance prediction accuracy and efficiency. Also instead of using whole signals, the study uses fewer time frame signal. The main contribution of this paper is the application of support vector regression (SVR), optimized through Observer-Teacher-Learner-Based-Optimization (OTLBO) for parameter tuning. The optimization process is further refined using Multi-Objective Observer-Teacher-Learner-Based-Optimization (MOOTLBO), targeting feature selection and dimension reduction to improve model performance significantly. Hardanger Bridge serves as a case study, demonstrating the model's capability to accurately predict acceleration responses to wind. By inputting wind sensor data and analyzing acceleration outputs, the enhanced SVR model, through OTLBO and MOOTLBO, shows remarkable predictive accuracy, validated against six performance indices. This methodological advancement suggests that the machine learning-based model could potentially supplant traditional finite element models, offering a more adaptable, efficient, and accurate tool for structural health monitoring (SHM). This advancement addresses critical discrepancies in SHM systems, such as data gaps or sensor malfunctions, by providing a reliable prediction method that ensures the ongoing safety and integrity of bridge structures.

An adaptive learning method for long-term gesture recognition based on surface electromyography.

The surface electromyography (EMG) signal reflects the user's intended actions and has become the important signal source for human-computer interaction. However, classification models trained on EMG signals from the same day cannot be applied for different days due to the time-varying characteristics of the EMG signal and the influence of electrodes shift caused by device wearing for different days, which hinders the application of commercial prosthetics. This type of gesture recogni-tion for different days is usually referred to as long-term gesture recognition.
Approach. To address this issue, we propose a long-term gesture recognition method by optimizing feature extraction, dimensionality reduction, and classification model calibration in EMG signal recognition. Our method extracts differential common spa-tial patterns (CSP) features and then conduct dimensionality reduction with non-negative matrix factorization (NMF), effectively reducing the influence of the non-stationarity of the EMG signals. Based on clustering and classification self-training(CCST) scheme, we select samples with high confidence from unlabeled sam-ples to adaptively updates the model before daily formal use. 
Main results. We verify the feasibility of our method on a dataset consisting of 30 days of gesture data. The proposed gesture recognition scheme achieves accuracy over 90%, similar to the performance of daily calibration with labeled data. However, our method needs only one repetition of unlabeled gestures samples to update the classifi-cation model before daily formal use. 
Significance. From the results we can conclude that the proposed method can not only ensure superior performance, but also greatly facilitate the daily use, which is especially suitable for long-term application.&#xD.

BalancerGNN: Balancer Graph Neural Networks for imbalanced datasets: A case study on fraud detection

Fraud detection for imbalanced datasets is challenging due to machine learning models inclination to learn the majority class. Imbalance in fraud detection datasets affects how graphs are built, an important step in many Graph Neural Networks (GNNs). In this paper, we introduce our BalancerGNN framework to tackle with imbalanced datasets and show its effectiveness on fraud detection. Our framework has three major components: (i) node construction with feature representations, (ii) graph construction using balanced neighbor sampling, and (iii) GNN training using balanced training batches leveraging a custom loss function with multiple components. For node construction, we have introduced (i) Graph-based Variable Clustering (GVC) to optimize feature selection and remove redundancies by analyzing multi-collinearity and (ii) Encoder-Decoder based Dimensionality Reduction (EDDR) using transformer-based techniques to reduce feature dimensions while keeping important information intact about textual embeddings. Our experiments on Medicare, Equifax, IEEE, and auto insurance fraud datasets highlight the importance of node construction with features representations. BalancerGNN trained with balanced batches consistently outperforms other methods, showing strong abilities in identifying fraud cases, with sensitivity rates ranging from 72.87% to 81.23% across datasets while balancing specificity. Additionally, BalancerGNN achieves impressive accuracy rates, ranging from 73.99% to 94.28%. These outcomes underscore the crucial role of graph representation and neighbor sampling techniques in optimizing BalancerGNN for fraud detection models in real-world applications.

Differentiation between invasive ductal carcinoma and ductal carcinoma in situ by combining intratumoral and peritumoral ultrasound radiomics

BackgroundThis study aimed to develop and validate an ultrasound radiomics model for distinguishing invasive ductal carcinoma (IDC) from ductal carcinoma in situ (DCIS) by combining intratumoral and peritumoral features.MethodsRetrospective analysis was performed on 454 patients from Chengzhong Hospital. The patients were randomly divided in accordance with a ratio of 8:2 into a training group (363 cases) and validation group (91 cases). In addition, 175 patients from Yanghu Hospital were used as the external test group. The peritumoral ranges were set to 2, 4, 6, 8, and 10 mm. Mann–Whitney U-test, recursive feature elimination, and a least absolute shrinkage and selection operator were used to in the dimension reduction of the radiomics features and clinical knowledge, and machine learning logistic regression classifiers were utilized to construct the diagnostic model. The area under the curve (AUC) of the receiver operating characteristics, accuracy, sensitivity, and specificity were used to evaluate the model performance.ResultsBy combining peritumoral features of different ranges, the AUC of the radiomics model was improved in the validation and test groups. In the validation group, the maximum increase in AUC was 9.7% (P = 0.031, AUC = 0.803) when the peritumoral range was 8 mm. Similarly, when the peritumoral range was only 8 mm in the test group, the maximum increase in AUC was 4.9% (P = 0.005, AUC = 0.770). In this study, the best prediction performance was achieved when the peritumoral range was only 8 mm.ConclusionsThe ultrasound-based radiomics model that combined intratumoral and peritumoral features exhibits good ability to distinguish between IDC and DCIS. The selection of peritumoral range size exerts an important effect on the prediction performance of the radiomics model.

Large data density peak clustering based on sparse auto-encoder and data space meshing via evidence probability distribution

The development of big data analysis technology has brought new development opportunities to the production and management of various industries. Through the mining and analysis of various data in the operation process of enterprises by big data technology, the internal associated data of the enterprises and even the entire industry can be obtained. As a common method for large-scale data statistical analysis, clustering technology can effectively mine the relationship within massive heterogeneous multidimensional data, complete unlabeled data classification, and provide data support for various model analysis of big data. Common big data density clustering methods are time-consuming and easy to cause errors in data density allocation, which affects the accuracy of data clustering. Therefore we propose a novel large data density peak clustering based on sparse auto-encoder and data space meshing via evidence probability distribution. Firstly, the sparse auto-encoder in deep learning is used to achieve feature extraction and dimensionality reduction for input high-dimensional data matrix through training. Secondly, the data space is meshed to reduce the calculation of the distance between the sample data points. When calculating the local density, not only the density value of the grid itself, but also the density value of the nearest neighbors are considered, which reduces the influence of the subjective selection truncation distance on the clustering results and improves the clustering accuracy. The grid density threshold is set to ensure the stability of the clustering results. Using the K-nearest neighbor information of the sample points, the transfer probability distribution strategy and evidence probability distribution strategy are proposed to optimize the distribution of the remaining sample points, so as to avoid the joint error of distribution. The experimental results show that the proposed algorithm has higher clustering accuracy and better clustering performance than other advanced clustering algorithms on artificial and real data sets.

A pixel-level assessment method of the aging status of silicone rubber insulators based on hyperspectral imaging technology and IPCA-SVM model

Acidic environments are a significant factor in the aging and failure of silicone rubber insulators. Addressing the effective assessment of insulators’ aging state to prevent power transmission accidents has been a critical and urgent issue for the power grid. Therefore, hyperspectral imaging (HSI) technology was employed in this paper, capturing spectral line data of silicone rubber in six aging states in both visible and near-infrared regions, respectively. To reduce data redundancy, genetic algorithm (GA) and band weighting were introduced to improve traditional principal component analysis (PCA), with performance compared using overall accuracy (OA) and Kappa, against 12 other feature extraction or dimensionality reduction methods. The improved principal component analysis − support vector machine (IPCA-SVM) model proposed effectively minimizes irrelevant information in hyperspectral original data, exceeding 93% accuracy and improving OA by 8.26% compared to all bands data. Finally, the IPCA-SVM model was used for pixel-level assessment of the surface aging state of silicone rubber insulators, demonstrating its reliability. This method effectively characterizes the aging state of composite insulators, providing a solid foundation for the safe and stable operation of power grids.

Joint Sparse Local Linear Discriminant Analysis for Feature Dimensionality Reduction of Hyperspectral Images

Although linear discriminant analysis (LDA)-based subspace learning has been widely applied to hyperspectral image (HSI) classification, the existing LDA-based subspace learning methods exhibit several limitations: (1) They are often sensitive to noise and demonstrate weak robustness; (2) these methods ignore the local information inherent in data; and (3) the number of extracted features is restricted by the number of classes. To address these drawbacks, this paper proposes a novel joint sparse local linear discriminant analysis (JSLLDA) method by integrating embedding regression and locality-preserving regularization into the LDA model for feature dimensionality reduction of HSIs. In JSLLDA, a row-sparse projection matrix can be learned, to uncover the joint sparse structure information of data by imposing a L2,1-norm constraint. The L2,1-norm is also employed to measure the embedding regression reconstruction error, thereby mitigating the effects of noise and occlusions. A locality preservation term is incorporated to fully leverage the local geometric structural information of the data, enhancing the discriminability of the learned projection. Furthermore, an orthogonal matrix is introduced to alleviate the limitation on the number of acquired features. Finally, extensive experiments conducted on three hyperspectral image (HSI) datasets demonstrated that the performance of JSLLDA surpassed that of some related state-of-the-art dimensionality reduction methods.

Enhanced Tomato Disease Classification Using Hybrid Deep Learning Approach

Tomato disease classification is important for agricultural productivity and food security, yet achieving high accuracy stays challenging due to the complexity of disease symptoms. This study proposed a model classification utilizing DenseNet121 combined with an Autoencoder for feature extraction and Simple neural network (SNN) for classifier. The dataset contains 10 classes of tomato diseases. Experimental results shows that the proposed approach achieved a high accuracy of 99%, outperforming the performance of several approaches. The integration of the Autoencoder enhances feature extraction and dimensionality reduction, important for improved classification outcomes. This study highlights the efficacy of the proposed work in achieving state-of-the-art accuracy in tomato disease detection. Future work may focus on refining the model using advanced augmentation techniques and expanding the dataset to enhance its applicability across diverse agricultural conditions.

An AI-assisted microfluidic paper-based multiplexed surface-enhanced raman scattering (SERS) biosensor with electrophoretic removal and electrical modulation for accurate acute myocardial infarction (AMI) diagnosis and prognosis

SERS detects single molecules with exceptional sensitivity. To counter the issue of selectivity faced by point-of-care, herein, an externally applied electric field that allows electrical modulation and electromigrates unbound SERS tags without multiple washing steps is successfully developed and demonstrated to improve the biosensor's selectivity and sensitivity in multiplexed detection of cTnI, HDL, and LDL in human serum at a low LoD. Ultra-sensitive detectors can detect signals from non-specifically absorbed species, and these species can cover up overlapping analyte peaks, amplifying the effect of non-specific binding. Even though antifouling molecules can prevent non-specific adsorption at the sensor interface, this approach does not completely eliminate it. Our significant findings show that an electrically regulated device can electromigrate non-specifically bound species without cross-reacting with endogenous albumin proteins. Stability, repeatability, and reproducibility were good, with an RSD of 10%. Artificial intelligence was employed to interpret and analyze high-dimensional fingerprint SERS spectra using feature selection and dimensionality reduction for accurate acute myocardial infarction diagnosis and prognosis. These machine learning methods allow quantification of cTnI, HDL, and LDL biomarkers with low RMSE. Machine learning classifiers showed strong AUROC values of 0.950 ± 0.111 and 0.884 ± 0.139 for early and recurrent AMI detection, respectively. A high negative predictive value (NPV) of ≥99% indicates an effective early AMI rule-out. In short, this work demonstrated that a simple, low-cost, electrophoretic modulated biosensor with machine learning can diagnose, rule out, and predict recurring AMI.

Template Attack Based on the Weighted Quadruple Residual Network

ABSTRACTTemplate attack, as a classic side‐channel attack method, has attracted extensive attention due to its ease of deployment and small number of hyperparameters. However, traditional template attacks (TA) suffer from high preprocessing complexity and poor performance in certain masked applications. To address these issues, this paper proposes a novel template attack method leveraging a weighted quadruple residual network. Specifically, we improve the quadruple network model by assigning different weights to the negative samples selected by the model, which balances the influence of the samples on the model's feedback. Aiming at the fact that the quadruple network ignores the label characteristics of samples in the side‐channel, the distance metric in the quadruple network model is improved, which can help the model select more suitable samples. The residual network designed in this paper enables the quadruplet network model to reduce feature dimensionality, thereby facilitating more effective analysis attacks in TA. Experimental results demonstrate a notable reduction in leaked traces required by the optimized template attack on the synchronous ASCAD dataset, achieving a 62.5% decrease compared to an advanced template attack scheme. On the AES_HD dataset, the energy trace needed for a successful attack diminishes by 46.5% in comparison to the advanced template attack scheme.

Fault diagnosis of wind turbine gears based on OCSSA-VMD and WOA-CNN-BiLSTM

Abstract The accuracy of wind turbine gearbox fault diagnosis will be compromised if the fault feature data is not adequately extracted during operation. To enhance fault identification efficiency and mitigate human interference in parameter setting, this paper introduces an optimized mode decomposition algorithm OCSSA-VMD, derived from variational mode decomposition (VMD) and further optimized by osprey-Cauchy-sparrow search algorithm (OCSSA). This algorithm offers two key advantages: (1) automatic optimization of parameters such as the number of modes k and penalty factor α; (2) reduction of feature dimensionality through mean impact value (MIV) algorithm based on minimum envelope entropy principle, resulting in a multi-fault feature vector set from 13 time-domain features in the intrinsic mode function (IMF) optimal component of wind turbine gearbox vibration data. Additionally, a fault diagnosis model WOA-CNN-BiLSTM is proposed based on whale optimization algorithm (WOA) and convolutional neural network-bidirectional long-short-term-memory (CNN-BiLSTM), which demonstrates improved fault classification accuracy to 98.3333% and diagnosis accuracy to 98.3853% under conditions of insufficient data when compared with other models.

FinSafeNet: securing digital transactions using optimized deep learning and multi-kernel PCA(MKPCA) with Nyström approximation.

With the swift advancement of technology and growing popularity of internet in business and communication, cybersecurity posed a global threat. This research focuses new Deep Learning (DL) model referred as FinSafeNet to secure loose cash transactions over the digital banking channels. FinSafeNet is based on a Bi-Directional Long Short-Term Memory (Bi-LSTM), a Convolutional Neural Network (CNN) and an additional dual attention mechanism to study the transaction data and influence the observation of various security threats. One such aspect is, relying these databases in most of the cases imposes a great technical challenge towards effective real time transaction security. FinSafeNet draws attention to the attack and reproductive phases of Hierarchical Particle Swarm Optimization (HPSO) feature selection technique simulating it in a battle for extreme time performance called the Improved Snow-Lion optimization Algorithm (I-SLOA). Upon that, the model then applies the Multi-Kernel Principal Component Analysis (MKPCA) accompanied by Nyström Approximation for handling the MKPCA features. MKPCA seeks to analyze and understand a non-linear structure of data whereas, Nyström Approximation reduces the burden on computational power hence allows the model to work in situations where large sizes of datasets are available but with no loss of efficiency of the model. This causes FinSafeNet to work easily and still be able to make accurate forecasts. Besides tackling feature selection and dimensionality reduction, the model presents advanced correlation measures as well as Joint Mutual Information Maximization to enhance variable correlation analysis. These improvements further help the model to detect the relevant features in transaction data that may present a threat to the security of the system. When tested on commonly used database for testing banks performance, for instance the Paysim database, FinSafeNet significantly improves upon the previous and fundamental approaches, achieving accuracy of 97.8%.

Research on Sleep Staging Based on Support Vector Machine and Extreme Gradient Boosting Algorithm.

To develop a sleep-staging algorithm based on support vector machine (SVM) and extreme gradient boosting model (XB Boost) and evaluate its performance. In this study, data features were extracted based on physiological significance, feature dimension reduction was performed through appropriate methods, and XG Boost classifier and SVM were used for classification. One hundred and twenty training sets and 80 test sets were randomly composed of the first 200 groups of data from the SHH1 database. The polysomnography (PSG) data of 20 real individuals in the clinic were selected as the experimental data. The C3 electroencephalogram (EEG), left and right electrooculogram (EOG), electromyogram (EMG), and other signals were analyzed. Finally, the stages were adjusted based on human sleep laws. The standard staging of the database and the doctor's diagnosis staging was used as the standard. The SHHS1 database test results were as follows: the average accuracy was 83.24%, the precision and recall of Stage Wake and Stage 2 NREM sleep (N2) were over 80%, and the precision, F1-Score and recall of Stage 3 NREM sleep (N3) and Rapid Eye Movement (REM) were more than 70%. The clinical data test results were as follows: the average accuracy rate was 76.37%; for Wake and N3, the precision reached 85%; for Wake, N2, and REM, the recall rate reached over 70%; for Wake, the F-1 Score reached over 90%. This study shows that the sleep staging results of the algorithm for the database and clinical data were similar. The staging results meet the requirements at the medical level.

COMPUTATIONALLY EFFICIENT DEEP FEDERATED LEARNING WITH OPTIMIZED FEATURE SELECTION FOR IOT BOTNET DETECTION

Internet of Things (IoT) is a technology that has revolutionized various fields, offering numerous benefits, such as remote patient monitoring, enhanced energy efficiency, and automation of routine tasks in homes. However, unsecured IoT devices are susceptible to botnet-based attacks such as distributed denial of Service (DDoS). Conventional machine learning models used for detecting these attacks compromise data privacy, prompting the adoption of federated learning (FL) to improve privacy. Yet, most FL-based cyberattack detection models proposed for IoT environments do not address computational complexity to suit their deployment on resource-constrained IoT edge devices. This paper introduces an FL model with low computational complexity, designed for detecting IoT botnet attacks. The study employs feature selection and dimensionality reduction to minimize computational complexity while maintaining high accuracy. First, an extreme gradient boosting model, trained with repeated stratified k-fold cross-validation, is used to select the optimal features of the botnet dataset based on feature importance. Principal component analysis is then used to reduce the dimensionality of these features. Finally, a differentially private multi-layer perceptron is trained locally by four FL clients and aggregated through federated averaging (FedAvg) to form a global Mirai botnet attack detection model. The model achieved an accuracy, precision, recall, and F1-score of 99.93%, an area under the curve of 1.0, and 8,612 floating-point operations, contributing to 87.34% reduction in computational complexity compared to the previous work. The proposed model is well-suited for detecting botnet attacks in smart homes, smart grids, and environments where resource-constrained IoT edge devices are deployed.

Predicting the productivity of fractured horizontal wells using few-shot learning

Predicting the productivity of multistage fractured horizontal wells plays an important role in exploiting unconventional resources. In recent years, machine learning (ML) models have emerged as a new approach for such studies. However, the scarcity of sufficient real data for model training often leads to imprecise predictions, even though the models trained with real data better characterize geological and engineering features. To tackle this issue, we propose an ML model that can obtain reliable results even with a small amount of data samples. Our model integrates the synthetic minority oversampling technique (SMOTE) to expand the data volume, the support vector machine (SVM) for model training, and the particle swarm optimization (PSO) algorithm for optimizing hyperparameters. To enhance the model performance, we conduct feature fusion and dimensionality reduction. Additionally, we examine the influences of different sample sizes and ML models for training. The proposed model demonstrates higher prediction accuracy and generalization ability, achieving a predicted R2 value of up to 0.9 for the test set, compared to the traditional ML techniques with an R2 of 0.13. This model accurately predicts the production of fractured horizontal wells even with limited samples, supplying an efficient tool for optimizing the production of unconventional resources. Importantly, the model holds the potential applicability to address similar challenges in other fields constrained by scarce data samples.