Small Number Of Features Research Articles

A study of feature importance for king salmon health classification with feature selection

King salmon is important for aquaculture in New Zealand, contributing significant economic value. Fish health is a priority for the industry, and the change in the health status of king salmon needs to be accurately detected at the earliest possible stage. Many factors affect the health of king salmon, such as temperature. Identifying the key features that influence health prediction is a crucial step toward achieving this goal. This study utilizes trial data collected by the Cawthron Institute, which includes diverse information on king salmon, such as blood biochemistry and hematology. We explore the data by employing statistical methods and feature selection techniques in machine learning to identify the most relevant features for king salmon health prediction, aiming to classify individuals as healthy or unhealthy with a small number of features. The results show that although the most efficient feature selection techniques on different datasets vary, overall, feature selection approaches can successfully identify relevant and informative features for king salmon health classification. Through the incorporation of a few selected features, the learned classifiers could still achieve statistically equal or better classification performance. This study not only contributes to the understanding of the health indicators of king salmon but also provides crucial insights into health prediction, which will be beneficial to the improvement of the health of king salmon, leading to the development of more effective management strategies for aquaculture.

Automated non-invasive quantification of myocardial work using 3D echocardiography and validation with invasive catheter measurements

Abstract Background Myocardial work is an emerging technique for the assessment of left ventricular (LV) performance. Using blood pressure as an afterload determinate offers further insight into LV mechanics than can be detected with load-sensitive parameters of ejection fraction or global longitudinal strain (GLS). Observational studies have signalled benefit in identifying vulnerable ventricles in cardiomyopathy, valvular, and coronary disease. While the reference standard for calculating myocardial work is through catheterisation, this is impractical for routine assessment. Currently, a vendor specific non-invasive index of myocardial work is commercially available, but requires subjective manual analysis and identification of valvular events. With increasing availability of 3D echocardiography (3DE) and automated analysis techniques, there is potential to derive more reproducible estimates of myocardial work. Purpose We sought to test whether global myocardial work could be accurately, non-invasively, and automatically estimated using 3DE, in combination with standard non-invasive clinical measurements. Methods Patients undergoing left heart catheterisation as part of routine care were prospectively recruited, and 3DE was performed within one hour of invasive LV pressure measurement. A validated deep learning model was used to automatically analyse the 3DE image volumes to produce GLS curves, from which a smaller number of features (principal components) were extracted. The invasive pressure traces were separated into cycles, which were each paired with the strain trace that had the minimal difference in R-R interval (and excluded if the relative R-R interval difference exceeded 25%). For each patient, an invasively-derived estimate of myocardial work was calculated as the area enclosed by the median pressure-strain (PS) loop. A multivariate regression model was used to relate a set of routine clinical parameters to the invasively-derived myocardial work, and subsequently assessed using leave-one-out cross-validation. Results A total of 131 out of 153 patients were included for analysis (2 excluded due to R-R interval mismatch, 20 excluded due to poor image or pressure trace quality), with basic demographics presented in Table I. The final parameter set consisted of 6 variables: systolic and diastolic brachial cuff pressures, peak GLS, R-R interval, and two strain curve principal components. When fit into a leave-one-out model, this resulted in a Pearson r-value of 0.86, demonstrating a very high correlation between predicted and invasively-derived myocardial work estimates, with a bias of 0 mmHg% and 95% limits of agreement (LOA) of ± 423 mmHg% (Figure 1). Conclusion This study demonstrates that fully automated and reliable estimates of myocardial work can be derived using 3DE and deep learning combined with routine clinical parameters. Automated and non-invasive approaches may facilitate increased clinical uptake of myocardial work assessment.

Analysis of Handwriting for Recognition of Parkinson’s Disease: Current State and New Study

One of the symptoms of Parkinson’s disease (PD) is abnormal handwriting caused by motor dysfunction. The development of tablet technology opens up opportunities for an effective analysis of the writing process of people suffering from Parkinson’s disease, aimed at supporting medical diagnosis using machine learning methods. Several approaches have been used and presented in the literature that discuss the analysis and understanding of images created during the writing of single words or sentences. In this study, we propose an analysis based on a sequence of sentences, which allows us to assess the evolution of writing over time. The study material consisted of handwriting image samples acquired in a group of 24 patients with PD and 24 healthy controls. The parameterization of the handwriting image samples was carried out using domain knowledge. Using the exhaustive search method, we selected the relevant features for the SVM algorithm performing binary classification. The results obtained were assessed using quality measures, including overall accuracy, which was 91.67%. The results were compared with competitive works on the same subject and seem to be better (a higher level of accuracy with a much smaller number of features than those presented by others).

A Machine Learning-Based Wrapper Method for Feature Selection

This paper presents a two-stage feature selection scheme using machine learning techniques. In the first stage a wrapper method is adopted to select various combinations of subsets of features from the original dataset. The performance of the model is evaluated by three classifiers: K-Nearest Neighbor (KNN), Support Vector Machines (SVM), and Random Forest (RF). In the second and final stage, a sequential backward feature selection Method is applied. The proposed method is demonstrated on eighteen datasets and the average classification accuracy of eighteen datasets achieved is 89.81%, 87.55%, and 89.82% using the KNN, SVM, and RF classifiers, respectively with a maximum reduced size of the subset being ten only. Comparing the proposed method to eight other feature selection methods, the former achieves better classification accuracy in terms of selecting the most useful but a smaller number of features.

Top-[formula omitted] discriminative feature selection with uncorrelated and [formula omitted]-norm equation constraints

Supervised feature selection (FS) as an interpretable dimensionality reduction technique has received increasing attention, where linear discriminative analysis (LDA)-based method can select informative features discriminatively and obtain promising performance. When original data has more features than samples, however, LDA-based method generally encounters degradation since the appearance of irreversible scatter matrix. This situation is called the small sample size (SSS) problem. To overcome it and enhance the discriminant power of selected feature subsets, in this paper, we design an elegant LDA-based FS model referred to as Top-k Discriminative FS (TDFS), which is constructed by seamlessly integrating the ℓ2,0-norm equation constraint into uncorrelated LDA model. More concretely, the ℓ2,0-norm equation constraint can explicitly characterize the number of selective features k to ensure the sparsity of projected matrix and select top features. The uncorrelated LDA model aims to improve discriminative ability based on uncorrelated data in projected subspace. Given the formidable nature of solving this non-convex model, a novel optimization algorithm is further developed and the SSS problem can be efficaciously addressed during the optimization process. We first decompose projection matrix into a discrete selection matrix and its corresponding nonzero projection matrix, then concurrently optimize above two matrices by employing a column-by-column update scheme, during which the reversibility of scatter matrix in selective feature subspace can be easily guaranteed to solve SSS problem. The extensive experiments on four synthetic data sets and eight real-world data sets show that the proposed method outperforms eight competitors validated by three classifiers. Moreover, although the theoretical analysis proves that our algorithm has quartic time complexity on the number of selected features k, the running time experiments verify that TDFS is still efficient and applicable in scenarios where only a small number of features need to be selected. From above perspectives, our algorithm shows desirable performance to achieve discriminative FS.

Feature subset selection using heuristic and metaheuristic approaches for diabetes prediction on a binary encoded dataset

The Machine Learning (ML) models are prone to a curse of dimensionality. The dataset with a greater number of features involves more computational cost and it may lead to low performance in the context of prediction accuracy. Therefore, in this research work we have predicted diabetes with more accuracy by using a smaller number of features. The heuristic methods Sequential Forward Selection (SFS), Sequential Backward Selection (SBS) and metaheuristic evolutionary methods — Whale Optimization Algorithm (WOA) and Genetic Algorithm (GA) are used for performing feature subset selection. The Gini index is also used as a filter evaluator. The performance of the feature subsets is analyzed by applying three different types of ML models, Random Forest (RF), Multi-Layer Perceptron (MLP) and K-Nearest Neighbor (KNN). We have predicted type-2 diabetes with an accuracy of 96.82%. Also, we have reduced the number of features up to 67.44% i.e., identified 32.56% most relevant features.

Modeling brain sex in the limbic system as phenotype for female-prevalent mental disorders

BackgroundSex differences exist in the prevalence and clinical manifestation of several mental disorders, suggesting that sex-specific brain phenotypes may play key roles. Previous research used machine learning models to classify sex from imaging data of the whole brain and studied the association of class probabilities with mental health, potentially overlooking regional specific characteristics.MethodsWe here investigated if a regionally constrained model of brain volumetric imaging data may provide estimates that are more sensitive to mental health than whole brain-based estimates. Given its known role in emotional processing and mood disorders, we focused on the limbic system. Using two different cohorts of healthy subjects, the Human Connectome Project and the Queensland Twin IMaging, we investigated sex differences and heritability of brain volumes of limbic structures compared to non-limbic structures, and subsequently applied regionally constrained machine learning models trained solely on limbic or non-limbic features. To investigate the biological underpinnings of such models, we assessed the heritability of the obtained sex class probability estimates, and we investigated the association with major depression diagnosis in an independent clinical sample. All analyses were performed both with and without controlling for estimated total intracranial volume (eTIV).ResultsLimbic structures show greater sex differences and are more heritable compared to non-limbic structures in both analyses, with and without eTIV control. Consequently, machine learning models performed well at classifying sex based solely on limbic structures and achieved performance as high as those on non-limbic or whole brain data, despite the much smaller number of features in the limbic system. The resulting class probabilities were heritable, suggesting potentially meaningful underlying biological information. Applied to an independent population with major depressive disorder, we found that depression is associated with male–female class probabilities, with largest effects obtained using the limbic model. This association was significant for models not controlling for eTIV whereas in those controlling for eTIV the associations did not pass significance correction.ConclusionsOverall, our results highlight the potential utility of regionally constrained models of brain sex to better understand the link between sex differences in the brain and mental disorders.

Machine learning approach for intrusion detection system using dimensionality reduction

As cyberspace has emerged, security in all the domains like networks, cloud, and databases has become a greater concern in real-time distributed systems. Existing systems for detecting intrusions (IDS) are having challenges coping with constantly changing threats. The proposed model, DR-DBMS (dimensionality reduction in database management systems), creates a unique strategy that combines supervised machine learning algorithms, dimensionality reduction approaches and advanced rule-based classifiers to improve intrusion detection accuracy in terms of different types of attacks. According to simulation results, the DR-DBMS system detected the intrusion attack in 0.07 seconds and with a smaller number of features using the dimensionality reduction and feature selection techniques efficiently.

Search space division method for wrapper feature selection on high-dimensional data classification

Feature selection (FS) is an essential pre-processing technique for high-dimensional data. Wrapper-based FS techniques are known for their superior performance over filter FS. However, when the dimensionality of data is very high the wrapper techniques become computationally very expensive. To solve this problem of scalability, this paper proposes the concept of search space division (SSD) which leads to smaller search spaces and hence reduced computational cost. The proposed SSD approach is generic in nature and can be integrated with any wrapper-based FS technique. The SSD approach divides search space into multiple parts and the wrapper-based FS is independently applied to each part. To facilitate the interaction of features over the complete search space, all feature subsets obtained from each part are combined and the wrapper-based FS is again applied to get the final feature subset. Moreover, a new wrapper FS technique named Binary Rao (BRAO) Algorithm has been proposed. BRAO is based on the metaphor-less, parameter-less meta-heuristic optimization algorithm Rao-1. The proposed combination of SSD and BRAO algorithm named SSDR has produced better classification accuracies with a smaller number of features in a much shorter time as compared to other well-regarded wrapper FS techniques as well as classical FS techniques over thirteen benchmark high-dimensional datasets.

Tensor completion algorithm-aided structural color design.

In recent years, structural color has developed rapidly due to its distinct advantages, such as low loss, high spatial resolution and environmental friendliness. Various inverse design methods have been extensively investigated to efficiently design optical structures. However, the optimization method for the inverse design of structural color remains a formidable challenge. Traditional optimization approaches, such as genetic algorithms require time-consuming repetitions of structural simulations. Deep learning-assisted design necessitates prior simulations and large amounts of data, making it less efficient for systems with a small number of features. This study proposes a tensor completion algorithm capable of swiftly and accurately predicting missing datasets based on partially obtained datasets to assist in structural color design. Transforming the complex physical problem of structural color design into a spatial structure relationship problem linking geometric parameters and spectral data. The method utilizes tensor multilinear data analysis to effectively capture the complex relationships associated with geometric parameters and spectral data in higher-order data. Numerical and experimental results demonstrate that the algorithm exhibits high reliability in terms of speed and accuracy for diverse structures, datasets of varying sizes, and different materials, significantly enhancing design efficiency. The proposed algorithm offers a viable solution for inverse design problems involving complex physical systems, thereby introducing a novel approach to the design of photonic devices. Additionally, numerical experiments illustrate that the structural color of cruciform resonators with diamond can overcome the high loss issues observed in traditional dielectric materials within the blue wavelength region and enhance the corrosion resistance of the structure. We achieve a wide color gamut and a high-narrow reflection spectrum nearing 1 by this structure, and the theoretical analysis further verifies that diamond holds great promise in the realm of optics.

Supervised machine learning improves general applicability of eDNA metabarcoding for reservoir health monitoring

Effective and standardized monitoring methodologies are vital for successful reservoir restoration and management. Environmental DNA (eDNA) metabarcoding sequencing offers a promising alternative for biomonitoring and can overcome many limitations of traditional morphological bioassessment. Recent attempts have even shown that supervised machine learning (SML) can directly infer biotic indices (BI) from eDNA metabarcoding data, bypassing the cumbersome calculation process of BI regardless of the taxonomic assignment of eDNA sequences. However, questions surrounding the general applicability of this taxonomy-free approach to monitoring reservoir health remain unclear, including model stability, feature selection, algorithm choice, and multi-season biomonitoring. Here, we firstly developed a novel biological integrity index (Me-IBI) that integrates multitrophic interactions and environmental information, based on taxonomy-assigned eDNA metabarcoding data. The Me-IBI can better distinguish the actual health status of the Three Gorges Reservoir (TGR) than physicochemical assessments and have a clear response to human activity. Then, taking this reliable Me-IBI as a supervised label, we compared the impact of selecting different numbers of features and SML algorithms on the stability and predictive performance of the model for predicting ecological conditions in multiple seasons using taxonomy-free eDNA metabarcoding data. We discovered that even with a small number of features, different SML algorithms can establish a stable model and obtain excellent predictive performance. Finally, we proposed a four-step strategy for standardized routine biomonitoring using SML tools. Our study firstly explores the general applicability problem of the taxonomy-free eDNA-SML approach and establishes a solid foundation for the large-scale and standardized biomonitoring application.

LASSO and Elastic Net Tend to Over-Select Features

Machine learning methods have been a standard approach to select features that are associated with an outcome and to build a prediction model when the number of candidate features is large. LASSO is one of the most popular approaches to this end. The LASSO approach selects features with large regression estimates, rather than based on statistical significance, that are associated with the outcome by imposing an L1-norm penalty to overcome the high dimensionality of the candidate features. As a result, LASSO may select insignificant features while possibly missing significant ones. Furthermore, from our experience, LASSO has been found to select too many features. By selecting features that are not associated with the outcome, we may have to spend more cost to collect and manage them in the future use of a fitted prediction model. Using the combination of L1- and L2-norm penalties, elastic net (EN) tends to select even more features than LASSO. The overly selected features that are not associated with the outcome act like white noise, so that the fitted prediction model may lose prediction accuracy. In this paper, we propose to use standard regression methods, without any penalizing approach, combined with a stepwise variable selection procedure to overcome these issues. Unlike LASSO and EN, this method selects features based on statistical significance. Through extensive simulations, we show that this maximum likelihood estimation-based method selects a very small number of features while maintaining a high prediction power, whereas LASSO and EN make a large number of false selections to result in loss of prediction accuracy. Contrary to LASSO and EN, the regression methods combined with a stepwise variable selection method is a standard statistical method, so that any biostatistician can use it to analyze high-dimensional data, even without advanced bioinformatics knowledge.

Combined Method for Informative Feature Selection for Speech Pathology Detection

The task of detecting vocal abnormalities is characterized by a small amount of available data for training, as a consequence of which classification systems that use low-dimensional data are the most relevant. We propose to use LASSO (least absolute shrinkage and selection operator) and BSS (backward stepwise selection) methods together to select the most significant features for the detection of vocal pathologies, in particular amyotrophic lateral sclerosis. Features based on fine-frequency cepstral coefficients, traditionally used in speech signal processing, and features based on discrete estimation of the autoregressive spectrum envelope are used. Spectral features based on the autoregressive process envelope spectrum are extracted using the generative method, which involves calculating a discrete Fourier transform of the report sequence generated using the autoregressive model of the input voice signal. The sequence is generated by the autoregressive model so as to account for the periodic nature of the Fourier transform. This improves the accuracy of the spectrum estimation and reduces the spectral leakage effect. Using LASSO in conjunction with BSS allowed us to improve the classification efficiency using a smaller number of features as compared to using the LASSO method alone.

Small immunological clocks identified by deep learning and gradient boosting.

The aging process affects all systems of the human body, and the observed increase in inflammatory components affecting the immune system in old age can lead to the development of age-associated diseases and systemic inflammation. We propose a small clock model SImAge based on a limited number of immunological biomarkers. To regress the chronological age from cytokine data, we first use a baseline Elastic Net model, gradient-boosted decision trees models, and several deep neural network architectures. For the full dataset of 46 immunological parameters, DANet, SAINT, FT-Transformer and TabNet models showed the best results for the test dataset. Dimensionality reduction of these models with SHAP values revealed the 10 most age-associated immunological parameters, taken to construct the SImAge small immunological clock. The best result of the SImAge model shown by the FT-Transformer deep neural network model has mean absolute error of 6.94 years and Pearson ρ = 0.939 on the independent test dataset. Explainable artificial intelligence methods allow for explaining the model solution for each individual participant. We developed an approach to construct a model of immunological age based on just 10 immunological parameters, coined SImAge, for which the FT-Transformer deep neural network model had proved to be the best choice. The model shows competitive results compared to the published studies on immunological profiles, and takes a smaller number of features as an input. Neural network architectures outperformed gradient-boosted decision trees, and can be recommended in the further analysis of immunological profiles.

GBJOF: Gradient Boosting Integrated with Jaya Algorithm to Optimize the Features in Malware Analysis

Malware analysis is used to identify suspicious file transferring in the network. It can be identified efficiently by using the reverse engineering hybrid approach. Implementing a hybrid approach depends on the feature selection because the dataset contains static and dynamic parameters. The given dataset contains 85 attributes with 10 different class labels. Since it has high dimensional and multi-classification data, existing approaches of ML could be more efficient in reducing the features. The model combines the enhanced JAYA genetic algorithm with a gradient boosting technique to identify the efficiency and a smaller number of features. Many existing approaches for feature selection either implement correlation analysis or wrapper techniques. The major disadvantages of these issues are that they are facing fitting problems with a very small number of features. With the Usage of the genetic approach, this paper has achieved 95% accuracy with 12 features, approximately 7% greater than ML approaches.

Reduced Deep Convolutional Activation Features (R-DeCAF) in Histopathology Images to Improve the Classification Performance for Breast Cancer Diagnosis.

Breast cancer is the second most common cancer among women worldwide, and the diagnosis by pathologists is a time-consuming procedure and subjective. Computer-aided diagnosis frameworks are utilized to relieve pathologist workload by classifying the data automatically, in which deep convolutional neural networks (CNNs) are effective solutions. The features extracted from the activation layer of pre-trained CNNs are called deep convolutional activation features (DeCAF). In this paper, we have analyzed that all DeCAF features are not necessarily led to higher accuracy in the classification task and dimension reduction plays an important role. We have proposed reduced DeCAF (R-DeCAF) for this purpose, and different dimension reduction methods are applied to achieve an effective combination of features by capturing the essence of DeCAF features. This framework uses pre-trained CNNs such as AlexNet, VGG-16, and VGG-19 as feature extractors in transfer learning mode. The DeCAF features are extracted from the first fully connected layer of the mentioned CNNs, and a support vector machine is used for classification. Among linear and nonlinear dimensionality reduction algorithms, linear approaches such as principal component analysis (PCA) represent a better combination among deep features and lead to higher accuracy in the classification task using a small number of features considering a specific amount of cumulative explained variance (CEV) of features. The proposed method is validated using experimental BreakHis and ICIAR datasets. Comprehensive results show improvement in the classification accuracy up to 4.3% with a feature vector size (FVS) of 23 and CEV equal to 0.15.

A novel systematic pipeline for increased predictability and explainability of growth patterns in children using trajectory features

ObjectiveLongitudinal patterns of growth in early childhood are associated with health conditions throughout life. Knowledge of such patterns and the ability to predict them can lead to better prevention and improved health promotion in adulthood. However, growth analyses are characterized by significant variability, and pattern detection is affected by the method applied. Moreover, pattern labelling is typically performed based on ad hoc methods, such as visualizations or clinical experience. Here, we propose a novel pipeline using features extracted from growth trajectories using mathematical, statistical and machine-learning approaches to predict growth patterns and label them in a systematic and unequivocal manner. Methods: We extracted mathematical and clinical features from 9577 children growth trajectories embedded with machine-learning predictions of the growth patterns. We experimented with two sets of features (CAnonical Time-series Characteristics and trajectory features specific to growth), developmental periods and six machine-learning classifiers. Clinical experts provided labels for the detected patterns and decision rules were created to associate the features with the labelled patterns. The predictive capacity of the extracted features was validated on two heterogenous populations (The Applied Research Group for Kids and the 2004 Pelotas Birth Cohort, based in Canada and Brazil, respectively). Results: Features predictive ability measured by accuracy and F1 score was ≥ 80% and ≥ 0.76 respectively in both cohorts. A small number of features (n = 74) was sufficient to distinguish between growth patterns in both cohorts. Slope, intercept of the trajectory, age at peak value, start value and change of the growth measure were among the top identified features. Conclusion: Growth features can be reliably used as predictors of growth patterns and provide an unbiased understanding of growth patterns. They can be used as tool to reduce the effort to repeat analysis and variability concerning anthropometric measures, time points and analytical methods, in the context of the same or similar populations.

AWC‐NIDS: Attack‐wise customized network intrusion detection system using machine learning, concurrency, and distributed systems

SummaryWith digitization and modern network applications, information security has gained a tremendous importance. Therefore, accurate and efficient detection systems are crucial for maintaining proactive security in computer networks. Machine learning (ML) has shown great potential as a promising solution since it can teach a machine to distinguish malicious and normal network activities. However, recently proposed methods are suffering from at least one of the following: detection accuracy, false alarm rate, and computational complexity issues. The main reason behind this problem is the complexity of the model in terms of attack types. From the ML perspective, intrusion detection is a classification problem where each attack type is identified by a set of different features, and features are used for classifying network activities. Thus, training an ML algorithm to detect more than one attack type leads to a more complex model; the increasing number of used features contributes positively to the model complexity, and may result in relatively lower detection accuracy or a higher false positive rate. To tackle this problem, this study proposes an attack‐wise customized network intrusion detection system (AWC‐NIDS) based on ML, concurrency, and distributed systems to achieve accurate and efficient network‐wide intrusion detection. Since CICIDS2017 contains many modern attacks, it was used for model development and performance evaluation. The experimental results showed that the proposed methodology achieved high classification performance for all datasets with a small number of features. However, it was observed that the lowest accuracy was achieved for the comprehensive dataset (which contains all attack types); for the single attack‐type datasets, the obtained accuracy was above 99%. This finding proves the concept of attack‐wise customization for intrusion detection and shows the significance of the proposed methodology. In conclusion, this framework is promising for implementing robust and accurate cybersecurity systems for traditional and modern networking.

Compact-Fusion Feature Framework for Ethnicity Classification

In computer vision, ethnicity classification tasks utilize images containing human faces to extract ethnicity labels. Ethnicity is one of the soft biometric feature categories useful in data analysis for commercial, public, and health sectors. Ethnicity classification begins with face detection as a preprocessing process to determine a human’s presence; then, the feature representation is extracted from the isolated facial image to predict the ethnicity class. This study utilized four handcrafted features (multi-local binary pattern (MLBP), histogram of gradient (HOG), color histogram, and speeded-up-robust-features-based (SURF-based)) as the basis for the generation of a compact-fusion feature. The compact-fusion framework involves optimal feature selection, compact feature extraction, and compact-fusion feature representation. The final feature representation was trained and tested with the SVM One Versus All classifier for ethnicity classification. When it was evaluated in two large datasets, UTKFace and Fair Face, the proposed framework achieved accuracy levels of 89.14%, 82.19%, and 73.87%, respectively, for the UTKFace dataset with four or five classes and the Fair Face dataset with four classes. Furthermore, the compact-fusion feature with a small number of features at 4790, constructed based on conventional handcrafted features, achieved competitive results compared with state-of-the-art methods using a deep-learning-based approach.

Combinatorial online high‐order interactive feature selection based on dynamic graph convolution network

Traditional feature selection algorithms assume that the sample and feature space is known before learning, while most of the data is feature streams or data streams in reality. Currently, streaming feature selection algorithms can retain relevant features by removing redundant and irrelevant features based on the interaction between features, but they ignore the specific number of features that have interaction. Most of the existing studies only consider the case of interaction between two features, which is not quite in line with most realistic scenarios, i.e., the number of features with interaction is unknown. This paper concentrates on the high-order interactions between stream feature, and proposes a Combinatorial High-Order Interactive Feature Selection based on Dynamic GCN and Sparse learning (CHOIFS-DGS). Based on previous definitions of feature interaction, this paper proposes some new metrics to measure the degree of interaction between a newly arrived feature and an already selected feature. CHOIFS-DGS consists of three main parts, namely: low-order online feature selection based on interaction measure, high-order online feature selection based on dynamic GCN, and Intra-group sparse feature selection. In the experimental analysis section, this study employs two different classifiers and eleven publicly released data sets, including gene data related to diseases and data from two classification challenges (NeurIPS 2003 feature selection challenge and WCCI 2006 Performance Prediction Challenge). The experimental results demonstrate that the proposed CHOIFS-DGS model significantly improves classification accuracy on all eleven data sets, while using a relatively smaller number of features, thus fulfilling the role of key feature selection. Furthermore, the CHOIFS-DGS algorithm consists of three components: LO-OIFS, HO-OIFS, and Group-Sparse. By applying these three sub-modules separately for extracting data features and comparing the results with CHOIFS-DGS, it is found that the performance of CHOIFS-DGS is lower than that of the individual sub-modules only in three data sets, while significantly better in the remaining eight data sets. This indicates that the integrated use of the three sub-modules can enhance model accuracy. Finally, in the ablation experiment, to verify the necessity of considering higher-order interactions among features, the results of the HO-OIFS module were compared with those of the other modules. The results show that the model’s accuracy significantly improves after incorporating the HO-OIFS module, thereby demonstrating that considering higher-order interactions between features is essential.