Feature selection method based on evolutionary hypergraph learning, electronic equipment and storage medium

Abstract

The invention provides a feature selection method based on evolutionary hypergraph learning, electronic equipment and a storage medium, and belongs to the field of artificial intelligence and machine learning. The method comprises the following steps: firstly, introducing a hypergraph structure to model a high-order feature interaction relationship, and comprehensively evaluating feature importance in combination with a genetic algorithm; a feature importance oriented gene pool is constructed based on an evaluation result, effective knowledge migration is realized among different populations through a specially designed cross strategy, and the search efficiency and the feature subset quality are improved. According to the method, experimental verification is carried out on a plurality of data sets with different dimensions, and a result shows that the method can stably select a higher-quality feature subset, and the balance accuracy of the model in a classification task is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning technology, specifically relating to a feature selection method, electronic device and storage medium based on evolutionary hypergraph learning. Background Technology

[0002] High-dimensional data is prevalent in many real-world fields, such as genomics, biomedicine, and disease screening. For example, in genomics, gene expression data is typically obtained through high-throughput sequencing, a time-consuming and costly process that limits the number of usable samples. However, gene expression data often contains tens of thousands of genes, naturally exhibiting sparse, high-dimensional characteristics. Directly using this type of high-dimensional data for model learning in disease screening often leads to classifier overfitting, thus affecting prediction accuracy.

[0003] Feature selection methods can eliminate redundant or irrelevant features, thereby reducing data dimensionality and improving the classification performance of the model. However, the feature selection problem has been proven to be NP-hard; as the feature dimensionality increases, its search space expands exponentially, posing a significant challenge to existing feature selection methods when handling ultra-high-dimensional data. Evolutionary computation, due to its advantages in global search, is widely used in feature selection tasks. However, most existing evolutionary computation-based feature selection methods do not consider the complex interactions between features, which may lead to the accidental removal of some potentially useful features.

[0004] In high-dimensional data, one or more features may interact with other features, and these interactions can be second-, third-, or even multi-order. Some features may seem insignificant when viewed alone, but become crucial after interacting with other features; while some features that appear important on their own may become redundant in the interaction. The interactions between features are often highly coupled and non-linear, posing a challenge to the selection of informative features.

[0005] While some methods attempt to model feature interactions using traditional graph structures, these methods struggle to effectively capture inherent higher-order feature relationships. Hypergraphs, as a generalization of graphs, can connect multiple strongly correlated nodes through hyperedges, thereby revealing more complex higher-order feature interactions. Therefore, this invention employs hypergraphs to model the interactions between features, thereby guiding the evolutionary process to more accurately identify informative features. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a feature selection method based on evolutionary hypergraph learning. Specifically, all features are first divided into multiple groups, and a genetic algorithm (GA) population is initialized for each group to search for discriminative features within that group. Furthermore, to characterize the high-order feature interactions within each group, a hypergraph is introduced to model the correlations between features, thereby guiding the population to search for optimal solutions. During the evolutionary learning process, beneficial knowledge is transferred between different populations to improve the performance of underperforming individuals in each population. Finally, the best-performing individuals from all populations are selected and integrated to form the final feature subset.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a feature selection method based on evolutionary hypergraph learning, comprising: S1. Cluster all features and divide them into several feature subsets; S2. Construct a feature hypergraph for each feature subset based on the Spearman correlation coefficient between features; S3. Determine the importance score of each node in the feature hypergraph based on the Fisher score and structural centrality of the nodes in the feature hypergraph; S4. Create an initial population for each feature subset, calculate the fitness of each individual, and initialize the gene pool based on the importance score of the node. S5. Use a group-based crossover strategy to evolve the current population, generate a new generation of population, and update the information in the gene pool. S6. Determine whether the current population meets the termination condition. If so, select the individual with the highest fitness from all feature subsets, and the corresponding feature set is the final selected feature; otherwise, return to step S5.

[0008] Furthermore, in step S1, based on the Euclidean distance between each feature, the k-means clustering algorithm is used to divide all features into groups. A feature subset is used to group features with similar distributions and similar variation patterns into the same cluster, thereby achieving a structured organization of the feature space.

[0009] Furthermore, to determine the optimal number of clusters... The silhouette coefficient method is introduced to address different... The clustering results corresponding to the values ​​are evaluated, and the optimal one is automatically selected. The value is used to rationally divide the feature space.

[0010] Furthermore, in step S2, the method for constructing the feature hypergraph is as follows: S2.1, Represent any feature subset as And Each feature in the graph is mapped to a node, forming a node set. ,in Indicates the first One characteristic, Representation of features The corresponding mapping node, ; S2.2 Calculate the feature subset The sample sorting for each feature is as follows: Given a feature The sample size is , sample values Sort in ascending order to obtain the sorted vector. ; S2.3, For feature subsets Any two features and Calculate the difference in sample ranking between the two: Calculate the sum of the differences of squares: Calculation features and Spearman correlation coefficient between : This yields the feature subset. Spearman correlation coefficient matrix in: S2.4, For Each node in Select those that meet the correlation coefficient And significance level nodes ,and Together they form a superedge: After calculating all the hyperedges of the hypergraph, the hyperedges and the remaining isolated nodes together constitute a feature subset. The feature hypergraph.

[0011] Furthermore, in step S3, the importance score of each node in the feature hypergraph is calculated, as follows: S3.1 To evaluate the importance of the interactions represented by hyperedges in the feature hypergraph, calculate the weight of each hyperedge; hyperedge The weight of a node is defined as the average Fisher Score of all nodes it contains: in, Indicates the superedge The number of nodes contained therein; Indicates that it is contained in the hyperedge Nodes in; Indicates that it is contained in the hyperedge Node corresponding features Fisher Score; S3.2 To measure the importance of nodes in higher-order interactions, the node importance score is calculated by combining two dimensions: structure centrality and classification ability. in, Represents a node Importance score Represents a node The degree of (i.e., the number of superedges it participates in). Represents all containing nodes The set of superedges This represents the weight of the hyperedge. The higher the degree of a node and the stronger the discriminative power of its superedge, the higher its value as a feature. Specifically, for isolated nodes not in any hyperedge, according to... We calculate the importance score to obtain the importance score of all nodes in the hypergraph.

[0012] Furthermore, based on Fisher Score, the discriminative power of the features corresponding to each node in the feature hypergraph is calculated, assuming the set of category labels is... ,feature The Fisher Score calculation formula is as follows: in, Indicates the first The number of samples in each class For the first The feature in the first The mean of the class, For the first The mean of each feature across all samples This represents the within-class variance.

[0013] Furthermore, the specific process of step S4 includes: S4.1. Create an initial population for each feature subset. Each individual in the initial population is represented as a binary string. This encoding method is used to identify the selection status of the current individual in the feature subset. When the corresponding index in the individual is 1, it means that the corresponding feature is selected, and 0 means that it is not selected. S4.2 To evaluate the performance of individuals, considering both balanced accuracy and the proportion of selected features, the fitness function is designed as follows: in, This represents the fitness value of an individual. This represents the average recall rate across all categories; This indicates the proportion of the currently selected feature. Indicates the number of selected features. Indicates the total number of features; The adjustment parameters are used to control the weights of the two parts; S4.3 Each feature subset's population corresponds to an independent gene pool, used to store the importance score of each feature in that feature subset. The frequency with which the feature is selected by the optimal individual is recorded; the selection frequency of all features is initialized to 0 and updated generation by generation during evolution; the formula for calculating the selection frequency is: in, The total number of iterations; For indicator functions, if the first... The optimal individual in the generation selected the feature. ,but Otherwise, it is 0; Ultimately, features The total weight is: in, Representation of features Initial importance score calculated based on the hypergraph The result after normalization The larger the value, the more important the feature.

[0014] Furthermore, in step S5, to achieve information exchange between feature subsets, a group-based cross-strategy is introduced to promote knowledge transfer. The group-based cross-strategy is as follows: For each feature subset corresponding to the population, in each round of evolution, individuals with high fitness are selected as parents to participate in the next crossover based on their fitness. ​The parent individuals in the current population are divided into three subgroups according to their fitness from largest to smallest. The first subgroup consists of the top third of the individuals, the second subgroup consists of the middle third of the individuals, and the third subgroup consists of the bottom third of the individuals. The first subgroup's crossover strategy: Individuals within the first subgroup undergo single-point crossover, which is only performed within the subgroup; and an elite strategy is adopted, directly copying the best individuals from the previous generation to the next generation; The crossover strategy for the second subgroup: Individuals within the second subgroup are crossovered with the gene pool of their own population. Specifically, the number of features to be removed is determined. And the number of new features that need to be added Randomly remove 1. Select features from the gene pool that are not in the current individual and have not been recently removed as candidate features; 2. Randomly select 3 features from the candidate features each time, and calculate the total weight of the 3 features. Add the feature with the largest value to the individual, and repeat this process until the feature with the largest value is added. One feature; The crossover strategy for the third subgroup is similar to that in the second subgroup, except that it randomly selects another feature subset of the population before crossover and crossovers individuals in the third subgroup with the gene pool of the other feature subset of the population, thereby achieving cross-population knowledge transfer and helping to improve the performance of low-fit individuals. After the crossover is completed, the fitness value of the offspring is updated according to formula (9), and the individual with the highest fitness is selected. The characteristics selected by the individual are counted and the gene pool is updated according to formulas (10) and (11).

[0015] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the electronic device performs the feature selection method based on evolutionary hypergraph learning.

[0016] A third aspect of the present invention provides a storage medium comprising a computer program that, when run on an electronic device, causes the electronic device to perform the feature selection method based on evolutionary hypergraph learning.

[0017] The beneficial effects of this invention are: (1) This invention constructs a feature hypergraph by calculating the higher-order correlation between features, which effectively evaluates the importance of features for classification, increases the probability of key features being selected during the evolution process, makes the search more comprehensive, and the feature hypergraph can guide the evolutionary algorithm and effectively improve the classification accuracy. (2) This invention reduces the search space of each population by evolving separately in the divided feature subsets, and designs a group-based crossover strategy to promote the transfer of important knowledge in different populations, improve the inferior individuals in each population, further improve the classification accuracy, and effectively reduce the number of selected features. Attached Figure Description

[0018] Figure 1 This is a model diagram of the feature selection method based on evolutionary hypergraph learning in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, this embodiment of the invention provides a feature selection method based on evolutionary hypergraph learning, which includes the following steps: First, divide all features into Each feature subset is used to construct a feature hypergraph based on the Spearman correlation coefficient between features; Then, the importance score of each node (feature) is determined by combining the Fisher score and structural centrality of the nodes in the feature hypergraph; Next, a population is created for each feature subset, the fitness of each individual is calculated, and the gene pool is initialized based on the importance score of the node. Subsequently, individuals in each population are evaluated, and a tournament selection strategy is used to select the better individuals as parents for the next generation. In each generation, individuals are divided into three groups based on their fitness values: individuals in the first group undergo single-point crossover within the group, individuals in the second group undergo crossover with the gene pool of their own group, and individuals in the third group undergo crossover with the gene pools of other subsets. After the crossover and mutation operations, a new generation of population is generated, and the information in the gene pool is updated simultaneously. Finally, when the termination condition is met, the individual with the highest fitness is selected from all feature subsets, and the corresponding feature set is the final selected feature.

[0021] To verify the effectiveness of the feature selection method based on evolutionary hypergraph learning proposed in this invention, 15 real-world datasets were selected from open-source datasets as experimental subjects (see Table 1). These datasets cover a variety of common data types, including text data, face recognition images, biometric data, and other types of data such as speech recognition, and have good representativeness and diversity.

[0022] In these datasets, the number of features (ND) ranges from 166 to 15,154 dimensions, covering different feature space scales from small-to-medium scale to ultra-high-dimensional data, enabling a comprehensive evaluation of the search capabilities of this invention across different dimensions. Furthermore, the number of samples (NS) and the number of classes (NC) vary across datasets, with some datasets exhibiting class imbalance. This allows for further investigation into the adaptability and stability of the method under different sample sizes, number of classes, and class distributions, making it suitable for comprehensive evaluation of binary and multi-class classification tasks.

[0023] The dataset was split into 70% training set and 30% test set. To mitigate the impact of class imbalance on the experimental results, a stratified sampling strategy was used during the splitting process. Furthermore, to ensure consistent feature scale during training, all features were normalized before training.

[0024] During the training phase, to ensure the stability and fairness of model evaluation, 5-fold cross-validation was used for performance evaluation, and a k-nearest neighbor classifier (k-NN, k=5) was used as the evaluation classifier to maintain consistency with mainstream feature selection literature. Notably, stratified sampling was also employed during cross-validation to ensure a consistent class distribution across each fold, thereby further reducing the impact of class imbalance on model performance.

[0025] Finally, during the testing phase, the same k-nearest neighbor classifier (k=5) was used to evaluate the classification performance of the selected feature subset on the test set, ensuring the consistency and objectivity of the method evaluation.

[0026] Table 1. Experimental Dataset The method proposed in this invention was compared with five feature selection methods based on evolutionary computation (EC) and four non-evolutionary computation methods. A brief description of these methods is as follows: GA (Genetic Algorithm): GA simulates the natural evolutionary process by iteratively searching for the optimal subset of features through operations such as selection, crossover, and mutation.

[0027] Particle Swarm Optimization (PSO): PSO simulates the cooperative search behavior of a swarm of particles. Based on the experience of individual optimality and global optimality, it continuously updates the position and velocity of particles to search for the optimal subset of features.

[0028] SFE-PSO: This method first performs a fast coarse screening, and then applies a particle swarm optimization algorithm to perform fine feature selection in the dimensionality-reduced subspace.

[0029] GBFS: This method uses NSGA-II dynamic optimization of the graph structure and combines k-shell decomposition to preserve features with high centrality.

[0030] JSEMO: This method is based on NSGAII-SDR, which maintains the diversity of solutions by introducing Jaccard similarity and combines two crossover operators to improve search capabilities.

[0031] infFS_S: This method constructs a graph to model the relationships between features, then calculates the importance score based on the total weight of all paths originating from each feature, and selects features accordingly.

[0032] infFS_U: This method is a variant of infFS that constructs a feature map based solely on the statistical properties between features, without relying on any label information. A subset of high-quality features is then selected from the constructed map.

[0033] ReliefF: This method evaluates the importance of each feature in distinguishing neighboring samples by the distance relationship between samples.

[0034] SPCAF: This method models feature selection as a problem of minimizing reconstruction error and introduces norm regularization to enhance sparsity.

[0035] For multi-objective methods, including GBFS and JSEMO, the feature subset used in the testing phase is the solution corresponding to the best-performing individual in the Pareto front during training. All methods based on evolutionary computation (EC) are run independently 30 times using 30 different random seeds. To evaluate the performance of the proposed feature selection method, two evaluation metrics are used: Balanced Accuracy and Feature Subset Size.

[0036] In the experiment, the population size was set to 100, and the number of generations was set to 100. The crossover probability (cxpb) was set to 0.8, and the mutation probability (mutpb) was set to 0.2. The tournament selection size was set to 3. The weight coefficients in the fitness function were... The threshold is set to 0.9. During individual encoding, the selection threshold for each feature is set to 0.5. In the method of this invention, the relevance threshold for constructing the hypergraph is set to 0.8, and the significance coefficient is set to 0.05. All benchmark evolutionary computation feature selection methods used in the comparison employ the same parameter configuration.

[0037] The proposed method was run independently 30 times on each dataset under 30 different random seeds. A t-test at a significance level of 0.05 was also performed to assess the statistical significance of the performance difference between the proposed method and the baseline method used for pairwise comparison.

[0038] Tables 2 and 3 present the average balanced accuracy of the proposed method and various benchmark EC (evolutionary computation) feature selection methods on 15 datasets. All results are based on the average of 30 independent runs on the test set.

[0039] Table 2. Comparison of equilibrium accuracy with evolutionary computation-based feature selection methods (GA, PSO, SFE-PSO) Table 3. Comparison of balanced accuracy with evolutionary computation-based feature selection methods (GBFS, JSEMO) In Tables 2 and 3, the symbol "+" indicates that the present invention is statistically significantly superior to the corresponding benchmark method, the symbol "-" indicates that it is significantly inferior, and the symbol "=" indicates that there is no significant difference. The "All Features" column represents the balanced classification accuracy when all features are used without feature selection.

[0040] Compared to results using all features, this invention achieved higher balanced accuracy on 13 out of 15 datasets, demonstrating a significant advantage in improving classification performance. Furthermore, compared to EC benchmark methods, this invention outperforms other EC methods in terms of overall performance under statistical significance testing. In a total of 75 comparison cases, this invention significantly outperformed other EC methods in 52 of them.

[0041] Table 4 reports the experimental results of this invention and four non-EC methods on 15 datasets. These methods select feature subsets based on different feature scoring mechanisms. Since these methods are deterministic, the balanced accuracy of a single run is recorded. Notably, in ReliefF and SPCAF, the number of selected features is set to 30% of the total number of features; while infFS_S and infFS_U automatically determine the number of features through the algorithm. Statistical results show that this invention outperforms the four non-EC methods overall. In a total of 60 comparison cases, this invention achieves significantly higher classification accuracy in 52 cases.

[0042] Table 4. Comparison of balanced accuracy between feature selection methods and non-evolutionary computational methods Tables 5 and 6 further summarize the significance test results and average rankings of the EC method and the non-EC method, in order to further analyze the classification ability of the feature subset selected in this invention.

[0043] As shown in Table 5, the present invention demonstrates significantly better classification performance compared to three single-objective EC methods (GA, PSO, and SFE-PSO). In 15 comparative cases, the present invention significantly outperforms GA, PSO, and SFE-PSO in 11, 11, and 14 cases, respectively.

[0044] For the two multi-objective methods (GBFS and JSEMO), the individuals with the highest classification accuracy from their respective Pareto fronts were selected as representatives for comparison. Compared with GBFS, the present invention is significantly better in 9 cases, indicating its stronger classification performance; compared with the current state-of-the-art JSEMO method, the present invention performs better in 7 cases and is only slightly inferior in 5 cases, showing a slight overall advantage.

[0045] Therefore, from a statistical perspective, this invention significantly outperforms the three single-objective methods and surpasses the optimal solutions of the two multi-objective methods in terms of balanced accuracy. Furthermore, considering the average ranking of the six EC methods, this invention achieves the lowest (i.e., best) average ranking, indicating its optimal overall performance.

[0046] Table 5. Summary of the results of the EC-based feature selection algorithm on 15 test sets. Table 6 summarizes the significance comparison results between the present invention and four non-EC methods. Compared with infFS_S, infFS_U, ReliefF, and SPCAF, the present invention significantly outperforms the other methods in 13, 13, 11, and 15 cases, respectively, out of all 15 comparison cases. Based on the statistical significance test results, the present invention is significantly superior to these four non-EC methods.

[0047] Furthermore, based on the average ranking results, this invention achieved the lowest average ranking (1.5333) among the five methods, further demonstrating its significant advantage over existing non-EC methods in terms of balanced accuracy.

[0048] Table 6. Summary of Results of Non-EC-Based Feature Selection Algorithms on 15 Test Sets Ablation experiments were conducted on all datasets to further explore the impact of introducing hypergraph modeling and knowledge transfer modules in this invention. The experimental results are shown in Table 7, where the last column gives the average running time of each iteration of this invention.

[0049] Table 7. Ablation experiments of the main modules of this invention and the average running time of each iteration In this study, the hypergraph modeling module, denoted as "Invention-HG," was first removed. Specifically, the feature importance scores calculated based on the hypergraph were removed from the gene pool, retaining only the selection frequency of features during evolution, while other settings remained unchanged. The results in Table 7 show that the balance accuracy decreased for most datasets. This indicates that the feature importance obtained through hypergraph modeling can effectively guide the algorithm in selecting key features.

[0050] Next, the knowledge transfer module was removed, denoted as "Invention-KT" in Table 7. In this setting, the feature grouping mechanism was disabled, and all features were searched within a unified population. In the crossover operation, the interaction between populations and the gene pool was cancelled, and a new strategy was adopted: 50% of the crossovers were performed within the current population, and the other 50% were performed with the gene pool. The results in Table 7 show that, in most cases, the balancing accuracy decreased and the number of selected features increased.

[0051] These results demonstrate that dividing features into multiple subspaces and having them searched by different populations, while combining cross-population gene pool knowledge transfer mechanisms, can effectively enhance the search capabilities of weak individuals, thereby improving the overall feature selection effect.

[0052] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any limitation on the present invention. Those skilled in the art should fully understand that modifications to the technical solutions described in the foregoing embodiments or equivalent substitutions for any part or all of the technical features are entirely feasible. Such modifications or substitutions, as long as they do not depart from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. A feature selection method based on evolutionary hypergraph learning, characterized in that, include: S1. Cluster all features and divide them into several feature subsets; S2. Construct a feature hypergraph for each feature subset based on the Spearman correlation coefficient between features; S3. Determine the importance score of each node in the feature hypergraph based on the Fisher score and structural centrality of the nodes in the feature hypergraph; S4. Create an initial population for each feature subset, calculate the fitness of each individual, and initialize the gene pool based on the importance score of the node. S5. Use a group-based crossover strategy to evolve the current population, generate a new generation of population, and update the information in the gene pool. S6. Determine whether the current population meets the termination condition. If so, select the individual with the highest fitness from all feature subsets. The corresponding feature set is the final selected feature. Otherwise, return to step S5.

2. The feature selection method based on evolutionary hypergraph learning according to claim 1, characterized in that, In step S1, based on the Euclidean distance between each feature, the k-means clustering algorithm is used to divide all features into groups. A subset of features.

3. The feature selection method based on evolutionary hypergraph learning according to claim 2, characterized in that, To determine the optimal number of clusters The contour coefficient method is introduced for different The clustering results corresponding to the values ​​are evaluated, and the optimal one is automatically selected. value.

4. The feature selection method based on evolutionary hypergraph learning according to claim 2 or 3, characterized in that, In step S2, the method for constructing the feature hypermap is as follows: S2.1, Represent any feature subset as And Each feature in the graph is mapped to a node, forming a node set. ,in Indicates the first One characteristic, Representation of features The corresponding mapping node, ; S2.2 Calculate the feature subset The sample sorting for each feature is as follows: Given a feature The sample size is , sample values Sort in ascending order to obtain the sorted vector. ; S2.3, For feature subsets Any two features and Calculate the difference in sample ranking between the two: Calculate the sum of the differences of squares: Calculation features and Spearman correlation coefficient between : This yields the feature subset. Spearman correlation coefficient matrix in: S2.4, For Each node in Select those that meet the correlation coefficient And significance level nodes ,and Together they form a superedge: After calculating all the hyperedges of the hypergraph, the hyperedges and the remaining isolated nodes together constitute a feature subset. The feature hypergraph.

5. The feature selection method based on evolutionary hypergraph learning according to claim 4, characterized in that, In step S2, and in step S3, the importance score of each node in the feature hypergraph is calculated, as follows: S3.1, Super Edge The weight of a node is defined as the average Fisher Score of all nodes it contains: in, Indicates the superedge The number of nodes contained therein; Indicates that it is contained in the hyperedge Nodes in; Indicates that it is contained in the hyperedge Nodes in Corresponding features Fisher Score; S3.2 Calculate the importance score of each node: in, Represents a node Importance score Represents a node The degree of a is equal to the number of hyperedges it participates in. Represents all containing nodes The set of superedges The weight of the superedge.

6. The feature selection method based on evolutionary hypergraph learning according to claim 5, characterized in that, Let the set of category labels be ,feature The Fisher Score calculation formula is as follows: in, Indicates the first The number of samples in each class For the first The feature in the first The mean of the class, For the first The mean of each feature across all samples The variance is the within-class variance.

7. The feature selection method based on evolutionary hypergraph learning according to claim 5 or 6, characterized in that, The specific process of step S4 includes: S4.

1. Create an initial population for each feature subset. Each individual in the initial population is represented as a binary string. This encoding method is used to identify the selection status of the current individual in the feature subset. When the corresponding index in the individual is 1, it means that the corresponding feature is selected, and 0 means that it is not selected. S4.2 The fitness function is designed as follows:   in, This represents the fitness value of an individual. This represents the average recall rate across all categories; This indicates the proportion of the currently selected feature. Indicates the number of selected features. Indicates the total number of features; The adjustment parameters are used to control the weights of the two parts; S4.3 Each feature subset's population corresponds to an independent gene pool, used to store the importance score of each feature in that feature subset. The frequency with which the feature is selected by the optimal individual is recorded; the selection frequency of all features is initialized to 0 and updated generation by generation during evolution; the formula for calculating the selection frequency is: in, The total number of iterations; For indicator functions, if the first... The optimal individual in the generation selected the feature. ,but Otherwise, it is 0; Ultimately, features The total weight is: in, Representation of features Initial importance score calculated based on the hypergraph The result after normalization.

8. The feature selection method based on evolutionary hypergraph learning according to claim 7, characterized in that, In step S5, the cross-grouping strategy is as follows: For each feature subset corresponding to the population, in each round of evolution, individuals with high fitness are selected as parents to participate in the next crossover based on their fitness. The parent individuals in the current population are divided into three subgroups according to their fitness from largest to smallest. The first subgroup consists of the top third of the individuals, the second subgroup consists of the middle third of the individuals, and the third subgroup consists of the bottom third of the individuals. The first subgroup's crossover strategy: Individuals within the first subgroup undergo single-point crossover, which is only performed within the group; and an elite strategy is adopted, directly copying the best individuals from the previous generation to the next generation; The crossover strategy for the second subgroup: Individuals within the second subgroup are crossovered with the gene pool of their own population. Specifically, the number of features to be removed is determined. And the number of new features that need to be added Randomly remove Features; Select features from the gene pool that are not in the current individual and have not been recently removed as candidate features; Each time, three features are randomly selected from the candidate features, and the total weight of the three features is selected. The feature with the largest value is added to the individual, and this process is repeated until the maximum value is added. One feature; The crossover strategy for the third subgroup is similar to that in the second subgroup, except that individuals in the third subgroup crossover with the gene pool of a population from another feature subset. After the crossover is completed, the fitness value of the offspring is updated according to formula (9), and the individual with the highest fitness is selected. The characteristics selected by the individual are counted and the gene pool is updated according to formulas (10) and (11).

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor; characterized in that, When the processor executes the computer program, it causes the electronic device to perform the feature selection method based on evolutionary hypergraph learning as described in any one of claims 1-8.

10. A storage medium comprising a computer program, characterized in that, When the computer program is run on an electronic device, the electronic device performs the feature selection method based on evolutionary hypergraph learning as described in any one of claims 1-8.

Images