Characteristic selection method based on graph structure and cross breeding algorithm

Abstract

The invention discloses a feature selection method based on a graph structure and a cross breeding algorithm. Industrial data to be subjected to feature optimization are obtained, the data to be selected are converted into graph structure representation, each node represents one feature, the node weight is determined by the relevance between the features and categories, the edge weight is defined by the relevance between the features, and therefore a feature weighted graph reflecting the feature relation is constructed. The feature weighted graph is grouped, strong correlation edges are divided through threshold values, features are divided into redundant features and interactive features, and feature space is divided, so that the dimension of the feature space is reduced, and key information is reserved; and searching the divided feature space by using an improved cross breeding algorithm, and searching in the feature subset space to obtain an optimal feature subset. Through combination of a graph structure and a cross breeding algorithm, effective modeling of high-dimensional complex feature association in industrial big data is realized, the precision and efficiency of feature selection are significantly improved, and the calculation cost is reduced at the same time.

Description

Technical Field

[0001] This invention relates to the field of feature selection technology, and in particular to a feature selection method based on graph structure and hybridization breeding algorithm. Background Technology

[0002] With the rapid development of the Industrial Internet and intelligent manufacturing, industrial big data is characterized by its massive volume, high dimensionality, strong heterogeneity, and complex redundancy and correlation. In the industrial field, the efficient processing of multi-source data such as equipment sensors, process parameters, and production logs directly determines the effectiveness of intelligent models for predictive maintenance, quality control, and energy consumption optimization. Against this backdrop, feature selection, as a key step in data preprocessing, aims to filter out a subset of features strongly correlated with the target variable from massive, multi-dimensional features, thereby reducing computational complexity, suppressing redundant noise interference, and improving the model's generalization ability.

[0003] Traditional feature selection methods mainly include filtering, wrapping, and embedded methods. Filtering methods select features based on their correlation with the target variable, offering high computational efficiency but neglecting feature redundancy. Wrapping methods select features by evaluating the model performance of feature subsets, capturing the effects of feature combinations, but with high computational complexity. Embedded methods embed feature selection into the model training process but are prone to getting trapped in local optima. However, features in industrial data are often not isolated but form complex relational networks. Existing methods, lacking effective modeling of these implicit relationships, may result in redundant or conflicting feature subsets, or even omission of key feature combinations, limiting the practical application effectiveness of the model. Therefore, there is an urgent need for an efficient feature selection method that can effectively model the relationships between features and balance global search and local optimization capabilities. Summary of the Invention

[0004] This invention provides a feature selection method based on graph structure and hybridization breeding algorithm, which can effectively model the relationship between features and select features while taking into account both global search and local optimization capabilities.

[0005] This invention provides a feature selection method based on graph structure and hybridization breeding algorithm, including:

[0006] The data to be selected is transformed into a graph structure, where each node represents a feature, the node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between the features.

[0007] Traverse the edge set in the graph structure, retain edges whose absolute weight is greater than or equal to a set threshold, and extract connected subgraphs from the graph structure through depth-first search. Each connected subgraph constitutes a feature cluster.

[0008] For each feature cluster, if the edge weights of all feature pairs within the cluster are less than 0 and the absolute value of the edge weights is greater than or equal to the maximum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster interactive feature; if the edge weights of all feature pairs within the cluster are greater than 0 and the absolute value of the edge weights is greater than or equal to the minimum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster redundant feature.

[0009] In each feature cluster, all intra-cluster interactive features are used to construct interactive sub-clusters, and all intra-cluster redundant features are used to construct redundant sub-clusters. The interactive sub-clusters and the redundant sub-clusters are feature sub-clusters.

[0010] The improved hybridization breeding algorithm searches all the obtained feature subclusters and initializes multiple rice individuals. In each rice individual, 1 indicates that the feature subcluster is selected and 0 indicates that the feature subcluster is not selected.

[0011] Based on the fitness value of each individual rice plant, they are divided into maintainer lines, restorer lines, and sterile lines;

[0012] For maintaining rice individuals, through the expression The newly generated maintainer rice individuals were obtained through iterative search. ;in, This indicates the individuals currently remaining in the system. Indicates the current iteration number. and These represent the lower and upper bounds of the current search space, respectively. This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. A random number between 0 and 1;

[0013] For each male-sterile rice individual, a maintainer line rice individual is randomly selected for hybridization to generate newly generated male-sterile rice individuals. ;

[0014] For restorer line rice individuals, through the expression Obtain newly generated restorer line rice individuals through iterative search ;in, This represents the globally optimal individual. This indicates the average evolutionary trend of individual restorer line rice plants. and This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. and This represents two different rice individuals randomly selected from the maintainer line;

[0015] In newly generated rice individuals, if the selected feature subcluster is a redundant subcluster, one feature from the current cluster is selected using a roulette wheel selection method; if the selected feature subcluster is an interactive subcluster, multiple features from the current cluster are selected using a roulette wheel selection method, with the number of features selected being [number missing]. ,in, This represents the total number of features contained in the k-th feature sub-cluster;

[0016] Calculate the fitness of all selected feature combinations, and select the feature combination with the highest fitness as the optimal feature subset.

[0017] Specifically, the process of transforming the data to be selected into a graph structure, where each node represents a feature, the node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between features, including:

[0018] The data to be selected is transformed into a graph structure. ;in, Represents feature nodes, Represents the edge set of all feature nodes. This represents the weight corresponding to each feature node. The relationships between feature nodes are represented by calculating the correlation between features and categories using symmetric uncertainty, which serves as the node weights of the graph structure. The specific formula is: ;in, Features With category Mutual information between them and Features and categories The entropy.

[0019] Specifically, after converting the data to be selected into a graph structure, the process also includes:

[0020] The features are sorted from largest to smallest according to the node weights of the graph structure, and the top k features are retained.

[0021] For any two features among the retained first k features and If satisfied Then the features and The edge weight between them is set to 0; if the condition is met... Through formula The correlation index FR was calculated as a feature. and The edge weights between them; where, Conditional mutual information, Features The entropy.

[0022] Specifically, the Through formula Calculated; where, This represents the maximum difference between any two feature weights. The dimension representing the feature.

[0023] Specifically, if a feature within a cluster is marked as both a redundant feature and an interactive feature, the interactive feature is retained.

[0024] Specifically, the classification of rice individuals into maintainer lines, restorer lines, and sterile lines based on their fitness values ​​includes:

[0025] Based on the fitness value of each rice individual, they are dynamically divided into maintainer lines, restorer lines, and sterile lines. Specifically, if the current iteration number t satisfies: Then the proportions of the maintenance line, restorer line, and sterile line are 20%, 50%, and 30%, respectively; if the current iteration number t satisfies: Then the proportions of the maintenance line, restorer line, and sterile line are 40%, 40%, and 20%, respectively; if the current iteration number t satisfies: The proportions of the maintenance line, restorer line, and sterile line are 60%, 20%, and 20%, respectively.

[0026] Specifically, the Through formula Calculated; where, This indicates the total number of individual rice plants in the restorer line.

[0027] Specifically, it also includes:

[0028] If the number of self-pollinations of restorer line rice individuals reaches a certain threshold but is still not updated, then through the expression... Obtain the restored rice individuals generated by the reset process. ;in, and This represents two different rice individuals randomly selected from the maintainer line, and , For individual rice fitness value, For individual rice fitness value, and These are adaptive coefficients adjusted based on fitness. , To maintain the maximum fitness difference among rice individuals in the line, .

[0029] Specifically, it also includes:

[0030] The mutation operation is performed on maintainer line rice individuals and sterile line rice individuals by randomly selecting one of three operations: addition, deletion, and replacement. The addition operation is to traverse the feature subclusters A that are not covered by the rice individual and select the feature with the highest weight in the feature subcluster A. The deletion operation is to prioritize the feature subcluster B with the most selected features and remove the feature with the lowest weight from the feature subcluster B. The replacement operation is to sort the selected features by weight, remove the feature with the lowest weight d, and select the feature with the highest weight from the feature subcluster to which feature d belongs for replacement.

[0031] Specifically, it also includes:

[0032] When adding features, if feature subcluster A is an interactive subcluster, then the suboptimal feature is further selected from feature subcluster A.

[0033] When performing a deletion operation, if all feature subclusters have the same number of selected features, the feature with the lowest weight is deleted from the selected features.

[0034] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0035] 1. Obtain the data to be selected and transform it into a graph structure representation, where each node represents a feature. The node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between features, thus constructing a feature-weighted graph reflecting feature relationships. The feature-weighted graph is grouped, and strongly correlated edges are divided using a threshold, classifying features into redundant and interactive features, thus partitioning the feature space. This reduces the dimensionality of the feature space while retaining key information. An improved hybridization breeding algorithm is used to search the partitioned feature space, finding the optimal feature subset within the feature subset space. By combining graph structure and hybridization breeding algorithm, effective modeling of complex feature associations in industrial big data is achieved, significantly improving the accuracy and efficiency of feature selection while reducing computational costs. This approach is suitable for high-dimensional feature selection scenarios such as predictive maintenance and quality control in the Industrial Internet of Things (IIoT).

[0036] 2. A graph structure is used to organize features. Not only is SU used as the initial screening criterion, but FR is also proposed to evaluate the redundancy and interaction between features. By distinguishing between strongly interactive features and strongly redundant features, redundancy problems can be avoided when selecting feature subsets, while retaining feature interactions that are helpful for classification.

[0037] 3. An optimization strategy incorporating adaptive perturbation, self-crossing, and reset mechanisms was designed. By dynamically adjusting the search range and population ratio, the algorithm's exploration and development capabilities were effectively balanced. The mutation strategy introduced three operations: addition, deletion, and replacement, which significantly improved the algorithm's global search capability and solution quality. Attached Figure Description

[0038] Figure 1 A flowchart illustrating a feature selection method based on graph structure and hybridization breeding algorithm provided in an embodiment of the present invention. Detailed Implementation

[0039] This invention provides a feature selection method based on graph structure and hybridization breeding algorithm, which can effectively model the relationship between features and perform feature selection while taking into account both global search and local optimization capabilities.

[0040] The technical solutions in the embodiments of the present invention are designed to achieve the above-mentioned technical effects, and the overall concept is as follows:

[0041] The model acquires and preprocesses multi-source heterogeneous data from industrial manufacturing processes; constructs a feature weighted graph based on mutual information, divides the feature space using dynamic thresholds, identifies redundant and interactive features, and forms feature clusters; utilizes an improved hybridization breeding algorithm to perform global and local searches on the feature clusters, and dynamically adjusts the population ratios of maintainer lines, restorer lines, and sterile lines to enhance exploration and development capabilities; optimizes the feature subset through adaptive perturbation, hybridization, self-pollination, and reset mechanisms; and finally outputs the optimal feature subset and evaluates the model's performance by introducing a mutation operator that includes addition, deletion, and replacement operations.

[0042] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0043] like Figure 1 As shown, the feature selection method based on graph structure and hybridization breeding algorithm provided in this embodiment of the invention includes:

[0044] The data to be selected is transformed into a graph structure, where each node represents a feature, the node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between the features.

[0045] In this embodiment, the data to be selected is industrial data with multi-dimensional features.

[0046] This step involves transforming the data to be selected into a graph structure. Each node in the graph represents a feature, and the node weights are determined by the correlation between the feature and the category. The edge weights are defined by the correlation between the features, including:

[0047] The selected data undergoes data cleaning, missing value imputation, standardization, and noise reduction to eliminate noise and redundant information. The processed data is then transformed into a graph structure. ;in, Represents feature nodes, Represents the edge set of all feature nodes. This represents the weight corresponding to each feature node. Representing the relationships between feature nodes, the correlation between features and categories is calculated using symmetric uncertainty, and used as node weights in the graph structure. The specific formula is: ;in, Features With category Mutual information between them and Features and categories Entropy is used to measure uncertainty. , , .

[0048] To achieve initial screening and eliminate weakly correlated feature pairs while ensuring computational efficiency, after transforming the data to be selected into a graph structure, the following steps are also included:

[0049] Sort the features according to the node weights of the graph structure from largest to smallest, and retain the top k features;

[0050] For any two features among the retained first k features and If satisfied If the two features are weakly correlated, then the feature is considered to be... and The edge weight between them is set to 0; if the condition is met... Through formula The correlation index FR was calculated as a feature. and The edge weights between them; where, Conditional mutual information indicates consideration of joining. Afterwards, characteristics The classification information that can still be provided ; Conditional entropy is used to measure a given condition. The uncertainty of the value of category C Features The entropy. If This indicates redundancy among features; the larger the value, the higher the degree of redundancy. This indicates that there is interaction between features; the smaller the value, the more information that the two features together can provide for the classification task. This indicates that the features are unrelated.

[0051] Specifically, Through formula Calculated; where, This represents the maximum difference between any two feature weights. The dimension representing the feature.

[0052] Traverse the edge set in the graph structure, keeping edge weights whose absolute values ​​are greater than or equal to a set threshold. The edges are used to extract connected subgraphs from the graph structure through depth-first search, and each connected subgraph constitutes a feature cluster.

[0053] In this embodiment, the absolute value of the edge weight is used as the first... Percentage as a set threshold .

[0054] For each feature cluster, if the edge weights of all feature pairs within the cluster are less than 0 and the absolute value of the edge weights is greater than or equal to the maximum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster interactive feature; if the edge weights of all feature pairs within the cluster are greater than 0 and the absolute value of the edge weights is greater than or equal to the minimum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster redundant feature.

[0055] Specifically, if a feature within a cluster is marked as both a redundant feature and an interactive feature, the interactive feature is retained.

[0056] In each feature cluster, all intra-cluster interactive features are used to construct interactive sub-clusters, and all intra-cluster redundant features are used to construct redundant sub-clusters. Interactive sub-clusters and redundant sub-clusters are feature sub-clusters.

[0057] An improved hybridization breeding algorithm is used to search all obtained feature subclusters. Multiple rice individuals are initialized, and within each rice individual, its gene location value is converted to 1 and 0 using the sigmoid function. Here, 1 indicates that the feature subcluster is selected, and 0 indicates that the feature subcluster is not selected.

[0058] Based on the fitness value of each individual rice plant, they are divided into maintainer lines, restorer lines, and sterile lines;

[0059] This step is explained in detail, classifying rice individuals into maintainer lines, restorer lines, and sterile lines based on their fitness values, including:

[0060] Based on the fitness value of each rice individual, they are dynamically divided into maintainer lines, restorer lines, and sterile lines. Specifically, if the current iteration number t satisfies: To conduct a broader search, the proportions of restorer lines and sterile lines need to be increased, with the proportions of maintainer lines, restorer lines, and sterile lines being 20%, 50%, and 30%, respectively. If the current iteration number t satisfies: If the convergence begins, the proportions of the maintenance line, restorer line, and sterile line will be 40%, 40%, and 20%, respectively; if the current iteration number t satisfies: To further improve the accuracy of the solution, the proportions of the maintenance line, restorer line, and sterile line are 60%, 20%, and 20%, respectively.

[0061] For maintaining rice individuals, through the expression The newly generated maintainer rice individuals were obtained through iterative search. That is, adaptive perturbation is applied to the individuals in the maintenance system, and the search range is dynamically adjusted with the number of iterations; among which, This indicates the individuals currently remaining in the system. Indicates the current iteration number. and These represent the lower and upper bounds of the current search space, respectively. This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. A random number between 0 and 1;

[0062] For each male-sterile rice individual, a maintainer line rice individual is randomly selected for hybridization to generate newly generated male-sterile rice individuals. ;

[0063] For restorer line rice individuals, through the expression Obtain newly generated restorer line rice individuals through iterative search This involves increasing population diversity and generating more high-quality solutions by drawing on evolutionary trends within the population and information from the globally optimal individual; among which, This represents the globally optimal individual. This indicates the average evolutionary trend of individual restorer line rice plants. and This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. and This represents two different rice individuals randomly selected from the maintainer line. In the early stages of the search, the differences between the individuals are relatively large, and the range of Gaussian random numbers is wide, thus enhancing the algorithm's global exploration capability. In the later stages of the search, although Gaussian random numbers are still used, the differences between the individuals gradually decrease, and the algorithm's local exploration capability is strengthened.

[0064] Specifically, Through formula Calculated; where, The total number of individuals in the restoration line rice.

[0065] Furthermore, if the number of self-pollinations of restorer line rice individuals reaches a certain threshold but is still not updated, the expression... Obtain the restored rice individuals generated by the reset process. ;in, and This represents two different rice individuals randomly selected from the maintainer line, and , For individual rice fitness value, For individual rice fitness value, and These are adaptive coefficients adjusted based on fitness. , To maintain the maximum fitness difference among rice individuals in the line, .

[0066] In newly generated rice individuals, if the selected feature subcluster is a redundant subcluster, one feature from the current cluster is selected using a roulette wheel selection method; if the selected feature subcluster is an interactive subcluster, multiple features from the current cluster are selected using a roulette wheel selection method, with the number of features selected being [number missing]. ,in, This represents the total number of features contained in the k-th feature sub-cluster;

[0067] To further enhance the global optimization capability of the embodiments of the present invention, it also includes:

[0068] The mutation operation is performed on maintainer line rice individuals and sterile line rice individuals by randomly selecting one of the three operations: addition, deletion, and replacement. The addition operation is to traverse the feature subclusters A that are not covered by the rice individual and select the feature with the highest weight in feature subcluster A. The deletion operation is to prioritize the feature subcluster B with the most selected features and remove the feature with the lowest weight from feature subcluster B. The replacement operation is to sort the selected features by weight, remove the feature with the lowest weight d, and select the feature with the highest weight from the feature subcluster to which feature d belongs for replacement.

[0069] Specifically, the decision to perform a mutation operation is made through greedy selection. The specific decision-making process is as follows:

[0070]

[0071] in, It is a random variable in the range [0, 1]. , The probability of performing addition, deletion, and replacement operations is set to the same value of 1 / 3.

[0072] Further explanation of variation manipulation techniques for maintainer line and male-sterile line rice individuals:

[0073] When adding features, if feature subcluster A is an interactive subcluster, the suboptimal feature is further selected from feature subcluster A, thereby forming a complementary enhancement effect between features.

[0074] When performing a deletion operation, if all feature subclusters have the same number of selected features, the feature with the lowest weight is deleted from the selected features, thereby achieving local optimization.

[0075] Calculate the fitness of all selected feature combinations, and select the feature combination with the highest fitness as the current optimal feature subset.

[0076] Determine if the termination condition is met. If it is, output the current optimal feature subset; otherwise, continue iterating through the process of dynamically dividing the maintainer line, restorer line, and sterile line, updating the maintainer line, restorer line, and sterile line, resetting the individual rice of the restorer line, and performing variation operations on the individual rice of the maintainer line and sterile line until the termination condition is met.

[0077] In summary, the embodiments of the present invention achieve the selection and optimization of high-dimensional features in industrial big data. While significantly reducing computing costs, it improves the accuracy and efficiency of feature selection. The embodiments of the present invention are applicable to predictive maintenance, quality control and other scenarios in the Industrial Internet of Things. They can provide diversified feature subsets for industrial manufacturing, support intelligent decision-making and optimization, and have broad application prospects and practical value.

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Any aspects of this invention not described in detail in the embodiments are well-known techniques to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention and not to limit it. Although this invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the spirit and scope of this invention, and all such modifications and substitutions should be covered within the scope of the claims of this invention.

Claims

1. A feature selection method based on graph structure and hybridization breeding algorithm, characterized in that, include: The data to be selected is transformed into a graph structure, where each node represents a feature, the node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between the features. Traverse the edge set in the graph structure, retain edges whose absolute weight is greater than or equal to a set threshold, and extract connected subgraphs from the graph structure through depth-first search. Each connected subgraph constitutes a feature cluster. For each feature cluster, if the edge weights of all feature pairs within the cluster are less than 0 and the absolute value of the edge weights is greater than or equal to the maximum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster interactive feature; if the edge weights of all feature pairs within the cluster are greater than 0 and the absolute value of the edge weights is greater than or equal to the minimum value of the feature node weights in the feature pair, the feature pair is marked as an intra-cluster redundant feature. In each feature cluster, all intra-cluster interactive features are used to construct interactive sub-clusters, and all intra-cluster redundant features are used to construct redundant sub-clusters. The interactive sub-clusters and the redundant sub-clusters are feature sub-clusters. The improved hybridization breeding algorithm searches all the obtained feature subclusters and initializes multiple rice individuals. In each rice individual, 1 indicates that the feature subcluster is selected and 0 indicates that the feature subcluster is not selected. Based on the fitness value of each individual rice plant, they are divided into maintainer lines, restorer lines, and sterile lines; For maintaining rice individuals, through the expression The newly generated maintainer rice individuals were obtained through iterative search. ;in, This indicates the individuals currently remaining in the system. Indicates the current iteration number. and These represent the lower and upper bounds of the current search space, respectively. This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. A random number between 0 and 1; For each male-sterile rice individual, a maintainer line rice individual is randomly selected for hybridization to generate newly generated male-sterile rice individuals. ; For restorer line rice individuals, through the expression Obtain newly generated restorer line rice individuals through iterative search ;in, This represents the globally optimal individual. This indicates the average evolutionary trend of individual restorer line rice plants. and This represents a random number generated by a Gaussian distribution with a mean of 0 and a variance of 1. and This represents two different rice individuals randomly selected from the maintainer line; In newly generated rice individuals, if the selected feature subcluster is a redundant subcluster, one feature from the current cluster is selected using a roulette wheel selection method; if the selected feature subcluster is an interactive subcluster, multiple features from the current cluster are selected using a roulette wheel selection method, with the number of features selected being [number missing]. ,in, This represents the total number of features contained in the k-th feature sub-cluster; Calculate the fitness of all selected feature combinations, and select the feature combination with the highest fitness as the optimal feature subset.

2. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, The process of transforming the data to be selected into a graph structure, where each node represents a feature, the node weight is determined by the correlation between the feature and the category, and the edge weight is defined by the correlation between features, including: The data to be selected is transformed into a graph structure. ;in, Represents feature nodes, Represents the edge set of all feature nodes. This represents the weight corresponding to each feature node. The relationships between feature nodes are represented by calculating the correlation between features and categories using symmetric uncertainty, which serves as the node weights of the graph structure. The specific formula is: ;in, Features With category Mutual information between them and Features and categories The entropy.

3. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 2, characterized in that, After converting the data to be selected into a graph structure, the process also includes: The features are sorted from largest to smallest according to the node weights of the graph structure, and the top k features are retained. For any two features among the retained first k features and If satisfied Then the features and The edge weight between them is set to 0; if the condition is met... Through formula The correlation index FR was calculated as a feature. and The edge weights between them; where, Conditional mutual information, Features The entropy.

4. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 3, characterized in that, The Through formula Calculated; where, This represents the maximum difference between any two feature weights. The dimension representing the feature.

5. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, If a feature within a cluster is marked as both a redundant feature and an interactive feature, retain the interactive feature.

6. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, The rice plants are classified into maintainer lines, restorer lines, and sterile lines based on the fitness value of each individual, including: Based on the fitness value of each rice individual, they are dynamically divided into maintainer lines, restorer lines, and sterile lines. Specifically, if the current iteration number t satisfies: Then the proportions of the maintenance line, restorer line, and sterile line are 20%, 50%, and 30%, respectively; if the current iteration number t satisfies: Then the proportions of the maintenance line, restorer line, and sterile line are 40%, 40%, and 20%, respectively; if the current iteration number t satisfies: The proportions of the maintenance line, restorer line, and sterile line are 60%, 20%, and 20%, respectively.

7. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, The Through formula Calculated; where, This indicates the total number of individual rice plants in the restorer line.

8. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, Also includes: If the number of self-pollinations of restorer line rice individuals reaches a certain threshold but is still not updated, then through the expression... Obtain the restored rice individuals generated by the reset process. ;in, and This represents two different rice individuals randomly selected from the maintainer line, and , For individual rice fitness value, For individual rice fitness value, and These are adaptive coefficients adjusted based on fitness. , To maintain the maximum fitness difference among rice individuals in the line, .

9. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 1, characterized in that, Also includes: The mutation operation is performed on maintainer line rice individuals and sterile line rice individuals by randomly selecting one of three operations: addition, deletion, and replacement. The addition operation is to traverse the feature subclusters A that are not covered by the rice individual and select the feature with the highest weight in the feature subcluster A. The deletion operation is to prioritize the feature subcluster B with the most selected features and remove the feature with the lowest weight from the feature subcluster B. The replacement operation is to sort the selected features by weight, remove the feature with the lowest weight d, and select the feature with the highest weight from the feature subcluster to which feature d belongs for replacement.

10. The feature selection method based on graph structure and hybridization breeding algorithm as described in claim 9, characterized in that, Also includes: When adding features, if feature subcluster A is an interactive subcluster, then the suboptimal feature is further selected from feature subcluster A. When performing a deletion operation, if all feature subclusters have the same number of selected features, the feature with the lowest weight is deleted from the selected features.

Images