Dynamic clustering and mixed agent combined assisted high-dimensional data evolution feature selection method
By combining dynamic clustering with a hybrid proxy model, the problem of balancing computational efficiency, accuracy, and feature simplification in high-dimensional data feature selection is solved, achieving efficient and accurate feature selection that is applicable to various datasets, especially high-dimensional data scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-10
AI Technical Summary
Existing high-dimensional data feature selection methods struggle to balance computational efficiency, accuracy, and feature simplification. Traditional particle swarm optimization algorithms are prone to getting trapped in local optima, static clustering strategies are prone to losing potentially effective features, and existing methods are insufficient in reducing the feature space.
A high-dimensional data evolution feature selection method jointly assisted by dynamic clustering and hybrid proxy is adopted. Strongly correlated features are clustered by FCFC strategy, weakly correlated features are integrated by filtering and ranking, a hybrid proxy model is constructed, and iterative optimization and incremental adjustment are carried out to form a complete optimization closed loop.
It achieves feature selection with high computational efficiency, high accuracy, and good feature size control in high-dimensional data scenarios, adapts to different datasets, is compatible with mainstream classifiers, and has great potential for engineering applications.
Smart Images
Figure CN121834281A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of high-dimensional data feature screening, and particularly relates to a high-dimensional data evolutionary feature selection method assisted by dynamic clustering and mixed agents. BACKGROUND
[0002] With the deep application of big data technology in the fields of biomedical, network security, image recognition, etc., the processing demand of high-dimensional data (the number of features can reach thousands or even tens of thousands) is increasingly urgent. Feature selection, as a core data preprocessing link, aims to filter out a subset of features with strong discrimination ability from the original high-dimensional feature set, thereby eliminating redundant and noisy features, reducing model calculation complexity, improving classification accuracy and generalization ability, and avoiding the problem of data sparsity caused by "dimension disaster".
[0003] Current mainstream feature selection methods are mainly divided into three categories: filter, wrapper and embedded. Filter methods select features based on inherent properties of features (such as correlation, information gain), which have high computational efficiency but ignore the interaction between features, and are prone to miss key weakly correlated features; embedded methods integrate feature selection and model training (such as L1 regularization, random forest feature importance evaluation), which balance efficiency and performance, but are strongly dependent on specific models and have limited generalization ability; wrapper methods use classification model performance as evaluation criteria to achieve feature selection by searching for the optimal feature subset, which can fully consider feature interaction and have better classification accuracy, but require repeated calls to the classifier for evaluation, which is extremely costly in high-dimensional data scenarios and difficult to be practical.
[0004] To solve the computational efficiency problem of wrapper methods, some scholars have introduced evolutionary algorithms such as particle swarm optimization (PSO) and genetic algorithm into high-dimensional feature selection, forming wrapper feature selection methods based on evolutionary algorithms. This kind of method searches for the optimal feature subset through population iteration, among which PSO algorithm is widely used due to its simplicity and fast convergence speed, but traditional PSO and its variants still have significant defects: first, the high-dimensional feature space leads to a sharp expansion of the search space, and the algorithm is prone to local optimum, such as self-adaptive particle swarm optimization (SaPSO) which improves search ability through parameter self-adaptive adjustment, but does not effectively reduce the feature space, and still has the problem of insufficient convergence efficiency in ten-thousand-dimensional feature data; second, static feature clustering strategies (such as subgroup segmentation clustering of HPSO-SSM) directly discard weakly correlated features after fixed clustering of strongly correlated features, which easily loses potential effective features and leads to reduced classification accuracy; third, the evaluation of population individuals relies on real classifiers, even if a single agent model is introduced, there is still the contradiction between insufficient prediction accuracy and limited efficiency improvement.
[0005] In addition, the existing feature selection methods generally have the problem of difficult balance between precision, efficiency and feature simplification: the filtering method has high efficiency, but the quality of the feature subset is poor; the traditional packaging method can improve the precision, but the calculation cost is high, and the number of selected features is often excessive; some optimization algorithms (such as SFE-PSO) reduce the number of features by strict screening, but at the cost of sacrificing the classification accuracy. SUMMARY
[0006] The purpose of the present application is to provide a high-dimensional data evolutionary feature selection method assisted by dynamic clustering and hybrid agent combination to solve the above technical problems.
[0007] To achieve the above purpose, the present application provides a high-dimensional data evolutionary feature selection method assisted by dynamic clustering and hybrid agent combination, comprising the following steps: S1, obtaining a high-dimensional original data set and class labels, and dividing the high-dimensional original data set into a strongly correlated feature group and a weakly correlated feature group, then clustering the strongly correlated feature group by the FCFC strategy to obtain an initial feature cluster, and sorting the weakly correlated feature group by using an integrated filtering sorting mechanism to construct an initial population; S2, based on the initial population constructed in step S1, constructing a pair comparison agent model by using a representative individual guided pairing strategy and a training sample construction strategy of feature space conversion, constructing a sample agent model by combining a representative measurement based sample screening and scale calculation rule, and forming a hybrid agent model; S3, approximating the fitness of the initial population individuals by using the sample agent model constructed in step S2, initializing the individual optimal solution and the global optimal solution, and entering the iterative optimization process, if the current iteration meets the incremental adjustment condition, then calculating the weakly correlated feature addition scale according to the iteration progress, integrating the weakly correlated features meeting the requirements into the existing feature cluster or generating a new cluster, reinitializing the population and reconstructing the hybrid agent model, and repeating until the maximum iteration number is reached, and outputting the optimal feature subset; S4, verifying the results based on the classification performance, feature scale and running time of the output optimal feature subset, and feeding back the update frequency of the optimization hybrid agent model and the incremental adjustment parameters of the feature clustering.
[0008] Therefore, the present application adopts the above-mentioned high-dimensional data evolutionary feature selection method assisted by dynamic clustering and hybrid agent combination, which has the following beneficial effects: 1. Strong strategy synergy: taking "dynamic clustering + hybrid agent" as the dual core, each sub-strategy (incremental clustering, double agent cooperation, population reinitialization) supports each other, forms a complete optimization closed loop, and solves multiple pain points of high-dimensional feature selection; 2. Wide adaptation to scenarios: can process data sets with different feature dimensions, sample sizes and class numbers, especially suitable for high-dimensional data scenarios, and has strong compatibility with mainstream classifiers, and has great potential for engineering application; 3. Performance balance is good: balance is achieved in classification accuracy, computational cost, and feature size control, and its effectiveness is verified by experiments and ablation experiments; 4. The design is specific: all key strategies are designed around the core needs of integer feature selection, from clustering, proxy models to evolutionary algorithms, adapting to problem characteristics layer by layer, avoiding the adaptability defects of general algorithms.
[0009] The technical solutions of the application will be further described in detail below with the help of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0010] Fig. 1 The flow chart of the dynamic clustering and mixed agent assisted high-dimensional data evolutionary feature selection method described in the application; Fig. 2 The feature dynamic clustering process chart of the dynamic clustering and mixed agent assisted high-dimensional data evolutionary feature selection method described in the application; Fig. 3 The construction process chart of a single training sample based on a pair comparison proxy model. DETAILED DESCRIPTION
[0011] In order to make the purpose, technical solutions and advantages of the embodiments of the application clearer and more apparent, the embodiments of the application will be further described in detail below with the help of the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the application and not to limit the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of the application. The examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout.
[0012] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.
[0013] The embodiments of the application will be described in detail below with the help of the accompanying drawings.
[0014] As shown in Figs. 1-3 The dynamic clustering and mixed agent assisted high-dimensional data evolutionary feature selection method includes the following steps: S1. Obtain the high-dimensional original dataset and class labels, and divide the high-dimensional original dataset into strongly correlated feature groups and weakly correlated feature groups. Then, use the FCFC strategy to cluster the strongly correlated feature groups to obtain the initial feature clusters. At the same time, use an integrated filtering sorting mechanism to sort the weakly correlated feature groups and construct the initial population. Step S1 specifically includes the following steps: S11. Collect high-dimensional raw dataset and corresponding class tags This constitutes a complete sample set. By analyzing the correlation between features and class labels, the original feature space is divided into strongly correlated feature groups. and weakly correlated feature groups ; S12. Use feature clustering methods to cluster strongly correlated feature groups. Clustering is performed to obtain Initial feature clusters , They represent the first Each feature class forms a feature space. ; Meanwhile, an integrated filtering ranking mechanism is used to sort weakly correlated feature groups. Sorting: Combining information gain, ReliefF algorithm, and mRMR algorithm for weakly correlated feature groups The features in the dataset are scored from multiple perspectives, and the final feature ranking result is generated by combining the median reduction method. ; S13, Based on the initial number of feature clusters An initialization strategy incorporating historical information is adopted, generating a scale of [missing information]. The encoding length is initial population Its encoded value corresponds to the index of the feature in the feature cluster.
[0015] S2. Based on the initial population constructed in step S1, a pair comparison surrogate model is constructed through a pairing strategy guided by representative individuals and a training sample construction strategy based on feature space transformation. A sample surrogate model is constructed by combining sample screening based on representativeness measure and size calculation rules to form a hybrid surrogate model. Step S2 specifically includes the following steps: S21. Construct a proxy model based on pairwise comparisons; S211. Using relative distribution density, performance volatility, and search direction consistency as indicators for evaluating the importance of multiple criteria, define the key characteristics of the multiple criteria. : ; in, ; ; ; wherein, denotes the size of the population ; and denote the relative distribution density of individual and individual , respectively (the smaller the value, the denser the area where individual is located); and denote the performance fluctuation of individual and individual , respectively (the smaller the value, the less the fluctuation of individual relative to other individuals on the fitness value of the whole population); and denote the consistency of search direction of individual and individual , respectively (the larger the value, the better the aggregation of individual in the search direction); denote the individual index; denotes the Hamming distance between individual and individual ; denotes the average fitness of all individuals in the population except individual ; denotes the fitness of individual ; denotes the angle between the search direction of individual and other individuals in the population; denotes the angle between the search direction of individual and other individuals in the population; S212, pairing strategy guided by representative individuals: according to the multi-criteria iconic individuals in the population are divided into an iconic reserve set and a non-iconic reserve set , and symmetric uncertainty is introduced to measure the similarity between individuals in the two reserve sets, and the individual pair with the smallest symmetric uncertainty is selected as the pairing sample; S213, constructing training samples by feature space conversion: converting the integer encoding of the pair of individuals to the feature space, if the features are the same after conversion, the corresponding dimension value is 0; if the feature is only selected by the dominant individual, the value is 1, only selected by the inferior individual, the value is -1; the feature corresponding to the dimension value is 0, which is not selected by any individual, and binary labels are given to the training samples (if the dominant individual is superior to the inferior individual, the label is 1, otherwise 0); S214, training the SVM base model based on the training samples, and obtaining the pair comparison agent model; S22, constructing a sample agent model; S221, using a clustering algorithm to divide the complete sample set into clusters, wherein the value is equal to the total number of categories of the complete sample set ; S222, calculating the sample representativeness measure value : ; In the formula, represents the number of categories contained in the sample cluster output by step S221; represents the value of the center sample of the th category in the sample cluster on the th feature; represents the value of the sample on the th feature; represents the feature dimension; S223, determining the sample agent model size : ; In the formula, represents the number of samples of the complete sample set ; and respectively represent the preset lower limit and upper limit of the proportion coefficient hyperparameter; and both belong to , and can be set according to the preference of the decision maker. In the embodiment, the is set to 0.5, and the is set to 0.1.
[0016] S23, integrating the pair comparison agent model and the sample agent model to form a hybrid agent model.
[0017] S3, approximate evaluation of the fitness of the initial population individuals using the sample agent model constructed in step S2, initialization of the individual optimal solution and the global optimal solution, and entering an iterative optimization process; if the current iteration meets the incremental adjustment condition, calculating the weakly correlated feature addition scale according to the iteration progress, integrating the weakly correlated features meeting the requirements into the existing feature clusters or generating new clusters, reinitializing the population and reconstructing the hybrid agent model, and repeating until the maximum number of iterations is reached, and outputting the optimal feature subset; Step S3 specifically includes the following steps: S31, initialization stage; Approximate evaluation of the fitness of each individual in the initial population using the sample agent model in the hybrid agent model; Initialization of the individual optimal solution and the global optimal solution , wherein is initialized as the individual itself, and is initialized as the individual with the highest fitness value in the initial population; S32, iterative optimization stage; S321, initialization of the iteration loop: setting the current iteration number , the maximum number of iterations , the iteration interval number threshold , the feature subset similarity threshold , and the progressive evaluation proportion coefficient ; S322, judging whether the feature incremental adjustment condition is met and ; if yes, triggering the incremental adjustment described in step S323, otherwise executing step S324; wherein denotes the optimal feature subset corresponding to the optimal individual of the th generation population and the Jaccard coefficient of the th generation optimal feature subset , which is expressed as follows: ; In the formula, and respectively represent the number of elements in the intersection and the union of and . The larger the value is, the higher the degree of similarity between the current optimal feature subset and the previous optimal feature subset is.
[0018] S323, calculating the feature scale that needs to be added from the weakly correlated feature group to the current feature clustering result: ; In the formula, Indicates the first The scale of the added features; For the first The number of iterations of the evolutionary algorithm corresponding to the adjustment of feature clustering results; For the first After the second adjustment, the weakly correlated feature group The number of remaining features; and and The values are respectively 0 and weakly correlated feature groups. The initial size of the features included in the clustering is shown. It can be seen that when adjusting the feature clustering results, the size of the added features, the number of remaining features in the current weakly correlated feature group, and the current iteration number are adjusted in tandem.
[0019] S324, Update feature clusters; S3241. Determine the representative feature set: from the current global optimal solution In this process, representative features of each existing feature cluster are extracted to form a set of representative features. ,and , Indicates the current number of clusters; S3242, Traverse the features to be added: from Before the election Features And iterate through the added features and the set of representative features. The first in Features correlation ; S3243, based on Value pairs represent the set of features The features in the dataset are sorted in descending order and stored in a set. The corresponding correlation values are stored in a set. ; S3244, Cluster allocation judgment: S32441, Traversal ,set up The Middle Features The feature clusters that were previously included in the feature clustering results are as follows: ; S32442, Calculation With class tags correlation as well as With class tags correlation ; S32443, if Then the features Into In the middle, mark Execute step S325; otherwise, proceed by feature Generate a new feature cluster and label it. ; S325. Based on the updated feature clusters, reinitialize the population. And reconstruct the hybrid agent model; S326. Use the pairwise comparison surrogate model in the reconstructed hybrid surrogate model to predict the current population individual and the current global optimum. The relationship between the superiority and inferiority of the solutions is analyzed, and the prediction performance is improved to be better than the current global optimum. Individuals identified as dominant individuals are identified until the traversal is complete, and all dominant individuals form a dynamic set of high-quality candidate solutions. ; S327. Using the sample proxy model in the reconstructed hybrid proxy model for approximate evaluation. Fitness value of an individual; S328, Select High fitness ranking Individuals are stored in a high-quality subset. middle; S329, Utilizing a complete sample set Evaluation of the trained classification model Fitness value of an individual; S3210, Update and ; Update individual optimal solution If the approximate fitness value of an individual's new location is greater than its current fitness value... If the fitness value is lower, or the fitness values are equal but the number of features is lower, then update. ; Update the global optimal solution From the updated Select the individual with the largest approximate fitness value and calculate its true fitness value. If it is better than the current... If the number of features is even smaller, then update; S3211, Implement the hybrid agent model management strategy; S32111, Update the pairwise comparison surrogate model: After the population update, from New training samples are generated by random pairing, and the pairwise comparison surrogate model is retrained when the cumulative number of evaluations of the true fitness reaches a preset value. S32112, Update the sample proxy model: Calculate relative prediction error : ; In the formula, Indicates the first The global extreme point obtained in the next iteration; and They represent The predicted fitness value and the actual fitness value; when When the value is greater than the set threshold, the current sample proxy model... Imputation: Randomly select misclassified samples from the remaining samples. Add to And the upper and lower limits of the filled sample size are respectively and ,in Set as , Set as The number of category labels included; S3212. The positions of all individuals in the population are updated using the integer particle swarm optimization rule to obtain the optimal feature subset.
[0020] In S32112, if Less than Then randomly select from the remaining sample set. One sample was added to ;like Greater than Then from Random deletion One sample.
[0021] The steps described in step S3210 During the update process, for those not selected For individuals, a fitness value prediction strategy based on distance-weighted interpolation is used to approximate the fitness value. The specific steps are as follows: First, calculate the individual to be evaluated. and The Middle Advantage reserve set at the next iteration Each individual in Euclidean distance between Subsequently, based on Euclidean distance... Sort each individual in ascending order and select the top-ranked one. Using each individual as a reference, the weighted average fitness value of the reference individuals is calculated as the value of the individual to be evaluated. The approximate fitness value is calculated using the following formula: ; In the formula, Indicates weight, and , Represents a constant (a small constant to prevent division by zero); Indicates the concentration of advantageous reserves. Individual; Indicates the individual to be evaluated Approximate fitness value; Dynamic adjustment The value balance prediction capability and computational cost are calculated using the following formula: .
[0022] S4. Based on the classification performance, feature size, and runtime verification results of the optimal feature subset, feedback is used to optimize the update frequency of the hybrid proxy model and the incremental adjustment parameters of feature clustering.
[0023] Simulation Experiment Datasets: To verify the effectiveness of the DCCDSA-PSO described in this invention, it was applied to 18 typical datasets. These datasets cover different sample sizes, feature dimensions, and number of classes (60-1687 samples, 301-15154 features, 2-10 classes), exhibiting high diversity and enabling a comprehensive evaluation of the algorithm's effectiveness and applicability. To reduce bias in the feature selection process, a double five-fold cross-validation strategy is employed: the outer loop divides the dataset into five mutually exclusive subsets, which are used sequentially as the test set, with the remainder serving as the training set; the inner loop further divides the training set into five subsets during the feature selection search process to evaluate the fitness values of individuals.
[0024] Table 1. Basic information about the dataset
[0025] Experiment 1: Exploring the dynamic feature clustering strategy and progressive evaluation ratio coefficient The effect of different values on the performance of DCCDSA-PSO.
[0026] In this experiment, The values are {0.2, 0.4, 0.6, 0.8}. The values are {10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%}. To evaluate the classification performance of DCCDSA-PSO under different values, the GFE01, MFD, Colon, Lung, and Leukemia_big datasets were selected for testing. These datasets cover different application domains and have varying numbers of features, effectively reflecting the algorithm's performance in diverse scenarios.
[0027] Table 2. Results of parameter sensitivity analysis
[0028] As shown in Table 2, when the parameter and At %, DCCDSA-PSO achieved the highest Ac value on these datasets. In summary, it is recommended to adjust the parameters... and Set them to 0.8 and 50% respectively. Experiment 2 verifies the effectiveness of the proposed dynamic feature clustering strategy.
[0029] In this experiment, the dynamic clustering optimization strategy in DCCDSA-PSO was replaced with a static feature clustering strategy as a comparison method, and the algorithm was labeled as DCCDSA-PSO-D.
[0030] Table 3 Comparison between DCCDSA-PSO and DCCDSA-PSO-D (average classification accuracy) ), the average number of selected features ( ), and average runtime ( ))
[0031] As shown in Table 3: (1) In terms of classification accuracy, DCCDSA-PSO achieved the highest accuracy on 17 datasets. Value. Specifically, for CNS, compared to DCCDSA-PSO-D, CSA-PSO achieves [value]. The value increased by 13.08%. (2) In terms of the average number of selected features, CSA-PSO achieved smaller values for semeion, Prostate, Colon, SRBCT, Drive_Face, and Ovarian. Value. The same was obtained on Orlraws10P. Value. Although DCCDSA-PSO-D achieved slightly smaller values on the remaining datasets. Value, but the present invention obtains The value is significantly lower than that of DCCDSA-PSO. (3) In terms of runtime, DCCDSA-PSO achieved the lowest values in Prostate, DLBCL, CNS, and Ovarian. The value was relatively large on the remaining 14 datasets. values. This is because the incremental feature clustering guided by the correlation is performed in the evolutionary process, the structure and number of feature clusters change, the search space increases, and the computational cost of the algorithm increases. DCCDSA-PSO-D obtains smaller values on most of the data sets, however, only obtains smaller values on QSAR. This strategy improves the classification performance of DCCDSA-PSO while keeping similar computational cost.
[0032] Experiment 3 verifies the effectiveness of the proposed hybrid agent model.
[0033] In the evolutionary process, the individual is evaluated based on the complete training sample set instead of the approximate evaluation of the hybrid agent model in DCCDSA-PSO as a variant algorithm, and it is named as DCCDSA-PSO-S.
[0034] Table 4 Comparison of DCCDSA-PSO and DCCDSA-PSO-S The unit: minute
[0035] From Table 4, it can be seen that: (1) In terms of classification accuracy, DCCDSA-PSO obtains higher values on 11 data sets. Obtains slightly lower values on the remaining data sets, because the classification accuracy is relatively high due to the true evaluation of the individual evaluation of DCCDSA-PSO-S. (2) In terms of the number of selected features, DCCDSA-PSO obtains smaller values on 11 data sets. Although DCCDSA-PSO-S obtains smaller values on the remaining 7 data sets, the values obtained by DCCDSA-PSO-S are relatively low on most of the data sets. Especially for Ovarian, the value of DCCDSA-PSO increases by 4.71% compared with DCCDSA-PSO-S in the case of similar values. (3) In terms of running time, DCCDSA-PSO obtains smaller values on 14 data sets. This is because the strategy uses the hybrid agent model to approximate the evaluation of the individual, which greatly reduces the computational cost of the algorithm compared with the true evaluation of DCCDSA-PSO-S. In summary, on most of the data sets, DCCDSA-PSO improves the classification performance of the algorithm while significantly reducing the computational cost of the algorithm.
[0036] Experiment 4 evaluates the average classification accuracy performance on the 15 datasets.
[0037] In the experiment, the classification accuracy was used to compare the proposed DCCDSA-PSO with the 6 comparative algorithms shown in Table 5. In addition, the test with the significance level of 0.5 was used to compare the statistical difference between DCCDSA-PSO and each comparative algorithm.
[0038] Table 5 Parameter settings
[0039] Table 6 The classification accuracy (%) obtained by DCCDSA-PSO and the 6 comparative algorithms (“+” means that DCCDSA-PSO is significantly better than the corresponding comparative algorithm, “=” means that DCCDSA-PSO has no significant difference with the corresponding comparative algorithm, and “-” means that the corresponding comparative algorithm is significantly better than DCCDSA-PSO.)
[0040] From Table 6, it can be seen that DCCDSA-PSO obtained the highest classification accuracy on the 15 datasets, which were 94.67%, 83.68%, 89.89%, 98.18%, 98.10%, 81.82%, 96.33%, 99.07%, 95.46%, 94.62%, 71.30%, 98.40%, 96.03%, 100.0% and 95.24%, respectively. Further, from the results of the test, it can be seen that: (1) Compared with SFE-PSO, DCCDSA-PSO obtained significantly better classification accuracy on 17 and 1 datasets, respectively, and had no significant difference in the classification accuracy on 1 dataset. (2) Compared with SaPSO and HPSO-SSM, DCCDSA-PSO obtained significantly better classification accuracy on 14 datasets. (3) Except for QSAR, WarpAR10P, Driv_Face and Orlraws10P 4 datasets, DCCDSA-PSO obtained significantly better classification accuracy on the remaining datasets than HFS-C-P. (4) On Prostate, WarpAR10P, CNS, Lung 4 datasets, DCCDSA-PSO was significantly better than PSOEMT. (5) On 5 datasets, DCCDSA-PSO showed very significant results in classification accuracy compared with MSa-HEFS. In summary, DCCDSA-PSO can identify high-quality feature subsets on most datasets and show significant competitive advantages.
[0041] Experiment 5: Evaluation of the average number of selected features compared to conventional methods (…) Performance on Table 7 shows the results obtained from DCCDSA-PSO and 6 comparison algorithms. value
[0042] From Table 7, we can conclude that: (1) DCCDSA-PSO achieves the minimum values on semeion and Colon. Value. (2) For all datasets, DCCDSA-PSO achieved the third smallest ranking. value, The algorithm is represented on all datasets. The average value. Although SFE-PSO achieved the minimum. Values, but as shown in Table 6, obtained on all datasets. The values are all lower than those of DCCDSA-PSO. The main reason is that SFE-PSO removes a large number of features in the first stage, causing the subsequent algorithm to search in the reduced feature space, thus selecting the fewest features and losing some key information. (3) Compared with SaPSO, HPSO-SSM, PSOPT and MSa-HEFS, DCCDSA-PSO achieves significantly smaller values. As shown in Table 6, CSA-PSO also achieved good results on most datasets. In summary, DCCDSA-PSO effectively controls feature size while maintaining high classification accuracy, fully demonstrating its superiority.
[0043] Experiment 6 evaluates the average runtime of the proposed method compared to traditional methods. Performance on Table 8 shows the results obtained from DCCDSA-PSO and 6 comparison algorithms. value( This indicates the rank of the values obtained by each algorithm on the same dataset. "" indicates that a certain algorithm achieves the following results on all datasets. (The average of the ranking values)
[0044] As shown in Table 8, (1) except for QSAR, MicroMass, and Leukemia_big, DCCDSA-PSO achieved the minimum [value] on the remaining 15 datasets. (2) On all datasets, DCCDSA-PSO achieved the smallest value compared to all comparison algorithms. values. (3) Compared with SaPSO, HPSO-SSM and PSO-EMT, DCCDSA-PSO obtains significantly smaller values. DCCDSA-PSO obtains larger values than MSa-HEFS, but its obtained values are 2.54% higher than those of MSa-HEFS. MSa-HEFS obtains values close to those of DCCDSA-PSO, because the regression proxy model adopted by MSa-HEFS can significantly reduce the computational cost of evaluating solutions in the evolutionary process. At the same time, although the ranking of the indicator values obtained by HFS-C-P is relatively high, its values are still significantly higher than those of DCCDSA-PSO. The reason is that HFS-C-P still relies on the complete training sample set to evaluate the fitness value of individuals, which inevitably increases the computational overhead. In contrast, DCCDSA-PSO combines the dynamic clustering technique with the double proxy model auxiliary strategy, not only significantly reduces the computational cost, but also shows strong competitiveness in terms of classification performance, becoming an efficient FS algorithm.
[0045] In summary, the simulation experiment compares the algorithm proposed in the present application with SaPSO, HPSO-SSM, PSO-EMT, HFS-C-P, SFE-PSO and MSa-HEFS, six evolutionary feature selection methods, on 18 data sets. The results show that the proposed algorithm achieves better classification performance with lower computational cost on most data sets, especially showing stronger advantages in high-dimensional data. In addition, the effectiveness of each key strategy on the overall performance is further verified through two groups of ablation experiments, thereby proving the effectiveness of the present application.
[0046] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and not to limit them. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.
Claims
1. A dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method, characterized in that: The method comprises the following steps: S1, obtaining a high-dimensional original data set and class labels, dividing the high-dimensional original data set into a strongly correlated feature group and a weakly correlated feature group, clustering the strongly correlated feature group by an FCFC strategy to obtain an initial feature cluster, and sorting the weakly correlated feature group by using an integrated filter sorting mechanism to construct an initial population; S2, based on the initial population constructed in step S1, constructing a pair comparison agent model by using a representative individual guided pairing strategy and a training sample construction strategy based on feature space conversion, constructing a sample agent model by combining sample screening based on a representative measure and scale calculation rules, and forming a hybrid agent model; S3, approximating the fitness of individuals in the initial population by using the sample agent model constructed in step S2, initializing the optimal solution of individuals and the global optimal solution, and entering an iterative optimization process; if the current iteration meets the incremental adjustment condition, the weakly correlated feature addition scale is calculated according to the iteration progress, the weakly correlated features meeting the requirements are integrated into the existing feature cluster or a new cluster is generated, the population is reinitialized and the hybrid agent model is reconstructed, and the cycle is repeated until the maximum iteration number is reached, and the optimal feature subset is output; S4, verifying the results based on the classification performance, feature scale and running time of the output optimal feature subset, and feeding back the update frequency of the optimized hybrid agent model and the incremental adjustment parameters of feature clustering.
2. The dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method according to claim 1, characterized in that: Step S1 specifically comprises the following steps: S11, collecting a high-dimensional original data set and corresponding class labels , to constitute a complete sample set , through correlation analysis of features and class labels, dividing the original feature space into a strong correlation feature group and a weak correlation feature group ; S12, using a feature clustering method to cluster the strongly correlated feature groups to obtain initial feature clusters , respectively represent the feature classes, forming a feature space ; Meanwhile, an integrated filtering ranking mechanism is used for weakly correlated feature groups. Sorting: Combining information gain, ReliefF algorithm, and mRMR algorithm for weakly correlated feature groups The features in the dataset are scored from multiple perspectives, and the final feature ranking result is generated by combining the median reduction method. ; S13, Based on the initial number of feature clusters An initialization strategy incorporating historical information is adopted, generating a scale of [missing information]. The encoding length is initial population Its encoded value corresponds to the index of the feature in the feature cluster.
3. The dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method of claim 2, wherein: Step S2 specifically comprises the following steps: S21, constructing a pair comparison agent model; S211. Using relative distribution density, performance volatility, and search direction consistency as indicators for evaluating the importance of multiple criteria, define the key characteristics of the multiple criteria. : ； Wherein, ； ； ； In the formula, represents the size of the population; and respectively represent the relative distribution density of individuals and individuals ; and respectively represent the performance fluctuation of individuals and individuals ; and respectively represent the search direction consistency of individuals and individuals ; all represent the individual index; represents the Hamming distance between individuals and individuals ; represents the mean fitness of all individuals in the population except for individual ; represents the fitness of individual ; represents the included angle between the search direction of individual and other individuals in the population; respectively represent the included angle between the search direction of individual and other individuals in the population; S212, Representative Individual Guided Pairing Strategy: According to Multi-criteria Landmark Divide individuals in population into landmarked reserve set and non-landmarked reserve set and introduce symmetric uncertainty to measure similarity between individuals in two reserve sets, select individual pair with minimum symmetric uncertainty as pairing sample; S213, constructing training samples by feature space conversion: converting the integer coding of the pair of individuals into the feature space, if the converted features are the same, the corresponding dimension value is 0; if the feature is only selected by the dominant individual, the value is 1, only selected by the inferior individual, the value is -1; the feature not selected by any individual corresponds to the dimension value 0, and the binary label is assigned to the training sample; S214, training the SVM base model based on the training sample to obtain the pair comparison agent model; S22, constructing a sample agent model; S221. Use clustering algorithms to analyze the complete sample set. Divided into Clusters, of which The value is equal to the complete sample set. Total number of categories; S222, calculating a sample representativeness measure : ； In the formula, represents the number of categories contained in the sample cluster output by step S221; represents the center sample of the th category in the sample cluster on the th feature; represents the value of the sample on the th feature; represents the feature dimension; S223, determining a sample agent model size : ； wherein represents the number of samples of the complete sample set ; and represent the preset lower and upper limits of the scale factor hyperparameters, respectively. S23, integrating the pair comparison agent model and the sample agent model to form a hybrid agent model.
4. The dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method of claim 3, wherein: Step S3 specifically comprises the following steps: S31, initialization stage; The fitness value of each individual in the initial population is approximated by using the sample agent model in the hybrid agent model; Initialize individual best solution and global best solution wherein is initialized to each individual itself, is initialized to the individual with the highest fitness value in the initial population; S32, iterative optimization stage; S321, iteration loop initialization: set the current iteration number , maximum iteration number , iteration interval number threshold , feature subset similarity threshold , and progressive evaluation proportion coefficient ; S322, judging whether a feature increment adjustment condition is satisfied and If yes, triggering the increment adjustment in step S323, otherwise executing step S324; wherein, denotes the optimal feature subset corresponding to the optimal individual of the first generation population denotes the optimal feature subset corresponding to the optimal individual of the first generation population denotes the Jaccard coefficient of the optimal feature subset of the first generation and the optimal feature subset of the second generation denotes the Jaccard coefficient of the optimal feature subset of the first generation and the optimal feature subset of the second generation denotes the Jaccard coefficient of the optimal feature subset of the first generation and the optimal feature subset of the second generation ； S323、Calculate the current feature clustering result needs to add features from the weakly related feature group Feature scale: ； In the formula, indicates the feature scale of the th addition; is the number of iterations of the evolutionary algorithm corresponding to the adjustment of the feature clustering result; th adjustment; is the number of remaining features in the weakly correlated feature group th adjustment; th adjustment. S324, updating the feature cluster; S3241、determine representative feature set: extract representative feature of each existing feature cluster to form set representative feature set from current global optimal solution represents the current cluster number S3242, Traverse the features to be added: from Before the election Features And iterate through the added features and the set of representative features. The first in Features correlation ; S3243、based on value pairs representing the set of features features in descending order and store to the set corresponding relevance values are stored to the set ; S3244, cluster allocation judgment: S32441, traversing , set the first characteristics of the feature cluster result before the feature cluster is adjusted as ; S32442, calculation correlation with class labels correlation with class labels and correlation with class labels correlation with class labels ; S32443, if Then the features Into In the middle, mark Execute step S325; otherwise, proceed by feature Generate a new feature cluster and label it. ; S325, re-initialize the population based on the updated feature cluster and reconstruct the hybrid proxy model; S326、using the pair-wise comparison in the reconstructed hybrid surrogate model to predict the current population individuals and the current global optimal solution the prediction performance is better than the current global optimal solution the individual with the best prediction performance is determined as the dominant individual until the end of the traversal, and all dominant individuals constitute a dynamic high-quality candidate solution set ; S327, approximating evaluation using the sample agent model in the reconfigured hybrid agent model the fitness value of the individual S328, selecting top-ranked fitness values individuals to the elite subset values; S329, utilizing a classification model trained based on a complete sample set evaluation of a trained classification model fitness value of the individual; S3210, updating and ; updating the individual best solution if the approximate fitness value of the new position of the individual is greater than the fitness value of the current , or the fitness value is equal but the number of features is smaller, then updating ; Update the global optimal solution From the updated Select the individual with the largest approximate fitness value and calculate its true fitness value. If it is better than the current... If the number of features is even smaller, then update; S3211, executing a hybrid agent model management strategy; S32111, updating the pairwise comparison agent model: after the population is updated, new training samples are generated by randomly pairing the population in the cumulative evaluation of the real fitness reaches a preset value, the pairwise comparison agent model is retrained. S32112, updating the sample agent model: calculating the relative prediction error : ； wherein denotes the global extremum point obtained at the th iteration; and denote the predicted fitness value and the real fitness value of the th individual, respectively; When the value is greater than a set threshold, the current sample agent model is filled: randomly screening the classification error samples from the remaining samples to supplement to , and the upper and lower limits of the filling sample size are and , wherein is set as , is set as the number of category labels contained S3212, updating the positions of all individuals in the population by using the integer particle swarm optimization rule to obtain the optimal feature subset.
5. The dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method of claim 4, wherein: In S32112, if is less than , then samples are randomly selected from the remaining sample set and added to ; if is greater than , then samples are randomly removed from .
6. The dynamic clustering and hybrid agent assisted high-dimensional data evolutionary feature selection method of claim 4, wherein: The step S3210 is described During the updating process, for the individuals not selected into the population, the fitness value of each individual is predicted by using the distance-weighted interpolation-based fitness value approximation calculation strategy, and the specific steps are as follows: First, the individual to be evaluated is calculated With the first iteration of the advantage reserve set Each individual in The Euclidean distance between ; then, according to the Euclidean distance, each individual in Is sorted in ascending order, and the top Individuals are selected as reference objects, and the weighted average fitness value of the reference objects is calculated as the approximate fitness value of the individual to be evaluated , The formula is as follows: ； In the formula, Indicates weight, and , Represents a constant; Indicates the concentration of advantageous reserves. Individual; Indicates the individual to be evaluated Approximate fitness value; Dynamic adjustment The value balances the prediction capability and the calculation cost, and the calculation formula is as follows: 。