High-dimensional evolutionary feature selection method fusing filtering and packaging strategies

Abstract

The invention relates to the field of artificial intelligence, particularly discloses a high-dimensional evolutionary feature selection method fusing filtering and packaging strategies, and aims at high-dimensional, small-sample and unbalanced complex data to realize minimum discriminant feature subset screening through a multi-stage hierarchical search framework from coarse to fine and from global to local. The method comprises the steps of generating a high-quality initial population by adopting core-marginal probability sampling and preferentially entering evolution based on double filtering type integrated score sorting of symmetric uncertainty SU and ReliefF; an environment-aware unimproved counter is introduced in the global search stage, the feature adding probability and the feature removing number are dynamically adjusted to balance exploration and development, and repeated evaluation is reduced in combination with fitness cache; and when search stops, binary particle swarm optimization neighborhood refinement is triggered, local high-strength discrete search is realized by Sigmoid mapping, and global optimum is backfilled and updated. Experiments prove that the method obtains a high F1-score and a significant compression feature number on a plurality of high-dimensional data sets.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies. Background Technology

[0002] In the current era of big data and artificial intelligence, with the rapid advancements in high-throughput sequencing technology, high-resolution imaging equipment, and internet information collection methods, the scale of data accumulated across various industries is experiencing explosive growth. This data is not only characterized by its large sample size, but more significantly by its extremely high feature dimensionality. For example, in genomics research, a single sample often contains tens of thousands of gene expression level features; in facial recognition tasks, image pixel features or extracted deep features can easily number in the tens of thousands. However, not all features among these massive amounts of data are beneficial for subsequent learning tasks (such as disease diagnosis and classification, object recognition, etc.). Instead, the data is filled with a large number of redundant features (features are highly correlated and contain repetitive information), irrelevant features (features that have no discriminatory power against the target variable), and even noisy features (which interfere with model learning). The existence of these ineffective features has led to the famous "curse of dimensionality" problem.

[0003] The "curse of dimensionality" not only significantly increases the computational complexity of machine learning algorithms during training and prediction, leading to unbearable storage and time costs, but more seriously, it sparsifies data distribution, making it difficult for models to capture the true structure of data in high-dimensional space. This introduces excessive interference information, severely reducing the model's generalization performance and causing overfitting. Therefore, how to effectively remove redundant and irrelevant features from high-dimensional data and select a low-dimensional feature subset that maximizes the retention of discriminative information from the original data—i.e., feature selection—has become a core problem urgently needing to be solved in the fields of machine learning and data preprocessing. Existing feature selection methods are generally divided into three categories based on evaluation criteria and their integration with learning algorithms: filtering, encapsulation, and embedded methods. Filtering methods mainly rely on the statistical properties of the data itself (such as information gain, correlation coefficient, mutual information, etc.) to evaluate the importance of features. Their advantages are high computational efficiency and independence from specific classifiers. However, because they ignore the interdependencies between features and the interaction between features and the classifier, they often fail to obtain the optimal feature subset. While embedded methods integrate feature selection into model training (e.g., L1 regularization, decision tree feature importance), resulting in higher efficiency, their results are highly dependent on specific model structures, exhibiting poor generality and a tendency to overfit on high-dimensional, small-sample data. Encapsulated methods directly use the predictive performance of a specific learning algorithm as the evaluation criterion for feature subsets, theoretically achieving the best classification accuracy. However, due to the need for frequent classifier training and evaluation, their computational cost is extremely high. Especially in high-dimensional spaces, as the feature dimension increases, the search space grows exponentially. Traditional full search or heuristic search strategies (e.g., sequential forward / backward selection) easily get trapped in local optima, making it difficult to find the globally optimal feature subset within an acceptable timeframe.

[0004] In recent years, evolutionary computation-based feature selection methods have become a research hotspot for solving high-dimensional feature selection problems due to their powerful global search capabilities and independence from problem gradient information. Genetic algorithms (GA), particle swarm optimization, and ant colony optimization have been widely applied in feature selection. However, directly applying these evolutionary algorithms to high-dimensional feature selection still faces significant challenges: First, the search range in high-dimensional space is extremely wide, and traditional random initialization strategies often result in low initial population quality, requiring the algorithm to consume a significant amount of time for ineffective searches in the early stages, leading to slow convergence. Second, existing algorithms lack real-time awareness of the search environment, typically using fixed parameters to control search behavior, making it difficult to achieve a dynamic balance between "global exploration" and "local development," and easily leading to premature convergence to local optima. Third, single evolutionary operators often lack the ability to refine local modifications in the later stages of the search, making it difficult to further fine-tune and optimize the discovered subset of superior features. To address these issues, there is an urgent need to develop a novel hybrid high-dimensional evolutionary feature selection method that effectively combines the advantages of filtering and encapsulation methods and possesses adaptive search capabilities. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies. This method utilizes evolutionary computation theory, integrates complementary filtering feature evaluation indicators with encapsulation search strategies, and employs an environment-aware mechanism for adaptive global optimization and binary particle swarm neighborhood refinement search. It constructs a multi-stage hierarchical search framework that progresses from coarse to fine and from global to local, thereby addressing the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies includes the following steps: Step 1: The original features are evaluated from multiple perspectives and sorted according to symmetric uncertainty and ReliefF. The sorting results are divided into core feature set and marginal feature set according to the initial important feature ratio parameter. Subpopulations are generated by core-margin probability sampling and then merged to form a temporary population. The temporary population is sorted according to the fitness function and the initial population that enters the evolutionary process is selected based on the best results.

[0007] Step 2: Perform an environment-aware adaptive global search on the initial population. For each individual in the feature subset, generate a random number and compare it with the feature addition probability parameter. If the random number is less than the feature addition probability parameter, perform a feature addition operation on the feature subset; otherwise, perform a feature removal operation on the feature subset. During fitness evaluation, use a fitness cache with a hash table structure to avoid duplicate evaluations. Monitor the global optimal solution update state through a non-improvement counter. After each iteration, adaptively adjust the feature addition probability parameter and the feature removal quantity parameter based on the value of the non-improvement counter. The feature removal quantity parameter is updated using an exponential decay strategy with each iteration.

[0008] Step 3: When the global search stalls, a neighborhood refinement stage based on binary particle swarm optimization is triggered. A local feature subspace is constructed with the current global optimal solution as the center, and particle swarm co-evolution search is performed in the subspace. The particle velocity is mapped to the feature selection probability using the Sigmoid function to obtain a locally better feature subset, and the locally better feature subset is mapped back to the original feature space to update the global optimal solution.

[0009] As a further aspect of this invention, step 1 involves evaluating and ranking the original features from multiple perspectives based on symmetric uncertainty and ReliefF. The ranking results are then divided into core feature sets and marginal feature sets according to the initial important feature ratio parameter. Subpopulations are generated using core-margin probability sampling and merged to form a temporary population. The temporary populations are then ranked based on a fitness function, and the best-performing initial population is selected to enter the evolutionary process. This includes the following specific content: To address the problem of blind searching in high-dimensional space, a population initialization strategy based on dual-filter ensemble is designed. This strategy does not directly perform random sampling in the full-dimensional space but introduces two complementary filter evaluation metrics—symmetric uncertainty (SU) and the ReliefF algorithm. Symmetric uncertainty, based on information theory, excels at capturing the nonlinear correlation between features and categories; the ReliefF algorithm, based on distance metrics, excels at capturing local interactions and discriminative abilities between features. By utilizing symmetric uncertainty and the ReliefF algorithm, the original features are pre-screened and scored. Based on the ranking results, a "core-edge" probability sampling mechanism is proposed. This mechanism divides the features into a "core feature set" (high ranking) and a "edge feature set" (low ranking), and assigns different selection probabilities (e.g., 80% probability of selecting the core and 20% probability of selecting the edge). This generates a high-quality initial population that contains both highly important features and maintains a certain degree of diversity, effectively compressing the search space of subsequent evolutionary algorithms and providing an excellent starting point for the algorithm.

[0010] As a further aspect of this invention, step 2 involves performing an environment-aware adaptive global search on the initial population. For each feature subset individual, a random number is generated and compared with a feature addition probability parameter. If the random number is less than the feature addition probability parameter, a feature addition operation is performed on the feature subset; otherwise, a feature removal operation is performed on the feature subset. A fitness cache with a hash table structure is used during fitness evaluation to avoid duplicate evaluations. The global optimal solution update state is monitored by a non-improvement counter. After each iteration, the feature addition probability parameter and the feature removal quantity parameter are adaptively adjusted based on the value of the non-improvement counter. The feature removal quantity parameter is updated using an exponential decay strategy with each iteration generation. This includes the following specific content: In the main evolutionary cycle, this invention proposes an environment-aware adaptive global search strategy. The adaptive global search strategy no longer relies on a fixed crossover mutation probability but introduces a "non-improvement counter" (…). This allows for real-time monitoring of the global optimal solution's status.

[0011] Based on the search stalling situation, the adaptive global search strategy dynamically adjusts the probability and magnitude of "feature addition" and "feature removal" operations. Specifically, when the adaptive global search strategy gets stuck in a local optimum (the number of consecutive unimproved attempts exceeds the limit), it automatically increases the "feature addition probability" (…). This enhances the exploration efforts; when the adaptive global search strategy is in the optimization phase, the search range is gradually narrowed. Simultaneously, regarding the "number of features removed" (…), the search scope is increased. The adaptive global search strategy employs an exponential decay strategy, significantly removing redundancy in the initial stage and fine-tuning it later. This mechanism enables the adaptive global search strategy to intelligently switch between wide-area exploration and deep exploration based on the current search environment, significantly improving search efficiency and the ability to escape local optima.

[0012] As a further aspect of this invention, step 3, when the global search stagnates, triggers a neighborhood refinement stage based on binary particle swarm optimization (BPSO). A local feature subspace is constructed centered on the current global optimum, and particle swarm co-evolutionary search is performed within this subspace. The Sigmoid function is used to map particle velocities to feature selection probabilities to obtain a locally better subset of features. This locally better subset of features is then mapped back to the original feature space to update the global optimum. This includes the following specific content: To address the difficulty in improving accuracy in the later stages of conventional evolutionary algorithms, this invention triggers a neighborhood refinement strategy based on binary particle swarm optimization (BPSO) when the global search stagnates. The neighborhood refinement strategy constructs a reduced local neighborhood search space centered on the current global optimum, and performs a high-intensity refined search within this space through particle swarm co-evolution. By mapping particle velocities to feature selection probabilities using the Sigmoid function, the particle swarm repeatedly verifies feature combinations within the local space, uncovers potential better solutions, further eliminates redundant features, and improves classification accuracy.

[0013] The technical effects and advantages of this invention's high-dimensional evolutionary feature selection method, which integrates filtering and encapsulation strategies, are as follows: This invention utilizes a filtering ensemble initialization that combines SU and ReliefF strategies, fully leveraging the global nonlinear correlation and local discriminative power of features to provide a high-quality search starting point for the algorithm. Combined with an environment-aware adaptive global search strategy, the algorithm can flexibly switch between "exploration" and "development," effectively avoiding getting trapped in local optima. Experiments demonstrate that on multiple high-dimensional medical datasets, the method of this invention achieves nearly 100% classification accuracy, significantly outperforming existing mainstream algorithms. While maintaining high accuracy, this invention greatly compresses the size of the feature subset. The average feature compression rate exceeds 60%, reducing thousands of dimensions to tens of dimensions; this not only significantly reduces the computational overhead of subsequent classification tasks but also greatly improves the interpretability of the model, contributing to the discovery of truly critical pathogenic genes or biomarkers in fields such as biomedicine. The "environment-aware" mechanism introduced in this invention allows the algorithm to move beyond fixed parameter settings and dynamically adjust its strategy based on real-time feedback during the search process (i.e., whether it is improving the solution). Combined with a local search strategy, this mechanism endows the algorithm with strong robustness, enabling it to adapt to high-dimensional datasets with different distributions, dimensions, and sample sizes, exhibiting stable optimization performance. Although this invention employs an encapsulation strategy, it effectively controls computational costs through multiple optimization measures. Filtering initialization quickly eliminates some extremely irrelevant features and narrows the initial search range; the fitness caching mechanism avoids repeated evaluation of the same feature subsets; and the local search strategy is limited to a very small subspace. These designs enable this invention to efficiently process large-scale high-dimensional data within an acceptable timeframe. Attached Figure Description

[0014] Figure 1 This is a flowchart of a high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies according to the present invention.

[0015] Figure 2 For the present invention in different A diagram illustrating the F1-score values ​​obtained under the given conditions. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0017] like Figure 1 As shown, the present invention provides a high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies, comprising the following steps: Step 1: The original features are evaluated from multiple perspectives and sorted according to symmetric uncertainty and ReliefF. The sorting results are divided into core feature set and marginal feature set according to the initial important feature ratio parameter. Subpopulations are generated by core-margin probability sampling and then merged to form a temporary population. The temporary population is sorted according to the fitness function and the initial population that enters the evolutionary process is selected based on the best results.

[0018] Step 2: Perform an environment-aware adaptive global search on the initial population. For each individual in the feature subset, generate a random number and compare it with the feature addition probability parameter. If the random number is less than the feature addition probability parameter, perform a feature addition operation on the feature subset; otherwise, perform a feature removal operation on the feature subset. During fitness evaluation, use a fitness cache with a hash table structure to avoid duplicate evaluations. Monitor the global optimal solution update state through a non-improvement counter. After each iteration, adaptively adjust the feature addition probability parameter and the feature removal quantity parameter based on the value of the non-improvement counter. The feature removal quantity parameter is updated using an exponential decay strategy with each iteration.

[0019] Step 3: When the global search stalls, a neighborhood refinement stage based on binary particle swarm optimization is triggered. A local feature subspace is constructed with the current global optimal solution as the center, and particle swarm co-evolution search is performed in the subspace. The particle velocity is mapped to the feature selection probability using the Sigmoid function to obtain a locally better feature subset, and the locally better feature subset is mapped back to the original feature space to update the global optimal solution.

[0020] Further, in step 1, the original features are evaluated from multiple perspectives and ranked based on symmetric uncertainty and ReliefF. The ranked results are then divided into core feature sets and edge feature sets according to the initial important feature ratio parameter. Subpopulations are generated using core-edge probability sampling and merged to form a temporary population. The temporary populations are ranked based on the fitness function, and the best are selected to enter the initial population for the evolutionary process. This includes: In high-dimensional feature selection tasks, if features of thousands or tens of thousands of dimensions are directly and randomly initialized, the generated population will be extremely sparse and of low quality, resulting in extremely slow algorithm convergence. Therefore, a filtered ensemble initialization is first performed. This process includes three sub-steps: feature multi-perspective evaluation, core-edge partitioning, and probability sampling to generate the population.

[0021] Step 11, multi-perspective feature evaluation, selecting two complementary filtering algorithms: Symmetric Uncertainty (SU) and ReliefF. First, calculate the SU score for all features (SU is an information-theoretic metric used to measure the non-linear correlation between features and category labels. It is based on the concepts of information entropy and mutual information, effectively identifying features with strong dependencies on the target variable, while being insensitive to the specific distribution of the features), and then sort the features according to their SU scores from highest to lowest to obtain the feature sequence. Then, the ReliefF weights of all features are calculated (ReliefF is a distance-based metric that calculates a feature's ability to distinguish between different classes by finding the nearest neighbors of a sample within the same class and the nearest neighbors of samples in different classes; ReliefF is more sensitive to the interactions between features and local data structures, and can capture feature weights that are difficult to detect with a single statistic), and the features are sorted from highest to lowest weight to obtain the feature sequence. .

[0022] Step 12, core-edge segmentation: To guide the population to concentrate in high-importance areas while retaining some exploratory ability, this embodiment sets an initial selection of important feature ratio parameters. (Based on experimental experience, the optimal settings are selected) (for 0.1); for the Sequence, extract the top-ranked items The features are defined as the "SU core feature set"; the remaining features are defined as the "SU edge feature set"; for The sequence, also extracting the top ranks The features are defined as the "ReliefF core feature set"; the remaining features are defined as the "ReliefF edge feature set".

[0023] Step 13 involves probabilistic sampling to generate a population. This method abandons simple truncation selection (i.e., selecting only the first N elements) and instead employs a probabilistic sampling mechanism to maintain diversity. Firstly, based on... The sequence randomly extracts features from the "SU core feature set" with a relatively high probability (preferably 80%), and simultaneously randomly extracts features from the "SU edge feature set" with a relatively low probability (preferably 20%), combining them into an individual. This process is repeated multiple times to generate a subpopulation. ;based on The sequences, using the same 80% / 20% probability sampling strategy, generate subpopulations. ;Will and The individuals are merged to form a larger temporary population. The fitness value of each individual in the population on the training set is calculated (in this embodiment, the F1-score obtained by combining the KNN classifier with 5-fold cross-validation is used as the fitness function). The merged population is then sorted in descending order based on the fitness values, and the top-performing individuals are selected. Each individual serves as the initial population that ultimately enters the evolutionary process. .

[0024] Further, in step 2, an environment-aware adaptive global search is performed on the initial population. For each individual in the feature subset, a random number is generated and compared with the feature addition probability parameter. If the random number is less than the feature addition probability parameter, a feature addition operation is performed on the feature subset; otherwise, a feature removal operation is performed on the feature subset. A fitness cache with a hash table structure is used during fitness evaluation to avoid duplicate evaluations. The global optimal solution update state is monitored by a non-improvement counter. After each iteration, the feature addition probability parameter and the feature removal quantity parameter are adaptively adjusted based on the value of the non-improvement counter. The feature removal quantity parameter is updated using an exponential decay strategy with each iteration. This includes: after obtaining the initial population, entering the iterative evolutionary global search phase. To address the problem of high-dimensional space searches easily getting trapped in local optima, this embodiment introduces an environment-aware mechanism. By monitoring the update state of the global optimal solution (i.e., the environment), the search strategy is adaptively adjusted. This includes three core components: individual evolution operations, fitness evaluation and cache update, and adaptive parameter adjustment.

[0025] Step 21, Individual Evolution Operations: This invention defines two basic search operators to update each individual (feature subset) in the population; for each individual... Generate a random number between [0, 1]. And add probability parameters to it along with the current features. Compare: If If it has never been added to the current individual, then the add operation will be performed. Randomly select one feature from the selected feature set. Add to this individual to generate a new individual. This operation aims to broaden the search scope and find key features that may have been missed; otherwise, a removal operation is performed, removing the individual from the current search. From the already selected feature set, the number of random selections is... feature set Eliminate and generate new individuals. This operation aims to remove redundancy and streamline the feature subset.

[0026] Step 22, Fitness Evaluation and Cache Update: To save computational resources for high-dimensional data, this invention introduces a fitness cache with a hash table structure. ), in assessing new individuals First, check if its feature combination already exists. If the fitness value exists, directly read the corresponding fitness value; if it does not exist, call the fitness evaluation function (such as performing 5-fold cross-validation on the training set for a k-NN classifier), calculate its F1-score, and store the result. Comparing new individuals With the original individual The fitness, if If a better solution is found, update the individuals in the population; check if the global optimal solution has been updated. If a better one is found than the current one A better solution will be updated. And will "not improve the counter" Reset to 0; otherwise, Add 1.

[0027] Step 23, adaptive parameter tuning, aims to address the poor adaptability problem caused by static parameters in evolutionary algorithms; after each iteration, according to... Value dynamically adjusted and Set a threshold for the number of times improvement will not occur. ( For example, 10), when When the search stalls, the algorithm determines that it has fallen into a local optimum. At this point, the algorithm increases... ,in For example, in 1.1, forcibly increasing the probability of "feature addition" prompts the algorithm to attempt to introduce new features to escape the current predicament; when When the search is normal, the algorithm reduces ,in For example, 0.9, and set a lower limit (such as 0.1) to gradually stabilize the search behavior; adopt an exponential decay strategy. As the iteration number iter increases, According to the formula:

[0028] Update, (of which) The attenuation factor is, for example, 0.99. Remove the initial value of the number of features to randomly select the number of features. The design (which removes the minimum number of randomly selected features during exponential decay) ensures that in the early stages of the search, the algorithm can remove a large number of redundant features at once (i.e., "large step back"), achieving rapid dimensionality reduction; as iterations proceed, The convergence to the minimum value allows for minor feature pruning in the later stages of the search.

[0029] Further, in step 3, when the global search stagnates, a neighborhood refinement stage based on binary particle swarm optimization is triggered. A local feature subspace is constructed centered on the current global optimum, and particle swarm co-evolutionary search is performed within this subspace. The Sigmoid function is used to map particle velocities to feature selection probabilities to obtain a locally better feature subset. This locally better feature subset is then mapped back to the original feature space to update the global optimum. This includes the following specific content: Although the global search stage can approximate the optimal solution through an adaptive strategy, in high-dimensional spaces, there are often still better microstructures near the global optimum. To further explore its potential, this embodiment designs a neighborhood refinement stage when the global search stagnates (i.e., when the global search stagnates...). Reaching a specific threshold It is triggered when ).

[0030] Step 31: Construct a local neighborhood search space, which is based on the current global optimal solution. Built around the core, the specific approach is to... All features included are used as a basis; if If the number of features in the original feature set is too small (resulting in insufficient search space), some unselected features are randomly selected from the original feature set to supplement it, forming a feature subspace that is much smaller than the original dimension but contains key information; all local search operations will be strictly restricted to this feature subspace.

[0031] Step 32, Binary Particle Swarm Optimization (BPSO): Within the feature subspace, the algorithm initializes a small-scale particle swarm (e.g., 10 particles). Each particle's position vector is a binary string with a length equal to the number of features in the subspace; each bit represents whether the corresponding feature is selected (1 for selected, 0 for unselected). The particle swarm is iteratively updated according to the BPSO equations of motion, with each particle updating its position based on its historical best position (individual extreme value). The velocity is updated using the group's historical best position (global extreme value G). Since position is a discrete binary value, a sigmoid function is introduced as a transfer function to map velocity values ​​to probability values. The probability that a particle takes a value of 1 in a certain dimension is determined by... Decision, if random number If the feature is selected (position set to 1), it is not selected (position set to 0); repeat the above process until the maximum number of iterations of the local search is reached. And update the local optimal solution in real time.

[0032] Step 33: After the local search is completed, check whether the feature subset corresponding to the globally optimal position found by the local search is better than the feature subset before entering the local search. If so, then map this better subset back to the original high-dimensional feature space, replacing the original one. Then, the algorithm updates the corresponding optimal fitness value. Afterward, the algorithm returns to the global search main loop and resets... Then, proceed to the next round of adaptive global search.

[0033] In this embodiment, a comparative experiment was designed to comprehensively and objectively verify the effectiveness of the method proposed in this invention: I. The experiment selected 12 representative high-dimensional benchmark datasets, covering face recognition (ORL), gene expression analysis (Colon, Lung, Lymphoma, GLIOMA, TOX-171, Prostate-GE, Leukemia, ALLAML, Carcinom), and other high-dimensional classification tasks. These datasets exhibit a wide range of feature dimensions, from 1024 (ORL) to 9182 (Carcinom), while the sample size is relatively small (e.g., GLIOMA has only 50 samples but 4434 features), representing a typical high-dimensional, small-sample "curse of dimensionality" scenario. Four current mainstream and advanced evolutionary feature selection algorithms were selected as benchmarks: PSO-EMT based on multi-task learning, NOCSA based on crow search, BEGJO based on leopard optimization, and FWPSO (Feature-Weighted Particle Swarm Optimization). The population size for all algorithms was set to 10, and the maximum number of iterations was set to 400. The unique parameters of this invention, FW-HEFS, are set as follows: The evaluation metrics used were the F1-score (for class imbalance) and the number of features. Each experiment was run independently 30 times and the average value was taken.

[0034] II. Analysis of Experimental Results 1. Classification performance comparison (F1-score): The experimental results in Table 1 show that the FW-HEFS of this invention achieved the highest F1-score on the vast majority of datasets, demonstrating excellent classification performance.

[0035] Table 1. F1-scores (%) of FW-HEFS and comparison algorithms As shown in Table 1, on the Lung dataset, FW-HEFS achieved an F1 score of 97.74%, significantly higher than NOCSA's 77.67% and FWPSO's 93.81%; on the Lymphoma dataset, FW-HEFS reached 99.46%; and on the Leukemia and WarpIE10P datasets, it achieved a perfect classification result of 100%. Statistics show that FW-HEFS has an average F1 score as high as 96.98%. Compared to PSO-EMT, the performance improvement is approximately 4.1%; compared to BEGJO, approximately 3.4%; compared to FWPSO, approximately 7.3%; and compared to NOCSA, a remarkable 17.9% improvement. This demonstrates that the dual-filter initialization and adaptive search strategy can effectively guide the algorithm to find more discriminative feature combinations.

[0036] 2. Comparison of feature dimensionality reduction capabilities (number of features): As shown in Table 2, the present invention performs particularly well in feature compression, and can significantly reduce the number of features while maintaining high accuracy.

[0037] Table 2 Number of Selected Features in FW-HEFS and Comparison Algorithms As shown in Table 2, on the Carcinom dataset (9182 dimensions), FW-HEFS selected only 95.2 features, achieving a compression rate exceeding 99%; on the Colon dataset (2000 dimensions), FW-HEFS selected only 23.8 features. The overall average number of features selected by FW-HEFS is approximately 40.2. In comparison, PSO-EMT selects an average of 130.0 features, NOCSA an average of 107.8 features, and FWPSO an average of 80.9 features. FW-HEFS reduces the number of features by approximately 69% compared to PSO-EMT, 63% compared to NOCSA, and 50% compared to FWPSO. This is attributed to the adaptive feature removal mechanism and refined pruning of local search, effectively removing a large number of redundant features.

[0038] 3. Key characteristic scaling parameters for the initialization phase. The experiment analyzed the impact of variations from 0.1 to 0.6 on performance. For example... Figure 2 The data trend shown indicates that when When set to 0.1, the algorithm achieved the highest F1-score on the vast majority of datasets. This is because a smaller [value]... This value ensures that the initial population focuses on high-quality core features, reduces noise interference in the initial stage, and provides a better starting point for subsequent searches.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0040] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies, characterized in that, Includes the following steps: Step 1: Based on symmetric uncertainty and ReliefF, the original features are evaluated from multiple perspectives and sorted separately. The sorting results are divided into core feature set and marginal feature set according to the initial important feature ratio parameter. Subpopulations are generated by core-margin probability sampling and then merged to form a temporary population. The temporary population is sorted based on the fitness function and the initial population that enters the evolution process is selected based on the best results. Step 2: Perform an environment-aware adaptive global search on the initial population. For each feature subset individual, generate a random number and compare the random number with the feature addition probability parameter. Monitor the global optimal solution update state through a non-improvement counter. After each iteration, adaptively adjust the feature addition probability parameter and the feature removal quantity parameter according to the value of the non-improvement counter. The feature removal quantity parameter is updated with an exponential decay strategy as the number of iterations increases. Step 3: When the global search stalls, a neighborhood refinement stage based on binary particle swarm optimization is triggered. A local feature subspace is constructed with the current global optimal solution as the center, and particle swarm co-evolution search is performed in the subspace. The particle velocity is mapped to the feature selection probability using the Sigmoid function to obtain a locally better feature subset, and the locally better feature subset is mapped back to the original feature space to update the global optimal solution.

2. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that... The adaptive adjustment of the feature addition probability parameter and the feature removal quantity parameter based on the value of the non-improvement counter includes: preset a non-improvement count threshold; when the non-improvement counter is greater than or equal to the non-improvement count threshold, the search stalls; at this time, the feature addition probability parameter is increased to forcibly increase the probability of feature addition, so that the algorithm can introduce new features to break out of the current predicament; when the non-improvement counter is less than the non-improvement count threshold, the search is normal, and the feature addition probability parameter is decreased.

3. The high-dimensional evolutionary feature selection method according to claim 2, characterized in that... The adaptive adjustment of the feature removal quantity parameter includes: adopting an exponential decay strategy to update according to the formula as the number of iterations increases. In the early stage of the search, redundant features can be removed at once to reduce dimensionality. As the iteration progresses, the number of features removed converges to the minimum value.

4. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that... The comparison of the random number with the feature addition probability parameter specifically includes: if the random number is less than the feature addition probability parameter, performing a feature addition operation on the feature subset; otherwise, performing a feature removal operation on the feature subset, and using a fitness cache with a hash table structure to avoid duplicate evaluation during fitness evaluation.

5. The high-dimensional evolutionary feature selection method according to claim 4, characterized in that... The fitness cache is a hash table structure. Before evaluating a new individual, it checks whether its feature combination already exists in the fitness cache. If it exists, the corresponding fitness value is read directly. If it does not exist, the fitness value is calculated and written to the fitness cache.

6. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that, In the core-edge probability sampling, the probability of extracting features from the core feature set of the symmetric uncertainty is 80%, and the probability of extracting features from the edge feature set of the symmetric uncertainty is 20%. The same probability sampling is used to generate corresponding subpopulations for the ReliefF core feature set and the ReliefF edge feature set.

7. The high-dimensional evolutionary feature selection method that integrates filtering and encapsulation strategies according to claim 1, characterized in that, The non-improvement counter is reset to 0 when a solution better than the current global optimum is found, and incremented by 1 when no improvement is found. The value of the non-improvement counter triggers adaptive adjustment of the search strategy.

8. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that, The neighborhood refinement stage constructs a local feature subspace based on the features contained in the current global optimal solution. When the number of features contained in the global optimal solution is insufficient, resulting in insufficient search space, unselected features are randomly selected from the original feature set to supplement the local feature subspace.

9. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that, In the adaptive global search with environmental awareness, after the fitness evaluation of each newly generated feature subset individual is completed, its fitness is compared with the fitness of the corresponding original individual, and the original individual is replaced by the new individual to update the population when the new individual has better fitness.

10. The high-dimensional evolutionary feature selection method according to claim 1, characterized in that, During the neighborhood refinement phase, the velocity of each particle is updated based on its historical best position and the group's historical best position.

Images