Text feature selection and classification method and system based on feature cluster subspace evolution

Abstract

The application discloses a text feature selection and classification method and system based on feature cluster subspace evolution. The method first performs vectorization and normalization processing on the text corpus, and constructs a word co-occurrence relationship matrix; then initializes a population containing a three-layer structure based on the co-occurrence relationship, divides the feature clusters according to the feature cluster division vector, and calculates the weighted center vector; the cluster attribution is updated through the distance between the word and the cluster center to realize subspace reconstruction, the feature weight is adaptively updated according to the normalized distance, the weight is converted into the selection probability in the cluster, and the candidate feature subset is obtained through Bernoulli sampling; the discriminant scatter fitness function is constructed according to the inter-class separation degree and the intra-class compactness to evaluate the subset, the optimal feature subset is obtained by driving the population iteration, and finally the text feature is reconstructed by using the optimal feature subset, and the classification model is trained to complete the classification. The application can retain the text semantic structure, reduce feature redundancy, and improve the text classification precision and stability.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data mining technology, specifically to a method and system for text feature selection and classification based on feature cluster subspace evolution, belonging to the direction of high-dimensional data feature modeling and intelligent optimization computing technology. Background Technology

[0002] In text classification tasks, it is often necessary to select a subset of discriminative features from a high-dimensional term feature space to improve the performance of the classification model and reduce computational complexity. To address this issue, various feature selection methods have been proposed in the prior art, mainly including filtering methods, wrapper methods, and embedded methods.

[0003] Text feature selection typically employs statistically based filtering methods, such as feature ranking based on information gain, chi-square test, or term frequency-inverse document frequency. These methods evaluate individual terms independently to filter features. However, these methods rely solely on the statistical relationship between a single feature and the category label, neglecting the semantic connections between terms. This can easily lead to the removal of semantically relevant features with low statistical contributions, thereby disrupting the overall semantic expression of the text and causing a decline in classification model performance.

[0004] Furthermore, for optimizing high-dimensional text feature spaces, there is also a wrapper approach that uses feature subsets as search objects and evaluates them in conjunction with the performance of classification models to obtain better feature combinations. However, this type of method usually performs a combination search directly in the original high-dimensional space without modeling the structural relationships between features. When the feature dimension is high, it is prone to the problem of search space explosion, which leads to a significant increase in computational cost. At the same time, the search process is prone to getting trapped in local optima, thus affecting the quality of the final feature subset.

[0005] Furthermore, in some existing patented technologies, regarding the problem of feature correlation modeling, patent document CN106570178A proposes a high-dimensional text data feature selection method based on graph clustering. This method constructs a weighted undirected graph and uses a community detection algorithm to cluster features, then removes redundant features according to the "maximum correlation, minimum redundancy" principle. While this method utilizes the graph structure relationship between features to some extent, its clustering process is sensitive to threshold parameters, and the removal of redundant features mainly relies on statistical correlation within local clusters, lacking a comprehensive understanding of global category discrimination information. This may lead to the accidental deletion of some features with discriminative complementarity.

[0006] Regarding the joint optimization of feature grouping and classifier parameters, patent document CN108363810A proposes a method combining Adaptive Particle Swarm Optimization (APSO) and clustering. First, APSO is used to optimize the initial K-means cluster centers to achieve feature clustering. Then, APSO is used to optimize the support vector machine parameters. This method introduces adaptive mechanisms in both the clustering and classification stages. However, feature selection mainly relies on the distance between cluster centers, failing to fully consider the distinguishing ability of features between different categories. Furthermore, APSO has a high iterative computational cost in high-dimensional, large-scale data, affecting its efficiency in practical applications.

[0007] Furthermore, regarding the problem of linking feature grouping and selection, patent document CN110147810A proposes a class-aware feature selection framework. This framework calculates the class relevance score and class discriminative score between features and classes, assigns features to the most relevant class clusters, and then selects a subset of features proportionally. This method considers both intra-class relevance and inter-class discriminative power of features, making it suitable for imbalanced datasets. However, its feature allocation depends on the quality of the initial class cluster partitioning, and the strategy of selecting features proportionally to class size may compress the space of strongly discriminative features in smaller classes.

[0008] In summary, existing text feature selection methods generally suffer from the following problems in high-dimensional sparse data scenarios: First, the lack of unified modeling of the semantic relationship structure between features leads to the destruction of semantic consistency in the feature selection process; second, the search process is not constrained by structural information, resulting in high computational complexity and easy getting trapped in local optima in high-dimensional space; third, the lack of an effective linkage mechanism between feature grouping and feature selection makes it difficult to form a stable and discriminative feature subspace structure. These problems are particularly prominent in large-scale text classification tasks. Summary of the Invention

[0009] The purpose of this application is to provide a method and system for text feature selection and classification based on feature cluster subspace evolution, which can improve the accuracy and stability of text classification.

[0010] Firstly, this application provides a text feature selection and classification method based on feature cluster subspace evolution, including: Data preprocessing: The text corpus is segmented, vectorized, and normalized to construct a term co-occurrence relation matrix; Population initialization: The population is initialized based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. Among them, the feature cluster partitioning vector is used to record the feature cluster number to which each term belongs, so as to realize the semantic grouping of the text feature space; The feature weight vector is used to quantify the importance of each term in the corresponding feature cluster, providing a continuous score basis for feature selection; The feature selection state vector is used to identify whether each term is selected in binary state, in order to generate a subset of candidate features that can be used for classification; Feature cluster subspace construction: Divide the terms into several feature clusters based on the feature cluster partition vector, and calculate the weighted center vector of each feature cluster; Feature cluster subspace evolution: Based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, update the feature cluster number to which the term belongs, so as to complete the feature cluster partitioning vector update; The feature weight vector is adaptively updated based on the normalized distance between the feature vector of a term and the weighted center vector of its feature cluster. A feature selection state vector is generated based on the updated feature weight vector, and a subset of candidate features is obtained based on the feature selection state vector, where each feature corresponds to a term. Fitness evaluation: A discriminant divergence fitness function is constructed using inter-class separation and intra-class compactness to evaluate candidate feature subsets and obtain the corresponding fitness values; Iterative optimization: The population is driven to evolve iteratively based on fitness values ​​to obtain the optimal feature subset; Classification execution: The text feature representation is reconstructed using the optimal feature subset, the classification model is trained, and the text classification is completed.

[0011] In one possible implementation, constructing the term co-occurrence relation matrix includes: representing the text corpus data as a text feature matrix. ,in for The Middle Feature vectors of each term, Each term's feature vector corresponds to the statistical distribution of that term across all document samples. For the number of document samples, For the number of terms; Normalization is performed to obtain the normalized text feature matrix. ,in For the first Normalized feature vectors of each term; according to the formula Calculate the term co-occurrence matrix , of which Line 1 Column elements Indicates the first The term and the first The co-occurrence intensity of each term across all documents.

[0012] In one possible implementation, initializing the population based on the term co-occurrence relation matrix includes: The initial clustering result is obtained based on the term co-occurrence relation matrix. The initial clustering result is obtained by performing similarity modeling on the term co-occurrence relation matrix and performing clustering analysis on it based on the co-occurrence strength between terms, so that terms form several feature clusters according to their statistical association structure in the initial stage. Based on the term co-occurrence relation matrix, the feature cluster partitioning vector is initialized, and a perturbation mechanism is introduced on this basis, as shown in the formula: ; in, Indicates the first The first generation (initial generation) of the population The individual The feature cluster number to which each term belongs; , The total number of individuals in the population; Indicates the first The initial clustering results of each term, Indicates the probability of cluster perturbation. Indicates the interval A uniformly random integer within the range represents a randomly assigned cluster number. Indicates the total number of feature clusters; Through the above perturbation mechanism, while preserving the initial cluster structure based on term co-occurrence relationships, a certain proportion of random cluster allocation is introduced, so that different individuals have differentiated cluster partitioning results in the initial stage, thereby improving the structural diversity of the population, expanding the search range of subsequent feature subspaces, and reducing the risk of getting trapped in local optima during the evolution process.

[0013] The feature weight vector is initialized as follows: each element in the vector represents the weight value of a term, and all elements are initialized to the value specified in the original text. Independent random numbers that follow a uniform distribution within the interval; The initialization formula for the feature selection state vector is: ; in, This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise; where Indicates the first The first generation (initial generation) of the population Among the individuals, the first The selection status identifier for each term is determined by the weight. With threshold The comparison yields the result if Greater than , , indicates that it is selected, otherwise it is not. , indicates that it was not selected; among them, Indicates the first The first generation of the population The individual The weight value of each term.

[0014] In subsequent evolution, the feature selection state vector can be updated by Bernoulli random sampling based on the intra-cluster normalized probability.

[0015] In one possible implementation, the formula for calculating the weighted center vector of each feature cluster is: ; in, Indicates the first The first generation of the population The first individual The weighted center vector of each feature cluster; Indicates the first The first generation of the population The individual The weight value of each term; Indicates the first The first generation of the population The first individual The set of selected terms within each feature cluster. , Indicates the first The first generation of the population The first individual A feature cluster, To indicate the first The first generation of the population A subset of candidate features for each individual Complete set of features ; When the selected feature does not exist within a certain feature cluster, the cluster center remains unchanged or is reset using a preset strategy.

[0016] In one possible implementation, the cluster affiliation of the updated feature is determined by the following formula: ; in, Indicates the first The first generation of the population The individual The feature cluster number to which each term belongs. This represents the distance metric function, which can be Euclidean distance or other similarity metrics.

[0017] In one possible implementation, the adaptive update of the feature weight vector is expressed by the following formula: ; ; in, Indicates the first The first generation of the population The individual Normalized distance of each term; and They represent the first The first generation of the population The first individual The and the first The weighted center vector of each feature cluster; For the first Normalized feature vectors of each term; Indicates the first The first generation of the population The first individual A feature cluster; Indicates the first The first generation of the population The individual The weight values ​​of each term. The normalized distance is used to characterize the degree of deviation of the term from the cluster center; the weight update is implemented through an exponential mapping function, so that the feature weights monotonically decrease as the normalized distance increases.

[0018] In one possible implementation, the inter-class separation degree is calculated using the following formula: ; in, Indicates the first The first generation of the population Inter-class separability of an individual, and Numbers indicating different categories, and They represent in Upper Class and First The mean feature vector of all document samples in the class; The intra-class compactness is calculated using the following formula: ; ; in, Indicates the first The first generation of the population Intra-class compactness of each individual For the first A collection of indexes for all document samples of the class. ; Indicates the first A document sample in Normalized eigenvectors on, Indicates in Upper The mean feature vector of all document samples in the class; wherein, the feature vector of a document sample on the candidate feature subset is composed of the statistical representation of the document sample on the terms in the candidate feature subset; The fitness function is: ; in, Indicates the first The first generation of the population The fitness value of each individual To prevent constants with zero denominators, this fitness function establishes a ratio between inter-class separation and intra-class compactness, providing a unified measure of the discriminative power of feature subsets. This fitness function serves as a surrogate measure of classification performance, allowing evaluation of the discriminative power of feature subsets without the need to train a classification model.

[0019] In one possible implementation, generating a feature selection state vector based on the updated feature weight vector includes: converting the updated feature weight vector into intra-cluster selection probabilities, performing Bernoulli sampling based on the intra-cluster normalized selection probabilities, and generating the feature selection state vector.

[0020] In one possible implementation, the classification model uses a support vector machine, a naive Bayesian model, or a neural network, taking the feature vector reconstructed from the optimal feature subset of the document as input to complete the text classification.

[0021] Secondly, this application provides a text feature selection classification system based on feature cluster subspace evolution, including: The data preprocessing module is used to perform word segmentation, vectorization, and normalization on the text corpus, and to construct a word co-occurrence relation matrix; The population initialization module is used to initialize the population based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. The feature cluster subspace construction module is used to divide the term into several feature clusters based on the feature cluster partitioning vector and calculate the weighted center vector of each feature cluster. The feature cluster subspace evolution module is used to update the feature cluster number to which the term belongs based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, so as to complete the feature cluster partitioning vector update; based on the normalized distance between the feature vector of the term and the weighted center vector of the feature cluster to which it belongs, the feature weight vector is adaptively updated; based on the updated feature weight vector, a feature selection state vector is generated, and a candidate feature subset is obtained based on the feature selection state vector. The fitness evaluation module is used to construct a discriminant divergence fitness function based on inter-class separation and intra-class compactness, evaluate candidate feature subsets, and obtain the corresponding fitness values. The iterative optimization module is used to drive the iterative evolution of the population according to the fitness value to obtain the optimal feature subset; The classification execution module is used to reconstruct the text feature representation using the optimal feature subset, train the classification model, and complete the text classification.

[0022] The system is used to implement the above method.

[0023] Thirdly, this application provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is used to invoke the computer program to execute the method described above.

[0024] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed on an electronic device, causes the electronic device to perform the method described above.

[0025] Fifthly, this application provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.

[0026] The specific implementation methods of the second to fourth aspects of this application can refer to the implementation methods of the first aspect, and will not be elaborated here.

[0027] Beneficial effects:

[0028] (1) This application can preserve the semantic structure of the text and avoid the destruction of semantic consistency. In view of the shortcomings of existing methods that ignore the semantic association of terms and only evaluate a single feature independently, this application constructs a feature cluster subspace based on the co-occurrence relationship of terms, and changes the feature selection from "independent screening" to "subspace collaborative optimization". During the screening process, the semantic structure of the text is fully preserved, and features that are semantically related but have low statistical contribution will not be deleted by mistake, thus ensuring the semantic integrity of the feature expression.

[0029] (2) This application can reduce the search complexity in high-dimensional space and avoid local optima. In view of the shortcomings of existing methods that directly search in the original high-dimensional space, are prone to space explosion and local optima, this application decomposes the original high-dimensional space into multiple structured subspaces by feature cluster partitioning, and carries out constrained search in the subspaces, which significantly reduces the computational complexity. At the same time, global optimization is driven by population evolution, which effectively avoids getting trapped in local optima and improves the quality of feature subsets.

[0030] (3) This application can achieve effective linkage between feature grouping and selection, thereby improving discriminability and stability. In view of the shortcomings of existing methods, such as the disconnect between feature grouping and feature selection and the difficulty in forming a stable discriminative subspace, this application couples and links the three-layer structure of cluster partitioning, weight update and feature selection, adaptively updates the cluster affiliation and feature weights, and uses the discriminative divergence function of inter-class separation degree + intra-class compactness for global evaluation. The resulting optimal feature subset has low redundancy, strong discriminative ability and high classification stability, which can significantly improve the accuracy and generalization performance of text classification.

[0031] (4) By using the precise center calculation mechanism based on the selected feature subset, the noise interference of unselected redundant features on the subspace center representation is effectively eliminated, ensuring that the feature cluster center is always accurately focused on the high discriminative region during the evolution process, forming a self-reinforcing positive evolution cycle, and improving the consistency and discrimination accuracy of the subspace structure.

[0032] (5) By calculating the intra-cluster normalized distance, the distortion problem of cross-cluster importance comparison caused by the difference in distribution scale between different semantic feature clusters is eliminated, so that the allocation of feature weights can be kept fair and comparable between different semantic groups, and the rationality of feature evaluation is improved.

[0033] (6) By updating the weights through the exponential mapping function, a wide and stable high-weight interval and a fast low-weight suppression interval are provided, which makes the weights of the core discriminative features stable and the weights of the noise features decisively suppressed, thus optimizing the response surface of the weight adjustment and enhancing the population stability in the evolutionary search process.

[0034] (7) This application has strong versatility and is suitable for high-dimensional sparse text scenarios. This solution can be stably applied to large-scale high-dimensional sparse text classification tasks such as sentiment analysis and topic classification. It has good adaptability to different classification models (support vector machine, Naive Bayes, neural network), and has a wide range of applications and high practical value. Attached Figure Description

[0035] Figure 1 This is a basic framework diagram of an algorithm according to one embodiment of this application. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be further described in detail below with reference to the embodiments and accompanying drawings.

[0037] Text data suffers from high-dimensional sparsity, strong semantic correlation between terms, and a large number of redundant features. Existing text feature selection methods often rely on single statistical indicators or independent feature evaluations when processing high-dimensional sparse data, lacking overall modeling of the semantic relationship structure between terms. This leads to the potential destruction of the underlying semantic structure and reduced classification performance during feature selection. To address this, this application proposes a text feature selection and classification method based on the evolution of feature cluster subspaces.

[0038] The core idea of ​​this application is to transform the word relationships in the high-dimensional text feature space from "independent evaluation" to "subspace structure expression". By constructing a feature cluster partitioning mechanism to characterize the semantic correlation between words, and on this basis, to establish a multi-subspace co-evolution framework, the feature selection process not only depends on the importance of individual words, but also on their structural position in their respective feature subspace (semantic subspace) and their overall distribution contribution, thereby achieving structured optimization of feature subsets.

[0039] Specifically, this application first vectorizes the text data, mapping the original text into a high-dimensional term feature space; then, it constructs a cluster partitioning representation at the feature level, dividing the terms into multiple feature subspaces, and on this basis, constructs a multi-layer individual structure that includes cluster partitioning, feature weights, and selection states; during the evolution process, it establishes the distribution structure of terms in the subspace by calculating the central representation of each feature cluster, and realizes the dynamic migration of features between different subspaces based on the relationship between terms and cluster centers, thereby forming a subspace structure that is constantly updated with the evolution process.

[0040] Building upon this foundation, this application further constructs a weight update mechanism based on intra-cluster distribution, whereby feature weights are determined by their relative positions in the subspace. Continuous weights are transformed into discrete selection states through probability mapping, thereby achieving dynamic generation of feature subsets. Simultaneously, by constructing a discriminative divergence evaluation function based on inter-class separation and intra-class compactness, the classification ability of different feature subsets is uniformly measured, thereby driving the population evolution process and achieving global optimization search of the feature subspace.

[0041] Through the above process, this application forms a feature selection mechanism that transforms from "independent feature selection" to "subspace co-evolution", which makes term selection not only consider individual contributions, but also comprehensively consider their overall role in the semantic structure, thereby effectively maintaining the semantic consistency and discriminative ability of text data while reducing feature dimensionality.

[0042] Compared with existing technologies, this application constructs a subspace representation mechanism based on feature cluster partitioning, transforming the text feature matrix into multiple feature subspaces with semantic structure constraints. Furthermore, it introduces a multi-subspace collaborative dynamic evolution strategy to achieve dynamic reconstruction and optimization of features across different subspaces. Further, by coupling feature weight updates and probability selection mechanisms during the evolution process, it achieves adaptive filtering of high-dimensional text features, thereby obtaining a feature subset with strong discriminative power for subsequent classification model training. Thus, this application achieves structured modeling and optimization of high-dimensional text feature spaces, effectively avoiding the destruction of semantically related features during feature selection, improving the expressive power and classification performance of feature subsets, and can be applied to sentiment analysis, text topic classification, and other high-dimensional text data modeling and classification tasks, demonstrating significant practical application value.

[0043] The working principle and process of the present application solution will be described in detail below based on specific embodiments.

[0044] Example 1: This embodiment provides a text feature selection and classification method based on feature cluster subspace evolution, including: S1. Data preprocessing: The text corpus is segmented, vectorized, and normalized to construct a term co-occurrence relation matrix.

[0045] The feature subspace optimization in this embodiment is used to automatically filter discriminative term subsets in a high-dimensional text feature space. Its core design lies in achieving adaptive reconstruction of terms in multiple feature subspaces through feature cluster partitioning and dynamic jumping mechanism between clusters, thereby improving the discriminativeness of feature representation.

[0046] In the input phase, the raw text corpus data from the text classification task is received and vectorized to construct a unified feature space representation. Let the raw text corpus data contain... There are 10 document samples, and each document sample corresponds to a category label.

[0047] Through term extraction and statistical representation, the original text corpus data is represented as a text feature matrix: ; in, This represents the text feature matrix, where each column vector corresponds to the statistical distribution of a term across all document samples. column vector Indicates the first The feature vector of the nth term, i.e., the nth word The statistical representation of a term in all document samples can be constructed using term frequency-inverse document frequency (TF-IDF) values ​​in a practical implementation. Indicates the size of the dictionary (number of entries). Indicates the number of document samples.

[0048] The text feature matrix is ​​normalized to obtain the normalized text feature matrix: ; in, This represents the normalization function, used to scale each column vector to reduce the impact of high-frequency words on subsequent calculations.

[0049] The data input to the feature subspace optimization module is the normalized text feature matrix mentioned above. With category labels ,in Indicates document sample category tags, Indicates the number of categories.

[0050] Define the complete set of features as: Each index corresponds to a unique term in the dictionary.

[0051] Furthermore, to characterize the semantic relevance and contextual co-occurrence relationships between terms, a term co-occurrence relationship matrix is ​​constructed: ; in, This represents the term co-occurrence relation matrix, i.e., the co-occurrence relation matrix between terms, where the nth term... Line 1 Column elements Indicates the first The term and the first The co-occurrence strength of a term in all documents is obtained by multiplying the transpose of the standardized feature matrix by itself, thus reflecting the degree of co-occurrence of terms in the corpus. This matrix can be used to characterize semantic similarity or topic relevance.

[0052] Through the above process, the text feature space and the statistical correlation structure between terms are obtained, which provides a foundation for subsequent semantic-driven feature grouping.

[0053] S2. Population initialization: The population is initialized based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. Among them, the feature cluster partitioning vector is used to record the feature cluster number to which each term belongs, so as to realize the semantic grouping of the text feature space; The feature weight vector is used to quantify the importance of each term in the corresponding feature cluster, providing a continuous score basis for feature selection; The feature selection state vector is used to identify whether each term is selected in binary state, in order to generate a subset of candidate features that can be used for classification.

[0054] After obtaining feature representations and term co-occurrence relationships, a population search model based on semantic structure constraints is constructed. Define the... The generation population is: ; in, Indicates the first The first generation of the population Individual, Indicates population size.

[0055] Each individual represents a candidate feature subspace, constructed using a three-layer structure, including a feature cluster partitioning vector, a weight vector, and a selection state vector. This structure simultaneously describes the semantic affiliation, importance, and selection of a term into the final feature subset. It can be represented as: ; The first layer is the feature cluster partitioning vector: ; in Indicates the first The first generation of the population The individual The feature cluster (semantic cluster, term cluster) to which each term belongs is numbered.

[0056] The feature clusters are not initialized randomly, but are based on the term co-occurrence relation matrix. The construction process begins. First, the term co-occurrence relation matrix is ​​constructed. The graph is transformed into an undirected weighted graph, where each term corresponds to a node in the graph, and the edge weight between any two terms is determined by their co-occurrence strength in the corpus. Based on this, a clustering method based on similarity metrics is used to divide the terms, so that terms with higher co-occurrence strength are grouped into the same cluster, resulting in the initial cluster division: ; This division allows terms with high co-occurrence frequency or semantic association in the text to be grouped into the same cluster, thus forming a feature grouping structure with thematic consistency.

[0057] Based on this, to ensure population diversity, a local perturbation mechanism is introduced when initializing individuals: ; in This represents the cluster perturbation probability, and its value range is... This is used to control the degree of deviation of an individual from the basic feature cluster division during initialization. Indicates the interval Uniform random sampling on the surface allows for the introduction of structural differences while maintaining semantic consistency.

[0058] The second layer is the feature weight vector: ; Its initialization method is as follows: ; Indicates the first The first generation of the population The individual The weight value of the term, i.e., the weight of the term. The importance weight of each term in the initial state, which is determined by the interval uniform distribution on Obtained through random sampling, it is used to maintain unbiased exploration capability for each feature in the early stages of the search.

[0059] The third layer is the feature selection state vector: ; Its initialization method is as follows: ; in, Indicates the first The first generation of the population Among the individuals, the first The selection status flag for the term is used to indicate the selection status of the term. Whether a term is selected into the current feature subset is determined by weighting. With threshold The comparison results show that the value is 1 when the weight is greater than the threshold, and 0 otherwise, thus forming the initial feature selection result. Take experience points. .

[0060] Therefore, we define the candidate feature subset corresponding to an individual: ; Through the above construction method, each individual not only represents a subset of candidate features, but also implies a semantic grouping structure based on term co-occurrence relationship, which enables the subsequent evolution process to be carried out in coordination at the feature cluster level and the feature selection level, thereby improving the search efficiency and expressive ability of the high-dimensional sparse feature space of the text.

[0061] S3. Feature Cluster Subspace Construction: Divide the terms into several feature clusters based on the feature cluster partition vector, and calculate the weighted center vector of each feature cluster.

[0062] After establishing the individual structure, the text feature space is subspaced based on the first-level cluster partitioning. In text classification tasks, features correspond to terms or lexical units (n-grams, sequences of consecutive terms (or characters), and their statistical representation in the corpus reflects their distribution across different documents. For the first... The individual, in the first Definition of the middle generation The feature clusters are: ; This set represents the number of individuals in the current set. A cluster is a set of term indexes that can be understood as a set of terms that have similar distribution patterns or potential semantic relationships in the corpus.

[0063] Furthermore, this step imposes a key technical limitation: the set of terms actually involved in the calculation of the center vector is not all terms within the cluster. Instead, it only takes the set of terms within the cluster that have been selected and preliminarily identified as having discriminative value, and defines it as: ; in Indicates the first Among the individuals, the first The set of terms that actually participate in the construction of the text classification model within a cluster.

[0064] In text classification scenarios, this process corresponds to dividing terms into multiple feature subspaces according to their distribution patterns in the document set, with each subspace containing terms having similar contextual distribution features.

[0065] Based on the above cluster structure, calculate the center vector of each cluster: ; in, Indicates the first Among the individuals, the first The center vector of each feature cluster is obtained by weighting and normalizing the normalized feature vectors of the selected terms within the cluster. This center vector characterizes the overall response features and average distribution pattern of the terms in the document feature space. The purpose of this technique is to address the noise interference from unselected terms on the representation of the subspace center during evolution. In early iterations, each feature cluster contains a large number of redundant or low-discriminative terms that, although assigned to the cluster, are not selected. If these terms (i.e., the entire set) are... Including cluster center vectors in the calculation will cause the cluster center vector to shift towards weakly discriminative or noisy regions in the feature space, thereby obscuring the semantic direction commonly pointed to by truly discriminative terms. This embodiment strictly limits the terms participating in the calculation to... and intersection This ensures that each subspace reconstruction focuses on the selected features with high confidence, enabling the cluster center to accurately represent the true distribution of the current discriminative subspace, thereby driving subsequent term migration and weight updates to be more accurate, forming a self-reinforcing positive evolutionary cycle.

[0066] S4. Evolution of the subspace of characteristic clusters: S4.1 Update the feature cluster number to which the term belongs based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, so as to complete the feature cluster partitioning vector update.

[0067] After completing feature cluster partitioning and cluster center calculation, a cluster center-based feature redistribution mechanism (inter-cluster jump-driven feature subspace reconstruction mechanism) is introduced to enable dynamic migration of terms between different subspaces. In text classification scenarios, this mechanism is used to adaptively adjust the semantic attribution of terms based on their distribution characteristics in the document space.

[0068] For the For any term in an individual, its cluster affiliation is redefined based on the distance relationship between its normalized feature vector and the cluster center vectors, specifically as follows: ; in, Indicates the first The first generation of the population The individual The feature cluster number to which each term belongs. The distance metric function between feature vectors can be Euclidean distance or cosine distance in practice, and is used to measure the similarity between the term distribution and the cluster center distribution.

[0069] Through the above-mentioned update method, namely the term redistribution mechanism based on cluster centers, each term undergoes adaptive dynamic migration between different clusters, i.e., multiple subspaces, based on its distribution characteristics in the document feature space and its similarity with different feature cluster centers. This allows terms with similar text distribution patterns to gradually gather into the same subspace, thereby forming a semantic structure that changes continuously with the evolution process.

[0070] After updating the cluster affiliation, rebuild the cluster sets. This forms a new feature subspace partitioning structure and provides an updated structural basis for subsequent feature weight adjustment and selection mechanisms.

[0071] S4.2. Based on the normalized distance between the feature vector of a term and the weighted center vector of its feature cluster, the feature weight vector is adaptively updated.

[0072] After reconstructing (updating) the feature cluster structure, the importance of each term in the text classification task is reassessed based on its distribution within its cluster. In text data, each feature corresponds to a term, and the feature vector of a term... Reflecting the statistical distribution of the term across all documents, the distance between the term's feature vector and the cluster center can be used to characterize its consistency with the overall semantic distribution of the cluster.

[0073] For the The individual For each term, first calculate its normalized distance relative to the center of its cluster: ; in, Indicates the first The first generation of the population The individual The normalized distance of the term is the term of the first term. The relative distance ratio of a term within its own cluster reflects the degree of deviation of the term from the current feature cluster. The smaller the distance, the closer its distribution characteristics are to the overall distribution pattern of the cluster.

[0074] Based on this, the feature weights are updated: ; in, Indicates the first The first generation of the population The individual The weight value of each term is the updated term weight. This update method transforms distance information into continuous weights through exponential mapping, so that terms that are closer to the cluster center and more representative in the document distribution receive higher weights, while terms that deviate more have lower weights, thus enabling the weights to reflect the representativeness of a term in its respective feature subspace.

[0075] Through the above mechanism, adaptive adjustment of feature weights based on local information of cluster structure is realized, enabling the second-layer weight structure to respond to the semantic segmentation results of the first layer.

[0076] S4.3. Generate a feature selection state vector based on the updated feature weight vector, and obtain a subset of candidate features based on the feature selection state vector, where each feature corresponds to a term.

[0077] After obtaining the updated feature weights, the continuous weights are further mapped to discrete selection states (feature selection probabilities), thereby generating a new subset of term features. In text classification tasks, this process corresponds to selecting the set of terms from the dictionary that are most discriminative for classification.

[0078] For the Word features in an individual First, normalized intra-cluster selection probabilities are constructed within its own cluster: ; in, Indicates terms The probability of a term being selected within its feature cluster, which reflects the relative importance of the term in that feature subspace.

[0079] The feature selection state is updated based on the intra-cluster selection probability: ; in, This indicates the updated feature selection status. This represents the Bernoulli random sampling operator, which maps continuous weights to discrete selection results.

[0080] Through the above mechanism, feature selection based on relative importance is completed within each feature cluster, forming a candidate feature subset based on intra-cluster competition. This ensures that the final generated feature subset is constrained by the statistical characteristics of terms in the document distribution space, while retaining a certain degree of randomness to maintain the diversity of the search process, thereby achieving effective coupling between the second-layer weight structure and the third-layer selection structure.

[0081] S5. Fitness Evaluation: Construct a discriminant divergence fitness function based on inter-class separation and intra-class compactness, evaluate candidate feature subsets, and obtain the corresponding fitness values.

[0082] After obtaining the subset of term features corresponding to each individual, it is necessary to evaluate its discriminative ability in the text classification task. Let the first... The individual in the first The corresponding feature subset is This subset represents the set of terms selected from the original dictionary for text representation.

[0083] First, define the inter-class separation degree: ; in, Indicates category In feature subset The mean feature vector, composed of the statistical representations of all documents belonging to the category on the selected terms, is used to characterize the central position of the category in the feature space. This metric measures the degree of separability of different categories of text in the current feature subspace, i.e., the degree of distinction; a higher value indicates a more significant difference between categories.

[0084] Further define intra-class compactness: ; in, Indicates belonging to a category A collection of document sample indexes. Indicates the first Documents in the feature subset The normalized feature vectors on the feature space. This metric measures the degree of clustering of documents within the same category in the feature space; a smaller value indicates a more compact intra-class structure.

[0085] Based on the above two indicators, construct the individual fitness function: ; in, Indicates the first The individual in the first Fitness value of the generation To prevent constants with zero denominators, this embodiment uses discriminative divergence, constructed based on inter-class separation and intra-class compactness, as a surrogate metric, rather than directly using classification accuracy as the optimization objective. This avoids the high computational complexity caused by frequent training of the classification model during evolution. Furthermore, this metric is consistent with the classification decision boundary, thus effectively guiding the feature subset to optimize in a direction favorable to classification.

[0086] S6. Iterative optimization: Drive the population to iteratively evolve according to the fitness value to obtain the optimal feature subset.

[0087] By maximizing the fitness function, the search for discriminative term subsets is optimized.

[0088] During population renewal, an elite retention strategy based on fitness ranking is adopted, prioritizing the retention of individuals with high fitness rankings, and performing individual update operations to generate new candidate individuals, thereby obtaining the next generation of the population. The retention ratio is usually set to a certain proportion of the population size, such as 10%–30%, to ensure that excellent individuals are inherited while maintaining population diversity. New individuals are generated by randomly perturbing or locally adjusting at least one of the feature cluster affiliation, feature weight distribution, and feature selection status of existing individuals, in order to maintain population diversity and enhance search capabilities. Through multiple generations of iterative evolution, the optimal individual is eventually obtained: ; The corresponding optimal feature subset (optimal term subset) is: .

[0089] S7. Classification Execution: Reconstruct the text feature representation using the optimal feature subset, train the classification model, and complete the text classification.

[0090] After obtaining the optimal subset of term features, the original text data is reconstructed based on this subset, that is, only the term indexes belonging to the subset are retained. Based on the feature dimensions, the reduced-dimensional text representation matrix is ​​obtained: ; in, This represents the text feature matrix constrained by the optimal feature subset, which serves as the input to the subsequent classification model. This operation physically compresses the original high-dimensional sparse "complete term space" into a low-dimensional dense "discriminative concept space," generating the reconstructed text feature representation matrix. Since the feature selection process in this scheme relies entirely on the discriminant divergence fitness function, which is independent of the classifier, the resulting optimal feature subset... and its reconstructed representation It has good versatility and does not overfit to any specific classifier.

[0091] Based on the above feature representations, train a text classification model: ; in, This represents a classification mapping function, whose input is the dimensionality-reduced text feature vector, and whose output is the corresponding category label.

[0092] In practice, the classification model can be trained using support vector machines, Naive Bayes, or neural network models to complete the text classification task.

[0093] Through the above-mentioned fitness assessment, population evolution and classification model construction process, discriminative terms are automatically selected from high-dimensional sparse text features, and the classification model is trained based on the optimized feature subspace, thereby improving the accuracy and computational efficiency of text classification.

[0094] To verify the effectiveness of the proposed method in text classification tasks, the following experimental environment was configured.

[0095] In terms of hardware, a high-performance computing server was used as the experimental platform, with an Intel Core 12700 processor, 32GB of memory, and a Tesla V100 graphics processing unit with 16GB of video memory to accelerate feature processing and model training.

[0096] In terms of software, the system uses Ubuntu 22.04 operating system, Python 3.10 as the programming language, the scikit-learn library to implement traditional classification models such as support vector machines, the PyTorch 2.0 framework to implement neural network models, and NumPy and SciPy to complete matrix operations and data processing.

[0097] In terms of dataset construction, this embodiment uses a self-built Chinese sentiment classification text dataset for validation. The dataset originates from text on a social media platform and, after noise reduction, duplicate sample removal, and manual annotation, forms a standard dataset. The dataset contains 6000 text samples, each consisting of a natural language sentence with an average length of approximately 35 Chinese characters. Based on the sentiment expressed in the text, the samples are divided into three categories: positive sentiment, neutral sentiment, and negative sentiment, with 2000 samples in each category to ensure a balanced distribution of categories.

[0098] In the data preprocessing process, the text is first segmented into word sequences, transforming the original text sequence into a word sequence. Then, a vocabulary is constructed, and the TF-IDF method is used to vectorize the text, resulting in a high-dimensional sparse feature matrix. After processing, the feature space dimension is approximately 7000, meaning each text sample is represented by 7000-dimensional word features, thus forming a normalized data matrix. .

[0099] In terms of data partitioning, the dataset is divided into training and testing sets, with the training set accounting for 80% and the testing set accounting for 20%, meaning there are 4800 training samples and 1200 testing samples. The training set is used for feature subspace search and classification model training, while the testing set is used to evaluate the model's generalization performance.

[0100] In terms of experimental setup, to verify the effectiveness of the proposed method, three sets of comparative experiments were set up. The first set was the baseline method, which directly used the original TF-IDF features as input to train the classification model without any feature selection. The second set was the traditional feature selection method, which used information gain to rank the features and selected the top 3500 features to construct a subspace. The third set was the proposed method, which automatically searched for the optimal feature subset through a multi-subspace co-evolution mechanism and trained the classification model based on this subset.

[0101] In the method of this application, the population size is set to 30, the number of evolutionary iterations is set to 50 generations, and the number of feature clusters is set to 20, thereby achieving a balance between search efficiency and search capability.

[0102] For classification model training, to verify the adaptability of feature subsets to different models, a lightweight neural network was selected as the classifier for experiments. The neural network model adopted a three-layer fully connected structure, with the input layer dimension consistent with the feature subset size, the number of hidden layer neurons set to 512 and 128 respectively, and the output layer corresponding to the three-class classification task. The Adam optimizer was used during model training, with a learning rate set to... The batch size is set to 32, and the number of training rounds is set to 50.

[0103] In terms of experimental evaluation metrics, classification accuracy and F1 score were used as the main evaluation metrics to comprehensively reflect the classification performance of the model on different categories. All experiments were repeated 10 times under the same random seed conditions, and the average value was taken as the final result to reduce the impact of randomness on the experimental results.

[0104] The experimental results are shown in Table 1.

[0105] Table 1. Experimental results of different methods on the emotion classification task. ; As shown in Table 1, without feature selection, the baseline method exhibits a certain degree of redundancy in the high-dimensional sparse feature space, limiting model performance, with a classification accuracy of 0.83. After using the information gain method for feature selection, the classification performance is improved to some extent, with an accuracy of 0.85, indicating that traditional feature selection methods can remove redundant features and improve model performance to a certain extent.

[0106] In comparison, the proposed method achieves the best performance with the same dataset and model configuration, achieving a classification accuracy of 0.90 and an F1 score of 0.89. Furthermore, the standard deviation of the experimental results is significantly reduced, indicating that the model performs more stably across different training epochs. This result demonstrates that the proposed method, by introducing a multi-subspace co-evolution mechanism based on feature clusters, enables terms to dynamically adjust between different feature subspaces, thereby effectively improving the discriminative power of the feature subset. Further analysis shows that the feature subset selected by the proposed method, while maintaining a low dimensionality, better preserves the semantic structure information of the text, resulting in a more distinct distribution of different sentiment categories in the feature space, thus improving the generalization performance of the classification model. Regarding the size of the feature subset, the proposed method adaptively determines the feature dimension during the evolution process. Multiple experiments show that the final feature subset size is approximately 25%–40% of the original feature space, i.e., approximately 1750 to 2800 dimensions. Compared to the 3500-dimensional features selected by the information gain method, the proposed method achieves superior classification performance while further reducing the feature dimension, indicating that the selected features have higher information density and discriminative power.

[0107] In summary, the method proposed in this application can significantly improve classification performance in text classification tasks while reducing feature dimensionality, and has good practical application value.

[0108] Example 2: This embodiment provides a text feature selection and classification system based on feature cluster subspace evolution, including: The data preprocessing module is used to perform word segmentation, vectorization, and normalization on the text corpus, and to construct a word co-occurrence relation matrix; The population initialization module is used to initialize the population based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. The feature cluster subspace construction module is used to divide the term into several feature clusters based on the feature cluster partitioning vector and calculate the weighted center vector of each feature cluster. The feature cluster subspace evolution module is used to update the feature cluster number to which the term belongs based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, so as to complete the feature cluster partitioning vector update; based on the normalized distance between the feature vector of the term and the weighted center vector of the feature cluster to which it belongs, the feature weight vector is adaptively updated; based on the updated feature weight vector, a feature selection state vector is generated, and a candidate feature subset is obtained based on the feature selection state vector. The fitness evaluation module is used to construct a discriminant divergence fitness function based on inter-class separation and intra-class compactness, evaluate candidate feature subsets, and obtain the corresponding fitness values. The iterative optimization module is used to drive the iterative evolution of the population according to the fitness value to obtain the optimal feature subset; The classification execution module is used to reconstruct the text feature representation using the optimal feature subset, train the classification model, and complete the text classification.

[0109] The system is used to implement the method described in Embodiment 1.

[0110] Example 3: This embodiment provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to execute the method as described in Embodiment 1.

[0111] Example 4: This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the method described in Embodiment 1.

[0112] Example 5: This embodiment provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described in Embodiment 1.

[0113] The specific implementation of the system, electronic device, computer-readable storage medium, and computer program product provided in this application can be referred to the specific embodiments of the above methods, and will not be repeated here.

[0114] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0115] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A text feature selection and classification method based on feature cluster subspace evolution, characterized in that, include: Data preprocessing: The text corpus is segmented, vectorized, and normalized to construct a term co-occurrence relation matrix; Population initialization: The population is initialized based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. Among them, the feature cluster partitioning vector is used to record the feature cluster number to which each term belongs, so as to realize the semantic grouping of the text feature space; The feature weight vector is used to quantify the importance of each term in the corresponding feature cluster, providing a continuous score basis for feature selection; The feature selection state vector is used to identify whether each term is selected in binary state, in order to generate a subset of candidate features that can be used for classification; Feature cluster subspace construction: Divide the terms into several feature clusters based on the feature cluster partition vector, and calculate the weighted center vector of each feature cluster; Feature cluster subspace evolution: Based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, update the feature cluster number to which the term belongs, so as to complete the feature cluster partitioning vector update; The feature weight vector is adaptively updated based on the normalized distance between the feature vector of a term and the weighted center vector of its feature cluster. The updated feature weight vector is converted into intra-cluster selection probabilities, and a feature selection state vector is generated through Bernoulli sampling, thereby obtaining a subset of candidate features, where each feature corresponds to a term. Fitness evaluation: A discriminant divergence fitness function is constructed based on inter-class separation and intra-class compactness to evaluate candidate feature subsets and obtain the corresponding fitness values; Iterative optimization: The population is driven to evolve iteratively based on fitness values ​​to obtain the optimal feature subset; Classification execution: The text feature representation is reconstructed using the optimal feature subset, the classification model is trained, and the text classification is completed.

2. The method according to claim 1, characterized in that, The construction of the term co-occurrence relation matrix includes: representing the text corpus data as a text feature matrix. ,in for The Middle Feature vectors of each term Each term's feature vector corresponds to the statistical distribution of that term across all document samples. For the number of document samples, For the number of terms; Normalization is performed to obtain the normalized text feature matrix. ,in For the first Normalized feature vectors of each term; according to the formula Calculate the term co-occurrence matrix , of which Line 1 Column elements Indicates the first The term and the first The co-occurrence intensity of each term across all documents.

3. The method according to claim 2, characterized in that, The initialization of the population based on the term co-occurrence relation matrix includes: The initial cluster partitioning results are obtained based on the term co-occurrence relation matrix; Based on the term co-occurrence relation matrix, the feature cluster partitioning vector is initialized using the following formula: ； in, Indicates the first The first generation of the population The individual The feature cluster number to which each term belongs; , The total number of individuals in the population; Indicates the first The initial clustering results of each term. Indicates the probability of cluster perturbation. Indicates the interval A uniformly random integer within the range represents a randomly assigned cluster number. Indicates the total number of feature clusters; The feature weight vector is initialized as follows: each element in the vector represents the weight value of a term, and all elements are initialized to the value specified in the original text. Independent random numbers that follow a uniform distribution within the interval; The initialization formula for the feature selection state vector is: ； in, Indicates the first The first generation of the population Among the individuals, the first The selection status identifier for each term is determined by the weight. With threshold The comparison yields the result if Greater than , , indicates that it is selected, otherwise it is not. , indicates that it was not selected; among them, Indicates the first The first generation of the population The individual The weight value of each term.

4. The method according to claim 3, characterized in that, The formula for calculating the weighted center vector of each feature cluster is as follows: ； in, Indicates the first The first generation of the population The first individual The weighted center vector of each feature cluster; Indicates the first The first generation of the population The individual The weight value of each term; Indicates the first The first generation of the population The first individual The set of selected terms within each feature cluster. , Indicates the first The first generation of the population The first individual A feature cluster, To indicate the first The first generation of the population A subset of candidate features for each individual Complete set of features ; .

5. The method according to claim 4, characterized in that, The formula for the updated term's feature cluster number is as follows: ； in, Indicates the first The first generation of the population The individual The feature cluster number to which each term belongs. This represents the distance metric function.

6. The method according to claim 5, characterized in that, The adaptive update of the feature weight vector is formulated as follows: ； ； in, Indicates the first The first generation of the population The individual Normalized distance of each term; and They represent the first The first generation of the population The first individual The and the first The weighted center vector of each feature cluster; For the first Normalized feature vectors of each term; Indicates the first The first generation of the population The first individual A feature cluster; Indicates the first The first generation of the population The individual The weight value of each term.

7. The method according to claim 6, characterized in that, The inter-class separation degree is calculated using the following formula: ； in, Indicates the first The first generation of the population Inter-class separability of an individual, and Numbers indicating different categories, and They represent in Upper Class and First The mean feature vector of all document samples in the class; The intra-class compactness is calculated using the following formula: ； ； in, Indicates the first The first generation of the population Intra-class compactness of individual entities, For the first The index set of all document samples in the class. , Total number of categories; Indicates the first A document sample in Normalized eigenvectors on, Indicates in Upper The mean feature vector of all document samples in the class; wherein, the feature vector of a document sample on the candidate feature subset is composed of the statistical representation of the document sample on the terms in the candidate feature subset; The fitness function is: ； in, Indicates the first The first generation of the population The fitness value of each individual To prevent constants with a denominator of zero.

8. The method according to any one of claims 1 to 7, characterized in that, The classification model uses support vector machine, naive Bayes or neural network, and takes the feature vector reconstructed from the optimal feature subset of the document as input to complete the text classification.

9. A text feature selection classification system based on feature cluster subspace evolution, characterized in that, To implement the method described in any one of 1 to 8, comprising: The data preprocessing module is used to perform word segmentation, vectorization, and normalization on the text corpus, and to construct a word co-occurrence relation matrix; The population initialization module is used to initialize the population based on the term co-occurrence relation matrix. Each individual in the population includes a three-layer structure: feature cluster partitioning vector, feature weight vector, and feature selection state vector. The feature cluster subspace construction module is used to divide the term into several feature clusters based on the feature cluster partitioning vector and calculate the weighted center vector of each feature cluster. The feature cluster subspace evolution module is used to update the feature cluster number to which a term belongs based on the distance between the feature vector of the term and the weighted center vector of each feature cluster, so as to complete the feature cluster partitioning vector update; based on the normalized distance between the feature vector of the term and the weighted center vector of the feature cluster to which it belongs, the feature weight vector is adaptively updated; the updated feature weight vector is converted into intra-cluster selection probability, and a feature selection state vector is generated through Bernoulli sampling; and a candidate feature subset is obtained based on the feature selection state vector. The fitness evaluation module is used to construct a discriminant divergence fitness function based on inter-class separation and intra-class compactness, evaluate candidate feature subsets, and obtain the corresponding fitness values. The iterative optimization module is used to drive the iterative evolution of the population according to the fitness value to obtain the optimal feature subset; The classification execution module is used to reconstruct the text feature representation using the optimal feature subset, train the classification model, and complete the text classification.

Images