A feature selection method based on label correlation and weighted features, a terminal device, and a storage medium

Abstract

The application relates to a feature selection method based on label correlation and weighted features, a terminal device and a storage medium. The method comprises the following steps: a maximum entropy model is used to construct a target function for feature selection; the size of the mapping parameter corresponding to each feature in the feature space is obtained by solving the target function; and feature selection is performed based on the size of the mapping parameter. The application uses the linear correlation between labels to construct the generic feature relationship shared by the labels in the output space, and uses the feature weight obtained by label distribution mutual information analysis to weight the features. In addition, the sparse norm is used to select the common features with strong classification ability in the entire label space, so that the robustness of the algorithm is improved.

Description

Technical Field

[0001] This invention relates to the field of big data processing, and in particular to a feature selection method, terminal device, and storage medium based on label association and weighted features. Background Technology

[0002] In the era of big data, the dimensionality of the feature space of samples exhibits an exponential growth trend, and labeled distribution data with fine-grained descriptive paradigms also face this problem. Since feature selection can reduce feature dimensionality and improve classification performance by removing redundant features to find effective features, dedicated feature selection algorithms for labeled distribution data have been proposed in recent years. However, existing algorithms have failed to fully utilize the label association information within the dedicated algorithm framework to mine shared class features, and have ignored the potential supervision information between the feature space and the label distribution space. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes a feature selection method, terminal device, and storage medium based on label association and weighted features.

[0004] The specific plan is as follows:

[0005] A feature selection method based on label association and weighted features includes the following steps:

[0006] The objective function for feature selection is constructed using a maximum entropy model:

[0007]

[0008] Where θ represents the mapping parameter matrix of the correlation between features and labels, L(θ) represents the loss function, Φ(θ) represents the sparsity term, and Γ(θ) represents the label association term. The term represents the weighted feature term, and α, β, and γ represent the balance parameters of the sparse term, the labeled association term, and the weighted feature term, respectively.

[0009] By solving the objective function, the magnitude of the mapping parameters corresponding to each feature in the feature space can be obtained;

[0010] Feature selection is based on the magnitude of the mapping parameters.

[0011] Furthermore, the sparse term uses the L21 norm to learn features.

[0012] Furthermore, in solving for the sparse term, it can be rewritten as follows:

[0013] Φ(θ)=tr(θTDθ)

[0014] Where tr(.) denotes the trace of the matrix, the superscript T denotes the transpose of the matrix, and D denotes a diagonal matrix with diagonal elements as... The diagonal matrix, d kk Let θ represent the k-th row and k-th column of the diagonal matrix D. k Let represent the k-th row of the mapping parameter matrix θ, k∈[1,d] represent the feature index, d represent the number of features, ||.||2 represents the L2 norm, and ε represents a small positive number.

[0015] Furthermore, the formula for calculating the associated items is as follows:

[0016]

[0017] Where, p ij Indicates sample xi The predicted value of the descriptive power of the j-th label, p il Indicates sample x i The predicted value of the descriptive power of the l-th label, R jl Let i represent the Pearson propensity relationship between the j-th label and the l-th label, where i represents the sample index, n represents the sample number, j and l both represent the label index, and c represents the label number.

[0018] Furthermore, the weighted feature terms are measured using the mutual information of the labeled distribution;

[0019] Mutual information I between continuous eigenvectors and label distribution vectors cc ( x·m ,d ·l The following formula is used for calculation:

[0020]

[0021] Where, x ·m Let d represent the vector formed by the values ​​of the m-th continuous feature of all samples. ·l Let t represent the vector formed by the values ​​of all samples at the l-th label, where t represents the number of nearest neighbors and n represents the number of samples. This indicates that the condition |x is satisfied. im -x i′m | <z i The number of samples, i′=1,2,...,n, x im Let x represent the value of the i-th sample in the m-th continuous feature. i′m Let z represent the value of the i′-th sample in the m-th continuous feature. i Let x represent the i-th sample. i The distance to its t-th nearest sample. This indicates that the condition |d is satisfied. il -d i′l | <z i The number of samples i′=1,2,...,n, d il Let x represent the i-th sample.i The descriptive power of the l-th label, d i′l Let x represent the i′-th sample. i′ In the descriptive power of the l-th label, ψ(g) represents the digamma function;

[0022] Mutual information I between discrete feature vectors and label distribution vectors dc (x ·r ,d ·l The following formula is used for calculation:

[0023]

[0024] Where, x ·r Let represent the vector formed by the values ​​of the r-th discrete feature of all samples. In vector x ·r Above and sample x i Let s be the number of samples with the same value. i Indicates the relationship with sample x i Among samples with the same discrete eigenvalues, sample x i The distance to the nearest t-th feature is then h. i Indicates the relationship with sample x i The distance is less than s i The number of samples;

[0025] Mutual information I based on continuous feature vectors and label distribution vectors cc (x ·m ,d ·l Mutual information I between discrete feature vectors and label distribution vectors dc (x ·r ,d ·l Construct a mutual information matrix M, where the elements M kj This represents the mutual information value between the k-th feature and the j-th label;

[0026] Based on the mutual information matrix M, a weighted feature term is constructed.

[0027]

[0028] Where diag(g) represents taking the diagonal elements of the matrix, ||.||2 represents the L2 norm, and the superscript T represents the transpose of the matrix.

[0029] Furthermore, the loss function is constructed based on KL divergence.

[0030] Furthermore, the objective function is solved using the quasi-Newton method.

[0031] A feature selection terminal device based on tag association and weighted features includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described above in the embodiments of the present invention.

[0032] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above in the embodiments of the present invention.

[0033] This invention employs the above technical solution, utilizing the linear correlation between labels to construct shared class feature relationships in the output space, and using feature weights obtained from label distribution mutual information analysis to assign weights to the features. Furthermore, a sparse paradigm is used to select common features with strong classification capabilities across the entire label space, improving the algorithm's robustness. Attached Figure Description

[0034] Figure 1 The diagram shown is a flowchart of Embodiment 1 of the present invention. Detailed Implementation

[0035] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention.

[0036] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0037] Example 1:

[0038] Let the feature space of multi-labeled data be X = [x1; x2; ...; x...]. n ]∈R n×d ,x i =[x i1 ,x i2 ,…,x id ]∈R 1×d Let x represent the feature set of the i-th sample. ik Let F represent the feature value of the i-th sample in the k-th feature set, and let F = {f1, f2, ..., fk}. d}, where i∈[1,n] represents the sample index, k∈[1,d] represents the feature index, n represents the number of samples, d represents the number of features, and the label set is Y={y1,y2,…,y c}, j∈[1,c] represents the label index, and c represents the label number. The purpose of label distribution learning is to learn from the given dataset DT={(x1,D1),(x2,D2),K(xn D n Find a mapping parameter θ∈R that describes the correlation between each feature and the label. d×c D i =[d i1 ,d i2 ,…,d ic ]∈R 1×c Let di represent the label distribution vector of the i-th sample. j Let d represent the descriptive power of the i-th sample with the j-th label, satisfying d ij ∈[0,1] and ∑d ij =1. The higher the descriptive power, the greater the weight of the label in describing the sample. A machine learning paradigm that conforms to this constraint is called label distribution learning. The purpose of label distribution feature selection is to select the optimal feature subset from the above mapping parameters θ. The difference between label distribution learning and multi-label learning is that the target space of label distribution learning is a real number space that conforms to the label distribution constraint.

[0039] To achieve the goal of label distribution learning, this embodiment provides a feature selection method based on label association and weighted features, such as... Figure 1 As shown, the method includes the following steps:

[0040] S1: Construct the objective function for feature selection.

[0041] In this embodiment, the maximum entropy model is used as the prediction model for label distribution learning (LDL), and the Kullback-Leibler (KL) divergence is used to measure the difference between the predicted distribution and the actual distribution. The maximum entropy model is defined as shown in formula (1).

[0042]

[0043] Where Z is the normalization factor, ensuring that the output satisfies the constraints of the labeled distribution, p(y ij |x i ) represents sample x i The descriptiveness prediction for the j-th label. The KL divergence is defined as shown in Equation (2).

[0044]

[0045] Combining the maximum entropy model and KL divergence, the loss function of LDL is constructed as Equation (3):

[0046]

[0047] In this embodiment, in order to select features with stronger classification performance, a sparse term, a label association term, and a weighted feature term are added to the loss function to find these features, and the objective function is constructed as shown in formula (4).

[0048]

[0049] Where Φ(θ) is the sparse term, and Γ(θ) is the label association term. The term is a weighted feature term. α, β, and γ are the balance parameters of these three terms, which control the influence of the three regularization terms. The three regularization terms will be introduced in detail below.

[0050] (1) Sparse terms

[0051] Learning common features and class features enables the selection of more effective feature subsets during the feature selection process. In this embodiment, these two types of features are learned based on sparse learning and label association, respectively.

[0052] Table 1

[0053]

[0054] The purpose of sparse learning on features is to improve the robustness of feature selection algorithms and, in the process, remove redundant features to select effective features. Sparse learning can be implemented in various ways, such as l0 norm, l1 norm, etc. 2,1 Norms, etc. The l0 norm has a relatively simple form, but its mathematical form is non-convex, making it inconvenient for optimization. The l1 norm can achieve partial sparsity of matrix elements, which can be used to select class features in feature selection, and it is also easier to solve than the l0 norm. However, its partial sparsity effect makes the performance of removing redundant features not significant. As shown in the comparison of sparsity effects of the two learning paradigms in Table 1, the score of the corresponding feature is represented by the row inner product of the mapping parameter matrix θ. A row with an inner product of 0 represents a redundant feature. The matrix under the l1 norm is difficult to achieve row sparsity, and therefore it is not good at removing redundant features. In comparison, l 2,1 The norm can achieve row sparsity, which can better filter out redundant features and select more effective common features across the entire label space. Therefore, this embodiment chooses to use l. 2,1 The norm is used to learn common features, and sparse terms are constructed as shown in equation (5).

[0055] Φ(θ)=||θ|| 2,1 (5)

[0056] Because of l 2,1 Since the norm itself is not smooth and cannot be solved directly, equation (5) is rewritten as an equivalent form of equation (6), where D∈R d×d diagonal elements are The diagonal matrix is ​​given, and ε is a small positive number to prevent the denominator from being 0. This allows us to directly solve the sparse problem.

[0057] Φ(θ)=tr(θ T Dθ), (6)

[0058] (2) Mark related items

[0059] Learning classifiable features can also improve the performance of feature selection algorithms. In real-world scenarios, the presence or absence of certain features greatly influences, or even determines, the appearance of a label. These features are called classifiable features. For example, in a landscape image sample, the feature value of the "sand and gravel" feature has a significant impact on whether the landscape image sample will be labeled "desert." A feature may also be a classifiable feature for multiple labels, and a label may have multiple classifiable features. For example, the presence of the features "flower" and "butterfly" means that a landscape image is likely to be labeled "grassland" or "garden," meaning these two labels share classifiable features. However, for the label "desert," obviously "flower" and "butterfly" are not its classifiable features; that is, "desert" and the aforementioned two labels do not share classifiable features. This embodiment will utilize label association to construct such relationships. Currently, some studies on LDL have constructed label association terms, such as using the Pearson correlation coefficient to dynamically describe the correlation between labels and applying this correlation to the distance of the correlation parameter column. However, this approach leads to the label association columns in the parameter matrix becoming similar, failing to reflect the differences between labels. The problem of convergence of label association columns in the parameter matrix is ​​improved by constraining the association to the model output. However, this term is designed for linear regression output models and is not suitable for maximum entropy output models. Considering the above problems, this embodiment combines the Pearson correlation coefficient to design a label association term to constrain the output of the maximum entropy model, as shown in Equation (7).

[0060]

[0061] Where p ij The representation of sample x i The predicted value of the descriptive power of the j-th label, R jl The Pearson tendency relationship between the j-th and l-th labels is represented by the formula shown in (8).

[0062]

[0063] pear(θ ·j ,θ ·l ) represents the Pearson correlation coefficient between the j-th and l-th columns in θ, and the calculation formula is shown in (9).

[0064]

[0065] In the formula This represents the mean of the j-th column of parameter θ.

[0066] Based on the above sparse terms and labeled association terms, we can learn the common features and shared class feature relationships.

[0067] (3) Weighted characteristic terms

[0068] Many existing studies on feature selection utilize mutual information analysis. Mutual information can be used to describe the dependence between features and labels, and this metric can be used to select features with better classification performance. Mutual information is the difference between the information entropy and conditional entropy of two random variables, as shown in Equation (10).

[0069] I(X,Y)=H(Y)-H(Y|X), (10)

[0070] in The information entropy of variable Y, This represents the conditional entropy of Y given the variable X.

[0071] This formula requires calculating the probability, joint probability, and conditional probability of each possible value of the variable. For discrete variables, the probability calculation can be replaced by the frequency of the value. In multi-label learning, this formula can be used directly to calculate the mutual information between discrete features and labels. However, for label distribution learning, the values ​​of the labels are continuous, making it difficult to directly calculate the probability of a certain value. In addition, features themselves are divided into continuous and discrete types, and features with continuous values ​​will have the same problem. Therefore, it is difficult to use this method directly to calculate the mutual information on the label distribution data. In this embodiment, the calculation of mutual information between continuous variables is implemented using a method based on the k-nearest neighbor idea. The continuous feature vector and the label distribution vector are substituted into the formula as shown in (11).

[0072]

[0073] In the formula, t represents the number of nearest neighbors; n represents the number of samples; for each sample x i ,by Indicates sample x i With sample x i′ The distance; in z i x represents i The distance to its t-th nearest sample, then This indicates that the condition |x is satisfied. im -x i′m | <z i The number of samples i′=1,2,...,n This indicates that the condition |d is satisfied. il -d i′l | <zi The number of samples i′=1,2,...,n; ψ(g) represents the digamma function, and its recursive calculation method is shown in equation (12).

[0074]

[0075] Where C = 0.5772156... represents Euler's constant.

[0076] For the calculation of mutual information between discrete feature vectors and label distribution vectors, the calculation formula in this embodiment is set as shown in equation (13).

[0077]

[0078] In the formula This indicates that in the discrete eigenvector x ·r Above and sample x i Let s be the number of samples with the same value. i Indicates the relationship with sample x i Samples with the same discrete eigenvalues ​​and x i The distance to the nearest t-th sample, then h i Indicates the relationship with sample x i The distance is less than s i The number of samples. For the t values ​​in equations (11) and (13), the smaller the value, the greater the statistical error and the smaller the systematic error, and the larger the value, the opposite. In this embodiment, t = 3 is taken.

[0079] The mutual information matrix M∈R can be calculated based on equations (11) and (13). d×c M kj This represents the mutual information value between the k-th feature and the j-th label. This value measures the dependence between them; a larger value indicates a stronger correlation. Let ||M| represent the mutual information value between the k-th feature and the j-th label. k ||2 represents the mutual information weight of the k-th feature in the entire label space. The higher the weight, the stronger the classification ability of the feature. In order to select the feature with stronger classification ability and remove redundant features, the weighted feature terms are constructed using the mutual information matrix as shown in Equation (14).

[0080]

[0081] In the formula, diag(g) represents taking the elements from the diagonal of the matrix, where the diagonal elements represent the weight differences between the corresponding feature rows of the parameter matrix and the mutual information matrix. This term allows the feature weight ranking relationship obtained from the mutual information matrix to affect the parameter matrix, so that features with higher mutual information weights are ranked higher in the feature ranking, thus selecting features with stronger classification ability and filtering redundant features.

[0082] Substituting the common feature learning of formula (6), the shared category feature relationship learning of formula (7), and the weighted feature relationship learning of formula (14) into formula (4), the final objective function can be obtained as shown in formula (15).

[0083]

[0084] S2: By solving the objective function, the magnitude of the mapping parameters corresponding to each feature in the feature space is obtained.

[0085] Equation (15) is a convex function, which can be solved directly using a convex optimization algorithm. The quasi-Newton method BFGS is widely used in solving convex optimization problems because of its strong numerical stability and the fact that it does not require the calculation of the inverse matrix of the Hessian matrix as in Newton's method, making it highly efficient. This embodiment also uses this algorithm to solve the problem.

[0086] For the objective function shown in equation (15), its iterative form can be approximated by its second-order Taylor expansion, as shown in equation (16).

[0087]

[0088] Where Δ=θ (l+1) -θ (l) This indicates the amount of parameter change between two iterations. and H(θ) (l) Let p represent the gradient and Hessian matrix at the current position, respectively. (l) The current position indicates the search direction, and the calculation formula is shown in (17).

[0089]

[0090] Newton's method uses the search direction to update the parameters in the iterative equation (17).

[0091] θ (l+1) =θ (l) +α (l) p (l) (18)

[0092] In the formula α (l) This represents the search step size at the current position. This step size is found using the line search method from the step sizes that satisfy the strong Wolfe condition. The strong Wolfe condition is the value that simultaneously satisfies inequalities (19) and (20).

[0093]

[0094]

[0095] Among them 0 <c1<c2<1。

[0096] Equation (17) has a high time cost in calculating the inverse of the Hessian matrix. To avoid this problem, this embodiment uses an iterative update matrix B to approximate the inverse of the Hessian matrix. Its iterative formula is shown in Equation (21).

[0097] B (l+1) =(I-ρ (l) s (l) (u (l) ) T B (l) (I-ρ (l) u (l) (s (l) ) T )+ρ (l) s (l) (s (l) ) T , (twenty one)

[0098] Where s (l) =θ (l+1) -θ (l) ,

[0099] The formula for calculating the search direction is replaced by equation (22).

[0100]

[0101] The calculation process of the BFGS algorithm is related to the first gradient of the objective function. In this embodiment, the first gradient of the objective function can be expressed as equation (23).

[0102]

[0103] Where p iz Indicates sample x i The z-th predicted label, the label association term with respect to θ sz gradient U(θ) sz Expand as shown in equation (24).

[0104]

[0105] The BFGS algorithm optimizes the objective function according to the above iterative formula to obtain the optimal parameter θ. This process of solving and returning the parameter is described in pseudocode as Algorithm 1 in Table 2. After obtaining the parameter θ, the weights of each feature are sorted in descending order and the feature with higher weight is selected to complete the feature selection task of label distribution learning.

[0106] Table 2

[0107]

[0108]

[0109] S3: Feature selection based on the size of the mapping parameters.

[0110] The larger the mapping parameter value, the more important the feature. You can select the top N features with the largest mapping parameters according to your needs.

[0111] Example 2:

[0112] The present invention also provides a feature selection terminal device based on tag association and weighted features, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the method embodiment described above in Embodiment 1 of the present invention.

[0113] Furthermore, as an executable solution, the feature selection terminal device based on label association and weighted features can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The feature selection terminal device based on label association and weighted features may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described composition of the feature selection terminal device based on label association and weighted features is merely an example and does not constitute a limitation on the feature selection terminal device based on label association and weighted features. It may include more or fewer components than described above, or combine certain components, or different components. For example, the feature selection terminal device based on label association and weighted features may also include input / output devices, network access devices, buses, etc., and this embodiment of the invention does not limit this.

[0114] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the feature selection terminal device based on tag association and weighted features, connecting all parts of the device via various interfaces and lines.

[0115] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the feature selection terminal device based on tag association and weighted features by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0116] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the embodiments of the present invention.

[0117] If the module / unit integrated into the feature selection terminal device based on tag association and weighted features is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media, etc.

[0118] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A feature selection method based on label association and weighted features, characterized in that, include: A maximum entropy model is used to construct an objective function for feature selection on multi-labeled data containing multiple labels: in, The mapping parameter matrix represents the correlation between features and labels. Represents the loss function. Represents sparse terms, Indicates the associated item. Represents the weighted characteristic terms, , and These represent the balance parameters for sparse terms, labeled association terms, and weighted feature terms, respectively. By solving the objective function, the magnitude of the mapping parameters corresponding to each feature in the feature space of the multi-labeled data is obtained, where the multi-labeled data is an image; Feature selection for multi-labeled data based on the size of the mapping parameters; Among them, the sparse term uses the L21 norm to learn features; In solving for the sparse term, it is rewritten as follows: in, The trace of the matrix is ​​represented by the superscript T, which indicates the transpose of the matrix. This represents a diagonal matrix whose diagonal elements are... diagonal matrix, Represents a diagonal matrix The k Line 1 k List, Represents the mapping parameter matrix The k OK, Indicates the feature number, Represents the characteristic number, Represents the L2 norm. Represents a small positive number; The formula for calculating the associated item is: in, Indicates sample No. The predicted value of the descriptiveness of each label. Indicates sample No. The predicted value of the descriptiveness of each label. Indicates the first The first mark and the first Pearson tendency relationship of 1 marker, Indicates the sample number. n Indicates the number of samples. and All of these represent the serial number of the marker. c Indicates the number of tags; Among them, the weighted feature term is measured by the mutual information of the label distribution; Mutual information between continuous feature vectors and label distribution vectors The following formula is used for calculation: in, Indicates that all samples are in the first... m A vector consisting of the values ​​of continuous features. Indicates that all samples are in the first... l A vector consisting of the values ​​of each tag. Represents the nearest neighbor number. Indicates the number of samples. This indicates that the condition is met. The number of samples, Indicates the first i The sample at the th m The values ​​of continuous features, Indicates the first The sample at the th m The values ​​of continuous features, Indicates the first i Sample To its first The distance to the nearest sample, This indicates that the condition is met. The number of samples, Indicates the first i Sample In the The descriptiveness of each tag, Indicates the first Sample In the The descriptiveness of each tag, Represents the digamma function; Mutual information between discrete feature vectors and label distribution vectors The following formula is used for calculation: in, Let represent the vector formed by the values ​​of the r-th discrete feature of all samples. In vector Above and sample Let the number of samples with the same value be denoted as . Representation and Sample Among samples with the same discrete feature values ​​and samples The most recent The distance of each feature, then Representation and Sample The distance is less than The number of samples; Mutual information between continuous feature vectors and label distribution vectors Mutual information between discrete feature vectors and label distribution vectors Construct a mutual information matrix , the elements Indicates the first The first feature and the second The mutual information value of each tag; Based on mutual information matrix Construct weighted feature terms : in, This indicates taking the elements from the diagonal of the matrix. The superscript T denotes the L2 norm, and the superscript T denotes the transpose of the matrix. The loss function is constructed based on KL divergence.

2. The feature selection method based on label association and weighted features according to claim 1, characterized in that: The objective function is solved using the quasi-Newton method.

3. A feature selection terminal device based on label association and weighted features, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 2.

4. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 2.

Images