An Unsupervised Feature Selection Method, Device, Equipment and Storage Medium

Abstract

The present invention discloses an unsupervised feature selection method, apparatus, device, and storage medium, which relate to the field of dimensionality reduction in machine learning, and solve the problem that the finally selected features in feature selection are not the optimal features due to the influence of bad features such as noise and redundancy contained in the original data. The unsupervised feature selection method specifically includes: obtaining the initial feature data of each of the n target objects, determining the first similarity between any two target objects, and constructing a first similarity matrix; constructing a second similarity matrix based on the first similarity matrix, performing iterative learning on the second similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix, and determining target feature data according to the initial feature data and the target projection matrix. The present invention integrates the two processes of constructing a similarity matrix and feature selection into a unified framework, and can obtain an optimal similarity matrix.

Description

Technical Field

[0001] The present invention relates to the field of dimensionality reduction in machine learning, and in particular, to an unsupervised feature selection method, apparatus, device, and storage medium. Background Art

[0002] Feature selection is a process of selecting some of the most representative features from the original features to reduce the data dimension, and the final feature set is a subset of the original features. This method can not only greatly reduce the time required to perform learning tasks, but also improve the corresponding performance. Therefore, feature selection is widely applied to various machine learning tasks, such as classification and clustering. Since graph models can simulate the manifold structure of data and have good data expression ability, the unsupervised feature selection technology based on graphs has received extensive attention and research in recent years.

[0003] The prior art combines the feature selection and the local structure learning process, and can well utilize the learned information to mitigate the adverse effects of noise and redundant features in the original data. However, the constructed similarity matrix is still constructed based on the original data fundamentally. Since the original data inevitably contains features such as noise and redundancy, the quality of the finally obtained similarity matrix will be affected to a certain extent, resulting in the finally selected features not being optimal. Summary of the Invention

[0004] The present invention provides an unsupervised feature selection method, apparatus, device, and storage medium, which solves the problem that the finally selected features in feature selection are not optimal features due to the influence of adverse features such as noise and redundancy contained in the original data.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides an unsupervised feature selection method, and the method includes:

[0007] Obtain the initial feature data of each of the n target objects input by the user, where the initial feature data includes d types of feature data, and n and d are positive integers greater than 1;

[0008] Determine the first similarity of any two of the n target objects according to the initial feature data of the any two of the n target objects;

[0009] Construct a first similarity matrix according to the first similarity of the any two of the target objects and a regularization parameter;

[0010] Based on the initial feature data of multiple candidate objects input by the user, determine the second similarity between any two candidate objects, and construct a second similarity matrix according to the second similarity between any two candidate objects and the regularization parameter. The multiple candidate objects are completely different or partially different from the n target objects;

[0011] Using a loss function, iteratively learn the second similarity matrix based on the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix. The target similarity matrix is the matrix with the highest similarity to the first similarity matrix;

[0012] Based on the initial feature data and the target projection matrix, determine target feature data. The target feature data includes k types of feature quantities, and k is a positive integer less than d.

[0013] Combined with the first aspect, in another possible implementation, the using a loss function to iteratively learn the second similarity matrix based on the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix includes:

[0014] Using formula (1), determine the target similarity matrix and the corresponding target projection matrix. Formula (1) satisfies the following conditions:

[0015]

[0016] where A is the first similarity matrix, S is the second similarity matrix, W is the target projection matrix, and ||·|| F represents the Frobenius norm of the matrix; k represents the number of features in the target projection matrix, and both λ and α are parameters used to adjust the balance. L S = D S -(S + S T ) / 2 represents the Laplacian matrix of the second similarity matrix S, where D S represents the diagonal matrix, and the i-th diagonal element is Σ j (s ij + s ji ) / 2.

[0017] Combined with the first aspect, in another possible implementation, the target projection matrix satisfies the conditions of the following formula (2):

[0018] ||W|| 2,0 = k (2).

[0019] Combined with the first aspect, in another possible implementation, the constructing the first similarity matrix according to the first similarity between any two target objects and the regularization parameter includes:

[0020] Using formula (3), construct the first similarity matrix, and formula (3) satisfies the following conditions:

[0021]

[0022] where a ij represents the similarity between the i-th target object and the j-th target object, γ≥0 represents a regularization parameter, and the regularization term is used to avoid trivial solutions.

[0023] In a second aspect, the present invention provides an unsupervised feature selection device, and the device includes:

[0024] An acquisition module, configured to acquire initial feature data of each target object in n target objects input by a user, where the initial feature data includes d types of feature data, and n and d are positive integers greater than 1;

[0025] A first determination module, configured to determine a first similarity degree between any two target objects according to the initial feature data of any two target objects among the n target objects;

[0026] A first construction module, configured to construct a first similarity matrix according to the first similarity degree between any two target objects and a regularization parameter;

[0027] A second construction module, configured to determine a second similarity degree between any two candidate objects according to the initial feature data of a plurality of candidate objects input by a user, and construct a second similarity matrix according to the second similarity degree between any two candidate objects and the regularization parameter, where the plurality of candidate objects are completely different or partially different from the n target objects;

[0028] An iterative learning module, configured to use a loss function to perform iterative learning on the second similarity matrix according to the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix, where the target similarity matrix is the matrix with the highest similarity degree to the first similarity matrix;

[0029] A second determination module, configured to determine target feature data according to the initial feature data and the target projection matrix, where the target feature data includes k types of feature quantities, and k is a positive integer less than d.

[0030] Combined with the second aspect, in another possible implementation manner, the iterative learning module is specifically configured to:

[0031] Use formula (1) to determine the target similarity matrix and the corresponding target projection matrix, and formula (1) satisfies the following conditions:

[0032]

[0033] Among them, A is the first similarity matrix, S is the second similarity matrix, W is the target projection matrix, and ||·|| F represents the Frobenius norm of the matrix; k represents the number of eigenvalues in the target projection matrix, and both λ and α are parameters used to adjust the balance. L S = D S -(S + S T ) / 2 represents the Laplacian matrix of the second similarity matrix S, where DS represents the diagonal matrix, and the i-th diagonal element is ∑ j (s ij + s ji ) / 2.

[0034] Combined with the second aspect, in another possible implementation manner, the target projection matrix satisfies the conditions of the following formula (2):

[0035] ||W|| 2,0 = k (2).

[0036] Combined with the second aspect, in another possible implementation manner, the first construction module is specifically used for:[[]]

[0037] Using formula (3), construct the first similarity matrix, and the formula (3) satisfies the following conditions:

[0038]

[0039] where a ij represents the similarity between the i-th target object and the j-th target object, γ ≥ 0 represents the regularization parameter, and the regularization term is used to avoid the trivial solution.

[0040] In a third aspect, the present invention provides an electronic device, which includes a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the above-mentioned unsupervised feature selection method.

[0041] In a fourth aspect, the present invention provides a computer-readable storage medium, including computer instructions. When the computer instructions run on an electronic device, the electronic device is caused to execute the unsupervised feature selection method described in any one of claims 1-4. Description of the Drawings

[0042] Figure 1 is one of the flowcharts of an unsupervised feature selection method provided by an embodiment of the present invention;

[0043] Figure 2 is the second flowchart of an unsupervised feature selection method provided by an embodiment of the present invention;

[0044] Figure 3 It is a comparison chart of the effects of different sparse learning strategies for feature selection;

[0045] Figure 4 It is the clustering accuracy of the embodiments of the present invention on 6 benchmark datasets such as JAFFE and Madelon. Specific implementation manners

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0047] Hereinafter, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present disclosure, unless otherwise stated, the meaning of "a plurality" is two or more. Additionally, the use of "based on" or "according to" means open and inclusive, because a process, step, calculation, or other action "based on" or "according to" one or more of the stated conditions or values may in practice be based on additional conditions or values beyond those stated.

[0048] With the prosperity of the information age and the maturity of big data technology, a large amount of data has been generated in many fields and industries. Through analysis, it is found that the development trend of these data is towards larger-scale high-dimensional data. In practical applications, these high-dimensional data inevitably contain various bad features such as redundancy and noise. The existence of a large number of bad features will lead to problems such as the curse of dimensionality. In addition, when directly processing these high-dimensional data, we will inevitably encounter various difficulties, such as a long learning task time and a high storage cost. Therefore, how to process these high-dimensional data has attracted the attention of a wide range of scholars, and scholars have proposed a variety of different data processing technologies. These technologies can be generally classified into two categories: feature extraction and feature selection.

[0049] Among them, feature extraction combines the original features in a certain form to construct new features so as to achieve the purpose of dimensionality reduction. This type of technology does not simply select features, so the resulting result is not a subset of the original features and lacks meaningful explanations. According to the usage of data labels, feature selection methods can be divided into: supervised, semi-supervised, and unsupervised. For supervised and semi-supervised cases, the corresponding performance will be affected by the label information used to a certain extent. However, in practical applications, the obtained high-dimensional data usually has no label information, and the cost of annotating these data is quite high. In addition, graph-based models have good data representation capabilities and can simulate the manifold structure of data; therefore, in recent years, unsupervised feature selection techniques based on graphs have received extensive attention and research.

[0050] In life, high-dimensional data exists in various different fields. Directly using this high-dimensional data will not only increase the storage cost, but also introduce irrelevant information such as redundancy and noise, thereby reducing the performance of the learning task.

[0051] The prior art proposes a technique that combines the feature selection and local structure learning processes. This technique can make good use of the learned information to reduce the impact of noise and redundant bad features in the original data. The disadvantage of this technique is that the quality of the constructed similarity matrix is not high, resulting in the final selected features not being optimal. The reason is that although the above technique combines the two processes of constructing the similarity matrix and selecting features and executes them simultaneously, fundamentally speaking, the similarity matrix is still constructed based on the original data. However, the original data inevitably contains noise and redundant features. Therefore, the quality of the obtained similarity matrix will be affected to a certain extent, resulting in the features selected by the feature selection process not being optimal.

[0052] To further improve the quality of the similarity matrix and select the optimal features at the same time, instead of directly using the similarity matrix constructed based on the original data, we re-learn a new similarity matrix that can adaptively maintain the manifold structure of the data, that is, it can adaptively adjust the allocation of neighbors. The newly learned similarity matrix can be continuously updated and iterated along with the feature selection process. After several iterations, a similarity matrix that is most suitable for the feature selection process can be learned, and ultimately better features will also be selected. In addition, directly using the l 2,0 norm to sparsify the projection matrix so that it has exactly a few non-zero rows, and at the same time, feature selection can be realized in a combined form.

[0053] In order to solve the problem that the final selected features in feature selection are not the optimal features due to the influence of bad features such as noise and redundancy contained in the original data, the embodiments of the present invention provide an unsupervised feature selection method, device, equipment, and storage medium.

[0054] The execution entity of the unsupervised feature selection method provided by the embodiments of the present invention is an unsupervised feature selection device. The unsupervised feature selection device may be a computer device, or the CPU in the device, or the client of the device. The embodiments of the present invention do not make any limitations. In the embodiments of the present invention, the unsupervised feature selection method in the embodiments of the present invention is introduced by taking the unsupervised feature selection device as a computer device as an example.

[0055] Figure 1 As shown in FIG. Figure 1 one of the flowcharts of an unsupervised feature selection method provided by the embodiments of the present invention, the method may include the following steps 101 to 106:

[0056] Step 101, obtain the initial feature data of each of the n target objects input by the user. The initial feature data includes d types of feature data, and n and d are positive integers greater than 1.

[0057] It can be understood that the target objects may be objects of the same type with different attributes. For example, the target objects may be different students, different employees, different images, or different items. The d types of feature data may be the attribute values of different attributes of the target objects. For example, when the target objects are students, the d types of feature data may include the attribute values of attributes such as the height, weight, age, and hobbies of the students.

[0058] Optionally, in some embodiments, when the user inputs the initial feature data of multiple objects in the Excel table of the computer device, the computer device may, in response to the user's input operation, obtain the initial feature data of the multiple objects, and determine n target objects from the multiple objects according to the pre-stored screening rules. The screening rules may include the preset ranges of different types of feature data, and the preset ranges are the ranges where the representative feature data is located. In this way, by initially screening out a part of the representative target objects from the objects input by the user, the data volume for subsequent feature selection can be reduced, thereby improving the efficiency of feature selection.

[0059] Step 102, determine the first similarity of any two of the n target objects according to the initial feature data of any two of the n target objects.

[0060] Taking any two target objects including a first object and a second object as an example, the computer device may respectively determine the similarities of the feature data of the same type of the first object and the second object, and then use the sum of all the similarities as the first similarity of the first object and the second object.

[0061] Step 103, construct a first similarity matrix according to the first similarity of any two target objects and the regularization parameter.

[0062] Optionally, construct a first similarity matrix according to the first similarity between any two target objects and a regularization parameter, including:

[0063] Use formula (3) to construct the first similarity matrix, and formula (3) satisfies the following conditions:

[0064]

[0065] where a ij represents the similarity between the i-th target object and the j-th target object, γ≥0 represents the regularization parameter, and the regularization term is used to avoid trivial solutions.

[0066] Specifically, input the initial feature data where d represents the number of features of the corresponding target object in the initial feature data, and n represents the total number of target objects. Based on the original data, use the above formula (3) to obtain the first similarity matrix for reflecting the first similarity between different target objects

[0067] Step 104: Determine the second similarity between any two candidate objects according to the initial feature data of multiple candidate objects input by the user, and construct a second similarity matrix according to the second similarity between any two candidate objects and the regularization parameter. The multiple candidate objects are completely different or partially different from the n target objects.

[0068] It can be understood that the computer device can determine multiple candidate objects that are completely different or partially different from the n target objects from the multiple objects output by the user, and construct a second similarity matrix based on the multiple candidate objects. The specific process of constructing the second similarity matrix can refer to the relevant description of constructing the first similarity matrix in the above step 103, which will not be elaborated here.

[0069] Specifically, on the basis of the first similarity matrix A obtained in the above steps, learn another second similarity matrix S, and this second similarity matrix can adaptively adjust the allocation of neighbors. The second similarity matrix S is calculated according to formula (4):

[0070]

[0071] where ||·|| F represents the Frobenius norm of the matrix.

[0072] Step 105: Use the loss function to perform iterative learning on the second similarity matrix according to the first similarity matrix to obtain the target similarity matrix and the corresponding target projection matrix. The target similarity matrix is the matrix with the highest similarity to the first similarity matrix.

[0073] Step 106: Determine target feature data according to the initial feature data and the target projection matrix. The target feature data includes k types of feature quantities, where k is a positive integer less than d.

[0074] Specifically, a low-dimensional embedding matrix is obtained by introducing a graph embedding process. m represents the dimension of the low-dimensional embedding subspace. The graph embedding function can be described as:

[0075]

[0076] where C represents a constant matrix, B is a constraint matrix used to avoid trivial solutions, and L S = D S -(S + S T ) / 2 represents the Laplacian matrix of the second similarity matrix S, where D S represents the diagonal matrix, and the i-th diagonal element is ∑ j (s ij + s ji ) / 2.

[0077] For the initial feature data The traditional feature selection method is realized by solving the following objective function:

[0078]

[0079] where is the label matrix, is the projection matrix, β is the sparse adjustment parameter, and ||·|| 2,1 represents the l 2,1 norm of the matrix. Here, the l 2,p norm of the projection matrix W can be expressed as: where w i represents the i-th row of the projection matrix W, and ||w i ||2 represents the 2-norm of the vector w i .

[0080] In reality, the generated data is often unlabeled, and labeling this data is both time-consuming and laborious. Therefore, people tend to adopt an unsupervised strategy. In the embodiments of the present invention, the low-dimensional embedding matrix F generated by the above graph embedding process can be used to replace the label matrix Y. In addition, although the second term in Equation (4) can be approximated as a convex problem through l 2,1 norm regularization and is more convenient to solve, the selected feature subset is still not optimal. There is literature indicating that a combination composed of several optimal single features may not be the best, and directly selecting features in the form of a combination has better effects than selecting single features. Figure 2FIG. 2 is a second flowchart of an unsupervised feature selection method provided by an embodiment of the present invention. Figure 3 FIG. is a comparison chart of the effects of different sparse learning strategies for feature selection; combined Figure 2 with Figure 3 As shown in FIG., the projection matrix sparsified by the l 2,1 -norm almost retains all features. Next, the corresponding scores of these features need to be calculated, sorted according to the scores, and finally the screening is completed. By constraining the projection matrix with the l 2,0 -norm, the projection matrix can have exactly k non-zero rows, corresponding to k features respectively. In the subsequent learning process, if there is a better set, the set will be directly replaced. Therefore, the l 2,0 -norm constraint can be selected to directly select features in the form of a combination. The specific feature selection framework is as follows:

[0081]

[0082] where k represents the number of finally selected features.

[0083] In summary, the construction of the similarity matrix and the feature selection are integrated into a unified framework by using formula (1) to ensure that the learned similarity matrix is optimal for feature selection. At the same time, by introducing the sparse learning strategy with the l 2,0 -norm constraint, the target projection matrix has k non-zero rows. Finally, a feature combination can be directly selected instead of a single feature.

[0084] The present invention is mainly to select the most representative features from the original features, and the finally selected feature subset can be used to replace the high-dimensional data in various learning tasks. This strategy can prevent the curse of dimensionality by reducing the number of features, and at the same time can reduce the time of learning tasks, reduce overfitting, and enhance the generalization ability of the model.

[0085] Optionally, the target projection matrix satisfies the condition of the following formula (2):

[0086] ||W|| 2,0 = k (2)

[0087] Figure 4 FIG. shows the clustering accuracy of the embodiments of the present invention on 6 benchmark data sets such as JAFFE and Madelon. As Figure 4 shown, a large number of experiments were carried out by performing clustering tasks on 6 public benchmark data sets, and the results prove the superiority of the present invention.

[0088] An embodiment of the present invention provides an unsupervised feature selection device, which may include: an acquisition module, configured to acquire initial feature data of each of n target objects input by a user, where the initial feature data includes d types of feature data, and n and d are positive integers greater than 1;

[0089] A first determination module, configured to determine a first similarity between any two of the n target objects according to the initial feature data of any two of the n target objects;

[0090] A first construction module, configured to construct a first similarity matrix according to the first similarity between any two of the target objects and a regularization parameter;

[0091] A second construction module, configured to determine a second similarity between any two of a plurality of candidate objects according to the initial feature data of the plurality of candidate objects input by the user, and construct a second similarity matrix according to the second similarity between any two of the candidate objects and the regularization parameter, where the plurality of candidate objects are completely different or partially different from the n target objects;

[0092] An iterative learning module, configured to perform iterative learning on the second similarity matrix according to the first similarity matrix by using a loss function to obtain a target similarity matrix and a corresponding target projection matrix, where the target similarity matrix is the matrix with the highest similarity to the first similarity matrix;

[0093] A second determination module, configured to determine target feature data according to the initial feature data and the target projection matrix, where the target feature data includes k types of feature data, and k is a positive integer less than d.

[0094] Optionally, the iterative learning module is specifically configured to:

[0095] Use formula (1) to determine the target similarity matrix and the corresponding target projection matrix, where formula (1) satisfies the following conditions:

[0096]

[0097] where A is the first similarity matrix, S is the second similarity matrix, W is the target projection matrix, and ||·|| F represents the Frobenius norm of the matrix; k represents the number of features in the target projection matrix, that is, k represents the number of non-zero rows of the target projection matrix, that is, k is the number of finally selected features, and both λ and α are parameters used to adjust the balance, and L S = D S - S + S T ) / 2 represents the Laplacian matrix of the second similarity matrix S, where D S represents the diagonal matrix, and the i-th diagonal element is ∑ j (s ij + sji ) / 2。

[0098] Optionally, the target projection matrix satisfies the condition of the following formula (2):

[0099] ||W|| 2,0 = k (2);

[0100] That is:

[0101] Optionally, the first construction module is specifically used for:

[0102] Using formula (3), construct the first similarity matrix, and formula (3) satisfies the following conditions:

[0103]

[0104] where a ij represents the similarity between the i-th sample and the j-th sample, γ≥0 represents the regularization parameter, and the regularization term is used to avoid trivial solutions.

[0105] Of course, the unsupervised feature selection device provided by the embodiments of the present invention includes but is not limited to the above modules.

[0106] The unsupervised feature selection device provided by the embodiments of the present invention is used to execute the above-mentioned unsupervised feature selection method, so it can achieve the same effect as the above-mentioned unsupervised feature selection method.

[0107] This embodiment also provides a device, which may include a processor and a memory. Among them, the memory stores a computer program, and the computer program is adapted to be loaded and executed by the processor to execute the above-mentioned unsupervised feature selection method of this embodiment.

[0108] The above device may be a computer terminal, a mobile terminal or a server, and the above device may also participate in constituting the device provided by the embodiments of the present invention.

[0109] This embodiment also provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are loaded and executed by the processor to execute the above-mentioned unsupervised feature selection method of this embodiment.

[0110] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An unsupervised feature selection method, characterized in that, Including: Obtain the initial feature data of each of the n target objects input by the user, where the initial feature data includes d types of feature data, and n and d are positive integers greater than 1; the target object is an image; Determine the first similarity between any two of the n target objects according to the initial feature data of the any two target objects; Construct a first similarity matrix according to the first similarity between the any two target objects and the regularization parameter; Determine the second similarity between any two candidate objects according to the initial feature data of the multiple candidate objects input by the user, and construct a second similarity matrix according to the second similarity between the any two candidate objects and the regularization parameter, where the multiple candidate objects are completely different or partially different from the n target objects; Use a loss function to perform iterative learning on the second similarity matrix according to the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix, where the target similarity matrix is the matrix with the highest similarity to the first similarity matrix; Determine target feature data according to the initial feature data and the target projection matrix, where the target feature data includes k types of feature quantities, and k is a positive integer less than d.

2. The unsupervised feature selection method according to claim 1, wherein The using a loss function to perform iterative learning on the second similarity matrix according to the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix includes: Use formula (1) to determine the target similarity matrix and the corresponding target projection matrix, and the formula (1) satisfies the following conditions: (1) Among them, is the first similarity matrix, is the second similarity matrix, is the target projection matrix, is the low-dimensional embedding matrix, , m represents the dimension of the low-dimensional embedding subspace, represents the Frobenius norm of the matrix; k represents the number of eigenvalues in the target projection matrix, and are both parameters used to adjust the balance, represents the Laplacian matrix of the second similarity matrix , where represents the diagonal matrix, and the i-th diagonal element is .

3. The unsupervised feature selection method according to claim 1 or 2, characterized in that The constructing a first similarity matrix according to the first similarity between the any two target objects and the regularization parameter includes: Use formula (3) to construct the first similarity matrix, and the formula (3) satisfies the following conditions: (3) Among them, represents the similarity between the i-th target object and the j-th target object, represents the regularization parameter, and the regularization term is used to avoid trivial solutions.

4. An unsupervised feature selection device, characterized in that, Including: An acquisition module for obtaining the initial feature data of each of the n target objects input by the user, where the initial feature data includes d types of feature data, and n and d are positive integers greater than 1; the target object is an image; A first determination module for determining the first similarity between any two of the n target objects according to the initial feature data of the any two target objects; A first construction module for constructing a first similarity matrix according to the first similarity between the any two target objects and the regularization parameter; A second construction module for determining the second similarity between any two candidate objects according to the initial feature data of the multiple candidate objects input by the user, and constructing a second similarity matrix according to the second similarity between the any two candidate objects and the regularization parameter, where the multiple candidate objects are completely different or partially different from the n target objects; An iterative learning module for using a loss function to perform iterative learning on the second similarity matrix according to the first similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix, where the target similarity matrix is the matrix with the highest similarity to the first similarity matrix; A second determination module, configured to determine target feature data according to the initial feature data and the target projection matrix, where the target feature data includes k types of feature quantities, and k is a positive integer less than d.

5. The unsupervised feature selection device according to claim 4, characterized in that, The iterative learning module is specifically configured to: Use formula (1) to determine the target similarity matrix and the corresponding target projection matrix, and formula (1) satisfies the following conditions: (1) Among them, is the first similarity matrix, is the second similarity matrix, is the target projection matrix, is the low-dimensional embedding matrix, , m represents the dimension of the low-dimensional embedding subspace, represents the Frobenius norm of the matrix; k represents the number of eigenvectors in the target projection matrix, and are both parameters used to adjust the balance, represents the Laplacian matrix of the second similarity matrix , where represents the angular matrix, and the i-th diagonal element is .

6. The unsupervised feature selection device according to claim 5, wherein The first construction module is specifically configured to: Use formula (3) to construct the first similarity matrix, and formula (3) satisfies the following conditions: (3) Among them, represents the similarity between the i-th target object and the j-th target object, represents the regularization parameter, and the regularization term is used to avoid trivial solutions.

7. An electronic device, characterized in that, The device includes a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement an unsupervised feature selection method according to any one of claims 1-3.

8. A computer-readable storage medium, characterized in that, Including computer instructions, when the computer instructions run on an electronic device, the electronic device is caused to execute the unsupervised feature selection method according to any one of claims 1-3.

Images