A multi-label feature selection method, device and medium based on dynamic flow labels

Abstract

The application relates to a dynamic flow label-based multi-label feature selection method, equipment and medium. In the method, implicit label importance of a dynamic flow label is formalized into a multinomial distribution on a dynamic label space by nearest neighbor reconstruction on a feature space; label correlation is seamlessly integrated into a feature-related objective function, and the influence of label correlation on features is quantified; and a feature conversion technology is constructed to fuse generic features corresponding to each dynamic flow label. The application applies joint learning of potential label importance and label correlation to a dynamic flow label-oriented multi-label feature selection method, and provides a high-quality feature sequence for model training.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a multi-label feature selection method, device and medium based on dynamic flow labels. Background Art

[0002] Multi-label feature selection has attracted widespread attention as an effective dimensionality reduction technique because it can solve the problem of samples belonging to multiple labels and the high dimensionality of the feature space. Compared with traditional feature selection, multi-label feature selection can better adapt to the needs of practical application scenarios.

[0003] In traditional multi-label feature selection research, a basic assumption is to obtain all labels in advance before feature selection. However, in many practical application scenarios, it is obviously unrealistic and costly to obtain all labels. Instead, it is more common that labels may arrive one by one in a streaming mode, and we cannot know the exact number of labels before learning. For example, in event detection tasks, new events that were excluded by early detection systems need to be checked in a timely manner. In many social networking systems, such as Facebook and MySpace posts, hot topics (tags) change every day.

[0004] Currently, several algorithms have attempted to perform feature selection in label streaming environments. The main idea behind these methods is to obtain categorical features for each independent stream label and fuse them together using fusion strategies. While these methods can achieve satisfactory prediction results, they often assume that the streaming labels are unrelated, an idealized assumption that is inconsistent with reality. In practice, we observe that labels are interrelated and influence each other. For example, a photo labeled "grass" and "tent" is likely to be labeled "outdoors." Similarly, if there are labels for "elephant" and "lion," the label "Africa" ​​is likely to be associated with them. While many studies on multi-label learning have demonstrated that leveraging label correlations can provide valuable additional information for learning tasks, they typically exploit label correlations in a static manner and are not directly applicable to the streaming label environment.

[0005] At the same time, it is difficult to capture the different importance of labels when the number of labels increases dynamically over time. Typically, the importance of each label is essentially different. For example, a news document can be annotated with labels such as "sports", "world cup" and "football" at the same time, and the importance of these labels in describing the semantics of the document varies depending on the length of the topic. Similar situations also occur in other tasks such as facial expression recognition and age estimation. Therefore, it is crucial to distinguish and reveal the relative importance of different labels. Currently, many methods have been developed to facilitate feature selection by extracting valuable information such as label importance. However, these methods estimate the relative importance of labels under the assumption that the complete label space is obtained, which cannot be generalized well to streaming data. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a multi-label feature selection method, device and medium based on dynamic flow labels.

[0007] The specific plan is as follows:

[0008] A multi-label feature selection method based on dynamic flow labels includes the following steps:

[0009] S1: Based on the neighborhood, positive neighborhood set, and negative neighborhood set of each sample in the multi-label data in the feature space, calculate the detailed description of the current arrival label relative to each sample;

[0010] S2: Combine the detailed description of each historical arrival label to calculate the relative label importance of the current arrival label relative to each sample;

[0011] S3: Calculate the invisible tag importance of the current arrival tag by combining the relative tag importance of each historical arrival tag;

[0012] S4: Calculate the label correlation and feature correlation influencing factor between the current arrival label and each historical arrival label for each feature in the feature space;

[0013] S5: Based on the label relevance and feature relevance influencing factors, the enhanced feature relevance and feature redundancy of the current arrival label are calculated;

[0014] S6: Taking the maximum difference between the enhanced feature correlation and feature redundancy as the objective function, constructing the optimal feature set, randomly selecting a feature in the feature space after each iteration and adding it to the optimal feature set, outputting the optimal feature set that satisfies the objective function;

[0015] S7: Constructing an initial mapping strategy based on the mapping of each feature in the preferred feature set corresponding to each arrival tag to the feature space; and constructing a final mapping strategy based on the initial mapping strategy of each arrival tag and the importance of the invisible tag.

[0016] S8: sorting the values of each element in the vector after the current arriving label is mapped by the final mapping strategy in descending order, and selecting the top N features with the largest values based on the sorting result to form a selected feature set;

[0017] S9: repeating steps S1-S8 until the current arriving label is the last arriving label in the multi-label data, and then fusing the selected feature sets corresponding to all arriving labels to obtain a final feature set output.

[0018] Further, the relative label importance degree of the current arriving label L t with respect to the sample x i The calculation formula of the fine description degree of the sample x i is as follows:

[0019]

[0020] wherein, represents the positive neighborhood set of the sample x i , and θ F (x i ) represents the neighborhood of the sample x i in the feature space, |. | represents the number of elements in the set, represents the positive label, represents the negative label.

[0021] Further, the relative label importance degree of the current arriving label L t with respect to the sample x i The calculation formula of the fine description degree of the sample x t is as follows:

[0022]

[0023] wherein, represents the fine description degree of the current arriving label L i with respect to the sample x k , A represents the set of all labels currently arriving, |. | represents the number of elements in the set, and k represents the serial number of the historical arriving label, represents the fine description degree of the historical arriving label L i with respect to the sample x t .

[0024] Further, the calculation formula of the implicit label importance Imp(L t ) of the current arriving label L t is as follows:

[0025] ​

[0026] Among them, i represents the serial number of the sample, n represents the total number of samples, Indicates the current arrival label L t Relative to sample x i The relative importance of the label, A represents the set of all currently arrived labels, |.| represents the number of elements in the set, k represents the sequence number of the historically arrived labels, Represents the historical arrival label L k Relative to sample x i The relative importance of labels.

[0027] Furthermore, regarding the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The label correlation Lco(F i , L t , L k ) is calculated as:

[0028] Lco(F i , L t , L k )=I(F i ;L t |L k )-I(F i ;L t )

[0029] Among them, I(F i ;L t |L k ) indicates that the historical arrival label L k Find the characteristic F under the condition i For the current arrival label L t The contribution of I(F i ;L t ) represents feature F i For the current arrival label L t degree of contribution.

[0030] Furthermore, regarding the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The feature correlation factor ω(F i , L t , L k ) is calculated as:

[0031]

[0032] Among them, L SH(.) represents the entropy.

[0033] Further, the enhanced feature relevance Rel(F t , L i , L t ) of the current arriving label L S is calculated as follows:

[0034] Rel(F i , L t , L S ) = I(F i ; L t ) x ω(F i , L t , L S )

[0035] The feature redundancy Red(F t , L i ) of the current arriving label L t is calculated as follows:

[0036]

[0037] wherein I(F i ; L t ) represents the contribution degree of the feature F i to the current arriving label L t , S represents the preferred feature set, |. | represents the number of elements in the set, F j represents the features in the preferred feature set, F i represents the features waiting for selection outside the preferred feature set, I(F i , F j | L t ) represents the contribution degree of the feature F i to the feature F j under the condition of the current arriving label L t , and ω(F i , L t , L S ) represents the feature relevance influence factor between the current arriving label L t and the historical arriving label L k .

[0038] Further, the final mapping strategy of the current arriving label L t is calculated as follows:

[0039]

[0040] wherein L S represents the historical label set, and Lk Indicates the historical arrival label, Imp(L k ) represents the historical arrival label L k Implicit label importance, Imp(L t ) indicates the current arrival label L t The implicit label importance, φ k Represents the historical arrival label L k The initial mapping strategy, φ t Indicates the current arrival label L t Initial mapping strategy.

[0041] A terminal device for multi-label feature selection based on dynamic flow labels 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, the steps of the method described above are implemented.

[0042] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described above in an embodiment of the present invention.

[0043] The present invention adopts the above technical solution to apply the joint learning of potential tag importance and tag correlation to the multi-label feature selection method for dynamic flow tags, providing high-quality feature sequences for model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Shown is a flow chart of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0045] To further illustrate various embodiments, the present invention provides accompanying drawings. These drawings form part of the present disclosure and are primarily used to illustrate the embodiments and, in conjunction with the relevant description in the specification, to explain the operating principles of the embodiments. By referring to these drawings, those skilled in the art will be able to understand other possible implementations and the advantages of the present invention.

[0046] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0047] Example 1:

[0048] The embodiment of the present invention provides a multi-label feature selection method based on dynamic flow labels, such as Figure 1 As shown, the method includes the following steps:

[0049] S1: Based on the neighborhood, positive neighborhood set, and negative neighborhood set of each sample in the multi-label data in the feature space, calculate the detailed description of the current arrival label relative to each sample.

[0050] The embodiment sets each row in the multi-label data as a sample, and the feature space is represented by a set F. Then, for a sample x i , the domain representation on the feature space is θ F (x i ) = {x j | Δ F (x i , x j ) ≤ θ, x j ∈ U}, where θ F (x i ) is the neighborhood granularity centered at the sample x i with a radius of θ, and θ is the standard deviation of the sample x i calculated under all features.

[0051] The embodiment is based on dynamic flow labels, and each label arrives in turn. The coarse description degree of a sample x i to the currently arrived label L t is defined by the consistency degree of the sample in the neighborhood relative to the class label as follows:

[0052]

[0053] wherein, L represents the positive label, and L represents the negative label.

[0054] In order to solve the problem that the degree to which a sample belongs to a label cannot be accurately described, it is necessary to give a clear measurement method to calculate the parameter τ, so as to clearly measure the degree to which a sample belongs to a label. In the embodiment, a calculation method of the fine description degree is proposed. By further studying the basic information of the class label of the sample in the neighborhood, the calculation formula of the fine description degree is defined by combining the positive neighborhood set and the negative neighborhood set .

[0055]

[0056] S2: The relative label importance degree of the currently arrived label to each sample is calculated by combining the fine description degrees of each historical arrived label.

[0057] In the embodiment, the relative label importance degree is calculated by normalization. The calculation formula of the relative label importance degree of the currently arrived label L t to the sample x i is as follows:

[0058] ​​

[0059] Among them, A represents the set of all currently arrived tags, k represents the sequence number of the historically arrived tags, Represents the historical arrival label L k Relative to sample x i The degree of detailed description.

[0060] S3: Calculate the invisible tag importance of the current arrival tag by combining the relative tag importance of each historical arrival tag.

[0061] The importance of the invisible label indicates the importance of the label relative to all samples. Set the current arrival label L t Implicit label importance Imp(L t ) is calculated as:

[0062]

[0063] Among them, i represents the serial number of the sample, n represents the total number of samples, Represents the historical arrival label L k Relative to sample x i The relative importance of labels.

[0064] S4: Calculate the label correlation and feature correlation influencing factor between the current arrival label and each historical arrival label for each feature in the feature space.

[0065] The impact factor based on label correlation and feature correlation can be used to quantify the impact of label relationship on feature correlation.

[0066] (1) Tag relevance

[0067] About the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The label correlation Lco(F i , L t , L k ) is calculated as:

[0068] Lco(F i , L t , L k )=I(F i ;L t |L k )-I(F i ;L t )

[0069] Among them, I(F i ;L t |L k) indicates that the historical arrival label L k Find the characteristic F under the condition i For the current arrival label L t The contribution of I(F i ;L t ) represents feature F i For the current arrival label L t The contribution of Lco(F i , L t , L k ) can be positive, zero or negative. If I(F i ;L t |L k )-I(F i ;L t )>0, indicating that when the label L is known k When the feature pair label L t The amount of information provided increases, in other words, the label L t and L k Relative to feature F i has a positive label correlation. On the contrary, if I(F i ;L t |L k )-I(F i ;L t )<0, indicating that when the label L is known k When the feature pair label L t The amount of information provided is reduced, and the label L t and L k is considered relative to feature F i Has negative label correlation. When I(F i ;L t |L k )-I(F i ;L t )=0, it means the introduction of label L k Will not affect the feature pair label L t The amount of information provided, in other words, the label L t and L k Relative to feature F i is independent.

[0070] (2) Feature correlation influencing factors

[0071] The feature correlation factor quantifies the degree to which the importance of a feature is affected by the label correlation, and sets the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The feature correlation factor ω(F i , Lt , L k ) is calculated as:

[0072]

[0073] Among them, L S represents the historical tag set, and H(.) represents the entropy. The first term in the formula is used to control ω(F i , L t , L k ) is greater than or equal to 0, and the label L involved in the latter item t and L k The entropy of is considered as a normalization factor.

[0074] S5: Based on the tag relevance and feature relevance influencing factors, the enhanced feature relevance and feature redundancy of the current arrival tag are calculated.

[0075] In this embodiment, the current arrival label L t Enhanced feature correlation Rel(F i , L t , L S ) is calculated as:

[0076] Rel(F i , L t , L S )=I(F i ;L t )×ω(F i , L t , L S )

[0077] Current arrival label L t The feature redundancy Red(F i , L t ) is calculated as:

[0078]

[0079] Among them, S represents the preferred feature set, which is initially an empty set; F j represents the features in the preferred feature set, F i For the features waiting to be selected outside the preferred feature set, I(F i , F j |L t ) indicates that at the current arrival tag L t Under the condition of i For feature F j degree of contribution.

[0080] Enhanced feature correlation Rel(F i , Lt , L S ) and the influence of the correlation between the reached labels and the current label on the feature correlation. If two labels L t and L k are completely independent in terms of a feature F i , i.e. ω(F i , L t , L S ) = 1, then Rel(F i , L t , L S ) degenerates to I(F i ; L t ).

[0081] S6: Maximize the difference between the enhanced feature correlation and the feature redundancy as the objective function, construct the preferred feature set, randomly select a feature in the feature space to add to the preferred feature set after each iteration, and output the preferred feature set that meets the objective function.

[0082] Let the difference between the enhanced feature correlation and the feature redundancy be the function score J(F i ), then the objective function is: max{J(F i )}.

[0083] In the solution of the objective function, the constructed preferred feature set is initially empty, and a feature F i is randomly selected from the feature space as a waiting candidate feature in each iteration. The function score corresponding to the feature F i is solved, and the feature F i is added to the preferred feature set after the solution, completing one iteration. In the next iteration, another feature is selected to complete the above operation, and after all features in the feature space are selected, the feature with the highest function score and the preferred feature set to which the feature is just added are selected as the preferred feature set that meets the objective function.

[0084] S7: Construct an initial mapping strategy based on the mapping of each feature in the preferred feature set corresponding to each reached label to the feature space; construct a final mapping strategy by combining the initial mapping strategies of each reached label and the importance degree of the hidden label.

[0085] In this embodiment, the initial mapping strategy is set to φ t : S t → F, where S t represents the preferred feature set of the current reached label, and F represents the original feature space. The initial mapping strategy is the mapping of the features in the preferred feature set to the features in the feature space.

[0086] φ t = [vt (F1), v t (F2), ..., v t (F d )], where φ t Indicates the current arrival label L t The initial mapping strategy, v t (F i ) represents feature F i The mapped value, i∈[1, d], d represents the dimension of the original feature space. i ∈S t When v t (F i )=1, otherwise v t (F i )=0.

[0087] Since the initial mapping strategy does not consider the real-time fusion of the mapped feature vectors. In addition, the initial mapping strategy directly converts the tag-specific features obtained by each arriving tag into 0 and 1, ignoring the importance of potentially hidden tags. In order to solve these two problems, this embodiment is based on the initial mapping strategy φ k Created a new mapping strategy as the final mapping strategy The calculation formula is as follows:

[0088]

[0089] Among them, Imp(L k ) indicates the arrival label L k The label importance of φ k Represents the historical arrival label L k Initial mapping strategy.

[0090] S8: Sort the values ​​of each element in the vector after the current arrival label is mapped by the final mapping strategy in descending order, and select the top N features with the largest values ​​based on the sorting results to form the selected feature set.

[0091] Current arrival label L t By final mapping strategy The content of the mapped vector is {Iv t (F1), Iv t (F2), ..., Iv t (F d )}, the sorted result is Satisfaction if Characteristics Comparison Features More important; on the contrary, if Characteristics Comparison Features Therefore, N (e.g., 20) features with the largest values ​​can be selected from the sorted results to form the selected feature set.

[0092] S9: Repeat steps S1-S8 until the current arrival label is the last arrival label in the multi-label data, and fuse the selected feature sets corresponding to all arrival labels to obtain the final feature set output.

[0093] The embodiments of the present invention provide a novel multi-label feature selection method for processing dynamic flow labels, while utilizing label correlation and learning implicit label importance information to provide additional information for learning tasks.

[0094] The embodiments of the present invention have the following beneficial effects:

[0095] 1. By jointly learning the implicit tag importance of streaming labels and the influence of label correlation on feature importance, a new learning framework is established to solve multi-label feature selection for dynamic streaming labels.

[0096] 2. By leveraging the neighborhood structure of samples in the feature space, we explore the implicit label importance of dynamic flow labels. This is seamlessly embedded into the feature selection model, providing valuable additional information for the learning task.

[0097] 3. Use label correlation to improve generalization performance, especially quantify the influence of the existing labels on the relationship between features and newly arrived labels to facilitate model training.

[0098] Example 2:

[0099] The present invention also provides a multi-label feature selection terminal device based on dynamic flow labels, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiment of embodiment 1 of the present invention are implemented.

[0100] Furthermore, as an executable solution, the multi-label feature selection terminal device based on dynamic flow labels can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The multi-label feature selection terminal device based on dynamic flow labels can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the composition structure of the multi-label feature selection terminal device based on dynamic flow labels is merely an example of a multi-label feature selection terminal device based on dynamic flow labels, and does not constitute a limitation on the multi-label feature selection terminal device based on dynamic flow labels. It can include more or fewer components than the above, or a combination of certain components, or different components. For example, the multi-label feature selection terminal device based on dynamic flow labels can also include input and output devices, network access devices, buses, etc., and the embodiment of the present invention does not limit this.

[0101] Furthermore, as an executable solution, the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices. The general-purpose processor may be a microprocessor or any conventional processor. The processor serves as the control center of the multi-label feature selection terminal device based on dynamic flow labels, and utilizes various interfaces and lines to connect various parts of the multi-label feature selection terminal device based on dynamic flow labels.

[0102] The memory can be used to store the computer program and / or module, and the processor realizes the various functions of the multi-label feature selection terminal device based on dynamic flow labels by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created based on the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0103] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method of the embodiment of the present invention are implemented.

[0104] If the module / unit integrated in the terminal device for selecting multi-label features based on dynamic flow labels is implemented in the form of 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, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM) and software distribution medium, etc.

[0105] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. A multi-label feature selection method based on dynamic flow labels, characterized in that: The following steps are involved: S1: Based on the neighborhood, positive neighborhood set, and negative neighborhood set of each sample in the multi-label data in the feature space, calculate the detailed description of the current arrival label relative to each sample; S2: Combine the detailed description of each historical arrival label to calculate the relative label importance of the current arrival label relative to each sample; S3: Calculate the invisible tag importance of the current arrival tag by combining the relative tag importance of each historical arrival tag; S4: Calculate the label correlation and feature correlation influencing factor between the current arrival label and each historical arrival label for each feature in the feature space; S5: Based on the label relevance and feature relevance influencing factors, the enhanced feature relevance and feature redundancy of the current arrival label are calculated; S6: Taking the maximum difference between the enhanced feature correlation and feature redundancy as the objective function, constructing the optimal feature set, randomly selecting a feature in the feature space after each iteration and adding it to the optimal feature set, outputting the optimal feature set that satisfies the objective function; S7: constructing an initial mapping strategy based on mapping each feature in the preferred feature set corresponding to each arrival tag to the feature space; Combining the initial mapping strategy of each arriving tag and the importance of the invisible tag, the final mapping strategy is constructed; S8: Sort the values ​​of the elements in the vector after the current arrival label is mapped by the final mapping strategy in descending order, and select the top N features with the largest values ​​based on the sorting results to form the selected feature set; S9: Repeat steps S1-S8 until the current arrival label is the last arrival label in the multi-label data, and fuse the selected feature sets corresponding to all arrival labels to obtain the final feature set output; Among them, the current arrival label L t Relative to sample x i Descriptive level The calculation formula is: in, Represents sample x i The positive neighborhood set of Represents sample x i The negative neighborhood set of θ F (x i ) represents the sample x i In the neighborhood of the feature space, |.| means the number of elements in the set. represents the positive label, represents a negative label.

2. The multi-label feature selection method based on dynamic flow labels according to claim 1 is characterized in that: Current arrival label L t Relative to sample x i The relative importance of labels The calculation formula is: in, Indicates the current arrival label L t Relative to sample x i A represents the set of all currently arrived tags, |.| represents the number of elements in the set, and k represents the sequence number of the historically arrived tags. Represents the historical arrival label L k Relative to sample X i The degree of detailed description.

3. The multi-label feature selection method based on dynamic flow labels according to claim 1 is characterized in that: Current arrival label L t Implicit label importance Imp(L t ) is calculated as: Among them, i represents the serial number of the sample, n represents the total number of samples, Indicates the current arrival label L t Relative to sample x i The relative importance of the label, A represents the set of all currently arrived labels, |.| represents the number of elements in the set, k represents the sequence number of the historically arrived labels, Represents the historical arrival label L k Relative to sample x i The relative importance of labels.

4. The multi-label feature selection method based on dynamic flow labels according to claim 1, characterized in that: About the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The label correlation Lco(F i ,L t ,L k ) is calculated as: Lco(F i ,L t ,L k )=I(F i ;L t |L k )-I(F i ;L t ) Among them, I(F i ;L t |L k ) indicates that the historical arrival label L k Find the characteristic F under the condition i For the current arrival label L t The contribution of I(F i ;L t ) represents feature F i For the current arrival label L t degree of contribution.

5. The multi-label feature selection method based on dynamic flow labels according to claim 1 is characterized in that: About the feature F in the feature space i , the current arrival label L t With the history of the arrival label L k The feature correlation factor ω(F i ,L t ,L k ) is calculated as: Among them, L S represents the historical label set, and H(.) represents entropy.

6. The multi-label feature selection method based on dynamic flow labels according to claim 1, characterized in that: Current arrival label L t Enhanced feature correlation Rel(F i ,L t ,L S ) is calculated as: Rel(F i ,L t ,L S )=I(F i ;L t )×ω(F i ,L t ,L S ) Current arrival label L t The feature redundancy Red(F i ,L t ) is calculated as: Among them, I(F i ;L t ) represents feature F i For the current arrival label L t The contribution degree of S represents the preferred feature set, |.| represents the number of elements in the set, and F j represents the features in the preferred feature set, F i For the features waiting to be selected outside the preferred feature set, I(F i ,F j |L t ) indicates that at the current arrival tag L t Under the condition of i For feature F j The contribution degree of ω(F i ,L t ,L S ) indicates the current arrival label L t With the history of the arrival label L k The feature correlation influencing factor between them.

7. The multi-label feature selection method based on dynamic flow labels according to claim 1, characterized in that: Current arrival label L t The final mapping strategy The calculation formula is: Among them, L S Represents the historical tag set, L k Indicates the historical arrival label, Imp(L k ) represents the historical arrival label L k Implicit label importance, Imp(L t ) indicates the current arrival label L t The implicit label importance, φ k Represents the historical arrival label L k The initial mapping strategy, φ represents the current arrival label L t Initial mapping strategy.

8. A terminal device for multi-label feature selection based on dynamic flow labels, characterized by: The method comprises a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images