Complementary label space based multi-label feature selection guided denoising method

Abstract

This application relates to the field of data mining technology and discloses a multi-label feature selection-guided denoising method based on complementary label space. The method includes: obtaining a set of complementary labels and a set of observed labels; obtaining a feature set; performing the (k+1)th iteration optimization on the k-th clean label recognition matrix based on the complementary label set, the k-th clean label set, and the k-th noisy label set to obtain the (k+1)-th clean label recognition matrix; performing the (k+1)-th iteration optimization on the k-th clean label set based on the (k+1)-th clean label recognition matrix, the k-th noisy label set, and the complementary label set to obtain the (k+1)-th clean label set; processing the k-th feature selection matrix based on the (k+1)-th clean label set and the feature set to obtain the (k+1)-th feature selection matrix; and guiding the sorting of features in the feature set based on the (k+1)-th feature selection matrix. This scheme can meet the practical needs of feature selection for data objects.

Description

Technical Field

[0001] This application relates to the field of data mining technology, specifically to, but not limited to, a multi-label feature selection guided denoising method based on complementary label space. Background Technology

[0002] In multi-label feature selection scenarios, each data object is typically assigned multiple labels. However, in practical applications, noisy labels are common among the multiple labels of a data object, which can mislead the feature selection for that data object.

[0003] To address the aforementioned technical challenges, related technologies provide methods for obtaining supervisory information from observation labels, classifying observation labels based on this supervisory information to remove noisy labels and obtain clean labels, establishing a mapping relationship between the features of data objects and their labels based on the clean labels, and finally selecting features of data objects based on this mapping relationship.

[0004] However, the accuracy of the mapping relationship established in the above way is insufficient and cannot meet the actual needs of feature selection for data objects. Summary of the Invention

[0005] Based on the above technical problems, this application provides a multi-label feature selection guided denoising method based on complementary label space, which can meet the actual needs of feature selection for data objects.

[0006] The technical solution provided in this application is as follows: This application provides a multi-label feature selection-guided denoising method based on complementary label space, including: Obtain a complementary label set and an observation label set; wherein, the complementary label set corresponds to the complementary label space; the observation label space corresponding to the observation label set is associated with the complementary label space; Obtain the feature set associated with the complementary label set and the observation label set; Based on the complementary label set, the kth clean label set, and the kth noisy label set, the kth clean label recognition matrix is ​​optimized in the (k+1)th iteration to obtain the (k+1)th clean label recognition matrix; where k is an integer greater than or equal to 1; the kth clean label set, the kth noisy label set, and the kth clean label recognition matrix are all related to the kth iteration optimization; Based on the (k+1)th clean label recognition matrix, the kth noisy label set, and the complementary label set, the (k+1)th iteration optimization is performed on the kth clean label set to obtain the (k+1)th clean label set. Based on the (k+1)th clean label set and the feature set, the k-th feature selection matrix is ​​processed to obtain the (k+1)-th feature selection matrix; wherein, the k-th feature selection matrix is ​​related to the k-th iteration optimization; The features in the feature set are sorted based on the (k+1)th feature selection matrix.

[0007] The multi-label feature selection-guided denoising method based on complementary label space provided in this application has at least the following beneficial effects: The multi-label feature selection-guided denoising method based on complementary label space provided in this application obtains a complementary label set and an observation label set. The complementary label set corresponds to the complementary label space, and the observation label set is associated with the corresponding observation label space and the complementary label space. Thus, the diversity of labels can be improved through the complementary label set and the observation label set. Furthermore, a feature set associated with the complementary label set and the observation label set is obtained. Thus, the richness and diversity of the associations between the features in the feature set and the labels in the two label sets can be improved through the feature set, the complementary label set, and the observation label set. Simultaneously, based on the complementary label set, the kth clean label set, and the kth noisy label set, the k+1th iteration optimization of the kth clean label recognition matrix is ​​performed to obtain the k+1th clean label recognition matrix. Thus, by utilizing the complementary labels contained in the complementary label set, a targeted extension of the k+1th clean label recognition matrix in the complementary label dimension can be achieved, thereby improving the clean label recognition accuracy of the k+1th clean label recognition matrix. Breadth; Based on this, based on the (k+1)th clean label recognition matrix, the kth noisy label set, and the complementary label set, the (k+1)th iteration optimization is performed on the kth clean label set to obtain the (k+1)th clean label set. In this way, not only is closed-loop iterative optimization of the kth clean label set achieved, but the breadth of clean labels in the (k+1)th clean label set can also be improved. At the same time, based on the (k+1)th clean label set and the feature set, the kth feature selection matrix is ​​processed to obtain the (k+1)th feature selection matrix. Since the kth feature selection matrix indirectly contains the correlation between the labels in the complementary label set and the features in the feature set, the breadth of the correlation between the labels and features in the (k+1)th feature selection matrix can be improved. Furthermore, based on the (k+1)th feature selection matrix, the features in the feature set are guided and sorted, which can achieve comprehensive and high-precision guided sorting of the features in the feature set, thereby enabling targeted denoising processing of the features in the feature set and meeting the feature selection requirements for any data object's corresponding feature set. Attached Figure Description

[0008] Figure 1A flowchart illustrating the multi-label feature selection-guided denoising method based on complementary label space provided in this application embodiment; Figure 2 A schematic diagram of the observation label set and complementary label set provided in the embodiments of this application; Figure 3 A schematic diagram of the architecture of Correlation and LabelDependency Feature Selection (CLDFS) provided in the embodiments of this application; Figure 4A A 3D waterfall plot comparing Hamming loss for various label feature selection methods on the Science dataset; Figure 4B A 3D waterfall plot showing the 0-1 loss comparison of various label feature selection methods on the Science dataset; Figure 4C A 3D waterfall plot of macro F1 scores for various label feature selection methods on the Science dataset; Figure 4D A 3D waterfall plot of the micro-F1 scores for various label feature selection methods on the Science dataset. Detailed Implementation

[0009] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0010] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0011] In multi-label feature selection scenarios, each data object is typically assigned multiple labels. However, in practical applications, noisy labels are common among the multiple labels of a data object, which can mislead the feature selection for that data object.

[0012] To address the aforementioned technical challenges, related technologies provide methods for obtaining supervisory information from observation labels, classifying the observation labels based on this supervisory information to remove noisy labels and obtain clean labels, establishing a mapping relationship between the features of data objects and their labels based on the clean labels, and then selecting features of data objects based on this mapping relationship.

[0013] However, the accuracy of the mapping relationship established in the above way is insufficient and cannot meet the actual needs of feature selection for data objects.

[0014] Based on the above technical problems, this application provides a multi-label feature selection guided denoising method based on complementary label space. Figure 1This is a flowchart illustrating the multi-label feature selection-guided denoising method based on complementary label space provided in an embodiment of this application. Figure 1 As shown, the method may include the following steps: Step 101: Obtain the set of complementary labels and the set of observation labels.

[0015] Among them, the complementary label set is associated with the complementary label space; the observation label space corresponding to the observation label set is associated with the complementary label space.

[0016] In some embodiments, the observation label set may include a set of original labels obtained by analyzing and labeling a set of data objects; for example, the observation label set may include incorrectly labeled noise labels; for example, the observation label set may be represented by an observation label matrix.

[0017] In some embodiments, the observation label space may include a vector space spanned by all observation labels in the observation label set.

[0018] In some embodiments, the complementary label set may include a set of labels that are mutually exclusive with or do not intersect with the true labels of the data object set; for example, the complementary labels in the complementary label set are used to label which categories the data objects do not belong to; specifically, the complementary label set and the observation label set may form a complement relationship.

[0019] In some embodiments, the complementary label space includes a representation space spanned by all complementary labels in the complementary label set as a basis. This space carries negative supervision information that the data object does not belong to the corresponding category. It is strictly mutually exclusive and complementary to the effective label space corresponding to the observation label space, and is used to define the boundary of the non-target category.

[0020] Figure 2 This is a schematic diagram illustrating the observation label set and complementary label set provided in an embodiment of this application. (See diagram below.) Figure 2 As shown, the observation label set can include the observation label set for the image data of "panda eating bamboo", which can specifically include Panda, Bamboo, Dog, Tiger, Fish and Elephant. The complementary label set can include Not Panda, Not Bamboo, Not Dog, Not Tiger, Not Fish and Not Elephant. Among them, the noise label can include Dog and Tiger, the clean label can include Panda, Bamboo, Not Fish and Not Elephant, and the remaining labels can be negative labels.

[0021] Step 102: Obtain the feature set associated with the complementary label set and the observation label set.

[0022] In some embodiments, the feature set may include a set of features of at least one dimension of at least one data object; exemplarily, the data object may include Figure 2 The image data shown contains pandas and bamboo.

[0023] In some embodiments, the feature set may include features of the data object in at least one aspect of the pixel dimension, audio dimension, and text dimension; for example, the feature set may be pre-established, and this application embodiment does not limit this.

[0024] In some embodiments, the feature set is associated with the complementary label set and the observation label set, which may include features in the feature set and have an association relationship with at least some of the labels contained in the complementary label set and the observation label set; for example, the initial relationship corresponding to the above association relationship can be established in advance, and can also be continuously adjusted through subsequent iterative optimization processes.

[0025] Step 103: Based on the complementary label set, the kth clean label set, and the kth noisy label set, perform the k+1th iteration optimization on the kth clean label recognition matrix to obtain the k+1th clean label recognition matrix.

[0026] Where k is an integer greater than or equal to 1; the k-th clean label set, the k-th noisy label set, and the k-th clean label recognition matrix are all related to the k-th iteration.

[0027] In some embodiments, the k-th clean label set and the k-th noisy label set can be the set of clean labels and the set of noisy labels obtained by the k-th iteration optimization, respectively; for example, the k-th clean label set and the k-th noisy label set can be proper subsets of the label set composed of the complementary label set and the observation label set.

[0028] In some embodiments, the k-th clean label recognition matrix can be the clean label recognition matrix obtained by optimization through the k-th iteration.

[0029] In some embodiments, the (k+1)th clean label set can be obtained in the following way: The k-th clean label set and the k-th noisy label set are compared to determine the degree of difference between the k-th clean label set and the k-th noisy label set. Then, based on the degree of difference and the k-th clean label set, the k-th clean label features of the clean labels in the complementary label set are tracked and captured. Finally, the k-th clean label recognition matrix is ​​augmented or supplemented based on the k-th clean label features to obtain the k+1-th clean label recognition matrix.

[0030] Step 104: Based on the (k+1)th clean label recognition matrix, the kth noisy label set, and the complementary label set, perform the (k+1)th iteration optimization on the kth clean label set to obtain the (k+1)th clean label set.

[0031] In some embodiments, the (k+1)th clean label set can be obtained in the following way: The k-th noisy label set is identified using the (k+1)-th clean label identification matrix to obtain the first identification result. The complementary label set is identified using the (k+1)-th clean label identification matrix to obtain the second identification result. Then, based on the symmetric difference between the first and second identification results, the k-th clean label set is supplemented or corrected to obtain the (k+1)-th clean label set.

[0032] Step 105: Based on the (k+1)th clean label set and feature set, process the kth feature selection matrix to obtain the (k+1)th feature selection matrix.

[0033] In some embodiments, the k-th feature selection matrix is ​​associated with the k-th iteration optimization.

[0034] In some embodiments, the k-th feature selection matrix may include the feature selection matrix calculated through the k-th iteration optimization process; for example, the feature selection matrix may include the mapping relationship between features in the feature set and clean labels in the clean label set.

[0035] In some embodiments, the (k+1)th feature selection matrix can be obtained in the following way: The k-th feature selection matrix is ​​analyzed to determine the k-th mapping relationship between the k-th clean label set represented by the k-th feature selection matrix and the features in the feature set. Then, based on the k-th mapping relationship, the k+1-th mapping relationship between the (k+1)-th clean label matrix and the features in the feature set is constructed. Next, the k-th mapping relationship is corrected or supplemented based on the (k+1)-th mapping relationship, and the corrected or supplemented k-th mapping relationship is converted into the (k+1)-th feature selection matrix.

[0036] Step 106: Guide the sorting of features in the feature set based on the (k+1)th feature selection matrix.

[0037] In some embodiments, when the degree of difference between the (k+1)th feature selection matrix and the kth feature selection matrix is ​​less than or equal to a degree threshold, the features in the feature set can be guided to be sorted based on the (k+1)th feature selection matrix.

[0038] In some embodiments, guided sorting of features in a feature set can be achieved in the following ways: Based on the (k+1)th feature selection matrix, the kth relaxed diagonal matrix is ​​updated to obtain the (k+1)th relaxed diagonal matrix; based on the (k+1)th relaxed diagonal matrix, the features in the feature set are sorted; wherein, the kth relaxed diagonal matrix corresponds to the kth feature selection matrix.

[0039] For example, the (k+1)th relaxed diagonal matrix can be the diagonal elements of the (k+1)th feature selection matrix, and all off-diagonal elements of the (k+1)th feature selection matrix can be 0, while the diagonal elements can have non-zero values; for example, the dimension of the (k+1)th relaxed diagonal matrix can be the same as the dimension of the (k+1)th feature selection matrix.

[0040] For example, the (k+1)th relaxed diagonal matrix can be calculated using equation (1): (1); in, Choose the matrix for the (k+1)th feature. To select the i-th diagonal element of the matrix for the (k+1)-th feature, For the (k+1)th relaxed diagonal matrix The The diagonal element is used to represent the first diagonal element. The adaptive regularization weights corresponding to each feature is an integer greater than or equal to 1, used to characterize the element index of the (k+1)th feature selection matrix. It is a very small positive number used to implement zero-reduction protection and to reduce the abnormal probability of the denominator being 0 when all diagonal elements of the (k+1)th feature selection matrix are 0.

[0041] Specifically, the guided sorting of features in the feature set based on the (k+1)th relaxed diagonal matrix can be achieved in the following way: Based on the values ​​of the diagonal elements in the (k+1)th relaxed diagonal matrix, the features in the feature set are sorted in ascending or descending order to obtain the feature ranking result. Features with an importance less than or equal to a threshold in the feature ranking result are then removed, thereby achieving noise reduction and filtering of features in the feature set. The importance of a feature can be represented by the value of the corresponding diagonal element in the (k+1)th relaxed diagonal matrix. For example, the importance of a feature can increase as the value of the corresponding diagonal element in the (k+1)th relaxed diagonal matrix increases.

[0042] As can be seen from the above, the multi-label feature selection-guided denoising method based on complementary label space provided in this application obtains a complementary label set and an observation label set. The complementary label set corresponds to the complementary label space, and the observation label set is associated with the corresponding observation label space and the complementary label space. Thus, the diversity of labels can be improved through the complementary label set and the observation label set. Furthermore, a feature set associated with the complementary label set and the observation label set is obtained. Thus, the richness and diversity of the association between the features in the feature set and the labels in the above two label sets can be improved through the feature set, the complementary label set, and the observation label set. At the same time, based on the complementary label set, the kth clean label set, and the kth noisy label set, the k+1th iteration optimization of the kth clean label recognition matrix is ​​performed to obtain the k+1th clean label recognition matrix. Thus, by using the complementary labels contained in the complementary label set, the k+1th clean label recognition matrix can be specifically extended in the complementary label dimension, thereby improving the k+1th clean label recognition matrix for clean labels. The breadth of recognition is increased. Based on this, the (k+1)th clean label recognition matrix, the kth noisy label set, and the complementary label set are used to perform the (k+1)th iteration optimization on the kth clean label set to obtain the (k+1)th clean label set. This not only achieves closed-loop iterative optimization of the kth clean label set but also improves the breadth of clean labels in the (k+1)th clean label set. At the same time, based on the (k+1)th clean label set and the feature set, the kth feature selection matrix is ​​processed to obtain the (k+1)th feature selection matrix. Since the kth feature selection matrix indirectly contains the correlation between the labels in the complementary label set and the features in the feature set, it can improve the breadth of the correlation between the labels and features in the (k+1)th feature selection matrix. Furthermore, based on the (k+1)th feature selection matrix, the features in the feature set are guided and sorted, which can achieve comprehensive and high-precision guided sorting of the features in the feature set. This enables targeted denoising of the features in the feature set and meets the feature selection requirements for any data object's corresponding feature set.

[0043] Based on the foregoing embodiments, the multi-label feature selection-guided denoising method based on complementary label space provided in this application optimizes the k+1th clean label recognition matrix by performing the (k+1)th iteration based on the complementary label set, the kth clean label set, and the kth noisy label set to obtain the (k+1)th clean label recognition matrix. This can be achieved in the following way: The k+1th iteration of the clean label recognition matrix is ​​optimized using equation (2): (2); in, This is the (k+1)th clean label recognition matrix. For the k-th clean label recognition matrix, A set of complementary tags, Let k be the set of noise labels. Let k be the set of clean labels. , and These are pre-set parameters.

[0044] As can be seen from the above, in the multi-label feature selection guided denoising method based on complementary label space provided in this application embodiment, the k+1th iteration optimization is performed on the k-th clean label recognition matrix through equation (2). Since equation (2) realizes the comprehensive integration of the k+1th clean label recognition matrix, the k-th clean label recognition matrix, the complementary label set, the k-th noisy label set, and the k-th clean label set, it can improve the recognition accuracy of the k+1th clean label recognition matrix for clean labels in complementary labels and improve the breadth of clean label recognition by the k+1th clean label recognition matrix.

[0045] Based on the foregoing embodiments, the multi-label feature selection-guided denoising method based on complementary label space provided in this application, which performs the (k+1)th iteration optimization on the (k+1)th clean label set based on the (k+1)th clean label set, the (k)th noisy label set, and the complementary label set, to obtain the (k+1)th clean label set, can be achieved in the following way: By performing the (k+1)th iteration optimization on the k-th clean label set using equation (3), we obtain the (k+1)-th clean label set: (3); in, Let k+1 be the set of clean labels. This is the (k+1)th clean label recognition matrix. Choose a matrix for the k-th feature; Let k be the set of clean labels. For the (k+1)th noise label set, For feature set, For the set of observation labels, A set of complementary tags, , and These are pre-set parameters.

[0046] As can be seen from the above, in the multi-label feature selection guided denoising method based on complementary label space provided in this application embodiment, the k+1th iteration optimization is performed on the k-th clean label set by equation (3) to obtain the k+1th clean label set. Equation (3) realizes the comprehensive integration of the k+1th clean label set, the k-th feature selection matrix, the k-th clean label set, the k+1th noisy label set, the feature set, the observation label set and the complementary label set, so that the k+1th clean label set can contain the clean labels in the complementary label set, thereby expanding the comprehensiveness and diversity of the k+1th clean label set.

[0047] Based on the foregoing embodiments, the multi-label feature selection-guided denoising method based on complementary label space provided in this application can also perform the following operations: Based on the observed label set, the kth clean label set, and the k+1th clean label recognition matrix, the kth noisy label matrix is ​​iterated to obtain the k+1th noisy label set, as shown in equation (4): (4); in, Let k be the set of noise labels, and let k be related to the k-th iteration optimization. These are pre-set parameters.

[0048] In some embodiments, the k-th noise label set is related to the k-th iteration optimization and may include the k-th noise label set being the set of noise labels calculated by the k-th iteration optimization.

[0049] As can be seen from the above, in the multi-label feature selection guided denoising method based on complementary label space provided in this application embodiment, the observation label set, the kth clean label set, the (k+1)th clean label recognition matrix, and the kth noisy label set are iterated to obtain the (k+1)th noisy label set. Thus, in the process of iterating the (k+1)th noisy label set, the influence of the observation label set, the kth clean label set, the (k+1)th clean label recognition matrix, and the kth noisy label set is added, so that the (k+1)th noisy label set can contain the noise labels in the observation label set, thereby improving the breadth and diversity of noise labels in the (k+1)th noisy label set.

[0050] Based on the foregoing embodiments, in the multi-label feature selection-guided denoising method based on complementary label space provided in this application, the k-th feature selection matrix is ​​processed based on the (k+1)-th clean label set and feature set to obtain the (k+1)-th feature selection matrix, which can be achieved in the following way: By iteratively processing the k-th feature selection matrix using equation (5), the (k+1)-th feature selection matrix is ​​obtained: (5); in, Choose the matrix for the (k+1)th feature. Choose a matrix for the k-th feature. For feature set, Let k+1 be the set of clean labels. Let k be the relaxed diagonal matrix corresponding to the feature set. and ,as well as These are pre-set parameters.

[0051] As can be seen from the above, in the multi-label feature selection guided denoising method based on complementary label space provided in this application embodiment, the k-th feature selection matrix is ​​iteratively processed by equation (3) to obtain the (k+1)-th feature selection matrix. Thus, equation (3) achieves a comprehensive integration of the (k+1)-th feature selection matrix, the k-th feature selection matrix, the feature set, the (k+1)-th clean label set, and the k-th relaxed diagonal matrix corresponding to the feature set, thereby improving the comprehensiveness and diversity of the data carried by the (k+1)-th feature selection matrix and improving the accuracy of the (k+1)-th feature selection matrix.

[0052] The following section will explain the origin of the iterative optimization process for the (k+1)th feature selection matrix.

[0053] First, construct the sparse regression model as shown in equation (6): The observed label matrix Y is explicitly decomposed to obtain the clean label matrix T and the noisy label matrix R; simultaneously, W is used as the feature selection matrix to establish the mapping relationship between features and clean labels: (6); In addition to the useful information contained in the observation label matrix, the complementary label matrix also contains a wealth of valuable information; therefore, the following model can also be constructed: Negative label information in the complementary label matrix is ​​identified by a clean label recognition matrix Q; for example, the clean label recognition matrix Q can be obtained by initializing the complement of the noisy label matrix R.

[0054] Specifically, the complementary label matrix serves as the complement of the observation label matrix. It can accurately indicate the label to which each feature should not be assigned, while at the same time, the complementary label matrix It is incomplete and contains noisy labels; therefore, it can be solved by using complementary label matrices. Coupled with the clean label matrix T and the clean label recognition matrix Q, a system is constructed using... The parameterized regularization term ensures that the complementary label matrix is ​​consistent with the clean label estimate corrected by the clean label recognition matrix Q, thereby utilizing the "definitely not belonging" labels in the complementary label matrix to optimize the estimate for T and reduce the influence of noisy labels. Specifically, as shown in equation (7): (7); Among them, the complementary label matrix is ​​the same as the observation label matrix, but the values ​​of corresponding elements are complementary; for the observation label matrix, if its elements If the value of is 1, it indicates that the i-th feature has the j-th label, and if If the value of is 0, it indicates that the i-th feature does not have the j-th label; j is an integer greater than or equal to 1.

[0055] Next, complementary learning is performed; specifically, another regularization term can be introduced to couple the learning processes of the noise label recognition matrix R and the clean label recognition matrix Q. Specifically, the above regularization term emphasizes the numerical complementarity between R and Q by constraining them, establishing a collaborative learning mechanism between R and Q. That is, Q can guide the learning of the noise matrix R from the clean label information learned from the complementary label space, while R can in turn correct Q's estimation of the clean labels. This complementary design improves the inherent consistency of the label space judgment; specifically, the above complementary relationship can be expressed by equation (8): (8); in, For dimension A matrix whose elements are all 1; and All are integers greater than or equal to 1.

[0056] Specifically, after introducing the above regularization term, equation (7) can be transformed into equation (9): (9); In equation (9), through Complementary label supervision items The combined effect of these elements allows for more effective utilization of complementary tag matrices. The reliable negative label information provided enables this clean label matrix T to be integrated with the complementary label space. Maintaining structural complementarity enhances the robustness of clean label recovery.

[0057] Furthermore, an objective function can be constructed and optimized; wherein the objective function can be as shown in equation (10): (10); In equation (10), This is used to apply sparse constraints to the feature selector W obtained from the feature matrix X and the clean label matrix T, thereby automatically identifying and removing features that contribute little to multi-label classification or are heavily contaminated by noise; and, for multi-label feature selection tasks, it employs... It can also balance the weight distribution of the same feature across different categories, enabling group selection and helping to reduce the suboptimal feature subset problem caused by the correlation between labels. It represents the Frobenius norm.

[0058] In the above process, the regularization term works in conjunction with the label denoising module: on the one hand, the denoised labels can provide more reliable supervision signals for feature selection; on the other hand, sparse feature subsets can reduce the propagation of noisy labels in the feature space, thereby further enhancing the robustness and generalization ability of label recovery.

[0059] In multi-label feature selection, mislabeling of labels in the observation label space can interfere with the selection of feature subsets. Therefore, by explicitly decomposing the observation labels into clean labels and noisy labels, and by using the recognition matrix to constrain the clean label matrix in the complementary label space rich in effective information, while combining regularization learned for the complementary label matrix, W can select the optimal feature subset that is closer to the clean labels.

[0060] The above process is the architecture flow of CLDFS that can be provided for the embodiments of this application. Figure 3 This is a schematic diagram of the CLDFS architecture provided in the embodiments of this application, as shown below. Figure 3 As shown, the architecture of CLDFS can include an input module, a core reconstruction mapping module, a regularization module, a label decomposition engine module, a mirror-enhanced supervision module, and a complementary collaborative learning constraint module.

[0061] For example, the input module is used to input the feature matrix into the core reconstruction mapping module to decompose the feature matrix by W. The obtained T is also used to input the observation label matrix into the label decomposition engine module to trigger the module to execute. The operation; at the same time, the core reconstruction mapping module needs to rely on the T output by the label decomposition engine module when decomposing the feature matrix.

[0062] For example, the regularization module can perform regularization processing on W. The mirror-enhanced supervision module is used to perform operations on the complementary label matrix, T, and Q. The processing and complementary collaborative learning constraint module is used for execution. deal with.

[0063] In other words, Figure 3The data processing flow performed by the architecture shown can be consistent with the data processing flow represented by equation (10).

[0064] However, since all variables in equation (10) are jointly nonconvex, the global optimal solution of the function cannot be obtained; furthermore, applying the feature selection term... Norm constraints introduce non-smoothness. To address this issue, the following four iterative update rules can be used: First, the non-smooth terms Relaxation is quadratic Then, the non-negativity constraints of the four variables W, T, R, and Q are incorporated into the objective function, and Ψ, Φ, Ω, and Θ are the Lagrange multipliers corresponding to W, T, R, and Q, respectively, to obtain the Lagrange function as shown in equation (11): (11); By calculating the partial derivatives of equation (11) with respect to W, T, R, and Q in turn, we can obtain equation (12): (12); Since some terms in the above partial derivatives do not have closed-form solutions, a solution framework can be constructed based on the Karush–Kuhn–Tucker (KKT) conditions. By analyzing the optimality conditions, the equivalent equation system shown in equation (13) can be derived: (13); in, Characterization matrix The element in the i-th row and j-th column, Characterization matrix The element in the i-th row and j-th column, Characterization matrix The element in the i-th row and j-th column, Characterization matrix The element in the i-th row and j-th column, Characterizing Lagrange multipliers The element in the i-th row and j-th column of the corresponding matrix, Characterizing Lagrange multipliers The element in the i-th row and j-th column of the corresponding matrix, Characterizing Lagrange multipliers The element in the i-th row and j-th column of the corresponding matrix, Characterizing Lagrange multipliers The elements of the matrix in the i-th row and j-th column of the corresponding matrix, where i and j are both integers greater than or equal to 1.

[0065] Based on the above derivation, we can obtain the calculation formulas shown in equations (2) to (5).

[0066] It should be noted that the set of observation label matrix, complementary label matrix, and features represented by X can constitute the set of data samples on which the above derivation process is based; and the observation label matrix can correspond to the set of observation labels in the aforementioned embodiments, the complementary label matrix can correspond to the set of complementary labels in the aforementioned embodiments, X can correspond to the set of features in the aforementioned embodiments, the clean label matrix can correspond to the set of clean labels in the aforementioned embodiments, and the noise label matrix can correspond to the set of noise labels in the aforementioned embodiments.

[0067] Table 1 ; To fully verify the effectiveness of the method, this application selected four cross-domain benchmark datasets, the specific types and domain distributions of which are shown in Table 1. The experiment adopted a comparative analysis framework, using four metrics—Macro-F1, Micro-F1, Hamming Loss (HL), and 0-1 Loss (ZOL)—to comprehensively evaluate the performance of the proposed scheme.

[0068] This paper compares the performance of the proposed solution with that of seven classic multi-label feature selection methods. These seven methods include: Fuzzy Logic Inference System (FLIS), Sparse Low-Dimensional Multi-Data Source (SLMDS), Nearest Class Mean based Feature Selection (NCMFS), Robust Low-Rank Embedding Feature Selection (RLEFS), Dual Sparse Multi-View Feature Selection (DSMFS), Multi-label Informed Feature Selection (MIFS), and Semi-Supervised Feature Selection (SSFS).

[0069] Figure 4A A 3D waterfall plot comparing Hamming loss for various label feature selection methods on the Science dataset; Figure 4B A 3D waterfall plot showing the 0-1 loss comparison of various label feature selection methods on the Science dataset;Figure 4C A 3D waterfall plot of macro F1 scores for various label feature selection methods on the Science dataset; Figure 4D A 3D waterfall plot of the micro-F1 scores for various label feature selection methods on the Science dataset.

[0070] exist Figures 4A to 4D In the graph, the X-axis represents the percentage of selected features, the Y-axis represents the name of different feature selection methods, and the Z-axis corresponds to the performance evaluation results of different feature selection methods.

[0071] from Figure 4A and Figure 4B As can be seen, the proposed solution on the Science dataset exhibits lower waterfall peak values ​​for both Hamming loss and 0-1 loss compared to other methods, indicating superior performance in minimizing loss. Figure 4C and Figure 4D As can be seen, the technical solution provided in this application has a higher waterfall peak in macro F1 and micro F1 scores on the Science dataset than other comparative methods, proving that it performs best on label-based classification metrics.

[0072] To further analyze the algorithm performance, we used the Friedman test and the Bonferroni-Dun test to examine the statistical significance and significant differences between the CLDFS provided in this application and the above-mentioned comparative schemes. Table 2 shows the performance ranking results of the comparative schemes provided in the embodiments of this application and the scheme of this application for individual indicators.

[0073] In Table 2, the micro-averaged support vector machine is a multi-label SVM that evaluates the SVM classifier with Micro-F1 as the core evaluation metric. It calculates F1 uniformly by summarizing the true positives, false positives, and false negatives of all labels. On the other hand, the macro-averaged support vector machine (Macro-F1based SVM) is an SVM classifier with Macro-F1 as the core evaluation metric. It calculates Macro-F1 for each label separately and then takes the arithmetic mean to obtain the output.

[0074] Table 2 ; As can be seen from Table 2, compared with the comparative schemes, the technical solution provided in this application ranks first in all indicator dimensions.

[0075] This application also provides comparative results of ablation experiments. Tables 3 to 5 show the comparative results of ablation experiments provided in this application.

[0076] Table 3 ; In Tables 3 to 5, CLDFS represents the complete model, while CLDFSY and CLDFSZ are its ablation variants. CLDFSY represents a variant that removes the complementary co-learning constraint module used for noise identification between the observation label and the complementary label in CLDFS, while CLDFSZ represents a variant that further removes the mirror enhancement supervision module based on CLDFSY.

[0077] Table 4 ; Table 5 ; As shown in Tables 3 to 5, on datasets including Birds, Genbase, and Health, CLDFS outperforms any of its ablation variants in the micro-F1, macro-F1, HL, and ZOL dimensions. This indicates that the complementary learning term focuses on enhancing the recognition ability and robustness of label noise, while the mirror-enhanced supervision term aims to enhance the discriminative power and stability of feature learning. The two work together to enable the complete CLDFS method to select the optimal feature subset with stronger discriminative power and higher robustness from noisy data.

[0078] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0079] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0080] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0081] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0082] It should be noted that the aforementioned computer-readable storage media can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0083] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0084] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware nodes. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-label feature selection-guided denoising method based on complementary label space, characterized in that, The method includes: Obtain a complementary label set and an observation label set; wherein, the complementary label set corresponds to the complementary label space; the observation label space corresponding to the observation label set is associated with the complementary label space; Obtain the feature set associated with the complementary label set and the observation label set; Based on the complementary label set, the kth clean label set, and the kth noisy label set, the kth clean label recognition matrix is ​​optimized in the (k+1)th iteration to obtain the (k+1)th clean label recognition matrix; where k is an integer greater than or equal to 1; the kth clean label set, the kth noisy label set, and the kth clean label recognition matrix are all related to the kth iteration optimization; Based on the (k+1)th clean label recognition matrix, the kth noisy label set, and the complementary label set, the (k+1)th iteration optimization is performed on the kth clean label set to obtain the (k+1)th clean label set. Based on the (k+1)th clean label set and the feature set, the k-th feature selection matrix is ​​processed to obtain the (k+1)-th feature selection matrix; wherein, the k-th feature selection matrix is ​​related to the k-th iteration optimization; The features in the feature set are sorted based on the (k+1)th feature selection matrix.

2. The method according to claim 1, characterized in that, The process of performing the (k+1)th iteration optimization on the k-th clean label recognition matrix based on the complementary label set, the k-th clean label set, and the k-th noisy label set to obtain the (k+1)-th clean label recognition matrix includes: The (k+1)th iteration optimization is performed on the k-th clean label recognition matrix using the following formula: ； in, This is the (k+1)th clean label recognition matrix. For the k-th clean label recognition matrix, A set of complementary tags, Let k be the set of noise labels. Let k be the set of clean labels. , and These are pre-set parameters.

3. The method according to claim 1, characterized in that, The process involves performing the (k+1)th iteration optimization on the (k)th clean label set based on the (k+1)th clean label set, the (k)th noisy label set, and the complementary label set, to obtain the (k+1)th clean label set, including: The (k+1)th clean label set is obtained by performing the (k+1)th iteration optimization on the k-th clean label set using the following formula: ； in, This is the (k+1)th clean label recognition matrix. Let k+1 be the set of clean labels. Choose a matrix for the k-th feature; Let k be the set of clean labels. For the (k+1)th noise label set, For feature set, For the set of observation labels, A set of complementary tags, , and These are pre-set parameters.

4. The method according to claim 3, characterized in that, The method further includes: Based on the observed label set, the k-th clean label set, and the (k+1)-th clean label identification matrix, the k-th noisy label set is iterated to obtain the (k+1)-th noisy label set; specifically as shown in the following formula: ； in, Let k be the set of noise labels, and let k be the set of noise labels related to the kth iteration optimization. These are pre-set parameters.

5. The method according to claim 1, characterized in that, The process of processing the k-th feature selection matrix based on the (k+1)-th clean label set and the feature set to obtain the (k+1)-th feature selection matrix includes: The k+1th feature selection matrix is ​​obtained by iteratively processing the k-th feature selection matrix using the following formula: ； in, Choose the matrix for the (k+1)th feature. Choose a matrix for the k-th feature. For feature set, Let k+1 be the set of clean labels. This is the k-th relaxed diagonal matrix corresponding to the feature set; and ,as well as These are pre-set parameters.

Images