Partial multi-label feature selection method based on label confidence
Patent Information
- Application Number
- CN202410504595.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-25
- Publication Date
- 2026-09-22
- Estimated Expiration
- 2044-04-25
AI Technical Summary
[0004]1、现有方法忽略了模糊特征与冗余特征的影响，模糊特征是引起标签歧义的特征，这两种特征阻碍了现有方法对噪音标签的识别
[0034](1)通过引入平均相同标签样本距离与聚类中心距离来综合评估一个标签的置信度，有效解决了部分多标签问题中的噪音标签问题；
Smart Images
Figure QLYQS_2 
Figure QLYQS_9 
Figure QLYQS_10
Abstract
Description
Technical Field
[0001] This invention relates to the field of feature selection technology, specifically a partial multi-label feature selection method based on label confidence. Background Technology
[0002] In partially multi-label learning, an instance can be associated with multiple labels simultaneously, but only some of these labels may be correct, and the number of correct labels is unknown. This situation is very common in real-world data collection; for example, in online image labeling, a set of candidate labels can be obtained from numerous annotators, but due to potential carelessness on the part of the annotators, only some labels are valid. Existing partially multi-label learning methods can be divided into two categories: two-stage methods and iterative methods. The two-stage method first denoises the labels, transforming the problem into a traditional multi-label learning problem. The iterative method combines the label denoising process with the model training process.
[0003] However, the following problems exist in both frameworks:
[0004] 1. Existing methods ignore the influence of ambiguous and redundant features. Ambiguous features are those that cause label ambiguity, and these two features hinder existing methods from recognizing noisy labels.
[0005] 2. Existing methods mainly rely on the original feature space to eliminate noise in the labels. However, this original space is not optimal for each label and is prone to producing suboptimal solutions. Summary of the Invention
[0006] The purpose of this invention is to provide a partial multi-label feature selection method based on label confidence to solve the problems existing in the background art.
[0007] To achieve the above objectives, this invention provides a partial multi-label feature selection method based on label confidence, comprising the following steps:
[0008] S1. Based on the input feature matrix X and noise label matrix L, find the optimal feature space F corresponding to each label. l Calculate the average distance d between samples with the same label avg Distance d from cluster center C ;
[0009] S2. Calculate the label confidence matrix T;
[0010] S3. Update matrices S and W using the objective function;
[0011] S4. Repeat step S3 until the objective function converges to the minimum value;
[0012] S5. Return W, and calculate each row of W. iThe L2 paradigm scores features and selects those with higher scores as the selected features.
[0013] Preferably, in S1, for sample x i The average distance d between the first label and the same label samples avg The definition is as follows:
[0014]
[0015] Where K represents the number of neighbors with the same label, i represents the sample index, and j represents the neighbor sample index. Indicates sample x i In its corresponding optimal feature space F l The representation vector in the text, Indicates sample x i In its corresponding optimal feature space F l The j-th sample with the same label, function This indicates the optimal feature space F corresponding to label 1. l Calculate sample x i The K nearest neighbors with the same label.
[0016] Preferably, in S1, for sample x i The cluster center distance d of the l-th label C The definition is as follows:
[0017]
[0018] Among them, the function This indicates that for all positive samples with label 1, the optimal feature space F is... l Clustering is performed on the data using density peak clustering, and the cluster centers are returned.
[0019] Preferably, the specific steps of S2 are as follows:
[0020] S21, regarding sample x i The confidence score for the first label is calculated as follows:
[0021]
[0022] Where λ is used to adjust d avg and d C Harmonic parameters of relative importance and They represent d respectively avg and d C The normalized value;
[0023] S22. Normalize the label confidence matrix T using the function τ(T, L, θ), as follows:
[0024]
[0025] Where θ is the threshold parameter;
[0026] S23, Update the tag matrix
[0027] in, This indicates that the matrix elements are multiplied element by element.
[0028] Preferably, the specific process of S3 is as follows:
[0029]
[0030] Wherein, matrix W is the sum of the feature matrix X and the correct label matrix L. * The linear fitting parameters between the feature matrix X and the noise labels in the noise label matrix L are given by matrix S. α, β, and γ are regularization parameters used to adjust the relative importance of each term in the objective expression.
[0031] (1) Update the objective expression while fixing parameters W and S respectively;
[0032] (2) After completing one target update, the features corresponding to the top 50% of the largest elements in the first column of the parameter W matrix are taken as the optimal feature space F of the corresponding label 1. l .
[0033] Therefore, the present invention employs the above-mentioned partial multi-label feature selection method based on label confidence, which has the following beneficial effects:
[0034] (1) By introducing the average distance between samples with the same label and the distance between cluster centers to comprehensively evaluate the confidence of a label, the problem of noisy labels in some multi-label problems is effectively solved;
[0035] (2) By proposing the construction of the optimal feature space corresponding to the label, positive and negative samples are more easily separated linearly in this space, laying a good foundation for the label de-noising process.
[0036] The technical solution of the present invention will be further described in detail below through embodiments. Detailed Implementation
[0037] Example
[0038] A partial multi-label feature selection method based on label confidence includes the following steps:
[0039] S1. Based on the input feature matrix X and noise label matrix L, find the optimal feature space F corresponding to each label. l Calculate the average distance d between samples with the same labelavg Distance d from cluster center C ;
[0040] (1) For sample x i The average distance d between the first label and the same label samples avg The definition is as follows:
[0041]
[0042] Where K represents the number of neighbors with the same label, i represents the sample index, and j represents the neighbor sample index. Indicates sample x i In its corresponding optimal feature space F l The representation vector in the text, Indicates sample x i In its corresponding optimal feature space F l The j-th sample with the same label, function This indicates the optimal feature space F corresponding to label 1. l Calculate sample x i The K nearest neighbors with the same label.
[0043] (2) For sample x i The distance d from the cluster center of the first label C The definition is as follows:
[0044]
[0045] Among them, the function This indicates that for all positive samples with label 1, the optimal feature space F is... l Clustering is performed on the data using density peak clustering, and the cluster centers are returned.
[0046] S2. Calculate the label confidence matrix T, as follows:
[0047] S21, regarding sample x i The confidence score of the l-th label is calculated as follows:
[0048]
[0049] Where λ is used to adjust d avg and d C Harmonic parameters of relative importance and They represent d respectively avg and d C The normalized value;
[0050] S22. Normalize the label confidence matrix T using the function τ(T, L, θ), as follows:
[0051]
[0052] Where θ is the threshold parameter;
[0053] S23, Update the tag matrix
[0054] in, This indicates that the matrix elements are multiplied element by element.
[0055] S3. Update matrices S and W using the objective function:
[0056]
[0057] Wherein, matrix W is the sum of the feature matrix X and the correct label matrix L. * The linear fitting parameters between the feature matrix X and the noise labels in the noise label matrix L are given by matrix S. α, β, and γ are regularization parameters used to adjust the relative importance of each term in the objective expression.
[0058] (1) Update the objective expression while fixing parameters W and S respectively;
[0059] (2) After completing one target update, the features corresponding to the top 50% of the largest elements in the first column of the parameter W matrix are taken as the optimal feature space F of the corresponding label 1. l .
[0060] S4. Repeat step S3 until the objective function converges to the minimum value;
[0061] S5. Return W, and calculate each row of W. i The L2 paradigm scores features and selects those with higher scores as the selected features.
[0062] Therefore, the present invention adopts the above-mentioned partial multi-label feature selection method based on label confidence, which comprehensively estimates the confidence of a label by using two distance metrics, solves the problem of noisy labels in the partial multi-label problem, and selects the corresponding optimal feature subset for each label.
[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A partial multi-label feature selection method based on label confidence, characterized in that, The method is used for online image tagging and includes the following steps: S1. Based on the input feature matrix X and noise label matrix L, find the optimal feature space corresponding to each label. Calculate the average distance between samples with the same label Distance from cluster center ; S2. Calculate the label confidence matrix T; S3. Update matrices S and W using the objective function; S4. Repeat step S3 until the objective function converges to the minimum value; S5, Return ,calculate each line The L2 paradigm scores features and selects those with higher scores as the selected features. In S1, for the sample The Average distance of identical label samples for each label The definition is as follows: in, This represents the number of neighbors with the same label, where i represents the sample index and j represents the neighbor sample index. Indicates sample In its corresponding optimal feature space The representation vector in the text, Indicates sample In its corresponding optimal feature space The j-th sample with the same label, function Indicates in the label The corresponding optimal feature space Above the calculated sample nearest neighbor Samples with the same positive label; In S1, for the sample The Distance between cluster centers of each label The definition is as follows: Among them, the function This indicates all tags Positive samples in the optimal feature space Clustering is performed on the data using density peak clustering, and the cluster centers are returned. The specific steps of S2 are as follows: S21, regarding the sample The The confidence score for each label is calculated as follows: , if in, For use in adjustment and Harmonic parameters of relative importance and They represent and The normalized value; S22, via function The label confidence matrix T is normalized as follows: in, For threshold parameters; S23, Update the tag matrix ; in, This indicates element-wise multiplication of the matrix elements; The specific process of S3 is as follows: Among them, matrix It is the feature matrix X and the updated label matrix The linear fitting parameters between the feature matrix X and the noise labels in the noise label matrix L are given by matrix S. , , These are all regularization parameters, used to adjust the relative importance of each term in the objective function; (1) Update the objective expression by fixing parameters W and S respectively; (2) After completing one objective update, the parameter W matrix is updated to the first... The features corresponding to the top 50% of the largest elements in the column are used as the corresponding labels. optimal feature space .