Label Propagation Method for Weakly Labeled Datasets

Abstract

The present invention relates to the field of label propagation, and discloses a label propagation method for a weakly labeled data set. The present invention constructs a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge; in the label propagation graph, category labels and labeled samples are obtained to construct a probability transfer matrix and a labeling matrix; when the labeled sample receives label information propagated from adjacent samples, the label information is filtered to complete the label information of the labeled sample; the completed label information is propagated to the unlabeled data; and it is judged whether the label matrix converges, and if so, the label propagation is completed. The original label propagation algorithm is improved by adopting the idea of ​​feedback, the missing label information of the weakly labeled data is filled, and the correctness of label allocation is improved.

Description

Technical Field

[0001] The present application relates to the field of label propagation, and in particular to a label propagation method for a weakly labeled dataset. Background Art

[0002] Research shows that in multi-label data, incomplete labels and label noise are very common. For example, image samples in scene classification tasks are prone to noisy and missing labels, such as a small number of labels describing useless information, and some labels are not labeled. Generally speaking, a "weakly labeled dataset" refers to a multi-label dataset with incomplete labels and certain noisy labels. When collecting and manually labeling large-scale data in the real world, weak labels are difficult to avoid. The weak label nature of the dataset has reduced the performance of the multi-label learning algorithm to a certain extent, and also affected its application in real life.

[0003] Therefore, how to improve the accuracy of label assignment in the absence of weak label data has become a technical problem that needs to be solved urgently.

[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of the present invention is to provide a label propagation method for a weakly labeled data set, aiming to solve the technical problem of improving the accuracy of label assignment when weakly labeled data is missing.

[0006] To achieve the above object, the present invention provides a label propagation method for a weakly labeled data set, the method comprising:

[0007] Constructing a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge;

[0008] Obtaining category labels and labeled samples in the label propagation graph to construct a probability transfer matrix and a labeling matrix;

[0009] When the labeled sample receives label information transmitted from a neighboring sample, filtering the label information to complete the label information of the labeled sample;

[0010] Propagate the completed label information to the unlabeled data;

[0011] Determine whether the label matrix converges, and if so, complete label propagation.

[0012] Optionally, before the step of filtering the label information to complete the label information of the labeled sample when the labeled sample receives the label information propagated from the neighboring sample, the step further includes:

[0013] Determine that there are n weakly labeled data X={(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x n ,y n )}, where sample x i The corresponding weak label is α i =(α i1 ,α i2 ,...,α ic ), c is the size of the label space;

[0014] When the user is a sample x i When labeling the jth class, α ij =1, for other unmarked labels α ij =0;

[0015] Assume that T represents the similarity measurement matrix on the dataset X, and the weak label matrix corresponding to all samples is Y = (Y 1 ,Y 2 ,...,Y n ), then for each labeled sample x i Assign initial value α i , while the unlabeled sample x u The label of is determined by the result of label propagation.

[0016] Optionally, when the labeled sample receives label information propagated from a neighboring sample, the step of filtering the label information to complete the label information of the labeled sample includes:

[0017] When the labeled sample receives label information transmitted from the neighboring sample, useful information and useless information are determined according to the label information;

[0018] Retain the useful information and delete the useless information;

[0019] The useful information after retention is supplemented into the label of the labeled sample itself.

[0020] Optionally, after the step of filtering the label information to complete the label information of the labeled sample when the labeled sample receives the label information propagated from the neighboring sample, the step further includes:

[0021] In the t+1th label propagation, the labeled sample xi The tags can be represented as follows:

[0022]

[0023] Among them, Y i 0 is the label state of the label matrix at the beginning. For each labeled sample x i , execute as follows:

[0024]

[0025] Among them, 0<θ<1 is used to determine the sample x i The label portion absorbed from the information from neighboring samples.

[0026] Optionally, the steps in the t+1th label propagation further include:

[0027] Unlabeled sample x u During label propagation, the following is performed:

[0028]

[0029] Optionally, the step of constructing a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, wherein the labeled samples and the unlabeled samples are connected by an undirected edge, comprises:

[0030] Obtain a target sample, and establish a label propagation graph based on the target sample;

[0031] The label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and unlabeled samples are connected by an undirected edge;

[0032] Initialize the weight corresponding to the undirected edge.

[0033] Optionally, the step of initializing the weights corresponding to the undirected edges includes:

[0034] Initialize the (l+u)×(l+u)-dimensional probability transfer matrix T ij ,in:

[0035]

[0036] In addition, to achieve the above object, the present invention also proposes a label propagation device for a weakly labeled data set, the device comprising:

[0037] An initialization module, used to construct a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge;

[0038] A matrix construction module, used for obtaining category labels and labeled samples in the label propagation graph to construct a probability transfer matrix and a labeling matrix;

[0039] A label completion module, configured to filter the label information to complete the label information of the labeled sample when the labeled sample receives label information propagated from a neighboring sample;

[0040] The label propagation module is used to propagate the completed label information to the unlabeled data;

[0041] The convergence judgment module is used to judge whether the label matrix converges. If so, label propagation is completed.

[0042] In addition, to achieve the above-mentioned purpose, the present invention also proposes a computer device, which includes: a memory, a processor, and a label propagation program for a weakly labeled data set stored in the memory and executable on the processor, wherein the label propagation program for the weakly labeled data set is configured to implement the label propagation method for the weakly labeled data set as described above.

[0043] In addition, to achieve the above objectives, the present invention also proposes a medium on which a label propagation program for a weak label data set is stored. When the label propagation program for the weak label data set is executed by a processor, the steps of the label propagation method for the weak label data set described above are implemented.

[0044] The present invention constructs a label propagation graph, which includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge; in the label propagation graph, category labels and labeled samples are obtained to construct a probability transfer matrix and a labeling matrix; when the labeled sample receives label information propagated from adjacent samples, the label information is filtered to complete the label information of the labeled sample; the completed label information is propagated to the unlabeled data; it is determined whether the label matrix converges, and if so, the label propagation is completed. The original label propagation algorithm is improved by adopting the idea of ​​feedback, and the missing label information of the weakly labeled data is filled, thereby improving the accuracy of label assignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a structural schematic diagram of a label propagation device for a weak label data set in a hardware operating environment involved in an embodiment of the present invention;

[0046] Figure 2 It is a flowchart of the first embodiment of the label propagation method for weakly labeled data sets of the present invention;

[0047] Figure 3This is a flow chart of a classic label propagation algorithm of the first embodiment of the label propagation method for a weakly labeled data set of the present invention;

[0048] Figure 4 This is a diagram of the architecture of a label propagation algorithm for weak labels in the first embodiment of the label propagation method for a weak label data set of the present invention. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0050] Reference Figure 1 , Figure 1 The figure is a schematic diagram of the structure of a label propagation device for a weakly labeled data set in a hardware operating environment involved in an embodiment of the present invention.

[0051] like Figure 1 As shown, the label propagation device of the weak label data set may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM) memory, or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0052] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the label propagation device for weak label data sets, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0053] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a data storage module, a network communication module, a user interface module, and a label propagation program for a weak label data set.

[0054] exist Figure 1In the label propagation device for the weak label data set shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the label propagation device for the weak label data set of the present invention can be set in the label propagation device for the weak label data set, and the label propagation device for the weak label data set calls the label propagation program of the weak label data set stored in the memory 1005 through the processor 1001, and executes the label propagation method for the weak label data set provided by the embodiment of the present invention.

[0055] The embodiment of the present invention provides a label propagation method for a weakly labeled data set, referring to Figure 2 , Figure 2 Schematic diagram of the flow chart of the first embodiment of the label propagation method for weakly labeled datasets of the present invention.

[0056] In this embodiment, the label propagation method of the weakly labeled data set includes the following steps:

[0057] Step S10: construct a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and unlabeled samples are connected by an undirected edge.

[0058] It should be noted that studies have shown that incomplete labels and label noise are very common in multi-label data. For example, image samples in scene classification tasks are prone to noisy and missing labels, such as a small number of labels describing useless information, and some labels are not labeled. Generally speaking, a "weakly labeled dataset" refers to a multi-label dataset with characteristics such as incomplete labels and certain noisy labels. When collecting and manually labeling large-scale data in the real world, weak labels are difficult to avoid. The weak label nature of the dataset has reduced the performance of the multi-label learning algorithm to a certain extent, and also affected its application in real life.

[0059] It can be understood that the label propagation algorithm is a graph-based semi-supervised learning method. Its main idea is to use the label information of the labeled samples to predict the labels of unlabeled samples. The assumption based on the label propagation algorithm is the smoothness assumption, that is, for data samples with high similarity, their labels are also similar. This assumption makes it possible to propagate labels between labeled data and unlabeled data, and the possibility of label propagation depends on the similarity between samples. The greater the similarity between two samples, the greater the possibility that the labels between the two samples are similar, and thus the greater the possibility that labels between samples are propagated.

[0060] Label propagation is performed on a weighted undirected relationship graph. The weight of the edge in the graph can be used to intuitively obtain the similarity between samples, and then analyze the degree of label propagation on this edge. To this end, a graph needs to be constructed before label propagation. In the graph, labeled samples and unlabeled samples are connected by an undirected edge, and the weight of the edge represents the similarity between samples.

[0061] In specific implementations, a label propagation process can be represented by matrix multiplication. During the propagation process, each sample propagates its own label to another sample with the corresponding probability in the probability transfer matrix T. The more similar the two samples are, the easier it is to assign their own labels to the label distribution of the other party, that is, the easier it is to propagate. After each label propagation, the label of the labeled sample is set to the initial state because the labels of the labeled samples are known, and by default they are complete and correct labels and cannot be changed. With the continuous propagation of the labeled data labels, the classification decision boundary is continuously pushed from the high-density data area to the low-density data area, which means that the sample space is divided into different categories. The label propagation algorithm process is mainly represented by the probability transfer matrix and the label matrix. The changes in the label during the propagation process can be shown in the label matrix. Therefore, the probability transfer matrix and the label matrix should be constructed before the algorithm starts to ensure the label propagation process. The overall architecture of the label propagation algorithm is roughly as follows Figure 3 shown.

[0062] Furthermore, the step of constructing a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge, includes: obtaining a target sample, and establishing a label propagation graph based on the target sample; the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge; and initializing the weights corresponding to the undirected edges.

[0063] Furthermore, the step of initializing the weights corresponding to the undirected edges comprises:

[0064] Initialize the (l+u)×(l+u)-dimensional probability transfer matrix T ij ,in:

[0065]

[0066] In the specific implementation, the second step of label propagation is improved. The labeled sample receives the label information propagated from the neighboring samples, filters the label information to its own label, and then propagates the completed label to the unlabeled data. Each time the received label is processed, the information is "absorbed", the useful label information is retained, and the useless label information is discarded. Assuming that all category labels have been labeled, in this way, the original labeled sample labels can be guaranteed during the propagation, and the missing labels can be filled as much as possible. The overall architecture of the label propagation algorithm for weak labels is roughly as follows: Figure 4 shown.

[0067] In the specific implementation, the algorithm uses the undirected graph constructed above and the label iterative propagation process to analyze and solve the problem of sample x in the weakly labeled dataset. i The label of the neighboring sample x u The problem of propagation. Suppose there are n weakly labeled data X={(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x n ,y n )}. Among them, the sample x i The corresponding weak label is α i =(α i1 ,α i2 ,...,α ic ), c is the size of the label space. When the user is a sample x i When labeling the jth class, α ij =1, for other unmarked labels α ij =0. Assume that T represents the similarity measurement matrix on the dataset X, and the weak label matrix corresponding to all samples is Y=(Y 1 ,Y 2 ,...,Y n ), then for each labeled sample x i Assign initial value α i , while the unlabeled sample x u The label of x is determined by the result of label propagation. In each iteration, the labeled sample "absorbs" part of the label information propagated from the neighboring samples and retains the label information at the initial state. In this way, in the t+1th label propagation, the labeled sample x i The tags can be represented as follows:

[0068]

[0069] Among them, Y i 0 is the label state of the label matrix at the beginning. At this time, for each labeled sample xi , execute the following formula:

[0070]

[0071] Among them, 0<θ<1 is used to determine the sample x i The label part "absorbed" from the information transmitted by the neighboring samples. The larger θ is, the less likely it is to add new labels to the labeled samples, and the closer the algorithm is to the traditional label propagation algorithm; conversely, the easier it is to add new labels to the labeled samples.

[0072] And the unlabeled sample x u Then you need to execute the following formula every time the label is propagated:

[0073]

[0074] Step S20: Obtain category labels and labeled samples in the label propagation graph to construct a probability transfer matrix and a labeling matrix.

[0075] Step S30: When the labeled sample receives label information transmitted from a neighboring sample, the label information is filtered to complete the label information of the labeled sample.

[0076] Furthermore, the step of filtering the label information to complete the label information of the labeled sample when the labeled sample receives label information propagated from neighboring samples includes: when the labeled sample receives label information propagated from neighboring samples, judging useful information and useless information according to the label information; retaining the useful information and deleting the useless information; and completing the label of the labeled sample itself with the retained useful information.

[0077] Furthermore, after the step of filtering the label information to complete the label information of the labeled sample when the labeled sample receives the label information transmitted from the neighboring sample, the method further includes:

[0078] In the t+1th label propagation, the labeled sample x i The tags can be represented as follows:

[0079]

[0080] Among them, Y i 0 is the label state of the label matrix at the beginning. At this time, for each labeled sample x i , execute as follows:

[0081]

[0082] Among them, 0<θ<1 is used to determine the sample xi The label portion absorbed from the information from neighboring samples.

[0083] Furthermore, the steps in the t+1th label propagation further include:

[0084] Unlabeled sample x u During label propagation, the following is performed:

[0085]

[0086] Step S40: propagate the completed label information to the unlabeled data.

[0087] Step S50: determine whether the label matrix converges, and if so, complete label propagation.

[0088] In the specific implementation, in the early stage of label propagation, in addition to labeling unlabeled samples, the labels of labeled samples must also be completed; and in the later stage, when the labeled label matrix no longer changes, the labeling of unlabeled samples is mainly completed. Therefore, in the process of label propagation, the labels of each sample are continuously updated until the algorithm converges, that is, the label matrix Y no longer changes. The specific algorithm flow is shown in the following table:

[0089]

[0090] In step 4, the labels in the sample will be propagated to other samples. In order to make the probability distribution of the labels consistent, the probability transfer matrix needs to be normalized before that. In step 5, since the labels of the labeled samples already exist, the existing algorithm will restore the labels of the labeled samples to their initial state after each label propagation in order not to destroy the label composition of the labeled samples. To address the problem of weak labels in the labeled samples, the improved algorithm fills in the missing labels of the labeled samples and "absorbs" the label information propagated from the neighboring samples. The specific method of "absorbing" labels is to increase the labels in the labeled samples after each label propagation according to the information of the adjacent samples to fill in the sample label composition.

[0091] The time complexity of the improved label propagation algorithm needs further discussion, that is, the speed of the algorithm. The previous algorithm may only need to go through a few label propagations before obtaining the optimal result; while the improved algorithm may need to iterate many times before finally converging, and each iteration requires a label propagation, which will affect the speed of the algorithm. Although the speed of the algorithm will drop significantly, it will not increase linearly with the increase in the number of iterations. Generally speaking, the speed of label propagation is related to the initial state of the known labels. The more labeled samples there are, the faster the convergence speed. The improved algorithm also needs to complete the missing labels during label propagation. The greater the label completeness, the shorter the convergence time of the algorithm. The completeness of the labels of the improved algorithm is getting larger and larger, so when the algorithm iterates, the convergence speed of label propagation will be faster. In the algorithm of this article, there is also the existence of the label propagation stop parameter t and the upper limit control of the number of algorithm iterations. The larger the value of t, the more iterations, and vice versa. The more iterations, the greater the completeness of the labels, and the subsequent label propagation is more inclined to the label propagation before the improvement, and the algorithm convergence speed is getting faster and faster, rather than increasing linearly with the number of iterations.

[0092] This embodiment constructs a label propagation graph, which includes labeled samples and unlabeled samples, and the labeled samples and unlabeled samples are connected by an undirected edge; in the label propagation graph, category labels and labeled samples are obtained to construct a probability transfer matrix and a labeling matrix; when the labeled sample receives label information propagated from neighboring samples, the label information is filtered to complete the label information of the labeled sample; the completed label information is propagated to the unlabeled data; it is determined whether the label matrix converges, and if so, the label propagation is completed. The original label propagation algorithm is improved by adopting the idea of ​​feedback, and the missing label information of the weakly labeled data is filled, thereby improving the accuracy of label assignment.

[0093] In addition, an embodiment of the present invention further proposes a medium on which a label propagation program for a weak label data set is stored. When the label propagation program for the weak label data set is executed by a processor, the steps of the label propagation method for the weak label data set described above are implemented.

[0094] The embodiments or specific implementations of the label propagation device for the weak label data set of the present invention can refer to the above-mentioned method embodiments, which will not be repeated here.

[0095] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0096] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0097] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0098] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A label propagation method for a weakly labeled dataset, characterized in that: The method comprises: Constructing a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge; Obtaining category labels and labeled samples in the label propagation graph to construct a probability transfer matrix and a labeling matrix; When the labeled sample receives label information transmitted from a neighboring sample, the label information is filtered to complete the label information of the labeled sample, including: when the labeled sample receives label information transmitted from a neighboring sample, useful information and useless information are determined according to the label information; the useful information is retained and useless information is deleted; and the retained useful information is completed into the label of the labeled sample itself; The completed label information is propagated to the unlabeled data, including: In the t+1th label propagation, the labeled sample x i The tags can be represented as follows: Where, T=(T ij ) represents the similarity measure matrix on the labeled dataset X, Y i 0 is the label matrix Y = (Y1, Y2, ..., Y n ) Initial label state, for each labeled sample x i , execute as follows: Among them, α i =(α i1 ,α i2 ,...,α ic ) is the weak label of sample xi, c is the size of the label space; 0<θ<1 is used to determine sample x i The label part absorbed from the information transmitted by the neighboring samples; the specific method of absorbing labels is to increase the labels in the labeled samples after each label propagation according to the information of the adjacent samples to fill the sample label composition; Determine whether the label matrix converges, and if so, complete label propagation.

2. The method according to claim 1, characterized in that Before the step of filtering the label information to complete the label information of the labeled sample when the labeled sample receives the label information propagated from the neighboring sample, the method further includes: Determine that there are n weakly labeled data X={(x1,y1),(x2,y2),...,(x n ,y n )}; When the user is a sample x i When labeling the jth class, α ij =1, for other unmarked labels α ij =0; For each labeled sample x i Assign initial value α i , while the unlabeled sample x u The label of is determined by the result of label propagation.

3. The method according to claim 1, characterized in that The steps in the t+1th label propagation also include: Unlabeled sample x u During label propagation, the following is performed:

4. The method according to any one of claims 1 to 3, characterized in that The step of constructing a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, wherein the labeled samples and the unlabeled samples are connected by an undirected edge, comprises: Obtain a target sample, and establish a label propagation graph based on the target sample; The label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and unlabeled samples are connected by an undirected edge; Initialize the weight corresponding to the undirected edge.

5. The method according to claim 4, characterized in that The step of initializing the weights corresponding to the undirected edges comprises: Initialize the (l+u)×(l+u)-dimensional probability transfer matrix T ij ,in:

6. A label propagation device for a weakly labeled data set, characterized in that: The device comprises: An initialization module, used to construct a label propagation graph, wherein the label propagation graph includes labeled samples and unlabeled samples, and the labeled samples and the unlabeled samples are connected by an undirected edge; A matrix construction module, used for obtaining category labels and labeled samples in the label propagation graph to construct a probability transfer matrix and a labeling matrix; The label completion module is used to filter the label information to complete the label information of the labeled sample when the labeled sample receives the label information transmitted from the neighboring sample, including: when the labeled sample receives the label information transmitted from the neighboring sample, judging useful information and useless information according to the label information; retaining the useful information and deleting the useless information; and completing the label of the labeled sample itself with the retained useful information; The label propagation module is used to propagate the completed label information to the unlabeled data, including: In the t+1th label propagation, the labeled sample x i The labels can be represented as follows: Where, T=(T ij ) represents the similarity measure matrix on the labeled dataset X, Y i 0 is the label matrix Y = (Y1, Y2, ..., Y n ) Initial label state, for each labeled sample x i , execute as follows: Among them, α i =(α i1 ,α i2 ,…,α ic ) is the sample x i weak label, c is the size of the label space; 0<θ<1 is used to determine the sample x i The label part absorbed from the information transmitted by the neighboring samples; the specific method of absorbing labels is to increase the labels in the labeled samples after each label propagation according to the information of the adjacent samples to fill the sample label composition; The convergence judgment module is used to judge whether the label matrix converges. If so, label propagation is completed.

7. A label propagation device for a weakly labeled data set, characterized in that: The device comprises: a memory, a processor, and a label propagation program for a weakly labeled data set stored in the memory and executable on the processor, wherein the label propagation program for the weakly labeled data set is configured to implement the steps of a label propagation method for a weakly labeled data set as described in any one of claims 1 to 5.

8. A medium, characterized in that The medium stores a label propagation program for a weak label data set, and when the label propagation program for the weak label data set is executed by a processor, the steps of the label propagation method for a weak label data set according to any one of claims 1 to 5 are implemented.

Images