Partial label feature selection method, device, equipment and medium for dynamic flow labels
The method addresses the challenge of feature selection in dynamic and noisy partial label learning environments by evaluating label confidence, disambiguating labels, and constructing an objective function to select optimal features, improving classification performance.
Patent Information
- Application Number
- CN202411146635.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-21
AI Technical Summary
Existing partial label learning methods struggle with selecting suitable feature subsets in dynamic and noisy environments, as they are designed for static label spaces and are prone to mislead feature selection due to inaccurate label information.
A method and device for partial label feature selection in dynamic flow labels that evaluates label confidence using neighborhood information, disambiguates labels, analyzes dynamic relationships, and constructs an objective function to select optimal features based on maximum conditional relevance and minimum conditional redundancy.
Improves classification performance by dynamically disambiguating labels, reducing noise impact, and selecting high-quality features, thereby enhancing the generalization performance of partial label learning systems.
Smart Images

Figure CN118656797B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data mining, and in particular to a method, device, equipment and medium for selecting partial label features of dynamic flow labels. Background Art
[0002] Partial label learning (PLL), as a weakly supervised learning paradigm, aims to learn a multi-class classifier from ambiguous instances, where each instance is associated with a set of candidate labels, only one of which is valid. Compared with ordinary supervised learning problems, PLL avoids the impractical overhead of manually labeling each instance with the true label. In recent years, due to the increasing demand for identifying true labels, the task of PLL naturally arises in different application areas such as network mining, eco-informatics, and natural language processing.
[0003] Due to the limited effective supervision information provided by the training set, the generalization performance of partial label learning systems often falls short of expectations. From the perspective of data generation, the presence of noisy, irrelevant, and redundant features can lead to an ambiguous labeling process, making it difficult for the algorithm to distinguish between potential true labels and incorrect labels. In the face of these problems, effective feature selection is a key step, as it not only achieves a more compact feature representation, but also increases the density of instances. In addition, it enhances the algorithm's ability to distinguish between correct and incorrect labels.
[0004] However, whether for single-label or multi-label scenarios, traditional feature selection methods are not suitable for partial label learning because inaccurate label information will confuse the importance of features and mislead the selection process. The inherent ambiguity in label information makes it particularly difficult to select a subset of features that are both informative and discriminative.
[0005] As an effective preprocessing strategy, partial label feature selection has received extensive attention in improving the generalization performance of partial label learning. Existing methods assume that the learning task is performed in a static and known label space environment. However, in reality, the environment is open and labels may arrive dynamically one after another. This makes it particularly challenging to select a suitable feature subset from partial label data. Currently, a variety of partial label feature selection methods have been introduced, but most of them only consider the static labels and cannot work well when the labels arrive dynamically.
[0006] In view of this, this application is filed. Summary of the invention
[0007] The present invention provides a method, device, equipment and medium for partial label feature selection of dynamic flow labels, which can at least partially improve the above-mentioned problems.
[0008] To achieve the above object, the present invention adopts the following technical solutions:
[0009] A partial label feature selection method for dynamic flow labels, comprising:
[0010] When the judgment label arrives dynamically, in the dynamic label space, the confidence of each label in the candidate label set as the true value label is evaluated according to the neighborhood information of the sample, and whether to perform label disambiguation processing is determined according to the preset constraint criteria;
[0011] Analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features to obtain feature enhancement items;
[0012] In combination with the feature enhancement item, an objective function considering related items and redundant items is constructed to select the optimal feature for each arrived label;
[0013] According to the objective function, feature sets corresponding to different labels are fused to generate a final feature set.
[0014] The present invention also provides a partial label feature selection device for dynamic flow labels, which includes:
[0015] A label dynamic disambiguation unit is used to evaluate the confidence of each label in the candidate label set as a true value label according to the neighborhood information of the sample in the dynamic label space when the label is dynamically arrived, and to determine whether to perform label disambiguation processing according to the preset constraint standard;
[0016] An enhancement item generation unit is used to analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features to obtain feature enhancement items;
[0017] A partial label feature selection unit, used to construct an objective function that considers related items and redundant items in combination with the feature enhancement items, and select an optimal feature for each arrived label;
[0018] The feature fusion unit is used to fuse the feature sets corresponding to different labels according to the objective function, and fuse them to generate a final feature set.
[0019] The present invention also provides a partial label feature selection device for dynamic flow labels, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the partial label feature selection method for dynamic flow labels as described in any one of the above.
[0020] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program can be executed by a processor of a device where the computer-readable storage medium is located to implement a partial label feature selection method for dynamic flow labels as described in any one of the above.
[0021] In summary, the partial label feature selection method for dynamic flow labels mainly includes four steps: First, the confidence that the candidate label is the true label is dynamically adjusted according to the probability distribution of the sample to achieve dynamic disambiguation of the label. Secondly, the dynamic relationship between the newly arrived label and the disambiguated label is analyzed to evaluate its impact on the feature quality. Next, the feature enhancement term is defined to quantify the above impact and reduce the error of the feature evaluation caused by the noisy label. On this basis, a feature evaluation criterion based on maximum conditional relevance and minimum conditional redundancy is designed, which simultaneously weighs the importance of the feature relative to the current label and the disambiguated label. Finally, a feature fusion strategy is constructed to fuse the feature sets corresponding to the dynamic flow labels to obtain a final set of high-quality features. It aims to handle the situation where labels arrive dynamically, eliminate label noise, and evaluate the feature quality through the maximum conditional relevance and minimum conditional redundancy strategies, thereby improving the classification performance of the learning model; that is, to select high-quality features in a dynamic and noisy label environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 1 is a flow chart of a partial label feature selection method for dynamic flow labels provided by the first embodiment of the present invention;
[0023] Figure 2 It is a module schematic diagram of a partial label feature selection device for dynamic flow labels provided in the second embodiment of the present invention. DETAILED DESCRIPTION
[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments. 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.
[0025] refer to Figure 1 As shown, the first embodiment of the present invention discloses a partial label feature selection method for dynamic flow labels, which can be performed by a partial label feature selection device for dynamic flow labels (hereinafter referred to as a selection device), and in particular, is performed by one or more processors in the selection device to implement the following method:
[0026] S1, when it is determined that the label arrives dynamically, in the dynamic label space, the confidence of each label in the candidate label set as the true value label is evaluated according to the neighborhood information of the sample, and whether to perform label disambiguation processing is determined according to the preset constraint criteria;
[0027] Specifically, step S1 includes: constructing a label confidence matrix for the currently known label set , calculate the true confidence of each label of each sample, where each element Representation Tags become The estimated confidence of the true value label is initialized as follows:
[0028]
[0029] in, express The base number, is the set of candidate labels, is the label confidence matrix the number of rows, is the label confidence matrix The number of columns and label confidence The following conditions must be met: , , ;
[0030] Calculate the probability that each label in the sample candidate label is true, and evaluate the probability that the current arrival label is true based on the candidate label information of the neighboring samples. The calculation formula for the true value probability is:
[0031]
[0032] Among them, if the condition " "Established, Returns 1, otherwise returns 0. represents the potential true label, is an instance The candidate label set, is the number of classes corresponding to each label, is a smoothing parameter, , For example The nearest neighbor sample set of ;
[0033] While calculating the probability of the true value label, the probability that each label in the candidate label is a noise label is considered according to the neighborhood information of the sample. The calculation formula is:
[0034]
[0035] in, is to remove the potential true label The final candidate label set Other candidate tags in ;
[0036] According to the preset constraint criteria, determine whether to disambiguate the candidate label. The formula is:
[0037]
[0038] Among them, through the constraint parameters , determine the constraint judgment result The size of ,Label will be regarded as Uncertain noise labels, do not label Perform disambiguation processing if ,Label yes uncertain noise labels and eliminate them in time uncertainty;
[0039] Repeat the above steps until no new tags arrive.
[0040] In this embodiment, the first step is to evaluate the confidence of each label in the candidate label set as a true value label through the neighborhood information of the sample in the dynamic label space, and perform label disambiguation; that is, to process the newly arrived labels, disambiguate each label when it arrives for the first time, use the sample neighborhood information to evaluate the true probability of the candidate label, and determine the disambiguation or delay the disambiguation according to the label arrival situation.
[0041] To handle candidate label sets with noisy labels, the true confidence of each label needs to be calculated for each sample. First, a label confidence matrix is constructed for the currently known label set. , where each element Representation Tags become The estimated confidence of the true value label is initialized as:
[0042]
[0043] in express The cardinality of the label. Satisfies: (1) , (2) , (3) The second constraint means that the original label of each sample is in its set of candidate labels, while the third constraint ensures that the confidence of any non-candidate label must be 0.
[0044] In order to more accurately eliminate noise labels, the label confidence is updated using the sample neighborhood information. First, the probability of each label in the sample candidate label being true is calculated, and the probability of the current arrival label being true is evaluated based on the candidate label information of the neighborhood sample. The true value probability calculation formula is as follows:
[0045]
[0046] Among them, if the condition " " is established, then Returns 1 if yes, otherwise returns 0; represents the potential true label; is an instance The candidate label set of is the number of classes corresponding to each label. In addition, is a smoothing parameter used to avoid the situation where the probability is equal to 0. .
[0047] The above formula does not take into account the situation where there are noisy labels in the sample neighborhood information. To solve this problem, it is necessary to consider the probability of each label in the candidate label being a noisy label while calculating the probability of the true value label based on the sample neighborhood information. The calculation formula is as follows:
[0048]
[0049] in, Yes Remove back Other candidate tags in .
[0050] Obtain the probability of each candidate label of the sample and update the label confidence according to the formula. The formula is as follows:
[0051]
[0052] in the formula and Used to balance the instance itself and its If the above confidence update formula is used directly, the result may fall into local optimization, ignoring that the order of label arrival may have a significant impact on the disambiguation result. To solve this problem, a more detailed constraint standard is proposed to determine whether to disambiguate the candidate label.
[0053]
[0054] By constraining parameters , it can be determined The size of , then the label will be regarded as Uncertain noise labels, do not label To disambiguate. , then the label yes The uncertain noise labels will be eliminated in time After that, the iterative process of candidate label elimination is repeated until no new labels arrive. Finally, when all labels arrive, the candidate labels may not be able to completely eliminate the uncertainty, and the label with the highest confidence will be selected as the true label of each instance.
[0055] S2, analyzes the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantifies the impact of labels arriving at different times on features to obtain feature enhancement items;
[0056] Specifically, step S2 includes: calculating the mutual information difference and analyzing it, and the calculation formula is:
[0057]
[0058] in, is the mutual information difference, The value of can be positive, negative or zero. When it is a positive number, it means that when the new label When known, Ability to add a known tag Provide additional information, and Relative to the feature Positively correlated, when When it is a negative number, it means that when the label Known, For known tags The amount of information provided is reduced. and Negatively correlated, when When it is equal to zero, it means For label The amount of information provided is not limited by the label The impact of tags and tags Relative to the feature is independent, for The known label set at time, For new label Arrival conditions, features For the reached label The amount of information provided, feature For the reached label the amount of information provided;
[0059] Define feature enhancement items and quantify the impact of label relationships on feature evaluation based on the feature enhancement items. For new labels, , the additional information provided by the feature for the known label after disambiguation is called feature enhancement, and its definition formula is:
[0060]
[0061] in, Feature enhancement items.
[0062] In this embodiment, in a dynamic partial label environment, the relationship between labels is also changing dynamically as labels arrive. This relationship plays different roles in feature evaluation, so it is necessary to analyze and quantify the impact of label relationships on feature evaluation. The second step is to analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features; that is, quantitative analysis of the impact that label relationships may have on feature evaluation. Under the premise of dynamic label arrival, the relationship between different labels is also changing dynamically, and feature enhancement items are defined using conditional mutual information.
[0063] The influence of quantified label relationship on feature evaluation is analyzed by calculating the mutual information difference. The specific formula is as follows:
[0064]
[0065] in Can be positive, negative or zero. When it is a positive number, it means that when the new label When known, Can be a known tag Provide additional information, i.e. and Relative to the feature On the contrary, if If it is a negative number, it means that when the label Known, For known tags The amount of information provided is reduced, i.e. and Negatively correlated. When it is equal to zero, it means For label The amount of information provided is not limited by the label The impact of label and tags Relative to the feature Is independent.
[0066] Define feature enhancement terms, which quantify the impact of label relationships on feature evaluation. , the additional information provided by the feature for the known label after disambiguation is called feature enhancement, which is defined as:
[0067]
[0068] Obviously, The minimum value of is 0, that is, the minimum amount of information provided by the feature enhancement item should be greater than or equal to 0.
[0069] S3, constructing an objective function that considers related items and redundant items in combination with the feature enhancement items, and selecting the optimal feature for each arrived label;
[0070] Specifically, step S3 includes: defining feature related items, and after disambiguating known labels, using conditional mutual trust to represent feature related items, the formula is:
[0071]
[0072] Combine the feature enhancement items to construct the final feature related items, where the feature set , is the selected feature set, , the labels evaluated according to the maximum conditional relevance criterion Corresponding features The score of the feature is:
[0073]
[0074] The above formula is described in detail as the objective function formula:
[0075]
[0076] Define feature redundancy and use all selected features With non-selected features It is measured by the conditional mutual information between for The formula for the new arrival label at time is:
[0077]
[0078] Combine the feature related items and feature redundant items to obtain the final feature score objective function, the formula is:
[0079]
[0080] Among them, the characteristics The score is evaluated by the MCRCR criteria, and the label Corresponding.
[0081] In this embodiment, a new feature selection strategy, the maximum conditional correlation and minimum conditional redundancy criteria, is proposed in combination with feature enhancement items to reduce the negative impact of noise labels on feature evaluation. In the third step, based on the quantitative terms describing the correlation, an objective function that considers the correlation terms and redundancy terms is constructed to select the optimal features for each arrived label; that is, a feature set that provides the most information is selected, and the features are evaluated according to the maximum conditional correlation and minimum conditional redundancy criteria to select a set of features with maximum conditional correlation and minimum conditional redundancy with the label.
[0082] Define feature dependencies and use conditional mutual trust to represent them after knowing the labels for disambiguation:
[0083]
[0084] Combine the proposed feature enhancement items to construct the final feature correlation items. Given a feature set , is the selected feature set, , the label evaluated by the maximum conditional relevance criterion Corresponding features The score of is defined as follows:
[0085]
[0086] The above formula can be described in detail as the following objective function formula:
[0087]
[0088] Define feature redundancy by passing all selected features With non-selected features The conditional mutual information between them is used to measure, given the feature set , is the selected feature set, , , , for The new arrival label at time , then the feature condition redundancy is:
[0089]
[0090] Combine the feature correlation terms and feature redundancy terms to obtain the final feature score objective function. The score is evaluated by the MCRCR criteria, and the label Correspondingly, the algorithm provides the most useful information features for each flow label, and at the same time helps the selected features contribute to the known labels after disambiguation. The algorithm pursues features that can continuously adjust feature scores and reduce the adverse effects of noise annotations. The objective function is defined as follows:
[0091]
[0092] S4, according to the objective function, the feature sets corresponding to different labels are merged to generate a final feature set.
[0093] Specifically, step S4 includes: calculating a set of feature scores for each flow label by a feature score formula, and superimposing the scores of each feature when the label flows in, fusing to generate a final feature set, wherein when When, characteristics For the current arrival tag set The score is:
[0094]
[0095] in, Used for calculation Time to arrive label The corresponding feature score, The larger the value of For the current arrival label The more important.
[0096] In this embodiment, the last step is to merge the feature sets corresponding to different labels into a final feature set according to the constructed objective function; that is, to merge the different optimal feature sets obtained by different labels, and to merge the category features that are only applicable to a single label into a feature set applicable to all labels that have been reached according to the constructed feature evaluation function.
[0097] The feature score formula can be used to obtain a set of feature scores for each flow label. However, as labels arrive dynamically, each feature has a specific score for different labels. Therefore, it is necessary to superimpose the scores of each feature as labels flow in. Given a feature set , For the selected feature set , For the time The known label set at time, for The new arrival class label at time, , then the feature For the current arrival tag set The score is:
[0098]
[0099] Used for calculation Time to arrive label The corresponding feature score. The larger the value of For the current arrival label The more important. The feature selected in each round should be the one with the highest score. In other words, the feature selection based on MCRCR is to maximize Values:
[0100]
[0101] In summary, when labels arrive dynamically, the newly arrived labels are initially disambiguated, the label confidence is estimated using sample neighborhood information, and the label probability is balanced according to the set smoothing parameter. After the new label is disambiguated, feature selection is performed on the label. In the process of evaluating the feature score, the dynamic relationship between labels is considered and the score is corrected. When no new labels arrive, the final label disambiguation is completed and the different features of different labels are fused according to the constructed feature evaluation function, and finally a feature set applicable to all labels is obtained.
[0102] See also Figure 2 The second embodiment of the present invention further provides a partial label feature selection device for dynamic flow labels, which includes:
[0103] The label dynamic disambiguation unit 201 is used to evaluate the confidence of each label in the candidate label set as a true value label according to the neighborhood information of the sample in the dynamic label space when the label is determined to be dynamically arrived, and determine whether to perform label disambiguation processing according to the preset constraint standard;
[0104] An enhancement item generating unit 202 is used to analyze the dynamic relationship between the feature and the disambiguated labels and the newly arrived labels, and quantify the influence of the labels arriving at different times on the feature to obtain the feature enhancement item;
[0105] A partial label feature selection unit 203, used to construct an objective function that considers related items and redundant items in combination with the feature enhancement items, and select an optimal feature for each arrived label;
[0106] The feature fusion unit 204 is used to fuse the feature sets corresponding to different tags according to the objective function to generate a final feature set.
[0107] The third embodiment of the present invention also provides a partial label feature selection device for dynamic flow labels, including a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the partial label feature selection method for dynamic flow labels as described in any one of the above.
[0108] The fourth embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which can be executed by a processor of a device where the computer-readable storage medium is located to implement the partial label feature selection method for dynamic flow labels as described in any one of the above items.
[0109] Exemplarily, the above-mentioned devices and process steps can be implemented by a computer program. The computer program can be divided into one or more units. The one or more units are stored in the memory and executed by the processor to complete the present invention.
[0110] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0111] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the present invention by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.
[0112] Wherein, if the electronic device or printer integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.
[0113] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.
[0114] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A partial label feature selection method for dynamic flow labels, characterized in that: include: When the judgment label arrives dynamically, in the dynamic label space, the confidence of each label in the candidate label set as the true value label is evaluated according to the neighborhood information of the sample, and whether to perform label disambiguation processing is determined according to the preset constraint criteria; Analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features to obtain feature enhancement items; In combination with the feature enhancement item, an objective function considering related items and redundant items is constructed to select the optimal feature for each arrived label; according to the objective function, feature sets corresponding to different labels are fused to generate a final feature set; When the judgment label arrives dynamically, in the dynamic label space, the confidence of each label in the candidate label set as the true value label is evaluated according to the neighborhood information of the sample, and whether to perform label disambiguation processing is determined according to the preset constraint criteria, specifically: Construct a label confidence matrix for the currently known label set Calculate the true confidence of each label for each sample, where each element P ij represents the potential true label l j Become x i The estimated confidence of the true value label is initialized as follows: Among them, |S i | indicates S i The cardinality, S i is the set of candidate labels, n is the label confidence matrix P t-1 The number of rows, q t-1 is the label confidence matrix P t-1 The number of columns, and the label estimation confidence P ij The following conditions must be met: Calculate the probability that each label in the sample candidate label is true, and evaluate the probability that the current arrival label is true based on the candidate label information of the neighboring samples. The calculation formula for the probability of the true value label is: Among them, if the condition l j ∈S a Established, j ∈S a ]] returns 1, otherwise returns 0, l j represents the potential true label, S a is instance x a The candidate label set, N is the number of classes corresponding to each label, λ is a smoothing parameter, λ=1, N(x i ) is an instance x i The nearest neighbor sample set of ; While calculating the probability of the true value label, the probability that each label in the candidate label is a noise label is considered according to the neighborhood information of the sample. The calculation formula is: Among them, l k is to remove the potential true label l j The candidate label set S a Other candidate tags in ; According to the preset constraint criteria, determine whether to disambiguate the candidate label. The formula is: m ij =Max(P t-1 )-P ij Among them, the constraint judgment result μ is determined by the constraint parameter η, 0<η<1 ij The size of ij ≤η, potential true label l j will be considered as instance x i The uncertain noise label is not yet known for the potential true label l j Perform disambiguation processing, if μ ij >η, potential true label l j is instance x i uncertain noise labels and timely eliminate potential true labels l j uncertainty; Repeat the above steps until no new tags arrive.
2. The partial label feature selection method for dynamic flow labels according to claim 1 is characterized in that: Analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features, specifically: Calculate the mutual information difference and analyze it. The calculation formula is: Among them, Lc(f i , l t , l k ) is the mutual information difference, where the feature set F = {f1, f2, ..., f q }, is the selected feature set, f i ∈FA,l t is the new arrival label at time t; Lc(f i , l t , l k ) is a positive, negative or zero value. i , l t , l k ) is a positive number, indicating that when the new label l t When f is known, i Can be a known label l k Provide additional information, t and l k Relative to feature f i Positively correlated, when Lc(f i , l t , l k ) is a negative number, which means that when the new label l t It is known that f i For a known label l k The amount of information provided is reduced, t and l k Negatively correlated, when Lc(f i , l t , l k ) is equal to zero, indicating that f i For label l k The amount of information provided is not limited by the new label t The impact of new labels t and label l k Relative to feature f i is independent, is the known label set at time t-1, I(f i ; l k |l t ) is the new label l t Under the condition of arrival, the feature f i The label l has been reached k The amount of information provided, I(f i ; l k ) Feature f i The label l has been reached k the amount of information provided; Define feature enhancement items and quantify the impact of label relationships on feature evaluation according to the feature enhancement items. For the new label l t , the additional information provided by the feature for the known label after disambiguation is called feature enhancement, and its definition formula is: Men(f i ,l t ,l k )=max{Lc(f i ,l t ,l k ),0} Among them, Men(f i , l t , l k ) is a feature enhancement item.
3. The partial label feature selection method for dynamic flow labels according to claim 2 is characterized in that: Combined with the feature enhancement items, an objective function that considers related items and redundant items is constructed to select the optimal feature for each arriving label. Specifically, feature related items are defined, and after disambiguation of known labels, conditional mutual trust is used to represent feature related items. The formula is: Combined with the feature enhancement items, the final feature correlation items are constructed, and the label l evaluated according to the maximum conditional correlation criterion t The corresponding feature f i The score of the feature is: J MCR (f i ,l t )=Mre(f i ,l t ,l k )+Men(f i ,l t ,l k ) The above formula is described in detail as the objective function formula: Define feature redundancy and use all selected features f j With non-selected features f i The conditional mutual information between them is used to measure it, and the formula for the feature conditional redundancy term is: Combine the feature related items and feature redundant items to obtain the final feature score objective function, the formula is: Among them, the feature f i The score of is evaluated by the MCRCR criterion and the label l t Corresponding.
4. The partial label feature selection method for dynamic flow labels according to claim 3 is characterized in that: According to the objective function, the feature sets corresponding to different labels are fused to generate the final feature set, which is specifically: A set of feature scores for each flow label is calculated through the feature score formula, and the scores of each feature are superimposed when the label flows in, and the final feature set is generated by fusion. When the characteristic f i For the current arrival tag set The score is: in, Used to calculate the arrival label at time t-1 The corresponding feature score, The larger the value of i For the current arrival label The more important.
5. A partial label feature selection device for dynamic flow labels, characterized in that: include: A label dynamic disambiguation unit is used to evaluate the confidence of each label in the candidate label set as a true value label according to the neighborhood information of the sample in the dynamic label space when the label is dynamically arrived, and to determine whether to perform label disambiguation processing according to the preset constraint standard; An enhancement item generation unit is used to analyze the dynamic relationship between features and disambiguated labels and newly arrived labels, and quantify the impact of labels arriving at different times on features to obtain feature enhancement items; A partial label feature selection unit, used to construct an objective function that considers related items and redundant items in combination with the feature enhancement items, and select an optimal feature for each arrived label; A feature fusion unit, used to fuse feature sets corresponding to different labels according to the objective function to generate a final feature set; When the judgment label arrives dynamically, in the dynamic label space, the confidence of each label in the candidate label set as the true value label is evaluated according to the neighborhood information of the sample, and whether to perform label disambiguation processing is determined according to the preset constraint criteria, specifically: Construct a label confidence matrix for the currently known label set Calculate the true confidence of each label for each sample, where each element P ij represents the potential true label l j Become x i The estimated confidence of the true value label is initialized as follows: Among them, |S i | indicates S i The cardinality, S i is the set of candidate labels, n is the label confidence matrix P t-1 The number of rows, q t-1 is the label confidence matrix P t-1 The number of columns, while estimating the label confidence P ij The following conditions must be met: Calculate the probability that each label in the sample candidate label is true, and evaluate the probability that the current arrival label is true based on the candidate label information of the neighboring samples. The calculation formula for the probability of the true value label is: Among them, if the condition l j ∈S a Established, j ∈S a ]] returns 1, otherwise returns 0, l j represents the potential true label, S a is instance x a The candidate label set, N is the number of classes corresponding to each label, λ is a smoothing parameter, λ=1, N(x i ) is an instance x i The nearest neighbor sample set of ; While calculating the probability of the true value label, the probability that each label in the candidate label is a noise label is considered according to the neighborhood information of the sample. The calculation formula is: Among them, l k is to remove the potential true label l j The candidate label set S a Other candidate tags in ; According to the preset constraint criteria, determine whether to disambiguate the candidate label. The formula is: m ij =Max(P t-1 )-P ij Among them, the constraint judgment result μ is determined by the constraint parameter η, 0<η<1 ij The size of ij ≤η, potential true label l j will be considered as instance x i The uncertain noise label is not yet known for the potential true label l j Perform disambiguation processing, if μ ij >η, potential true label l j is instance x i uncertain noise labels and timely eliminate labels l j uncertainty; Repeat the above steps until no new tags arrive.
6. A partial label feature selection device for dynamic flow labels, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the partial label feature selection method for dynamic flow labels as described in any one of claims 1 to 4.
7. A computer-readable storage medium, characterized in that: A computer program is stored, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement the partial label feature selection method for dynamic flow labels as described in any one of claims 1 to 4.
