Partial mark feature selection method based on particle ball prototype disambiguation and related device

Abstract

The invention discloses a partial mark feature selection method based on particle ball prototype disambiguation and a related device, and the method comprises the steps: obtaining a data set, constructing initial coarse-grained particle balls, dynamically splitting the coarse-grained particle balls, and calculating the feature value of each divided particle ball and the mark of each particle ball; calculating a prototype vector of each mark of all the pellets under the mark, and calculating the similarity between the feature vector of each sample and the prototype vector; determining a weight voting matrix based on the similarity ratio, and iteratively updating the marked space confidence coefficient based on the weight voting matrix to obtain a marked space confidence coefficient matrix; introducing a global mark distribution matrix, establishing a disambiguation-feature selection bidirectional optimization model, and constructing constraint terms to be embedded into the model; and carrying out end-to-end global optimization on the model by adopting an alternating iteration updating strategy, and outputting a feature sorting result to obtain a feature subset with the highest identification capability.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a partial-label feature selection method and related device based on granule sphere prototype disambiguation. Background Art

[0002] Partial-label learning (PLL) is an important weakly-supervised machine learning framework. Under this framework, the task of partial-label learning is to learn a multi-class classifier based on training examples, where each example is associated with a set of candidate labels, but only one label is the true one. Different from traditional multi-class classification problems that deal with precise supervision information, partial-label learning deals with imprecise supervision information, which greatly reduces the annotation cost and makes it widely used in many fields in the real world, such as ecological informatics, multimedia content analysis, and natural language processing.

[0003] Since the supervision information obtained by the partial-label classification model from training examples is limited, its generalization performance usually cannot reach the expectation. Analyzing from the data generation mechanism, the noise, redundancy, and irrelevant features existing in the high-dimensional feature space will have a coupling effect with the ambiguity contained in the candidate label set: on the one hand, the existence of irrelevant features will exacerbate the ambiguity of candidate labels, making it difficult for the model to distinguish between true labels and false positive labels; on the other hand, the ambiguity of candidate labels will in turn interfere with the accurate evaluation of feature importance. By performing effective feature selection, not only can irrelevant or redundant features be eliminated to reduce the computational complexity of the learning process, but also discriminative features can be selected to provide more critical feature support for subsequent classification tasks.

[0004] The existing partial-label feature selection methods have achieved certain effects, but there are still some deficiencies: The existing methods adopt a two-stage framework of disambiguation first and then feature selection. The feature selection result cannot reverse-optimize the disambiguation process, resulting in the model falling into a local optimum, and there is still residual noise information in the data after disambiguation. And the subsequent feature selection is based on this data, resulting in error accumulation and affecting the model performance; in addition, when the model evaluates the feature importance, it usually uses the information of relevant labels (1 label). However, the relevant labels in partial-label data are uncertain. Due to the existence of false positive labels, being labeled as 1 does not necessarily represent the true label. Directly relying on such uncertain information leads to deviation in feature weight allocation. False positive labels will mislead the model to assign high weights to irrelevant features, resulting in the model being unable to accurately identify discriminative features. Summary of the Invention

[0005] In view of the above situation, it is necessary to provide a partial-label feature selection method and related device based on granule sphere prototype disambiguation for the problem of inaccurate partial-label feature recognition in the prior art.

[0006] On the one hand, the present invention provides a partial-label feature selection method based on granule prototype disambiguation, including: Obtain real partial-label datasets from different fields; Use the information of the entire dataset to construct initial coarse-grained granules, and dynamically split the coarse-grained granules by a partitioning method, calculate the values of the features of each granule after partitioning, and the labels of each granule; Based on the granules after dynamic splitting, for each label, calculate the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label, and use the cosine similarity to calculate the similarity between the feature vector of each sample and the prototype vector of each label; Determine the weight voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weight voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, obtain the label space confidence matrix, and use it as the pseudo-label distribution matrix to complete label disambiguation; Use linear projection to map the high-dimensional feature space to the label distribution space, introduce the global label distribution matrix, establish a disambiguation-feature selection two-way optimization model by minimizing the error loss with the pseudo-label distribution matrix, and at the same time construct a constraint term based on the similarity between samples and the contribution of irrelevant labels and embed it into the model; Adopt an alternating iterative update strategy to globally optimize the model end-to-end, and output the feature ranking result to obtain the most discriminative feature subset.

[0007] Further, in the above partial-label feature selection method based on granule prototype disambiguation, the step of dynamically splitting the coarse-grained granules by a partitioning method, calculating the values of the features of each granule after partitioning, and the labels of each granule includes: The coarse-grained granules Iteratively and dynamically split the granules by a partitioning method. After iterations, until all granules meet the quality requirements, thus splitting the entire dataset into disjoint granule sets ; For each split granule, calculate the mean value of the feature values of each sample in each granule to obtain the values of each feature of each granule; for each label , determine the label of each granule according to the proportion of 1 of each sample in the granule on this label.

[0008] Further, in the above partial-label feature selection method based on granule prototype disambiguation, the step of calculating the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label includes: For each label , the prototype vector of each label is obtained by calculating the weighted mean of the feature vectors of all granules under the label , and the calculation formula is as follows: , ; where, represents the number of granules under label , represents the -th granule under label , is the feature vector of granule , is the weight of the granule, is the number of candidate labels of granule .

[0009] Furthermore, in the above partial label feature selection method based on granule prototype disambiguation, the step of iteratively updating the label space confidence based on the weight voting matrix includes: For sample , the weight of its corresponding label is determined by the similarity ratio between the sample and the prototype corresponding to the label, and the weight voting matrix is defined as follows: ; where, is the similarity between sample and prototype vector , represents the number of labels, n represents the number of samples, is the candidate label set related to , is the k-th label, means returning 1 if the expression in the parentheses is true, otherwise returning 0; Initialize the label space confidence matrix , specifically as follows: ; Based on the weight voting matrix iteratively update the label space confidence , until the maximum number of iterations is reached, and the iteration formula is: ; where, is the balance parameter, which is used to balance the importance of the original label space and the predicted label space.

[0010] Furthermore, in the above partial label feature selection method based on granule prototype disambiguation, the disambiguation-feature selection two-way optimization model is: ; wherein, represents a feature matrix, is a weight matrix for feature pair marking, is a global marking distribution matrix, is a mapping matrix, is a balance parameter; The steps of constructing a constraint term based on the similarity between samples and the contribution of irrelevant markings and embedding it into the model include: Using the error marking information determined by the marking space to explore the contribution of irrelevant markings between samples, and constructing a constraint term as follows: ; where and respectively represent the irrelevant marking vectors of samples and , is an irrelevant marking matrix, represents the Hadamard product, X i , X j respectively represent the i-th row feature matrix and the j-th row feature matrix; Embedding the constraint term constructed based on the similarity between samples and the contribution of irrelevant markings into the model, the final objective function to be optimized is obtained as follows: ; wherein, is the graph Laplacian matrix of the sample similarity matrix , represents the similarity between samples and , is a diagonal matrix, and each element is , and L2 represents the graph Laplacian function.

[0011] Furthermore, for the above-mentioned partial marking feature selection method based on granule-ball prototype disambiguation, wherein the steps of performing end-to-end global optimization on the model by adopting an alternating iteration update strategy and outputting a feature ranking result to obtain the most discriminative feature subset include: Adopting an alternating iteration update strategy for optimization, fixing , and updating , the optimization problem is reformulated as follows: ; Fixing , and updating , the optimization problem is reformulated as follows: ; Then, the accelerated proximal gradient descent method is adopted to solve this optimization problem; By calculating Evaluate the feature importance and output the feature subset sorted in descending order.

[0012] On the other hand, the present invention also provides a partial-label feature selection device based on granule prototype disambiguation, including: An acquisition module, configured to acquire real partial-label data sets from different fields; A dynamic splitting granule module, which constructs initial coarse-grained granules by using the information of the entire data set, and dynamically splits the coarse-grained granules by a partitioning method, calculates the values of the features of each granule after partitioning, and the labels of each granule; A labeled prototype vector generation module, configured to, based on the granules after dynamic splitting, for each label, calculate the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label, and calculate the similarity between the feature vector of each sample and the prototype vector of each label by using the cosine similarity; A pseudo-label distribution matrix generation module, configured to determine a weight voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weight voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, obtain a label space confidence matrix, and use it as a pseudo-label distribution matrix to complete label disambiguation; A disambiguation-feature selection two-way optimization model construction module, which projects the high-dimensional feature space into the label distribution space by linear projection, introduces a global label distribution matrix, establishes a disambiguation-feature selection two-way optimization model by minimizing the error loss with the pseudo-label distribution matrix, and at the same time constructs a constraint term based on the similarity between samples and the contribution of irrelevant labels and embeds it into the model; A feature subset output module, which globally optimizes the model end-to-end by using an alternating iterative update strategy and outputs the feature sorting result to obtain the most discriminative feature subset.

[0013] On the other hand, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the partial-label feature selection method based on granule prototype disambiguation described in any one of the above.

[0014] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, it implements the partial-label feature selection method based on granule prototype disambiguation described in any one of the above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: By dynamically splitting the granule balls to calculate the prototype vectors of the labels, the weighted mean strategy constrained by the granule balls combined with the maximum margin convergence strategy significantly reduces the computational overhead and improves the disambiguation efficiency; By minimizing the error loss between the global label distribution matrix and the pseudo-label distribution matrix, a disambiguation-feature selection two-way optimization model is established, and the alternating iteration update strategy is used to perform end-to-end global optimization on the model, avoiding the local optimum of the traditional two-stage method; By considering the similarity between samples and the contribution of irrelevant labels, the model's ability to identify noise and redundant features is enhanced, and the performance and accuracy of the model are improved. Description of the Drawings

[0016] Figure 1 It is a schematic flowchart of the partial-label feature selection method based on granule prototype disambiguation provided by the first embodiment of the present invention; Figure 2 It is a schematic framework diagram of the partial-label feature selection method based on granule prototype disambiguation provided by the first embodiment of the present invention; Figure 3 It is a schematic module diagram of the partial-label feature selection device based on granule prototype disambiguation provided by the second embodiment of the present invention; Figure 4 It is a schematic structural diagram of the electronic device in the embodiment of the present invention. Detailed Embodiments

[0017] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0018] Referring to the following description and drawings, the embodiments of the present invention will be clear. In these descriptions and drawings, some specific embodiments in the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto. On the contrary, the embodiments of the present invention include all changes, modifications, and equivalents falling within the spirit and connotation of the appended claims.

[0019] Please refer to Figure 1 and Figure 2 , which is a partial-label feature selection method based on granule prototype disambiguation in the first embodiment of the present invention, including steps S11 to S16: Step S11, obtaining a real partial-label data set from different fields.

[0020] Specifically, given the partial-label data set , is a candidate tag set related to , representing the feature matrix, where and respectively represent the number of samples and the feature dimension, represents the tag matrix, where represents the number of tags.

[0021] Step S12: Construct initial coarse-grained granules using the information of the entire data set, and dynamically split the coarse-grained granules by a partitioning method, and calculate the feature values of each granule after partitioning and the tags of each granule.

[0022] Specifically, the steps of dynamically splitting the coarse-grained granules by a partitioning method and calculating the feature values of each granule after partitioning and the tags of each granule include: S121: Iteratively and dynamically split the coarse-grained granule by a partitioning method. After iterations until all granules meet the quality requirements, thus splitting the entire data set into disjoint granule sets ; S122: For each split granule, calculate the mean value of the feature values of each sample in each granule to obtain the feature values of each granule; for each tag , determine the tag of each granule according to the proportion of 1s of each sample in the granule on this tag.

[0023] Iteratively and dynamically split the coarse-grained granule by a partitioning method. The number of samples in the further split granules determines the feasibility of this partitioning. To ensure the feasibility of voting on the tags of the samples in the granules and improve the classification accuracy, it is required that the number of samples in each granule , so after iterations, the entire data set is split into disjoint granule sets .

[0024] Step S13: Based on the dynamically split granules, for each tag, calculate the weighted mean of the feature vectors of all granules under this tag to obtain the prototype vector of each tag, and calculate the similarity between the feature vector of each sample and the prototype vector of each tag using the cosine similarity.

[0025] Specifically, the weight of each granule is the reciprocal of the number of its candidate tags. The steps of calculating the weighted mean of the feature vectors of all granules under this tag to obtain the prototype vector of each tag include: ​For each label , the prototype vector of each label is obtained by calculating the weighted mean of the feature vectors of all granules under the label , and the calculation formula is as follows: , ; wherein, represents the number of granules under the label , represents the -th granule under the label , is the feature vector of the granule , is the weight of the granule, is the number of candidate labels of the granule .

[0026] The cosine similarity is used to measure the similarity between the sample and the prototype vector , and the calculation formula is as follows: ; wherein and are the and norms of, that is, their Euclidean lengths ( ). ).

[0027] Step S14, determine the weight voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weight voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, obtain the label space confidence matrix, and use it as the pseudo-label distribution matrix to complete label disambiguation.

[0028] For the sample , the weight of its corresponding label is determined by the similarity ratio between the sample and the prototype corresponding to the label, and the weight voting matrix is defined as follows: ; wherein, is the similarity between the sample and the prototype vector , represents the number of labels, n represents the number of samples, is the candidate label set related to , is the k-th label, Returns 1 if the expression in parentheses is true, and 0 otherwise.

[0029] Initialize the label space confidence matrix , specifically as follows: .

[0030] Based on the weight voting matrix Iteratively update the label space confidence , until the maximum number of iterations is reached. The iteration formula is: ; where is a balance parameter used to balance the importance of the original label space and the predicted label space.

[0031] Furthermore, to avoid unnecessary iterative calculations and reduce the accumulation of errors, a maximum margin convergence strategy is designed to allow the iterative process to be terminated early. The termination conditions are as follows: ; where represents the gap between the two highest confidences in the current iteration, is a preset threshold.

[0032] Step S15: Use linear projection to map the high-dimensional feature space to the label distribution space, introduce the global label distribution matrix, establish a disambiguation-feature selection bi-directional optimization model by minimizing the error loss with the pseudo-label distribution matrix, and at the same time construct a constraint term based on the similarity between samples and the contribution of irrelevant labels and embed it into the model.

[0033] Use linear projection to map the high-dimensional feature space to the label distribution space, explore the relationship between features and label distributions, introduce the global label distribution matrix, and establish a disambiguation-feature selection bi-directional optimization model by minimizing the error loss with the pseudo-label distribution matrix obtained by disambiguation as follows: ; where is the weight matrix of features to labels, is the global label distribution matrix, that is, the true label distribution matrix for global prediction, is the mapping matrix, is a balance parameter with a value between 0 and 1.

[0034] Adding regularization in this model can make the rows of sparse, better distinguish informative features and irrelevant features, and can also control the complexity of the model.

[0035] Explore the similarity between samples using the true information in the feature space and construct the constraint terms as follows: ; where is the sample similarity matrix of the graph Laplacian matrix, represents the sample and similarity between, which is calculated by the heat kernel, is a diagonal matrix, each element is , X i , X j represent the i-th row feature matrix and the j-th row feature matrix respectively.

[0036] Explore the contribution of irrelevant labels using the mislabeled information determined by the labeled space and construct the constraint terms as follows: ; where and represent the sample and irrelevant label vectors, is the irrelevant label matrix, which is an anti-logical matrix, represents the Hadamard product (element-wise multiplication), and L2 represents the graph Laplacian function.

[0037] Embed the constraint terms constructed by considering the similarity between samples and the contribution of irrelevant labels into the model to obtain the final objective function to be optimized as follows: .

[0038] Step S16, adopt an alternating iterative update strategy to perform end-to-end global optimization on the model, and output the feature ranking result to obtain the most discriminative feature subset.

[0039] Since the final objective function to be optimized is convex but exhibits non-smoothness and there are two parameter variables, an alternating iterative update strategy is adopted for optimization.

[0040] Fix , update , and the optimization problem is reformulated as follows:

[0041] Fix , update , and the optimization problem is reformulated as follows:

[0042] Then the accelerated proximal gradient descent method is used to solve these two optimization problems.

[0043] By calculating Evaluate the feature importance and output the feature subset sorted in descending order.

[0044] The following comparative experiments are used to illustrate the technical effects of the embodiments of the present invention: In the experiment, the partial-label feature selection method based on granule prototype disambiguation described in the embodiments of the present invention is denoted as ours. Ours is compared with five existing partial-label feature selection algorithms. Moreover, all partial-label feature selection algorithms are combined with 2 partial-label base classifiers (PL-KNN, IPAL) to evaluate the ability of the five feature selection algorithms in improving the generalization performance. The classification accuracies of the algorithms on 5 real datasets are shown in Table 1. In addition, the performance differences between the method described in the embodiments of the present invention and other comparative algorithms are analyzed from a statistical perspective, and the excellent / equal / inferior statistical results of the paired tests are shown in Table 2.

[0045] Table 1 Classification accuracies of the algorithms on real datasets (mean standard deviation)

[0046] Table 2 Excellent / equal / inferior statistics of the algorithms on real datasets (paired tests, significance level is 0.05)

[0047] As can be seen from the table, compared with other comparative algorithms, the method described in the embodiments of the present invention achieves significantly superior performance on 80% of the real datasets. In 60 cases (5 datasets 6 comparative algorithms 2 base classifiers), the method described in the embodiments of the present invention is the best in 57 cases, and the performance is 95% higher than that of other comparative algorithms. This proves the superiority and feasibility of the method described in the embodiments of the present invention in feature selection, and it helps to improve the generalization performance of the model.

[0048] Please refer to Figure 3 , for the partial-label feature selection device based on granule prototype disambiguation in the second embodiment of the present invention, including: An acquisition module 21, configured to acquire real partial-label datasets from different fields; A dynamic splitting granule sphere module 22, configured to construct an initial coarse-grained granule sphere by using the information of the entire dataset, and dynamically split the coarse-grained granule sphere by a partitioning method, and calculate the feature values of each granule sphere after partitioning and the labels of each granule sphere; The labeled prototype vector generation module 23 is used to calculate the weighted mean of the feature vectors of all the granules under each label based on the granules after dynamic splitting, so as to obtain the prototype vector of each label, and calculate the similarity between the feature vector of each sample and the prototype vector of each label by using the cosine similarity; The pseudo-label distribution matrix generation module 24 is used to determine the weight voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weight voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, so as to obtain the label space confidence matrix, and use it as the pseudo-label distribution matrix to complete label disambiguation; The disambiguation-feature selection two-way optimization model construction module 25 projects the high-dimensional feature space into the label distribution space by using linear projection, and introduces the global label distribution matrix, and establishes a disambiguation-feature selection two-way optimization model by minimizing the error loss with the pseudo-label distribution matrix. At the same time, a constraint term is constructed based on the similarity between samples and the contribution of irrelevant labels and embedded into the model; The feature subset output module 26 globally optimizes the model end-to-end by using an alternating iterative update strategy, and outputs the feature ranking result to obtain the most discriminative feature subset.

[0049] The partial-label feature selection device provided by the embodiment of the present invention has the same implementation principle and the same technical effects as the foregoing method embodiment. For the sake of brief description, for the parts not mentioned in the device embodiment, reference may be made to the corresponding content in the foregoing method embodiment.

[0050] On the other hand, the present invention also proposes an electronic device. Please refer to Figure 4 As shown in the figure, the electronic device in the embodiment of the present invention includes a processor 10, a memory 20, and a computer program 30 stored on the memory and executable on the processor. When the processor 10 executes the computer program 30, it implements the partial-label feature selection method based on granule prototype disambiguation as described above.

[0051] Among them, the electronic device may be, but is not limited to, computer devices such as personal computers and mobile phones. In some embodiments, the processor 10 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips, and is used to run the program code stored in the memory 20 or process data, etc.

[0052] Among them, the memory 20 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory, etc.), magnetic memory, magnetic disk, optical disc, etc. The memory 20 can be an internal storage unit of the electronic device in some embodiments, such as the hard disk of the electronic device. The memory 20 can also be an external storage device of the electronic device in other embodiments, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a FlashCard, etc. equipped on the electronic device. Further, the memory 20 can also include both the internal storage unit and the external storage device of the electronic device. The memory 20 can be used not only to store application software installed on the electronic device and various types of data, etc., but also to temporarily store the data that has been output or will be output.

[0053] Optionally, the electronic device may further include a user interface, a network interface, a communication bus, etc. The user interface may include a display, an input unit such as a keyboard. Optionally, the user interface may further include a standard wired interface and a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch liquid crystal display, and an OLED (Organic Light-Emitting Diode) toucher, etc. Among them, the display can also be appropriately referred to as a display screen or a display unit, which is used to display the information processed in the electronic device and to display the visual user interface. The network interface may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface), which is usually used to establish a communication connection between this device and other electronic devices. The communication bus is used to realize the connection and communication between these components.

[0054] It should be noted that Figure 4 The structure shown does not constitute a limitation on the electronic device. In other embodiments, the electronic device may include fewer or more components than those shown in the figure, or combine some components, or have different component arrangements.

[0055] The present invention also proposes a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the above-mentioned partial label feature selection method based on granule prototype disambiguation. Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus (such as a computer-based system, a system including a processor, or other systems that can obtain and execute instructions from the instruction execution system, apparatus), or in combination with these instruction execution systems, apparatuses. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in combination with an instruction execution system, apparatus.

[0056] More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection part (electronic device) having one or more wirings, a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other appropriate processing as necessary, and then stored in a computer memory.

[0057] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0058] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0059] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention shall be subject to the appended claims.

Claims

1. A partial-label feature selection method based on granulocyte prototype disambiguation, characterized in that Including: Obtain a real partially labeled dataset from different fields; Utilize the information of the entire dataset to construct initial coarse-grained granules, and dynamically split the coarse-grained granules by a partitioning method, calculate the values of the features of each granule after partitioning and the label of each granule; Based on the granules after dynamic splitting, for each label, calculate the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label, and calculate the similarity between the feature vector of each sample and the prototype vector of each label using cosine similarity; Determine the weight voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weight voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, obtain the label space confidence matrix, and use it as the pseudo-label distribution matrix to complete label disambiguation; Adopt linear projection to map the high-dimensional feature space to the label distribution space, introduce the global label distribution matrix, establish a disambiguation-feature selection two-way optimization model by minimizing the error loss with the pseudo-label distribution matrix, and at the same time construct a constraint term based on the similarity between samples and the contribution of irrelevant labels and embed it into the model; Adopt an alternating iterative update strategy to globally optimize the model end-to-end, and output the feature ranking result to obtain the most discriminative feature subset.

2. The partial label feature selection method based on granule prototype disambiguation according to claim 1, wherein The step of dynamically splitting the coarse-grained granules by a partitioning method, calculating the values of the features of each granule after partitioning and the label of each granule includes: Coarse-grained granules are iteratively and dynamically split by a partitioning method until, after iterations, all granules meet the quality requirements, thus splitting the entire dataset into disjoint granule sets ; For each split granulocyte, the value of each feature of each granulocyte is obtained by calculating the mean of the feature values of each sample within each granulocyte; for each label , the label of each granulocyte is determined according to the proportion of 1s on this label in each sample within the granulocyte.

3. The partial label feature selection method based on granulosphere prototype disambiguation according to claim 1, characterized in that The step of calculating the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label includes: For each label , the prototype vector of each label is obtained by calculating the weighted mean of the feature vectors of all the grains under the label , and the calculation formula is as follows: ， ； Among them, represents the number of granulocytes under the label, represents the label under which the th granulocyte, is the feature vector of the granulocyte is the weight of the granulocyte, is the granulocyte number of candidate labels.

4. The partial label feature selection method based on granule prototype disambiguation according to claim 1, wherein The step of iteratively updating the label space confidence based on the weight voting matrix includes: For a sample , the weight of its corresponding label is determined by the proportion of the similarity between the sample and the corresponding prototype of the label, and the weight voting matrix is defined as follows: ； Among them, is the sample and the prototype vector between the similarities represents the number of markers, n represents the number of samples, is related to the set of candidate markers is the k-th marker, represents that if the expression in the parentheses is true, it returns 1, otherwise it returns 0; Initialize the confidence matrix of the tag space , which is as follows: ； Based on the weight voting matrix Iteratively update the confidence of the labeled space , until the maximum number of iterations is reached. The iteration formula is: ； Among them, is a balance parameter used to balance the importance of the original token space and the predicted token space.

5. The partial label feature selection method based on granulosphere prototype disambiguation according to claim 4, characterized in that The disambiguation-feature selection two-way optimization model is: ； Among them, represents the feature matrix, is the weight matrix for feature pair marking, is the global marking distribution matrix, is the mapping matrix, is the balance parameter; The step of constructing a constraint term based on the similarity between samples and the contribution of irrelevant labels and embedding it into the model includes: Explore the contribution of irrelevant labels between samples using the mislabel information determined by the label space, and construct the constraint term as follows: ； Among them and respectively represent the irrelevant marker vectors of the samples and , is the irrelevant marker matrix, represents the Hadamard product, X i , X j respectively represent the feature matrix of the i-th row and the feature matrix of the j-th row; Embed the constraint term constructed based on the similarity between samples and the contribution of irrelevant labels into the model to obtain the final objective function to be optimized as follows: ； Among them, is the sample similarity matrix of the graph Laplacian matrix, represents the sample and the similarity between them, is a diagonal matrix, and each element is , and L2 represents the graph Laplacian function.

6. The partial label feature selection method based on granule prototype disambiguation according to claim 5, characterized in that, The step of adopting an alternating iterative update strategy to globally optimize the model end-to-end, and output the feature ranking result to obtain the most discriminative feature subset includes: Optimize using an alternating iterative update strategy, fixing , updating , the optimization problem is reformulated as follows: ； Fixed , updated , the optimization problem is reformulated as follows: ； Then use the accelerated proximal gradient descent method to solve this optimization problem; By calculation Evaluate the feature importance and output the feature subset sorted in descending order.

7. A partial-label feature selection device based on granulocyte prototype disambiguation, characterized in that, Including: An acquisition module for obtaining a real partially labeled dataset from different fields; A dynamic splitting granule module that utilizes the information of the entire dataset to construct initial coarse-grained granules, and dynamically splits the coarse-grained granules by a partitioning method, calculates the values of the features of each granule after partitioning and the label of each granule; A label prototype vector generation module for, based on the granules after dynamic splitting, for each label, calculating the weighted mean of the feature vectors of all granules under this label to obtain the prototype vector of each label, and calculating the similarity between the feature vector of each sample and the prototype vector of each label using cosine similarity; The pseudo-label distribution matrix generation module is used to determine the weighted voting matrix based on the similarity ratio, and iteratively update the label space confidence based on the weighted voting matrix until the maximum number of iterations is reached or the maximum margin convergence strategy is satisfied, obtaining the label space confidence matrix and using it as the pseudo-label distribution matrix to complete label disambiguation; The disambiguation-feature selection two-way optimization model construction module projects the high-dimensional feature space into the label distribution space by linear projection, introduces the global label distribution matrix, and establishes the disambiguation-feature selection two-way optimization model by minimizing the error loss with the pseudo-label distribution matrix. At the same time, a constraint term is constructed based on the similarity between samples and the contribution of irrelevant labels and embedded into the model; The feature subset output module performs end-to-end global optimization on the model using the alternating iterative update strategy and outputs the feature ranking result to obtain the most discriminative feature subset.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the partial-label feature selection method based on granule-sphere prototype disambiguation as described in any one of claims 1 to 6.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the partial-label feature selection method based on granule-sphere prototype disambiguation as described in any one of claims 1 to 6.

Images