A partial label feature selection method and system

Abstract

The present invention provides a method and system for selecting partial label features, which includes: obtaining a number of candidate features, and a number of adjacent labels and a number of label labels corresponding thereto, to obtain an initial confidence corresponding to the candidate features; obtaining an updated confidence of the candidate features, and then obtaining a final confidence; obtaining a candidate feature matrix corresponding to the candidate features, obtaining a candidate cluster center and a total cluster center, and then obtaining a closeness with the candidate features, selecting a number of stage features from a number of candidate features by closeness, obtaining a pessimistic dependency and an optimistic dependency of the stage features, and obtaining an importance evaluation value of the stage features based on the final confidence, pessimistic dependency and optimistic dependency, to obtain the optimal features. The above method improves the ability to identify identifiable features, overcomes the interference of false positive labels on unimodal dependency, and significantly improves the robustness of feature importance evaluation.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a partial label feature selection method and system. Background Art

[0002] Partially labeled learning (PLL), a key paradigm in weakly supervised learning, primarily aims to learn a multi-class classifier based on training examples, where each example is associated with a set of candidate labels. Unlike traditional multi-label learning, in partial labeling, the candidate label set contains both true and ambiguous labels, with only a single true label. This imprecise supervision significantly reduces the cost of data annotation, enabling its successful application in practical fields such as natural language processing, multimedia content analysis, and ecological informatics.

[0003] Existing partial-label feature selection methods are mostly based on preset static granularity partitioning strategies, which do not fully consider the inherent hierarchical structural characteristics of the data, which directly affects the recognition of identifiable features. In addition, existing methods usually use single-modal analysis to quantify the dependency measurement of feature discriminability. Since false positive labels will distort the reliability of single-modal dependency, it is impossible to achieve robust feature importance evaluation, thereby affecting the performance of the model. Summary of the Invention

[0004] In view of this, an object of the present invention is to provide a method and system for selecting partial label features.

[0005] In a first aspect, an embodiment of the present application provides a method for selecting partial label features, comprising the following steps:

[0006] Obtain a partially labeled dataset, the partially labeled dataset including several candidate features, obtain a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including several neighboring labels, the full label set including several labeled labels, and obtain an initial confidence score of the candidate feature based on the neighboring labels and the labeled labels;

[0007] Obtaining an updated confidence of the candidate feature, and obtaining a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence;

[0008] Constructing a feature space including several sub-dimensional spaces, mapping the candidate features to the feature space to obtain a candidate feature matrix corresponding to the candidate features, obtaining candidate cluster centers corresponding to the candidate feature matrix, and obtaining a total cluster center of all the candidate feature matrices;

[0009] Based on the candidate cluster center and the total cluster center, the closeness of each candidate feature is obtained, and several stage features are selected from the several candidate features according to the closeness, and the pessimistic dependency and optimistic dependency of the stage feature are obtained. Based on the final confidence, the pessimistic dependency and the optimistic dependency, the importance evaluation value of the stage feature is obtained, and the optimal feature is selected from the several stage features based on the importance evaluation value.

[0010] Furthermore, the formula for obtaining the initial confidence is:

[0011] ,

[0012] in, represents the initial confidence corresponding to the i-th candidate feature, represents the kth label in the full label set, represents the set of neighboring labels corresponding to the i-th candidate feature, represents the cardinality of the neighboring label set corresponding to the i-th candidate feature.

[0013] Furthermore, the step of obtaining the updated confidence of the candidate feature is specifically as follows:

[0014] Selecting one of the candidate features as a reference feature, and selecting the remaining candidate features as pending features;

[0015] Obtaining a Euclidean distance between the reference feature and the undetermined feature, selecting a plurality of discriminant units from the plurality of undetermined features based on the Euclidean distance, and obtaining a confidence kernel between the reference feature and the discriminant unit using the Euclidean distance;

[0016] A multivariate decision combination is performed on a plurality of the confidence kernels to obtain an updated confidence corresponding to the reference feature.

[0017] Furthermore, the step of selecting a plurality of discrimination units from a plurality of the undetermined features based on the Euclidean distance is specifically as follows:

[0018] Comparing the Euclidean distance with a distance threshold, and selecting the undetermined feature corresponding to the Euclidean distance smaller than the distance threshold as a discrimination unit;

[0019] The formula for obtaining the confidence kernel is:

[0020] ,

[0021] in, Indicates the confidence kernel between the i-th candidate feature as the benchmark feature and the j-th candidate feature as the discriminant unit. represents the weight parameter, represents the distance attenuation coefficient, represents the Euclidean distance between the i-th candidate feature and the j-th candidate feature, Represents the exponential function with base e.

[0022] Furthermore, the step of obtaining candidate cluster centers corresponding to the candidate feature matrix includes:

[0023] Selecting one of the candidate feature matrices as a reference feature matrix, and selecting the remaining candidate feature matrices as pending feature matrices;

[0024] Obtaining a Mahalanobis distance between the reference feature matrix and the undetermined feature matrix, selecting a plurality of adjacent feature matrices from a plurality of the undetermined feature matrices based on the Mahalanobis distance, and combining the reference feature matrix and the adjacent feature matrix into a neighborhood topology unit;

[0025] The neighborhood topology units are clustered based on the DBSCAN algorithm to obtain candidate cluster centers corresponding to the benchmark feature matrix.

[0026] Furthermore, the formula for obtaining the tightness is:

[0027] ,

[0028] in, represents the closeness corresponding to the i-th candidate feature, represents the candidate cluster center corresponding to the i-th candidate feature, represents the total cluster center, represents the neighborhood correlation coefficient.

[0029] Furthermore, the step of selecting a plurality of stage features from a plurality of candidate features according to the closeness includes:

[0030] Selecting one of the candidate features as a first intermediate feature, and selecting the remaining candidate features as second intermediate features, and obtaining closeness differences between the first intermediate feature and all the second intermediate features respectively;

[0031] comparing the closeness difference with a difference threshold, selecting a second intermediate feature corresponding to the closeness difference less than the difference threshold as a feature to be merged, and merging the first intermediate feature and the feature to be merged into a dynamic granularity level;

[0032] The compactness mean of the dynamic granularity layer is obtained, the dynamic granularity layer with the largest compactness mean is selected as the stage granularity layer, and several candidate features in the stage granularity layer are all selected as stage features.

[0033] Furthermore, the step of obtaining the pessimistic dependency and the optimistic dependency of the stage characteristics includes:

[0034] Selecting one of the stage features as a feature to be identified, selecting the remaining stage features as reference features, selecting the neighboring tags corresponding to the feature to be identified as a tag set to be identified, and selecting the neighboring tags corresponding to the reference feature as a reference tag set;

[0035] Comparing the tag set to be identified with all the reference tag sets, if a neighboring tag in the tag set to be identified exists in all the reference tag sets, then selecting the neighboring tag as the positive region; if a neighboring tag in the tag set to be identified exists in some of the reference tag sets, then selecting the neighboring tag as the boundary region;

[0036] A pessimistic dependency of the feature to be identified is obtained based on the determined positive region, and an optimistic dependency of the feature to be identified is obtained based on the determined positive region and the boundary region.

[0037] Furthermore, the step of obtaining the importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency includes:

[0038] Obtaining a dynamic fusion coefficient through the final confidence level;

[0039] The pessimistic dependency and the optimistic dependency are weighted as an importance evaluation value based on the dynamic fusion coefficient.

[0040] In a second aspect, an embodiment of the present application provides a partial label feature selection system, which is applied to the partial label feature selection method described in the first aspect above, and the system includes:

[0041] An acquisition module is configured to obtain a partially labeled dataset, the partially labeled dataset including a plurality of candidate features, obtain a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including a plurality of neighboring labels, the full label set including a plurality of labeled labels, and obtain an initial confidence score for the candidate feature based on the neighboring labels and the labeled labels;

[0042] An updating module, configured to obtain an updated confidence of the candidate feature, and obtain a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence;

[0043] a partitioning module, configured to construct a feature space comprising a plurality of sub-dimensional spaces, map the candidate features to the feature space, obtain a candidate feature matrix corresponding to the candidate features, obtain candidate cluster centers corresponding to the candidate feature matrix, and obtain a total cluster center of all the candidate feature matrices;

[0044] An execution module is used to obtain the closeness of each candidate feature based on the candidate cluster center and the total cluster center, select a number of stage features from the candidate features according to the closeness, obtain the pessimistic dependency and optimistic dependency of the stage feature, obtain the importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency and the optimistic dependency, and select the optimal feature from the number of stage features based on the importance evaluation value.

[0045] In a third aspect, an embodiment of the present application provides a computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the biased label feature selection method as described in the first aspect above is implemented.

[0046] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the partial label feature selection method described in the first aspect above is implemented.

[0047] Compared with the prior art, the beneficial effects of the present invention are: by constructing a tightness-oriented dynamic granularity hierarchical structure, the feature evaluation granularity can be adaptively adjusted as the data stream distribution changes, thereby improving the ability to identify identifiable features; a pessimistic-optimistic bimodal dependency metric is proposed to quantify feature discrimination, effectively overcoming the interference of false positive labels on unimodal dependency, and significantly improving the robustness of feature importance evaluation; a multi-granularity reconstruction mechanism based on neighborhood topological units fully considers supervisory information in the granulation process, enabling the model to better identify noise and redundant features, thereby improving the accuracy and overall performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the partial label feature selection method according to the first embodiment of the present invention;

[0049] Figure 2 This is a structural block diagram of a partial label feature selection system in a second embodiment of the present invention;

[0050] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0051] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0052] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0054] See also Figure 1 The first embodiment of the present invention provides a partial label feature selection method, comprising the following steps:

[0055] S10: Obtain a partially labeled dataset, the partially labeled dataset including several candidate features, obtain a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including several neighboring labels, the full label set including several labeled labels, and obtain an initial confidence of the candidate feature based on the neighboring labels and the labeled labels;

[0056] In some embodiments, the partially labeled dataset corresponds to a painting, and the candidate features in the partially labeled dataset are feature vectors representing the names of the artists. This essentially involves selecting the actual artist name corresponding to the painting from among the multiple author names, and the neighboring labels are label vectors representing composition, size, color, shape, line, etc. It should be noted that the neighboring label set is included in the full label set, which may contain labels unrelated to the painting.

[0057] The formula for obtaining the initial confidence is:

[0058] ,

[0059] in, represents the initial confidence corresponding to the i-th candidate feature, represents the kth label in the full label set, represents the set of neighboring labels corresponding to the i-th candidate feature, represents the cardinality of the neighboring label set corresponding to the i-th candidate feature.

[0060] S20: Obtaining an updated confidence of the candidate feature, and obtaining a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence;

[0061] The step S20 includes:

[0062] S210: selecting one of the candidate features as a reference feature, and selecting the remaining candidate features as pending features;

[0063] S220: Obtaining a Euclidean distance between the reference feature and the feature to be determined, selecting a plurality of discriminant units from the plurality of features to be determined based on the Euclidean distance, and obtaining a confidence kernel between the reference feature and the discriminant unit using the Euclidean distance;

[0064] The Euclidean distance is obtained using an existing algorithm and will not be further described here. After obtaining the Euclidean distance, the Euclidean distance is compared with a distance threshold, and the undetermined features corresponding to the Euclidean distance less than the distance threshold are selected as the discrimination units. Several of the discrimination units are combined into a discrimination region.

[0065] The formula for obtaining the confidence kernel is:

[0066] ,

[0067] in, Indicates the confidence kernel between the i-th candidate feature as the benchmark feature and the j-th candidate feature as the discriminant unit. represents the weight parameter, represents the distance attenuation coefficient, represents the Euclidean distance between the i-th candidate feature and the j-th candidate feature, Represents the exponential function with base e.

[0068] S230: performing a multivariate decision combination on a plurality of the confidence kernels to obtain an updated confidence corresponding to the reference feature;

[0069] The calculation formula for the updated confidence is:

[0070] ,

[0071] in, represents the updated confidence corresponding to the i-th candidate feature, represents a multivariate decision operation, represents the jth candidate feature, Indicates the discrimination area.

[0072] It is understandable that the reference feature is also a candidate feature, expressed only differently. Therefore, the updated confidence corresponding to the reference feature is the updated confidence corresponding to the candidate feature. The above operation is repeated for all candidate features to obtain the updated confidence for all candidate features. After obtaining the updated confidence, the initial confidence and the updated confidence are weighted together to obtain the final confidence by assigning weighting coefficients to each.

[0073] S30: constructing a feature space including several sub-dimensional spaces, mapping the candidate features to the feature space to obtain a candidate feature matrix corresponding to the candidate features, obtaining candidate cluster centers corresponding to the candidate feature matrix, and obtaining a total cluster center of all the candidate feature matrices;

[0074] It can be understood that the candidate feature matrix is ​​a d-dimensional matrix, whose dimension corresponds to the number of the sub-dimensional spaces.

[0075] The step S30 includes:

[0076] S310: selecting one of the candidate feature matrices as a reference feature matrix, and selecting the remaining candidate feature matrices as pending feature matrices;

[0077] S320: Obtaining a Mahalanobis distance between the reference feature matrix and the undetermined feature matrix, selecting a plurality of adjacent feature matrices from a plurality of the undetermined feature matrices based on the Mahalanobis distance, and combining the reference feature matrix and the adjacent feature matrix into a neighborhood topology unit;

[0078] The Mahalanobis distance is obtained by an existing algorithm and will not be described in detail here. After obtaining the Mahalanobis distance, the Mahalanobis distance is compared with a proximity threshold, and the undetermined feature matrix corresponding to the Mahalanobis distance that is smaller than the proximity threshold is selected as the proximity feature matrix.

[0079] S330: performing clustering processing on the neighborhood topology units based on the DBSCAN algorithm to obtain candidate cluster centers corresponding to the benchmark feature matrix;

[0080] The DBSCAN algorithm is a conventional algorithm for clustering processing and will not be described in detail here. It can be understood that the reference feature matrix is ​​also the candidate feature matrix, only the expression method is different. Therefore, the candidate cluster center corresponding to the reference feature matrix is ​​the candidate cluster center corresponding to the candidate feature matrix. The above operation is repeated for all the candidate feature matrices to obtain the candidate cluster centers of all the candidate feature matrices. The total cluster center can also be obtained using the DBSCAN algorithm, the only difference being the number of samples processed.

[0081] S40: obtaining the closeness of each candidate feature based on the candidate cluster center and the total cluster center, selecting a plurality of stage features from the plurality of candidate features according to the closeness, obtaining the pessimistic dependency and the optimistic dependency of the stage feature, obtaining an importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency, and selecting the optimal feature from the plurality of stage features based on the importance evaluation value;

[0082] The formula for obtaining the tightness is:

[0083] ,

[0084] in, represents the closeness corresponding to the i-th candidate feature, represents the candidate cluster center corresponding to the i-th candidate feature, represents the total cluster center, represents the neighborhood correlation coefficient.

[0085] The step S40 includes:

[0086] S410: selecting one of the candidate features as a first intermediate feature, and selecting the remaining candidate features as second intermediate features, and obtaining closeness differences between the first intermediate feature and all the second intermediate features respectively;

[0087] S420: Compare the closeness difference with a difference threshold, select a second intermediate feature corresponding to the closeness difference smaller than the difference threshold as a feature to be merged, and merge the first intermediate feature and the feature to be merged into a dynamic granularity level;

[0088] The dynamic granularity level includes the first intermediate feature and several features to be merged, which are distinguished only in the name, but actually include several candidate features. All the candidate features are divided into several dynamic granularity levels.

[0089] S430: Obtain the average density of the dynamic granularity layer, select the dynamic granularity layer with the largest average density as the stage granularity layer, and select several candidate features in the stage granularity layer as stage features;

[0090] S440: selecting one of the stage features as a feature to be identified, selecting the remaining stage features as reference features, selecting neighboring tags corresponding to the feature to be identified as a tag set to be identified, and selecting neighboring tags corresponding to the reference feature as a reference tag set;

[0091] S450: Compare the tag set to be identified with all the reference tag sets. If a neighboring tag in the tag set to be identified exists in all the reference tag sets, select the neighboring tag as a positive region. If a neighboring tag in the tag set to be identified exists in some of the reference tag sets, select the neighboring tag as a boundary region.

[0092] Suppose there is one label set to be identified and two reference label sets, the label set to be identified is {composition, size, color, texture}, one of the reference label sets is {composition, color, shape}, and the other reference label set is {composition, size, line}. If the composition in the label set to be identified exists in all the reference labels, then the composition is the positive area, and if the size and color exist in the two reference label sets respectively, then the size and color are the boundary areas.

[0093] S460: Obtaining a pessimistic dependency of the feature to be identified based on the determined positive region, and obtaining an optimistic dependency of the feature to be identified based on the determined positive region and the boundary region;

[0094] It can be understood that the feature to be identified is the candidate feature, and the pessimistic dependency and optimistic dependency of the feature to be identified are the pessimistic dependency and optimistic dependency of the candidate feature. The above operation is repeated for all the candidate features to obtain the pessimistic dependency and optimistic dependency of all the candidate features.

[0095] The formula for obtaining the pessimistic dependency is:

[0096] ,

[0097] in, represents the pessimistic dependency corresponding to the i-th candidate feature, represents the positive region of the i-th candidate feature, represents the set of adjacent labels in the label set to be identified and the reference label set, Represents the cardinality of the set of adjacent tags in the tag set to be identified and the reference tag set.

[0098] The formula for obtaining the optimistic dependency is:

[0099] ,

[0100] in, represents the optimistic dependency corresponding to the i-th candidate feature, represents the boundary region of the i-th candidate feature.

[0101] S470: Obtaining a dynamic fusion coefficient using the final confidence level;

[0102] The formula for obtaining the dynamic fusion coefficient is:

[0103] ,

[0104] in, represents the dynamic fusion coefficient corresponding to the i-th candidate feature, Indicates the operation parameters, represents the final latest degree corresponding to the i-th candidate feature.

[0105] S480: Weighting the pessimistic dependency and the optimistic dependency into an importance evaluation value based on the dynamic fusion coefficient;

[0106] The formula for obtaining the importance evaluation value is:

[0107] ,

[0108] in, represents the importance evaluation value corresponding to the i-th candidate feature. After obtaining the importance evaluation value, the stage feature with the highest importance evaluation value is selected from the plurality of stage features as the optimal feature.

[0109] By constructing a tightness-oriented dynamic granularity hierarchy structure, the feature evaluation granularity can be adaptively adjusted as the data stream distribution changes, thereby improving the ability to identify identifiable features; a pessimistic-optimistic bimodal dependency metric is proposed to quantify feature discrimination, effectively overcoming the interference of false positive labels on unimodal dependency and significantly improving the robustness of feature importance assessment; a multi-granularity reconstruction mechanism based on neighborhood topological units fully considers supervisory information in the granulation process, enabling the model to better identify noise and redundant features, thereby improving the accuracy and overall performance of the model.

[0110] The partial label feature selection method described in this example is denoted as ours and compared with five existing feature selection algorithms. All algorithms are combined with two partial label base classifiers (PL-KNN and IPAL) to evaluate the ability of different selection algorithms to improve generalization performance. The classification accuracy of the algorithms on five real datasets is shown in Table 1:

[0111] Table 1

[0112]

[0113] , the performance differences between the method described in the embodiment of the present invention and other comparative algorithms were analyzed from a statistical perspective. The superiority / parity / inferiority statistical results of the paired t-test at a significance level of 0.05 are shown in Table 2:

[0114] Table 2

[0115] .

[0116] As can be seen from the table, compared with other comparison algorithms, the method described in the embodiment of the present invention achieved significantly superior performance on 80% of the real data sets. Among 60 cases, the method described in the embodiment of the present invention was the best in 58 cases, and the performance was 97% higher than other comparison algorithms. This proves the superiority and feasibility of the method described in the embodiment of the present invention in feature selection, and it helps to improve the generalization performance of the model.

[0117] See also Figure 2 The second embodiment of the present invention provides a partial label feature selection system, which is applied to the partial label feature selection method described in the above embodiment. The details that have been explained will not be repeated here. As used below, the terms "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and conceivable.

[0118] The system comprises:

[0119] An acquisition module 10 is configured to acquire a partially labeled dataset, the partially labeled dataset including a plurality of candidate features, acquire a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including a plurality of neighboring labels, the full label set including a plurality of labeled labels, and acquire an initial confidence score for the candidate feature based on the neighboring labels and the labeled labels;

[0120] An updating module 20 is configured to obtain an updated confidence of the candidate feature, and obtain a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence;

[0121] The update module 20 includes:

[0122] The first unit is configured to select one of the candidate features as a reference feature and select the remaining candidate features as pending features;

[0123] A second unit is configured to obtain a Euclidean distance between the reference feature and the feature to be determined, select a plurality of discriminant units from the plurality of features to be determined based on the Euclidean distance, and obtain a confidence kernel between the reference feature and the discriminant unit using the Euclidean distance;

[0124] The second unit is further configured to compare the Euclidean distance with a distance threshold, and select a pending feature corresponding to the Euclidean distance that is smaller than the distance threshold as a discrimination unit;

[0125] A third unit is configured to perform a multivariate decision combination on a plurality of the confidence kernels to obtain an updated confidence corresponding to the reference feature;

[0126] A partitioning module 30 is configured to construct a feature space comprising a plurality of sub-dimensional spaces, map the candidate features to the feature space, obtain a candidate feature matrix corresponding to the candidate features, obtain candidate cluster centers corresponding to the candidate feature matrix, and obtain a total cluster center of all the candidate feature matrices;

[0127] The division module 30 includes:

[0128] A fourth unit is configured to select one of the candidate feature matrices as a reference feature matrix, and select the remaining candidate feature matrices as pending feature matrices;

[0129] a fifth unit, configured to obtain a Mahalanobis distance between the reference feature matrix and the undetermined feature matrix, select a plurality of adjacent feature matrices from a plurality of the undetermined feature matrices based on the Mahalanobis distance, and combine the reference feature matrix and the adjacent feature matrix into a neighborhood topology unit;

[0130] The sixth unit is configured to perform clustering processing on the neighborhood topology units based on the DBSCAN algorithm to obtain candidate cluster centers corresponding to the benchmark feature matrix;

[0131] An execution module 40 is configured to obtain a closeness of each candidate feature based on the candidate cluster center and the total cluster center, select a plurality of stage features from the plurality of candidate features based on the closeness, obtain a pessimistic dependency and an optimistic dependency of the stage feature, obtain an importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency, and select an optimal feature from the plurality of stage features based on the importance evaluation value;

[0132] The execution module 40 includes:

[0133] a seventh unit, configured to select one of the candidate features as a first intermediate feature, select the remaining candidate features as second intermediate features, and obtain closeness differences between the first intermediate feature and all the second intermediate features;

[0134] an eighth unit, configured to compare the closeness difference with a difference threshold, select a second intermediate feature corresponding to the closeness difference less than the difference threshold as a feature to be merged, and merge the first intermediate feature and the feature to be merged into a dynamic granularity level;

[0135] a ninth unit configured to obtain a compactness mean value of the dynamic granularity layer, select the dynamic granularity layer having the largest compactness mean value as the stage granularity layer, and select a plurality of candidate features in the stage granularity layer as stage features;

[0136] a tenth unit, configured to select one of the stage features as a feature to be identified, select the remaining stage features as reference features, select adjacent tags corresponding to the feature to be identified as a tag set to be identified, and select adjacent tags corresponding to the reference feature as a reference tag set;

[0137] an eleventh unit, configured to compare the tag set to be identified with all the reference tag sets, and if a neighboring tag in the tag set to be identified exists in all the reference tag sets, select the neighboring tag as a positive region; if a neighboring tag in the tag set to be identified exists in some of the reference tag sets, select the neighboring tag as a boundary region;

[0138] A twelfth unit is configured to obtain a pessimistic dependency of the feature to be identified based on the determined positive region, and to obtain an optimistic dependency of the feature to be identified based on the determined positive region and the boundary region;

[0139] A thirteenth unit is used to obtain a dynamic fusion coefficient according to the final confidence level;

[0140] The fourteenth unit is configured to weight the pessimistic dependency and the optimistic dependency into an importance evaluation value based on the dynamic fusion coefficient.

[0141] The present invention also provides a computer, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the biased label feature selection method described in the above technical solution is implemented.

[0142] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the partial label feature selection method described in the above technical solution is implemented.

[0143] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0144] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A partial label feature selection method for screening the author of a painting, characterized by: The following steps are involved: Obtain a partially labeled dataset, the partially labeled dataset corresponding to the painting, the partially labeled dataset including several candidate features, the candidate features being feature vectors representing the author's name, obtaining a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including several neighboring labels, the neighboring labels being label vectors representing composition, size, color, shape, and line, the full label set including several marked labels, the neighboring label set being included in the full label set, and obtaining an initial confidence score for the candidate feature based on the neighboring labels and the marked labels; Obtaining an updated confidence of the candidate feature, and obtaining a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence; Constructing a feature space including several sub-dimensional spaces, mapping the candidate features to the feature space to obtain a candidate feature matrix corresponding to the candidate features, obtaining candidate cluster centers corresponding to the candidate feature matrix, and obtaining a total cluster center of all the candidate feature matrices; Obtaining the closeness of each candidate feature based on the candidate cluster center and the total cluster center, selecting a plurality of stage features from the plurality of candidate features based on the closeness, obtaining a pessimistic dependency and an optimistic dependency of the stage feature, obtaining an importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency, and selecting an optimal feature from the plurality of stage features based on the importance evaluation value; The step of selecting a plurality of stage features from a plurality of candidate features according to the closeness comprises: Selecting one of the candidate features as a first intermediate feature, and selecting the remaining candidate features as second intermediate features, and obtaining closeness differences between the first intermediate feature and all the second intermediate features respectively; comparing the closeness difference with a difference threshold, selecting a second intermediate feature corresponding to the closeness difference less than the difference threshold as a feature to be merged, and merging the first intermediate feature and the feature to be merged into a dynamic granularity level; Obtaining a compactness mean value of the dynamic granularity layer, selecting the dynamic granularity layer with the largest compactness mean value as the stage granularity layer, and selecting a plurality of candidate features in the stage granularity layer as stage features; The step of obtaining the pessimistic dependency and the optimistic dependency of the stage characteristics includes: Selecting one of the stage features as a feature to be identified, selecting the remaining stage features as reference features, selecting the neighboring tags corresponding to the feature to be identified as a tag set to be identified, and selecting the neighboring tags corresponding to the reference feature as a reference tag set; Comparing the tag set to be identified with all the reference tag sets, if a neighboring tag in the tag set to be identified exists in all the reference tag sets, then selecting the neighboring tag as the positive region; if a neighboring tag in the tag set to be identified exists in some of the reference tag sets, then selecting the neighboring tag as the boundary region; A pessimistic dependency of the feature to be identified is obtained based on the determined positive region, and an optimistic dependency of the feature to be identified is obtained based on the determined positive region and the boundary region.

2. The partial label feature selection method according to claim 1, characterized in that: The formula for obtaining the initial confidence is: ， in, represents the initial confidence corresponding to the i-th candidate feature, represents the kth label in the full label set, represents the set of neighboring labels corresponding to the i-th candidate feature, represents the cardinality of the neighboring label set corresponding to the i-th candidate feature.

3. The partial label feature selection method according to claim 1, characterized in that: The step of obtaining the updated confidence of the candidate feature is specifically as follows: Selecting one of the candidate features as a reference feature, and selecting the remaining candidate features as pending features; Obtaining a Euclidean distance between the reference feature and the undetermined feature, selecting a plurality of discriminant units from the plurality of undetermined features based on the Euclidean distance, and obtaining a confidence kernel between the reference feature and the discriminant unit using the Euclidean distance; A multivariate decision combination is performed on a plurality of the confidence kernels to obtain an updated confidence corresponding to the reference feature.

4. The partial label feature selection method according to claim 3, characterized in that: The step of selecting a plurality of discrimination units from a plurality of the undetermined features based on the Euclidean distance is specifically as follows: Comparing the Euclidean distance with a distance threshold, and selecting the undetermined feature corresponding to the Euclidean distance smaller than the distance threshold as a discrimination unit; The formula for obtaining the confidence kernel is: ， in, Indicates the confidence kernel between the i-th candidate feature as the benchmark feature and the j-th candidate feature as the discriminant unit. represents the weight parameter, represents the distance attenuation coefficient, represents the Euclidean distance between the i-th candidate feature and the j-th candidate feature, Represents the exponential function with base e.

5. The partial label feature selection method according to claim 1, characterized in that: The step of obtaining candidate cluster centers corresponding to the candidate feature matrix includes: Selecting one of the candidate feature matrices as a reference feature matrix, and selecting the remaining candidate feature matrices as pending feature matrices; Obtaining a Mahalanobis distance between the reference feature matrix and the undetermined feature matrix, selecting a plurality of adjacent feature matrices from a plurality of the undetermined feature matrices based on the Mahalanobis distance, and combining the reference feature matrix and the adjacent feature matrix into a neighborhood topology unit; The neighborhood topology units are clustered based on the DBSCAN algorithm to obtain candidate cluster centers corresponding to the benchmark feature matrix.

6. The partial label feature selection method according to claim 1, characterized in that: The formula for obtaining the tightness is: ， in, represents the closeness corresponding to the i-th candidate feature, represents the candidate cluster center corresponding to the i-th candidate feature, represents the total cluster center, represents the neighborhood correlation coefficient.

7. The partial label feature selection method according to claim 1, characterized in that: The step of obtaining the importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency includes: Obtaining a dynamic fusion coefficient through the final confidence level; The pessimistic dependency and the optimistic dependency are weighted as an importance evaluation value based on the dynamic fusion coefficient.

8. A partial label feature selection system, applied to the partial label feature selection method according to any one of claims 1 to 7, characterized in that: The system comprises: An acquisition module is configured to obtain a partially labeled dataset, the partially labeled dataset corresponding to the painting, the partially labeled dataset including a plurality of candidate features, the candidate features being feature vectors representing the author's name, obtain a neighboring label set and a full label set corresponding to the candidate features, the neighboring label set including a plurality of neighboring labels, the neighboring labels being label vectors representing composition, size, color, shape, and line, the full label set including a plurality of marked labels, the neighboring label set being included in the full label set, and obtain an initial confidence score for the candidate feature based on the neighboring labels and the marked labels; An updating module, configured to obtain an updated confidence of the candidate feature, and obtain a final confidence corresponding to the candidate feature based on the initial confidence and the updated confidence; a partitioning module, configured to construct a feature space comprising a plurality of sub-dimensional spaces, map the candidate features to the feature space, obtain a candidate feature matrix corresponding to the candidate features, obtain candidate cluster centers corresponding to the candidate feature matrix, and obtain a total cluster center of all the candidate feature matrices; an execution module, configured to obtain a closeness of each candidate feature based on the candidate cluster center and the total cluster center, select a plurality of stage features from the plurality of candidate features based on the closeness, obtain a pessimistic dependency and an optimistic dependency of the stage feature, obtain an importance evaluation value of the stage feature based on the final confidence, the pessimistic dependency, and the optimistic dependency, and select an optimal feature from the plurality of stage features based on the importance evaluation value; The execution module includes: a seventh unit, configured to select one of the candidate features as a first intermediate feature, select the remaining candidate features as second intermediate features, and obtain closeness differences between the first intermediate feature and all the second intermediate features; an eighth unit, configured to compare the closeness difference with a difference threshold, select a second intermediate feature corresponding to the closeness difference less than the difference threshold as a feature to be merged, and merge the first intermediate feature and the feature to be merged into a dynamic granularity level; a ninth unit configured to obtain a compactness mean value of the dynamic granularity layer, select the dynamic granularity layer having the largest compactness mean value as the stage granularity layer, and select a plurality of candidate features in the stage granularity layer as stage features; a tenth unit, configured to select one of the stage features as a feature to be identified, select the remaining stage features as reference features, select adjacent tags corresponding to the feature to be identified as a tag set to be identified, and select adjacent tags corresponding to the reference feature as a reference tag set; an eleventh unit, configured to compare the tag set to be identified with all the reference tag sets, and if a neighboring tag in the tag set to be identified exists in all the reference tag sets, select the neighboring tag as a positive region; if a neighboring tag in the tag set to be identified exists in some of the reference tag sets, select the neighboring tag as a boundary region; The twelfth unit is configured to obtain a pessimistic dependency of the feature to be identified based on the determined positive region, and to obtain an optimistic dependency of the feature to be identified based on the determined positive region and the boundary region.

Images