Partially multi-label classification method, device, equipment and medium based on weighted manifold disambiguation and dynamic local label correlation

Abstract

A partial multi-label classification method, device, equipment and medium based on weighted manifold disambiguation and dynamic local label correlation relates to the field of partial multi-label learning technology. The partial multi-label classification method can process complex data sets efficiently and accurately. First, the candidate labels are converted into a logical label matrix, the relationship between the feature space and the label distribution is constructed, and the loss function is designed accordingly. Secondly, by calculating the semantic similarity, the k-nearest neighbor method is used to determine the neighbors of the training instance to form a similarity graph weight matrix. Next, the initial local label correlation matrix is ​​calculated and dynamically updated, combined with the similarity graph weight matrix to guide label distribution learning. Then, the objective function is constructed, and the model parameters are iteratively optimized until convergence to achieve label disambiguation. Finally, the trained model can be used to predict all categories to which the instance to be tested belongs, with high accuracy and robustness, effectively coping with the challenges of noisy labels, and providing a high-performance solution to the partial multi-label classification problem.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-label learning classification, and in particular to a partial multi-label classification method, device, equipment and medium based on weighted manifold disambiguation and dynamic local label correlation. Background Art

[0002] Partial Multi-Label Learning (PML) is an important branch of the multi-label learning field. In the PML problem, each training instance is associated with a set of candidate labels, but only some of them are correct, and the rest are noise labels. This learning scenario is very common in the real world. For example, in fields such as text classification, bioinformatics, and multimedia content analysis, you may encounter situations where only some labels are correct.

[0003] Traditional multi-label learning methods usually assume that all provided labels are accurate, but in the PML problem, this assumption often does not hold. Therefore, the goal of PML is to identify the correct subset of labels from these noisy candidate labels. Existing PML learning methods mainly rely on the smoothness assumption that associated labels tend to appear on similar instances at the same time, thereby exploiting the topological structure and label correlation in the feature space of training instances to recover the correct label distribution.

[0004] However, existing methods have limitations in dealing with negative information between features and labels, and the label correlations constructed based on candidate label sets are easily affected by noisy labels, which can significantly reduce the prediction performance of the model. In addition, most existing works ignore the complex relationship between features and labels in the feature space, and how to effectively learn accurate label distributions from noisy labels. Therefore, researchers have been seeking more effective PML learning methods to improve classification performance in noisy environments and accurately identify the correct label set from the candidate label set. Summary of the invention

[0005] The present invention provides a method, device, equipment and medium for partial multi-label classification based on weighted manifold disambiguation and dynamic local label correlation to improve at least one of the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation, which includes S1 to S5.

[0007] S1. Obtain a multi-labeled data set to be processed, convert candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix. And construct a loss function according to the feature space matrix and the label distribution matrix.

[0008] S2, calculate the semantic similarity of the logical label matrix, and then -The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples.

[0009] S3, calculate the initial local label correlation matrix according to the logical label matrix. And build a local label correlation dynamic learning mechanism according to the initial local label correlation matrix and the label distribution matrix to obtain a new local label correlation matrix. Then, build a manifold regularization term induced label distribution matrix based on the similarity graph weight matrix of the training instance and the new local label correlation matrix. study.

[0010] S4. Construct an objective function of the partial multi-label classifier according to the loss function and the manifold regularization term, and perform iterative update processing until the objective function value converges, or the iteration reaches a preset maximum number of iterations, completes label disambiguation to obtain the optimal model parameters, and obtains the trained partial multi-label classifier.

[0011] S5. Obtain a test instance and input the trained partial multi-label classifier to predict all possible categories of the test instance.

[0012] In an optional implementation, step S1 specifically includes steps S11 to S15.

[0013] S11. Obtain the multi-labeled dataset to be processed, denoted as ,in Indicates training examples, express The candidate tag set, is the total sample size.

[0014] S12. For each over-labeled sample , constructed using the hard thresholding method Logical label vector of :For any ,if ,but ,otherwise ;in, represents the number of categories, For the Category tags, express No. The logical value of the candidate marker.

[0015] S13. Construct the logical label matrix of all training instances and the feature space matrix .in, represents the transpose of a matrix, represents the set of real numbers, is the total sample size, Represents the dimension of the feature space.

[0016] S14. Obtaining a marker distribution matrix according to the feature space matrix : , where express The distribution of markers, represents the feature mapping function, is the weight matrix of the model, Represents the offset vector.

[0017] S15. According to the feature space matrix and the label distribution matrix, a least squares regression model is used to construct a loss function. The loss function formula is:

[0018]

[0019] in, represents the feature matrix after mapping, express dimensional all-1 vector, For vector Norm, is the Frobenius norm of the matrix.

[0020] In an optional implementation, step S2 specifically includes steps S21 to S25.

[0021] S21. Constructing semantic similarity matrix using cosine similarity . , .in, Indicates Examples and Examples The corresponding similarity, is the total sample size, and Respectively represent Examples and Examples The corresponding logical label vector, represents the transpose of a matrix, For vector Norm.

[0022] S22. Calculate the distance matrix between training instances , .in, Indicates Examples and Examples The distance between.

[0023] S23. Construct a weighted distance matrix of training instances , .in, Indicates Examples and Examples The weighted distance between .

[0024] S24. For each training instance , select the closest one by weighted distance training instances, Neighborhood .

[0025] S25, calculating the similarity graph weight matrix of the training instance according to the neighborhood . It is constructed by solving the following linear least squares problem:

[0026]

[0027] in, Represents a training instance for elements in the neighborhood of . For the Examples and Examples The similarity weight between them.

[0028] In an optional implementation, step S3 specifically includes steps S31 to S33.

[0029] S31, using the logical labeling matrix to calculate the initial local labeling correlation matrix , ,in, Indicates Class and The initial local label correlation of each class, and Respectively represent Class and A logical vector of classes, Indicates the kernel size.

[0030] S32, based on the initial local label correlation matrix and the label distribution matrix Perform dynamic learning to obtain a new local label correlation matrix , . The learning formula is:

[0031]

[0032] in, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, express The transpose of express The diagonal elements of express -dimensional all-1 vector.

[0033] S33. Similarity graph weight matrix based on training instances and the new local label correlation matrix , construct the manifold regularization term induced label distribution matrix Wherein, the formula of the manifold regularization term is:

[0034]

[0035] in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the total sample size, and Respectively represent Examples and Examples The distribution of markers, For the Examples and Examples The similarity weight between For vector Norm, represents the transpose of a matrix, Indicates the number of categories, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, The new local label correlation matrix The Laplace matrix of is a diagonal matrix whose diagonal elements are , Represents the rank of the matrix.

[0036] In an optional implementation, step S4 specifically includes step S41 to step S42.

[0037] S41. Constructing an objective function of a partial multi-label classifier according to the loss function and the manifold regularization term.

[0038] The objective function formula is:

[0039]

[0040] in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the similarity graph weight matrix, , and represents the regularization parameter, is the weight matrix of the model, represents the offset vector, is the label distribution matrix, is the new local label correlation matrix, represents the feature matrix after mapping, express dimensional all-1 vector, is the total sample size, represents the transpose of a matrix, is the Frobenius norm of the matrix, represents the rank of the matrix, The new local label correlation matrix The Laplace matrix of is a diagonal matrix, is the initial local label correlation matrix, express dimensional all-1 vector, express diagonal elements of .

[0041] S42, using an alternating iterative update strategy to optimize the objective function until convergence or reaching the maximum number of iterations, completing label disambiguation to obtain the optimal model parameters of the partial multi-label classifier and , get the trained partial multi-label classifier. , where represents the feature mapping function, Indicates training examples, and is the optimal model parameter obtained by training, Represents the transpose of a matrix.

[0042] In the second aspect, the present invention provides a partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, characterized in that it includes a training data acquisition module, a similarity graph weight calculation module, a manifold regularization term construction module, an iteration module and a recognition module.

[0043] The training data acquisition module is used to acquire the multi-labeled data set to be processed, convert the candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix. And a loss function is constructed based on the feature space matrix and the label distribution matrix.

[0044] The similarity graph weight calculation module is used to calculate the semantic similarity of the logical label matrix, and then calculate the semantic similarity and -The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples.

[0045] The manifold regularization term construction module is used to calculate the initial local label correlation matrix according to the logical label matrix. According to the initial local label correlation matrix and the label distribution matrix, a local label correlation dynamic learning mechanism is constructed to obtain a new local label correlation matrix. Then, based on the similarity graph weight matrix of the training instance and the new local label correlation matrix, a manifold regularization term induced label distribution matrix is ​​constructed. study.

[0046] An iterative module is used to construct an objective function of a partial multi-label classifier according to the loss function and the manifold regularization term, and perform iterative update processing until the objective function value converges, or the iteration reaches a preset maximum number of iterations, completes label disambiguation to obtain the optimal model parameters, and obtains the trained partial multi-label classifier.

[0047] The recognition module is used to obtain the instance to be tested and input it into the trained partial multi-label classifier to predict all possible categories of the instance to be tested.

[0048] In a third aspect, the present invention provides a partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any paragraph of the first aspect.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any paragraph of the first aspect.

[0050] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0051] The partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation first constructs the relationship between the label distribution matrix and the sample feature matrix through the nonlinear kernel function technology, and uses the semantic similarity of the training instance to construct the similarity graph weight matrix of the training instance and its neighbors. Based on the similarity graph weight matrix and the local label correlation matrix, the label distribution matrix learning is guided to complete the partial multi-label disambiguation. Then, a local label correlation matrix dynamic learning mechanism is constructed to eliminate the influence of noise labels on the model. Finally, a model update strategy is constructed so that the model can effectively predict the label set of partial multi-label data.

[0052] The partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation applies kernel technology and manifold learning to the partial multi-label learning algorithm based on weighted manifold disambiguation and dynamic local label correlation, providing a high-performance classification algorithm for the partial multi-label environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the specific implementation methods of the present invention. It should be understood that the following drawings only show certain specific implementation methods of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 It is a flowchart of the multi-label classification method.

[0055] Figure 2 It is the background knowledge graph of the multi-label classification method.

[0056] Figure 3 It is a schematic diagram of the framework of the multi-label classification method. DETAILED DESCRIPTION

[0057] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] Example 1, please refer to Figures 1 to 3 The first embodiment of the present invention provides a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation, which can be performed by a partial multi-label classification device based on weighted manifold disambiguation and dynamic local label correlation (hereinafter referred to as: classification device). In particular, it is performed by one or more processors in the classification device to implement steps S1 to S5.

[0059] S1, obtain a partially multi-labeled data set to be processed, convert the candidate labels of the partially multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix. And construct a loss function based on the feature space matrix and the label distribution matrix. Based on the above embodiment, in an optional embodiment of the present invention, step S1 includes steps S11 to S14.

[0060] S11. Note that the obtained multi-label dataset is ,in Indicates training examples, represents the set of real numbers, represents the dimension of the feature space, express The candidate tag set To have The output space of possible class labels, represents the number of categories, is the total sample size.

[0061] S12. For each over-labeled sample (in ), constructed using the hard thresholding method Logical label vector of , for any ,if ,but ,otherwise .in, For the Category tags, express No. The logical value of the candidate marker.

[0062] The construction formula of the logical label vector is: .

[0063] S13. Construct the logical label matrix of all training instances and the feature space matrix .in, represents the transpose of a matrix, represents the set of real numbers, is the total sample size, Represents the dimension of the feature space.

[0064] S14. Obtaining a marker distribution matrix according to the feature space matrix .

[0065]

[0066] in, Representation instance The distribution of markers, Represents mapping the feature space to a The characteristic mapping function of the high-dimensional Hilbert space, is the weight matrix of the model, represents the offset vector, express The transpose of .

[0067] Specifically, each row in the multi-labeled dataset is set as a sample. is the logical label matrix of all samples, is the feature space matrix of all samples. Assume that the instance sample (ie: feature space matrix ) and the label distribution matrix There is a nonlinear mapping relationship between them, and then the label distribution matrix is ​​obtained through this nonlinear mapping relationship .

[0068] S15. In order to obtain the weight matrix of the model, a least squares regression model is used to construct a loss function according to the feature space matrix and the label distribution matrix.

[0069] The loss function formula is:

[0070]

[0071] in, represents the feature matrix after mapping, express dimensional all-1 vector, For vector Norm, is the Frobenius norm of the matrix.

[0072] S2, calculate the semantic similarity of the logical label matrix, and then - The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples. Based on the above embodiment, in an optional embodiment of the present invention, step S2 specifically includes steps S21 to S25.

[0073] S21. Constructing semantic similarity matrix using cosine similarity .

[0074]

[0075]

[0076] in, Indicates Examples and Examples The corresponding similarity, is the total sample size, and Respectively represent Examples and Examples The corresponding logical label vector, represents the transpose of a matrix, For vector Norm.

[0077] Specifically, since the logical tag matrix is ​​binary coded, the cosine function is used to construct the semantic similarity matrix To measure the semantic relevance of each group of samples.

[0078] The calculation formula is:

[0079]

[0080] S22. Calculate the distance matrix between training instances .

[0081]

[0082]

[0083] in, Indicates Examples and Examples The distance between.

[0084] Specifically, the commonly used Euclidean distance is used to calculate the distance matrix between instance samples .

[0085] The calculation formula is:

[0086]

[0087] S23. Construct a weighted distance matrix of training instances .

[0088]

[0089]

[0090] in, Indicates Examples and Examples The weighted distance between .

[0091] Specifically, considering that two instance samples with unrelated candidate tag sets have a low probability of becoming domains regardless of whether their feature spaces are similar. Therefore, the weighted distance matrix of the instance samples is calculated based on semantic similarity: .

[0092] The calculation formula is:

[0093]

[0094] S24. For each training instance , select the closest one by weighted distance training instances, Neighborhood .

[0095] S25, calculating the similarity graph weight matrix of the training instance according to the neighborhood .

[0096]

[0097] It is constructed by solving the following linear least squares problem:

[0098]

[0099] in, Represents a training instance for elements in the neighborhood of . For the Examples and Examples The similarity weight between them.

[0100] In this embodiment, the weighted distance of each sample is calculated by Specifically, the weighted distance of each sample is sorted from small to large to obtain . For sample From near to far A set of neighbors. Any number in .

[0101] Based on the above method, samples were obtained Neighborhood , assuming that the sample It can be represented by a linear combination of its neighborhood samples, and the sample is constructed by solving a linear least squares problem Similarity graph weight matrix with its neighborhood samples , .

[0102] The objective function is:

[0103]

[0104] in, Represents a training instance for elements in the neighborhood of .

[0105] In this embodiment, the objective function is optimized using the quadprog function of MATLAB. .

[0106] S3, calculate the initial local label correlation matrix according to the logical label matrix. And build a local label correlation dynamic learning mechanism according to the initial local label correlation matrix and the label distribution matrix to obtain a new local label correlation matrix. Then, build a manifold regularization term induced label distribution matrix based on the similarity graph weight matrix of the training instance and the new local label correlation matrix. study.

[0107] On the basis of the above embodiment, in an optional embodiment of the present invention, step S3 specifically includes steps S31 to S33.

[0108] S31, using the logical labeling matrix to calculate the initial local labeling correlation matrix .

[0109]

[0110]

[0111] in, Indicates Class and The initial local label correlation of each class, and Respectively represent Class and A logical vector of classes, Indicates the kernel size.

[0112] Specifically, the label distribution matrix is ​​learned based on the correlation between labels and labels to complete label disambiguation. In addition, in order to reduce the impact of noise labels on label distribution learning, the label correlation matrix is ​​dynamically learned during the model learning process. Specifically, the initial local label correlation matrix is ​​calculated using the logical label matrix , .

[0113] The calculation formula is:

[0114]

[0115] in, and Respectively represent Class and A logical vector of classes, Indicates the kernel size.

[0116] S32, based on the initial local label correlation matrix and the label distribution matrix Perform dynamic learning to obtain a new local label correlation matrix , .

[0117] The learning formula is:

[0118]

[0119] in, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, express The transpose of express The diagonal elements of express -dimensional all-1 vector.

[0120] Specifically, by constructing a correlation matrix with the initial local label The loss function and consistency with the label distribution correlation learn the local label correlation matrix , .

[0121] S33. Similarity graph weight matrix based on training instances and the new local label correlation matrix , construct the manifold regularization term induced label distribution matrix study.

[0122] The formula of the manifold regularization term is:

[0123]

[0124] in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the similarity graph weight matrix, is the label distribution matrix, is the total sample size, and Respectively represent Examples and Examples The distribution of markers, For the Examples and Examples The similarity weight between For vector Norm, represents the matrix transpose, Indicates the number of categories, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, The new local label correlation matrix The Laplace matrix of is a diagonal matrix whose diagonal elements are , Represents the rank of the matrix.

[0125] In this embodiment, based on the smoothness assumption, the label distribution space should maintain the linear relationship between samples in the domain in the sample feature space, and the correlation between label distributions should be consistent with the local label correlation. Therefore, a manifold regularization term is constructed to induce label distribution matrix learning.

[0126] S4, constructing the objective function of the partial multi-label classifier according to the loss function and the manifold regularization term, and performing iterative update processing until the objective function value converges, or the iteration reaches a preset maximum number of iterations, completing label disambiguation to obtain the optimal model parameters, and obtaining the trained partial multi-label classifier. Based on the above embodiment, in an optional embodiment of the present invention, step S4 specifically includes steps S41 to S42.

[0127] S41. Based on the loss function and manifold regularization term of steps S1 to S3, construct an objective function of a partial multi-label classifier.

[0128] The objective function formula is:

[0129]

[0130] in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the similarity graph weight matrix, , and represents the regularization parameter, is the weight matrix of the model, represents the offset vector, is the label distribution matrix, is the new local label correlation matrix, represents the feature matrix after mapping, express dimensional all-1 vector, is the total sample size, represents the transpose of a matrix, is the Frobenius norm of the matrix, represents the rank of the matrix, The new local label correlation matrix The Laplace matrix of is a diagonal matrix, is the initial local label correlation matrix, express dimensional all-1 vector, express diagonal elements of .

[0131] S42, using an alternating iterative update strategy to optimize the objective function until convergence or reaching the maximum number of iterations, completing label disambiguation to obtain the optimal model parameters of the partial multi-label classifier and , get the trained partial multi-label classifier.

[0132]

[0133] In the formula, represents the feature mapping function, Indicates training examples, and is the optimal model parameter obtained by training, Represents the transpose of a matrix.

[0134] Specifically, a commonly used alternating iterative update strategy is used to update the variables in the objective function. Iterate and update until the objective function converges or the maximum number of iterations is reached. By optimizing this formula, the optimal model parameters can be obtained. and , determine a partial multi-label classifier.

[0135] by For example, the expression of the partial multi-label classifier is:

[0136]

[0137] It should be noted that the classifier can adopt an existing classification algorithm. Then, training is performed through steps S1 to S4 to obtain the multi-label classifier. Among them, the existing classification algorithm can be a probability model-based method, an optimization-based method, a feature representation-based method, a rule-based learning, an ensemble learning method, transfer learning / transfer learning, or a label propagation algorithm, which is not specifically limited in the present invention.

[0138] S5, obtaining a test instance and inputting it into the trained partial multi-label classifier to predict all possible categories of the test instance and output them.

[0139] In this embodiment, the trained multi-label classifier is used to classify the test instance. A rough estimate of the marker distribution is made, and the estimation model formula is as follows: Specifically, by comparing The label distribution vector The value of each element and the threshold value The category label to which it belongs is used to obtain the final classification result output.

[0140] The judgment formula is:

[0141]

[0142] in, The test instance output by the model The distribution of markers, represents the feature mapping function, and is the optimal model parameter obtained by training, represents the transpose of a matrix, For the Category tags, for No. element values, is the threshold parameter (in this embodiment, )、 Indicates the number of categories.

[0143] In summary, the partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation is the first to explore the impact of semantic correlation on the sample topological structure, and proposes a new learning framework for PML. In addition, this method invents a new local label correlation learning mechanism, which aims to eliminate the impact of noise labels on label correlation. The partial multi-label classification method further integrates the neighbor information of the training instance and the label distribution matrix learning based on label correlation. Finally, the partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation constructs a model update strategy so that the model can adapt to multi-label data in a noisy environment and correctly classify unknown samples.

[0144] The partial multi-label learning method based on weighted manifold disambiguation and dynamic local label correlation proposed in this application effectively solves the problems of insufficient mining of negative information between features and labels and the influence of noise labels in traditional partial multi-label learning. Through the nonlinear kernel function technology, the relationship between the sample feature matrix and the label distribution matrix is ​​constructed. At the same time, the semantic similarity of the training instance and the local label correlation matrix are used to guide the learning of the label distribution matrix, thereby realizing the noise disambiguation of partial multi-label data. In addition, the method also constructs a dynamic learning mechanism of the local label correlation matrix to eliminate the negative impact of noise labels on model performance, and improves the accuracy of the model's prediction of the label set of partial multi-label data through the model update strategy. This innovative method provides a high-performance classification algorithm for partial multi-label environments, which has important practical application value and theoretical significance, especially in the application of text classification, ecological informatics and multimedia content analysis. It shows significant advantages.

[0145] The following is a detailed description of the above solution using a specific case. In the multi-labeled dataset to be processed, each sample contains sample features (e.g., images, text, audio, or video data) and candidate tags (e.g., tags that are manually or otherwise added to the sample features). The candidate tags are divided into correct tags and irrelevant tags (e.g., noise tags).

[0146] Preferably, the multi-labeled dataset to be processed can be some labeled sample images, such as Figure 2 As shown, the image sample is labeled by the labeler as "mountain", "lake", "tree", "flower", "cloud", "bird", "house" and "road", which are the candidate label sets of the image sample, but "bird", "house" and "road" in these candidate label sets are not elements contained in the image and are noise labels.

[0147] like Figure 3 As shown, the partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation of an embodiment of the present invention first trains the partial multi-label classifier through the partial multi-label data set to be processed according to steps S1 to S4.

[0148] Preferably, the multi-label classifier can be selected according to the actual data type, for example: when the sample feature is an image, the multi-label classifier can be CNNs, object detection models or image segmentation models. When the sample feature is a video, the multi-label classifier can be 3D CNNs, two-stream CNNs or RNNs. When the sample feature is text, the multi-label classifier can be RNNs, Transformers or word embedding. When the sample feature is audio, the multi-label classifier can be CNNs, RNNs, MFCCs or WaveNet.

[0149] After the multi-label classifier is trained, the instance to be tested (unlabeled, the data type can be: picture, text, video or audio) is input into the trained multi-label classifier, and all possible labels (ie: categories) of the instance to be tested are output.

[0150] It is understandable that when the instance to be tested is an image, the label may be an object contained in the image or the style of the image, etc. When the instance to be tested is audio, the label may be the style of the audio. When the instance to be tested is text, the label may be the emotion or theme contained in the text. When the instance to be tested is bioinformatics data, the label may be protein function prediction and gene expression analysis, etc. When the instance to be tested is user behavior data, the label may be the user's preference. The embodiment of the present invention does not limit the specific data type of the instance to be tested, nor does it limit its label (i.e., the category to which it belongs).

[0151] The partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation of the embodiment of the present invention predicts the label set that each to-be-predicted instance may belong to based on the relationship between the features and labels learned in the training data, performs classification prediction among the numerous labels containing noise labels, and labels the unlabeled to-be-predicted data with the correct label for user reference. This provides a reference for social activities such as data classification and doctor diagnosis, and the efficiency of labeling data through the partial multi-label classification method is far better than manual work, with good technical effects.

[0152] Embodiment 2: The present invention provides a partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, characterized in that it includes a training data acquisition module, a similarity graph weight calculation module, a manifold regularization term construction module, an iteration module and a recognition module.

[0153] The training data acquisition module is used to acquire the multi-labeled data set to be processed, convert the candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix. And a loss function is constructed based on the feature space matrix and the label distribution matrix.

[0154] The similarity graph weight calculation module is used to calculate the semantic similarity of the logical label matrix, and then calculate the semantic similarity and -The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples.

[0155] The manifold regularization term construction module is used to calculate the initial local label correlation matrix according to the logical label matrix. According to the initial local label correlation matrix and the label distribution matrix, a local label correlation dynamic learning mechanism is constructed to obtain a new local label correlation matrix. Then, based on the similarity graph weight matrix of the training instance and the new local label correlation matrix, a manifold regularization term induced label distribution matrix is ​​constructed. study.

[0156] An iterative module is used to construct an objective function of a partial multi-label classifier according to the loss function and the manifold regularization term, and perform iterative update processing until the objective function value converges, or the iteration reaches a preset maximum number of iterations, completes label disambiguation to obtain the optimal model parameters, and obtains the trained partial multi-label classifier.

[0157] The recognition module is used to obtain the instance to be tested and input it into the trained partial multi-label classifier to predict all possible categories of the instance to be tested.

[0158] Embodiment 3, the present invention provides a partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any paragraph of Embodiment 1.

[0159] Embodiment 4: The present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any paragraph of Embodiment 1.

[0160] It is understandable that the classification device may be an electronic device with computing capabilities, such as a portable notebook computer, a desktop computer, a server, a smart phone, or a tablet computer.

[0161] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0162] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0163] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0164] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0165] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0166] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0167] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0168] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation, characterized in that: Include: Obtain a multi-labeled data set to be processed, convert candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix; A loss function is constructed according to the feature space matrix and the label distribution matrix; wherein, in the multi-label data set to be processed, each sample includes a sample feature and a candidate label; the sample is an image, text, audio or video; the candidate label is a label that is manually or otherwise marked on the sample feature; Calculate the semantic similarity of the logical tag matrix, and then - The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples; The initial local label correlation matrix is ​​calculated according to the logical label matrix; and a local label correlation dynamic learning mechanism is constructed according to the initial local label correlation matrix and the label distribution matrix to obtain a new local label correlation matrix; then, a manifold regularization term induced label distribution matrix is ​​constructed based on the similarity graph weight matrix of the training instance and the new local label correlation matrix. study; Constructing an objective function of a partial multi-label classifier according to the loss function and the manifold regularization term, performing iterative updating processing until the objective function value converges, or the iteration reaches a preset maximum number of iterations, completing label disambiguation to obtain optimal model parameters, and obtaining a trained partial multi-label classifier; Obtain a test instance and input it into the trained partial multi-label classifier to predict all possible categories of the test instance.

2. According to claim 1, a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation is characterized in that: Obtain a multi-labeled data set to be processed, convert candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix; And construct a loss function according to the feature space matrix and the label distribution matrix, specifically including: Get the multi-label dataset to be processed, denoted as ,in Indicates training examples, express The candidate tag set, is the total sample size; For each over-labeled sample , constructed using the hard thresholding method Logical label vector of :For any ,if ,but ,otherwise ;in, represents the number of categories, For the Category tags, express No. The logical value of the candidate marker; Construct the logical label matrix of all training instances and the feature space matrix ;in, represents the transpose of a matrix, represents the set of real numbers, is the total sample size, Represents the dimension of the feature space; According to the feature space matrix, obtain the label distribution matrix : , where express The distribution of markers, represents the feature mapping function, is the weight matrix of the model, represents the offset vector; According to the feature space matrix and the label distribution matrix, a least squares regression model is used to construct a loss function; the loss function formula is: in, represents the feature matrix after mapping, express dimensional all-1 vector, For vector Norm, is the Frobenius norm of the matrix.

3. The method for partial multi-label classification based on weighted manifold disambiguation and dynamic local label correlation according to claim 1, characterized in that: Calculate the semantic similarity of the logical label matrix, then find the neighbor samples of the training instance according to the semantic similarity and the k-nearest neighbor method, and calculate the similarity graph weight matrix between the training instance and its neighbor samples, specifically including: Constructing semantic similarity matrix using cosine similarity ; , ;in, Indicates Examples and Examples The corresponding similarity, is the total sample size, and Respectively represent Examples and Examples The corresponding logical label vector, represents the transpose of a matrix, For vector norm; Compute the distance matrix between training instances ; , ;in, Indicates Examples and Examples The distance between Construct a weighted distance matrix for training instances ; , ;in, Indicates Examples and Examples The weighted distance between For each training instance , select the closest one by weighted distance training instances, Neighborhood ; Calculate the similarity graph weight matrix of the training instance based on the neighborhood ; It is constructed by solving the following linear least squares problem: in, Represents a training instance for The elements in the neighborhood of ; For the Examples and Examples The similarity weight between them.

4. The method for partial multi-label classification based on weighted manifold disambiguation and dynamic local label correlation according to claim 1, characterized in that: The initial local label correlation matrix is ​​calculated according to the logical label matrix; and a local label correlation dynamic learning mechanism is constructed according to the initial local label correlation matrix and the label distribution matrix to obtain a new local label correlation matrix; then, a manifold regularization term induced label distribution matrix is ​​constructed based on the similarity graph weight matrix of the training instance and the new local label correlation matrix. Learning, including: Calculate the initial local label correlation matrix using the logical label matrix ; , ,in, Indicates Class and The initial local label correlation of each class, and Respectively represent Class and A logical vector of classes, represents the kernel size; Based on the initial local label correlation matrix and the label distribution matrix Perform dynamic learning to obtain a new local label correlation matrix , ; The learning formula is: in, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, express The transpose of express The diagonal elements of express dimensional all-1 vector; Similarity graph weight matrix based on training instances and the new local label correlation matrix , construct the manifold regularization term induced label distribution matrix Learning; wherein the manifold regularization term formula is: in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the total sample size, and Respectively represent Examples and Examples The distribution of markers, For the Examples and Examples The similarity weight between For vector Norm, represents the transpose of a matrix, Indicates the number of categories, Indicates Class and New local label correlations for classes, and Respectively represent Class and The label distribution vector of each class, The new local label correlation matrix The Laplace matrix of is a diagonal matrix whose diagonal elements are , Represents the rank of the matrix.

5. The method for partial multi-label classification based on weighted manifold disambiguation and dynamic local label correlation according to claim 1, characterized in that: The objective function of the partial multi-label classifier is constructed according to the loss function and the manifold regularization term, and an iterative update process is performed until the objective function value converges or the iteration reaches a preset maximum number of iterations, and the label disambiguation is completed to obtain the optimal model parameters, and the trained partial multi-label classifier is obtained, which specifically includes: The objective function of the partial multi-label classifier is constructed according to the loss function and the manifold regularization term; the objective function formula is: in, , It is a substitution symbol used to connect the two formulas. It has no actual meaning. express dimensional identity matrix, is the similarity graph weight matrix, , and represents the regularization parameter, is the weight matrix of the model, represents the offset vector, is the label distribution matrix, is the new local label correlation matrix, represents the feature matrix after mapping, express dimensional all-1 vector, is the total sample size, represents the transpose of a matrix, is the Frobenius norm of the matrix, represents the rank of the matrix, The new local label correlation matrix The Laplace matrix of is a diagonal matrix, is the initial local label correlation matrix, express dimensional all-1 vector, express The diagonal elements of ; The objective function is optimized by using an alternating iterative update strategy until convergence or the maximum number of iterations is reached, and label disambiguation is completed to obtain the optimal model parameters of the partial multi-label classifier. and , obtain the trained partial multi-label classifier; , where represents the feature mapping function, Indicates training examples, and is the optimal model parameter obtained through training.

6. A partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, characterized in that: Include: The training data acquisition module is used to obtain the multi-labeled data set to be processed, convert the candidate labels of the multi-labeled data into a logical label matrix, and then construct a feature space matrix and a label distribution matrix; A loss function is constructed according to the feature space matrix and the label distribution matrix; wherein, in the multi-label data set to be processed, each sample includes a sample feature and a candidate label; the sample is an image, text, audio or video; the candidate label is a label that is manually or otherwise marked on the sample feature; The similarity graph weight calculation module is used to calculate the semantic similarity of the logical label matrix, and then calculate the semantic similarity and - The nearest neighbor method finds the nearest neighbor samples of the training instance and calculates the similarity graph weight matrix between the training instance and its nearest neighbor samples; A manifold regularization term construction module is used to calculate an initial local label correlation matrix based on the logical label matrix; and to construct a local label correlation dynamic learning mechanism based on the initial local label correlation matrix and the label distribution matrix to obtain a new local label correlation matrix; and then to construct a manifold regularization term induced label distribution matrix based on the similarity graph weight matrix of the training instance and the new local label correlation matrix. study; An iteration module is used to construct an objective function of a partial multi-label classifier according to the loss function and the manifold regularization term, and perform iterative update processing until the objective function value converges or the iteration reaches a preset maximum number of iterations, completes label disambiguation to obtain optimal model parameters, and obtains a trained partial multi-label classifier; The recognition module is used to obtain the instance to be tested and input it into the trained partial multi-label classifier to predict all possible categories of the instance to be tested.

7. A partial multi-label learning device based on weighted manifold disambiguation and dynamic local label correlation, characterized in that: It includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a partial multi-label classification method based on weighted manifold disambiguation and dynamic local label correlation as described in any one of claims 1 to 5.

Images