An online data classification method based on semi-supervised subspace representation model

Abstract

The application discloses a data online classification method based on a semi-supervised subspace representation model, which comprises the following steps: firstly, pre-processing a large-scale high-dimensional data set; then, selecting part of samples from the labeled samples of each category to construct a small-scale dictionary, and taking the rest of the samples as a data set to be classified; next, for each sample to be classified, a dictionary representation model with affine, non-negative and sparse constraints is established based on a global dictionary; then, an alternating iterative optimization algorithm is designed by using an augmented Lagrange multiplier method to solve the optimal coefficient representation vector of each sample to be classified; finally, the correlation degree of the small-scale dictionary atom labels corresponding to the non-zero elements in the coefficient vector is analyzed to determine the category of the sample to be classified. The application reduces the calculation complexity by using a small-scale dictionary, improves the classification accuracy by using semi-supervised information and sparse constraints, and supports online processing of large-scale high-dimensional data.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and data mining technology, specifically relating to an online data classification method based on a semi-supervised subspace representation model. Background Technology

[0002] In the fields of data mining and artificial intelligence, particularly in the development of large language models and deep learning, academia and industry are undergoing a paradigm shift from "model-centric" to "data-centric." The necessity of dataset quality assessment is mainly reflected in three aspects. First, data quality determines the upper and lower bounds of model generalization; the classic computer science law of "garbage in, garbage out" still applies in the era of deep learning. Research shows that label noise, distribution bias, and information redundancy in datasets directly lead to overfitting, decision bias, and decreased robustness. Second, the high cost of training and computing power is a constraint. With the emergence of multimodal large models and large-scale image datasets, the computing power cost of training with full data is extremely high. Selecting high-quality, high-information-density core datasets through dataset quality assessment methods can maintain or even improve model performance while reducing the amount of data by more than 50%. Finally, there are rigid requirements for security and compliance. In high-reliability fields such as healthcare, autonomous driving, and military equipment, data defects (such as outliers, artifacts, and mislabeling) can lead to catastrophic consequences. Meanwhile, data privacy, toxicity, and compliance assessments are also an indispensable part of data quality assessment methodologies.

[0003] In dataset quality assessment, the completeness, accuracy, and consistency of the dataset are important evaluation metrics. Specifically, for dataset accuracy, the data samples need to realistically reflect the real-world entities or events they represent, free from noise or errors. Dataset accuracy is primarily reflected in feature accuracy (i.e., undistorted sensor readings or physical measurements) and label accuracy (i.e., the correct labels for classification tasks are not incorrectly assigned).

[0004] In datasets, accuracy assessment is one of the most technically challenging aspects of quality evaluation methods. It can be broadly categorized into four types based on different learning approaches: decisional confidence learning methods, deep learning-based methods, training dynamics-based variability detection methods, and representation learning methods. For decisional confidence learning methods, the model jointly analyzes the probability distribution of noisy labels and the model's predicted probability distribution, identifying inconsistent samples by estimating the joint transpose matrix of the noisy and true labels. Common deep learning methods for sample accuracy assessment include Co-teaching and DivideMix. These methods calculate and sort the loss values ​​of all samples during training, classifying samples with loss values ​​below a given threshold as potential label noise or inaccurate samples, and either removing them or transferring them to semi-supervised learning for pseudo-label correction. Training dynamics-based variability detection methods record the average confidence level and standard deviation (variability) of the model's predicted probability of a sample's true label during model training, thus creating a data map. Samples with high variability are classified as ambiguous / inaccurate samples. Representation learning methods utilize self-supervised or contrastive learning models to map data into a high-dimensional feature embedding space, and use the geometric distribution and density features of samples in the feature space to determine their accuracy.

[0005] In representation learning methods, subspace representation methods, due to their high-dimensional learning capabilities, are frequently used in high-dimensional data processing, often combined with spectral analysis to address high-dimensional data clustering problems. This method assumes that data from a high-dimensional space is distributed across multiple low-dimensional subspaces; therefore, its main working principle is to learn the low-dimensional subspaces of samples of the same class in the high-dimensional data. Subspace representation methods are primarily implemented by applying regularization constraints to the learned coefficient matrix in a dictionary representation learning model. There are two main approaches to learning the optimal coefficient matrix: sparse learning, such as sparse subspace learning methods, and low-rank representation, such as robust subspace segmentation algorithms based on low-rank representation. Based on these two typical methods, many different variants have been developed in recent years, such as scalable sparse subspace learning algorithms and implicit low-rank representation methods. Due to its high-dimensional learning capability, subspace representation methods can effectively handle high-dimensional datasets, especially for data such as images and videos.

[0006] Therefore, combining subspace representation learning models with the problem of judging the accuracy of samples in datasets is a promising application of machine learning methods. However, dictionary-based subspace representation learning models often use the original data as a dictionary for self-representation learning. This leads to an excessively large and redundant coefficient matrix when processing large-scale image datasets due to the large dictionary scale, significantly increasing computation time and storage costs, and making it unsuitable for direct online data processing. Furthermore, most existing subspace representation methods are unsupervised learning methods, and directly using them for judging the accuracy of data samples results in a high misclassification rate, severely impacting dataset quality assessment. Therefore, for traditional subspace representation models, how to efficiently evaluate the accuracy of large-scale image dataset samples online is a crucial problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to provide an online data classification method based on a semi-supervised subspace representation model, which solves the problems of high storage cost, high computation time and missing label information faced by traditional classification methods for large-scale high-dimensional datasets. It can be effectively applied to scenarios such as large-scale image dataset classification and quality assessment.

[0008] To achieve the above objectives, the present invention employs the following technical solution: An online data classification method based on a semi-supervised subspace representation model includes: Step 1: Preprocess the given image dataset; where some images in the dataset have known labels, these are denoted as labeled samples. Step 2: For the preprocessed image dataset, select a portion of the labeled samples from each category as a sub-dictionary for that category; merge the sub-dictionaries corresponding to each category to form a small-scale dictionary; the remaining samples other than those used to construct the small-scale dictionary are used as samples to be classified. Step 3: For the samples to be classified, construct a sparse dictionary representation model based on semi-supervised information; Step 4: Use an alternating iterative optimization algorithm based on augmented Lagrange multipliers to solve the sparse dictionary representation model and obtain the optimal coefficient representation vector; Step 5: For the optimal coefficient representation vector learned by the sample to be classified, the correlation analysis is performed by combining it with the correspondence with the labeled samples in the small-scale dictionary, and the predicted label of the sample to be classified is obtained according to the label of the labeled samples in the small-scale dictionary, so as to achieve online classification.

[0009] Furthermore, the given image dataset undergoes preprocessing, including: First, each sample is converted into a vector form using feature extraction methods; then, principal component analysis is used to initially remove noise information; finally, normalization methods are used to process the samples.

[0010] Furthermore, the sparse dictionary representation model based on semi-supervised information is as follows: ; in, For the first One sample to be classified; Indicates the first A vector of coefficients to be optimized; For small-scale dictionaries; Represents a column vector whose elements are all 1s; express Norm; express Norm; superscript T indicates transpose.

[0011] Furthermore, by transforming the formula of the objective function of the sparse dictionary representation model based on semi-supervised information, we can obtain the following equivalent problem: ; A column vector was introduced. The coefficients represent vectors Perform equivalent substitution, that is Therefore, the above problem can be equivalently transformed into the following problem: ; Constructing an augmented Lagrangian function transforms the constrained optimization problem into one involving column vectors. Unconstrained optimization problem: ; in, Represents the Lagrange multipliers; This is the penalty parameter.

[0012] Furthermore, an alternating iterative optimization algorithm based on the augmented Lagrange multiplier method: When the coefficients represent vectors Once fixed, update the intermediate variables. Then the unconstrained optimization problem can be further simplified to the following problem: ; about Differentiate and set to zero, then The optimal solution can be obtained by the following formula: .

[0013] Furthermore, the alternating iterative optimization algorithm based on the augmented Lagrange multiplier method further includes: When intermediate variables After fixing, update the coefficient representation vector. Then the unconstrained optimization problem can be further simplified to: ; Introducing vectors Its definition is: ; but The analytical solution can be calculated using the following formula: ; in, This represents the final coefficient representation vector. In the optimal support set The value of ; Describe a vector whose set is the optimal support set. The elements in the vector are used as the indices. In addition to the elements having values, All other elements in the sequence are 0; The coefficient representation vector In the optimal support complement The value of is set to 0 here; For threshold The soft threshold projection operator is defined. , The superscript + indicates that the operator performs a positive projection operation; optimal support set ; This indicates that a hard threshold truncation operator is used first. Preservation vector Previous Find the maximum value, set the remaining elements to 0, and then take the indices of the non-zero elements in the resulting vector; this gives the optimal support set. ; Therefore, the coefficient representation vector can be obtained. The optimal solution in the current iteration step, i.e., by Given the non-zero part, by Give the complete vector of the zero-value portion.

[0014] Furthermore, for vectors Each element of is calculated using the following formula: ; in, Indicates The first element in the result vector after operator processing Each element value; For vectors of Each element value; A soft threshold, used for... Perform a contraction operation; It is an orthographic projection operator; This means "defined as"; The maximum index value is defined as follows: ; in, Representing vectors of Each element value.

[0015] Furthermore, based on the atomic positions corresponding to the non-zero elements of the superior coefficient representation vector in the small-scale dictionary, the label information corresponding to these atoms is extracted; by statistically analyzing the categories corresponding to these non-zero elements, the category with the strongest correlation is determined as the predicted label of the sample to be classified.

[0016] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the online data classification method based on a semi-supervised subspace representation model.

[0017] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the online data classification method based on a semi-supervised subspace representation model.

[0018] Compared with the prior art, the present invention has the following technical features: This invention preprocesses the evaluation dataset through feature extraction, followed by dimensionality reduction using principal component analysis. The preprocessed dataset is then applied to an online data classification method based on a semi-supervised subspace representation model. Dictionary representation learning and association analysis are used to learn and determine the label information for each data sample, achieving real-time online data classification. To effectively address the problem of high error rates in data sample accuracy determination by existing unsupervised learning methods during dataset quality assessment, this invention proposes an online data classification method based on a semi-supervised subspace representation model. By introducing the subspace representation model, it not only improves the processing capability for high-dimensional data but also provides online data representation capabilities. Furthermore, this invention designs a small-scale dictionary construction method to reduce computational complexity and redundant information in the dictionary representation process, and the combination of semi-supervised information effectively improves the classification accuracy of data samples. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of online data classification based on a semi-supervised subspace representation model; Figure 2 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0020] This invention provides an online data classification method based on a semi-supervised subspace representation model. It introduces a small-scale dictionary construction method to replace all the original data, thereby reducing the computational complexity and redundant information in the dictionary representation process. Furthermore, to improve the accuracy of data sample discrimination, this invention introduces the label categories of data samples in the dictionary during the small-scale dictionary construction process. The label category of the corresponding sample is determined by the learned coefficient matrix and prior knowledge of the dictionary data samples. In addition, combined with the augmented Lagrangian function optimization method, this invention designs an efficient optimization algorithm based on alternating iteration to solve the proposed model. This invention solves the problems of high computational and storage costs, information redundancy, and high misclassification rates encountered by existing subspace representation learning methods when evaluating the accuracy of dataset samples. The specific method of this invention is as follows: Step 1 involves preprocessing the given large-scale image dataset: First, each sample is converted into a vector form using feature extraction methods; then, Principal Component Analysis (PCA) is used to initially remove noise and reduce feature redundancy; finally, normalization is applied to ensure that all sample vectors are consistent in numerical scale, improving the consistency of feature representation. This step outputs a preprocessed large-scale image dataset (containing a small number of samples with known labels and a large number of samples with unknown labels), providing the foundational data for dictionary construction in Step 2. In this embodiment, the large-scale image dataset is a face image dataset, ultimately achieving face image classification; as... Figure 1 The example shown is the classification process for a face image dataset, where each sample is a face image. Some face images have identity labels, and the final goal is to predict the identity labels of face images without labels. In other application examples, emotion labels, such as happy or depressed, can also be used to predict the emotion labels of the samples.

[0021] Step 2: In the preprocessed large-scale image dataset, there are some samples with known labels, which are denoted as labeled samples. In order to reduce dictionary scale and computational complexity, this invention does not use all labeled samples, but selects some samples from the labeled samples of each category as a sub-dictionary for that category. The sub-dictionaries corresponding to each category are merged to form a small-scale dictionary. The remaining samples other than those used to construct the small-scale dictionary are used as samples to be classified. (The remaining unlabeled samples and the labeled samples not selected into the sub-dictionary) together constitute the dataset to be classified.

[0022] For preprocessed large-scale image datasets ,in Represents the total number of samples. Represents the feature dimension of the sample. Represent the real number space; assume a large-scale image dataset contains There are several categories, and each category contains samples with known labels; then, a predetermined number of samples are selected from the samples of each category to form a sub-dictionary for each category. The selection of samples can be achieved through methods such as random sampling or cluster center selection; all sub-dictionaries are merged into a single global small-scale dictionary, as follows: (1) in, Indicates the first Sub-dictionaries corresponding to each category; This represents the number of samples in the global small-scale dictionary, which is much smaller than the total number of samples. This ensures the small-scale nature of the dictionary.

[0023] This step outputs a small-scale dictionary. (Including the label information of the corresponding atoms) and the dataset to be classified (i.e., the set of all samples that were not selected into the dictionary).

[0024] Step 3: For each sample to be classified, construct a sparse dictionary representation model based on semi-supervised information to solve for the optimal coefficient representation vector.

[0025] Based on the fundamental assumptions of subspace representation learning, each sample to be classified Both can be derived from small-scale dictionaries The atoms in the equation are approximated by linear combinations, that is, in the ideal case, they are represented as: (2) in, This is the coefficient representation vector under ideal conditions.

[0026] Although traditional subspace representation methods use the entire dataset To be used as a dictionary, that is The method achieves good representation performance on small datasets. However, it often suffers from information redundancy and high time cost when dealing with large-scale image datasets. In addition, due to the presence of noise in the data and the fact that the representation is not completely accurate, it is necessary to minimize the representation error in practice.

[0027] Based on the aforementioned small-scale dictionary construction method, and introducing affine constraints (ensuring the sum of coefficients is 1), nonnegativity constraints (ensuring all coefficient elements are nonnegative), and sparsity constraints (controlling the number of non-zero coefficients), combined with the sample label information in the known dictionary, this invention proposes a dictionary representation model based on semi-supervised information. For each sample to be classified... The optimization problem is as follows: (3) in, Indicates the first There are vectors of coefficients to be optimized, each corresponding to a labeled sample. Small-scale dictionary Linear combinations under the following conditions Represents a column vector whose elements are all 1s; express Norm; in the objective function Indicates minimization; The superscript T indicates transpose, and the same applies below.

[0028] For the optimization problem (3), here we use Norms are used to measure dictionary representation loss. In general subspace representation learning problems, not all subspaces cross zeros; that is, some subspaces are affine subspaces. Therefore, based on the definition of affine subspaces, for coefficient representation vectors... The first constraint condition, as in problem (3), is given; furthermore, in order to ensure the coefficient vector This indicates that the association analysis can be performed correctly and the classification results can be obtained. A second constraint is introduced here. This ensures that each element in the coefficient vector is not less than 0.

[0029] For the optimization problem (3) mentioned above, the coefficient vector can be obtained by performing subspace representation learning on each data sample, and then the classification result can be obtained through association analysis. Since different data samples can be represented and classified separately, the proposed model can perform real-time online classification of data. However, the optimization solution of problem (3) may fall into the predicament of trivial solutions. This is because of the inherent defects of dictionary representation, that is, for a given first-order dictionary... For a data sample of a certain category, the data sample can be derived from the first category. Data samples from one class and data samples from other classes are represented together. In this case, the learned coefficient representation vector is... This is not the optimal representation. To avoid such trivial solutions, constraints need to be placed on the learned coefficient representation vector. Similar to sparse subspace clustering algorithms, this invention introduces a pre-defined sparsity parameter. This is used to apply sparsity constraints to the coefficient representation vector, limiting the maximum number of non-zero numbers in the coefficient representation vector.

[0030] Therefore, combining the optimization problem (3) with the sparse constraints, the sparse dictionary representation model based on semi-supervised information proposed in this invention has the following final optimization problem: (4) in, express Norm.

[0031] It can be seen that the proposed sparse dictionary representation model based on semi-supervised information can achieve online representation of samples, and the coefficient constraints are... The introduction of this ensures that the learned coefficient vector It has a clearer correlation, thereby improving classification accuracy.

[0032] Step 4: Alternating iterative optimization algorithm based on augmented Lagrange multiplier method.

[0033] Due to the introduction of sparsity constraints, among them The norm is non-convex, which makes it impossible to directly optimize and solve problem (2). To this end, this invention combines the augmented Lagrange multiplier method and designs a simple and effective alternating iterative optimization algorithm based on the augmented Lagrange multiplier method to solve the above problem.

[0034] First, by transforming the objective function of optimization problem (4), we can obtain the following equivalent problem: (5) To introduce the augmented Lagrange multiplier method, this invention introduces a column vector. The coefficients represent vectors Perform equivalent substitution, that is Used in the optimization process to handle variables Separating the problems facilitates a step-by-step solution; thus, problem (5) can be equivalently transformed into the following problem: (6) To address the equality constraints in problem (6) above, we construct the corresponding augmented Lagrangian function, transforming the constrained optimization problem into one concerning column vectors. The unconstrained optimization problem takes the following form: (7) here, Denotes a Lagrange multiplier with dimension . ; The penalty parameter is a positive scalar used to control the weight of the penalty term. Its value increases gradually during iteration to enforce the constraint.

[0035] Since the objective function of problem (7) contains cross terms, this invention adopts the idea of ​​alternating direction multiplier method to update variables separately. and To solve: Step 4.1, when the coefficients represent vectors Once fixed, update the intermediate variables. Then the optimization problem (7) can be further simplified to the following problem: (8) As can be seen, question (8) concerns variables. It is a convex quadratic optimization problem. Therefore, problem (8) is about By taking the derivative and setting it to zero, the optimal solution to problem (8) can be obtained by the following formula: (9) Step 4.2, when intermediate variables After fixing, update the coefficient representation vector. Then the optimization problem (7) can be further simplified to: (10) For convenience, we introduce vectors here. Its definition is: (11) The optimization problem (10) describes a problem with affine constraints, nonnegativity constraints, and... The sparsely constrained Euclidean projection problem; according to the literature (A. Kyrillidis, S. Becker, V. Cevher, and C. Koch, Sparseprojections onto the simplex, in Proc. Int. Conf. Mach. Learn., 2013, pp.235–243.), its analytical solution can be determined by the following steps: First, define the vector. Support set; command That is, retain the non-negative part; For vectors The 1 element; given set its supplement About sets Defined by; set The potential (number of elements) is Define a vector Support set as ; To support set operators; Represents all vectors that satisfy the condition. The The index of the condition that the element is not equal to 0 The set that makes up the whole; finally, define A vector, except for the set The elements in the vector are used as the indices. In addition to the elements having values, All other elements in the sequence are 0.

[0036] Based on the above definitions, the analytical solution to problem (10) can be calculated using the following formula: (12) in, This represents the final coefficient representation vector. In the optimal support set The values ​​that can be taken on the value, i.e., the non-zero part; Describe a vector whose set is the optimal support set. The elements in the vector are used as the indices. In addition to the elements having values, All other elements in the sequence are 0; The coefficient representation vector In the optimal support complement The value of is set to 0 here; Indicates complement; For threshold The soft threshold projection operator is used to subtract a threshold from each element of the vector and then take the positive value; this scheme defines... , The superscript + indicates that the operator performs a positive projection operation. Optimal support set ; This indicates that a hard threshold truncation operator is used first. Preservation vector Previous Find the maximum value, set the remaining elements to 0, and then take the indices of the non-zero elements in the resulting vector; this gives the optimal support set. .

[0037] Furthermore, for vectors Each element of is calculated using the following formula: (13) in, Indicates The first element in the result vector after operator processing Each element value; For vectors of Each element value; A soft threshold, used for... Perform a contraction operation; It is an orthographic projection operator; This means "defined as"; Indicates the maximum index value. The definition is as follows: ; in, Representing vectors of Each element value The operator for finding the maximum value.

[0038] Therefore, the coefficient representation vector can be obtained through formula (12) and its corresponding threshold calculation formula. The optimal solution in the current iteration step, i.e., by Given the non-zero part, by Give the complete vector of the zero-value part It will be used as the input for the next iteration or as the output as the final coefficient representation vector.

[0039] Step 4.3, updating parameters At a given learning rate Use the formula below To perform the calculation; for Lagrange multipliers Its update calculation formula is: .

[0040] Step 4.4: Repeat steps 4.1 to 4.3 above until the algorithm converges (e.g., reaching the maximum number of iterations or within two adjacent iterations). and If the difference is less than a set threshold, then the current sample to be classified is obtained. The optimal coefficient representation vector .

[0041] Step 5: Association analysis and sample classification based on dictionary labels.

[0042] For the samples to be classified The learned optimal coefficient representation vector Association analysis is performed by combining the correspondence with labeled samples in the small-scale dictionary, and the sample to be classified is obtained based on the labels of the labeled samples in the small-scale dictionary. The predicted labels are used to classify the samples.

[0043] Specifically, the vector is represented by the dominance coefficient. Non-zero elements in a small-scale dictionary The positions of the atoms (each atom corresponds to a labeled sample) are analyzed, and the label information (i.e., the category of the labeled sample to which the atom belongs) is extracted from these atoms. By statistically analyzing the categories corresponding to these non-zero elements, the category with the highest proportion (i.e., the strongest correlation) is determined as the sample to be classified. Predicted labels.

[0044] By iteratively repeating the calculation process in steps 3 and 5, all samples to be classified are solved independently and correlation analysis is performed sequentially. Since the optimization process for each sample to be classified is independent, this method supports online classification in a streaming manner. Once all samples to be classified have been labeled, the final classification result for the entire dataset is output.

[0045] Based on the above optimization steps, this invention designs the above algorithm flow to iteratively solve the optimization problem (4) until convergence, thereby obtaining the optimal coefficient representation vector for each data sample. Then, the correlation analysis between this representation vector and the dictionary is performed, thus directly obtaining the final classification result. Since each data sample can be individually represented in a subspace and analyzed for correlation, the method proposed in this invention can achieve online processing of large-scale high-dimensional datasets, effectively improving classification efficiency. Furthermore, the design of the small-scale dictionary construction strategy and the introduction of semi-supervised information effectively reduce the problem of information redundancy in the dictionary and the decline in sample label accuracy caused by the lack of prior knowledge.

[0046] Figure 1 This paper demonstrates the online classification process of a semi-supervised subspace representation model for large-scale high-dimensional datasets. The proposed semi-supervised subspace representation model first constructs a small-scale dictionary using partially known sample label information, and then learns the dictionary representation based on sparse constraints for the remaining data samples. For each data sample, after obtaining the coefficient representation vector, the label information of the data sample is obtained based on the vector element values ​​and the label information of the corresponding dictionary sample. Each data sample can be classified in real time through the above steps without affecting each other.

[0047] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An online data classification method based on a semi-supervised subspace representation model, characterized in that, include: Step 1: Preprocess the given image dataset; The image dataset contains some samples with known labels, which are denoted as labeled samples. Step 2: For the preprocessed image dataset, select a subset of samples from the labeled samples of each category as a sub-dictionary for that category; By merging the sub-dictionaries corresponding to each category, a small-scale dictionary is formed. The remaining samples, excluding those used to construct a small-scale dictionary, are used as samples to be classified. Step 3: For the samples to be classified, construct a sparse dictionary representation model based on semi-supervised information; Step 4: Use an alternating iterative optimization algorithm based on the augmented Lagrange multiplier method to solve the sparse dictionary representation model and obtain the optimal coefficient representation vector; Step 5: For the optimal coefficient representation vector learned by the sample to be classified, the correlation analysis is performed by combining it with the correspondence with the labeled samples in the small-scale dictionary, and the predicted label of the sample to be classified is obtained according to the label of the labeled samples in the small-scale dictionary, so as to achieve online classification.

2. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, Preprocessing a given image dataset includes: First, each sample is transformed into a vector form using feature extraction methods; then, principal component analysis is used to initially remove noise information; finally, normalization methods are used to process the samples.

3. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, The sparse dictionary representation model based on semi-supervised information is as follows: ； in, For the first One sample to be classified; Indicates the first A vector of coefficients to be optimized; For small-scale dictionaries; Represents a column vector whose elements are all 1s; express Norm; express Norm; superscript T indicates transpose.

4. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, By transforming the formula of the objective function of the sparse dictionary representation model based on semi-supervised information, we can obtain the following equivalent problem: ； A column vector was introduced. The coefficients represent vectors Perform equivalent substitution, that is ; Therefore, the above problem can be equivalently transformed into the following problem: ； Constructing an augmented Lagrangian function transforms the constrained optimization problem into one involving column vectors. Unconstrained optimization problem: ； in, Represents the Lagrange multipliers; This is the penalty parameter.

5. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, Alternating Iterative Optimization Algorithm Based on Augmented Lagrange Multiplier Method: When the coefficients represent vectors Once fixed, update the intermediate variables. Then the unconstrained optimization problem can be further simplified to the following problem: ； about Differentiate and set to zero, then The optimal solution can be obtained by the following formula: 。 6. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, The alternating iterative optimization algorithm based on the augmented Lagrange multiplier method further includes: When intermediate variables After fixing, update the coefficient representation vector. Then the unconstrained optimization problem can be further simplified to: ； Introducing vectors Its definition is: ； but The analytical solution can be obtained by the following formula: ； in, This represents the final coefficient representation vector. In the optimal support set The value of ; Describe a vector whose set is the optimal support set. The elements in the vector are used as the indices. In addition to the elements having values, All other elements in the sequence are 0; The coefficient representation vector In the optimal support complement The value of is set to 0 here; For threshold The soft threshold projection operator is defined. , The superscript + indicates that the operator performs a positive projection operation; optimal support set ; This indicates that a hard threshold truncation operator is used first. Preservation vector Previous Find the maximum value, set the remaining elements to 0, and then take the indices of the non-zero elements in the resulting vector; this gives the optimal support set. ; Therefore, the coefficient representation vector can be obtained. The optimal solution in the current iteration step, i.e., by Given the non-zero part, by Give the complete vector of the zero-value portion.

7. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, For vectors Each element of is calculated using the following formula: ； in, Indicates The first element in the result vector after operator processing Each element value; For vectors of Each element value; A soft threshold, used for... Perform a contraction operation; It is an orthographic projection operator; This means "defined as"; The maximum index value is defined as follows: ； in, Representing vectors of Each element value.

8. The online data classification method based on a semi-supervised subspace representation model according to claim 1, characterized in that, Based on the atomic positions corresponding to the non-zero elements of the superiority coefficient representation vector in the small-scale dictionary, the label information corresponding to these atoms is extracted; by statistically analyzing the categories corresponding to these non-zero elements, the category with the strongest correlation is determined as the predicted label of the sample to be classified.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes a computer program, it implements the online data classification method based on a semi-supervised subspace representation model as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When a computer program is executed by a processor, it implements the online data classification method based on a semi-supervised subspace representation model as described in any one of claims 1-8.

Images