Deep multi-label classification method based on graph structure

Abstract

The invention discloses a deep multi-label classification method based on a graph structure, which can perform efficient compression and feature reconstruction on input data through a deep network layer under the condition that noise exists in original data, break through redundant information interference of the original data, accurately capture essential feature representation contained in the data, and perform classification on the basis of the compressed feature data. And constructing a graph Laplacian matrix capable of completely retaining key node associated information, and finally performing label prediction classification on the obtained graph Laplacian matrix. The method can be applied to a classification experiment of dimension reduction data, the lowest rank expression of a graph structure is finally obtained, interference of data noise on the classification process is better resisted, and therefore the classification performance is improved.

Description

Technical Field

[0001] This invention relates to the field of graph structure data processing and dimensionality reduction technology, specifically a graph structure-based deep multi-label classification method that combines graph structure learning and subspace learning. Background Technology

[0002] Currently, in the field of machine learning, data is typically characterized by high dimensionality and low value density, while also containing a large amount of noisy or missing data. This poses a significant challenge to the processing of graph-structured data. To overcome these problems, many classic and effective data dimensionality reduction methods have been proposed, such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Local Linear Preserving Projection (LPP).

[0003] Although existing classic data dimensionality reduction methods still achieve good results, they all ignore the limited ability to filter redundant information in real-world applications, making it difficult to effectively remove interfering components. Furthermore, they neglect the potential structural relationships between data and fail to utilize the global structural information of the data to improve model robustness.

[0004] Therefore, there is an urgent need for a deep multi-label classification method that can effectively resist data noise interference and fully explore the essential features and structural relationships of data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application proposes a graph-based deep multi-label classification method. This method uses a hybrid network to efficiently compress and reconstruct the input data, accurately capturing the essential feature representations contained in the data. Based on the compressed feature data, a graph Laplacian matrix that can completely preserve the association information of key nodes is constructed. Finally, the obtained graph Laplacian matrix is ​​used for label prediction and classification.

[0006] The technical solution adopted in this invention is as follows: A deep multi-label classification method based on graph structure includes the following steps: Step 1: Collect the data to be processed and form a dataset. ; Step 2: Use deep network layers to process the dataset The dataset is obtained by compression. ,based on Constructing an adjacency matrix S Laplace matrix L = D - S,D It is a diagonal matrix; Step 3: Introduce the classifier W According to the Laplace matrix L and classifierW Construct a classification method: Where Y represents the true category label, and It is a regularization parameter. Denotes the square of the F-norm; Based on the construction of the Lagrange function, the variables in the above formula are updated iteratively. Step 4: Find the optimal solution when the loop iteration stops. W And X, based on the optimal solution W X, classify the data, and output the classification result labels.

[0007] Furthermore, the deep network layer is constructed using convolutional layers and graph neural networks.

[0008] Furthermore, based on the compressed dataset Construct an adjacency matrix using cosine similarity. S .

[0009] Furthermore, the method for updating the variables in step 3 is as follows: Step 1, fix other variables and only consider W Variables, about W The derivative of the function and set it to 0 is expressed as: ;in, It is an identity matrix; Step 2, let x represent the input of the deep network layer, and x' represent the reconstructed value of x. The objective function for data reconstruction is: ;in, Hyperparameters for controlling the trade-off between feature compression and other components; Update the parameter set of deep network layers using gradient descent. : ;in, Indicates the learning rate. express The gradient.

[0010] Further, choose Norm loss .

[0011] Furthermore, set an iteration termination condition: until the objective function value is less than... Or the number of iterations is greater than 200.

[0012] A terminal, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the above-described graph-based deep multi-label classification method when running the executable instructions.

[0013] A computer-readable storage medium storing an executable program, wherein the executable program, when executed by a processor, implements the aforementioned graph-based deep multi-label classification method.

[0014] The beneficial effects of this invention are as follows: This invention mainly considers to efficiently compress and reconstruct the input data through deep network layers, accurately capture the essential feature representation contained in the data, and construct a graph Laplacian matrix that can completely retain the association information of key nodes based on the compressed feature data. Finally, the obtained graph Laplacian matrix is ​​used for label prediction and classification.

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0017] Combined with appendix Figure 1 The deep multi-label classification method based on graph structure proposed in this invention includes the following steps: Step 1: Collect data As the required dataset, for each data point d and n represent the dimension and number of data, respectively.

[0018] Step 2: Transfer the dataset After passing through deep network layers, it becomes Where r represents the dimension of the data after compression through the deep network layer, and then utilizes... Construct an adjacency matrix using cosine similarity. Laplace matrix L = D - S ,in, It is a symmetric positive definite matrix. It is a diagonal matrix, and the diagonal elements are the sum of the weights of the nodes.

[0019] In this embodiment, the deep network layer can be constructed using convolutional layers and graph neural networks.

[0020] Step 3: Introduce the classifier Where q represents the dimension of the label set, and a classification method is constructed based on the constructed Laplacian matrix L and the classifier W: The first term measures the difference between the predicted class probability and the true class label, Y. This represents the true category label, where q and n represent the dimension and number of the label set, respectively. and It is a regularization parameter. This represents the square of the F-norm.

[0021] Based on the construction of the Lagrange function, the variables are updated and solved: Step 1, fixing other variables and considering only the variable W, we take the derivative of the function with respect to W and set it to 0. The calculation yields: ;in, It is an identity matrix.

[0022] Step 2, let x represent the input of the deep network layer, and x' represent the reconstructed value of x. The objective function for data reconstruction is: ;in, Hyperparameters are used to control the trade-offs between feature compression and other components. Furthermore, we choose... Norm loss Gradient descent is used to update the parameter set of deep network layers. : ;in, Indicates the learning rate. express The gradient.

[0023] Step 4: Iterate through the solution process in Step 3 until the objective function value is less than... Alternatively, the number of iterations may exceed 200. If the above conditions are met, the loop exits; otherwise, the loop continues. When the loop exits, it means that the optimal solutions W and X have been found. Based on the optimal solutions W and X, the dimensionality-reduced data can be further used for conventional classification, outputting classification result labels. For example, the dimensionality-reduced data of this invention can be used for classification to mine hidden information in the data.

[0024] A terminal, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the above-described graph-based deep multi-label classification method when running the executable instructions.

[0025] A computer-readable storage medium storing an executable program, wherein the executable program, when executed by a processor, implements the aforementioned graph-based deep multi-label classification method.

[0026] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A deep multi-label classification method based on graph structure, characterized in that, Includes the following steps: Step 1: Collect the data to be processed and form a dataset. ; Step 2: Use deep network layers to process the dataset The dataset is obtained by compression. ,based on Constructing an adjacency matrix S ; Laplace matrix L = D - S,D It is a diagonal matrix; Step 3: Introduce the classifier W According to the Laplace matrix L and classifier W Construct a classification method: Where Y represents the true category label, and It is a regularization parameter. Denotes the square of the F-norm; Based on the construction of the Lagrange function, the variables in the above formula are updated iteratively. Step 4: Find the optimal solution when the loop iteration stops. W And X, based on the optimal solution W X, classify the data, and output the classification result labels.

2. The graph-based deep multi-label classification method according to claim 1, characterized in that, The deep network layers are constructed using convolutional layers and graph neural networks.

3. The graph-based deep multi-label classification method according to claim 1, characterized in that, Based on the compressed dataset Construct an adjacency matrix using cosine similarity. S .

4. The graph-based deep multi-label classification method according to claim 1, characterized in that, The method for updating the variables in step 3 is as follows: Step 1, fix other variables and only consider W Variables, about W The derivative of the function and set it to 0 is expressed as: ;in, It is an identity matrix; Step 2, let x represent the input of the deep network layer, and x' represent the reconstructed value of x. The objective function for data reconstruction is: ;in, Hyperparameters for controlling the trade-off between feature compression and other components; Update the parameter set of deep network layers using gradient descent. : ;in, Indicates the learning rate. express The gradient.

5. The graph-based deep multi-label classification method according to claim 4, characterized in that, choose Norm loss .

6. The graph-based deep multi-label classification method according to claim 4, characterized in that, Set the iteration termination condition: until the objective function value is less than... Or the number of iterations is greater than 200.

7. A terminal, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the graph-based deep multi-label classification method of claim 1 when running the executable instructions.

8. A computer-readable storage medium, characterized in that, The readable storage medium stores an executable program, which, when executed by a processor, implements the graph-based deep multi-label classification method of claim 1.

Images