CN113255798A - Classification model training method, device, equipment and medium - Google Patents
Classification model training method, device, equipment and medium Download PDFInfo
- Publication number
- CN113255798A CN113255798A CN202110613729.6A CN202110613729A CN113255798A CN 113255798 A CN113255798 A CN 113255798A CN 202110613729 A CN202110613729 A CN 202110613729A CN 113255798 A CN113255798 A CN 113255798A
- Authority
- CN
- China
- Prior art keywords
- training
- vertex
- graph
- classification model
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012549 training Methods 0.000 title claims abstract description 247
- 238000013145 classification model Methods 0.000 title claims abstract description 102
- 238000000034 method Methods 0.000 title claims abstract description 94
- 239000011159 matrix material Substances 0.000 claims abstract description 158
- 238000013528 artificial neural network Methods 0.000 claims abstract description 107
- 238000010586 diagram Methods 0.000 claims abstract description 70
- 230000008569 process Effects 0.000 claims abstract description 56
- 230000009466 transformation Effects 0.000 claims description 31
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000004590 computer program Methods 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 8
- 238000003860 storage Methods 0.000 claims description 8
- 230000007246 mechanism Effects 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 4
- 239000010410 layer Substances 0.000 description 92
- 230000006870 function Effects 0.000 description 12
- 239000000126 substance Substances 0.000 description 7
- 239000013598 vector Substances 0.000 description 7
- 238000004891 communication Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 208000035473 Communicable disease Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 102000004169 proteins and genes Human genes 0.000 description 1
- 108090000623 proteins and genes Proteins 0.000 description 1
- 239000004576 sand Substances 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000003892 spreading Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application discloses a classification model training method, a classification model training device, classification model training equipment and a classification model training medium, wherein the classification model training device comprises the following steps: constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information of each vertex of the graph data set; inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher diagram wavelet neural network in the classification model for supervised training, and determining corresponding supervised training loss in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in the classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain the trained classification model. Thus, the classification accuracy of the classification model can be improved.
Description
Technical Field
The present application relates to the field of classifier technologies, and in particular, to a classification model training method, apparatus, device, and medium.
Background
With the rapid development of information technologies such as cloud computing, internet of things, mobile communication, and intelligent terminals, novel applications represented by social networks, communities, and blogs are widely used. These applications are constantly generating large amounts of data that facilitate graphical modeling analysis. Wherein the vertices represent individuals or groups and the connecting edges represent the connections between them; the vertices are typically tagged with information representing the age, gender, location, hobbies, and religious beliefs of the modeled object, as well as many other possible characteristics. These features reflect individual behavioral preferences from various aspects, and ideally, each social network user is tagged with all of the tags associated with their own features. But this is not the case in reality. This is because, for the purpose of protecting personal privacy, more and more social network users are more cautious when sharing personal information, resulting in that the social network media can only collect part of the user's information. Therefore, it is important and urgent to predict the labels of the remaining users according to the label information of the known users. This problem is the vertex classification problem.
Currently, solving the vertex classification problem through a graph neural network has become a research hotspot. A graph neural network is typically composed of an input layer, one or more hidden layers, and an output layer. For example, referring to fig. 1, fig. 1 is a diagram of a neural network structure in the prior art, and fig. 1 shows a structure of a typical graph convolution neural network, which is composed of an Input layer (Input layer), two graph convolution layers (Gconv layers), and an Output layer (Output layer). Wherein the input layer readsn*dThe vertex feature matrix of the dimension, the graph convolution layer extracts the feature of the vertex feature matrix, the vertex feature matrix is transferred to the next graph convolution layer after being transformed by a nonlinear activation function such as ReLu, finally, an output layer, namely a task layer, completes specific tasks such as vertex classification, clustering and the like, and what is shown in figure 1 is a vertex feature matrixAnd the vertex classification task layer outputs the class label of each vertex. Currently, how to improve the classification accuracy is a problem to be solved.
Disclosure of Invention
In view of this, an object of the present application is to provide a method, an apparatus, a device and a medium for training a classification model, which can improve the classification accuracy of the classification model. The specific scheme is as follows:
in a first aspect, the present application discloses a classification model training method, including:
constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher graph wavelet neural network in a classification model for supervised training, and determining corresponding supervised training loss in the training process;
inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process;
determining a target training loss based on the supervised training loss and the unsupervised training loss;
and when the target training loss is converged, outputting the current classification model to obtain a trained classification model.
Optionally, the determining a corresponding supervised training loss in the training process includes:
in the training process, determining corresponding supervised training loss based on a first vertex label prediction result of a Teacher diagram wavelet neural network and the vertex label matrix;
correspondingly, the determining the corresponding unsupervised training loss in the training process includes:
and in the training process, determining corresponding unsupervised training loss based on a second vertex label prediction result and the first vertex label prediction result of the Student graph wavelet neural network.
Optionally, the method further includes:
in the training process, updating the vertex label matrix by using the first vertex label prediction result;
and when the target training loss is converged, outputting the current vertex label matrix to obtain the class prediction result of each vertex without class labels.
Optionally, the method further includes:
calculating a graph wavelet transform basis for the graph data set and a graph wavelet inverse transform basis using a chebyshev polynomial;
correspondingly, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network perform graph convolution operation based on the diagram wavelet transformation basis and the diagram wavelet inverse transformation basis in the training process.
Optionally, the method further includes:
obtaining a calculation formula of the image wavelet transformation base;
wherein the calculation formula is a formula defined based on spectrum theory.
Optionally, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network both include an input layer, a plurality of diagram convolution layers, and an output layer;
the graph volume layer is used for sequentially carrying out feature transformation and graph volume operation processing on input data of the layer in a training process.
Optionally, the method further includes:
in the training process, determining the convolution kernel of the corresponding graph convolution layer in the Student graph wavelet neural network by utilizing the convolution kernel of the graph convolution layer obtained by training the Teacher graph wavelet neural network based on an attention mechanism.
In a second aspect, the present application discloses a classification model training apparatus, comprising:
the training data construction module is used for constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
the classification model training module is used for inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher diagram wavelet neural network in a classification model for supervised training and determining corresponding supervised training loss in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain a trained classification model.
In a third aspect, the present application discloses an electronic device, comprising:
a memory for storing a computer program;
a processor for executing the computer program to implement the aforementioned classification model training method.
In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program which, when executed by a processor, implements the aforementioned classification model training method.
Therefore, the vertex feature matrix, the adjacency matrix and the vertex label matrix are constructed on the basis of the graph data set; the vertex label matrix comprises label information of each vertex of the graph data set, and then the vertex feature matrix, the adjacency matrix and the vertex label matrix are input into a Teacher graph wavelet neural network in a classification model for supervised training, and corresponding supervised training loss is determined in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain a trained classification model. Therefore, the vertex characteristic matrix and the adjacency matrix of the graph data set are input into the graph neural network for training, the graph topological structure and the vertex characteristic are utilized, supervised training and unsupervised training are utilized during training, the advantages of the supervised training and the unsupervised training are fully played, and the classification accuracy of the classification model can be improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a diagram of a prior art neural network architecture;
FIG. 2 is a flow chart of a classification model training method disclosed herein;
FIG. 3 is a flow chart of a particular classification model training method disclosed herein;
FIG. 4 is a diagram of a classification model architecture disclosed herein;
FIG. 5 is a diagram of a particular classification model architecture disclosed herein;
FIG. 6 is a flow chart of a particular classification model training method disclosed herein;
FIG. 7 is a schematic structural diagram of a classification model training apparatus according to the present disclosure;
fig. 8 is a block diagram of an electronic device disclosed in the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
Referring to fig. 2, an embodiment of the present application discloses a classification model training method, including:
step S11: constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
wherein the tag information represents a corresponding category tag or no category tag.
In a specific embodiment, a graph data set is assumed to be,VA set of vertices is represented that is,Vclustering into a small set of vertices with class labelsAnd most class label-free vertex setsTwo parts, and satisfy,;ERepresenting a set of connected edges. In addition to the label or labels,Geach vertex of (1)vAll havedA feature of all the vertexes constitutingVertex feature matrix of dimension, denotedX。GIs denoted as an adjacency matrixAElements ofRepresenting verticesiAndjthe weight of the connecting edge therebetween. Vertex set from existing labelsConstruction ofVertex label matrix of dimensionYWherein, in the step (A),showing the number of all the vertices in the graph,Cnumber of label classes, matrix elements, representing all verticesRepresenting verticesiWhether the class label of (1) isWhen it is at the vertexiWhen the category label exists, the corresponding second label is putjThe column elements are 1 and the remaining column elements are 0. Namely, the method comprises the following steps:
when the vertex is upiAnd if the label is not the category label, setting each column element corresponding to the row to be 0.
Step S12: and inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher graph wavelet neural network in a classification model for supervised training, and determining corresponding supervised training loss in the training process.
Step S13: and inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process.
In a specific implementation mode, in a training process, determining corresponding supervised training loss based on a first vertex label prediction result of a Teacher diagram wavelet neural network and a vertex label matrix; and determining corresponding unsupervised training loss based on the second vertex label prediction result and the first vertex label prediction result of the Student graph wavelet neural network.
Specifically, the first vertex label prediction result is compared with the vertex label matrix to calculate supervised training loss, and the second vertex label prediction result is compared with the first vertex label prediction result to calculate unsupervised learning loss.
Step S14: determining a target training loss based on the supervised training loss and the unsupervised training loss.
In a specific embodiment, the target training loss is calculated as follows:
wherein the content of the first and second substances,indicating a loss of supervised training,indicating a loss of unsupervised training,is a constant used to adjust the proportion of the unsupervised training loss in the target loss.Representing the result of the prediction of the first vertex label,representing the second vertex label prediction.
Wherein the content of the first and second substances,andare all made ofA matrix of dimensions, and,orEach column vector ofIndicating that all vertices belong to a classjProbability of, i.e. it isEach element representing a vertexiBelong to the categoryThe probability of (c).
It should be noted that, in the embodiments of the present application, the output layers of the Teacher diagram wavelet neural network and the Student diagram wavelet neural network may be defined as;
in order to map the wavelet transform basis,in order to map the basis of the inverse wavelet transform,is shown asLThe layer map is convolved with the convolution kernel matrix of the layer,is shown asLThe layer vertex feature transformation result, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network compriseLThe layer map is rolled up layer by layer.
Moreover, the supervised training loss function calculates the difference degree of the actual label probability distribution and the predicted label probability distribution of the vertex based on the cross entropy principle; unsupervised training loss function computationAndthe sum of the squares of the differences between the same coordinate elements.
Thus, when the whole network training is finished, the output results of the two networksAndthe agreement or difference is negligible. Output of wavelet neural network with Teacher diagramIs the output of the entire network model.
In the training process, the vertex label matrix is updated by using the first vertex label prediction result, specifically, for the vertex without class label, that is, for the vertex without class labelAnd updating the vertex label matrix by taking the class with the highest probability in the first vertex label prediction result as the latest class of the vertex.
Step S15: and when the target training loss is converged, outputting the current classification model to obtain a trained classification model.
And when the target training loss is converged, outputting the current vertex label matrix to obtain the class prediction result of each vertex without class labels.
In a specific embodiment, when the target training loss reaches a preset threshold or the number of iterations reaches a specified iteration maximum, the target training loss converges, and the training ends. The preset threshold value is usually a small value, and at this time, for a vertex without a class label, the class to which the vertex label belongs is obtained according to the current vertex label matrix.
That is, the present application fuses the prediction of unlabeled vertices into the training process: in the training process, the vertex label matrix is updated according to the training result of each time, and any one category label without a label vertex can be obtained after the training is finished.
In a specific embodiment, the network parameters of each layer of the graph wavelet neural network may be initialized according to a specific strategy, such as normal distribution random initialization, Xavier initialization or He initialization. In the training process, network parameters of each layer of the graph wavelet neural network can be modified and updated according to specific strategies such as SGD (i.e., Stochastic Gradient Descent), MGD (i.e., Momentum Gradient Descent), Nesterov Momentum, AdaGrad (i.e., Adaptive Gradient algorithm), RMSprop (i.e., Root Mean Square Gradient Descent algorithm), Adam (i.e., Adaptive Gradient Estimation) or BGD (i.e., Batch Gradient Descent) to optimize the loss function value.
Therefore, the vertex feature matrix, the adjacency matrix and the vertex label matrix are constructed on the basis of the graph data set; the vertex label matrix comprises label information of each vertex of the graph data set, and then the vertex feature matrix, the adjacency matrix and the vertex label matrix are input into a Teacher graph wavelet neural network in a classification model for supervised training, and corresponding supervised training loss is determined in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain a trained classification model. Therefore, the vertex characteristic matrix and the adjacency matrix of the graph data set are input into the graph neural network for training, the graph topological structure and the vertex characteristic are utilized, supervised training and unsupervised training are utilized during training, the advantages of the supervised training and the unsupervised training are fully played, and the classification accuracy of the classification model can be improved.
Referring to fig. 3, the embodiment of the present application discloses a specific classification model training method, including:
step S21: and obtaining a calculation formula of the wavelet transform base of the image.
Wherein the calculation formula is a formula defined based on spectrum theory.
It should be noted that the graph convolution operation defined by the fourier transform has poor locality in the vertex domain, and the basis of the graph wavelet transform is defined by using the spectrum theory, so that the locality of the graph convolution calculation is ensured.
Step S22: a graph wavelet transform basis for the graph data set and a graph wavelet inverse transform basis are calculated using chebyshev polynomials.
In a specific embodiment, the calculation formula of the graph wavelet transformation base isWherein, in the step (A),representing a set of slave graph dataGThe extracted graph wavelet transform basis in (1),Urepresenting a data set from a mapGLaplacian matrix ofCarrying out characteristic decomposition to obtain a matrix consisting of characteristic vectors;Dis a diagonal matrix with major diagonalsnEach element representsnThe degrees of each vertex and the remaining elements are zero.Is scaled torAnd scaling the matrix of, Is a pair of drawingsGPerforming characteristic decomposition on the Laplace matrix to obtain characteristic values; inverse transformation base of image waveletCan be prepared by mixingIn (1)Is replaced byAnd (6) obtaining. Because the characteristic decomposition calculation cost of the matrix is large, in order to avoid the cost, the Chebyshev polynomial is utilizedAnd is and,to approximate the computation of the graph wavelet transform basis, and the graph wavelet inverse transform basis.
Correspondingly, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network perform graph convolution operation based on the diagram wavelet transformation basis and the diagram wavelet inverse transformation basis in the training process.
It should be noted that, in the prior art, the graph fourier transform is inefficient during the graph convolution operation because the eigenvector matrix of the laplacian matrix is dense, whereas the graph convolution operation is performed based on the graph wavelet transform basis and the graph wavelet inverse transform basis, which are sparse, in the present embodiment, so that the operation efficiency of the graph convolution operation can be improved.
Step S23: constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset, the label information representing a corresponding class label or no class label.
Step S24: and inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher graph wavelet neural network in a classification model for supervised training, and determining corresponding supervised training loss in the training process.
Step S25: and inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process.
In a specific implementation mode, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network respectively comprise an input layer, a plurality of diagram convolution layers and an output layer; the graph volume layer is used for sequentially carrying out feature transformation and graph volume operation processing on input data of the layer in a training process. In particular, 1 input layer may be included,a map volume layer, and an output layer.
That is, in the embodiment of the present application, the graph convolution layer performs feature transformation on the input data of the layer first, and then performs graph convolution operation processing, so that the graph convolution layer is divided into two processing stages, namely, the feature transformation and the graph convolution operation, which can reduce network parameters, thereby reducing the amount of model operation and improving the model training efficiency.
wherein the content of the first and second substances,andare respectively the firstlThe layer diagram hides input and output data of the layer, an;Is as followslThe layer is the feature transformation matrix to be trained,is as followslAs a result of the transformation of the layer characteristics,Trepresenting the transpose operation of the matrix.
It should be noted that, in the prior art, the graph convolution layer definition does not usually distinguish between the feature transformation and the convolution operation, and if the graph convolution layer is not divided into two processing stages, namely the feature transformation and the graph convolution operation, in combination with the graph wavelet transformation base in the embodiment of the present application. The graph convolution layer is defined by the following formula:
wherein the content of the first and second substances,Xa matrix of the characteristics of the vertices is represented,mthe ordinal number of the graph convolution layer is shown,Fis a graph convolution kernel matrix that is,his an activation function. In the graph convolution layer defined in the above manner, the number of parameters included isWhereinnIn the representation diagramThe number of the top points is,pthe vertex feature dimension representing the input of the layer,qrepresenting the vertex feature dimension of the layer output. In the embodiment of the present application, the feature transformation is stripped from the graph convolution operation, and the number of parameters of each graph convolution layer becomes。
In addition, in a specific embodiment, in the training process, the convolution kernel of the corresponding graph convolution layer in the Student graph wavelet neural network is determined by using the convolution kernel of the graph convolution layer obtained by training the Teacher graph wavelet neural network based on an attention mechanism.
Specifically, the classification model may include a Teacher wavelet neural network, a Student wavelet neural network, and an attention network connecting each pair of map convolutional layers of the Teacher wavelet neural network and the Student wavelet neural network.
In this specification, the following meansIs the firstlThe graph of the layer is convolved with the kernel matrix, which is a diagonal matrix. From the point of view of the signal processing,diagonal elementsThe frequency can be regarded as a graph, and represents the importance of the feature vector corresponding to the frequency. Recording the Teacher diagram wavelet neural network and the Student diagram wavelet neural networklThe convolution kernel matrices of the layers are respectivelyT l AndS l convolution kernels of the layer by the Teacher diagram wavelet neural network respectivelyAnd convolution kernel of the layer of the Student graph wavelet neural networkDiagonalization is obtained, both arenA column vector of dimensions.
In this embodiment, attention transfer (attention transfer) may be performed based on an attention mechanism: the Teacher diagram wavelet neural network transfers the learned convolution kernel of each layer to the corresponding layer of the Student diagram wavelet neural network, namely, the Student diagram wavelet neural network learns to the Teacher diagram wavelet neural network, and the performance of the whole network is improved. Specifically, a single-layer feedforward neural network can be designed, and the input layer of the feedforward neural network is responsible for reading the Teacher diagram wavelet neural network and the Student diagram wavelet neural networklLayer convolution kernelAnd(ii) a With hidden layers for implementing attention functionsIn order to obtain attention weight between two vectors;
Further, attention is weighted by softmax functionThe normalization is performed to obtain a normalized attention weight of:
Wherein the content of the first and second substances,to representTo (1) aiThe number of the components is such that,to representTo (1) aiAnd (4) a component. Then there are:
wherein the content of the first and second substances,represents the first time that the Student graph wavelet neural network learns from the Teacher graph wavelet neural networklA layer convolution kernel.
It should be noted that the addition of the attention mechanism promotes the Student graph wavelet neural network to quickly utilize the knowledge mastered by the Teacher graph wavelet neural network, and the training speed is increased.
Step S26: determining a target training loss based on the supervised training loss and the unsupervised training loss.
Step S27: and when the target training loss is converged, outputting the current classification model to obtain a trained classification model.
For example, referring to fig. 4, fig. 4 is a diagram of a classification model structure disclosed in the embodiment of the present application. Teacher diagram wavelet neural network GWNTStudent graph wavelet neural network GWNSFurther, referring to fig. 5, fig. 5 is a specific classification model structure diagram disclosed in the embodiment of the present application. The classification model is composed of a Teacher diagram wavelet neural network GWNTAnd, Student graph wavelet neural network GWNSAnd an attention network connecting each pair of map convolutional layers of the two networks. GWNTSupervised learning is carried out according to the graph vertexes with the labels, and the prediction accuracy is high; GWNSIn GWNTUnder the guidance of (using the prediction results of) using unlabeled graph vertex marchingAnd performing unsupervised learning so as to improve the prediction accuracy and obtain a better vertex classification model. Attention network for GWNTTransferring the learned "knowledge" of each layer, i.e., the convolution kernel, to the GWNSCorresponding layer, i.e. GWNSTowards GWNTAnd (5) learning. GWNTAnd GWNSEach comprises 1 input layer,LOne map convolutional layer and 1 output layer. The input layer is mainly used for reading the data of the graph to be classified and comprises an adjacent matrix representing the topological structure of the graphAAnd vertex feature matrixX. In the graph convolution layer, the graph convolution operation is decomposed into two stages of feature transformation and graph convolution. The output layer is used for outputting the prediction result.
In addition, in the whole classification model, the network parameters of each layer comprise feature transformation matrixes(including the Teacher diagram wavelet neural networkAnd Student graph wavelet neural network) Convolution kernel (convolution kernel)And convolution kernel) And further updating the convolution kernel matrix with the convolution kernelAnd attention network parameters. In the initialization stage, the network parameters are initialized, and in the training process, the network parameters are updated.
For example, referring to fig. 6, the embodiment of the present application discloses a toolMethod for training classification model of body, flow chart, for given image data setGWith its adjacent matrixA、Vertex feature matrixXAnd vertex label matrixYAnd (3) as input, sending the input into a network for forward propagation, calculating the prediction results of all vertexes belonging to each category, updating a prediction result matrix, calculating the loss of a supervised learning part and the loss of an unsupervised learning part so as to obtain a total network loss function value, updating network parameters of each layer according to a certain strategy, and finishing training until the network error reaches a specified smaller value or the iteration times reaches a specified maximum value.
For example, a method according to an embodiment of the present application trains a classification model using a scientific paper set and predicts class labels of unlabeled scientific papers.
(1) Downloading a quotation network data set Citeseer, which comprises 3312 scientific papers divided into six categories and the quotation relation among 4732 papers; bag-of-words (bag of words) model is utilized to construct feature vector of each paperxFeature vectors of all documents form a feature matrixX. Constructing an adjacency matrix according to reference relations among papersA. The goal is to classify each document, randomly draw 20 instances per category as labeled data, 1000 instances as test data, and the rest as unlabeled data; building a vertex label matrixY。
(2) Defining a network structure: graph convolution layers, output layers, and network loss functions are defined based on the foregoing disclosure.
(3) The graph wavelet transform basis and the basis of the graph wavelet inverse transform are calculated using a chebyshev polynomial approximation.
(4) And initializing the network parameters according to a regularization initialization method.
(5) To be provided withA,XAndYand as network input, sending the data to the network for forward propagation. Wherein, the Teacher diagram wavelet neural network GWNTTo be provided withA,XAndYas input, the Student graph wavelet neural network GWNSTo be provided withAAndXas an input. Each network is based on the definition of the graph convolution layerDefining, combining the input feature matrixes of the layers, and calculating the output feature matrix of each layer; according to the definition of the output layer, the prediction results of all the vertexes belonging to each category are calculatedZ T OrZ S Calculating a supervised learning loss function value and an unsupervised learning function loss value according to the defined network loss function, and further obtaining a loss function value of the whole network; for the non-label vertex, the class with the highest probability is taken as the latest class of the vertex, and the vertex label matrix is updatedY。
(6) According to the optimization method, the gradient of the loss function relative to the network parameters is calculated and propagated backwards so as to optimize the network parameters, and the training is finished until the network prediction error reaches a specified smaller value or the iteration times reach a specified maximum value. In this case, for vertices without class labels, the vertex label matrix can be based onYGet the category to which it should belong.
Of course, the present application is not limited to the scientific citation classification problem listed in the examples, but may also be applied to any classification problem that facilitates the data represented by graph modeling, such as proteins, graphic images, and the like, and to research the laws of spreading of infectious diseases and thought points and the like in social networks over time, how groups in social networks form communities around specific interests or membership relationships, and the strength of community connections; according to the rule of 'people group by group', the social network finds people with similar interests and suggests or recommends new links or connections to the people; the question-answering system directs questions to the most experienced person; the advertisement system displays advertisements, etc., to individuals who are most interested in and willing to accept advertisements on a particular topic.
Referring to fig. 7, an embodiment of the present application discloses a classification model training apparatus, including:
the training data construction module 11 is used for constructing a vertex feature matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
the classification model training module 12 is configured to input the vertex feature matrix, the adjacency matrix, and the vertex label matrix into a Teacher wavelet neural network in a classification model for supervised training, and determine a corresponding supervised training loss in a training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain a trained classification model.
Therefore, the vertex feature matrix, the adjacency matrix and the vertex label matrix are constructed on the basis of the graph data set; the vertex label matrix comprises label information of each vertex of the graph data set, and then the vertex feature matrix, the adjacency matrix and the vertex label matrix are input into a Teacher graph wavelet neural network in a classification model for supervised training, and corresponding supervised training loss is determined in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; and when the target training loss is converged, outputting the current classification model to obtain a trained classification model. Therefore, the vertex characteristic matrix and the adjacency matrix of the graph data set are input into the graph neural network for training, the graph topological structure and the vertex characteristic are utilized, supervised training and unsupervised training are utilized during training, the advantages of the supervised training and the unsupervised training are fully played, and the classification accuracy of the classification model can be improved.
The classification model training module 12 is specifically configured to determine, in a training process, a corresponding supervised training loss based on a first vertex label prediction result of a Teacher graph wavelet neural network and the vertex label matrix; and determining corresponding unsupervised training loss based on the second vertex label prediction result and the first vertex label prediction result of the Student graph wavelet neural network.
The classification model training module 12 is further configured to: in the training process, updating the vertex label matrix by using the first vertex label prediction result; and when the target training loss is converged, outputting the current vertex label matrix to obtain the class prediction result of each vertex without class labels.
The apparatus also includes a graph wavelet transform basis calculation module for calculating a graph wavelet transform basis of the graph data set and a graph wavelet inverse transform basis using a chebyshev polynomial; correspondingly, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network perform graph convolution operation based on the diagram wavelet transformation basis and the diagram wavelet inverse transformation basis in the training process.
The device also comprises an image wavelet transformation base formula acquisition module, a calculation module and a calculation module, wherein the image wavelet transformation base formula acquisition module is used for acquiring a calculation formula of the image wavelet transformation base; wherein the calculation formula is a formula defined based on spectrum theory.
In a specific implementation mode, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network respectively comprise an input layer, a plurality of diagram convolution layers and an output layer;
the graph volume layer is used for sequentially carrying out feature transformation and graph volume operation processing on input data of the layer in a training process.
The classification model training module 12 is further configured to determine, based on an attention mechanism, a convolution kernel of a corresponding graph convolution layer in the Student graph wavelet neural network by using the convolution kernel of the graph convolution layer obtained by training the Teacher graph wavelet neural network in a training process.
Referring to fig. 8, an embodiment of the present application discloses an electronic device 20, which includes a processor 21 and a memory 22; wherein, the memory 22 is used for saving computer programs; the processor 21 is configured to execute the computer program and the classification model training method disclosed in the foregoing embodiment.
For the specific process of the above classification model training method, reference may be made to the corresponding contents disclosed in the foregoing embodiments, and details are not repeated here.
The memory 22 is used as a carrier for resource storage, and may be a read-only memory, a random access memory, a magnetic disk or an optical disk, and the storage mode may be a transient storage mode or a permanent storage mode.
In addition, the electronic device 20 further includes a power supply 23, a communication interface 24, an input-output interface 25, and a communication bus 26; the power supply 23 is configured to provide an operating voltage for each hardware device on the server 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and an external device, and a communication protocol followed by the communication interface is any communication protocol applicable to the technical solution of the present application, and is not specifically limited herein; the input/output interface 25 is configured to obtain external input data or output data to the outside, and a specific interface type thereof may be selected according to a specific application requirement, which is not specifically limited herein.
Further, a computer-readable storage medium for storing a computer program is disclosed in an embodiment of the present application, where the computer program is executed by a processor to implement the classification model training method disclosed in the foregoing embodiment.
For the specific process of the above classification model training method, reference may be made to the corresponding contents disclosed in the foregoing embodiments, and details are not repeated here.
The embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same or similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
The above detailed description is given to a classification model training method, apparatus, device and medium provided by the present application, and a specific example is applied in the present application to explain the principle and the implementation of the present application, and the description of the above embodiment is only used to help understanding the method and the core idea of the present application; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.
Claims (10)
1. A classification model training method is characterized by comprising the following steps:
constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher graph wavelet neural network in a classification model for supervised training, and determining corresponding supervised training loss in the training process;
inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process;
determining a target training loss based on the supervised training loss and the unsupervised training loss;
when the target training loss is converged, outputting a current classification model to obtain a trained classification model;
the device comprises a Teacher diagram wavelet neural network, a Student diagram wavelet neural network and a terminal, wherein the Teacher diagram wavelet neural network and the Student diagram wavelet neural network respectively comprise an input layer, a plurality of diagram convolution layers and an output layer;
and, the method further comprises: in the training process, determining the convolution kernel of the corresponding graph convolution layer in the Student graph wavelet neural network by utilizing the convolution kernel of the graph convolution layer obtained by training the Teacher graph wavelet neural network based on an attention mechanism.
2. The method of classification model training according to claim 1, wherein the determining a corresponding supervised training loss during the training process comprises:
in the training process, determining corresponding supervised training loss based on a first vertex label prediction result of a Teacher diagram wavelet neural network and the vertex label matrix;
correspondingly, the determining the corresponding unsupervised training loss in the training process includes:
and in the training process, determining corresponding unsupervised training loss based on a second vertex label prediction result and the first vertex label prediction result of the Student graph wavelet neural network.
3. The classification model training method according to claim 2, further comprising:
in the training process, updating the vertex label matrix by using the first vertex label prediction result;
and when the target training loss is converged, outputting the current vertex label matrix to obtain the class prediction result of each vertex without class labels.
4. The classification model training method according to claim 1, further comprising:
calculating a graph wavelet transform basis for the graph data set and a graph wavelet inverse transform basis using a chebyshev polynomial;
correspondingly, the Teacher diagram wavelet neural network and the Student diagram wavelet neural network perform graph convolution operation based on the diagram wavelet transformation basis and the diagram wavelet inverse transformation basis in the training process.
5. The classification model training method according to claim 4, further comprising:
obtaining a calculation formula of the image wavelet transformation base;
wherein the calculation formula is a formula defined based on spectrum theory.
6. The method for training classification models according to any one of claims 1 to 5, wherein the graph convolution layer is used for sequentially performing feature transformation and graph convolution operation processing on input data of the layer in a training process.
7. A classification model training apparatus, comprising:
the training data construction module is used for constructing a vertex characteristic matrix, an adjacency matrix and a vertex label matrix based on the graph data set; wherein the vertex label matrix comprises label information for each vertex of the graph dataset;
the classification model training module is used for inputting the vertex characteristic matrix, the adjacency matrix and the vertex label matrix into a Teacher diagram wavelet neural network in a classification model for supervised training and determining corresponding supervised training loss in the training process; inputting the vertex characteristic matrix and the adjacency matrix into a Student graph wavelet neural network in a classification model for unsupervised training, and determining corresponding unsupervised training loss in the training process; determining a target training loss based on the supervised training loss and the unsupervised training loss; when the target training loss is converged, outputting a current classification model to obtain a trained classification model;
the device comprises a Teacher diagram wavelet neural network, a Student diagram wavelet neural network and a terminal, wherein the Teacher diagram wavelet neural network and the Student diagram wavelet neural network respectively comprise an input layer, a plurality of diagram convolution layers and an output layer;
and the classification model training module is further configured to: in the training process, determining the convolution kernel of the corresponding graph convolution layer in the Student graph wavelet neural network by utilizing the convolution kernel of the graph convolution layer obtained by training the Teacher graph wavelet neural network based on an attention mechanism.
8. The classification model training device according to claim 7, wherein the classification model training module is specifically configured to determine, during a training process, a corresponding supervised training loss based on a first vertex label prediction result of a Teacher diagram wavelet neural network and the vertex label matrix; and determining corresponding unsupervised training loss based on the second vertex label prediction result and the first vertex label prediction result of the Student graph wavelet neural network.
9. An electronic device, comprising:
a memory for storing a computer program;
a processor for executing the computer program to implement the classification model training method according to any one of claims 1 to 6.
10. A computer-readable storage medium for storing a computer program which, when executed by a processor, implements the classification model training method according to any one of claims 1 to 6.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110613729.6A CN113255798A (en) | 2021-06-02 | 2021-06-02 | Classification model training method, device, equipment and medium |
PCT/CN2021/121905 WO2022252458A1 (en) | 2021-06-02 | 2021-09-29 | Classification model training method and apparatus, device, and medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110613729.6A CN113255798A (en) | 2021-06-02 | 2021-06-02 | Classification model training method, device, equipment and medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113255798A true CN113255798A (en) | 2021-08-13 |
Family
ID=77186018
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110613729.6A Pending CN113255798A (en) | 2021-06-02 | 2021-06-02 | Classification model training method, device, equipment and medium |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN113255798A (en) |
WO (1) | WO2022252458A1 (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112364372A (en) * | 2020-10-27 | 2021-02-12 | 重庆大学 | Privacy protection method with supervision matrix completion |
CN114048816A (en) * | 2021-11-16 | 2022-02-15 | 中国人民解放军国防科技大学 | Method, device and equipment for sampling graph neural network data and storage medium |
CN114943324A (en) * | 2022-05-26 | 2022-08-26 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
CN115240037A (en) * | 2022-09-23 | 2022-10-25 | 卡奥斯工业智能研究院(青岛)有限公司 | Model training method, image processing method, device and storage medium |
WO2022252458A1 (en) * | 2021-06-02 | 2022-12-08 | 苏州浪潮智能科技有限公司 | Classification model training method and apparatus, device, and medium |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR102492318B1 (en) * | 2015-09-18 | 2023-01-26 | 삼성전자주식회사 | Model training method and apparatus, and data recognizing method |
CN111552803B (en) * | 2020-04-08 | 2023-03-24 | 西安工程大学 | Text classification method based on graph wavelet network model |
CN111639755B (en) * | 2020-06-07 | 2023-04-25 | 电子科技大学中山学院 | Network model training method and device, electronic equipment and storage medium |
CN112464057A (en) * | 2020-11-18 | 2021-03-09 | 苏州浪潮智能科技有限公司 | Network data classification method, device, equipment and readable storage medium |
CN113255798A (en) * | 2021-06-02 | 2021-08-13 | 苏州浪潮智能科技有限公司 | Classification model training method, device, equipment and medium |
-
2021
- 2021-06-02 CN CN202110613729.6A patent/CN113255798A/en active Pending
- 2021-09-29 WO PCT/CN2021/121905 patent/WO2022252458A1/en active Application Filing
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112364372A (en) * | 2020-10-27 | 2021-02-12 | 重庆大学 | Privacy protection method with supervision matrix completion |
WO2022252458A1 (en) * | 2021-06-02 | 2022-12-08 | 苏州浪潮智能科技有限公司 | Classification model training method and apparatus, device, and medium |
CN114048816A (en) * | 2021-11-16 | 2022-02-15 | 中国人民解放军国防科技大学 | Method, device and equipment for sampling graph neural network data and storage medium |
CN114048816B (en) * | 2021-11-16 | 2024-04-30 | 中国人民解放军国防科技大学 | Method, device, equipment and storage medium for sampling data of graph neural network |
CN114943324A (en) * | 2022-05-26 | 2022-08-26 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
CN114943324B (en) * | 2022-05-26 | 2023-10-13 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
CN115240037A (en) * | 2022-09-23 | 2022-10-25 | 卡奥斯工业智能研究院(青岛)有限公司 | Model training method, image processing method, device and storage medium |
Also Published As
Publication number | Publication date |
---|---|
WO2022252458A1 (en) | 2022-12-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023000574A1 (en) | Model training method, apparatus and device, and readable storage medium | |
CN110084296B (en) | Graph representation learning framework based on specific semantics and multi-label classification method thereof | |
Ozdemir et al. | Feature Engineering Made Easy: Identify unique features from your dataset in order to build powerful machine learning systems | |
US11227190B1 (en) | Graph neural network training methods and systems | |
CN113255798A (en) | Classification model training method, device, equipment and medium | |
US20180336472A1 (en) | Projection neural networks | |
CN112116599B (en) | Sputum smear tubercle bacillus semantic segmentation method and system based on weak supervised learning | |
CN114048331A (en) | Knowledge graph recommendation method and system based on improved KGAT model | |
CN110781407A (en) | User label generation method and device and computer readable storage medium | |
CN116010684A (en) | Article recommendation method, device and storage medium | |
US20220138502A1 (en) | Graph neural network training methods and systems | |
CN112085615A (en) | Method and device for training graph neural network | |
CN113392317A (en) | Label configuration method, device, equipment and storage medium | |
Liu et al. | Research of animals image semantic segmentation based on deep learning | |
US20240177006A1 (en) | Data processing method and apparatus, program product, computer device, and medium | |
CN116664719A (en) | Image redrawing model training method, image redrawing method and device | |
CN111695046A (en) | User portrait inference method and device based on spatio-temporal mobile data representation learning | |
Wang et al. | Feature subspace transfer for collaborative filtering | |
CN112131261A (en) | Community query method and device based on community network and computer equipment | |
Hu | Deep learning for ranking response surfaces with applications to optimal stopping problems | |
US11455512B1 (en) | Representing graph edges using neural networks | |
CN112836007B (en) | Relational element learning method based on contextualized attention network | |
CN112055038A (en) | Method for generating click rate estimation model and method for predicting click probability | |
CN111309923B (en) | Object vector determination method, model training method, device, equipment and storage medium | |
CN116992151A (en) | Online course recommendation method based on double-tower graph convolution neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210813 |
|
RJ01 | Rejection of invention patent application after publication |