Classification Method for Unsupervised Domain Adaptation Based on Graph Convolutional Network

Abstract

The present application relates to an unsupervised domain adaptation classification based on a graph convolutional network. The method includes: obtaining sample data in a source domain and sample data in a target domain as training data; respectively updating the graph connection relationships of samples in the two domains according to the similarity between the sample data in the source domain and the target domain; inputting the sample data in the source domain and the target domain into a domain adaptation network for training, where the domain adaptation network is an unsupervised domain adaptation network based on a graph convolutional network, and the domain adaptation network includes: a cross-domain feature extraction model, a source domain feature extraction model, a classification model, a domain adversarial discrimination model, and a class alignment model; training the domain adaptation network to continuously update and iterate the parameters in the domain adaptation network, and when the domain adaptation network reaches the convergence condition, obtaining a domain adaptation classification model; inputting the data to be classified into the domain adaptation classification model for classification to obtain the classification result of the data to be classified. The performance of the unsupervised domain adaptation model based on graph convolution is improved.

Description

Technical Field

[0001] This application relates to the technical field of deep learning, and particularly to an unsupervised domain adaptation classification method based on a graph convolutional network. Background Art

[0002] For an unsupervised domain adaptation task, that is, using the information of the source domain to assist in completing the task on the target domain, where the samples in the source domain are labeled or partially labeled, and the samples in the target domain are unlabeled. The main challenge of unsupervised domain adaptation is how to align the data distributions of the source domain and the target domain.

[0003] For the unsupervised domain adaptation task, general deep learning methods usually transform the samples in the source domain and the target domain into the same common space. For example, Zeng et al. designed a Maximum Mean Discrepancy (MMD) loss on the shared parameter layer in the network to reduce the distribution difference between the source domain and the target domain in the common space. Ganin et al. designed a domain discriminator to distinguish which domain each sample comes from, and proposed a Gradient Reversal Layer (GRL) to maximize the domain classification loss to reduce the distribution difference between domains. Ding et al. proposed an Adaptive Exploration (AE) method to solve the domain transfer problem of person re-identification by maximizing the distance between all pedestrian images and minimizing the distance between similar pedestrian images. Although deep learning methods have made some progress in reducing domain differences, the label rate in the source domain still affects the prediction results of the unsupervised domain adaptation task. The lower the label rate in the source domain, the worse the prediction results of the target domain.

[0004] With the introduction of graph neural networks, the Graph Convolutional Network (GCN) proposed by Kipf et al. has achieved ideal results in semi-supervised classification tasks. In the domain adaptation task, given a small amount of labeled source data, the graph convolutional network can usually build a well-performing classifier by propagating sample information in the source network. For example, Dai et al. combined the graph convolutional network and an adversarial domain adaptation model to reduce the distribution difference and make accurate label predictions.

[0005] Existing unsupervised domain adaptation classification methods based on graph convolutional networks focus on the common information between the two domains, without utilizing the specific information of the domains. Moreover, the category-level distribution alignment problem is not further concerned, which may lead to negative alignment of the distributions of the same class samples across domains and may be unfavorable for the tasks of the target domain, thus resulting in low performance of the trained graph convolutional-based unsupervised domain adaptation classification model. Summary of the Invention

[0006] Based on this, in view of the above technical problems, it is necessary to provide an unsupervised domain adaptation classification method based on a graph convolutional network that can improve the performance of an unsupervised domain adaptation classification model based on graph convolution for training.

[0007] An unsupervised domain adaptation classification method based on a graph convolutional network, the method comprising:

[0008] Obtain sample data in the source domain and sample data in the target domain as training data;

[0009] Update the graph connection relationships of samples in the two domains respectively according to the similarity between the sample data in the source domain and the target domain;

[0010] Input the sample data in the source domain and the target domain into a domain adaptation network for training, the domain adaptation network is an unsupervised domain adaptation network based on a graph convolutional network, and the domain adaptation network includes: a cross-domain feature extraction model, a source domain feature extraction model, a classification model, a domain adversarial discrimination model, and a class alignment model;

[0011] Train the domain adaptation network to continuously update and iterate the parameters in the domain adaptation network. When the domain adaptation network reaches the convergence condition, obtain a domain adaptation classification model;

[0012] Input the data to be classified into the domain adaptation classification model for classification to obtain the classification result of the data to be classified.

[0013] In one embodiment, the cross-domain feature extraction model extracts the sample features common to the source domain and the target domain, the source domain feature extraction model extracts the sample features specific to the source domain, the classification model calculates the classification loss value, the domain adversarial discrimination model calculates the domain feature alignment loss value, and the class alignment model calculates the class feature alignment loss value.

[0014] In one embodiment, the total loss value is the sum of the feature difference loss value, the classification loss value, the domain feature alignment loss value, and the class feature alignment loss value. Among them, the feature difference loss value is the feature difference obtained by inputting the sample data in the source domain into the cross-domain feature extraction model and the source domain feature extraction model, and the classification loss value is based on the sample data in the source domain input into the classification model.

[0015] In one embodiment, the construction method of the domain adaptation network includes:

[0016] Input the sample data in the source domain and the sample data in the target domain into the cross-domain feature extraction model to obtain the common embedded feature representation of the source domain and the target domain;

[0017] The sample data in the source domain is input into the source domain feature extraction model to obtain the specific embedded feature representation of the source domain;

[0018] Calculate the difference between the common embedded feature representation of the source domain and the specific embedded feature representation to construct a feature difference loss function;

[0019] The sample data in the target domain is input into the source domain feature extraction model to obtain the target domain embedded feature representation with the source domain style, and then combined with the common embedded feature representation of the target domain through the attention mechanism to form the embedded feature representation of the target domain; at the same time, the common embedded feature representation of the source domain is combined with the specific embedded feature representation of the source domain through the attention mechanism to form the embedded feature representation of the source domain;

[0020] The obtained embedded feature representations of the source domain and the target domain are input into the classification model. The part with class labels in the embedded feature representation of the source domain constructs a classification loss function, and the remaining part of the source domain without class labels and the embedded feature representation of the target domain generate pseudo-class labels corresponding to their feature representations;

[0021] The obtained common embedded feature representations of the source domain and the target domain are input into the domain adversarial discriminant model to construct a domain feature alignment loss function;

[0022] The sample data of the source domain and the sample data of the target domain are grouped according to the classes in the class labels and pseudo-class labels. At the same time, the embedded feature representations of the samples in different groups are input into the class alignment model to construct a class feature alignment loss function.

[0023] In one embodiment, the cross-domain feature extraction model is composed of a shared network of a two-layer graph convolutional neural network. The sample data of the source domain and the sample data of the target domain are both input into the shared network to obtain their common embedded feature representation;

[0024] The source domain feature extraction model is composed of a two-layer graph convolutional neural network model.

[0025] In one embodiment, the feature difference loss function is:

[0026]

[0027] In the formula, represents the common embedded feature representation of the source domain, represents the specific embedded feature representation of the source domain, L m represents the feature difference loss function, and T represents the transpose operation.

[0028] In one embodiment, the classification loss function is:

[0029]

[0030] In the formula, represents the embedded feature representation of the class-labeled samples in the source domain, is the classification result measured by the classification model, is the class label of the source domain belonging to the k-th class, k ∈ [1, C], where C is the total number of classes of the samples, and n sl is the number of samples with class labels in the source domain, and L s represents the classification loss function.

[0031] In one embodiment, the domain feature alignment loss function is:

[0032]

[0033] In the formula, z ci represents the i-th common embedded feature representation, and G d (z i ) is the result measured by the domain adversarial discriminant model, is whether the input common feature representation belongs to the domain label of the source domain or the target domain, and n s is the total number of samples in the source domain, and n t is the total number of samples in the target domain, and L d represents the domain feature alignment loss function.

[0034] In one embodiment, the class feature alignment loss function is:

[0035]

[0036] In the formula, L c represents the class feature alignment loss function, represents the embedded feature representation of the source domain samples whose class label or pseudo-class label is the k-th class, represents the embedded feature representation of the target domain samples whose pseudo-class label is the k-th class, and are respectively and 's probability distributions, and C is the total number of classes of the samples.

[0037] In one embodiment, the expression for continuously updating the parameters in the domain adaptation network is:

[0038] min(L s + λL m - βL d + γL c )

[0039] In the formula, L mDenote the feature difference loss function as \(L\). s Denote the classification loss function as \(L\). d Denote the domain feature alignment loss function as \(L\). c Denote the class feature alignment loss function. \(\lambda\), \(\beta\), and \(\gamma\) are the balance factors between the corresponding loss functions respectively.

[0040] The above - mentioned classification method for unsupervised domain adaptation based on graph convolutional network obtains sample data in the source domain and sample data in the target domain as training data; updates the graph connection relationships of samples in the two domains respectively according to the similarity between the sample data in the source domain and the target domain; inputs the sample data in the source domain and the target domain into the domain - adaptive network for training. The domain - adaptive network is an unsupervised domain - adaptive network based on graph convolutional network, and the domain - adaptive network includes: a cross - domain feature extraction model, a source - domain feature extraction model, a classification model, a domain - adversarial discriminative model, and a class alignment model; trains the domain - adaptive network to continuously update and iterate the parameters in the domain - adaptive network. When the domain - adaptive network reaches the convergence condition, a domain - adaptive classification model is obtained; inputs the data to be classified into the domain - adaptive classification model for classification, and obtains the classification result of the data to be classified. It improves the performance of the unsupervised domain - adaptive classification model based on graph convolution and further improves the accuracy of data classification. Description of the Drawings

[0041] Figure 1 It is a schematic flow chart of a classification method for unsupervised domain adaptation based on graph convolutional network in an embodiment.

[0042] Figure 2 It is a schematic flow chart of the construction method of the domain - adaptive network in an embodiment. Detailed Embodiments

[0043] In order to make the objectives, technical solutions, and advantages of this application clearer, the following further details this application in combination with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain this application and are not used to limit this application.

[0044] The classification method for unsupervised domain adaptation based on graph convolutional network provided by this application can be applied to a terminal or a server. Among them, the terminal can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0045] In one embodiment, as Figure 1 shown, a classification method for unsupervised domain adaptation based on graph convolutional network is provided. Taking the application of this method to a terminal as an example, the method includes the following steps:

[0046] Step S220: Obtain the sample data X in the source domain s and the sample data X in the target domain t as training data.

[0047] Among them, the categories and the number of categories of the sample data in the source domain and the sample data in the target domain are the same. Some of the sample data in the source domain has category labels, while the sample data in the target domain does not have category labels. The category label refers to the label used to mark which category the sample belongs to. The type of the sample data can be text data, picture data, or audio data. According to the needs of the classification task, the type of the sample data is determined. For example, when training a classification model for classifying which subject a paper belongs to, the papers marked with which subject can be used as the sample data in the source domain, and the papers not marked with which subject can be used as the sample data in the target domain, and steps S240 to S280 are executed to obtain a domain adaptation classification model for classifying which subject a paper belongs to.

[0048] Step S240: Update the graph connection relationships A of the samples in the two domains respectively according to the similarity between the sample data in the source domain and the target domain s and A t .

[0049] Among them, the positive pointwise mutual information (PPMI) is used to calculate the similarity between the sample data. The calculation formula of PPMI is as follows:

[0050]

[0051] In the formula, where n is the number of samples in a domain, A ij is the weight coefficient of the connection between sample i and sample j, ppmi ij is the sample similarity between sample i and sample j, and the larger the value of ppmi ij , the higher the similarity.

[0052] Step S260: Input the sample data in the source domain and the target domain into the domain adaptation network for training. The domain adaptation network is an unsupervised domain adaptation network based on the graph convolutional network. The domain adaptation network includes: a cross-domain feature extraction model, a source domain feature extraction model, a classification model, a domain adversarial discriminative model, and a class alignment model.

[0053] Step S280: Train the domain adaptation network to continuously update and iterate the parameters in the domain adaptation network. When the domain adaptation network reaches the convergence condition, a domain adaptation classification model is obtained.

[0054] Step S300: Input the data to be classified into the domain adaptation classification model for classification to obtain the classification result of the data to be classified.

[0055] Among them, the data to be classified is the data that needs to be classified. The data to be classified can be many data or one data. For example, to classify which subject a certain paper belongs to, input the paper into the domain adaptation classification model, and output the subject to which the paper belongs.

[0056] The above-mentioned unsupervised domain adaptation classification method based on graph convolutional network obtains the sample data in the source domain and the sample data in the target domain as training data; respectively updates the graph connection relationships of the samples in the two domains according to the similarity between the sample data in the source domain and the target domain; inputs the sample data in the source domain and the target domain into the domain adaptation network for training. The domain adaptation network is an unsupervised domain adaptation network based on graph convolutional network, and the domain adaptation network includes: a cross-domain feature extraction model, a source domain feature extraction model, a classification model, a domain adversarial discriminant model, and a class alignment model; trains the domain adaptation network to continuously update and iterate the parameters in the domain adaptation network. When the domain adaptation network reaches the convergence condition, the domain adaptation classification model is obtained; inputs the data to be classified into the domain adaptation classification model for classification to obtain the classification result of the data to be classified. It improves the performance of the unsupervised domain adaptation classification model based on graph convolution and further improves the accuracy of data classification.

[0057] In one embodiment, the cross-domain feature extraction model extracts the sample features common to the source domain and the target domain, the source domain feature extraction model extracts the sample features specific to the source domain, the classification model calculates the classification loss value, the domain adversarial discriminant model calculates the domain feature alignment loss value, and the class alignment model calculates the class feature alignment loss value.

[0058] In one embodiment, the total loss value is the sum of the feature difference loss value, the classification loss value, the domain feature alignment loss value, and the class feature alignment loss value. Among them, the feature difference loss value is the feature difference obtained by inputting the sample data in the source domain into the cross-domain feature extraction model and the source domain feature extraction model, and the classification loss value is based on inputting the sample data in the source domain into the classification model.

[0059] Such as Figure 2As shown, in one embodiment, the domain adaptation network is constructed as follows: The sample data of the source domain and the sample data of the target domain are input into a cross-domain feature extraction model to obtain the common embedded feature representations of the source domain and the target domain; the sample data of the source domain are input into a source domain feature extraction model to obtain the specific embedded feature representations of the source domain; the difference between the common embedded feature representations and the specific embedded feature representations of the source domain is calculated to construct a feature difference loss function; the sample data of the target domain are input into the source domain feature extraction model to obtain the target domain embedded feature representations with the source domain style, and then combined with the common embedded feature representations of the target domain through an attention mechanism to form the embedded feature representations of the target domain; at the same time, the common embedded feature representations of the source domain are combined with the specific embedded feature representations of the source domain through the attention mechanism to form the embedded feature representations of the source domain; the obtained embedded feature representations of the source domain and the target domain are input into a classification model, the part with class labels in the embedded feature representations of the source domain constructs a classification loss function, and the remaining part without class labels in the source domain and the embedded feature representations of the target domain generate pseudo-class labels corresponding to their feature representations; the obtained common embedded feature representations of the source domain and the target domain are input into a domain adversarial discriminant model to construct a domain feature alignment loss function; the sample data of the source domain and the sample data of the target domain are grouped according to the classes in the class labels and pseudo-class labels, and at the same time, the embedded feature representations of the samples in different groups are input into a class alignment model to construct a class feature alignment loss function.

[0060] Among them, the common embedded feature representations of the source domain and the target domain are obtained by using a cross-domain feature extraction model, and then the specific embedded feature representations of the source domain and the specific embedded feature representations of the target domain with the source domain style are respectively obtained through the source domain feature extraction model. The common embedded feature representations and the specific embedded feature representations are respectively fused through an attention mechanism to obtain the embedded feature representations of the source domain and the target domain, thereby confusing the source domain and the target domain to narrow the distribution difference between the two domains. The embedded feature representations with class labels in the source domain are used to train the classification model, and at the same time, pseudo-class labels are generated for the samples without class labels in the two domains. The feature difference loss function makes the common embedded feature representations and the specific embedded feature representations mutually exclusive, and a domain feature alignment loss function and a class feature alignment loss function are respectively designed to eliminate the domain distribution difference and the distribution difference of the same class, improving the accuracy of the domain adaptation task under the condition of only relying on a small number of labeled samples in the source domain.

[0061] In one embodiment, the cross-domain feature extraction model is composed of a shared network of a two-layer graph convolutional neural network (GCN), and the sample data of the source domain and the sample data of the target domain are both input into the shared network to obtain their common embedded feature representations; the source domain feature extraction model is composed of a two-layer graph convolutional neural network model.

[0062] Among them, the graph convolutional neural network extracts the embedded feature representations of samples in different domains, mines the connection relationships between samples, and promotes the information transmission between samples. The calculation formula for the common embedded feature representation of the sample data in the source domain and the sample data in the target domain is as follows:

[0063]

[0064]

[0065] Among them, A s is the graph connection relationship between samples in the source domain, X s is the sample data in the source domain, θ0 is the network parameter of the first layer of the graph convolutional neural network, θ1 is the network parameter of the second layer of the graph convolutional neural network, A t is the graph connection relationship of samples in the target domain, X t is the sample data in the target domain, represents the common embedded feature representation of the target domain, represents the common embedded feature representation of the source domain.

[0066] The source domain feature extraction model is composed of a two-layer graph convolutional neural network (GCN). The source domain samples are input into this model to obtain the source domain specific embedded feature representation The target domain is input into this model to obtain the target domain specific embedded feature representation with the source domain style

[0067] In one embodiment, the feature difference loss function is:

[0068]

[0069] In the formula, represents the common embedded feature representation of the source domain, represents the specific embedded feature representation of the source domain, L m represents the feature difference loss function, and T represents the transpose operation.

[0070] Among them, the embedded feature representation Z of the source domain s is combined by the attention mechanism with the specific embedded feature representation of the source domain and the common embedded feature representation of the source domain. The embedded feature representation Z of the target domain t is combined by the attention mechanism with the target domain specific embedded feature representation with the source domain style and the target domain common embedded feature representation. Among them, the calculation method of the attention mechanism is as follows:

[0071]

[0072]

[0073] Wherein, w1 and w2 are column vectors, and w1 + w2 = 1.

[0074] In one embodiment, the classification loss function is:

[0075]

[0076] Wherein, represents the embedded feature representation of the labeled samples in the source domain, is the classification result measured by the classification model, is the class label of the source domain belonging to the k-th class, k ∈ [1, C], C is the total number of classes of the samples, n sl is the number of labeled samples in the source domain, L s represents the classification loss function.

[0077] In one embodiment, the domain feature alignment loss function is:

[0078]

[0079] Wherein, z ci represents the i-th common embedded feature representation, G d (z i ) is the result measured by the domain adversarial discriminant model, is whether the input common feature representation belongs to the domain label of the source domain or the domain label of the target domain, n s is the total number of samples in the source domain, n t is the total number of samples in the target domain, L d represents the domain feature alignment loss function.

[0080] Wherein, the domain label is an identifier used to identify which domain the common feature representation belongs to.

[0081] In one embodiment, the class feature alignment loss function is:

[0082]

[0083] Wherein, L c represents the class feature alignment loss function, represents the embedded feature representation of the source domain sample whose class label or pseudo-class label is the k-th class, represents the embedded feature representation of the target domain sample whose pseudo-class label is the k-th class, and are respectively and 's probability distributions, C is the total number of classes of the samples.

[0084] In one embodiment, the expression for continuously updating the parameters in the domain adaptation network is:

[0085] min(L s +λL m -βL d +γL c )

[0086] In the formula, L m represents the feature difference loss function, L s represents the classification loss function, L d represents the domain feature alignment loss function, L c represents the class feature alignment loss function, and λ, β, and γ are the balance factors between the corresponding loss functions respectively.

[0087] For the above unsupervised domain adaptation classification method based on graph convolutional network, the graph convolutional neural network is used to extract the embedded feature representations of samples in different domains, mine the connection relationships between samples, and promote the information transmission between samples. Secondly, the target domain passes through the source domain feature extraction model to obtain specific embedded feature representations with the source domain style, and then the difference loss is used to make the common embedded feature representations irrelevant to the specific embedded feature representations. In the present invention, a domain adversarial discriminant model is established through an adversarial mechanism to maximize the domain classification loss, eliminating the distribution difference of the common embedded features between domains. The common and specific embedded feature representations are fused through an attention mechanism into the source domain embedded feature representation and the target domain embedded feature representation. At the same time, a classification model is established to classify the samples with class labels, calculate the classification loss, and assign pseudo-class labels to the samples without class labels, where the classification loss ensures the effectiveness of the classification model. Finally, a class alignment model is established in the invention to eliminate the distribution difference between the same-class samples in different domains, aligning the sample distributions of the two domains at the class level. Further effectively improves the performance of the unsupervised domain adaptation classification model based on graph convolution.

[0088] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the arrows, these steps do not necessarily execute in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0089] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0090] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. An unsupervised domain adaptation classification method based on graph convolutional network, characterized in that, The method includes: Obtaining the sample data in the source domain and the sample data in the target domain as training data, and the type of the training data is text data; Respectively updating the graph connection relationships of the samples in the two domains according to the similarity between the sample data in the source domain and the target domain; Inputting the sample data in the source domain and the target domain into a domain adaptation network for training. The domain adaptation network is an unsupervised domain adaptation network based on a graph convolutional network, and the domain adaptation network includes: a cross-domain feature extraction model, a source domain feature extraction model, a classification model, a domain adversarial discriminant model, and a class alignment model; Training the domain adaptation network to continuously update and iterate the parameters in the domain adaptation network. When the domain adaptation network reaches the convergence condition, a domain adaptation classification model is obtained; Inputting the data to be classified into the domain adaptation classification model for classification to obtain the classification result of the data to be classified; The construction method of the domain adaptation network includes: Inputting the sample data of the source domain and the sample data of the target domain into the cross-domain feature extraction model to obtain the common embedding feature representations of the source domain and the target domain; Inputting the sample data of the source domain into the source domain feature extraction model to obtain the specific embedding feature representation of the source domain; Calculating the difference between the common embedding feature representation of the source domain and the specific embedding feature representation to construct a feature difference loss function; Inputting the sample data of the target domain into the source domain feature extraction model to obtain the target domain embedding feature representation with the source domain style, and then combining it with the common embedding feature representation of the target domain through an attention mechanism to obtain the embedding feature representation of the target domain; at the same time, combining the common embedding feature representation of the source domain with the specific embedding feature representation of the source domain through the attention mechanism to obtain the embedding feature representation of the source domain; Inputting the obtained embedding feature representations of the source domain and the target domain into the classification model. For the part with class labels in the embedding feature representation of the source domain, a classification loss function is constructed, and for the remaining part without class labels in the source domain and the embedding feature representation of the target domain, pseudo-class labels corresponding to their feature representations are generated; Inputting the obtained common embedding feature representations of the source domain and the target domain into the domain adversarial discriminant model to construct a domain feature alignment loss function; Grouping the sample data of the source domain and the sample data of the target domain according to the classes in the class labels and pseudo-class labels, and at the same time inputting the embedding feature representations of the samples in different groups into the class alignment model to construct a class feature alignment loss function.

2. The method according to claim 1, wherein The cross-domain feature extraction model extracts the common sample features of the source domain and the target domain, the source domain feature extraction model extracts the specific sample features of the source domain, the classification model calculates the classification loss value, the domain adversarial discriminant model calculates the domain feature alignment loss value, and the class alignment model calculates the class feature alignment loss value.

3. The method according to claim 2, wherein The total loss value is the sum of the feature difference loss value, the classification loss value, the domain feature alignment loss value, and the class feature alignment loss value. Among them, the feature difference loss value is the feature difference obtained by inputting the sample data of the source domain into the cross-domain feature extraction model and the source domain feature extraction model, and the classification loss value is based on the sample data of the source domain input into the classification model.

4. The method according to claim 1, wherein The cross-domain feature extraction model is composed of a shared network of a two-layer graph convolutional neural network. The sample data of the source domain and the sample data of the target domain are both input into the shared network to obtain their common embedded feature representations. The source domain feature extraction model is composed of a two-layer graph convolutional neural network model.

5. The method according to claim 1, wherein The feature difference loss function is: In the formula, represents the common embedded feature representation of the source domain, represents the specific embedded feature representation of the source domain, and L m represents the feature difference loss function, and T represents the transpose operation.

6. The method according to claim 1, wherein The classification loss function is: In the formula, represents the embedded feature representation of the class-labeled samples in the source domain, is the classification result measured by the classification model, is the class label of the k-th class in the source domain, where k ∈ [1, C] and C is the total number of classes of the samples, and n sl is the number of samples with class labels in the source domain, and L s represents the classification loss function.

7. The method according to claim 1, characterized in that, The domain feature alignment loss function is: where z ci represents the i-th common embedded feature representation, G d (z i ) is the result measured by the domain adversarial discriminative model, is the domain label indicating whether the input common feature representation belongs to the source domain or the target domain, n s is the total number of samples in the source domain, n t is the total number of samples in the target domain, L d represents the domain feature alignment loss function.

8. The method according to claim 1, characterized in that, The class feature alignment loss function is: In the formula, L c represents the class feature alignment loss function, represents the source domain embedding feature representation of the sample with the class label or pseudo-class label being the k-th class, represents the target domain embedding feature representation of the sample with the pseudo-class label being the k-th class, and are respectively and 's probability distributions, and C is the total number of classes of the samples.

9. The method according to claim 1, wherein The expression for continuously updating the parameters in the domain adaptation network is: min(L s +λL m -βL d +γL c ) Where, L m represents the feature difference loss function, L s represents the classification loss function, L d represents the domain feature alignment loss function, L c represents the class feature alignment loss function, and λ, β, and γ are the balance factors between the corresponding loss functions respectively.

Images