A heterogeneous domain adaptation learning method for cross-network heterogeneous node classification

Abstract

The application relates to the technical field of information classification, and discloses a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification, which comprises the following steps: S1, constructing PPMI matrices of source networks and target networks to obtain global structure information of the networks, and using a variational graph autoencoder to obtain node features of the source networks and the target networks, decoding the node features to restore the original PPMI matrices so as to retain global adjacent information between the nodes; S2, constructing a hypergraph, and more closely associating the source networks and the target networks by using labeled source nodes and target nodes; and S3, constructing a hypergraph autoencoder to obtain high-order features of nodes of the source networks and the target networks, and using the learned features to reconstruct the hypergraph to retain high-order relationships. The application has the advantages that several learning tasks based on real-world data sets are constructed, and extensive experiments are conducted. The performance of the method is better than that of some other existing methods, and the cross-network heterogeneous node classification task is better completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information classification, and particularly relates to a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification. BACKGROUND

[0002] With the rapid development of information technology, various systems have generated a large amount of data. The relationship between these data is complex and diverse, causing the problem of "information overload", which greatly increases the difficulty of users in obtaining effective information when performing network analysis and data mining tasks. Especially when dealing with data from different networks, the heterogeneity of data becomes a huge challenge. Cross-network node classification, as one of the key tasks of graph data analysis, aims to identify the categories of nodes in different networks with the help of graph structure information.

[0003] In practical applications, samples of different networks are often represented by different features, resulting in data heterogeneity. For example, users in social networks from two different countries may use different languages to describe, showing heterogeneous data attributes. In this context, cross-network heterogeneous node classification has become an important research field, and the core problem involved is how to effectively transfer the knowledge learned from a source network to another network with different feature distributions. In recent years, hypergraph learning has received extensive attention due to its excellent modeling ability on complex data, so researchers have introduced hypergraph learning into the node classification task. In a hypergraph, a hyperedge can connect multiple nodes; nodes in different networks can exhibit high-order relationships because they may share similar classes. Therefore, researchers have established a hypergraph representing the high-order relationship between the nodes of the source network and the target network, i.e. in the hypergraph, a value of 1 indicates that the source node and the target node belong to the same class, and a value of 0 indicates that they do not belong, so that the high-level feature representation between nodes can be utilized by the hypergraph convolution network.

[0004] Heterogeneous domain adaptive methods can be used to solve the problem of inconsistent data distribution between source and target domains. These methods mainly aim to find a latent common space to reduce the distribution difference between the source and target domains. Existing heterogeneous domain adaptation methods, such as learning a domain-invariant space or projecting source and target samples into a shared feature space to simultaneously minimize information loss and domain difference. Although traditional heterogeneous domain adaptive methods perform well in handling independent and identically distributed samples, they may not perform well when dealing with network structure data that is not independent and identically distributed.

[0005] Therefore, it is necessary to invent a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification to solve the above problems. SUMMARY

[0006] The application aims to provide a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification, aiming to solve cross-network node classification under heterogeneous settings, i.e., cross-network heterogeneous node classification.

[0007] In order to achieve the above-mentioned purpose, the application provides the following technical scheme:

[0008] A heterogeneous domain adaptive learning method for cross-network heterogeneous node classification comprises the following steps:

[0009] S1, constructing PPMI matrices of source network and target network to obtain global structure information of the network, and using a variational graph autoencoder to obtain node features of the source network and the target network, and decoding the node features to restore the original PPMI matrix to retain the global proximity information between nodes;

[0010] S2, constructing a hypergraph, and more closely associating the source network and the target network by using the labeled source nodes and the target nodes;

[0011] S3, constructing a hypergraph autoencoder to obtain high-order features of the source network and the target network nodes, and using the learned features to reconstruct the hypergraph to retain the high-order relationship;

[0012] S4, using a feature matching mechanism to minimize the difference between different features of each labeled node, and obtaining a feature representation with rich and consistent node information;

[0013] S5, using an adversarial domain adaptation module to reduce the domain difference between the labeled nodes of the source network and the target network;

[0014] S6, obtaining an overall objective function, and training a classifier based on the features of the labeled source nodes and the target nodes.

[0015] Further, the above step S1 comprises:

[0016] S1.1, for the source network, using a double-layer normalized graph convolutional network as an X s and P s based global information reasoning model:

[0017]

[0018] wherein X s is a source network node representation, P s is a source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node feature, is the i-th vector of X , and has

[0019] wherein μs =φ s (X s ,P s ) is the mean vector, μ s,i It is its i-th vector. It is the variance matrix, σ s,i It is its i-th vector; where φ s (·)and A two-layer normalized graph convolutional network is defined as follows:

[0020]

[0021] in For a symmetric normalized adjacency matrix, Φ s For P s The angle matrix, σ(·) is the ReLU activation function, Δ s,1 and Δ s,2 These are the network parameters for the first and second layers, respectively.

[0022] S1.2, Using the inner product between latent representations as the generative model:

[0023]

[0024] in To reconstruct the source PPMI matrix, sigmoid(·) is the logistic sigmoid function;

[0025] Therefore, the loss function of the variational graph autoencoder on the source network is obtained as follows:

[0026]

[0027] Where KL[q s (·)||p s (·)] is q s (·) and p s Kullback-Leibler divergence between (·) and (·), It is a Gaussian prior;

[0028] S1.3 For the target network, let q t (·) and p t (·) represent the global information reasoning model and the generative model, respectively, based on X. t and P t The loss function is as follows:

[0029]

[0030] Where X t For the target network node representation, Pt For the target network PPMI, To reconstruct the target PPMI matrix, The learned potential target node features, Gaussian priors;

[0031] S1.4 In summary, the loss term for analyzing the global information of the source network and the target network is:

[0032]

[0033] Furthermore, step S2 above includes:

[0034] S2.1 Given the node labels of the source network and part of the target network, construct a hypergraph to represent the higher-order relationships between these nodes:

[0035] use This represents a hypergraph where each entry H(i,j) indicates whether the i-th source node and the j-th target node belong to the same category.

[0036]

[0037] Where N s The number of nodes in the source network. To mark the number of target samples.

[0038] Furthermore, step S3 above includes:

[0039] S3.1 For the source network, a two-layer normalized hypergraph convolution operation g is used. s (·),as follows:

[0040]

[0041] in For the learned source node features, W s,1 and W s,2 These are the network parameters for the first and second layers, respectively.

[0042]

[0043] Where Ω s It is a diagonal matrix that stores the positive weights of the hyperedges.

[0044]

[0045] and

[0046]

[0047] are diagonal matrices, which help to avoid the problems of gradient vanishing and numerical instability in the convolution operation process;

[0048] S3.2、Similarly, the target network g t The two-layer normalized hypergraph convolution function of (·) is represented as:

[0049]

[0050] where is the labeled target data feature, is the learned labeled target node representation, W t,1 and W t,2 are the network parameters of the first and second layers, respectively,

[0051]

[0052] where Ω t is a diagonal matrix storing the positive weights of nodes,

[0053]

[0054] and

[0055]

[0056] are all diagonal matrices;

[0057] S3.3、The high-order relationship between the labeled source nodes and the target nodes is preserved by the decoder h(·) of the reconstructed hypergraph:

[0058]

[0059] where is the reconstructed hypergraph, and are the i-th and j-th vectors in and respectively;

[0060] S3.4、Therefore, the hypergraph reconstruction loss function is obtained as:

[0061]

[0062] Further, the above step S4 includes:

[0063] S4.1、In order to obtain the source network and target network node information rich and consistent feature representation, the following feature matching mechanism is adopted:

[0064]

[0065] where is the mean square error function. Thus, the difference between the two feature spaces of the labeled source nodes and target nodes can be reduced by minimizing this expression, thereby obtaining informative and consistent feature representations.

[0066] Further, the above step S5 comprises:

[0067] S5.1, to further mitigate the negative impact of domain shift, an adversarial domain adaptation is employed to jointly train an effective classifier and align the distributions of the labeled source nodes and target nodes;

[0068] respectively, as the node classifier and domain classifier. Define y (·) and f d (·) as the node classifier and domain classifier. Define whose node label and domain label are and Note that Y s is the one-hot label matrix of nodes, is the one-hot label matrix of the labeled target nodes, and are 0 and 1 vectors with dimensions N s and have:

[0069]

[0070] where and are the loss of the node label classifier and domain classifier, respectively, where and are the loss functions evaluated at the i-th node. Here, γ is a balance parameter, z i is the i-th vector of Z, y i and d i are the i-th element of vectors y and d, respectively;

[0071] S5.2, minimize to learn node label invariant features so that the classifier f y (·) can correctly distinguish different nodes into their respective classes. On the other hand, maximize to deceive the domain classifier f d (·) by generating node features that are indistinguishable across networks. By adding a gradient reversal layer f r (·), that is, The above expression can be re-expressed as:

[0072]

[0073] Further, the step S6 comprises:

[0074] S6.1, based on the above steps, the total objective function of the model is obtained,

[0075]

[0076] Wherein, alpha, beta and eta are balance parameters;

[0077] Training the classifier based on the features of the source node and the target node of the label;

[0078] S6.2, the performance of the method is evaluated by using Macro-F1 and Micro-F1 indicators: TP(c), FP(c) and FN(c) represent the number of true cases, false positives and false negatives related to label c.

[0079] Micro-F1 gives the same weight to each sample, and is defined as follows:

[0080]

[0081] Macro-F1 gives the same weight to each class, and is defined as follows:

[0082]

[0083] In the above technical solutions, the technical effects and advantages provided by the present application are:

[0084] 1. The present application obtains two types of features by using global information in each network and high-order relationships between network nodes, and then learns network invariant features by using feature matching and domain adaptation;

[0085] 2. The present application studies a more general setting: heterogeneous setting, that is, nodes have different features, which makes this method make a good contribution to the machine learning and data mining community;

[0086] 3. The present application constructs several learning tasks based on real-world data sets and conducts extensive experiments. The performance of the method is better than that of some other existing methods, and the cross-network heterogeneous node classification task is better completed. BRIEF DESCRIPTION OF DRAWINGS

[0087] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0088] Figure 1A flowchart of a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification. DETAILED DESCRIPTION

[0089] In order to enable those skilled in the art to better understand the technical solutions of the present application, the present application will be further described in detail below with reference to the accompanying drawings.

[0090] The present application provides a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification as shown in Figure 1 The present application provides a heterogeneous domain adaptive learning method for cross-network heterogeneous node classification as shown in

[0091] S1, constructing the PPMI matrix of the source network and the target network to obtain the global structure information of the network, and using a variational graph autoencoder to obtain the node features of the source network and the target network, and decoding the node features to restore the original PPMI matrix to retain the global proximity information between the nodes.

[0092] Specifically, step S1 includes:

[0093] S1.1, for the source network, a double-layer normalized graph convolutional network is used as an X s and P s based global information reasoning model:

[0094]

[0095] Wherein X s is the source network node representation, P s is the source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node feature, is the i-th vector of X , and has

[0096]

[0097] Wherein μ s = φ s (X s , P s ) is the mean vector, μ s,i is the i-th vector thereof, is the variance matrix, σ s,i is the i-th vector thereof; wherein φ s (·) and represent two layers of normalized graph convolutional network, which are defined as:

[0098]

[0099] Wherein is a symmetric normalized adjacency matrix, Φ sP s is the diagonal matrix, σ(·) is the ReLU activation function, Δ s,1 and Δ s,2 are the network parameters of the first and second layers, respectively;

[0100] S1.2, the inner product between latent representations is used as the generative model:

[0101]

[0102] wherein is the reconstructed source PPMI matrix, sigmoid(·) is the logistic sigmoid function;

[0103] Therefore, the loss function of the variational graph autoencoder on the source network is obtained as follows:

[0104]

[0105] wherein KL[q s (·)||p s (·)] is the Kullback-Leibler divergence between q s (·) and p s (·), is the Gaussian prior;

[0106] S1.3, for the target network, let q t (·) and p t (·) be the global information inference model and the generative model, respectively, based on X t and P t , the following loss function is obtained:

[0107]

[0108] wherein X t is the target network node representation, P t is the target network PPMI, is the reconstructed target PPMI matrix, is the learned latent target node feature, is the Gaussian prior;

[0109] S1.4, in summary, the loss term for analyzing the global information of the source network and the target network is:

[0110]

[0111] S2, a hypergraph is constructed, and the labeled source nodes and target nodes are used to more closely associate the source network and the target network.

[0112] Specifically, step S2 includes:

[0113] S2.1, given the node labels of the source network and part of the target network, a hypergraph is constructed to represent the high-order relationship between the nodes:

[0114] Let H represent the hypergraph, and each entry H(i,j) represents whether the ith source node and the jth target node belong to the same category:

[0115]

[0116] where N s is the number of source network nodes, is the number of labeled target samples.

[0117] S3, a hypergraph autoencoder is constructed to obtain high-order features of the source network and target network nodes, and the learned features are used to reconstruct the hypergraph to preserve the high-order relationship.

[0118] Specifically, step S3 includes:

[0119] S3.1, for the source network, a two-layer normalized hypergraph convolution operation g s (·) is used, as follows:

[0120]

[0121] where is the learned source node feature, W s,1 and W s,2 are the network parameters of the first layer and the second layer, respectively,

[0122]

[0123] where Ω s is a diagonal matrix storing the positive weights of the hyperedge,

[0124]

[0125] and

[0126]

[0127] are diagonal matrices, which help to avoid the problems of gradient disappearance and numerical instability in the convolution operation process;

[0128] S3.2, similarly, the two-layer normalized hypergraph convolution function of the target network g t (·) is represented as:

[0129] ​

[0130] where is the labeled source node feature, is the learned labeled target node feature, W t,1 and W t,2 are the network parameters of the first and second layers, respectively,

[0131]

[0132] where Ω t is a diagonal matrix storing the positive weights of nodes,

[0133]

[0134] and

[0135]

[0136] are all diagonal matrices;

[0137] S3.3, the high-order relationships between the labeled source nodes and target nodes are preserved by reconstructing the hypergraph decoder h(·):

[0138]

[0139] where is the reconstructed hypergraph, and are the i-th and j-th vectors in and , respectively;

[0140] S3.4, therefore, the hypergraph reconstruction loss function is obtained as:

[0141]

[0142] S4, a feature matching mechanism is adopted to minimize the difference between different features of each labeled node, so as to obtain a node information-rich and consistent feature representation.

[0143] Specifically, step S4 includes:

[0144] S4.1, in order to obtain a node information-rich and consistent feature representation of the source network and the target network, the following feature matching mechanism is adopted:

[0145]

[0146] where is the mean square error function. Therefore, the difference between the two feature spaces of the labeled source nodes and target nodes can be reduced by minimizing this formula, so as to obtain an information-rich and consistent feature representation.

[0147] S5, adopt an adversarial domain adaptation module to reduce the domain difference between the source network and the target network labeled nodes.

[0148] Specifically, step S5 includes:

[0149] S5.1, in order to further reduce the negative effects of domain bias, adopt adversarial domain adaptation to jointly train effective classifiers and align the distributions of labeled source nodes and target nodes.

[0150] respectively use f y (·) and f d (·) as node classifier and domain classifier. Define whose node label and domain label are and Note that Y s is the one-hot label matrix of the node, is the one-hot label matrix of the labeled target node, and are 0 and 1 vectors respectively, whose dimensions are N s and have:

[0151]

[0152] where and are the losses of node label classifier and domain classifier respectively, where and are the loss functions evaluated at the i-th node. Here, γ is a balance parameter, z i is the i-th vector of Z, y i and d i are the i-th elements of vectors y and d respectively.

[0153] S5.2, minimize to learn node label invariant features, so that the classifier f y (·) can correctly distinguish different nodes into their respective categories. On the other hand, maximize to deceive the domain classifier f d (·) by generating node features that are indistinguishable across networks. By adding a gradient reversal layer f r (·), that is The above formula can be restated as:

[0154]

[0155] S6. obtaining an overall objective function based on the labeled source and target node features to train the classifier.

[0156] Specifically, step S6 includes:

[0157] S6.1, based on the above steps, the overall objective function of the model is obtained as,

[0158]

[0159] Wherein, α, β and η are balance parameters;

[0160] Training the classifier based on the labeled source node and target node features;

[0161] S6.2, the performance of the proposed method is evaluated by using Macro-F1 and Micro-F1 indicators: TP(c), FP(c) and FN(c) represent the number of true positives, false positives and false negatives related to label c.

[0162] Micro-F1 gives the same weight to each sample, and is defined as follows:

[0163]

[0164] Macro-F1 gives the same weight to each class, and is defined as follows:

[0165]

[0166] The present application evaluates the effectiveness of the proposed method on ACMv9(A), Citationv1(C) and DBLPv7(D) three real data sets provided by ArnetMine. A, C, D are three citation networks from ACM, Microsoft Academic Graph and DBLP respectively. The citation network is modeled as an undirected network, where the node represents the paper, the edge represents the citation relationship, and each paper belongs to one of the following. The attribute of each node is the sparse bag-of-words feature extracted from the paper title. Six cross-network heterogeneous node classification tasks are constructed based on the three fields of C→A, D→A, A→C, D→C, A→D and C→D. Five nodes are selected for each class as labeled target nodes, and the remaining target data is used for testing.

[0167] Table 1 lists the statistical information of these data sets.

[0168] Table 1

[0169]

[0170] The Micro-F1 and Macro-F1 scores of the application on the data set are compared with the current most advanced heterogeneous domain adaptation method DDA (Yao, Y., Zhang, Y., Li, X., Ye, Y.: Discriminative distribution alignment: A unified framework for heterogeneous domain adaptation. Pattern Recognition 101, 107165 (2020)), SSAN (Li, S., Xie, B., Wu, J., Zhao, Y., Liu, C.H., Ding, Z.: Simultaneous semantic alignment network for heterogeneous domain adaptation. In: Proceedings of the 28th ACM International Conference on Multimedia, pp. 3866-3874 (2020)), CDSPP (Wang, Q., Breckon, T.P.: Cross-domain structure preserving projection for heterogeneous domain adaptation. Pattern Recognition 123, 108362 (2022));

[0171] The experimental results are shown in Table 2, algorithm 1 corresponds to the verification results of the algorithm of the application; algorithm 2 corresponds to the verification results of DDA; algorithm 3 corresponds to the verification results of SSAN; algorithm 4 corresponds to the verification results of CDSPP; from Table 2, it can be seen that the Micro-F1 and Macro-F1 score results of the application are always better than those of other algorithms.

[0172] Table 2

[0173]

[0174]

[0175] The above only describes some exemplary embodiments of the application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the application. Therefore, the above figures and description are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the application.

Claims

1. A heterogeneous domain adaptation learning method for cross-network heterogeneous node classification, characterized in that, comprising the following steps: S1, constructing PPMI matrices of source network and target network to obtain global structure information of the networks, and using variational graph autoencoder to obtain node features of the source network and the target network, and decoding the node features to restore the original PPMI matrices to retain the global proximity information between the nodes, wherein the nodes represent papers, and the attributes of each node are sparse bag-of-words features extracted from paper titles; S2, constructing a hypergraph to more closely associate the source network and the target network by using the labeled source nodes and target nodes; S3, constructing a hypergraph autoencoder to obtain high-order features of the nodes of the source network and the target network, and using the learned features to reconstruct the hypergraph to retain the high-order relationships; S4, using a feature matching mechanism to minimize the difference between different features of each labeled node to obtain feature representations that are rich in node information and consistent; S5, using an adversarial domain adaptation module to reduce the domain difference between the labeled nodes of the source network and the target network; S6, obtaining an overall objective function to train a classifier based on the features of the labeled source nodes and target nodes.

2. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 1, wherein the step S1 comprises: S1.2, using the inner product between latent representations as a generative model: S1.1, for the source network, adopt a double-layer normalized graph convolution network as the X s and P s global information inference model: where X s is the source network node representation, P s is the source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node feature, is the i-th vector of X and has where μ s = φ s (X s , P s ) is the mean vector, μ s,i is its i-th vector, is the covariance matrix, σ s,i is its i-th vector; where φ s (·) and denote two-layer normalized graph convolutional networks, defined as: wherein is a symmetric normalized adjacency matrix, Φ s is a diagonal matrix of P s , σ(·) is a ReLU activation function, Δ s,1 and Δ s,2 are network parameters of the first and second layers, respectively; Therefore, the loss function of the variational graph autoencoder on the source network is obtained as: wherein sigmoid(·) is a logisticsigmoid function for reconstructing the source PPMI matrix. S1.4, based on the above, the loss term for analyzing the global information of the source network and the target network is: where KL[q s (·)||p s (·)] is the Kullback-Leibler divergence between q s (·) and p s (·), is a Gaussian prior; S1.3, for the target network, set q t (·) and p t (·) are the global information inference model and the generation model, respectively, based on X t and P t , the following loss function is obtained: where X t is the target network node representation, P t is the target network PPMI matrix, is the reconstructed target PPMI matrix, is the learned latent target node feature, is the Gaussian prior; where the first term measures the difference between the reconstructed data and the original data using cross-entropy, and the second term measures the difference between the pre-specified prior distribution and the network-generated posterior distribution using KL divergence; 3. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 2, wherein the step S2 comprises: S2.1, given the node labels of the source network and part of the target network, a hypergraph is constructed to represent the high-order relationships between the nodes:

4. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 3, wherein the step S3 comprises: With representing a hypergraph, each entry H(i,j) represents whether the ith source node and the jth target node belong to the same class: where N s is the number of source network nodes, is the number of labeled target samples. and S3.1, for the source network, a two-layer normalized hypergraph convolutional network g s (·), as follows: wherein W are learned source node features, s,1 and W s,2 are network parameters of the first and second layers, respectively, where Ω s is a diagonal matrix storing the positive weights of the hyperedges, where are diagonal matrices, which helps to avoid the problems of gradient disappearance and numerical instability in the convolution operation; and S3.2, Likewise, the target network g t The two-layer normalized hypergraph convolution function of (·) is represented as: wherein is a labeled target data feature, is a learned labeled target node feature, W t,1 and W t,2 are network parameters of the first and second layers, respectively, where Ω t is a diagonal matrix storing the positive weights of the nodes, are diagonal matrices; S3.3, the high-order relationships between the labeled source nodes and target nodes are retained by the decoder h(·) of the reconstructed hypergraph: S3.4, therefore, the hypergraph reconstruction loss function is obtained as: wherein is the reconstructed hypergraph, and are the i-th and j-th vectors in and respectively.

5. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 4, wherein the step S4 comprises: S4.1, in order to obtain feature representations that are rich in node information and consistent for the source network and the target network, the following feature matching mechanism is used: wherein l(·) is a mean square error function; therefore, the difference between the two feature spaces of the labeled source nodes and target nodes can be reduced by minimizing this formula, thereby obtaining feature representations that are rich in information and consistent.

6. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 5, wherein the step S5 comprises: ​ S5.1, In order to further alleviate the negative impact of domain shift, an adversarial domain adaptation is adopted to jointly train an effective classifier and align the distribution of labeled source nodes and target nodes; Use f y (·) and f d (·) as node classifier and field classifier respectively; define Y and Note that Y s is a one-hot label matrix for nodes, is a one-hot label matrix for target nodes that are labeled, and are 0 and 1 vectors respectively with dimension N s and have: where and are the losses for the node label classifier and the domain classifier, respectively, where and are the corresponding loss functions evaluated at the ith node; here, γ is a balancing parameter, z i is the ith vector of Z, y i and d i are the ith elements of the vectors y and d, respectively; S5.2, Minimize to learn node label invariant features, so that the classifier f y (·) can correctly distinguish different nodes into their respective categories; maximize to deceive the domain classifier f d (·) by generating node features that are indistinguishable across networks; by adding a gradient reversal layer f r (·), i.e.

7. The heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to claim 6, wherein the step S6 comprises: S6.1, Based on the above steps, the overall objective function of the model is obtained as, wherein a, β and η are balance parameters; training a classifier based on the labeled source nodes and target nodes features; S6.2, Macro-F1 and Micro-F1 indicators are used to evaluate the performance: TP(c), FP(c) and FN(c) represent the number of true positives, false positives and false negatives related to label c; Micro-F1 gives the same weight to each sample, and is defined as follows: Macro-F1 gives the same weight to each class, and is defined as follows:

Images