Heterogeneous domain adaptive learning method oriented to cross-network heterogeneous node classification

Abstract

The invention relates to the technical field of information classification, and discloses a heterogeneous domain adaptive learning method oriented to cross-network heterogeneous node classification, comprising the following steps: S1, constructing PPMI matrixes of a source network and a target network to obtain global structure information of the networks, and obtaining node features of the source network and the target network by using a variational graph auto-encoder, decoding the PPMI matrix to restore the original PPMI matrix so as to retain global proximity information between the nodes; s2, constructing a hypergraph, and using the marked source nodes and target nodes to more closely associate the source network and the target network; and S3, constructing a hypergraph auto-encoder so as to obtain high-order features of source network and target network nodes, and using the learned features to reconstruct a hypergraph so as to retain a high-order relationship. The method has the advantages that several data set learning tasks based on the real world are constructed, and wide experiments are carried out. The performance of the method is better than that of some existing other methods, and a cross-network heterogeneous node classification task is better completed.

Description

Technical Field

[0001] The present invention relates to the technical field of information classification, and in particular to a heterogeneous domain adaptation learning method for cross-network heterogeneous node classification. Background Art

[0002] With the rapid development of information technology, various systems have generated massive amounts of data. The relationships between these data are complex and diverse, causing the problem of "information overload", which greatly increases the difficulty for users to obtain effective information when performing network analysis and data mining tasks. Especially when processing 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 from different networks are often represented by different features, resulting in data heterogeneity. For example, users in social networks from two different countries may be described in different languages, showing heterogeneous data properties. In this context, cross-network heterogeneous node classification has become an important research field. The core issue involved is how to effectively transfer the knowledge learned from a source network to another network with different feature distribution. In recent years, hypergraph learning has attracted widespread 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 show high-order relationships because they may share similar classes. Therefore, researchers established a hypergraph representing the high-order relationship between the source and target networks based on their node labels, that is, a value of 1 in the hypergraph indicates that the source node and the target node belong to the same category, and a value of 0 indicates that they do not belong, so that the high-level feature representation between the nodes of the hypergraph convolutional network can be utilized.

[0004] Heterogeneous domain adaptation methods can be used to solve the problem of inconsistent data distribution between the source domain and the target domain. These methods mainly aim to discover a potential common space to reduce the distribution difference between the source domain and the target domain. Existing heterogeneous domain adaptation methods, such as by learning a domain-invariant space or by projecting the source and target samples into a shared feature space to simultaneously minimize information loss and domain differences. Although traditional heterogeneous domain adaptation methods perform well in processing independent and identically distributed samples, they may perform poorly when processing non-independently distributed network structure data.

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

[0006] The purpose of the present invention is to provide a heterogeneous domain adaptation learning method for cross-network heterogeneous node classification, aiming to solve the cross-network node classification under heterogeneous settings, that is, cross-network heterogeneous node classification.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

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

[0009] S1. Construct the PPMI matrix of the source network and the target network to obtain the global structural information of the network, and use the variational graph autoencoder to obtain the node features of the source network and the target network, decode them to restore the original PPMI matrix to retain the global proximity information between the nodes;

[0010] S2, build a hypergraph, using the labeled source and target nodes to more closely associate the source network and the target network;

[0011] S3. Build a hypergraph autoencoder to obtain high-order features of source and target network nodes, and use the learned features to reconstruct the hypergraph to preserve high-order relationships.

[0012] S4, adopt feature matching mechanism to minimize the differences between different features of each marked node and obtain feature representation with rich and consistent node information;

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

[0014] S6. Obtain the overall objective function and train the classifier based on the labeled source node and target node features.

[0015] Furthermore, the above step S1 includes:

[0016] S1.1. For the source network, a two-layer normalized graph convolutional network is used as the X-based s and P s Global information reasoning model:

[0017]

[0018] Where X s is the source network node, P s is the source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node features, for The i-th vector of

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

[0020]

[0021] in is the symmetric normalized adjacency matrix, Φ s P s 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;

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

[0023]

[0024] in is 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:

[0026]

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

[0028] S1.3. For the target network, let q t (·) and p t (·) are global information inference model and generation model respectively, based on X t and P t , there is the following loss function:

[0029]

[0030] Where X t is the target network node, Pt PPMI for the target network, is the reconstructed target PPMI matrix, is the learned potential target node feature, is a Gaussian prior;

[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, the above step S2 includes:

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

[0035] use 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 is the number of source network nodes, is the number of labeled target samples.

[0038] Furthermore, the above step S3 includes:

[0039] S3.1. For the source network, two layers of normalized hypergraph convolution operation g are used s (·),as follows:

[0040]

[0041] in is the learned source node feature, W s,1 and W s,2 are the network parameters of the first and second layers respectively,

[0042]

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

[0044]

[0045] and

[0046]

[0047] They are all diagonal matrices, which help avoid the problems of gradient vanishing and numerical instability during convolution operations;

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

[0049]

[0050] in To mark the target data features, 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 the nodes,

[0053]

[0054] and

[0055]

[0056] They are all angle matrices;

[0057] S3.3. Preserve the high-order relations between the labeled source and target nodes by reconstructing the decoder h(·) of the hypergraph:

[0058]

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

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

[0061]

[0062] Furthermore, the above step S4 includes:

[0063] S4.1. In order to obtain information-rich and consistent feature representations of source and target network nodes, the following feature matching mechanism is adopted:

[0064]

[0065] in is the mean square error function. Therefore, the difference between the two feature spaces of the labeled source and target nodes can be reduced by minimizing this formula, thereby obtaining an informative and consistent feature representation.

[0066] Furthermore, the above step S5 includes:

[0067] S5.1. To further mitigate the negative impact of domain shift, adversarial domain adaptation is used to jointly train effective classifiers and align the distribution of labeled source and target nodes.

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

[0069]

[0070] in and are the losses of the node label classifier and the domain classifier, respectively. and is the corresponding loss function 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;

[0071] S5.2. Minimization To learn the node label invariant features, the classifier f y (·) can correctly distinguish different nodes into their respective categories. On the other hand, maximizing Deceiving domain classifiers by generating node features that are indistinguishable across networks d (·). By adding a gradient reversal layer f r (·),Right now The above formula can be restated as:

[0072]

[0073] Furthermore, the above step S6 includes:

[0074] S6.1. Based on the above steps, the overall objective function of the model is:

[0075]

[0076] Among them, α, β and η are equilibrium parameters;

[0077] Train a classifier based on the labeled source and target node features;

[0078] S6.2. The performance of the proposed method is evaluated 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 associated with label c.

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

[0080]

[0081] Macro-F1 assigns equal weight to each category and is defined as follows:

[0082]

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

[0084] 1. The present invention obtains two types of features by utilizing the global information in each network and the high-order relationships between network nodes, and then uses feature matching and domain adaptation to learn network invariant features;

[0085] 2. This paper studies a more general setting: heterogeneous setting, i.e. nodes have different characteristics, which makes this method a good contribution to the machine learning and data mining community;

[0086] 3. This paper constructs several real-world dataset learning tasks and conducts extensive experiments. The performance of this method is better than some other existing methods and can better complete the task of classifying heterogeneous nodes across networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0088] Figure 1The figure is a flow chart of the heterogeneous domain adaptation learning method for cross-network heterogeneous node classification according to the present invention. DETAILED DESCRIPTION

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

[0090] The present invention provides Figure 1 A heterogeneous domain adaptation learning method for cross-network heterogeneous node classification is shown, comprising the following steps:

[0091] S1. Construct the PPMI matrix of the source network and the target network to obtain the global structural information of the network, and use the variational graph autoencoder to obtain the node features of the source network and the target network, decode them 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 two-layer normalized graph convolutional network is used as the X-based s and P s Global information reasoning model:

[0094]

[0095] Where X s is the source network node, P s is the source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node features, for The i-th vector of

[0096]

[0097] where μ s =φ s (X s ,P s ) is the mean vector, μ s,i is its i-th vector, is the variance matrix, σ s,i is its i-th vector; where φ s (·)and represents a two-layer normalized graph convolutional network, defined as:

[0098]

[0099] in is the symmetric normalized adjacency matrix, Φ sP s 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. Using the inner product between potential representations as a generative model:

[0101]

[0102] in is to reconstruct the 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:

[0104]

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

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

[0107]

[0108] Where X t is the target network node, P t PPMI for the target network, is the reconstructed target PPMI matrix, is the learned potential target node feature, is a 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. Build a hypergraph to more closely associate the source network and the target network using labeled source and target nodes.

[0112] Specifically, step S2 includes:

[0113] S2.1. Given the node labels of the source network and part of the target network, construct a hypergraph to represent the high-order relationships between these nodes:

[0114] use 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:

[0115]

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

[0117] S3. Build a hypergraph autoencoder to obtain high-order features of source and target network nodes, and use the learned features to reconstruct the hypergraph to preserve high-order relationships.

[0118] Specifically, step S3 includes:

[0119] S3.1. For the source network, two layers of normalized hypergraph convolution operation g are used s (·),as follows:

[0120]

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

[0122]

[0123] Where Ω s is a diagonal matrix that stores the positive weights of hyperedges,

[0124]

[0125] and

[0126]

[0127] They are all diagonal matrices, which help avoid the problems of gradient vanishing and numerical instability during convolution operations;

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

[0129]

[0130] in To mark the target data features, 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 the nodes,

[0133]

[0134] and

[0135]

[0136] They are all angle matrices;

[0137] S3.3. Preserve the high-order relations between the labeled source and target nodes by reconstructing the decoder h(·) of the hypergraph:

[0138]

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

[0140] S3.4. Therefore, the hypergraph reconstruction loss function is:

[0141]

[0142] S4. A feature matching mechanism is used to minimize the differences between different features of each labeled node and obtain a feature representation that is rich in node information and consistent.

[0143] Specifically, step S4 includes:

[0144] S4.1. In order to obtain information-rich and consistent feature representations of source and target network nodes, the following feature matching mechanism is adopted:

[0145]

[0146] in is the mean square error function. Therefore, the difference between the two feature spaces of the labeled source and target nodes can be reduced by minimizing this formula, thereby obtaining an informative and consistent feature representation.

[0147] S5. Adopt an adversarial domain adaptation module to reduce the domain differences between the labeled nodes of the source and target networks.

[0148] Specifically, step S5 includes:

[0149] S5.1. To further mitigate the negative impact of domain shift, adversarial domain adaptation is adopted to jointly train effective classifiers and align the distributions of labeled source and target nodes.

[0150] Use f y (·) and f d (·) as a node classifier and a domain classifier. Definition Its node labels and domain labels are and Note that Y s is the one-hot label matrix of the nodes, is the one-hot labeling matrix of the labeled target node, and are 0 and 1 vectors, respectively, with dimensions N s and have:

[0151]

[0152] in and are the losses of the node label classifier and the domain classifier, respectively. and is the corresponding loss function 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.

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

[0154]

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

[0156] Specifically, step S6 includes:

[0157] S6.1. Based on the above steps, the overall objective function of the model is:

[0158]

[0159] Among them, α, β and η are equilibrium parameters;

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

[0161] S6.2. The performance of the proposed method is evaluated 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 associated with label c.

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

[0163]

[0164] Macro-F1 assigns equal weight to each category and is defined as follows:

[0165]

[0166] The effectiveness of the proposed method is evaluated on three real datasets provided by ArnetMine: ACMv9 (A), Citationv1 (C) and DBLPv7 (D). A, C and D are three citation networks from ACM, Microsoft Academic Graph and DBLP respectively. The citation network is modeled as an undirected network, where nodes represent papers and edges represent citation relationships. Each paper belongs to one of the following. The attributes of each node are sparse bag-of-words features 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 the marked target nodes, and the remaining target data is used for testing.

[0167] Table 1 lists the statistics of these datasets.

[0168] Table 1

[0169]

[0170] The Micro-F1 and MacroF1 scores of the present invention on the dataset are compared with the current most advanced heterogeneous domain adaptation methods DDA (Yao, Y., Zhang, Y., Li, X., Ye, Y.: Discriminative distribution alignment: Aunified framework for heterogeneous domain adaptation. Pattern Recognition 101, 107165 (2020)), SSAN (Li, S., Xie, B., Wu, J., Zhao, Y., Liu, CH, Ding, Z.: Simultaneous semantic alignment network for heterogeneous domain adaptation. In: Proceedings of the 28th ACM International Conference on Multimedia, pp. 3866–3874 (2020)), and CDSPP (Wang, Q., Breckon, TP: 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 result of the algorithm proposed in this invention; Algorithm 2 corresponds to the verification result of DDA; Algorithm 3 corresponds to the verification result of SSAN; Algorithm 4 corresponds to the verification result of CDSPP. It can be seen from Table 2 that the Micro-F1 and Macro-F1 score results of this invention are always better than those of other algorithms.

[0172] Table 2

[0173]

[0174]

[0175] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A heterogeneous domain adaptation learning method for cross-network heterogeneous node classification, characterized in that: The following steps are involved: S1. Construct the PPMI matrix of the source network and the target network to obtain the global structural information of the network, and use the variational graph autoencoder to obtain the node features of the source network and the target network, decode them to restore the original PPMI matrix to retain the global proximity information between the nodes; S2, build a hypergraph, using the labeled source and target nodes to more closely associate the source network and the target network; S3. Build a hypergraph autoencoder to obtain high-order features of source and target network nodes, and use the learned features to reconstruct the hypergraph to preserve high-order relationships. S4, adopt feature matching mechanism to minimize the differences between different features of each marked node and obtain feature representation with rich and consistent node information; S5, using adversarial domain adaptation module to reduce the domain differences between the labeled nodes of the source network and the target network; S6. Obtain the overall objective function and train the classifier based on the labeled source node and target node features.

2. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S1 includes: S1.

1. For the source network, a two-layer normalized graph convolutional network is used as the X-based s and P s Global information reasoning model: Where X s is the source network node, P s is the source network PPMI matrix, N s is the number of source network nodes, is the learned latent source node features, for The i-th vector of where μ s =φ s (X s ,P s ) is the mean vector, μ s,i is its i-th vector, is the variance matrix, σ s,i is its i-th vector; where φ s (·)and represents a two-layer normalized graph convolutional network, defined as: in is the symmetric normalized adjacency matrix, Φ s P s 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; S1.

2. Using the inner product between potential representations as a generative model: in is to reconstruct the source PPMI matrix, sigmoid(·) is the logistic sigmoid function; Therefore, the loss function of the variational graph autoencoder on the source network is obtained as: where KL[q s (·)||p s (·)] is q s (·) and p s (·) is the Kullback-Leibler divergence between is a Gaussian prior; S1.

3. For the target network, let q t (·) and p t (·) are global information inference model and generation model respectively, based on X t and P t , the loss function is as follows: Where X t is the target network node, P t For the target network PPMI, array is the reconstructed target PPMI matrix, is the learned potential target node feature, is a Gaussian prior; the first item uses cross entropy to measure the gap between the reconstructed data and the original data, and the second item uses KL degree to measure the gap between the pre-set prior distribution and the posterior distribution generated by the network; S1.

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

3. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S2 includes: S2.

1. Given the node labels of the source network and part of the target network, construct a hypergraph to represent the high-order relationships between these nodes: use Represents a hypergraph, where each entry H(i,j) indicates whether the i-th source node and the i-th target node belong to the same category: Where N s is the number of source network nodes, is the number of labeled target samples.

4. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S3 includes: S3.

1. For the source network, a two-layer normalized hypergraph convolutional network g is used s (·),as follows: in is the learned source node feature, W s,1 and W s,2 are the network parameters of the first and second layers respectively, Where Ω s is a diagonal matrix that stores the positive weights of hyperedges, where and They are all diagonal matrices, which help avoid the problems of gradient vanishing and numerical instability during convolution operations; S3.2, similarly, the target network g t The two-layer normalized hypergraph convolution function of (·) is expressed as: in To mark the target data features, 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, Where Ω t is a diagonal matrix storing the positive weights of the nodes, and They are all angle matrices; S3.

3. Preserve the high-order relations between the labeled source and target nodes by reconstructing the decoder h(·) of the hypergraph: in is the reconstructed hypergraph, and They are and The i-th and j-th vectors in ; S3.

4. Therefore, the hypergraph reconstruction loss function is:

5. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S4 includes: S4.

1. In order to obtain information-rich and consistent feature representations of source and target network nodes, the following feature matching mechanism is adopted: where l(·) is the mean square error function; therefore, the difference between the two feature spaces of the labeled source and target nodes can be reduced by minimizing this formula, thereby obtaining an informative and consistent feature representation.

6. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S5 includes: S5.

1. To further mitigate the negative impact of domain shift, adversarial domain adaptation is used to jointly train effective classifiers and align the distribution of labeled source and target nodes. Use f y (·) and f d (·) as node classifier and domain classifier; definition Its node labels and domain labels are and Note that Y s is the one-hot label matrix of the nodes, is the one-hot labeling matrix of the labeled target node, and are 0 and 1 vectors, respectively, with dimensions N s and have: in and are the losses of the node label classifier and the domain classifier, respectively. and is the corresponding loss function evaluated at the i-th node; here, γ is a balancing 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; S5.

2. Minimization To learn the node label invariant features, the classifier f y (·) can correctly distinguish different nodes into their respective categories; on the other hand, maximizing Deceiving domain classifiers by generating node features that are indistinguishable across networks d (·); by adding a gradient reversal layer f r (·),Right now The above formula can be restated as:

7. According to the heterogeneous domain adaptive learning method for cross-network heterogeneous node classification according to claim 1, the above step S6 comprises: S6.

1. Based on the above steps, the overall objective function of the model is: Among them, α, β and η are equilibrium parameters; Train a classifier based on the labeled source and target node features; S6.2, the performance of the proposed method is evaluated 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 associated with label c; Micro-F1 assigns the same weight to each sample, which is defined as follows: Macro-F1 assigns equal weight to each category and is defined as follows:

Images