A multi-domain data classification method and system based on optimal transmission

Abstract

The application relates to the technical field of multi-domain data interaction, and discloses a multi-domain data classification method and system based on optimal transmission, which comprises the following specific steps: acquiring data with attributes in different domains; preprocessing the data with attributes to construct graph data; considering the cost and constraint conditions of intra-domain and inter-domain transportation, setting an optimal transmission target function, and obtaining an optimal inter-domain sample transmission scheme; merging the domains by using the optimal inter-domain sample transmission scheme; inputting the merged domains into a graph neural network to obtain node representation; specifically, after the merged domains are input into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and the original attributes of the merged domains into the graph neural network to obtain node representation; and based on the node representation, a classification result is obtained. The application solves the problem that existing classification technology ignores the structural information in edges, and has the characteristics of being capable of improving data representation learning efficiency and classification accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of multi-domain data interaction, and more specifically, to a multi-domain data classification method and system based on optimal transmission. Background Technology

[0002] Most existing graph transfer learning methods are based on domain adaptation. These methods have achieved certain results in the field of graph transfer learning, but they still have some shortcomings. These methods do not consider the potential relationships between samples in different domains and cannot encode effective structural information. Furthermore, the lack of structural information makes it impossible to fully utilize the rich attribute data, which affects the learning of node representations and consequently affects the accuracy of data prediction.

[0003] A graph transfer learning method based on fine-grained domain adaptation is proposed, comprising: collecting samples from a source domain and a target domain, and labeling some or all samples in the source and target domains respectively to obtain labeled samples; constructing source domain graphs and target domain graphs for the labeled samples respectively; dividing the labeled samples in the source domain graph into training and validation sets, and treating the labeled samples in the target domain graph as the test set; training graph neural networks using the training set of the source domain graph and the target domain samples to obtain at least two graph neural networks with different parameters; selecting a graph neural network using the validation set of the source domain graph; predicting labels for the samples in the target domain using the selected graph neural network; and obtaining the evaluation result by comparing the true labels and predicted labels of all labeled samples in the target domain. The principle of this method is still based on traditional domain adaptation methods, which improve classification performance by reducing the data distribution differences between the two networks and transferring label information from the existing network to the new network. Graphs are a special type of data, containing structural information in the edges in addition to node attributes. This information is crucial for node representation learning and classification, a problem that traditional methods struggle to solve.

[0004] In summary, existing technologies have the problem of ignoring structural information in edges. Therefore, how to invent a multi-domain data classification method that takes edge structure into account is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the problem that existing classification techniques neglect structural information in edges, this invention provides a multi-domain data classification method and system based on optimal transmission, which can improve data representation learning efficiency and classification accuracy.

[0006] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0007] A multi-domain data classification method based on optimal transmission includes the following specific steps:

[0008] S1: Retrieve data with attributes from different domains;

[0009] S2: Preprocess the attributed data to construct the graph data;

[0010] S3: Considering the costs and constraints of intra-domain and inter-domain transport, set the optimal transport objective function and obtain the optimal transport scheme for inter-domain samples;

[0011] S4: Merge domains using the optimal inter-domain sample transfer scheme;

[0012] S5: Input the merged domain into the graph neural network to obtain node representations;

[0013] Specifically, by inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix of the merged domain and the original attributes into the graph neural network to obtain node representations;

[0014] S6: Obtain classification results based on node representation.

[0015] Preferably, in step S1, the attributed data is divided into a source domain and a target domain based on the presence or absence of tag information.

[0016] Furthermore, in step S2, the data with attributes is preprocessed to construct the graph data. The specific method is as follows:

[0017] Treat each sample in the attributed data as a node;

[0018] If there is a natural pointing relationship between samples, then use it as the basis to connect the two points as an edge; otherwise, calculate the distance between sample attributes, use it as the weight of the edge between the two points, and finally store the edge in the adjacency matrix.

[0019] Furthermore, in step S3, considering the costs and constraints of intra-domain and inter-domain transport, an optimal transmission objective function is set to obtain the optimal inter-domain sample transmission scheme. The specific steps are as follows:

[0020] S31: Set the optimal transmission target function:

[0021] The cost and constraints of moving data within and between domains are taken as input;

[0022] The specific transportation costs are as follows:

[0023]

[0024] Among them, C U Let C(·,·) represent the transportation cost within domain U, and let C(·,·) represent the distance metric between samples, defined on the graph as the shortest path between two nodes. Let represent the i-th sample within domain U; ​​M represents the inter-domain transport cost; and d(·,·) represents the distance metric between attributes, defined as the Euclidean distance between two attributes. Let represent the attribute of the i-th sample in the source domain S. This represents the attribute of the j-th sample within the target domain T;

[0025] The specific constraints are as follows:

[0026] μS = π1, μT = π T 1

[0027] Where μU represents the contribution of samples within domain U, π represents the optimal transmission scheme, and (·) T This represents the transpose of a matrix, where 1 represents a column vector with all elements equal to 1.

[0028] The objective function is as follows:

[0029]

[0030] Where π represents the transmission scheme, π* represents the optimal value of the transmission scheme, <·, ·> represent the inner product between inputs, α represents the adjustment parameter, and ||·|| represents the L2 norm;

[0031] S32: Optimize the objective function to obtain the optimal transmission scheme:

[0032] At step i, calculate the gradient of the objective function with respect to step i-1:

[0033]

[0034] Solving linear programming problems:

[0035]

[0036] The weights are obtained through line search, and the transmission scheme is updated accordingly.

[0037]

[0038] Where, π (i) For the transmission scheme in step i, τ (i) Let be the weight for the i-th step.

[0039] Furthermore, in step S4, the domains are merged using the optimal inter-domain sample transfer scheme. The specific method is as follows:

[0040] S41: Utilize the optimal transmission scheme to transfer attributes from the source domain to the target domain, serving as the optimal representation of the source domain data in the target domain, and coordinating with the target domain sample attributes:

[0041] The specific source domain attributes to be moved are:

[0042]

[0043] in, To transfer the attributes of the source domain S sample, use diag(·). -1 This represents retaining the diagonal elements of the matrix and finding its inverse, where π represents the optimal transmission scheme, 1 represents a column vector with all elements equal to 1, and X represents... T The attributes of the target domain T sample;

[0044] The specific steps for merging source and target domain attributes are as follows:

[0045]

[0046] in, The attributes of all samples after merging;

[0047] S42: Utilize the optimal transmission scheme to transfer the structure from the source domain to the target domain, serving as the optimal representation of the source domain data in the target domain, and coordinating the structure with the target domain samples:

[0048] The specific structure of the transport source domain is as follows:

[0049]

[0050] in, Let μ be the adjacency matrix of the source domain S samples after transportation. S Let π represent the contribution of samples within the source domain S, where π is the optimal transmission scheme, and A is the mean. T Let T be the adjacency matrix of the samples in the target domain T, (·) T Represents the transpose of a matrix;

[0051] The specific steps for merging the source and target domain structures are as follows:

[0052]

[0053] in, This is the adjacency matrix among all samples after merging.

[0054] Furthermore, in step S5, after inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and original attributes of the merged domain into the graph neural network to obtain node representations, specifically:

[0055] S51: Input the merged domain into the graph neural network. Each layer of the graph neural network will process each node... v From its neighboring nodes Receive message:

[0056]

[0057] Where AGGREGATE() is an aggregate function, For the neighbor messages of node v in the (k+1)th iteration, Let i be the representation of node i in the k-th iteration;

[0058] S52: Each node v updates its characteristics using the received message:

[0059]

[0060] Among them, UPDATE() is the update function;

[0061] After several layers, the final node representation is obtained.

[0062] Furthermore, in step S6, the classification result is obtained based on the node representation, specifically as follows:

[0063] S61: Based on node representation, a classification algorithm is used to train a classifier, and the trained classifier is used to classify the nodes and edges to be classified to obtain the classification results;

[0064] S62: Use cross-validation to validate the classification results of the classifier.

[0065] Furthermore, the aggregate function AGGREGATE() is specifically one or more functions for summation, averaging, or maximizing.

[0066] Furthermore, the update function UPDATE() is specifically one or more functions that calculate the sum, average, or maximum value.

[0067] A multi-domain data classification system based on optimal transmission includes the following modules: a data acquisition module, a preprocessing module, a relationship discovery module, a collaboration module, a graph neural network module, and a classification module;

[0068] The data acquisition module is used to acquire attributed data from different domains;

[0069] The preprocessing module is used to preprocess attributed data and construct graph data;

[0070] The relationship discovery module is used to consider the cost and constraints of intra-domain and inter-domain transport, set the optimal transport objective function, and obtain the optimal transport scheme for inter-domain samples.

[0071] The aforementioned collaborative module is used to merge domains using the optimal inter-domain sample transmission scheme;

[0072] The graph neural network module is used to input the merged domain into the graph neural network to obtain node representations; specifically, after the merged domain is input into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and original attributes of the merged domain into the graph neural network to obtain node representations.

[0073] The classification module is used to obtain classification results based on node representation.

[0074] The beneficial effects of this invention are as follows:

[0075] This invention processes data to construct graph data; optimal transmission can describe the relationship between attributes between samples and uncover potential geometric connections; combined with prior knowledge, optimal transmission is used to mine the relationship between samples, and the transmission plan is used to merge domains, so that all data can be processed as a whole, improving the efficiency of data representation learning; graph neural networks are used to encode attributes and structural information to obtain good node representations; based on the node representations, the attribute data to be classified is classified, and the classification results obtained are more accurate. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating a multi-domain data classification method based on optimal transmission according to the present invention.

[0077] Figure 2 This is a schematic diagram of a multi-domain data classification system based on optimal transmission according to the present invention. Detailed Implementation

[0078] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0079] Example 1

[0080] like Figure 1 As shown, a multi-domain data classification method based on optimal transmission includes the following specific steps:

[0081] S1: Retrieve data with attributes from different domains;

[0082] S2: Preprocess the attributed data to construct the graph data;

[0083] S3: Considering the costs and constraints of intra-domain and inter-domain transport, set the optimal transport objective function and obtain the optimal transport scheme for inter-domain samples;

[0084] S4: Merge domains using the optimal inter-domain sample transfer scheme;

[0085] S5: Input the merged domain into the graph neural network to obtain node representations;

[0086] Specifically, by inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix of the merged domain and the original attributes into the graph neural network to obtain node representations;

[0087] S6: Obtain classification results based on node representation.

[0088] Example 2

[0089] More specifically, in one embodiment, in step S1, the attributed data is divided into a source domain and a target domain based on the presence or absence of tag information.

[0090] Furthermore, in step S2, the data with attributes is preprocessed to construct the graph data. The specific method is as follows:

[0091] Treat each sample in the attributed data as a node;

[0092] If there is a natural pointing relationship between samples, then use it as the basis to connect the two points as an edge; otherwise, calculate the distance between sample attributes, use it as the weight of the edge between the two points, and finally store the edge in the adjacency matrix.

[0093] In this embodiment, the distance between sample attributes is specifically the cosine distance.

[0094] In one specific embodiment, step S3 involves considering the costs and constraints of intra-domain and inter-domain transport, setting an optimal transmission objective function, and obtaining an optimal inter-domain sample transmission scheme. The specific steps are as follows:

[0095] S31: Set the optimal transmission target function:

[0096] The cost and constraints of moving data within and between domains are taken as input;

[0097] The specific transportation costs are as follows:

[0098]

[0099] Among them, C U Let C(·,·) represent the transportation cost within domain U, and let C(·,·) represent the distance metric between samples, defined on the graph as the shortest path between two nodes. Let represent the i-th sample within domain U; ​​M represents the inter-domain transport cost; and d(·,·) represents the distance metric between attributes, defined as the Euclidean distance between two attributes. Let represent the attribute of the i-th sample in the source domain S. This represents the attribute of the j-th sample within the target domain T;

[0100] The specific constraints are as follows:

[0101] μS = π1, μT = π T 1

[0102] Where μU represents the contribution of samples within domain U, π represents the optimal transmission scheme, and (·) T This represents the transpose of a matrix, where 1 represents a column vector with all elements equal to 1.

[0103] The objective function is as follows:

[0104]

[0105] Where π represents the transmission scheme, π* represents the optimal value of the transmission scheme, <·, ·> represent the inner product between inputs, α represents the adjustment parameter, and ||·|| represents the L2 norm;

[0106] In this embodiment, the objective function is adopted, which takes into account both the attribute differences between samples and the structural relationships between samples, thereby learning more accurate relationships between samples. In addition, by adjusting the weight coefficients, the tendency of attributes and structures is balanced to flexibly adapt to different needs of different data.

[0107] S32: Optimize the objective function to obtain the optimal transmission scheme:

[0108] At step i, calculate the gradient of the objective function with respect to step i-1:

[0109]

[0110] Solving linear programming problems:

[0111]

[0112] The weights are obtained through line search, and the transmission scheme is updated accordingly.

[0113]

[0114] Where, π (i) For the transmission scheme in step i, τ (i) Let be the weight for the i-th step.

[0115] In this embodiment, by using the optimization objective function, only a small number of non-zero elements are added in each iteration, which can quickly converge and obtain a sparse solution, so that the sample will not find too many samples in another domain, thus avoiding the problem of excessive smoothing when learning the representation in the later stage, that is, excessive similarity between the sample representations.

[0116] In one specific embodiment, in step S4, the domains are merged using the optimal inter-domain sample transfer scheme. The specific method is as follows:

[0117] S41: Utilize the optimal transmission scheme to transfer attributes from the source domain to the target domain, serving as the optimal representation of the source domain data in the target domain, and coordinating with the target domain sample attributes:

[0118] The specific source domain attributes to be moved are:

[0119]

[0120] in, To transfer the attributes of the source domain S sample, use diag(·). -1 This represents retaining the diagonal elements of the matrix and finding its inverse, where π represents the optimal transmission scheme, 1 represents a column vector with all elements equal to 1, and X represents... T The attributes of the target domain T sample;

[0121] The specific steps for merging source and target domain attributes are as follows:

[0122]

[0123] in, The attributes of all samples after merging;

[0124] S42: Utilize the optimal transmission scheme to transfer the structure from the source domain to the target domain, serving as the optimal representation of the source domain data in the target domain, and coordinating the structure with the target domain samples:

[0125] The specific structure of the transport source domain is as follows:

[0126]

[0127] in, Let μ be the adjacency matrix of the source domain S samples after transportation. S Let π represent the contribution of samples within the source domain S, where π is the optimal transmission scheme, and A is the mean. T Let T be the adjacency matrix of the samples in the target domain T, (·) T Represents the transpose of a matrix;

[0128] The specific steps for merging the source and target domain structures are as follows:

[0129]

[0130] in, This is the adjacency matrix among all samples after merging.

[0131] In this embodiment, similarly, the optimal transmission scheme is used to transport the structure, unifying the structure between samples in the source and target domains. The purpose of merging domains is to enable samples to utilize the rich information in other domains and fully explore the relationships between samples. The premise of merging is that the differences between the data in the domains are as similar as possible to those before merging. Through the optimal transmission, the data in one domain can be "embedded" into the data in another domain, becoming a whole.

[0132] In one specific embodiment, in step S5, after inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and original attributes of the merged domain into the graph neural network to obtain node representations, specifically:

[0133] S51: Input the merged domain into the graph neural network. Each layer of the graph neural network will process each node... v From its neighboring nodes Receive message:

[0134]

[0135] Where AGGREGATE() is an aggregate function, For the neighbor messages of node v in the (k+1)th iteration, Let i be the representation of node i in the k-th iteration;

[0136] S52: Each node v updates its characteristics using the received message:

[0137]

[0138] Among them, UPDATE() is the update function;

[0139] After several layers, the final node representation is obtained.

[0140] In one specific embodiment, in step S6, the classification result is obtained based on the node representation, specifically as follows:

[0141] S61: Based on node representation, a classification algorithm is used to train a classifier, and the trained classifier is used to classify the nodes and edges to be classified to obtain the classification results;

[0142] S62: Cross-validation is used to validate the classification results of the classifier. The purpose of cross-validation is to prevent overfitting.

[0143] In one specific embodiment, the aggregation function AGGREGATE() is specifically one or more functions for summation, averaging, or maximization.

[0144] In one specific embodiment, the update function UPDATE() is specifically one or more functions such as summation, averaging, or maximization.

[0145] Example 3

[0146] like Figure 2 As shown, a multi-domain data classification system based on optimal transmission includes the following modules: data acquisition module, preprocessing module, relationship discovery module, collaboration module, graph neural network module, and classification module;

[0147] The data acquisition module is used to acquire attributed data from different domains;

[0148] The preprocessing module is used to preprocess attributed data and construct graph data;

[0149] The relationship discovery module is used to consider the cost and constraints of intra-domain and inter-domain transport, set the optimal transport objective function, and obtain the optimal transport scheme for inter-domain samples.

[0150] The aforementioned collaborative module is used to merge domains using the optimal inter-domain sample transmission scheme;

[0151] The graph neural network module is used to input the merged domain into the graph neural network to obtain node representations; specifically, after the merged domain is input into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and original attributes of the merged domain into the graph neural network to obtain node representations.

[0152] The classification module is used to obtain classification results based on node representation.

[0153] The relationship discovery module includes a first-step unit and a second-step unit;

[0154] The first unit is used to set the optimal transmission objective function, defining its input and mathematical expression:

[0155] The cost and constraints of data transfer within and between domains are taken as input; the transfer cost is:

[0156]

[0157] Among them, C U Let C(·,·) represent the transportation cost within domain U, and let C(·,·) represent the distance metric between samples, defined on the graph as the shortest path between two nodes. Let represent the i-th sample within domain U; ​​M represents the inter-domain transport cost; and d(·,·) represents the distance metric between attributes, defined as the Euclidean distance between two attributes. Let represent the attribute of the i-th sample within domain S. Represents the attribute of the j-th sample within domain T;

[0158] The constraints are:

[0159] μS = T1, μT = T T 1

[0160] Where μU represents the contribution of samples within domain U, T represents the optimal transmission scheme, and (·) T This represents the transpose of a matrix, where 1 represents a column vector with all elements equal to 1.

[0161] The objective function is:

[0162]

[0163] Where π represents the transmission scheme, π* represents the optimal value of the transmission scheme, <·, ·> represent the inner product between inputs, α represents the adjustment parameter, and ||·|| represents the L2 norm;

[0164] Used to optimize the objective function and obtain the optimal transmission scheme:

[0165] At step i, calculate the gradient of the objective function with respect to step i-1:

[0166]

[0167] Solving linear programming problems:

[0168]

[0169] The weights are obtained through line search, and the transmission scheme is updated accordingly.

[0170]

[0171] Where, τ( i ) represents the weight at step i. π ( i Let be the transmission scheme for step i.

[0172] The collaborative module includes attribute units and structural units;

[0173] Attribute Unit: Utilizing the optimal transmission scheme, attributes are transferred from the source domain to the target domain, serving as the optimal representation of source domain data in the target domain, and working in conjunction with target domain sample attributes.

[0174] First, transfer the source domain attributes:

[0175]

[0176] in, To transfer the attributes of the source domain S sample, dia.g(·) -1This represents retaining the diagonal elements of the matrix and finding its inverse, where π represents the optimal transmission scheme, 1 represents a column vector with all elements equal to 1, and X represents... T The attributes of the target domain T sample;

[0177] Merge source and target domain attributes:

[0178]

[0179] in, The attributes of all samples after merging;

[0180] Structural Unit: Utilizing the optimal transmission scheme, the structure is transferred from the source domain to the target domain, serving as the optimal representation of the source domain data in the target domain, and coordinating with the structural elements of the target domain samples.

[0181] First, move the source domain structure:

[0182]

[0183] in, Let μ be the adjacency matrix of the source domain S samples after transportation. S Let π represent the contribution of samples within the source domain S, where π is the optimal transmission scheme, and A is the mean. T Let T be the adjacency matrix of the samples in the target domain T, (·) T Represents the transpose of a matrix;

[0184] Merge the source and target domain structures:

[0185]

[0186] in, This is the adjacency matrix among all samples after merging.

[0187] The classification module includes a classification unit and a verification unit;

[0188] The classification unit, based on node representation, uses a classification algorithm to train a classifier, and then uses the trained classifier to classify the nodes and edges to be classified, thereby obtaining the classification result.

[0189] The validation unit uses cross-validation to validate the classification results of the classifier.

[0190] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A multi-domain data classification method based on optimal transport, characterized in that: Comprise the following specific steps: S1: acquire data with attribute in different domains; S2: pre-process the data with attribute, and construct graph data; S3: consider the cost and constraint conditions of intra-domain and inter-domain transfer, set the optimal transmission objective function, and obtain the optimal inter-domain sample transmission scheme, which comprises the following specific steps: S31: set the optimal transmission objective function: Take the cost and constraint conditions of intra-domain and inter-domain transfer as input; The cost of transfer is specifically: in, Representation domain The cost of internal transportation This represents the distance metric between samples, defined on a graph as the shortest path between two nodes. Representation domain The first i One sample; This represents the cost of inter-domain data transfer. This represents the distance metric between attributes, defined as the Euclidean distance between two attributes. Represents the source domain Inner i The attributes of each sample Represents the target domain Inner j The attributes of a sample; The constraint condition is specifically: wherein, representation domain contribution of the sample within the representation domain, represents an optimal transmission scheme, represents the transpose of a matrix, 1 is a column vector with all elements being 1. The objective function is specifically: wherein, denotes a transmission scheme, denotes an optimal value of a transmission scheme, denotes an inner product between inputs, denotes a tuning parameter, denotes an L2 norm; S32: optimize the objective function to obtain the optimal transmission scheme: At the first i step, the gradient of the objective function with respect to the first i-1 step is computed: Linear programming is used to solve the problem: Linear search is used to obtain the weight, and the transmission scheme is updated: wherein, is the first i step transmission scheme, is the first i step weight; S4: merge the domains using the optimal inter-domain sample transmission scheme, specifically: S41: use the optimal transmission scheme to transfer the attribute from the source domain to the target domain as the optimal representation of the source domain data in the target domain, and coordinate with the sample attribute of the target domain: The specific method of transferring the attribute of the source domain is: wherein, is the post-hoc source domain attributes of the sample, denotes the diagonal elements of the reserved matrix and inverts, is the optimal transport scheme, 1 is a column vector with all element values being 1, is the target domain attributes of the sample; The specific method of merging the attribute of the source domain and the target domain is: wherein, is the attribute of all samples after merging; S42: use the optimal transmission scheme to transfer the structure from the source domain to the target domain as the optimal representation of the source domain data in the target domain, and coordinate with the structure between the samples of the target domain: The specific method of transferring the structure of the source domain is: wherein, is the source domain after transfer is the adjacency matrix of the samples, is the source domain is the contribution of the samples, is the optimal transfer scheme, is the target domain is the adjacency matrix of the samples, denotes the transpose of a matrix; The specific method of merging the structure of the source domain and the target domain is: wherein is the adjacency matrix between all samples after merging; S5: input the merged domain into the graph neural network to obtain the node representation; Specifically, after inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and the original attribute of the merged domain into the graph neural network to obtain the node representation. S6: obtain the classification result based on the node representation.

2. The method of claim 1, wherein: In S1, the data with attribute is divided into source domain and target domain according to the presence or absence of label information.

3. The method of claim 2, wherein: In S2, the specific method of pre-processing the data with attribute to construct graph data is: Each sample in the data with attribute is taken as a node; If there is a natural directed relationship between samples, it is taken as a basis to connect two points as an edge; otherwise, the distance between sample attributes is calculated as the weight of the edge between two points, and finally the edge is stored in the adjacency matrix.

4. The method of claim 1, wherein: In step S5, after inputting the merged domain into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and the original attribute of the merged domain into the graph neural network to obtain the node representation, specifically: S51: input the merged domains into a graph neural network, each layer of the graph neural network will each node v receive messages from its neighbor nodes receive messages: wherein, is a polymerization function, is a node v in the first k+1 iteration of neighbor messages, is a node i in the first k iteration of representations; S52: Each node v Update its features using the received messages: wherein is an update function; After several layers, the final node representation is obtained.

5. The method of claim 4, wherein: In S6, based on the node representation, the classification result is obtained, specifically: S61: based on the node representation, the classifier is trained using a classification algorithm, and the trained classifier is used to classify the nodes and edges to be classified to obtain the classification result; S62: the classification result of the classifier is verified by cross-validation method.

6. The method of claim 4, wherein: The aggregation function Specifically, any one or more of a sum, average, or maximum value function.

7. The method of claim 4, wherein: The update function Specifically, any one or more of a sum, average, or maximum value function.

8. A multi-domain data classification system based on optimal transport, characterized by: For implementing the method of any one of claims 1-7, comprising the following modules: data acquisition module, preprocessing module, relationship discovery module, coordination module, graph neural network module, classification module; The data acquisition module is used to acquire data with attribute in different domains; The preprocessing module is used to pre-process the data with attribute to construct graph data; The relationship discovery module is used to consider the cost and constraint conditions of inter-domain transfer, set an optimal transmission objective function, and obtain an inter-domain sample optimal transmission scheme; The cooperation module is used to merge the domains by using the inter-domain sample optimal transmission scheme; The graph neural network module is used to input the merged domains into a graph neural network to obtain node representations; specifically, after the merged domains are input into the graph neural network, each layer of the graph neural network inputs the adjacency matrix and the original attributes of the merged domains into the graph neural network to obtain the node representations; The classification module is used to obtain a classification result based on the node representations.

Images