Cross-domain data mapping method based on graph neural network

Abstract

The invention discloses a cross-domain data mapping method based on a graph neural network, and the method comprises the following steps: obtaining data of a source domain and a target domain, and respectively constructing a source domain graph structure and a target domain graph structure; processing the two types of graph structures to generate a corresponding initial node feature matrix and an initial adjacent matrix; introducing an improved DiffPool, performing cross-domain mutual solution pooling structure generation processing on the initial node features of the source domain and the target domain and the adjacent matrix, and generating a first pooling structure pooling the source domain and a second pooling structure pooling the target domain; respectively generating a source domain pooling graph and a target domain pooling graph according to the pooling structure; performing graph feature coding processing on the two types of pooling graphs to obtain source domain and target domain feature representation of a unified representation space; and establishing a cross-domain mapping relation model based on the feature representation to obtain a mapping function from the source domain to the target domain. According to the method, unified alignment of the source domain and the target domain can be realized on node features and structure levels, and the accuracy and adaptability of cross-domain mapping are improved.

Description

Technical Field

[0001] This invention relates to the field of data mapping technology, and in particular to a cross-domain data mapping method based on graph neural networks. Background Technology

[0002] Most existing cross-domain data mapping techniques rely on feature-level similarity measurement, statistical alignment, or deep learning-based embedding space alignment methods. However, when processing graph-structured data, they often struggle to maintain consistency in node features, inter-node relationships, and hierarchical information. Existing graph neural networks typically encode only the source and target domains separately in cross-domain tasks, lacking a unified processing mechanism for differences in node arrangement, adjacency structure, and local and global structural relationships. This leads to problems such as feature misalignment, structural mismatch, or representation space offset during the cross-domain mapping process.

[0003] In particular, traditional graph pooling methods, such as the standard DiffPool, can only perform node aggregation within a single domain. They cannot perform cross-domain mutual decomposition of node order, structural relationships, and pooling strategies between the source and target domains. It is difficult to establish structural consistency and feature correspondence between the source and target domains. When facing cross-domain scenarios with large structural differences, the above methods lack node alignment capabilities, structural relationship unification capabilities, and cross-domain pooling allocation capabilities. They cannot generate reusable cross-domain pooling structures, resulting in insufficient generalization and low mapping accuracy of the final cross-domain mapping model. Therefore, there is an urgent need for a new cross-domain data mapping method that can achieve cross-domain mutual decomposition at the pooling level, achieve unified space generation at the feature representation level, and establish accurate correspondence at the mapping level. Summary of the Invention

[0004] One objective of this invention is to propose a cross-domain data mapping method based on graph neural networks. This invention adopts an improved DiffPool cross-domain graph mapping method to achieve unified alignment of the structure and features of the source domain and the target domain, and has the advantages of high mapping accuracy and strong adaptability.

[0005] A cross-domain data mapping method based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0006] Acquire source domain data and target domain data, and construct source domain graph structure and target domain graph structure;

[0007] Graph structure encoding is performed on the source domain graph structure and the target domain graph structure respectively to generate the initial node feature matrix and the initial adjacency matrix of the source domain and the target domain;

[0008] An improved differentiable graph pooling network, DiffPool, is used to perform cross-domain mutual pooling structure generation on the initial node feature matrices and initial adjacency matrices of the source and target domains. The first pooling structure of the source domain is generated based on the initial node feature matrix and initial adjacency matrix of the target domain, and the second pooling structure of the target domain is generated based on the initial node feature matrix and initial adjacency matrix of the source domain.

[0009] The first pooling structure is used to perform pooling operations on the initial node feature matrix and the initial adjacency matrix of the source domain to generate the source domain pooling graph. The second pooling structure is used to perform pooling operations on the initial node feature matrix and the initial adjacency matrix of the target domain to generate the target domain pooling graph.

[0010] Graph feature encoding is performed on the source domain pooling graph and the target domain pooling graph respectively to generate source domain feature representations and target domain specific representations located in the same representation space;

[0011] A cross-domain mapping relationship model is established based on the feature representations of the source domain and the target domain to obtain the mapping function from the source domain to the target domain.

[0012] Optionally, the construction of the source domain graph structure and the target domain graph structure specifically includes:

[0013] Data parsing is performed on the acquired source domain data, and a set of source domain nodes is established with each source domain data item as a source domain node. The association information between source domain nodes is collected to generate a set of source domain node relationships.

[0014] The source domain data is structurally integrated based on the source domain node set and the source domain node relationship set to generate a source domain graph structure;

[0015] Perform data parsing on the acquired target domain data, establish a target domain node set by treating each target domain data item as a target domain node, collect the association information between target domain nodes, and create a target domain node relationship set;

[0016] The target domain data is structurally integrated based on the target domain node set and the target domain node relationship set to generate the target domain graph structure.

[0017] Optionally, the execution graph structure encoding process specifically includes:

[0018] The source domain node set in the source domain graph structure is organized node by node to form a source domain node attribute set, and the source domain node attribute set is then subjected to unified structural processing.

[0019] Organize the set of relationships between source domain nodes in the source domain graph structure to generate a set of source domain connection relationships;

[0020] Graph structure encoding preprocessing is performed based on the set of source domain node attributes and the set of source domain connection relationships. The initial node feature matrix of the source domain is constructed based on the set of source domain node attributes, and the initial adjacency matrix of the source domain is constructed based on the set of source domain connection relationships. They are then integrated in a preset order to generate the initial node feature matrix and the initial adjacency matrix of the source domain.

[0021] The target domain node set in the target domain graph structure is organized node by node attributes. The attribute information of each target domain node is combined in a fixed order to form a target domain node attribute set, and the target domain node attribute set is subjected to unified structural processing.

[0022] Organize the relationship set of target domain node relationships in the target domain graph structure, and arrange the connection relationships between target domain nodes in a unified manner according to the node number order to generate a target domain connection relationship set.

[0023] Graph structure encoding preprocessing is performed based on the structured target domain node attribute set and target domain connection relationship set. The target domain initial node feature matrix is ​​constructed based on the target domain node attribute set, and the target domain initial adjacency matrix is ​​constructed based on the target domain connection relationship set. These are then integrated in a preset order to generate the target domain initial node feature matrix and the target domain initial adjacency matrix.

[0024] Optionally, the cross-domain mutual de-pooling structure generation process specifically includes:

[0025] The cross-domain node alignment unit of the improved differentiable graph pooling network DiffPool is used to construct the initial node feature matrix of the target domain into the target domain node feature input set according to a preset order, and to construct the initial adjacency matrix of the target domain into the target domain structural relationship input set according to the node number order.

[0026] The improved differentiable graph pooling network DiffPool includes a cross-domain node alignment unit, a cross-domain structural relationship unification unit, and a cross-domain mutual de-pooling allocation generation unit.

[0027] The cross-domain node alignment unit performs cross-domain node correspondence processing on the initial node feature matrix of the source domain based on the target domain node feature input set, generating the first pooled node alignment set of the source domain;

[0028] The cross-domain structural relation unification unit is used to perform cross-domain structural correspondence processing on the target domain structural relation input set and the initial adjacency matrix of the source domain to generate the first pooling structural alignment set of the source domain.

[0029] A cross-domain mutual de-pooling allocation generation unit is used to perform cross-domain pooling structure construction operations on the source domain first pooling node alignment set and the source domain first pooling structure alignment set to generate the first pooling structure;

[0030] The cross-domain node alignment unit constructs the initial node feature matrix of the source domain into a source domain node feature input set according to a preset order, and constructs the initial adjacency matrix of the source domain into a source domain structural relationship input set according to the node number order;

[0031] The cross-domain node alignment unit is used to perform cross-domain node correspondence processing on the source domain node feature input set and the target domain initial node feature matrix to generate a second pooled node alignment set;

[0032] A cross-domain structural relation unification unit is used to perform cross-domain structural correspondence processing on the source domain structural relation input set and the initial adjacency matrix of the target domain to generate a second pooling structural alignment set;

[0033] The cross-domain mutual pooling allocation generation unit is used to perform cross-domain pooling structure construction operations on the target domain second pooling node alignment set and the target domain second pooling structure alignment set to generate the second pooling structure.

[0034] Optionally, the generation of the source domain pooling graph and the target domain pooling graph specifically includes:

[0035] Based on the first pooling structure, the initial node feature matrix of the source domain is subjected to node-level aggregation processing. The node features belonging to the same pooling group are integrated according to the feature merging strategy defined in the first pooling structure, including weighted combination, mean aggregation or feature rearrangement of node feature vectors, to generate the source domain pooling node feature set.

[0036] Based on the first pooling structure, the initial adjacency matrix of the source domain is reconstructed at the relational level. The connection relationships between nodes in the same pooling group are integrated according to the structure merging strategy defined in the first pooling structure, including performing aggregation on connection weights, rearranging edge relationships, or performing cross-node merging on local structures, to generate a set of pooled node relationships in the source domain.

[0037] The source domain pooling node feature set and the source domain pooling node relationship set are systematically integrated according to a preset graph structure assembly method to construct a source domain pooling graph.

[0038] Based on the second pooling structure, node-level aggregation processing is performed on the initial node feature matrix of the target domain. The node features belonging to the same pooling group in the target domain are integrated according to the feature merging strategy specified in the second pooling structure to generate a pooled node feature set of the target domain.

[0039] Based on the second pooling structure, the initial adjacency matrix of the target domain is reconstructed at the relational level. The structural relationships between nodes in the same pooling group are reconnected, the connection weights are updated, or the structures are merged according to the relational merging rules defined in the second pooling structure, thereby generating a set of pooled node relationships in the target domain.

[0040] The target domain pooling node feature set and the target domain pooling node relationship set are structurally integrated according to a preset graph structure assembly method to construct the target domain pooling graph.

[0041] Optionally, the generation of the source domain feature representation and the target domain specific representation specifically includes:

[0042] The node feature content in the source pooling graph is organized sequentially, and the feature data of the nodes in the source pooling graph are structured and combined according to the node order. The node relationship content in the source pooling graph is organized accordingly, and the relationship between the nodes is structured and combined according to the node number order, so that the node feature content and node relationship content of the source pooling graph generate the feature set of the source pooling graph.

[0043] The node feature content in the target domain pooling graph is organized sequentially, and the feature data of the nodes in the target domain pooling graph is structured and combined according to the node order. The node relationship content in the target domain pooling graph is organized accordingly, and the relationship between the nodes is structured and combined according to the node number order, so that the node feature content and node relationship content of the target domain pooling graph generate the target domain pooling graph feature set.

[0044] Graph feature encoding is performed based on the feature set of the source domain pooled graph. The node feature content and node relationship content of the source domain pooled graph are processed sequentially, extracted hierarchically and combined in a structured manner to generate source domain encoding results.

[0045] Graph feature encoding is performed based on the feature set of the target domain pooled graph. The node feature content and node relationship content of the target domain pooled graph are processed sequentially, extracted hierarchically and combined in a structured manner to generate the target domain encoding result.

[0046] The source domain encoding results are organized according to the feature arrangement of a unified representation space, so that the source domain encoding results generate source domain feature representations;

[0047] The target domain encoding results are organized according to the feature arrangement of a unified representation space, so that the target domain encoding results generate target domain feature representations.

[0048] Optionally, the generation of the mapping function specifically includes:

[0049] The feature content contained in the source domain feature representation is organized according to the preset feature order, feature level and feature dimension to generate the source domain feature set;

[0050] The feature content contained in the target domain feature representation is organized according to the preset feature order, feature level and feature dimension to generate the target domain feature set;

[0051] The source domain feature set and the target domain feature set are paired according to their correspondence to establish a cross-domain feature correspondence set;

[0052] Based on the cross-domain feature correspondence set, a cross-domain relationship structure construction operation is performed. The source domain feature set and the target domain feature set are associated, combined, sequentially merged, or hierarchically integrated according to a preset correspondence method to generate a cross-domain relationship structure for establishing a cross-domain mapping model.

[0053] The feature association content in the cross-domain relationship structure is structurally combined to generate a cross-domain mapping relationship model structure;

[0054] In the cross-domain mapping relationship model structure, the relationship combination between the source domain feature representation and the target domain feature representation is extracted according to a preset method. The feature correspondence structure in the cross-domain mapping relationship model is then consolidated to generate the mapping result structure.

[0055] The mapping result structure is standardized according to the requirements of the feature mapping order and structure, generating a mapping function for performing the transformation from the source domain to the target domain.

[0056] The beneficial effects of this invention are:

[0057] This invention introduces an improved differentiable graph pooling network, DiffPool, and builds upon it a cross-domain node alignment unit, a cross-domain structural relationship unification unit, and a cross-domain mutual depooling allocation generation unit. This enables bidirectional, synchronous, and structured alignment of the source and target domains in terms of node feature arrangement, connection structure, and pooling hierarchy logic. By performing node-level and structure-level cross-domain unification before pooling, this invention effectively eliminates cross-domain representation offsets caused by inconsistent node numbers, different feature field orders, and significant differences in adjacency structures between different data domains, allowing the source and target domains to obtain comparable hierarchical structure expressions during the pooling stage. Simultaneously, through the cross-domain mutual depooling allocation mechanism, this invention moves beyond single-domain internal derivation of the pooling structure. Instead, it guides the generation of the pooling structure of the domain to be aligned by referencing the feature input set and structural relationship input set of the reference domain, achieving cross-domain mutual guidance of pooling structures. This generates a first and second pooling structure that can be shared across domains, providing a structurally consistent input foundation for subsequent graph feature encoding.

[0058] After obtaining a pooled graph with cross-domain comparability, this invention further generates source domain feature representations and target domain feature representations located in the same representation space through a unified graph feature encoding process. This enables the graph structure information, node feature information, and association relationships of different data domains to be structurally expressed under the same encoding system. Based on the above encoding results, this invention finally establishes a cross-domain mapping relationship model using cross-domain feature correspondence, cross-domain feature combination rules, and cross-domain relationship structure, and thereby generates a mapping result structure and mapping function to achieve accurate feature conversion between the source domain and the target domain. This invention significantly improves the stability, interpretability, and generalization ability of cross-domain mapping through a continuous cross-domain unified processing mechanism from the node level, structure level, pooling level to the representation level. It solves the technical problems of inconsistent structure, unstable feature alignment, and difficulty in convergence of the mapping process in existing cross-domain graph mapping methods, and has the beneficial effects of strong cross-domain adaptability, high mapping accuracy, and excellent structural expression consistency. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a cross-domain data mapping method based on graph neural networks proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the structure of the improved DiffPool for cross-domain mutual pooling structure generation processing proposed in this invention, based on a graph neural network-based cross-domain data mapping method. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 and Figure 2 A cross-domain data mapping method based on graph neural networks includes the following steps:

[0064] Acquire source domain data and target domain data, and construct source domain graph structure and target domain graph structure;

[0065] Graph structure encoding is performed on the source domain graph structure and the target domain graph structure respectively to generate the initial node feature matrix and the initial adjacency matrix of the source domain and the target domain;

[0066] An improved differentiable graph pooling network, DiffPool, is used to perform cross-domain mutual pooling structure generation on the initial node feature matrices and initial adjacency matrices of the source and target domains. The first pooling structure of the source domain is generated based on the initial node feature matrix and initial adjacency matrix of the target domain, and the second pooling structure of the target domain is generated based on the initial node feature matrix and initial adjacency matrix of the source domain.

[0067] The first pooling structure is used to perform pooling operations on the initial node feature matrix and the initial adjacency matrix of the source domain to generate the source domain pooling graph. The second pooling structure is used to perform pooling operations on the initial node feature matrix and the initial adjacency matrix of the target domain to generate the target domain pooling graph.

[0068] Graph feature encoding is performed on the source domain pooling graph and the target domain pooling graph respectively to generate source domain feature representations and target domain specific representations located in the same representation space;

[0069] A cross-domain mapping relationship model is established based on the feature representations of the source domain and the target domain to obtain the mapping function from the source domain to the target domain.

[0070] In this embodiment, the construction of the source domain graph structure and the target domain graph structure specifically includes:

[0071] Data parsing is performed on the acquired source domain data. Each source domain data item is used as a source domain node to establish a source domain node set. The association information between source domain nodes is collected. Based on the structural attributes, relational attributes, interaction attributes, or temporal attributes contained in the source domain data, connection information between source domain nodes is generated, and a source domain node relationship set is generated to represent the connection relationship between source domain nodes.

[0072] The source domain data is structurally integrated based on the source domain node set and the source domain node relationship set to generate a source domain graph structure that includes node set representation, node relationship representation, and node attribute representation.

[0073] Data parsing is performed on the acquired target domain data. Each target domain data item is used as a target domain node to establish a target domain node set. The association information between target domain nodes is collected. Based on the structural attributes, relational attributes, interaction attributes, or temporal attributes contained in the target domain data, connection information between target domain nodes is generated, and a target domain node relationship set is generated to represent the connection relationship between target domain nodes.

[0074] The target domain data is structurally integrated based on the target domain node set and the target domain node relationship set to generate a target domain graph structure that includes node set representation, node relationship representation, and node attribute representation.

[0075] In this embodiment, the execution graph structure encoding process specifically includes:

[0076] The source domain node set in the source domain graph structure is organized node by node attributes. The attribute information of each source domain node is combined in a fixed order to form a source domain node attribute set, and the source domain node attribute set is subjected to unified structural processing.

[0077] The relationships of the source domain nodes in the source domain graph structure are organized, and the connection relationships between the source domain nodes are arranged in a unified manner according to the node number order to generate a source domain connection relationship set.

[0078] Graph structure encoding preprocessing is performed based on the set of source domain node attributes and the set of source domain connection relationships. The initial node feature matrix of the source domain is constructed based on the set of source domain node attributes, and the initial adjacency matrix of the source domain is constructed based on the set of source domain connection relationships. They are then integrated in a preset order to generate the initial node feature matrix and the initial adjacency matrix of the source domain.

[0079] The target domain node set in the target domain graph structure is organized node by node attributes. The attribute information of each target domain node is combined in a fixed order to form a target domain node attribute set, and the target domain node attribute set is subjected to unified structural processing.

[0080] Organize the relationship set of target domain node relationships in the target domain graph structure, and arrange the connection relationships between target domain nodes in a unified manner according to the node number order to generate a target domain connection relationship set.

[0081] Graph structure encoding preprocessing is performed based on the structured target domain node attribute set and target domain connection relationship set. The target domain initial node feature matrix is ​​constructed based on the target domain node attribute set, and the target domain initial adjacency matrix is ​​constructed based on the target domain connection relationship set. These are then integrated in a preset order to generate the target domain initial node feature matrix and the target domain initial adjacency matrix.

[0082] In this embodiment, the cross-domain mutual de-pooling structure generation process specifically includes:

[0083] The cross-domain node alignment unit of the improved differentiable graph pooling network DiffPool is used to construct the initial node feature matrix of the target domain into the target domain node feature input set according to a preset order, and to construct the initial adjacency matrix of the target domain into the target domain structural relationship input set according to the node number order.

[0084] The improved differentiable graph pooling network DiffPool includes a cross-domain node alignment unit, a cross-domain structural relationship unification unit, and a cross-domain mutual de-pooling allocation generation unit.

[0085] The cross-domain node alignment unit is used to perform cross-domain node semantic unification processing on the initial node feature matrix of the target domain; the cross-domain structural relationship unification unit is used to perform cross-domain structural relationship unification processing on the initial adjacency matrix of the target domain; the cross-domain mutual de-pooling allocation generation unit is used to generate a cross-domain pooling allocation structure for source domain pooling hierarchy derivation based on the target domain node feature input set and the target domain structural relationship input set.

[0086] The generation of the cross-domain node alignment unit specifically includes: parsing the node number of the initial node feature matrix of the reference domain to obtain a set of reference domain node numbers representing the arrangement of nodes in the reference domain; parsing the node number of the initial node feature matrix of the domain to be aligned to obtain a set of node numbers of the domain to be aligned to represent the arrangement of nodes in the domain to be aligned; using the set of reference domain node numbers as the arrangement basis, constructing a node arrangement reference for unifying the node order across domains; adjusting the position of the initial node feature matrix of the reference domain according to the node arrangement reference to obtain the node rearrangement structure of the reference domain; adjusting the position of the initial node feature matrix of the domain to be aligned according to the node arrangement reference to obtain the node rearrangement structure of the domain to be aligned; comparing the node rearrangement structure of the reference domain with the node rearrangement structure of the domain to be aligned item by item according to the node number to obtain a set of node correspondences reflecting the cross-domain node correspondences; and based on the set of node correspondences, encapsulating and integrating the node arrangement reference, the node rearrangement structure of the reference domain, the node rearrangement structure of the domain to be aligned, and the set of node correspondences in a modular manner to generate the cross-domain node alignment unit.

[0087] The generation of the cross-domain structural relationship unification unit specifically includes: decomposing the node connection relationships of the initial adjacency matrix of the target domain to generate a target domain structural ordering set representing the structural order between nodes in the target domain; decomposing the connection relationships of the initial adjacency matrix of the source domain to generate a source domain structural ordering set representing the structural order between nodes in the source domain; extracting node arrangement benchmarks from the target domain structural ordering set to form a structural arrangement reference for cross-domain structural comparison; reorganizing the node connection order of the initial adjacency matrix of the target domain according to the structural arrangement reference to obtain a target domain structural alignment set for cross-domain comparison; reorganizing the node connection order of the initial adjacency matrix of the source domain according to the structural arrangement reference to obtain a source domain structural alignment set for cross-domain comparison; comparing the target domain structural alignment set and the source domain structural alignment set item by item according to node number to generate a cross-domain structural correspondence set describing the cross-domain structural association method; and modularly combining the structural arrangement reference, the target domain structural alignment set, the source domain structural alignment set, and the cross-domain structural correspondence set to generate the cross-domain structural relationship unification unit.

[0088] The construction of the cross-domain mutual de-pooling allocation generation unit specifically includes: sequentially reading attribute fields of the node feature input set of the reference domain and sequentially reading join terms of the structural relationship input set of the reference domain, and obtaining the reference domain feature structure reference using feature serialization and structural sequence extraction; sequentially reading attribute fields of the node feature input set of the domain to be aligned and sequentially reading join terms of the structural relationship input set of the domain to be aligned, and obtaining the domain feature structure reference through a serialization method consistent with the reference domain; comparing the reference domain feature structure reference and the domain feature structure reference of the domain to be aligned item by item according to the cross-domain feature correspondence, and generating a cross-domain node allocation reference based on the similarity of node field arrangement; and then comparing the reference domain feature structure reference with... The feature structure reference of the domain to be aligned is compared according to the structural association order, and a cross-domain structure allocation reference is generated based on the structural correspondence between the connected nodes. The cross-domain node allocation reference and the cross-domain structure allocation reference are jointly integrated according to the preset pooling hierarchy rules to generate a pooling allocation structure that describes the cross-domain pooling allocation logic. The pooling allocation structure is encapsulated in a modular way to generate a cross-domain mutual depooling allocation generation unit. When the reference domain is the target domain and the domain to be aligned is the source domain, the cross-domain mutual depooling allocation generation unit generates a cross-domain pooling allocation structure to construct the first pooling structure. When the reference domain is the source domain and the domain to be aligned is the target domain, the cross-domain mutual depooling allocation generation unit generates a cross-domain pooling allocation structure to construct the second pooling structure.

[0089] The cross-domain node alignment unit performs cross-domain node correspondence processing on the initial node feature matrix of the source domain based on the target domain node feature input set. It maps the node order and node arrangement contained in the target domain node feature input set to the initial node feature matrix of the source domain, performs cross-domain unification processing on the initial node feature matrix of the source domain, and generates the first pooled node alignment set of the source domain to represent the cross-domain node feature alignment result.

[0090] The cross-domain node matching process specifically includes: reading the node feature input set of the reference domain node by node, arranging the nodes in the reference domain according to the reading order to obtain the reference domain node arrangement order; performing feature position rearrangement processing on the initial node feature matrix of the domain to be aligned according to the reference domain node arrangement order, adjusting the arrangement position of each node in the domain to be aligned to obtain the node feature arrangement set of the domain to be aligned; comparing each node feature in the node feature arrangement set of the domain to be aligned item by item with the feature field position in the reference domain node feature input set, and performing position correction on the feature fields of the nodes to be aligned according to the field appearance order to obtain a unified set of node feature fields of the domain to be aligned; combining the reference domain node arrangement order and the unified set of node feature fields of the domain to be aligned according to the node number to generate a cross-domain node alignment set; when the reference domain is the target domain and the domain to be aligned is the source domain, a cross-domain node alignment set with a first pooling structure is generated; when the reference domain is the source domain and the domain to be aligned is the target domain, a cross-domain node alignment set with a second pooling structure is generated.

[0091] The cross-domain structural relationship unification unit performs cross-domain structural correspondence processing on the target domain structural relationship input set and the initial adjacency matrix of the source domain, maps the node numbering order and associated layout contained in the target domain structural relationship input set to the initial adjacency matrix of the source domain, performs cross-domain relationship unification processing on the initial adjacency matrix of the source domain, and generates the first pooled structural alignment set of the source domain to represent the cross-domain structural relationship alignment result.

[0092] The cross-domain structure correspondence processing specifically includes: reading the connection relationships item by item from the structural relationship input set of the reference domain, arranging the connection items between node pairs according to their appearance order to obtain the connection order of the reference domain nodes; performing connection relationship rearrangement processing on the initial adjacency matrix of the domain to be aligned according to the connection order of the reference domain nodes, reorganizing the connection positions between the nodes of the domain to be aligned, and obtaining the reorganized set of the domain structure to be aligned; comparing the connection relationships in the reorganized set item by item with the corresponding relationships in the structural relationship input set of the reference domain, and adjusting them according to the order of appearance of the connections to generate the adjusted set of the domain structure order; arranging the relationship items in the adjusted set according to the node number order to ensure that the relationship items are consistent in the overall structural layout, and obtaining the unified set of the domain structure to be aligned; combining the connection order of the reference domain nodes and the unified set of structures according to the node number to generate the cross-domain structure alignment set; generating the first pooled structure alignment set when the reference domain is the target domain and the domain to be aligned is the source domain; generating the second pooled structure alignment set when the reference domain is the source domain and the domain to be aligned is the target domain.

[0093] A cross-domain mutual de-pooling allocation generation unit is used to perform cross-domain pooling structure construction operations on the first pooling node alignment set and the first pooling structure alignment set of the source domain. The target domain node feature input set and the target domain structure relationship input set are cross-domain allocated and fused in the source domain pooling level to generate the first pooling structure used to represent the relationship of the source domain pooling level.

[0094] The cross-domain pooling structure construction operation specifically includes: reading the node arrangement in the cross-domain node alignment set of the reference domain item by item, using the positional relationship of the nodes in the alignment set as the node aggregation order to obtain a node aggregation reference; reading the connection relationship in the cross-domain structure alignment set of the reference domain item by item, using the association order between node pairs as the structure aggregation order to obtain a structure aggregation reference; and, based on the node aggregation reference, performing merging, rearrangement, or hierarchical combination on the node features in the initial node feature matrix of the domain to be pooled according to the aggregation order to obtain the domain to be pooled. Node aggregation results; based on the structural aggregation reference, the node connection relationships in the initial adjacency matrix of the domain to be pooled are consolidated, associated, or hierarchically combined according to the aggregation order to obtain the structural aggregation results of the domain to be pooled; the node aggregation results and structural aggregation results of the domain to be pooled are structurally integrated according to the preset pooling hierarchy requirements to obtain the cross-domain pooling structure; when the reference domain is the target domain and the domain to be pooled is the source domain, the cross-domain pooling structure is the first pooling structure; when the reference domain is the source domain and the domain to be pooled is the target domain, the cross-domain pooling structure is the second pooling structure;

[0095] The cross-domain node alignment unit constructs the initial node feature matrix of the source domain into a source domain node feature input set according to a preset order, and constructs the initial adjacency matrix of the source domain into a source domain structural relationship input set according to the node number order;

[0096] The cross-domain node alignment unit performs cross-domain node correspondence processing on the source domain node feature input set and the target domain initial node feature matrix, maps the node order and node arrangement contained in the source domain node feature input set to the target domain initial node feature matrix, performs cross-domain unification processing on the target domain initial node feature matrix, and generates the target domain second pooled node alignment set to represent the cross-domain node feature alignment result.

[0097] A cross-domain structural relationship unification unit is used to perform cross-domain structural correspondence processing on the source domain structural relationship input set and the initial adjacency matrix of the target domain. The node numbering order and associated layout contained in the source domain structural relationship input set are mapped to the initial adjacency matrix of the target domain. Cross-domain relationship unification processing is performed on the initial adjacency matrix of the target domain to generate the second pooling structure alignment set of the target domain used to represent the cross-domain structural relationship alignment result.

[0098] The cross-domain mutual depooling allocation generation unit performs cross-domain pooling structure construction operation on the target domain second pooling node alignment set and the target domain second pooling structure alignment set. The source domain node feature input set and the source domain structure relationship input set are cross-domain allocated and fused in the target domain pooling level to generate the second pooling structure used to represent the target domain pooling level relationship.

[0099] In this embodiment, the generation of the source domain pooling graph and the target domain pooling graph specifically includes:

[0100] Based on the first pooling structure, the initial node feature matrix of the source domain is subjected to node-level aggregation processing. The node features belonging to the same pooling group are integrated according to the feature merging strategy defined in the first pooling structure, including weighted combination, mean aggregation or feature rearrangement of node feature vectors, to generate a set of source domain pooled node features used to express the attribute content of the pooled nodes.

[0101] Based on the first pooling structure, the initial adjacency matrix of the source domain is reconstructed at the relational level. The connection relationships between nodes in the same pooling group are integrated according to the structure merging strategy defined in the first pooling structure. This includes performing aggregation on connection weights, rearranging edge relationships, or performing cross-node merging on local structures to generate a set of source domain pooled node relationships used to express the topology of node association after pooling.

[0102] The source domain pooling node feature set and the source domain pooling node relationship set are systematically integrated according to a preset graph structure assembly method to construct a source domain pooling graph.

[0103] Based on the second pooling structure, the initial node feature matrix of the target domain is subjected to node-level aggregation processing. The node features belonging to the same pooling group in the target domain are integrated according to the feature merging strategy specified in the second pooling structure to generate a target domain pooled node feature set used to express the node attribute content after pooling of the target domain.

[0104] Based on the second pooling structure, the initial adjacency matrix of the target domain is reconstructed at the relation level. The structural relationships between nodes in the same pooling group are reconnected, the connection weights are updated, or the structures are merged according to the relation merging rules defined in the second pooling structure, generating a set of target domain pooled node relationships to express the topological hierarchy of nodes after target domain pooling.

[0105] The target domain pooling node feature set and the target domain pooling node relationship set are structurally integrated according to a preset graph structure assembly method to construct the target domain pooling graph.

[0106] In this embodiment, the generation of the source domain feature representation and the target domain specific representation specifically includes:

[0107] The node feature content in the source pooling graph is organized sequentially, and the feature data of the nodes in the source pooling graph are structured and combined according to the node order. The node relationship content in the source pooling graph is organized accordingly, and the relationship between the nodes is structured and combined according to the node number order, so that the node feature content and node relationship content of the source pooling graph generate the feature set of the source pooling graph.

[0108] The node feature content in the target domain pooling graph is organized sequentially, and the feature data of the nodes in the target domain pooling graph is structured and combined according to the node order. The node relationship content in the target domain pooling graph is organized accordingly, and the relationship between the nodes is structured and combined according to the node number order, so that the node feature content and node relationship content of the target domain pooling graph generate the target domain pooling graph feature set.

[0109] Graph feature encoding is performed based on the feature set of the source pooling graph. The node feature content and node relationship content of the source pooling graph are processed sequentially, extracted hierarchically and combined in a structured manner to generate source domain encoding results that represent the overall features of the source pooling graph.

[0110] Graph feature encoding is performed based on the feature set of the target domain pooled graph. The node feature content and node relationship content of the target domain pooled graph are processed sequentially, extracted hierarchically and combined in a structured manner to generate the target domain encoding result that represents the overall features of the target domain pooled graph.

[0111] The source domain encoding results are organized according to the feature arrangement of a unified representation space, so that the source domain encoding results generate source domain feature representations with a unified structure and a unified dimension.

[0112] The generation of the feature representation specifically includes: performing sequential processing on the node feature content of the pooling graph to obtain a pooling graph node feature sequence describing the arrangement of node features; performing relationship parsing on the node relationship content of the pooling graph to obtain a pooling graph relationship sequence describing the relationship structure between nodes; performing hierarchical feature extraction processing on the node feature content of the pooling graph according to the pooling graph node feature sequence to obtain a preliminary feature combination result of the pooling graph; performing structural relationship integration processing on the node relationship content of the pooling graph according to the pooling graph relationship sequence to obtain a structural feature combination result of the pooling graph; and performing combination and ordering processing on the preliminary feature combination result and the structural feature combination result according to a preset representation dimension to generate a representation of the entire pooling graph. The encoding results of the graph structure and node feature information are processed according to the feature dimensions, arrangement rules, and feature types of the unified representation space to generate feature representations. When the object of feature representation generation is a source domain pooled graph, the pooled graph node feature sequence is generated from the source domain pooled node feature set, and the pooled graph relation sequence is generated from the source domain pooled node relation set. The final encoding result is then normalized using the unified representation space to generate the source domain feature representation. When the object of feature representation generation is a target domain pooled graph, the pooled graph node feature sequence is generated from the target domain pooled node feature set, and the pooled graph relation sequence is generated from the target domain pooled node relation set. The final encoding result is then normalized using the unified representation space to generate the target domain feature representation.

[0113] The target domain encoding results are organized according to the feature arrangement of a unified representation space, so that the target domain encoding results generate a target domain feature representation with a unified structure and unified dimension.

[0114] In this embodiment, the generation of the mapping function specifically includes:

[0115] The feature content contained in the source domain feature representation is organized according to the preset feature order, feature level and feature dimension to generate the source domain feature set;

[0116] The feature content contained in the target domain feature representation is organized according to the preset feature order, feature level and feature dimension to generate the target domain feature set;

[0117] The source domain feature set and the target domain feature set are paired according to their correspondence. A cross-domain feature correspondence set is established based on the feature arrangement, feature dimension order, and feature correspondence rules between the source domain feature representation and the target domain feature representation.

[0118] Based on the cross-domain feature correspondence set, a cross-domain relationship structure construction operation is performed. The source domain feature set and the target domain feature set are associated, combined, sequentially merged, or hierarchically integrated according to a preset correspondence method to generate a cross-domain relationship structure for establishing a cross-domain mapping model.

[0119] The feature association content in the cross-domain relationship structure is structurally combined to generate a cross-domain mapping relationship model structure that describes the mapping relationship between the feature representation of the source domain and the feature representation of the target domain.

[0120] The generation of the cross-domain mapping relationship model structure specifically includes: performing node-level feature correspondence parsing on the input cross-domain relationship structure to obtain node correspondence parsing results that characterize the cross-domain node feature association method; performing relation-level structure correspondence parsing on the cross-domain relationship structure to generate relation correspondence parsing results that characterize the relationship matching method between cross-domain nodes; performing joint ordering processing on the node correspondence parsing results and relation correspondence parsing results according to a preset cross-domain feature combination logic to obtain a cross-domain feature combination sequence; performing hierarchical association extraction processing on the cross-domain feature combination sequence to generate a cross-domain association hierarchy result that represents the association hierarchy between cross-domain features; and performing structured arrangement processing on the cross-domain association hierarchy result according to the model construction format to generate the cross-domain mapping relationship model structure.

[0121] In the cross-domain mapping relationship model structure, the relationship combination between the source domain feature representation and the target domain feature representation is extracted according to a preset method. The feature correspondence structure in the cross-domain mapping relationship model is then consolidated to generate a mapping result structure that represents the transformation relationship from the source domain to the target domain.

[0122] The generation of the mapping result structure specifically includes: performing item-by-item extraction of the feature correspondence content in the cross-domain mapping relationship model structure to obtain a feature pairing set used to characterize the feature pairing relationship; performing sequential parsing processing on the feature arrangement content in the cross-domain mapping relationship model structure to obtain a feature arrangement set used to describe the cross-domain feature arrangement method; comparing the feature pairing set and the feature arrangement set in order according to the structural requirements of the cross-domain mapping relationship model to obtain a feature mapping association set used to represent the mapping association method between features; and structuring the feature mapping association set according to the node number order and feature dimension order to obtain a mapping result structure with a complete cross-domain feature mapping structure.

[0123] The mapping result structure is standardized according to the requirements of the feature mapping order and structure, and a mapping function is generated to perform the transformation from the source domain to the target domain.

[0124] The generation of the mapping function specifically includes: performing mapping order parsing processing on the feature-corresponding content in the mapping result structure to obtain a mapping order set for describing the cross-domain feature transformation order; performing association structure extraction processing on the feature-related content in the mapping result structure to obtain a mapping combination set for describing the cross-domain feature combination method; performing ordering processing on the mapping order set and the mapping combination set according to the structural requirements of the mapping relationship model to obtain a mapping rule structure for describing the cross-domain transformation logic; and performing unification and organization processing on the mapping rule structure according to the feature mapping order requirements and structural requirements to generate a mapping function for performing source domain to target domain transformation.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to cross-domain data mapping tasks in multiple industrial fields. Specifically, the invention was validated in a multimodal data fusion production environment involving data acquired from multiple different sources. These data sources come from different devices and sensors in the smart manufacturing process, including image data, sensor data, and time-series data generated by IoT devices. Among these devices, there are significant differences in data structure and format between sensors, equipment, and control systems. How to achieve effective fusion of cross-domain data while ensuring data consistency is a key challenge in the current production environment.

[0127] In this environment, the source domain data comes from real-time production data collected by different types of sensors, including physical quantities such as temperature, pressure, humidity, and vibration. The target domain data comes from image data collected by industrial cameras and structured data obtained from robotic arms. These data differ not only in data format but also in data hierarchy, data granularity, and sampling frequency. Because there is no direct mapping relationship between the source and target domain data, traditional cross-domain data alignment methods cannot effectively unify the structure and characteristics of the source and target domain data, resulting in inaccurate data fusion results and affecting subsequent analysis and decision-making.

[0128] To address this issue, this invention provides a cross-domain data mapping method based on graph neural networks. By using an improved DiffPool graph pooling network, this invention can perform deep node alignment and structural unification processing on the source and target domain data. The source and target domain data are respectively processed by graph structure encoding to generate their respective initial node feature matrices and adjacency matrices. The graph structures of the source and target domains are processed by cross-domain node alignment units and cross-domain structural relationship unification units, thereby eliminating differences in node arrangement and structure between data sources and achieving alignment between the two data domains. In particular, during the construction stage of the cross-domain pooling structure, by utilizing the cross-domain mutual de-pooling allocation generation unit, this invention can effectively solve the pooling structure differences between the source and target domains, ensure the consistency of the cross-domain pooling levels, and provide stable input for subsequent graph feature encoding.

[0129] In this production scenario, the application of this invention is manifested in its ability to automatically and dynamically adapt to the characteristics of the source and target domains when performing real-time data fusion among multiple devices. By generating a cross-domain pooling structure, this invention not only achieves alignment at the data level but also ensures the consistency of node features and structural information in the same representation space, thereby improving the accuracy of data fusion. Traditional cross-domain mapping methods often rely on static rules and models, while this invention, by adaptively adjusting the pooling structure, can adjust the mapping relationship in real time for the feature changes of different devices and sensors, effectively improving the adaptability of cross-domain mapping.

[0130] Table 1. Performance comparison between the method of this invention and traditional cross-domain mapping methods

[0131]

[0132] As can be seen from the table results, the improved DiffPool cross-domain mapping method proposed in this invention is significantly better than traditional methods in several key technical indicators such as mapping accuracy, structure alignment capability, robustness, and pooling performance. At the same time, compared with traditional single-domain GNN, the improvement of this invention remains within a reasonable medium range, which is more in line with the actual performance in engineering practice.

[0133] In terms of cross-domain node correspondence accuracy, this invention achieves 86.2%, which is nearly 19.4 percentage points higher than the 66.8% of the traditional rule method and 11.9 percentage points higher than the 74.3% of the traditional GNN. This improvement comes from the role of the cross-domain node alignment unit, which makes the arrangement of nodes and field positions between cross-domains more consistent, thereby reducing the probability of mismatch caused by differences in node order.

[0134] The cross-domain structural alignment consistency has also been improved to 85.4%, which is significantly better than traditional methods. Traditional methods have difficulty solving the problem that the topological structures of the source domain and the target domain are fundamentally different. However, this invention improves the matching quality at the structural level by reconstructing the adjacency relationship in a unified arrangement through the structural relationship unification unit.

[0135] In terms of robustness, under heterogeneous data perturbation scenarios, the stability of this invention reaches 82.8%, which is 21.3% higher than traditional rule-based methods and 13.4% higher than traditional GNNs. In scenarios with missing nodes, this invention achieves 81.6%, which is also significantly better than traditional methods. This invention completes the dual alignment of nodes and structures before pooling, enabling the model to maintain stable feature representation even when the input is missing or perturbed.

[0136] In terms of pooling performance, the reliability of pooling structure generation in this invention reaches 86.1%, and the consistency of cross-domain pooling hierarchy reaches 84.7%, which is about 14 to 22 percentage points better than traditional methods. The core reason is that the cross-domain mutual depooling mechanism of this invention unifies the structural sequences of the two domains, so that the source domain and the target domain have highly consistent aggregation rules in the pooling stage, which helps to reduce the offset error of hierarchical pooling.

[0137] In terms of computational performance, the convergence time of the mapping method in this invention is 0.76 times that of the traditional method, which has a significant acceleration effect. Since the features and structure are aligned and unified, the data distribution in the graph feature encoding stage is more regular, reducing gradient oscillation and optimization path offset, thereby shortening the convergence time. At the same time, the computational resource utilization rate is maintained at a "low to medium" level, indicating that the computational burden is not significantly increased while improving performance.

[0138] In terms of overall performance, the cross-domain mapping score of this invention reaches 87.8, surpassing the traditional method's 65.9 and the traditional GNN's 73.1. The performance advantage of this invention mainly comes from the synergistic effect of three core improvement mechanisms: the cross-domain node alignment mechanism unifies the arrangement of cross-domain nodes, improving the accuracy of node feature correspondence; the cross-domain structural relationship unification mechanism reconstructs adjacency relationships, making the topological structures of different domains comparable; and the cross-domain mutual pooling mechanism eliminates hierarchical differences in the pooling stage, ensuring the consistency of pooling output in semantics and structure. These mechanisms work together to enable the model to maintain stable cross-domain mapping capabilities when facing heterogeneous scenarios, structural differences, and data perturbations, demonstrating the significant advantages of this invention in practical applications.

[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-domain data mapping method based on a graph neural network, characterized in that, The method comprises the following steps: obtaining source domain data and target domain data, and constructing a source domain graph structure and a target domain graph structure; performing graph structure coding processing on the source domain graph structure and the target domain graph structure respectively to generate initial node feature matrices and initial adjacency matrices of the source domain and the target domain; performing cross-domain mutual solution pooling structure generation processing on the initial node feature matrices and the initial adjacency matrices of the source domain and the target domain by using an improved differentiable graph pooling network DiffPool, generating a first pooling structure of the source domain based on the initial node feature matrix and the initial adjacency matrix of the target domain, and generating a second pooling structure of the target domain based on the initial node feature matrix and the initial adjacency matrix of the source domain; performing pooling operations on the initial node feature matrix and the initial adjacency matrix of the source domain according to the first pooling structure to generate a source domain pooled graph, and performing pooling operations on the initial node feature matrix and the initial adjacency matrix of the target domain according to the second pooling structure to generate a target domain pooled graph; performing graph feature coding processing on the source domain pooled graph and the target domain pooled graph respectively to generate source domain feature representations and target domain feature representations in the same representation space; establishing a cross-domain mapping relationship model based on the source domain feature representations and the target domain feature representations to obtain a mapping function from the source domain to the target domain.

2. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, The construction of the source domain graph structure and the target domain graph structure specifically comprises: performing data analysis on the obtained source domain data, establishing a source domain node set by taking each item of source domain data as a source domain node, collecting association information between the source domain nodes to generate a source domain node relationship set, and performing structural integration on the source domain data based on the source domain node set and the source domain node relationship set to generate a source domain graph structure; performing data analysis on the obtained target domain data, establishing a target domain node set by taking each item of target domain data as a target domain node, collecting association information between the target domain nodes to generate a target domain node relationship set, and performing structural integration on the target domain data based on the target domain node set and the target domain node relationship set to generate a target domain graph structure. The graph structure coding processing specifically comprises: performing node-by-node attribute arrangement on the source domain node set in the source domain graph structure to form a source domain node attribute set, and performing unified structural processing on the source domain node attribute set; 3. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, performing relationship arrangement on the source domain node relationship set in the source domain graph structure to generate a source domain connection relationship set; performing graph structure coding preprocessing based on the structured source domain node attribute set and the source domain connection relationship set, constructing a source domain initial node feature matrix based on the source domain node attribute set, constructing a source domain initial adjacency matrix based on the source domain connection relationship set, and integrating in a preset order to generate the source domain initial node feature matrix and the source domain initial adjacency matrix; performing node-by-node attribute arrangement on the target domain node set in the target domain graph structure to form a target domain node attribute set by combining attribute information of each target domain node in a fixed order, and performing unified structural processing on the target domain node attribute set; performing relationship arrangement on the target domain node relationship set in the target domain graph structure to uniformly arrange connection relationships between the target domain nodes in node number order to generate a target domain connection relationship set. ​ ​ The graph structure coding preprocessing is performed based on the structured target domain node attribute set and the target domain connection relationship set, an initial node feature matrix of the target domain is constructed based on the target domain node attribute set, an initial adjacency matrix of the target domain is constructed based on the target domain connection relationship set, and the initial node feature matrix of the target domain and the initial adjacency matrix of the target domain are integrated in a preset order to generate the initial node feature matrix of the target domain and the initial adjacency matrix of the target domain.

4. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, The cross-domain mutual solution pooling structure generation process specifically includes: The cross-domain node alignment unit of the improved differentiable graph pooling network DiffPool is used to construct the initial node feature matrix of the target domain into a target domain node feature input set in a preset order, and construct the initial adjacency matrix of the target domain into a target domain structure relationship input set according to the node number order; The improved differentiable graph pooling network DiffPool includes a cross-domain node alignment unit, a cross-domain structure alignment unit and a cross-domain mutual solution pooling distribution generation unit; The cross-domain node alignment unit performs cross-domain node corresponding processing on the initial node feature matrix of the source domain based on the target domain node feature input set to generate a first pooled node alignment set of the source domain; The cross-domain structure alignment unit is used to perform cross-domain structure corresponding processing on the target domain structure relationship input set and the initial adjacency matrix of the source domain to generate a first pooled structure alignment set of the source domain; The cross-domain mutual solution pooling distribution generation unit is used to perform cross-domain pooled structure construction operation on the first pooled node alignment set of the source domain and the first pooled structure alignment set of the source domain to generate a first pooled structure. The cross-domain node alignment unit constructs the initial node feature matrix of the source domain into a source domain node feature input set in a preset order, and constructs the initial adjacency matrix of the source domain into a source domain structure relationship input set according to the node number order; The cross-domain node alignment unit is used to perform cross-domain node corresponding processing on the source domain node feature input set and the initial node feature matrix of the target domain to generate a second pooled node alignment set; The cross-domain structure alignment unit is used to perform cross-domain structure corresponding processing on the source domain structure relationship input set and the initial adjacency matrix of the target domain to generate a second pooled structure alignment set; The cross-domain mutual solution pooling distribution generation unit is used to perform cross-domain pooled structure construction operation on the second pooled node alignment set of the target domain and the second pooled structure alignment set of the target domain to generate a second pooled structure.

5. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, The source domain pooled graph and the target domain pooled graph are generated specifically as follows: Based on the first pooled structure, node-level aggregation processing is performed on the initial node feature matrix of the source domain, and the node features belonging to the same pooled group are integrated according to the feature merging strategy defined in the first pooled structure, including weighted combination, mean aggregation or feature rearrangement of the node feature vector, to generate a source domain pooled node feature set; Based on the first pooled structure, relationship-level structure reconstruction is performed on the initial adjacency matrix of the source domain, and the connection relationship between nodes in the same pooled group is integrated according to the structure merging strategy defined in the first pooled structure, including aggregation of connection weights, rearrangement of edge relationships or cross-node merging of local structures, to generate a source domain pooled node relationship set; The source domain pooled node feature set and the source domain pooled relation set are sequentially integrated according to a preset graph structure assembly manner to construct a source domain pooled graph. Based on the second pooling structure, node-level aggregation processing is performed on the initial node feature matrix of the target domain, node features belonging to the same pooling group in the target domain are integrated according to the feature merging strategy defined in the second pooling structure, and a target domain pooled node feature set is generated. Based on the second pooling structure, relation-level topology reconstruction is performed on the initial adjacency matrix of the target domain, the structural relationship between nodes in the same pooling group is reconnected, the connection weight is updated, or the structure is merged according to the relation merging rule defined in the second pooling structure, and a target domain pooled node relation set is generated. The target domain pooled node feature set and the target domain pooled node relation set are structurally integrated according to a preset graph structure assembly manner to construct a target domain pooled graph.

6. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, The generation of the source domain feature representation and the target domain feature representation specifically includes: The node feature content in the source domain pooled graph is sequentially arranged, the feature data of the nodes of the source domain pooled graph is structurally combined according to the node order, the node relation content in the source domain pooled graph is correspondingly arranged, the association relationship between the nodes is structurally combined according to the node number order, and the node feature content and the node relation content of the source domain pooled graph generate a source domain pooled graph feature set; The node feature content in the target domain pooled graph is sequentially arranged, the feature data of the nodes of the target domain pooled graph is structurally combined according to the node order, the node relation content in the target domain pooled graph is correspondingly arranged, the association relationship between the nodes is structurally combined according to the node number order, and the node feature content and the node relation content of the target domain pooled graph generate a target domain pooled graph feature set; Based on the source domain pooled graph feature set, graph feature coding processing is performed, the node feature content and the node relation content of the source domain pooled graph are sequentially processed, hierarchically extracted and structurally combined, and a source domain coding result is generated; Based on the target domain pooled graph feature set, graph feature coding processing is performed, the node feature content and the node relation content of the target domain pooled graph are sequentially processed, hierarchically extracted and structurally combined, and a target domain coding result is generated; The source domain coding result is arranged according to the feature arrangement manner of the unified representation space, so that the source domain coding result generates a source domain feature representation; The target domain coding result is arranged according to the feature arrangement manner of the unified representation space, so that the target domain coding result generates a target domain feature representation.

7. The cross-domain data mapping method based on a graph neural network according to claim 1, characterized in that, The generation of the mapping function specifically includes: The feature content contained in the source domain feature representation is processed according to the preset feature order, feature level and feature dimension to generate a source domain feature set; The feature content contained in the target domain feature representation is processed according to the preset feature order, feature level and feature dimension to generate a target domain feature set; The source domain feature set and the target domain feature set are paired according to the corresponding relationship to establish a cross-domain feature correspondence relationship set; The cross-domain relationship structure construction operation is performed based on the cross-domain feature correspondence relationship set, the source domain feature set and the target domain feature set are associated, combined, sequentially fused or hierarchically integrated according to a preset correspondence mode, and a cross-domain relationship structure for cross-domain mapping model establishment is generated; The feature association content in the cross-domain relationship structure is structured and combined to generate a cross-domain mapping relationship model structure; In the cross-domain mapping relationship model structure, the relationship combination of the source domain feature representation corresponding to the target domain feature representation is extracted according to a preset mode, the feature correspondence structure in the cross-domain mapping relationship model is sorted, and a mapping result structure is generated; The mapping result structure is uniformly arranged according to the sequence requirement and structure requirement of feature mapping to generate a mapping function for executing source domain to target domain conversion.

Images