Method and system for automatically constructing heterogeneous table association relationship based on graph neural network

Abstract

The present application belongs to the technical field of database data integration and intelligent data processing, and particularly relates to a method for automatically constructing a heterogeneous table association relationship based on a graph neural network, comprising the following steps: obtaining multiple pieces of heterogeneous table data, and extracting the table header semantic features and data distribution features of each data column in each table; based on the table header semantic features and data distribution features, calculating the foreign key association probability between each data column of different tables. The method adopts a double-dimensional matching mode of table header semantic features and data distribution features, breaks through the technical limitations of traditional character matching, and can accurately identify the column-level semantic association of synonyms; then, global structure reasoning is performed through a graph neural network to denoise and optimize the initial association result, filter out false associations with low confidence, and effectively guarantee the reliability of the association relationship.

Description

Technical Field

[0001] This invention belongs to the field of database data integration and intelligent data processing technology, and in particular relates to a method and system for automatically constructing heterogeneous table association relationships based on graph neural networks. Background Technology

[0002] In enterprise digital operations scenarios, various business data are typically stored in different business systems such as sales, inventory, finance, and supply chain, resulting in a large number of business tables with heterogeneous structures and inconsistent field naming rules. When performing cross-table data queries, data integration governance, and intelligent data analysis, it is essential to rely on clear inter-table relationships (such as foreign key mapping relationships and JOIN join relationships) as the foundation for data association.

[0003] Traditional methods of building multi-table relationships mainly rely on data engineers manually sorting out business logic, manually configuring foreign key constraints in the database, or pre-defining mapping relationships between tables in the data warehouse. This approach is not only time-consuming and labor-intensive, but also difficult to adapt to the rapid iteration of enterprise business: when adding new business tables, adjusting table structures, or renaming fields, the relationships must be manually sorted out and maintained again, which is prone to problems such as lagging relationship configuration and incorrect mapping relationships, and cannot support the dynamically changing enterprise data environment.

[0004] In existing technologies, there are also automatic association discovery schemes based on column name character matching and regular expression rule matching. However, such schemes can only identify field associations with completely identical column names or data formats that conform to preset rules. They cannot handle semantic association scenarios with different names but the same meaning, such as "customer ID" and "payer number". The matching accuracy is low and a large number of false associations are easily generated. At the same time, such schemes can only identify direct associations between columns and cannot explore multi-hop indirect associations between tables, making it difficult to build a complete global table association network.

[0005] In recent years, large language models have been widely used in enterprise intelligent data question answering scenarios. Although large language models have powerful natural language understanding and logical reasoning capabilities, they cannot automatically infer semantic relationships between isolated heterogeneous tables that lack explicit associations, and they struggle to accurately execute complex cross-table query reasoning. This severely restricts the effectiveness and reliability of intelligent data question answering.

[0006] Therefore, how to automatically, efficiently, and accurately construct a network of relationships between heterogeneous tables, freeing us from the reliance on manual configuration, while adapting to the application requirements of intelligent data query, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned technical problems by providing a method and system for automatically constructing heterogeneous table relationships based on graph neural networks.

[0008] In view of this, the present invention provides a method for automatically constructing heterogeneous table association relationships based on graph neural networks, comprising the following steps: Obtain data from multiple heterogeneous tables, and extract the semantic features of the table headers and the data distribution features of each data column in each table. Based on the semantic features of the table header and the data distribution features, calculate the probability of foreign key associations between data columns in different tables; Each heterogeneous table is abstracted as a node in a graph network. The edge weights between corresponding nodes are determined based on the foreign key association probability of matching column pairs between different tables, and an initial cross-table semantic association graph is constructed. The initial cross-table semantic association graph is input into a graph neural network for graph structure reasoning, the global semantic representation of the table nodes is updated and the edge weights between nodes are optimized, and the optimized multi-table association network is output.

[0009] Furthermore, the extraction of the semantic features of the headers of each data column in each table specifically involves: A pre-trained language model is used to semantically vectorize the column names of the table, generating semantic feature vectors for the corresponding column headers.

[0010] Furthermore, the extraction of data distribution characteristics from each data column in each table specifically involves: Extract and encode at least one statistical feature from the value range, mean, variance, and quartile values ​​of numerical data within a column, or extract and encode regular format pattern features from text data within a column, to generate a data distribution feature vector for the corresponding column.

[0011] Furthermore, the calculation of the foreign key association probability between data columns in different tables specifically includes: Calculate the cosine similarity between the semantic feature vectors of the headers of different tables, and filter column pairs whose original cosine similarity is higher than a preset similarity threshold as potential matching column pairs; The foreign key association probability of the corresponding column pair is obtained by weighting the semantic matching degree of the column header and the matching degree of the data distribution based on the potential matching column pairs.

[0012] Furthermore, the step of determining the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables specifically involves: Take the maximum value of the foreign key association probability of all matching columns between the two tables, and use it as the edge weight between the corresponding nodes of the two tables.

[0013] Furthermore, the graph neural network employs a graph convolutional network or a graph attention network; The graph structure reasoning specifically involves: iteratively aggregating the feature information of multi-hop neighbor nodes of each node through a multi-layer network, mining the potential semantic associations between indirectly connected table nodes, updating the global semantic representation of each table node and optimizing the edge weights between nodes, and filtering out false association edges whose weights are lower than a preset threshold after optimization.

[0014] Furthermore, the method also includes a cross-table query adaptation step: When a cross-table query request is received, based on the optimized multi-table association network, the local table sub-networks that are semantically related to the query request are extracted using a personalized PageRank algorithm, and the joint semantic context of the corresponding cross-table query is generated and output.

[0015] An automatic system for constructing heterogeneous table relationships based on graph neural networks, applied to the above methods, includes: The feature extraction module is used to acquire multiple heterogeneous table data and extract the semantic features of the table headers and the data distribution features of each data column in each table. The association calculation module is used to calculate the foreign key association probability between data columns of different tables based on the semantic features of the table header and the data distribution features; The graph construction module is used to abstract each heterogeneous table into a node in a graph network, determine the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables, and construct an initial cross-table semantic association graph. The graph reasoning module is used to input the initial cross-table semantic association graph into the graph neural network for graph structure reasoning, update the global semantic representation of table nodes and optimize the edge weights between nodes, and output the optimized multi-table association network.

[0016] Furthermore, the association calculation module is specifically used to calculate the cosine similarity between the semantic feature vectors of the headers of different tables, filter out potential matching column pairs according to a preset similarity threshold, and calculate the foreign key association probability by weighting the semantic matching degree of the headers and the matching degree of the data distribution of the potential matching column pairs. The graph construction module is specifically used to take the maximum value of the foreign key association probability of all matching columns between two tables, and use it as the edge weight between the corresponding two table nodes.

[0017] Furthermore, the graph reasoning module uses a graph convolutional network or a graph attention network to iteratively aggregate the feature information of multi-hop neighbor nodes of each node through a multi-layer network, mine the potential semantic associations between indirectly connected table nodes, update the global semantic representation of table nodes and optimize the edge weights between nodes, and filter out false association edges whose weights are lower than a preset threshold after optimization.

[0018] Furthermore, it also includes a query adaptation module, which, upon receiving a cross-table query request, extracts the local table sub-networks semantically related to the query request based on the optimized multi-table relationship network and uses a personalized PageRank algorithm to generate and output the joint semantic context of the corresponding cross-table query.

[0019] The beneficial effects of this invention are: This application adopts a two-dimensional matching method of header semantic features and data distribution features, which breaks through the technical limitations of traditional character matching and can accurately identify column-level semantic relationships with different names and similar meanings. Then, it uses a graph neural network to perform global structural reasoning, and performs noise reduction and optimization on the initial association results to filter out false associations with low confidence, effectively ensuring the reliability of the association relationship.

[0020] This solution eliminates the need for manual configuration of JOIN relationships or foreign key constraints, automatically performing feature extraction, relationship calculation, and network updates for heterogeneous tables. When enterprises add new business tables or adjust field structures, the system can quickly and automatically iterate and maintain relationships, significantly reducing the manpower costs and maintenance cycles of data integration. It is adaptable to application scenarios where enterprise business table structures change frequently, improving data governance efficiency.

[0021] This application utilizes the multi-hop information aggregation capability of graph neural networks to infer potential indirect relationships between tables based on existing direct relationships. This breaks through the limitation of traditional solutions that can only identify direct relationships, and can construct a global multi-table relationship network with a wider coverage and more complete business logic, providing a complete relationship foundation for complex cross-table data analysis.

[0022] This application combines a personalized PageRank subgraph extraction mechanism to quickly filter out local table association networks relevant to the query semantics for specific cross-table query requests, generating accurate joint semantic context for large language models. This approach not only solves the pain point of large language models being unable to automatically identify inter-table relationships, but also avoids contextual redundancy caused by full table information, effectively improving the reasoning efficiency and accuracy of intelligent data question answering. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the automatic construction method for heterogeneous table associations based on graph neural networks proposed in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] To overcome the technical shortcomings of existing database data integration scenarios, such as reliance on manual configuration of JOIN relationships for heterogeneous table relationships, high false association rates in single-dimensional matching, and inability to automatically identify multi-hop indirect associations, resulting in low efficiency and insufficient accuracy in building multi-table association networks, this embodiment provides an automatic construction method for heterogeneous table association relationships based on graph neural networks. Through a two-level architecture of "column-level dual-feature coarse matching graph construction + graph neural network global inference optimization", it achieves automated and high-precision construction of heterogeneous table association relationships.

[0026] Please see Figure 1 The method for automatically constructing heterogeneous table relationships based on graph neural networks provided in this embodiment specifically includes the following steps: S1. Obtain data from multiple heterogeneous tables, and extract the semantic features of the table headers and the data distribution features of each data column in each table.

[0027] This embodiment is applied to an enterprise database data integration scenario. The multiple heterogeneous table data are structured data tables that are stored separately in different business systems, such as sales order tables, inventory detail tables, financial payment collection tables, supplier information tables, etc. The column naming rules, data formats, and data types of each table are different, and no unified foreign key association constraints are pre-configured.

[0028] The process of extracting semantic features of the table header is as follows: a pre-trained language model is used to perform semantic vectorization encoding on the column name text of each data column. In this embodiment, the SBERT pre-trained model can be used to convert column names in natural language form such as "customer ID" and "payer number" into semantic feature vectors of fixed dimensions (such as 768 dimensions) to capture the semantics of the business entity behind the column name and solve the semantic matching barrier of different names of the same entity in different tables.

[0029] The data distribution feature extraction process is as follows: features are extracted and encoded for data columns of different data types. For numerical data columns, at least one statistical feature is extracted from the range of values, mean, variance, and quartile values. For text data columns, format patterns are identified through regular expression matching, such as ID card format, mobile phone number format, email format, fixed-length numeric number format, etc., and the format matching results are encoded into a one-hot vector. All data distribution features are normalized to the [0,1] interval by min-max and then concatenated into a fixed-dimensional data distribution feature vector.

[0030] The semantic feature vector of the header of the same data column is concatenated with the normalized data distribution feature vector to obtain the comprehensive feature vector of the data column, which is used for subsequent association matching calculation.

[0031] S2. Based on the semantic features of the table header and the data distribution features, calculate the foreign key association probability between the data columns of different tables.

[0032] This step uses two-dimensional feature matching to evaluate whether two columns in different tables point to the same business entity and outputs the probability value of a foreign key relationship between the column pairs, with a value range of [0,1].

[0033] In practice, the cosine similarity between the semantic feature vectors of any two column headers in different tables is first calculated. The higher the similarity value, the higher the degree of semantic matching between the two columns. The preset similarity threshold is 0.7. Column pairs with an original cosine similarity of more than 0.7 between the semantic feature vectors of the two column headers are selected as potential matching column pairs, completing coarse-grained screening and reducing the amount of subsequent calculations.

[0034] For the selected potential matching column pairs, scores for header semantic matching and data distribution matching are calculated separately. The header semantic matching score is the value of the cosine similarity of the semantic feature vectors of the two column headers after linear mapping and normalization to the [0,1] interval (mapping formula: p = (sim + 1) / 2). The data distribution matching score is the value of the cosine similarity of the feature vectors of the two data distributions after normalization using the same mapping. A weight of 0.6 is assigned to the header semantic matching score, and a weight of 0.4 is assigned to the data distribution matching score. The foreign key association probability corresponding to the column pair is calculated by weighted summation.

[0035] S3. Abstract each heterogeneous table into a node in a graph network. Determine the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables, and construct an initial cross-table semantic association graph.

[0036] This step transforms scattered, independent tables into structured graph network data, providing an input foundation for subsequent global graph structure reasoning; In practice, each business table is abstracted as an independent node in a graph network. The initial features of a node are obtained by aggregating the comprehensive feature vectors of all data columns in the table through mean pooling. For any two tables with potentially matching column pairs, the foreign key association probabilities of all matching column pairs between the two tables are obtained, and the maximum value is taken as the weight of the edge between the two table nodes; if there are no matching column pairs between the two tables, no edge is established.

[0037] For example, there are two sets of matching column pairs between the sales order table and the financial payment table: the foreign key association probability of "Customer ID - Payer ID" is 0.92, and the foreign key association probability of "Contact Mobile Number - Customer Phone Number" is 0.85. The maximum value of 0.92 is taken as the edge weight between the sales order table node and the financial payment table node.

[0038] By combining all table nodes with their associated edges using the above method, an initial cross-table semantic association graph is constructed. This graph initially establishes semantic associations between tables, but issues such as spurious association noise and missing multi-hop indirect associations still exist, requiring further optimization through subsequent graph neural network inference.

[0039] S4. Input the initial cross-table semantic association graph into the graph neural network for graph structure reasoning, update the global semantic representation of the table nodes and optimize the edge weights between nodes, and output the optimized multi-table association network.

[0040] This step utilizes the global information transmission and aggregation capabilities of graph neural networks to denoise and complete the initial association graph, uncover multi-hop indirect associations, and improve the accuracy and completeness of the association network.

[0041] In its implementation, the graph neural network employs a two-layer graph attention network (GAT), which can also be replaced by a graph convolutional network (GCN). The network training phase utilizes supervised training: pre-labeled real table-to-table association pairs are used as labels (the labeled data can originate from existing database foreign key constraint definitions or business documents), and an initial cross-table semantic association graph is used as input. The network parameters are optimized using a cross-entropy loss function. This training is a one-time offline process; after training, the model parameters are fixed, and subsequent inference phases require no manual labeling or configuration to automatically complete the association inference for new tables.

[0042] During graph-structured reasoning, each layer of the GAT network performs information transfer and aggregation operations: each table node aggregates the feature information of its direct neighbors, and assigns differentiated attention weights to different neighbors through an attention mechanism to strengthen the information contribution of highly related neighbors; after two layers of network iteration, each node can aggregate the feature information of its second-order neighbors, thereby mining the potential semantic relationships between indirectly connected table nodes. For example, if the sales order table and the inventory details table are strongly related, and the inventory details table and the supplier information table are strongly related, the network can infer the potential semantic relationship between the sales order table and the supplier information table. Specifically, based on the updated global semantic representation of each node in the graph neural network, the cosine similarity between all non-directly connected table node pairs is calculated, and new related edges are established for node pairs with similarity exceeding a preset threshold for adding edges, thereby achieving automatic discovery of potential indirect relationships.

[0043] After updating the global semantic representation of each node using a graph neural network, the feature cosine similarity between any two nodes with an edge is recalculated. This cosine similarity is then normalized to the [0,1] interval using a linear mapping and fused with the original edge weights at a weighted ratio of 0.5:0.5 to obtain the optimized edge weights. Since the node representation updated by the graph neural network incorporates global structural information, the recalculated cosine similarity may be lower than the initial value. Therefore, some initial edge weights may drop below the filtering threshold after fusion and updating. A preset edge weight filtering threshold of 0.3 is used; edges with optimized edge weights below 0.3 are identified as false associations and removed. The final output is an optimized multi-table association network, which includes both direct associations and indirect associations derived through reasoning, with false associations effectively filtered out.

[0044] In an optional implementation, the method further includes a cross-table query adaptation step: When a user's natural language cross-table query request is received, based on the optimized multi-table association network, the local table sub-networks that are semantically related to the query request are extracted using a personalized PageRank algorithm, and the corresponding joint semantic context for the cross-table query is generated and output.

[0045] In practical implementation, the same pre-trained language model used for header feature extraction is first employed to semantically vectorize the user's cross-table query request, resulting in a query semantic vector. The cosine similarity between the query semantic vector and the global semantic representation of each table node is calculated. The top-K table nodes with the highest similarity are selected as seed nodes (K is a preset parameter). Initial probabilities are evenly distributed among the seed nodes, while the initial probabilities of the remaining nodes are set to zero, thus forming the initial probability distribution for the personalized PageRank algorithm. A damping coefficient of 0.85 and an iteration convergence threshold of 1e-6 are set, and the personalized PageRank algorithm is executed to calculate the association score between each table node and the query request. A score threshold of 0.1 is set, and table nodes with scores higher than the threshold and their associated edges are extracted to form a local table sub-network.

[0046] The table structure information, column-level association information, and inter-table semantic association information contained in the local table sub-network are integrated into a joint semantic context and provided to the large language model, enabling the large language model to perform cross-table query reasoning based on accurate inter-table associations.

[0047] For example, when a user queries "What are the corresponding supplier payment cycles for the top three best-selling products in East China last month?", the system automatically locates and extracts a local sub-network consisting of the sales order table, product information table, supplier information table, and financial payment table through the subgraph extraction steps described above. The system then uses the table structure and relationships as context input to the large language model, assisting the model in generating accurate cross-table query results. This approach uses only the local table network relevant to the query as context, ensuring complete relational information required for reasoning while significantly reducing context redundancy, thus improving the query reasoning efficiency and accuracy of the large language model.

[0048] Secondly, this embodiment also provides an automatic construction system for heterogeneous table association relationships based on graph neural networks, used to implement the automatic construction method for heterogeneous table association relationships described in any of the above claims, the system comprising: The feature extraction module is used to acquire multiple heterogeneous table data and extract the semantic features of the table headers and the data distribution features of each data column in each table. The association calculation module is used to calculate the foreign key association probability between data columns of different tables based on the semantic features of the table header and the data distribution features; The graph construction module is used to abstract each heterogeneous table into a node in a graph network, determine the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables, and construct an initial cross-table semantic association graph. The graph reasoning module is used to input the initial cross-table semantic association graph into the graph neural network for graph structure reasoning, update the global semantic representation of table nodes and optimize the edge weights between nodes, and output the optimized multi-table association network.

[0049] In one optional implementation, the association calculation module is specifically used to calculate the cosine similarity between the semantic feature vectors of the headers of different tables, filter out potential matching column pairs according to a preset similarity threshold, and calculate the foreign key association probability by weighting the semantic matching degree of the headers and the matching degree of the data distribution of the potential matching column pairs. The graph construction module is specifically used to take the maximum value of the foreign key association probability of all matching columns between two tables, and use it as the edge weight between the corresponding two table nodes.

[0050] In one optional implementation, the graph reasoning module uses a graph convolutional network or a graph attention network to iteratively aggregate the feature information of multi-hop neighbor nodes of each node through a multi-layer network, mine the potential semantic associations between indirectly connected table nodes, update the global semantic representation of the table nodes and optimize the edge weights between nodes, and filter out false association edges whose weights are lower than a preset threshold after optimization.

[0051] In an optional implementation, the system further includes a query adaptation module, which, upon receiving a cross-table query request, extracts local table sub-networks semantically related to the query request based on an optimized multi-table association network and a personalized PageRank algorithm, generates and outputs the joint semantic context of the corresponding cross-table query.

[0052] The specific execution process of each module of the system described in this embodiment is consistent with the corresponding steps in the aforementioned method embodiment, and has the corresponding technical effects, so it will not be repeated here.

[0053] It should be noted that the method described in this embodiment can be implemented by a computer program, which can be stored in a computer-readable storage medium and executed by the processor of a computing device to implement the above method steps.

[0054] In summary, this embodiment completes the initial association screening through column-level dual-feature coarse matching, and then uses a graph neural network for global structural reasoning, completion, and noise reduction. Compared with the traditional single-dimensional column name matching method, this method can significantly improve the accuracy of inter-table association identification, effectively reduce the false association rate, and completely eliminate the dependence on manual configuration of inter-table JOIN relationships. It can also adapt to frequent changes in table structure. At the same time, combined with the subgraph extraction mechanism, it provides accurate cross-table semantic context for large language models, providing reliable table association technology support for enterprise-level intelligent data question answering and data integration governance. It is suitable for various technical scenarios such as large-scale enterprise database integration and intelligent data query.

[0055] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for automatically constructing heterogeneous table relationships based on graph neural networks, characterized in that, Includes the following steps: Obtain data from multiple heterogeneous tables, and extract the semantic features of the table headers and the data distribution features of each data column in each table. Based on the semantic features of the table header and the data distribution features, calculate the probability of foreign key associations between data columns in different tables; Each heterogeneous table is abstracted as a node in a graph network. The edge weights between corresponding nodes are determined based on the foreign key association probability of matching column pairs between different tables, and an initial cross-table semantic association graph is constructed. The initial cross-table semantic association graph is input into a graph neural network for graph structure reasoning, the global semantic representation of the table nodes is updated and the edge weights between nodes are optimized, and the optimized multi-table association network is output.

2. The method for automatically constructing heterogeneous table association relationships based on graph neural networks according to claim 1, characterized in that, The extraction of the semantic features of the headers of each data column in each table specifically involves: A pre-trained language model is used to semantically vectorize the column names of the table, generating semantic feature vectors for the corresponding column headers.

3. The method for automatically constructing heterogeneous table associations based on graph neural networks according to claim 1, characterized in that, The extraction of data distribution characteristics from each data column in each table specifically involves: Extract and encode at least one statistical feature from the value range, mean, variance, and quartile values ​​of numerical data within a column, or extract and encode regular format pattern features from text data within a column, to generate a data distribution feature vector for the corresponding column.

4. The method for automatically constructing heterogeneous table associations based on graph neural networks according to claim 1, characterized in that, The calculation of the foreign key association probability between data columns in different tables specifically includes: Calculate the cosine similarity between the semantic feature vectors of the headers of different tables, and filter column pairs whose original cosine similarity is higher than a preset similarity threshold as potential matching column pairs; The foreign key association probability of the corresponding column pair is obtained by weighting the semantic matching degree of the column header and the matching degree of the data distribution based on the potential matching column pairs.

5. The method for automatically constructing heterogeneous table association relationships based on graph neural networks according to claim 1, characterized in that, The step of determining the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables is as follows: Take the maximum value of the foreign key association probability of all matching columns between the two tables, and use it as the edge weight between the corresponding nodes of the two tables.

6. The method for automatically constructing heterogeneous table associations based on graph neural networks according to claim 1, characterized in that, The graph neural network employs either a graph convolutional network or a graph attention network. The graph structure reasoning specifically involves: iteratively aggregating the feature information of multi-hop neighbor nodes of each node through a multi-layer network, mining the potential semantic associations between indirectly connected table nodes, updating the global semantic representation of each table node and optimizing the edge weights between nodes, and filtering out false association edges whose weights are lower than a preset threshold after optimization.

7. The method for automatically constructing heterogeneous table associations based on graph neural networks according to claim 1, characterized in that, The method also includes a cross-table query adaptation step: When a cross-table query request is received, based on the optimized multi-table association network, the local table sub-networks that are semantically related to the query request are extracted using a personalized PageRank algorithm, and the joint semantic context of the corresponding cross-table query is generated and output.

8. An automatic system for constructing heterogeneous table relationships based on graph neural networks, applied to the automatic construction method for constructing heterogeneous table relationships based on graph neural networks as described in any one of claims 1-7, characterized in that, include: The feature extraction module is used to acquire multiple heterogeneous table data and extract the semantic features of the table headers and the data distribution features of each data column in each table. The association calculation module is used to calculate the foreign key association probability between data columns of different tables based on the semantic features of the table header and the data distribution features; The graph construction module is used to abstract each heterogeneous table into a node in a graph network, determine the edge weights between corresponding nodes based on the foreign key association probability of matching column pairs between different tables, and construct an initial cross-table semantic association graph. The graph reasoning module is used to input the initial cross-table semantic association graph into the graph neural network for graph structure reasoning, update the global semantic representation of table nodes and optimize the edge weights between nodes, and output the optimized multi-table association network.

9. The automatic construction system for heterogeneous table associations based on graph neural networks according to claim 8, characterized in that, The association calculation module is specifically used to calculate the cosine similarity between the semantic feature vectors of the headers of different tables, filter out potential matching column pairs according to a preset similarity threshold, and calculate the foreign key association probability by weighting the semantic matching degree of the headers and the matching degree of the data distribution of the potential matching column pairs. The graph construction module is specifically used to take the maximum value of the probability of foreign key association between all matching columns between two tables, and use it as the edge weight between the corresponding two table nodes; The graph reasoning module uses a graph convolutional network or a graph attention network as the graph neural network. It is used to iteratively aggregate the feature information of multi-hop neighbor nodes of each node through a multi-layer network, mine the potential semantic associations between indirectly connected table nodes, update the global semantic representation of table nodes and optimize the edge weights between nodes, and filter out false association edges whose weights are lower than a preset threshold after optimization.

10. The automatic construction system for heterogeneous table associations based on graph neural networks according to claim 8, characterized in that, It also includes a query adaptation module, which, when a cross-table query request is received, extracts the local table sub-networks that are semantically related to the query request based on the optimized multi-table relationship network and a personalized PageRank algorithm, generates the joint semantic context of the corresponding cross-table query, and outputs it.

Images