An efficient data retrieval method and system based on AI

Abstract

This invention discloses an efficient data retrieval method and system based on AI. It integrates multi-dimensional data and entity associations to construct a multi-dimensional retrieval graph and uncover collaborative retrieval paths. Based on a GNN model, it implements node feature aggregation and edge embedding attention mechanisms to highlight core association dimensions, effectively capturing implicit associations between data and making the embedded vectors more closely aligned with multi-dimensional collaborative features. A real-time computing framework enables dynamic updates of the embedded vectors and graph structure, providing rapid responses to changes in multi-dimensional data. A secondary sorting mechanism integrating Euclidean distance, cosine similarity, and multi-dimensional collaborative weights prioritizes the output of resources with stronger multi-dimensional collaboration, reducing user filtering costs. Based on user feedback and a graph structure and model optimization mechanism, it continuously learns user demand characteristics in different scenarios, constantly optimizing retrieval performance.

Description

Technical Field

[0001] This invention relates to the field of data retrieval, specifically to an efficient data retrieval method and system based on AI. Background Technology

[0002] With the advent of the big data era, the scale of data is growing exponentially, and the types of data are becoming increasingly complex. Most existing data retrieval technologies rely on single-dimensional data, failing to adequately consider multi-dimensional and comprehensive data, resulting in search results that struggle to match users' deeper needs. Furthermore, existing methods largely depend on text matching and simple feature association, lacking the ability to uncover the multi-dimensional collaborative relationships between "users, interests, intent, and influence," thus failing to identify implicit connections between data and leading to one-sided search results that fail to cover users' potential needs. Therefore, designing an efficient AI-based data retrieval method and system to improve the efficiency and accuracy of data retrieval is essential. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by proposing an efficient data retrieval method and system based on AI. It mainly solves the problems that existing methods are unable to integrate multi-dimensional data to improve data retrieval efficiency and lack the ability to capture deep data correlations.

[0004] The technical solution to achieve the objective of this invention is: an efficient data retrieval method based on AI, comprising the following steps: Collect multi-dimensional and full-dimensional data and construct a retrieval graph by mapping nodes and edges according to the relationships between data types; The GNN model is used to aggregate the neighbor information of nodes in the retrieval graph, generate the association features between different types of nodes, and construct the multi-dimensional fusion embedding vector of users; at the same time, the neighbor association information of data resource nodes is aggregated to generate the embedding vector of data resource nodes; an edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weight of each associated edge; the path search capability of GNN is used to explore multi-dimensional collaborative paths. The system receives user search requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vectors and the data resource node embedding vectors, and selects the top-N similar data resources. The top-N results are then sorted a second time using multi-dimensional collaborative weights. Establish a dynamic graph structure update mechanism to automatically identify entities in newly added multi-dimensional data and add corresponding nodes and edges, and then recalculate the embedding vector; collect user feedback data on search results, and adjust the aggregation weights and edge association strength thresholds of the GNN model based on the feedback data. This application also proposes an efficient AI-based data retrieval system, including a data acquisition module, a path mining module, a retrieval ranking module, and an optimization feedback module. The data acquisition module collects multi-dimensional and multi-dimensional data and constructs a retrieval graph by mapping nodes and edges according to the relationships between data types. The path mining module uses a GNN model to aggregate neighbor information of nodes in the retrieval graph, generate association features between different types of nodes, and construct a multi-dimensional fusion embedding vector for the user. Simultaneously, it aggregates neighbor association information of data resource nodes to generate embedding vectors for data resource nodes. An edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weights of each associated edge. The system leverages the path search capabilities of GNNs. The system leverages multi-dimensional collaborative paths. The retrieval and ranking module receives user search requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vector and the data resource node embedding vector, selecting the top-N similar data resources. It then performs a secondary ranking of the top-N results using multi-dimensional collaborative weights. The optimization feedback module establishes a dynamic graph structure update mechanism, automatically identifying entities in newly added multi-dimensional data and adding corresponding nodes and edges, subsequently recalculating the embedding vectors. It collects user feedback data on the search results and adjusts the aggregation weights of the GNN model and the edge association strength threshold based on the feedback data. Compared with existing technologies, the beneficial effects of this invention are as follows: This invention integrates multi-dimensional and multi-dimensional data and entity associations to construct a multi-dimensional retrieval graph and mine collaborative retrieval paths; it implements node feature aggregation and edge embedding attention mechanisms based on a GNN model, highlighting core association dimensions and effectively capturing implicit associations between data, making the embedding vectors more closely aligned with multi-dimensional collaborative features; it combines a real-time computing framework to achieve dynamic updates of the embedding vectors and graph structure, enabling rapid responses to changes in multi-dimensional data; it integrates a secondary sorting mechanism that combines Euclidean distance, cosine similarity, and multi-dimensional collaborative weights, prioritizing the output of resources with stronger multi-dimensional collaboration, reducing user filtering costs; and it continuously learns user demand characteristics under different scenarios based on user feedback graph structure and model optimization mechanisms, constantly optimizing retrieval performance. Attached Figure Description

[0005] Figure 1 This is a flowchart of an efficient data retrieval method based on AI, as described in an embodiment of this application. Figure 2 This is a block diagram of an AI-based high-efficiency data retrieval system according to an embodiment of this application. Detailed Implementation

[0006] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0007] like Figure 1As shown in the figure, a specific embodiment of the present invention discloses an efficient data retrieval method based on AI, comprising the following steps: Collect multi-dimensional and full-dimensional data and construct a retrieval graph by mapping nodes and edges according to the relationships between data types; The GNN model is used to aggregate the neighbor information of nodes in the retrieval graph, generate the association features between different types of nodes, and construct the multi-dimensional fusion embedding vector of users; at the same time, the neighbor association information of data resource nodes is aggregated to generate the embedding vector of data resource nodes; an edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weight of each associated edge; the path search capability of GNN is used to explore multi-dimensional collaborative paths. The system receives user search requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vectors and the data resource node embedding vectors, and selects the top-N similar data resources. The top-N results are then sorted a second time using multi-dimensional collaborative weights. A dynamic graph structure update mechanism is established to automatically identify entities in newly added multi-dimensional data and add corresponding nodes and edges, and then recalculate the embedding vector; user feedback data on search results is collected, and the aggregation weights of the GNN model and the edge association strength threshold are adjusted in reverse based on the feedback data.

[0008] This embodiment provides an efficient AI-based data retrieval method. It integrates multi-dimensional data and entity associations to construct a multi-dimensional retrieval graph and uncover collaborative retrieval paths. Based on a GNN model, it implements node feature aggregation and edge embedding attention mechanisms to highlight core association dimensions, effectively capturing implicit relationships between data and making the embedding vectors more closely aligned with multi-dimensional collaborative features. A real-time computing framework enables dynamic updates of the embedding vectors and graph structure, providing rapid responses to changes in multi-dimensional data. A secondary sorting mechanism integrating Euclidean distance, cosine similarity, and multi-dimensional collaborative weights prioritizes the output of resources with stronger multi-dimensional collaboration, reducing user filtering costs. Based on user feedback and a graph structure and model optimization mechanism, it continuously learns user demand characteristics in different scenarios, constantly optimizing retrieval performance.

[0009] Collect multi-dimensional and full-dimensional data and construct a retrieval graph by mapping nodes and edges according to the relationships between data types; In some preferred embodiments, the collected multi-dimensional data includes user-dimensional data, interest-dimensional data, intent-dimensional data, influence-dimensional data, and data resource characteristics. User-dimensional data includes basic user attributes, behavioral history, and account preference settings. Interest-dimensional data includes topics followed by the user, preferred content formats, and intensity of interest. Intent-dimensional data includes search scenarios, search targets, and implicit needs. Influence-dimensional data includes the dissemination scope, authority, timeliness, and user reviews of data resources. Data resource data includes resource metadata, content characteristics, and associated resources.

[0010] The collected multi-dimensional data undergoes data processing, including deduplication, missing value imputation, outlier removal, and standardization. The processed multi-dimensional data is then mapped to corresponding nodes, with attributes of each dimension used as feature vectors for the nodes. User-dimensional data is mapped to user nodes, interest-dimensional data to interest nodes, intent-dimensional data to intent nodes, influence-dimensional data to influence nodes, and data resource data to data resource nodes. Subsequently, edges are established between nodes based on semantic matching and behavioral association analysis. For example, if a user clicks multiple times on content related to a specific interest topic, an edge is established between that user node and the corresponding interest node. Semantic matching between the interest topic and a specific intent is then performed. When the match degree is greater than a threshold, an edge association is constructed between the interest node and the intent node. The final constructed retrieval graph includes the following relationships between edges: user-interest edge, interest-intent edge, intent-data resource edge, data resource-influence edge, user-intent edge, and interest-data resource edge. Among them, the user-interest edge represents the user's attention relationship to a certain interest topic; the interest-intent edge represents the correspondence between the interest topic and the retrieval intent; the intent-data resource edge represents the matching relationship between the retrieval intent and the data resource; the data resource-influence edge represents the association relationship between the data resource and the influence indicator; the user-intent edge represents the retrieval intent indirectly associated by the user through the interest topic; and the interest-data resource edge represents the data resource indirectly associated by the interest topic through the retrieval intent.

[0011] The GNN model is used to aggregate the neighbor information of nodes in the retrieval graph, generate the association features between different types of nodes, and construct a multi-dimensional fusion embedding vector. At the same time, the neighbor association information of data resource nodes is aggregated to generate the embedding vector of data resource nodes. An edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weight of each associated edge. The path search capability of GNN is used to explore multi-dimensional collaborative paths.

[0012] In some preferred embodiments, a hybrid GNN model combining GAT and GCN is employed. GAT is used to capture the attention weights of a node's neighbors, while GCN is used to efficiently aggregate neighbor features. The input to the hybrid GNN model is the feature vector of the node. The intermediate layer includes three GAT layers and two GCN layers. The attention weights between nodes are obtained through the GAT layers, and then the node features between neighboring nodes are weighted and aggregated according to the GCN layers. The output layer outputs the node features through a fully connected layer. Specifically, for user nodes, only the features of interest nodes and intent nodes among their neighboring nodes are aggregated to generate user-interest-intent association features, resulting in a user fusion embedding vector. For data resource nodes, only the features of intent nodes, influence nodes, and interest nodes among their neighboring nodes are aggregated to generate a data embedding vector. Based on the attention mechanism of the GAT layer, the attention weight of each edge is calculated, and the final weight of each associated edge is set based on the attention weight to highlight the impact of the association on retrieval matching.

[0013] In some preferred embodiments, following multi-dimensional topology rules, a path search algorithm is used to mine collaborative paths including user nodes, interest nodes, intent nodes, data resource nodes, and influencing nodes. A hybrid path search algorithm combining depth-first search and breadth-first search is employed to uncover implicit associations while avoiding missing valid paths. Specifically, starting with a user node, a breadth-first search algorithm is used to traverse all its directly associated interest nodes, forming an initial set of user-interest node pairs. Subsequently, for each interest node, a breadth-first search algorithm is used to traverse all its directly associated intent nodes, forming a three-level node path of user-interest-intent. The three-level node path is pruned by removing all edges with weights less than a preset threshold to reduce subsequent invalid search overhead.

[0014] Based on the three-level node path, a depth-first search algorithm is used to traverse data resource nodes and influencing nodes layer by layer: for each intent node, all associated data resource nodes are traversed, and data resource nodes with a matching degree greater than or equal to 0.7 are selected to ensure a strong association between resources and intents; for the selected data resource nodes, all associated influencing nodes are traversed, and influencing nodes with an edge weight greater than or equal to 0.7 are selected to ensure that the data resources in the path have effective influence value; on this basis, a backtracking mechanism is introduced: if a path does not find a node that meets the conditions in subsequent levels, it backtracks to the previous level and switches to other nodes to continue the search; finally, an initial collaborative path set is generated.

[0015] The total weight of each collaborative path in the initial collaborative path set is calculated, and collaborative paths are filtered through two methods: total weight threshold and path deduplication. The total weight threshold refers to setting a minimum threshold for the total weight of paths based on historical retrieval data, and removing all collaborative paths in the initial collaborative path set whose total weight is less than the minimum threshold. Path deduplication refers to retaining only the collaborative path with the highest total weight when multiple collaborative paths have the same data resource nodes, so as to avoid the same resource being calculated repeatedly.

[0016] The system receives user search requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vectors and the data resource node embedding vectors, and selects the top-N similar data resources. Finally, it performs a secondary sorting of the top-N results using multi-dimensional collaborative weights.

[0017] In some preferred embodiments, core words are extracted from the user's search requirements and converted into word vectors. These vectors are then combined with the user's historical behavior data to obtain a query embedding vector, which is denoted as... Subsequently, the query embedding vector and the data resource node embedding vector in the collaborative path are calculated based on Euclidean distance and cosine similarity. The overall similarity is calculated using the Euclidean distance, which is the square root of the sum of the squares of the differences between the elements of the query embedding vector and the data resource node embedding vector. The formula is as follows: , in, The Euclidean distance is represented by the cosine similarity, which is calculated by dividing the dot product of the query embedding vector and the data resource node embedding vector by the product of the magnitudes of the query embedding vector and the data resource node embedding vector. The formula is as follows: , in, Represents the magnitude of a vector; due to cosine similarity The value range is [-1, 1], Euclidean distance For ∈[0,+∞), both need to be normalized to the interval [0,1]. The specific normalization method can be to use minimum-maximum normalization, and then obtain the comprehensive similarity through weighted summation; then select the top-N data resources according to the magnitude of the comprehensive similarity.

[0018] In some preferred embodiments, multi-dimensional collaborative weights are introduced to perform secondary sorting of the TOP-N results. The analytic hierarchy process (AHP) is used to analyze each dimension in the multi-dimensional dataset, and the weights of each dimension are calculated by constructing a judgment matrix. For the TOP-N data resources, the four corresponding collaborative dimension attributes are extracted from the collaborative path and standardized to [0,1] using Z-score. Then, the collaborative score of the data resources is obtained by weighted summation based on the multi-dimensional collaborative weights and the four collaborative dimension attributes. The final data resources are obtained by secondary sorting based on the collaborative score.

[0019] A dynamic graph structure update mechanism is established to automatically identify entities in newly added multi-dimensional data and add corresponding nodes and edges, and then recalculate the embedding vector; user feedback data on search results is collected, and the aggregation weights of the GNN model and the edge association strength threshold are adjusted in reverse based on the feedback data.

[0020] In some preferred embodiments, the retrieval graph is updated accordingly based on the dynamic changes of nodes or edges. The dynamic changes of nodes include adding, deleting, and updating attributes. Adding a node means that when new multi-dimensional data is collected, a corresponding node is generated, and attribute information is added to establish a relationship with related nodes. Deleting a node means that when a node becomes invalid, its status is marked as invalid, historical data is retained for model training, and invalid nodes are filtered out during retrieval. Updating attributes means that when the attributes of a node change, the node attributes in the graph database are updated in real time. The dynamic changes of edges are similar to those of nodes and will not be described in detail here.

[0021] Collect user feedback data, including positive and negative feedback; calculate the actual matching score between the user and the data resource and the error between the score obtained from the hybrid GNN model and the score obtained from the hybrid GNN model based on the feedback data; and use the gradient descent method to adjust the aggregation weights of the hidden layer of the hybrid GNN model to minimize the error.

[0022] After generating new model parameters after each optimization, the historical model parameters are retained. When the performance of the new model fails the validation, the hybrid GNN model is updated again using the saved historical model parameters.

[0023] Furthermore, such as Figure 2As shown, this application also proposes an efficient AI-based data retrieval system, including a data acquisition module, a path mining module, a retrieval ranking module, and an optimization feedback module. The data acquisition module collects multi-dimensional and multi-dimensional data and constructs a retrieval graph by mapping nodes and edges according to the relationships between data types. The path mining module uses a GNN model to aggregate neighbor information of nodes in the retrieval graph, generate association features between different types of nodes, and construct a multi-dimensional fusion embedding vector for the user. Simultaneously, it aggregates neighbor association information of data resource nodes to generate embedding vectors for data resource nodes. An edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weights of each associated edge. The system utilizes GNN path search. The system has the capability to discover multi-dimensional collaborative paths. The retrieval and ranking module receives user retrieval requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vector and the data resource node embedding vector, selecting the top-N similar data resources. It then performs a secondary ranking of the top-N results using multi-dimensional collaborative weights. The optimization feedback module establishes a dynamic graph structure update mechanism, automatically identifying entities in newly added multi-dimensional data and adding corresponding nodes and edges, subsequently recalculating the embedding vectors. It collects user feedback data on the retrieval results and adjusts the aggregation weights of the GNN model and the edge association strength threshold based on the feedback data.

[0024] Specifically, the data acquisition module collects multi-dimensional data, including user-dimensional data, interest-dimensional data, intent-dimensional data, influence-dimensional data, and data resource characteristics. User-dimensional data includes basic user attributes, behavioral history, and account preference settings. Interest-dimensional data includes user-focused topics, preferred content formats, and interest intensity. Intent-dimensional data includes search scenarios, search targets, and implicit needs. Influence-dimensional data includes the dissemination scope, authority, timeliness, and user reviews of data resources. Data resource data includes resource metadata, content characteristics, and associated resources. The collected multi-dimensional data undergoes data processing, employing methods such as deduplication, missing value imputation, outlier removal, and standardization for data cleaning. Each processed multi-dimensional data item is mapped to a corresponding node, and the attributes of each dimension are used as the feature vectors of the nodes. Subsequently, edges between nodes are established based on data semantic matching and behavioral association analysis, constructing the edges in the retrieval graph.

[0025] Specifically, the path mining module adopts a hybrid GNN model combining GAT and GCN. GAT is used to capture the attention weights of a node's neighbors, while GCN is used to efficiently aggregate neighbor features. The input of the hybrid GNN model is the feature vector of the node. The intermediate layer consists of three GAT layers and two GCN layers. The attention weights between nodes are obtained through the GAT layer, and then the node features between neighboring nodes are weighted and aggregated according to the GCN layer. The output layer outputs the node features through a fully connected layer. Following multi-dimensional topology rules, the module mines collaborative paths including user nodes, interest nodes, intent nodes, data resource nodes, and influencing nodes based on the path search algorithm. A hybrid path search algorithm combining depth-first search and breadth-first search is used to generate an initial set of collaborative paths, mining implicit associations while avoiding missing effective paths. On this basis, a backtracking mechanism is introduced. If a path does not find a node that meets the conditions in subsequent levels, it backtracks to the previous level and switches to other nodes to continue the search.

[0026] The total weight of each collaborative path in the initial collaborative path set is calculated, and collaborative paths are filtered through two methods: total weight threshold and path deduplication. The total weight threshold refers to setting a minimum threshold for the total weight of paths based on historical retrieval data, and removing all collaborative paths in the initial collaborative path set whose total weight is less than the minimum threshold. Path deduplication refers to retaining only the collaborative path with the highest total weight when multiple collaborative paths have the same data resource nodes, so as to avoid the same resource being calculated repeatedly.

[0027] Specifically, the retrieval and ranking module extracts core words from the user's search requirements and converts them into word vectors. Then, it combines the user's historical behavior data to obtain query embedding vectors. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vectors and the data resource node embedding vectors in the collaborative path, and selects the top-N data resources according to the magnitude of the comprehensive similarity. Multi-dimensional collaborative weights are introduced to perform secondary ranking of the top-N results. The hierarchical analysis method is used to analyze each dimension in the multi-dimensional full-dimensional analysis, and the weight of each dimension is calculated by constructing a judgment matrix. For the top-N data resources, the corresponding four collaborative dimension attributes are extracted from the collaborative path and standardized to [0,1]. Then, the collaborative score of the data resources is obtained by weighted summation based on the multi-dimensional collaborative weights and the four collaborative dimension attributes. Based on the collaborative score, secondary ranking is performed to obtain the final data resources.

[0028] Specifically, the optimization feedback module updates the retrieval graph according to the dynamic changes of nodes or edges; collects user feedback data, including positive and negative feedback; calculates the actual matching score between the user and the data resource and the error between the score obtained from the hybrid GNN model and the score obtained from the hybrid GNN model based on the feedback data; adjusts the aggregation weight of the hidden layer of the hybrid GNN model using the gradient descent method to minimize the error; after generating new model parameters after each optimization, the historical model parameters are retained; when the performance verification of the new model fails, the hybrid GNN model is updated again using the saved historical model parameters.

[0029] In this embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform steps of an AI-based efficient data retrieval method.

[0030] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0031] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0032] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An efficient data retrieval method based on AI, characterized in that, Includes the following steps: Collect multi-dimensional and full-dimensional data and construct a retrieval graph by mapping nodes and edges according to the relationships between data types; The GNN model is used to aggregate the neighbor information of nodes in the retrieval graph, generate the association features between different types of nodes, and construct the multi-dimensional fusion embedding vector of the user; at the same time, the neighbor association information of data resource nodes is aggregated to generate the embedding vector of data resource nodes; an edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weight of each associated edge. Leveraging the path search capabilities of GNNs, we can uncover multi-dimensional collaborative paths. The system receives user search requests and generates query embedding vectors using a GNN model. Based on Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vectors and the data resource node embedding vectors, and selects the top-N similar data resources. The top-N results are then sorted a second time using multi-dimensional collaborative weights. Establish a dynamic update mechanism for the graph structure, automatically identify entities in newly added multi-dimensional data and add corresponding nodes and edges, and then recalculate the embedding vector; Collect user feedback data on search results, and adjust the aggregation weights and edge association strength thresholds of the GNN model based on the feedback data.

2. The efficient data retrieval method based on AI as described in claim 1, characterized in that, The GNN model includes: A hybrid GNN model combining GAT and GCN is used to perform node feature aggregation. For user nodes, interest nodes and intent nodes in their neighbors are aggregated to generate a multi-dimensional fusion embedding vector for the user. For data resource nodes, intent nodes, influence nodes, and interest nodes in their neighbors are aggregated to generate an embedding vector for the data resource nodes. An edge embedding mechanism is introduced, and attention weights for each associated edge are calculated through an attention mechanism. Core associated edges are assigned higher weights to highlight their association value. Leveraging the path search capabilities of the GNN model and following multi-dimensional topology rules, a hybrid algorithm combining depth-first search and breadth-first search is employed to traverse node relationships and uncover complete multi-dimensional collaborative paths that include user nodes, interest nodes, intent nodes, data resource nodes, and influencing nodes. A backtracking mechanism is introduced during the search process to prune node chains that do not meet the criteria.

3. The efficient data retrieval method based on AI as described in claim 2, characterized in that, The sorting of the data resource nodes includes the following steps: Extract core words from user search requests and convert them into word vectors, then combine them with user historical behavior data to generate query embedding vectors; The similarity between the query embedding vector and the data resource node embedding vector is calculated using both Euclidean distance and cosine similarity. The Euclidean distance is normalized, and the comprehensive similarity is obtained by weighted summation. TOP-N similar data resources were obtained based on comprehensive similarity screening; the analytic hierarchy process was used to determine the collaborative weights of each dimension, and the four collaborative dimension attributes corresponding to the TOP-N similar data resources were extracted from the collaborative path. The attributes were standardized, and the collaborative score of the data resources was calculated by weighted summation. The TOP-N similar data resources were then sorted in a secondary order based on the collaborative score.

4. The efficient data retrieval method based on AI as described in claim 3, characterized in that, The dynamic update mechanism includes: A dynamic graph structure update mechanism is established to automatically identify entities in newly added multi-dimensional data, add corresponding nodes and edges, and synchronously update the relevant node information in the retrieval graph for node addition, deletion, and attribute update operations. Similarly, for edge addition, weight adjustment, and deletion operations, the relevant edge information in the retrieval graph is synchronously updated. Combined with a real-time computing framework, the changes in multi-dimensional data are monitored. When multi-dimensional data is added, modified, or deleted, incremental calculation of embedding vectors is performed only on the nodes involved in the change and their direct neighbor nodes, avoiding full graph recalculation. The updated embedding vectors are synchronized to the graph database and model cache to ensure the timeliness of the embedding vectors.

5. The efficient data retrieval method based on AI as described in claim 4, characterized in that, The method for constructing the retrieval graph includes the following steps: The collected multi-dimensional data includes user-dimensional data, interest-dimensional data, intent-dimensional data, influence-dimensional data, and data resource data. The collected multi-dimensional data undergoes deduplication, missing value imputation, outlier removal, and standardization. Each processed multi-dimensional data item is mapped to a corresponding type of node: user-dimensional data is mapped to user nodes, interest-dimensional data to interest nodes, intent-dimensional data to intent nodes, influence-dimensional data to influence nodes, and data resource data to data resource nodes. Based on semantic matching and behavioral association analysis, edge relationships are established between nodes, constructing a retrieval graph containing user-interest edges, interest-intent edges, intent-data resource edges, data resource-influence edges, user-intent edges, and interest-data resource edges.

6. The efficient data retrieval method based on AI as described in claim 5, characterized in that, The method of adjusting the GNN model based on feedback data includes the following steps: The system collects user feedback data on search results, including both positive and negative feedback. It standardizes the collected feedback data to quantify the actual matching evaluation of data resources by users. Based on the quantified feedback data, it calculates the actual matching score between users and data resources, compares the actual matching score with the collaborative score output by the GNN model, and obtains the error between the two. It then uses gradient descent to adjust the aggregation weights of the hidden layers of the GNN model based on the error, dynamically adjusting the edge association strength threshold. Finally, it retains the model parameters after each optimization, and updates the GNN model again using historical model parameters when the performance of the new model fails validation.

7. An AI-based high-efficiency data retrieval system, characterized in that, include: The data acquisition module collects multi-dimensional and multi-dimensional data and constructs a retrieval graph by mapping nodes and edges according to the relationships between data types. The path mining module uses a GNN model to aggregate neighbor information of nodes in the retrieval graph, generate association features between different types of nodes, and construct a multi-dimensional fusion embedding vector for users. At the same time, it aggregates the neighbor association information of data resource nodes to generate embedding vectors for data resource nodes. An edge embedding mechanism is introduced, and an attention mechanism is used to calculate the attention weight of each associated edge. Leveraging the path search capabilities of GNNs, we can uncover multi-dimensional collaborative paths. The retrieval and ranking module receives the user's retrieval request and generates a query embedding vector through a GNN model; based on the dual indicators of Euclidean distance and cosine similarity, it calculates the comprehensive similarity between the query embedding vector and the data resource node embedding vector, and selects the TOP-N similar data resources; and combines multi-dimensional collaborative weights to perform a secondary ranking of the TOP-N results. The feedback module was optimized, and a dynamic graph structure update mechanism was established to automatically identify entities in newly added multi-dimensional data and add corresponding nodes and edges, and then recalculate the embedding vectors. Collect user feedback data on search results, and adjust the aggregation weights and edge association strength thresholds of the GNN model based on the feedback data.

Images