A graph structure optimization learning method based on large language model inference enhancement

Abstract

The application discloses a graph structure optimization learning method based on large language model reasoning enhancement. The method comprises the following steps: constructing an original graph structure comprising a node set, an edge set and a node text attribute; constructing node pair instruction data based on the node text attribute; performing graph structure reasoning fine-tuning training on the large language model by using the node pair instruction data; obtaining node semantic representation based on the trained large language model; constructing an edge prediction network to generate a candidate edge set; fusing the original graph structure and the candidate edge set to form a to-be-optimized graph structure; performing reasoning judgment on the edge connection relationship in the to-be-optimized graph structure by using the large language model to obtain an optimized graph structure; and inputting the optimized graph structure into a graph neural network to perform a downstream graph learning task. The semantic reasoning capability of the large language model is introduced, and the graph structure quality, the node relationship modeling capability and the performance of a downstream node classification task are improved.

Description

Technical Field

[0001] This invention relates to the fields of graph structure learning, graph neural networks, and large language models, specifically a graph structure optimization learning method based on large language model reasoning enhancement. Background Technology

[0002] Graph-structured data is widely used in scenarios such as social networks, recommender systems, knowledge graphs, biological networks, and academic citation networks. As the scale of data continues to expand, removing irrelevant noise and mining potential relationships between nodes to improve the performance of downstream learning tasks becomes increasingly challenging.

[0003] Graph neural networks can effectively learn node representations through neighbor information aggregation mechanisms, achieving good results in tasks such as node classification, link prediction, and graph classification. Graph structure learning techniques optimize graph structures by reconstructing adjacency matrices, which can alleviate graph noise problems to some extent.

[0004] However, existing graph neural network methods generally rely on pre-constructed graph structures as supervision information. When there are a large number of noisy edges, missing edges, and erroneous connections in the graph structure, the node representation learning performance will degrade. Existing graph structure learning methods typically use node feature similarity, graph convolutional representation, or attention mechanisms to establish node connections, and their optimization process is highly dependent on the existing graph structure.

[0005] With the development of large language models, they possess powerful knowledge understanding, logical reasoning, and semantic association analysis capabilities. Large language models can learn rich prior knowledge from text and establish potential relationships between different entities.

[0006] However, existing graph structure learning methods often lack the ability to deeply understand the semantic information of node texts, making it difficult to discover potential implicit relationships between nodes using external knowledge.

[0007] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0008] To address the problems in related technologies, this invention proposes a graph structure optimization learning method based on large language model reasoning enhancement to overcome the aforementioned technical issues in existing related technologies. The specific technical solution adopted by this invention is as follows:

[0009] A graph structure optimization learning method based on large language model reasoning enhancement, characterized by the following steps:

[0010] S1. Define the original graph structure as follows: The set of nodes is denoted as Let N be the number of nodes and let the adjacency matrix be denoted as N. The set of node text attributes is denoted as Construct the original graph structure containing node connection relationships and text semantic information;

[0011] S2, Define the training node set as Randomly sample node pairs from the training node set. A node consistency instruction dataset is constructed based on node category labels;

[0012] S3. Supervised fine-tuning training of the pre-trained large language model is performed using the node consistency instruction dataset to obtain the graph structure inference model. ;

[0013] S4. Input the node text attributes into the graph structure reasoning model. Extract the set of semantic representations of nodes. And construct the node semantic feature space;

[0014] S5. Using node semantics to represent sets Train the edge prediction network to obtain the connection probability matrix between nodes. ;

[0015] S6. Based on the connection probability matrix Filter Top- for each node Candidate edges, and their relationship with the original adjacency matrix. Fusion generates candidate graph structure ;

[0016] S7. Structure the candidate graph Input graph structure reasoning model By leveraging the reasoning capabilities of a large language model, edge connectivity relationships are determined, and an optimized adjacency matrix is ​​generated. ;

[0017] S8. Construct an optimized graph structure And input the graph neural network to perform the node classification task;

[0018] S9. Output the node classification results and the optimized graph structure.

[0019] Furthermore, the node consistency instruction data construction method in S2 includes the following steps:

[0020] S2-1, Define Nodes The corresponding text attribute is The node category label is ;

[0021] S2-2, From the training node set Random sampling node pairs Node pairs constitute the samples;

[0022] S2-3, Define node consistency labels When the node category satisfies hour, otherwise ;

[0023] S2-4. Change the node text attribute and Fill the preset instruction template to generate a node consistency judgment task;

[0024] S2-5, Utilizing Node Consistency Tags Together with node category labels, supervised training samples are constructed to form a graph structure inference instruction dataset.

[0025] Furthermore, the edge prediction network construction method in S5 includes the following steps:

[0026] S5-1, Define the semantic representation vector of nodes. and , respectively corresponding to nodes With nodes ;

[0027] S5-2. Concatenate the semantic representations of nodes to obtain the node pair feature vector. ;

[0028] S5-3. Input the node pair feature vectors into the edge prediction network. The probability of edge existence is obtained. ;

[0029] S5-4, Define the real edge label as Construct the cross-entropy loss function ;

[0030] S5-5, Using the loss function Update the edge prediction network parameters to obtain the connection probability matrix. .

[0031] Furthermore, the candidate graph structure construction method in S6 includes the following steps:

[0032] S6-1, Define the connection probability matrix as follows: The original adjacency matrix is ​​defined as follows: Define the candidate connection matrix output by the edge prediction network as follows: For each node, select the Top-ranked node with the highest connection probability. Establish edge connections between candidate nodes to obtain the candidate connection matrix. ;

[0033] S6-2, Transfer candidate connection matrix With the original adjacency matrix By fusion, candidate graph structures are obtained. .

[0034] Furthermore, the graph structure optimization reasoning method in S7 includes the following steps:

[0035] S7-1. Define the set of node pairs in the candidate graph structure as follows: And extract the node text information and edge connection information corresponding to the node pairs;

[0036] S7-2. Construct reasoning hints from node text information and edge connection information, and input them into the graph structure reasoning model. ;

[0037] S7-3, Using Graph Structure Reasoning Models Determine the connection relationships between node pairs to obtain edge connection prediction results;

[0038] S7-4. Delete invalid edge connections in the candidate graph structure based on the edge connection prediction results;

[0039] S7-5. Based on the edge connection prediction results, retain valid edge connections and establish new edge connection relationships;

[0040] S7-6. Update the edge connection relationships and generate an optimized adjacency matrix. ;

[0041] S7-7. Construct an optimized graph structure based on the optimized adjacency matrix. The graph neural network is then used to perform downstream graph learning tasks.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) A graph structure optimization method based on large language model reasoning enhancement is proposed, which can make full use of the semantic knowledge in the node text attributes to optimize the edge connection relationship in the graph structure and improve the quality of the graph structure;

[0044] (2) By fusing the candidate edge information generated by the edge prediction network with the reasoning results of the large language model, potential associations can be discovered and noisy connections can be eliminated, thereby improving the ability to model node relationships;

[0045] (3) Using the optimized graph structure for graph neural network training can effectively improve the accuracy of downstream node classification tasks and enhance the robustness and generalization performance of the model in noisy environments and sparse graph structure scenarios. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the technical solution proposed in this invention. Detailed Implementation

[0048] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0049] According to embodiments of the present invention, an interpretable recommendation method based on multimodal retrieval path enhancement of a large language model is provided, such as... Figure 1 As shown, this method is applied to a recommendation system composed of user historical interaction data and a large language model, and is performed according to the following steps:

[0050] S1. Define the original graph structure as follows: The set of nodes is denoted as Let N be the number of nodes and let the adjacency matrix be denoted as N. The set of node text attributes is denoted as Construct the original graph structure containing node connection relationships and text semantic information;

[0051] S2, Define the training node set as Randomly sample node pairs from the training node set. A node consistency instruction dataset is constructed based on node category labels;

[0052] S2-1, Define Nodes The corresponding text attribute is The node category label is ;

[0053] S2-2, From the training node set Random sampling node pairs Node pairs constitute the samples;

[0054] S2-3, Define node consistency labels When the node category satisfies hour, otherwise ;

[0055] S2-4. Change the node text attribute and Fill the preset instruction template to generate a node consistency judgment task;

[0056] S2-5, Utilizing Node Consistency Tags Together with node category labels, supervised training samples are constructed to form a graph structure inference instruction dataset;

[0057] S3. Supervised fine-tuning training of the pre-trained large language model is performed using the node consistency instruction dataset to obtain the graph structure inference model. ;

[0058] S4. Input the node text attributes into the graph structure reasoning model. Extract the set of semantic representations of nodes. And construct the node semantic feature space;

[0059] S5. Using node semantics to represent sets Train the edge prediction network to obtain the connection probability matrix between nodes. ;

[0060] S5-1, Define the semantic representation vector of nodes. and , respectively corresponding to nodes With nodes ;

[0061] S5-2. Concatenate the semantic representations of nodes to obtain the node pair feature vector. ;

[0062] S5-3. Input the node pair feature vectors into the edge prediction network. The probability of edge existence is obtained. ;

[0063] S5-4, Define the real edge label as Construct the cross-entropy loss function ;

[0064] S5-5, Using the loss function Update the edge prediction network parameters to obtain the connection probability matrix. ;

[0065] S6. Based on the connection probability matrix Filter Top- for each node Candidate edges, and their relationship with the original adjacency matrix. Fusion generates candidate graph structure ;

[0066] S6-1, Define the connection probability matrix as follows: The original adjacency matrix is ​​defined as follows: Define the candidate connection matrix output by the edge prediction network as follows: For each node, select the Top-ranked node with the highest connection probability. Establish edge connections between candidate nodes to obtain the candidate connection matrix. ;

[0067] S6-2, Transfer candidate connection matrix With the original adjacency matrix By fusion, candidate graph structures are obtained. ;

[0068] S7. Structure the candidate graph Input graph structure reasoning model By leveraging the reasoning capabilities of a large language model, edge connectivity relationships are determined, and an optimized adjacency matrix is ​​generated. ;

[0069] S7-1. Define the set of node pairs in the candidate graph structure as follows: And extract the node text information and edge connection information corresponding to the node pairs;

[0070] S7-2. Construct reasoning hints from node text information and edge connection information, and input them into the graph structure reasoning model. ;

[0071] S7-3, Using Graph Structure Reasoning Models Determine the connection relationships between node pairs to obtain edge connection prediction results;

[0072] S7-4. Delete invalid edge connections in the candidate graph structure based on the edge connection prediction results;

[0073] S7-5. Based on the edge connection prediction results, retain valid edge connections and establish new edge connection relationships;

[0074] S7-6. Update the edge connection relationships and generate an optimized adjacency matrix. ;

[0075] S7-7. Construct an optimized graph structure based on the optimized adjacency matrix. And input the graph neural network to perform downstream graph learning tasks;

[0076] S8. Construct an optimized graph structure And input the graph neural network to perform the node classification task;

[0077] S9. Output the node classification results and the optimized graph structure;

[0078] S9-1. Input the optimized graph structure into the graph neural network model for node representation learning. The output node representation of the graph neural network is denoted as... ;

[0079] S9-2, Node-based representation Perform a node classification task, obtain the final node category prediction results, and output the node classification results and the optimized graph structure.

Claims

1. A graph structure optimization learning method based on large language model inference enhancement, characterized in that, The method includes the following steps: S1, define the original graph structure as , the node set is denoted as , the number of nodes is denoted as N, the adjacency matrix is denoted as , and the node text attribute set is denoted as , and an original graph structure containing node connection relationships and text semantic information is constructed; S2, define a training node set as randomly sample a node pair sample set from the training node set construct a node consistency instruction data set according to the node category label S3, performing supervised fine-tuning training on the pre-trained large language model by using the node consistency instruction data set to obtain a graph structure reasoning model ; S4. Input the node text attributes into the graph structure reasoning model. Extract the set of semantic representations of nodes. And construct the node semantic feature space; S5、Utilizing the set of node semantic representations training the edge prediction network to obtain the inter-node connection probability matrix ; S6. Based on the connection probability matrix Filter Top- for each node Candidate edges, and their relationship with the original adjacency matrix. Fusion generates candidate graph structure ; S7. Structure the candidate graph Input graph structure reasoning model By leveraging the reasoning capabilities of a large language model, edge connectivity relationships are determined, and an optimized adjacency matrix is ​​generated. ; S8. Construct an optimized graph structure And input the graph neural network to perform the node classification task; S9. Output the node classification results and the optimized graph structure.

2. The graph structure optimization learning method based on large language model inference enhancement according to claim 1, characterized in that, The method for constructing node consistency instruction data in S2 includes the following steps: S2-1, define node The corresponding text attribute is , and the node category label is ; S2-2, randomly sampling node pairs from the set of training nodes comprising randomly sampling node pairs from the set of training nodes to form the node pair sample; S2-3, define node consistency label When the node category satisfies , Otherwise ; S2-4, filling the node text attribute with into the preset instruction template, generating a node consistency judgment task; S2-5, Utilizing Node Consistency Tags Together with node category labels, supervised training samples are constructed to form a graph structure inference instruction dataset.

3. The graph structure optimization learning method based on large language model reasoning enhancement according to claim 1, characterized in that, The method for constructing the edge prediction network in S5 includes the following steps: S5-1, Define the semantic representation vector of nodes. and , respectively corresponding to nodes With nodes ; S5-2, splicing the node semantic representations to obtain a node pair feature vector ; S5-3, input the node pair feature vector into the edge prediction network , to obtain the edge existence probability ; S5-4, define the real edge label as , construct a cross-entropy loss function ; S5-5, utilizing a loss function updating the side prediction network parameters to obtain a connection probability matrix .

4. The graph structure optimization learning method based on large language model inference enhancement according to claim 1, characterized in that, The candidate graph structure construction method in S6 includes the following steps: S6-1, Define the connection probability matrix as follows: The original adjacency matrix is ​​defined as follows: Define the candidate connection matrix output by the edge prediction network as follows: For each node, select the Top-ranked node with the highest connection probability. Establish edge connections between candidate nodes to obtain the candidate connection matrix. ; S6-2, Transfer candidate connection matrix With the original adjacency matrix By fusion, candidate graph structures are obtained. .

5. The graph structure optimization learning method based on large language model reasoning enhancement according to claim 1, characterized in that, The S7 graph structure optimization reasoning method includes the following steps: S7-1, define a node pair set in a candidate graph structure as and extract node text information and edge connection information corresponding to the node pair. S7-2, construct the node text information and the edge connection information into inference prompt information, and input into a graph structure inference model ; S7-3, utilizing the graph structure reasoning model judging the connection relationship between the node pairs to obtain an edge connection prediction result; S7-4. Delete invalid edge connections in the candidate graph structure based on the edge connection prediction results; S7-5. Based on the edge connection prediction results, retain valid edge connections and establish new edge connection relationships; S7-6, update the side connection relationship, and generate an optimized adjacency matrix ; S7-7, constructing an optimized graph structure according to the optimized adjacency matrix and input the graph neural network to perform a downstream graph learning task.

Images