Block adjustment measure recommendation method and system based on graph neural network

Abstract

The invention discloses a block adjustment measure recommendation method and system based on a graph neural network, and relates to the technical field of oil field block adjustment measure recommendation, and the method comprises the steps: obtaining historical adjustment measure recommendation data, and constructing an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph and an adjustment measure co-occurrence graph; extracting text features corresponding to the evaluation indexes, the evaluation blocks and the adjustment measures by adopting a BERT pre-training language model, inputting the text features and the evaluation block-adjustment measure interaction diagram into a GCN model, and calculating high-order embedding features of the evaluation blocks and the adjustment measures; inputting the evaluation block collaboration graph and the adjustment measure co-occurrence graph into a LightGCN model, and calculating collaboration embedding features of the evaluation blocks and the adjustment measures; obtaining final embedded features of the evaluation blocks and the adjustment measures through feature fusion; and predicting a recommendation probability score through inner product calculation, and generating an adjustment measure recommendation list. The accuracy of the block adjustment measure recommendation task can be improved.

Description

Technical Field

[0001] This application relates to the technical field of oilfield block adjustment measure recommendation, and particularly to a block adjustment measure recommendation method and system based on a graph neural network. Background Art

[0002] As an important chemical fuel and fossil fuel, petroleum occupies a crucial position in the national economy. However, with the continuous exploitation of petroleum resources, the discovered main oilfields are gradually entering the late stage of development, resulting in increased oil production difficulty and cost. Improving oil recovery has become the key to the sustainable development of the petroleum industry. In the late stage of oilfield development, a series of oilfield block adjustment measures, such as fracturing, perforation repair, water injection conversion, etc., can slow down the reduction of reserves and the decline rate of production, and improve the economic benefits of the oilfield. However, with the extension of oilfield development time, various dynamic data increase, and the internal influencing factors of the reservoir are numerous and the relationships are intricate, resulting in a high difficulty in the design and selection of oil well measure adjustment plans, thus reducing the efficiency of block adjustment measure recommendation.

[0003] Currently, conventional block adjustment measure recommendation methods mainly rely on the experience of oilfield experts. It takes a large amount of time and cost to formulate a relatively precise and detailed block adjustment measure recommendation plan. Currently, existing block adjustment measure recommendation methods include regression statistical models, decision trees, fuzzy logic theory, knowledge graph-based methods, collaborative filtering-based methods, etc. However, these block adjustment measure recommendation methods only stay at the surface feature extraction and analysis of historical block evaluation data, and cannot capture implicit semantic information and structural relationships. Therefore, the accuracy of block adjustment measure recommendation tasks cannot be guaranteed. Based on this, there is an urgent need for a method that can fully mine complex historical block evaluation data and more accurately implement block adjustment measure recommendation tasks to improve the accuracy of block adjustment measure recommendation tasks. Summary of the Invention

[0004] The purpose of this application is to provide a block adjustment measure recommendation method and system based on a graph neural network, which can improve the accuracy of block adjustment measure recommendation tasks.

[0005] To achieve the above purpose, this application provides the following solutions.

[0006] In the first aspect, this application provides a block adjustment measure recommendation method based on a graph neural network. The block adjustment measure recommendation method based on a graph neural network specifically includes the following steps.

[0007] Obtain historical adjustment measure recommendation data; the historical adjustment measure recommendation data includes text data corresponding to evaluation indicators, evaluation blocks, and adjustment measures.

[0008] According to the historical adjustment measure recommendation data, an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph are respectively constructed.

[0009] According to the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph, using the BERT pre-trained language model, feature extraction is performed on the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures to obtain the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures.

[0010] The text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures and the evaluation block-adjustment measure interaction graph are respectively input into the GCN model to calculate the high-order embedding features of the evaluation blocks and the high-order embedding features of the adjustment measures.

[0011] The evaluation block collaboration graph and the adjustment measure co-occurrence graph are respectively input into the LightGCN model to calculate the collaborative embedding features of the evaluation blocks and the collaborative embedding features of the adjustment measures.

[0012] Feature fusion is respectively performed between the high-order embedding features of the evaluation blocks and the collaborative embedding features of the evaluation blocks, and between the high-order embedding features of the adjustment measures and the collaborative embedding features of the adjustment measures to obtain the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures.

[0013] According to the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, an inner product calculation is performed to predict the recommended probability score, and a recommended list of adjustment measures for each evaluation block is generated.

[0014] In a second aspect, the present application provides a block adjustment measure recommendation system based on a graph neural network, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the above-mentioned block adjustment measure recommendation method based on a graph neural network.

[0015] According to the specific embodiments provided by the present application, the present application has the following technical effects.

[0016] The present application provides a method and system for recommending block adjustment measures based on graph neural networks, which applies graph neural networks to the specific field and specific scenario of block adjustment measures recommendation. In the process of block adjustment measures recommendation, the BERT pre-trained language model, the GCN model and the LightGCN model are combined, and various graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph and the adjustment measure co-occurrence graph are constructed as inputs of the above models. On the one hand, the BERT pre-trained language model is used to capture the complex semantic features in the text data corresponding to the evaluation indicators, evaluation blocks and adjustment measures, which can capture the rich contextual information around the node text, which is conducive to understanding the position and relationship of the node in the entire text sequence, thereby helping to improve the accuracy of the block adjustment measure recommendation task. On the other hand, using the GCN model and the LightGCN model to respectively learn the embedded representations of evaluation blocks and adjustment measures in graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph will help to further obtain richer node features, capture more complex relationships between nodes, and explore the potential associations between evaluation blocks and adjustment measures, thereby obtaining more accurate high-order embedding features, collaborative embedding features, and fused final embedding features of evaluation blocks and adjustment measures, and thus being able to accurately predict the recommendation probability score and generate an accurate and reliable adjustment measure recommendation list, thereby improving the accuracy of the block adjustment measure recommendation task. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 This is an application environment diagram of a method for recommending block adjustment measures based on a graph neural network provided in one embodiment of the present application.

[0019] Figure 2 A flowchart of a method for recommending block adjustment measures based on a graph neural network is provided in accordance with an embodiment of the present application.

[0020] Figure 3 A schematic diagram of the structure of a block adjustment measure recommendation system based on a graph neural network provided in one embodiment of the present application. DETAILED DESCRIPTION

[0021] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0022] In the related art, there is a method for recommending oilfield production enhancement measures based on random forest and gradient boosting decision tree. The method includes the following steps: 1) Collect historical data and divide it into an implementation data set and an implementation effect data set with the moment of implementing the production enhancement measure as the boundary; 2) Perform data preprocessing on the implementation data set under each historical production enhancement measure, including: data cleaning, calculating the recovery ratio, adding production enhancement measure effect classification labels, and sampling the training set and test set; 3) Screen out important feature parameters from the training set and test set respectively; 4) Construct and train an implementation effect classification prediction model; 5) Recommend production wells that can increase production to the production enhancement measures. Due to the limitations of the characteristics of offshore oilfields, this method uses big data machine learning algorithms to mine the correlation between parameters in dimensions such as the formation, wellbore, and implementation design of production wells and specific production enhancement measures, realizes the recommendation of production wells under effective production enhancement measures, ensures the success of the recommended measures to the greatest extent, breaks through the technical difficulties of currently difficult to support multi-dimensional comprehensive analysis to achieve production increase technically, gives reliable and effective recommended measures to guide the production increase of offshore oilfields, promotes the intelligence of production increase, and ensures the safe, stable, and efficient operation of production wells. In the related art, there is also an automatic recommendation method for oil well production enhancement measures. The method includes: obtaining sample data of oil well production enhancement measures, digitally transforming the text data in the sample data through the One-hot encoding method; performing measure sensitive factor analysis on the sample data to obtain the main sensitive factors; using the main sensitive factors as the recommendation basis, and analyzing the measures taken by historical similar wells and the effects of the measures, and pushing the corresponding optimized measures. The analysis includes a matching stage and a ranking stage; realizing measure recommendation based on the collaborative filtering-based measure recommendation method. This method can more accurately and efficiently select the best production enhancement plan for common measures such as fracturing, acidizing, and perforating and re-completing intervals by fully mining a large amount of measure well monitoring data and combining machine learning algorithms.

[0023] However, although the above-mentioned existing methods for recommending block adjustment measures have achieved certain results in structured data analysis, they rely too much on static historical data and manual feature screening. In addition, there may be a situation where similar historical data is missing, resulting in the failure of the recommendation. Moreover, the above-mentioned existing methods for recommending block adjustment measures do not make full use of the attribute information in each block, ignore complex deep semantic associations and implicit interaction relationships, and have weak generalization in sparse data scenarios.

[0024] A Graph Neural Network (GNN) combines graph data with neural networks to perform end-to-end computations on graph data, capturing complex relationships in the graph data through interactions between nodes. This embodiment aims to provide a method and system for recommending block adjustment measures based on a graph neural network. By combining various graph structures such as an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph with various neural network models such as the BERT pre-trained language model, the GCN (Graph Convolutional Network) model, and the LightGCN (Light Graph Convolutional Network) model, and applying them to the scenario of recommending block adjustment measures, potential associations between evaluation blocks and measures are mined from a large amount of historical adjustment measure recommendation data, thereby improving the accuracy of the block adjustment measure recommendation task and solving the problems existing in the above-mentioned existing block adjustment measure recommendation methods.

[0025] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The method for recommending block adjustment measures based on a graph neural network provided by the embodiments of the present application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set separately, integrated on the server 104, placed on the cloud or other servers. The terminal 102 can send the historical adjustment measure recommendation data to the server 104. After receiving the historical adjustment measure recommendation data, for the historical adjustment measure recommendation data, the server 104 constructs an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph; uses the BERT pre-trained language model to extract the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and inputs them together with the evaluation block-adjustment measure interaction graph into the GCN model to calculate the high-order embedding features of the evaluation blocks and adjustment measures; inputs the evaluation block collaboration graph and the adjustment measure co-occurrence graph into the LightGCN model to calculate the collaborative embedding features of the evaluation blocks and adjustment measures; obtains the final embedding features of the evaluation blocks and adjustment measures through feature fusion; calculates the predicted recommendation probability score through inner product calculation, and generates an adjustment measure recommendation list. The server 104 can feedback the obtained adjustment measure recommendation list to the terminal 102. In addition, in some embodiments, the block adjustment measure recommendation method based on the graph neural network can also be implemented by the server 104 or the terminal 102 alone. For example, the terminal 102 can directly perform block adjustment measure recommendation processing on the historical adjustment measure recommendation data, or the server 104 can obtain the historical adjustment measure recommendation data from the data storage system and perform block adjustment measure recommendation processing on the historical adjustment measure recommendation data.

[0027] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0028] In an exemplary embodiment, as Figure 2 shown, a block adjustment measure recommendation method based on the graph neural network is provided. This method is executed by a computer device, and can be specifically executed by a computer device such as a terminal or a server alone, or jointly executed by a terminal and a server. In the embodiments of the present application, taking this method applied to Figure 1 the server 104 in

[0029] Step S1: Obtain historical adjustment measure recommendation data. Among them, the historical adjustment measure recommendation data includes the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures.

[0030] Step S2: Construct an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co - occurrence graph respectively based on the historical adjustment measure recommendation data.

[0031] In this embodiment, step S2 constructs an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co - occurrence graph respectively based on the historical adjustment measure recommendation data, and specifically includes the following steps.

[0032] Step S21: Determine the interaction relationship among the evaluation indicators, evaluation blocks, and adjustment measures according to the historical adjustment measure recommendation data.

[0033] Step S22: Construct an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co - occurrence graph respectively according to the historical adjustment measure recommendation data and the interaction relationship among the evaluation indicators, evaluation blocks, and adjustment measures.

[0034] In this embodiment, step S22 constructs an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co - occurrence graph respectively according to the historical adjustment measure recommendation data and the interaction relationship among the evaluation indicators, evaluation blocks, and adjustment measures, and specifically includes the following steps.

[0035] Step S221: Define the evaluation block - adjustment measure interaction graph , the evaluation block set and the adjustment measure set are respectively used as two parts of this evaluation block - adjustment measure interaction graph , then the adjacency matrix of the evaluation block - adjustment measure interaction graph is , when = 1, establish an edge between the evaluation block and the adjustment measure .

[0036] Step S222: Define the evaluation block collaboration graph , in this evaluation block collaboration graph , both of the two interacting parts are evaluation block nodes. When establishing the edges between each evaluation block node, use the content - based similarity calculation method, introduce the evaluation indicators of the evaluation blocks, and calculate the cosine similarity between the evaluation indicators of each evaluation block. Then the adjacency matrix of the evaluation block collaboration graph is , when , establish an edge between the evaluation block and the evaluation block , is the weight of this edge.

[0037] Step S223: Define the co-occurrence graph of adjustment measures , in this co-occurrence graph of adjustment measures , both of the two interacting parts are adjustment measure nodes. When establishing the edges between each adjustment measure node, according to the constructed evaluation block-adjustment measure interaction graph 's edge set, calculate the co-occurrence frequency between two adjustment measure nodes , then the adjacency matrix of the co-occurrence graph of adjustment measures is . When , establish an edge between adjustment measure and adjustment measure , being the weight of this edge.

[0038] Step S3: According to the evaluation block-adjustment measure interaction graph, evaluation block collaboration graph, and co-occurrence graph of adjustment measures, use the BERT pre-trained language model to extract features from the text data corresponding to evaluation indicators, evaluation blocks, and adjustment measures, and obtain the text features corresponding to evaluation indicators, evaluation blocks, and adjustment measures.

[0039] In this embodiment, step S3 uses the BERT pre-trained language model to extract features from the text data corresponding to evaluation indicators, evaluation blocks, and adjustment measures according to the evaluation block-adjustment measure interaction graph, evaluation block collaboration graph, and co-occurrence graph of adjustment measures, and obtain the text features corresponding to evaluation indicators, evaluation blocks, and adjustment measures, which specifically includes the following steps.

[0040] Step S31: Set the input sequence , where , = [node type], to are node text characters, .

[0041] Step S32: According to the input vector of each token in the input sequence , construct the input embedding matrix ; among them, the input vector of each token is represented by the following formula.

[0042] ; where represents the token embedding matrix, represents the position embedding matrix, represents the segment embedding matrix; the constructed input embedding matrix ​​, where is the total number of all nodes.

[0043] Step S33: Input the input embedding matrix into the multi-layer Transformer encoder of the BERT pre-trained language model for layer-by-layer processing, and output the text features corresponding to the evaluation metrics, evaluation blocks, and adjustment measures.

[0044] When the multi-layer Transformer encoder performs layer-by-layer processing, the multi-head self-attention is calculated using the following formula.

[0045] ; ; where is the attention of the th head. The self-attention weight matrices of the th head respectively have , , , is the weight matrix, is the input embedding matrix, is the function, is the multi-head self-attention, represents the concatenation operation.

[0046] Step S4: Input the text features corresponding to the evaluation metrics, evaluation blocks, and adjustment measures, and the evaluation block-adjustment measure interaction graph into the GCN model respectively, and calculate the high-order embedding features of the evaluation blocks and the high-order embedding features of the adjustment measures.

[0047] In this embodiment, Step S4 inputs the text features corresponding to the evaluation metrics, evaluation blocks, and adjustment measures, and the evaluation block-adjustment measure interaction graph into the GCN model respectively, and calculates the high-order embedding features of the evaluation blocks and the high-order embedding features of the adjustment measures, which specifically includes the following steps.

[0048] Step S41: Based on the evaluation block-adjustment measure interaction graph , use the GCN model to perform representation learning on the evaluation blocks and adjustment measures, and obtain the initial feature matrix of the evaluation blocks as according to the evaluation metric embedding; initialize the initial feature matrix of the adjustment measures as , where is the feature dimension of each evaluation block and adjustment measure.

[0049] Step S42: Use the GCN model to update the embedding representations of all nodes in each layer of graph convolution. The The embedding update formula of the layer evaluation block nodes is expressed by the following formula.

[0050] ; Among them, represents the high-order embedding feature of the layer evaluation block nodes of the layer, is the normalized adjacency matrix of the interaction between the evaluation block and the adjustment measures, is the embedding representation of the evaluation block after the layer update, is the feature transformation matrix of the layer, (·) is the function.

[0051] The embedding update formula of the layer adjustment measure nodes of the layer is expressed by the following formula.

[0052] ; Among them, represents the high-order embedding feature of the layer adjustment measure nodes of the layer, is the transpose of the normalized adjacency matrix of the interaction between the evaluation block and the adjustment measures, is the embedding representation of the adjustment measures after the layer update.

[0053] Step S43, after the graph convolution operation of the layer, the high-order embedding feature of the evaluation block and the high-order embedding feature of the adjustment measures are obtained.

[0054] Step S5, input the evaluation block co-occurrence graph and the adjustment measure co-occurrence graph into the LightGCN model respectively, and calculate the co-embedding feature of the evaluation block and the co-embedding feature of the adjustment measures.

[0055] In this embodiment, step S5 inputs the evaluation block co-occurrence graph and the adjustment measure co-occurrence graph into the LightGCN model respectively, and calculates the co-embedding feature of the evaluation block and the co-embedding feature of the adjustment measures, which specifically includes the following steps.

[0056] Step S51, based on the evaluation block co-occurrence graph and the adjustment measure co-occurrence graph , use the LightGCN model to process the evaluation block co-occurrence graph and the adjustment measure co-occurrence graph respectively. Let the initial feature embedding matrix of each evaluation block be , and the initial feature embedding matrix of each adjustment measure node be , where is the characteristic dimension for evaluating the block and the adjustment measure, then the co-occurrence graph of the graph convolution operation is defined as follows.

[0057] ; where represents the embedding representation of the evaluation block, is the normalized adjacency matrix of the co-occurrence graph of the evaluation block; The co-occurrence graph of the adjustment measure has its graph convolution operation defined as follows.

[0058] ; where represents the embedding representation of the adjustment measure, is the normalized adjacency matrix of the co-occurrence graph of the adjustment measure.

[0059] Step S52, After a single-layer graph convolution operation, obtain the embedding representation of the evaluation block and the embedding representation of the adjustment measure.

[0060] Step S53, Use a linear layer to map each embedding vector in the embedding representation of the evaluation block and the embedding representation of the adjustment measure to dimensions, obtaining the co-embedded feature of the evaluation block and the co-embedded feature of the adjustment measure.

[0061] Step S6, Perform feature fusion between the high-order embedding feature of the evaluation block and the co-embedded feature of the evaluation block, and between the high-order embedding feature of the adjustment measure and the co-embedded feature of the adjustment measure, respectively, to obtain the final embedding feature of the evaluation block and the final embedding feature of the adjustment measure.

[0062] In this embodiment, step S6 performs feature fusion between the high-order embedding feature of the evaluation block and the co-embedded feature of the evaluation block, and between the high-order embedding feature of the adjustment measure and the co-embedded feature of the adjustment measure, respectively, to obtain the final embedding feature of the evaluation block and the final embedding feature of the adjustment measure, which specifically includes the following steps.

[0063] Step S61, Perform splicing fusion between the high-order embedding feature of the evaluation block and the co-embedded feature of the evaluation block to obtain the final embedding feature of the evaluation block, which is represented by the following formula.

[0064] ; Among them, represents the final embedded feature of the evaluation block, is the concatenation operation of vectors, represents the high-order embedded feature of the evaluation block, represents the collaborative embedded feature of the evaluation block.

[0065] Step S62: Concatenate and fuse the high-order embedded feature of the adjustment measure and the collaborative embedded feature of the adjustment measure to obtain the final embedded feature of the adjustment measure, which is represented by the following formula.

[0066] ; Among them, represents the final embedded feature of the adjustment measure, represents the high-order embedded feature of the adjustment measure, represents the collaborative embedded feature of the adjustment measure.

[0067] Step S7: Calculate the inner product based on the final embedded feature of the evaluation block and the final embedded feature of the adjustment measure, predict the recommended probability score, and generate a recommended list of adjustment measures for each evaluation block.

[0068] In this embodiment, step S7 calculates the inner product based on the final embedded feature of the evaluation block and the final embedded feature of the adjustment measure, predicts the recommended probability score, and generates a recommended list of adjustment measures for each evaluation block, which specifically includes the following steps.

[0069] Step S71: Calculate the inner product based on the final embedded feature of the evaluation block and the final embedded feature of the adjustment measure, and predict the recommended probability score of each evaluation block for each adjustment measure, which is represented by the following formula.

[0070] ; Among them, represents the evaluation block and the adjustment measure The recommended probability score between them, represents the evaluation block The final embedded feature of, represents the adjustment measure The final embedded feature of, T represents the transpose.

[0071] Step S72: Generate a recommended list of adjustment measures for each evaluation block based on the recommended probability score of each evaluation block for each adjustment measure, which is represented by the following formula.

[0072] ; Among them, Indicates the evaluation block A list of recommended adjustment measures, Indicates that the highest recommendation probability score is selected adjustment measures, Represents a list of candidate adjustment measures.

[0073] Implement the above steps S1 to S7, by applying graph neural networks to the specific field and specific scenario of block adjustment measures recommendation. In the process of block adjustment measures recommendation, the BERT pre-trained language model, the GCN model and the LightGCN model are combined, and various graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph and the adjustment measure co-occurrence graph are constructed as the input of the above models. On the one hand, the BERT pre-trained language model is used to capture the complex semantic features in the text data corresponding to the evaluation indicators, evaluation blocks and adjustment measures, which can capture the rich contextual information around the node text, which is conducive to understanding the position and relationship of the node in the entire text sequence, thereby helping to improve the accuracy of the block adjustment measures recommendation task. On the other hand, using the GCN model and the LightGCN model to respectively learn the embedded representations of evaluation blocks and adjustment measures in graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph will help to further obtain richer node features, capture more complex relationships between nodes, and explore the potential associations between evaluation blocks and adjustment measures, thereby obtaining more accurate high-order embedding features, collaborative embedding features, and fused final embedding features of evaluation blocks and adjustment measures, and thus being able to accurately predict the recommendation probability score and generate an accurate and reliable adjustment measure recommendation list, thereby improving the accuracy of the block adjustment measure recommendation task.

[0074] In this embodiment, after performing inner product calculation according to the final embedded features of the evaluation block and the final embedded features of the adjustment measures in step S7, predicting the recommendation probability score, and generating a recommended list of adjustment measures for each evaluation block, the block adjustment measure recommendation method based on graph neural network also includes the following steps.

[0075] The BPR loss function is used to maximize the gap between the recommendation probability scores of positive samples and negative samples to improve the ability of the BERT pre-trained language model, the GCN model, and the LightGCN model to sort the adjustment measure recommendation list.

[0076] In order to make the technical solution of this embodiment clearer, the specific implementation process of this embodiment is described in detail below in the form of examples, which specifically includes the following implementation steps.

[0077] Step S1: Collect and organize the historical adjustment measure recommendation data of the evaluation block, obtain the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and determine the interaction relationships among the evaluation indicators, evaluation blocks, and adjustment measures.

[0078] Step S2: Based on the historical adjustment measure recommendation data in Step S1, construct an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph, and format the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures to be input.

[0079] The method for constructing an evaluation block - adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph in this embodiment includes the following steps.

[0080] (1) Define the evaluation block - adjustment measure interaction graph , the evaluation block set and the adjustment measure set are respectively used as the two parts of the evaluation block - adjustment measure interaction graph , then the adjacency matrix of the evaluation block - adjustment measure interaction graph is , when , an edge is established between the evaluation block and the adjustment measure to form the edge set of the evaluation block - adjustment measure interaction graph.

[0081] (2) Define the evaluation block collaboration graph , in this evaluation block collaboration graph , both of the two interacting parts are evaluation block nodes. When establishing the edges between different evaluation block nodes, consider using a content-based similarity calculation method, introduce the evaluation indicators of the evaluation blocks, and calculate the cosine similarity between the evaluation indicators of each evaluation block. Then the adjacency matrix of the evaluation block collaboration graph is , when , an edge is established between the evaluation block and the evaluation block , is the weight of this edge.

[0082] (3) Define the adjustment measure co-occurrence graph , in this adjustment measure co-occurrence graph , both of the two interacting parts are adjustment measure nodes. When establishing the edges between different adjustment measure nodes, according to the edge set of the evaluation block - adjustment measure interaction graph constructed above, calculate the co-occurrence frequency between adjustment measure pairs. Then the adjacency matrix of the adjustment measure co-occurrence graph is When happens, an edge is established between the adjustment measure and the adjustment measure , and is the weight of this edge.

[0083] Step S3: Based on Step S2, use the BERT pre-trained language model to extract features from the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and obtain the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures.

[0084] The specific method for using the BERT pre-trained language model to extract features from the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures in this embodiment is: on the basis of obtaining the text data of the nodes participating in the interaction, set the input sequence as , where , = [node type], to are the node text characters, .

[0085] The input vector of each token is superimposed by three parts, and the embedding dimension size is . The calculation formula is as follows.

[0086] ; Among them, represents the token embedding matrix, represents the position embedding matrix, represents the segment embedding matrix. Thus, the input embedding matrix is constructed, where is the number of all nodes, and it is input into the layer Transformer encoder of the BERT pre-trained language model for layer-by-layer processing, and the multi-head self-attention is calculated in the following way.

[0087] ; ; Among them, is the attention of the th head, and the self-attention weight matrices of the th head are respectively , , , is the weight matrix, is the input embedding matrix, is function, is multi-head self-attention, Represents a concatenation operation.

[0088] Concatenate all the self-attention matrices together and multiply them by the weight matrix Get multiple attention The final output is the last hidden state corresponding to each input word, that is, the evaluation index, evaluation block, and text features corresponding to the adjustment measures that integrate the context information.

[0089] The input embedding matrix will be input into the multi-layer Transformer encoder of the BERT pre-trained language model, and the input embedding matrix will be processed to learn the features and dependencies in the sequence. In step S3, the output of the BERT pre-trained language model corresponds to the last hidden state of each input word, that is, the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures that integrate the context information.

[0090] Step S4: Input the evaluation block-adjustment measure interaction diagram obtained in step S2 and the text features corresponding to the evaluation indicators, evaluation blocks and adjustment measures obtained in step S3 into the GCN model to calculate the high-order embedding features of the evaluation blocks and adjustment measures.

[0091] For the evaluation block-adjustment measure interaction diagram , the GCN model is used to learn the representation of the evaluation block and adjustment measures. The initial feature matrix of the evaluation block obtained by embedding the evaluation index is: ; The initial characteristic matrix of the initial adjustment measure is ,in The feature dimension of each evaluation block and adjustment measure. By stacking multiple layers of graph convolution, deeper graph structure information can be captured. Each layer of graph convolution updates the embedding representation of all nodes. The embedding update formula of the layer evaluation block node can be expressed as the following formula.

[0092] ; in, Indicates The high-order embedding features of the layer evaluation block nodes, is the normalized adjacency matrix of the interaction between the evaluation block and the adjustment measure, For the The embedding representation of the evaluation block after layer update, For the The feature transformation matrix of the layer, (·)for function.

[0093] Similarly, in The embedding update formula of the layer adjustment measure node can be expressed as the following formula.

[0094] ; Among them, represents the high-order embedding feature of the layer adjustment measure node at the th layer, is the transpose of the normalized adjacency matrix for evaluating the interaction between blocks and adjustment measures, is the embedding representation of the adjustment measure after the th layer update. After the graph convolution operation of the th layer, the high-order embedding feature of the evaluation block and the high-order embedding feature of the adjustment measure are obtained. These embedding representations capture the high-order graph structure information between nodes.

[0095] Step S5: Input the evaluation block collaboration graph and the adjustment measure co-occurrence graph obtained in Step S2 into the LightGCN model respectively, and calculate the collaborative embedding features of the evaluation block and the adjustment measure respectively.

[0096] For the evaluation block collaboration graph and the adjustment measure co-occurrence graph , a single-layer graph convolution network model, namely the LightGCN model, is used to process the evaluation block collaboration graph and the adjustment measure co-occurrence graph respectively to obtain the evaluation index and adjustment measure representations. Set the initial feature embedding matrix of each evaluation block as , and the initial feature embedding matrix of each adjustment measure node as , where is the feature dimension of the evaluation block and the adjustment measure. The graph convolution operation of the evaluation block collaboration graph is defined as follows.

[0097] ; Among them, represents the embedding representation of the evaluation block, is the normalized adjacency matrix of the evaluation block collaboration graph.

[0098] Similarly, the graph convolution operation of the adjustment measure co-occurrence graph is defined as follows.

[0099] ; Among them, represents the embedding representation of the adjustment measure, is the normalized adjacency matrix of the adjustment measure co-occurrence graph.

[0100] After a single-layer graph convolution operation, the embedded representation of the evaluation block is obtained and the embedded representation of the adjustment measures . Finally, the embedded representation of the evaluation block and the embedded representation of the adjustment measures are respectively mapped to dimensions through a linear layer to obtain the final representations of the co-occurrence graph of the evaluation block and the adjustment measures, including the collaborative embedding features of the evaluation block and the collaborative embedding features of the adjustment measures .

[0101] Step S6: Feature fusion is performed on the high-order embedding features of the evaluation block and the adjustment measures obtained in step S4 and the collaborative embedding features of the evaluation block and the adjustment measures obtained in step S5 respectively to obtain the final embedding features of the evaluation block and the adjustment measures.

[0102] For each node, a concatenation operation is used to fuse the final representations of the evaluation block and the adjustment measures obtained in step S4 and step S5 as shown in the following formula.

[0103] ; ; where is the concatenation operation of vectors, represents the final embedding features of the evaluation block, represents the final embedding features of the adjustment measures, which are used for subsequent recommended probability score prediction tasks.

[0104] Step S7: According to the final embedding features of the evaluation block and the adjustment measures in step S6, an inner product calculation is performed to predict the recommended probability score and generate a recommended list of adjustment measures for each evaluation block.

[0105] For any evaluation block and any adjustment measure , their final embedded representations, that is, the final embedding features are respectively and , and the recommended probability score between the evaluation block and the adjustment measure can be expressed by the following formula.

[0106] ; where represents the recommended probability score between the evaluation block and the adjustment measure , represents the final embedding features of the evaluation block , Indicates the adjustment measure of the final embedded feature, T Indicates transpose.

[0107] This recommended probability score indicates the evaluation block and the adjustment measure The potential interaction tendency between them. The above is to calculate the recommended probability score of a single evaluation block for a single adjustment measure. According to this method, the recommended probability scores of each evaluation block for all adjustment measures are calculated respectively to obtain the recommended probability score matrix . Specifically, given an evaluation block , the adjustment measure recommendation list can be expressed by the following formula.

[0108] ; Among them, Indicates the adjustment measure recommendation list of the evaluation block , Indicates selecting the one with the highest recommended probability score Adjustment measures, Indicates the adjustment measure candidate list, Indicates the evaluation block and the adjustment measure The recommended probability score between them.

[0109] Step S8: Use the BPR loss function to maximize the gap between the recommended probability scores of positive and negative samples, so as to improve the ability of each model such as the BERT pre-trained language model, GCN model, and LightGCN model to rank the adjustment measure recommendation list.

[0110] In the process of model iteration, the construction of positive and negative samples is the key link for calculating the BPR loss. For positive sample sampling, select the adjustment measures that have interacted with the evaluation block as positive samples, denoted as: . For negative sample sampling, select the adjustment measures that have not interacted with the evaluation block as negative samples, denoted as: .

[0111] For each evaluation block , given its set of positive sample adjustment measures that have interacted and the set of negative sample adjustment measures that have not interacted, the expression of the BPR loss is as follows.

[0112] ; Among them, Indicates the BPR loss, Represents the triples in the training data; is the sigmoid function, which is used to map the difference of recommendation probability scores to the probability space; Indicates the evaluation block Recommended probability scores for negative samples for sampling adjustment measures, is the model parameter Regularization term, To control the regularization strength.

[0113] This embodiment cleverly integrates multiple graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph to improve the quality and accuracy of the adjustment measure recommendation. The BERT pre-trained language model, the GCN model, and the LightGCN model are combined to learn the embedded representations of various graph structures, and the learned embedded representations are fused to enhance the expressive power of the evaluation block and adjustment measure node features, which helps to obtain a more appropriately sorted list of adjustment measure recommendations. This makes the model applicable to the problem of oilfield block adjustment measure recommendations, effectively alleviates the sparsity of data interaction, improves the accuracy of block adjustment measure recommendations, and provides a new method for oilfield block adjustment measure recommendations.

[0114] This embodiment fully considers the interactive sparsity of oilfield block adjustment measures data and designs a block adjustment measures recommendation method based on graph neural network. In order to capture the complex interactive relationship between the evaluation block and the adjustment measures and realize the efficient block adjustment measures recommendation task, the graph neural network is applied to the field of block adjustment measures recommendation to provide a reference for block adjustment measures recommendation. On the one hand, the BERT pre-trained language model is used to capture complex semantic features and can capture rich contextual information around the node text, which is conducive to understanding the position and relationship of the node in the entire text sequence, thereby improving the accuracy of the block adjustment measures recommendation task. On the other hand, using the GCN model and the LightGCN model to respectively learn the embedded representations of evaluation blocks and adjustment measures in graph structures such as the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph will help to further obtain richer node features, capture more complex relationships between nodes, and explore the potential associations between evaluation blocks and adjustment measures, thereby obtaining more accurate high-order embedding features, collaborative embedding features, and fused final embedding features of evaluation blocks and adjustment measures, and thus being able to accurately predict the recommendation probability score and generate an accurate and reliable adjustment measure recommendation list, thereby improving the accuracy of the block adjustment measure recommendation task.

[0115] In an exemplary embodiment, a block adjustment measure recommendation system based on a graph neural network is provided. The system may be a computer device, which may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 3As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store historical adjustment measure recommendation data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a method for recommending block adjustment measures based on a graph neural network.

[0116] Those skilled in the art can understand that Figure 3 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0117] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0118] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0119] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for recommending block adjustment measures based on a graph neural network, characterized in that, The block adjustment measure recommendation method based on graph neural network includes: Obtain historical adjustment measure recommendation data; the historical adjustment measure recommendation data includes evaluation indicators, evaluation blocks, and text data corresponding to adjustment measures; Construct an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph respectively according to the historical adjustment measure recommendation data; According to the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph, use the BERT pre-trained language model to extract features from the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and obtain the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures; Input the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and the evaluation block-adjustment measure interaction graph into the GCN model respectively to calculate the high-order embedding features of the evaluation blocks and the high-order embedding features of the adjustment measures; Input the evaluation block collaboration graph and the adjustment measure co-occurrence graph into the LightGCN model respectively to calculate the collaborative embedding features of the evaluation blocks and the collaborative embedding features of the adjustment measures; Fuse the features between the high-order embedding features of the evaluation blocks and the collaborative embedding features of the evaluation blocks, and between the high-order embedding features of the adjustment measures and the collaborative embedding features of the adjustment measures respectively to obtain the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures; Perform inner product calculation according to the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, predict the recommended probability score, and generate an adjustment measure recommendation list for each evaluation block.

2. The method for recommending block adjustment measures based on a graph neural network according to claim 1, characterized in that, Construct an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph respectively according to the historical adjustment measure recommendation data, specifically including: Determine the interaction relationship between the evaluation indicators, evaluation blocks, and adjustment measures according to the historical adjustment measure recommendation data; Construct an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph respectively according to the historical adjustment measure recommendation data and the interaction relationship between the evaluation indicators, evaluation blocks, and adjustment measures.

3. The method for recommending block adjustment measures based on a graph neural network according to claim 2, characterized in that, Construct an evaluation block-adjustment measure interaction graph, an evaluation block collaboration graph, and an adjustment measure co-occurrence graph respectively according to the historical adjustment measure recommendation data and the interaction relationship between the evaluation indicators, evaluation blocks, and adjustment measures, specifically including: Define the evaluation block - adjustment measure interaction graph , the set of evaluation blocks and the set of adjustment measures are respectively used as the two parts of this evaluation block - adjustment measure interaction graph , then the adjacency matrix of the evaluation block - adjustment measure interaction graph is , when = 1, establish an edge between the evaluation block and the adjustment measure . Define the evaluation block collaboration graph , in this evaluation block collaboration graph , both of the two interacting parts are evaluation block nodes. When establishing the edges between each evaluation block node, a content-based similarity calculation method is used, introducing various evaluation indicators of the evaluation block, and calculating the cosine similarity between the evaluation indicators of each evaluation block , then the adjacency matrix of the evaluation block collaboration graph is , when , an edge is established between the evaluation block and the evaluation block , is the weight of this edge; Define the co-occurrence graph of adjustment measures , in this co-occurrence graph of adjustment measures , both of the two interacting parts are adjustment measure nodes. When establishing the edges between each adjustment measure node, according to the edge set of the constructed evaluation block-adjustment measure interaction graph , calculate the co-occurrence frequency between two adjustment measure nodes , then the adjacency matrix of the co-occurrence graph of adjustment measures is , when , establish an edge between adjustment measure and adjustment measure , is the weight of this edge.

4. The method for recommending block adjustment measures based on a graph neural network according to claim 1, characterized in that, According to the evaluation block-adjustment measure interaction graph, the evaluation block collaboration graph, and the adjustment measure co-occurrence graph, use the BERT pre-trained language model to extract features from the text data corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, specifically including: Set the input sequence , where , = [node type], to are node text characters, ; According to the input sequence and each token in it , construct the input embedding matrix ; among them, each token in it is represented as: ； Among them, represents the token embedding matrix, represents the position embedding matrix, represents the segment embedding matrix; the constructed input embedding matrix , where is the number of all nodes; Embed the input into a matrix Input it into the multi-layer Transformer encoder of the BERT pre-trained language model for layer-by-layer processing, and output the text features corresponding to the evaluation metrics, evaluation blocks, and adjustment measures; When the multi-layer Transformer encoder processes layer by layer, use the following formula to calculate the multi-head self-attention: ； ； Among them, is the attention of the th head, and the self-attention weight matrices of the th head respectively have , , . is the weight matrix, is the input embedding matrix, is function, is the multi-head self-attention, represents the concatenation operation.

5. The method for recommending block adjustment measures based on a graph neural network according to claim 1, characterized in that, Input the text features corresponding to the evaluation indicators, evaluation blocks, and adjustment measures, and the evaluation block-adjustment measure interaction graph into the GCN model respectively to calculate the high-order embedding features of the evaluation blocks and the high-order embedding features of the adjustment measures, specifically including: Based on the evaluation block-adjustment measure interaction diagram , use the GCN model to perform representation learning on evaluation blocks and adjustment measures, and obtain the initial feature matrix of the evaluation block according to the evaluation index embedding as ; Initialize the initial feature matrix of the adjustment measure as , where is the feature dimension of each evaluation block and adjustment measure; Use the GCN model to update the embedding representations of all nodes through graph convolution at each layer. The embedding update formula for evaluating block nodes in the th layer is expressed as: ； Among them, represents the high-order embedding feature of the layer evaluation block node, is the normalized adjacency matrix of the interaction between the evaluation block and the adjustment measure, is the embedding representation of the evaluation block after the layer update, is the feature transformation matrix of the layer, (·) is the function; The embedding update formula of the layer adjustment measure node is expressed as: ； Among them, represents the high-order embedding feature of the layer adjustment measure node, is the transpose of the normalized adjacency matrix for evaluating the interaction between the evaluation block and the adjustment measure, is the embedding representation of the adjusted measure after the layer update; After performing graph convolution operations on the evaluation block, the high-order embedded features of the evaluation block and the high-order embedded features of the adjustment measures are obtained.

6. The method for recommending block adjustment measures based on a graph neural network according to claim 1, wherein, Input the co-occurrence graph of the evaluation blocks and the co-occurrence graph of the adjustment measures into the LightGCN model respectively, and calculate the collaborative embedding features of the evaluation blocks and the collaborative embedding features of the adjustment measures, specifically including: Based on the collaborative graph of evaluation blocks and the co-occurrence graph of adjustment measures , use the LightGCN model to process the collaborative graph of evaluation blocks and the co-occurrence graph of adjustment measures respectively. Let the initial feature embedding matrix of each evaluation block be , and the initial feature embedding matrix of each adjustment measure node be . Among them is the feature dimension of the evaluation block and the adjustment measure. Then the graph convolution operation of the collaborative graph of evaluation blocks is defined as: ； Among them, represents the embedded representation of the evaluation block, is the normalized adjacency matrix of the evaluation block collaboration graph; The co-occurrence graph of the adjustment measures The graph convolution operation is defined as: ； Among them, represents the embedded representation of the adjustment measure, is the normalized adjacency matrix of the adjustment measure collaboration graph; After a single-layer graph convolution operation, the embedded representation of the evaluation block is obtained and the embedded representation of the adjustment measures ; Use a linear layer to separately map the embedding representations of the evaluation block and the embedding representations of the adjustment measures so that each embedding vector in them is mapped to dimensions, obtaining the collaborative embedding features of the evaluation block and the collaborative embedding features of the adjustment measures .

7. The method for recommending block adjustment measures based on a graph neural network according to claim 1, wherein, Perform feature fusion between the high-order embedding features of the evaluation blocks and the collaborative embedding features of the evaluation blocks, and between the high-order embedding features of the adjustment measures and the collaborative embedding features of the adjustment measures respectively, to obtain the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, specifically including: Perform concatenation fusion between the high-order embedding features of the evaluation blocks and the collaborative embedding features of the evaluation blocks to obtain the final embedding features of the evaluation blocks, denoted as: ； Among them, represents the final embedded feature of the evaluation block, is the concatenation operation of vectors, represents the high-order embedded feature of the evaluation block, represents the collaborative embedded feature of the evaluation block; Perform concatenation fusion between the high-order embedding features of the adjustment measures and the collaborative embedding features of the adjustment measures to obtain the final embedding features of the adjustment measures, denoted as: ； Among them, represents the final embedded feature of the adjustment measure, represents the high-order embedded feature of the adjustment measure, represents the collaborative embedded feature of the adjustment measure.

8. The method for recommending block adjustment measures based on a graph neural network according to claim 1, wherein, Perform inner product calculation based on the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, predict the recommended probability scores, and generate a recommended list of adjustment measures for each evaluation block, specifically including: Perform inner product calculation based on the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, and predict the recommended probability scores of each evaluation block for each adjustment measure, denoted as: ； Among them, represents the recommendation probability score between the evaluation block and the adjustment measure, represents the final embedding feature of the evaluation block, represents the final embedding feature of the adjustment measure, T represents transpose; Generate a recommended list of adjustment measures for each evaluation block based on the recommended probability scores of each evaluation block for each adjustment measure, denoted as: ； Among them, represents the recommended list of adjustment measures for the evaluation block , represents selecting the adjustment measures with the highest recommended probability score represents the candidate list of adjustment measures.

9. The method for recommending block adjustment measures based on a graph neural network according to claim 1, wherein, After the step of performing inner product calculation based on the final embedding features of the evaluation blocks and the final embedding features of the adjustment measures, predicting the recommended probability scores, and generating a recommended list of adjustment measures for each evaluation block, the method for recommending block adjustment measures based on a graph neural network further includes: Use the BPR loss function to maximize the gap between the recommended probability scores of positive samples and negative samples, so as to improve the ability of the BERT pre-trained language model, the GCN model, and the LightGCN model to rank the recommended list of adjustment measures.

10. A system for recommending block adjustment measures based on a graph neural network, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the method for recommending block adjustment measures based on a graph neural network according to any one of claims 1-9.

Images