Graph structure reconstruction method and device, storage medium and program product

Abstract

The invention discloses a graph structure reconstruction method and device, a storage medium and a program product, and the method comprises the steps: determining a hyperbolic embedding vector corresponding to each graph node in a to-be-processed initial graph structure, and analyzing target priori knowledge semantically matched with each hyperbolic embedding vector; for each hyperbolic embedding vector, graph level representation corresponding to the hyperbolic embedding vector is extracted, and corresponding target priori knowledge is fused to obtain graph node fusion features; and inputting each graph node fusion feature into a reinforcement learning model to determine a corresponding target graph reconstruction action, and performing deletion operation on edge connection between the graph nodes in the initial graph structure according to the target graph reconstruction action to obtain a corresponding target reconstruction graph structure. Therefore, the topological representation characteristics of the graph and the adaptive ability of reinforcement learning can be utilized to reconstruct a graph structure meeting actual task requirements.

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a graph structure reconstruction method, device, storage medium and program product. Background Art

[0002] Graph structures are widely used in a variety of data processing scenarios, such as social network analysis, traffic networks, material molecular biology, and communication networks. In these applications, graph structures are often used to represent the relationship or interaction between entities, where nodes represent entities and edges represent the relationship or connection between nodes, making graph structures an indispensable part of data analysis and machine learning.

[0003] For example, in a social network scenario, nodes represent users, and edges represent friendships between users, so that a large social network graph can be constructed. By analyzing the social network graph, users' interests and hobbies can be accurately predicted, and personalized content recommendations can be provided, thereby significantly improving user experience and user stickiness of the platform. In the field of transportation, nodes represent transportation hubs, and edges represent road connections, so a transportation network topology graph can be constructed. Using this graph structure, traffic flow distribution can be effectively optimized, travel routes can be reasonably planned, transportation efficiency can be improved, and traffic congestion can be alleviated.

[0004] However, the complexity and scale of graph structured data are increasing, and traditional graph reconstruction methods are unable to cope with these challenges. Most traditional methods are based on the geometric characteristics of Euclidean space, using graph adjacency matrices or distance matrices to construct the relationship between nodes and edges. This method can achieve good results when dealing with small-scale, regular graph data, but its performance is limited when faced with large-scale, sparse, and high-dimensional complex graph structures. Especially in graph data with hierarchical structures, the embedding method of Euclidean space cannot fully express the distance and relationship between graph nodes, which leads to problems such as loss of key information and low reconstruction accuracy during graph structure reconstruction.

[0005] The industry has not yet proposed a better solution to the above problems. Summary of the invention

[0006] The present application provides a graph structure reconstruction method, device, storage medium and program product, which are used to at least solve the problem that the traditional Euclidean space embedding method has great limitations in processing complex graph data.

[0007] In the first aspect, an embodiment of the present application provides a graph structure reconstruction method, including: determining the hyperbolic embedding vector corresponding to each graph node in the initial graph structure to be processed, and parsing the target prior knowledge that semantically matches each of the hyperbolic embedding vectors; for each of the hyperbolic embedding vectors, extracting the graph hierarchical representation corresponding to the hyperbolic embedding vector, and fusing the corresponding target prior knowledge to obtain a graph node fusion feature; inputting each of the graph node fusion features into a reinforcement learning model to determine a corresponding target graph reconstruction action, and performing deletion and modification operations on the edge connections between the graph nodes in the initial graph structure according to the target graph reconstruction action to obtain a corresponding target reconstructed graph structure.

[0008] In a second aspect, an embodiment of the present application provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the graph structure reconstruction method of any embodiment of the present application.

[0009] In a third aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the graph structure reconstruction method of any embodiment of the present application are implemented.

[0010] In a fourth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the graph structure reconstruction method of any embodiment of the present application.

[0011] The beneficial effects of the embodiments of the present application are: By adopting hyperbolic embedding vectors, graph nodes can be represented more accurately in non-Euclidean space, and the complex nonlinear relationships between nodes in hierarchical graph data can be better captured. In addition, by integrating target prior knowledge, the construction of graph node features is not only based on the characteristics of the nodes themselves, but also can be combined with external domain semantic information. Furthermore, by introducing a reinforcement learning model to automatically delete and modify the connections between graph nodes, based on the training and feedback mechanism of reinforcement learning, the graph reconstruction action corresponding to the optimal graph reconstruction strategy can be selected to improve the accuracy of graph reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 A flowchart of an example of a graph structure reconstruction method according to an embodiment of the present application is shown; Figure 2 An operation flow chart of an example of determining a hyperbolic embedding vector of each graph node according to an embodiment of the present application is shown; Figure 3 A schematic diagram showing a structural connection of an example of a reinforcement learning model according to an embodiment of the present application is shown; Figure 4 A schematic diagram showing an effect of an example of a chemical variable relationship diagram structure according to an embodiment of the present application; Figure 5 A flowchart of an example of a method for reconstructing a chemical variable relationship graph based on graph reinforcement learning provided in an embodiment of the present application is shown; Figure 6 A schematic structural diagram of an example of a graph attribute-knowledge tree according to an embodiment of the present application is shown; Figure 7 A schematic diagram showing the effect of an example of a corpus of graph structure features and prior knowledge according to an embodiment of the present application; Figure 8 A visualization diagram showing an example of a prior knowledge weight vector; Fig. 9 A schematic diagram of the architecture of an example of graph reinforcement learning according to an embodiment of the present application is shown; Fig.10 A schematic diagram of a framework of an example of chemical variable relationship diagram reconstruction according to an embodiment of the present application is shown; Fig.11 A schematic diagram showing a comparison of experimental simulation effects of an example of the identification result of the system variable 1 according to an embodiment of the present application is shown; Fig.12 A schematic diagram showing a comparison of experimental simulation effects of an example of the identification result of the system variable 2 according to an embodiment of the present application is shown; Fig.13 It is a schematic structural diagram of an embodiment of an electronic device of the present application. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0015] It should be noted that GNN (Graph Neural Networks) has made breakthrough progress in the field of graph data analysis. Graph neural networks achieve feature learning of graph nodes and edges by iteratively aggregating feature information of nodes and their neighboring nodes. However, existing graph neural networks usually learn based on fixed graph structures and lack flexibility. When the structure of the graph changes dynamically, these models have difficulty adapting to the new structure. RL (Reinforcement Learning) mechanism is introduced to enhance the learning ability of dynamic graphs. Reinforcement learning can continuously optimize and adjust the graph structure through interaction with the environment. Despite this, existing reinforcement learning methods usually only focus on the optimization of the topological structure of the graph, ignoring the semantic information represented by prior knowledge in the graph, so there are still certain limitations when dealing with complex graph structures.

[0016] Figure 1 A flowchart of an example of a graph structure reconstruction method according to an embodiment of the present application is shown.

[0017] like Figure 1 As shown, in step S110, the hyperbolic embedding vector corresponding to each graph node in the initial graph structure to be processed is determined, and the target prior knowledge that semantically matches each hyperbolic embedding vector is parsed.

[0018] In some embodiments, for the graph structure to be processed, each graph node will be mapped to a low-dimensional hyperbolic space, which is completed by a hyperbolic geometry embedding algorithm, mapping each graph node to a vector in the hyperbolic space, aiming to capture the complex nonlinear relationship between nodes. Compared with the traditional Euclidean space, the hyperbolic space can better express the hierarchy and topological structure between nodes. It should be understood that the type of hyperbolic space can be diverse, such as the Poincare sphere model, negative surface or Lorentz model, etc., which are not limited here.

[0019] In addition, according to the specific business application scenarios of the graph structure (such as social networks, transportation networks, etc.), target prior knowledge is introduced, such as external knowledge such as user behavior and traffic flow, and it is semantically matched with the embedding vector of each graph node. Through semantic matching, it can help enhance the expressive power of the graph node embedding vector and further enrich the semantic information of the node, so that the representation of the node not only considers the structure of the graph itself, but also combines external information, thereby improving the accuracy and practicality of the graph node representation.

[0020] In step S120, for each hyperbolic embedding vector, the graph level representation corresponding to the hyperbolic embedding vector is extracted, and the corresponding target prior knowledge is fused to obtain the graph node fusion feature.

[0021] In some embodiments, for each hyperbolic embedding vector of a graph node, its hierarchical structure information in the graph is extracted. For example, local and global hierarchical representations of graph nodes in the graph structure are extracted through a graph convolutional network (GCN), a graph attention network (GAT), or other graph-based deep learning methods. Thus, the hierarchical representation of the graph nodes is extracted, so that each node is not only limited to its own characteristics, but also takes into account its position in the overall structure of the graph, effectively improving the richness of node features.

[0022] Furthermore, the target prior knowledge is fused with the hierarchical features of the graph nodes, which can be accomplished through simple weighted fusion, splicing, or more complex attention mechanisms, so that the features of the graph nodes not only reflect the graph structure, but also express richer semantic information.

[0023] In step S130, the fusion features of each graph node are input into the reinforcement learning model to determine the corresponding target graph reconstruction action, and the edge connections between the graph nodes in the initial graph structure are deleted or modified according to the target graph reconstruction action to obtain the corresponding target reconstructed graph structure.

[0024] In some embodiments, the reinforcement learning model determines the graph reconstruction operation by evaluating the fusion features of each graph node to optimize the connectivity of the graph structure, making the relationship between the nodes in the graph clearer and enhancing the practical application effect of the graph structure expression. It should be understood that the architecture of the reinforcement learning model can be diverse, and a policy gradient-based algorithm such as PPO (Proximal Policy Optimization) or DQN (Deep Q-Network) can be used to learn how to delete and modify the edge connections between graph nodes.

[0025] It should be noted that the reinforcement learning model executes graph reconstruction action decisions based on the current graph structure state and generates target graph reconstruction actions. Reconstruction actions include deleting unimportant edges, adding missing edges, or adjusting edge weights. Through continuous iterative training, the model can learn the optimal graph reconstruction strategy, and then delete or modify the edge connections in the initial graph structure according to the actions generated by the reinforcement learning model, thereby obtaining the target reconstructed graph structure, which can better express the relationship between nodes and improve the accuracy of graph analysis.

[0026] Through the embodiments of the present application, the optimal graph reconstruction strategy is automatically selected according to the characteristics of the graph structure through the reinforcement learning model, making the reconstruction process of the graph structure more intelligent and dynamic. Compared with the traditional manually designed graph reconstruction method, the model can adapt to the characteristics of different graph data and provide a more accurate reconstruction effect. In addition, by optimizing edge connections through action decisions, redundant edges and irrelevant information can be reduced, and the accuracy and interpretability of the graph structure can be improved. This is especially important for large-scale, sparse and high-dimensional graph data, and can ensure that the graph structure used in graph analysis tasks (such as node classification, graph matching, etc.) is more efficient and accurate.

[0027] In the hybrid graph reconstruction method based on hyperbolic space and semantic graph reinforcement learning provided in the embodiment of the present application, the graph data is embedded in the hypersurface space, and the geometric characteristics of the hypersurface space are used to enhance the representation ability of complex graph structures. At the same time, the semantic information of the extracted graph is combined to construct a semantic graph, and the geometric graph and the semantic graph are jointly optimized through the graph reinforcement learning framework, and the geometric distance and semantic association between the graph nodes are comprehensively considered, which significantly improves the accuracy and efficiency of the graph reconstruction task.

[0028] Figure 2 An operational flowchart of an example of determining a hyperbolic embedding vector for each graph node according to an embodiment of the present application is shown.

[0029] like Figure 2 As shown, in step S210, an initial graph embedding space corresponding to the initial graph structure in the Euclidean space is determined.

[0030] In some implementations, it is assumed that the initial graph structure to be processed is ,in is a collection of nodes, is the edge set. Each graph node Each contains a node feature vector, whose features can be the basic attributes of the node, historical behavior or other semantic information. Furthermore, the existing graph embedding algorithms (such as DeepWalk, Node2Vec, etc.) can be used to map the graph structure to the Euclidean space. Each graph node is embedded in a low-dimensional vector space in the Euclidean space, such as 128 or 256 dimensions, to obtain the initial graph embedding space representation, so that the relationship between nodes is mainly measured based on the Euclidean distance.

[0031] In step S220, the node features of each graph node in the initial graph embedding space are mapped to corresponding points in the Poincare space based on the Poincare sphere model, and a corresponding first hyperbolic embedding vector is obtained through a learnable linear transformation layer.

[0032] In some embodiments, the Poincare sphere model is a non-Euclidean space model that can effectively represent nonlinear relationships in high-dimensional, sparse, and complex graph structures, in which the embedding space of node features is transformed into a hyperbolic geometric space that can capture complex dependencies between graph nodes.

[0033] Specifically, by using the mapping function of the Poincare sphere model, the Euclidean embedding vector of each graph node is mapped to the corresponding point in the Poincare space, ensuring that the structural relationship between nodes is better expressed in the hyperbolic space. In particular, for graph data with a hierarchical structure, the Poincare space can effectively retain this hierarchical information. Then, through a learnable linear transformation layer, the mapping of each graph node is adjusted to produce a more reasonable hyperbolic embedding vector.

[0034] As a further preferred implementation, in order to implement the node aggregation update operation on the graph structure.

[0035] In step S230, each first hyperbolic embedding vector is remapped back to the Euclidean space, and graph nodes are aggregated through a graph attention algorithm to obtain updated graph nodes.

[0036] It should be noted that directly performing graph node aggregation operations in the Poincare hyperbolic space will introduce additional computational overhead, and most graph operations (such as convolution) are defined in the Euclidean space. Therefore, before information aggregation, the hyperbolic embedding vector of each graph node will be remapped back to the Euclidean space, which ensures that subsequent aggregation operations can take advantage of the computational advantages of the Euclidean space and reduce computational overhead.

[0037] In Euclidean space, node features are aggregated through the graph attention algorithm. The graph attention mechanism is used to assign dynamic and learnable attention weights to each edge. The updated feature of each node is the weighted sum of the features of its neighboring nodes, so that each node can adaptively select meaningful information from its neighboring nodes for aggregation, thereby improving the quality and efficiency of node feature updates.

[0038] In step S240, each updated graph node is remapped back to the Poincare hyperbolic space and nonlinearly transformed by an activation function to obtain a corresponding second hyperbolic embedding vector.

[0039] In some embodiments, the new feature vector of each graph node is converted into a new vector in the Poincare space using a mapping function of hyperbolic geometry, and the geometric characteristics of the graph structure are fully preserved. In order to further enhance the expressiveness of node features, a nonlinear activation function (such as ReLU, tanh, etc.) is used to transform the feature vector remapped to the Poincare space, introducing more nonlinear factors into the node features, making the representation of node features richer, thereby being able to capture more complex relationships and patterns.

[0040] Through the embodiments of the present application, by first mapping the node features to the Euclidean space for information aggregation, and then mapping them back to the Poincare hyperbolic space and performing nonlinear transformation, it is possible to effectively combine the computational advantages of the Euclidean space and the superior representation capabilities of the Poincare hyperbolic space for the complexity of the graph structure. In this way, the efficiency of graph node aggregation is ensured, and the expression capabilities of node features are improved, and the nonlinear and hierarchical relationships between nodes in the graph can be more accurately captured, thereby improving the overall graph embedding and task performance.

[0041] Regarding the specific operational details of parsing the hyperbolic embedding vector for semantically matching target prior knowledge, in some examples of the embodiments of the present application, for each hyperbolic embedding vector, the semantic similarity between the hyperbolic embedding vector and each prior knowledge weight word vector in the graph corpus is calculated to screen the matching target prior knowledge.

[0042] Here, the prior knowledge weight word vector is a word vector containing learning weights converted from the graph attribute-knowledge tree, and the graph attribute-knowledge tree is constructed based on the fusion of the task attribute tree and the graph structure training samples.

[0043] Regarding the construction details of the graph attribute-knowledge tree, it can be to extract high-order attributes and hierarchical information in the task, and represent them as nodes or branches in the tree, so as to obtain a task attribute tree. Afterwards, in the context of the graph structure training samples, high-order attributes related to the task are extracted. These attributes may include the hierarchical structure between nodes, the functional relationship between nodes, the similarity between nodes, or other meaningful task-related information, and may also involve the upstream and downstream relationships of nodes, the clustering information of nodes, etc. Then, the extracted high-order attributes are integrated with the hierarchical information of the nodes and represented as a tree structure in the graph. The root node in the tree structure represents the central node of the high-order attributes in the graph, and each original graph node is regarded as a leaf node of the tree. The branch part of the tree represents the affiliation of the node in the graph hierarchy.

[0044] The graph attribute-knowledge tree constructs a tree structure containing multi-level information by integrating the structural information of the graph and the high-order attributes of the task (or downstream task). The secondary root node is divided into a structural layer and a node layer according to the position and level of the node in the graph. The structural layer represents the order of the node relative to the central node, that is, the level or distance of the node in the graph structure. The node layer contains the original nodes representing the actual variables in the graph, which are directly related to the original data in the graph. In this way, the original graph nodes of the graph structure training samples are divided into different categories in the graph hierarchy according to their high-order features, and new nodes and connections are automatically generated to further supplement the hierarchical information in the original graph. Therefore, the graph attribute-knowledge tree not only retains the structural information of the original graph, but also embeds the task attributes in a hierarchical manner, so that the contextual semantic information of the task can be effectively integrated into the graph, improving the expressive power of the graph structure.

[0045] In addition, through the graph attribute-knowledge tree, the task attributes in the graph are mapped to word vectors with learning weights. These weighted word vectors reflect the hierarchical relationship between nodes and the importance of each node in the task. Each prior knowledge word vector is optimized through learning to ensure that they are closely related to the structure of the graph and the task attributes. By calculating the semantic similarity between the hyperbolic embedding vector and the prior knowledge weighted word vector, it is determined which prior knowledge matches the features of the current graph node, such as by calculating the cosine similarity or Manhattan distance to screen out prior knowledge with higher similarity, which can effectively enhance the representation ability of graph nodes, not only providing additional semantic support for graph nodes, but also helping the graph model to better understand the role and significance of each graph node in the graph structure.

[0046] Through the embodiments of the present application, the similarity between the hyperbolic embedding vector and the prior knowledge weighted word vector is calculated, and the high-order attributes of the node can be accurately matched semantically. Through semantic matching, each graph node can not only utilize its own structural information, but also combine external prior knowledge to further enhance the intelligence and task relevance of the graph data. Therefore, by effectively screening and integrating prior knowledge, the node features can be dynamically adjusted according to the needs of the task, so that the model can automatically introduce the most relevant contextual information when processing graph data, thereby improving the accuracy and efficiency of task execution.

[0047] The graph attribute-knowledge tree constructed by combining the graph structure and the task attribute tree can process and represent the high-level information of the graph nodes more finely. By introducing prior knowledge that matches the semantics of node features, it not only provides rich contextual support for graph nodes, but also improves the accuracy of node feature representation and the relevance of tasks, enhances the intelligence and flexibility of graph structure analysis, and helps to improve the in-depth analysis and reasoning accuracy of complex graph data.

[0048] Figure 3A schematic diagram of the structural connection of an example of a reinforcement learning model according to an embodiment of the present application is shown.

[0049] like Figure 3 As shown, the reinforcement learning model adopts a graph reinforcement learning agent, and the reinforcement learning model 300 includes a GAT layer 310 and a fully connected layer 320.

[0050] Specifically, feature extraction is performed on the fusion features of each graph node based on the GAT layer 310, and a heuristic graph reconstruction action group is output through the fully connected layer 320.

[0051] It should be understood that the number of GAT layers 310 can be one or more, and the features of each graph node and its adjacent nodes are weightedly aggregated through the self-attention mechanism. The fused feature vector of the graph node is passed as input to the GAT layer for further processing. The GAT layer uses the attention mechanism to aggregate features based on the relationship between each node and its neighbors. The local structure and neighbor information of the node in the graph will be effectively integrated, enhancing the contextual expression ability of the node features.

[0052] In the fully connected layer 320, the node features are further processed by a multi-layer perceptron or a linear transformation to map the features of the graph nodes to a higher-dimensional space. The heuristic strategy obtained through learning is used to generate a heuristic graph reconstruction action group. These actions include deletion and modification operations of the edges between nodes to guide the optimization of the graph structure under the action of different action strategies.

[0053] The deletion and modification operations of the edge connections corresponding to each heuristic graph reconstruction action in the heuristic graph reconstruction action group are executed in sequence to obtain the corresponding candidate reconstructed graph structures respectively.

[0054] For example, based on the graph reconstruction action, when the connection between nodes is considered no longer meaningful, the edge is deleted. In addition, when the model finds that the connection between certain nodes is underestimated or omitted, a new edge connection is added. After each graph reconstruction action is performed, a new candidate graph structure is generated, which represents a possible graph reconstruction method, aiming to optimize the connectivity and expressiveness of the graph through deletion and modification operations.

[0055] Each candidate reconstructed graph structure is evaluated according to the downstream task performance index corresponding to each candidate reconstructed graph structure to determine the corresponding target reconstructed graph structure.

[0056] In some embodiments, the performance indicators of downstream tasks, such as system identification accuracy or fault diagnosis accuracy, can be used as rewards for graph reconstruction, thereby evaluating the optimization effect of candidate reconstructed graph structures to select the best reconstructed graph structure. By comparing the task application performance of candidate graph structures, the optimization process of reconstructed graphs has clear goals, ensuring that the graph structure is not only optimized in form, but also can provide the best support for specific tasks. Thus, through the task-oriented graph structure optimization method, the accuracy and efficiency of graph analysis tasks are improved, supporting applications in complex tasks that require adaptive adjustment of graph structures.

[0057] Regarding the training details of the reinforcement learning model, the loss function of the GAT layer 310 adopts an enhanced design based on mutual information. Specifically, the loss function of the GAT layer 310 is determined by maximizing the sum of the mutual information between the graph feature representation of each candidate reconstructed graph structure and the graph feature representation of the global reconstructed graph structure. Here, the global reconstructed graph structure is constructed by fusing the graph feature representations of all candidate reconstructed graph structures, for example, the feature representations of all candidate graph structures are merged (such as by weighted averaging, splicing, etc.) to obtain a comprehensive global graph feature representation. The global graph structure represents the overall information of all candidate graph structures and provides a global perspective for graph reconstruction.

[0058] It should be noted that mutual information measures the degree of information sharing between the feature representations of two graph structures. Maximizing mutual information helps to improve the similarity between the candidate graph and the global graph, thereby ensuring that each candidate structure can be adjusted towards the goal of global optimization during the graph reconstruction process. Therefore, each candidate graph not only improves the quality of the local structure during the optimization process, but also keeps consistent with the global structure, avoiding overfitting of local data and improving the accuracy of the reconstruction action.

[0059] It should be noted that in the chemical process industry, chemical processes often involve multiple variables, including controlled variables, manipulated variables, and other measurable variables involved in the process. These variables each represent different physical states, such as temperature, pressure, flow rate, etc., and there are complex coupling relationships between them. The relationship between these chemical variables cannot be accurately described by mechanism relationships or manual experience alone. This makes it difficult for relevant industry personnel to understand and optimize the corresponding chemical processes, and also poses challenges to the use of graph learning for system identification and fault diagnosis of chemical processes. Unfortunately, the construction of chemical variable graphs (e.g., chemical knowledge graphs) in the chemical industry is often highly dependent on expert experience, lacks generalization, and is prone to subjective omissions.

[0060] In some business application scenarios combined with the embodiments of the present application, the initial graph structure is a chemical variable relationship graph structure, whose graph nodes are defined by controlled variables, manipulated variables or measurable variables, and the edge connections in the chemical variable relationship graph structure are defined by the mechanism relationship between the variables.

[0061] Figure 4 A schematic diagram showing the effect of an example of a chemical variable relationship graph structure according to an embodiment of the present application is shown.

[0062] like Figure 4 As shown in the figure, the chemical process is represented from a graph perspective and decomposed into nodes and edges. Specifically, by analyzing the mechanism model of the chemical process, the controlled variables (such as inlet condensate temperature and condenser vacuum), manipulated variables (such as pump frequency and fan frequency) and measurable variables (such as condenser inlet and outlet pressures) involved are represented as nodes in the chemical variable relationship graph, and the connections between these nodes can be initialized according to their original mechanism relationships.

[0063] Through the chemical variable relationship graph reconstruction method based on hyperbolic space and semantic graph reinforcement learning provided in the embodiment of the present application, the graph data is embedded into the hypersurface space, and the geometric characteristics of the hypersurface space are used to enhance the representation ability of complex graph structures. At the same time, the word2vec algorithm is combined to extract the semantic information of the graph, and the chemical variable relationship graph is automatically jointly optimized through the graph reinforcement learning framework, and it is evaluated according to the performance of downstream tasks.

[0064] Specifically, through the chemical variable relationship graph reconstruction method based on graph reinforcement learning, the problem of variable relationship discovery in the field of chemical process is effectively solved to find out the complex coupling relationship between chemical variables, provide knowledge reference, and improve the performance of downstream tasks. The exemplary operations are as follows: chemical process representation based on graph; node representation graph embedding based on hyperbolic space; constructing graph attribute-knowledge tree to enrich the representation of the original graph; using word2vec algorithm for semantic embedding to retain important feature information in the graph and integrate prior knowledge; integrating into semantic graph reinforcement learning framework for graph reconstruction. As a result, the topological representation characteristics of the graph and the adaptive ability of reinforcement learning can be used to reconstruct the graph structure that conforms to the actual chemical process coupling relationship, provide reference for industry personnel to understand the corresponding process, and improve the performance of graph tasks.

[0065] Figure 5 A flowchart of an example of a method for reconstructing a chemical variable relationship graph based on graph reinforcement learning provided in an embodiment of the present application is shown.

[0066] In an embodiment of the present application, a method for reconstructing a chemical variable relationship graph based on hyperbolic space and semantic graph reinforcement learning is provided, by embedding the topological features of the graph into the hypersurface space, and combining the semantic information of the graph, the topological graph and the semantic graph are jointly optimized using a graph reinforcement learning framework. This method can provide higher reconstruction accuracy and downstream task efficiency in the graph reconstruction task of large-scale complex graphs.

[0067] like Figure 5 The chemical variable relationship graph reconstruction method based on graph reinforcement learning shown in the figure aims to solve the problem of variable relationship discovery in the field of chemical processes, so as to find out the complex coupling relationship between chemical variables, provide knowledge reference, and improve the performance of downstream tasks.

[0068] It should be noted that the graph reinforcement learning framework constructed based on the embodiments of the present application is intended to provide solutions for graph construction tasks in different application scenarios. In order to prove the effectiveness of the present application, a system identification experiment was carried out based on real cooling tower process data and used as an example. It should be noted that this part of the embodiment is combined with the description of the cooling tower process application scenario, which is only used as an example and should not be regarded as a limitation on the scope of implementation of the embodiments of the present application, and downstream system identification tasks or other chemical processes, as well as downstream graph task applications such as fault diagnosis and fault tracing, all fall within the scope of implementation of the embodiments of the present application.

[0069] Aiming at the problem of complex coupling relationship of chemical variables in the chemical process industry, a chemical variable relationship graph reconstruction method based on hyperbolic space and semantic graph reinforcement learning is proposed. The specific process framework can be described as follows: In step S1, the chemical process is represented from a graph perspective and decomposed into nodes and edges. The chemical process is abstracted in the form of a graph, where nodes represent various variables in the chemical process, such as temperature, pressure, etc., and edges represent the potential relationship between these variables. The characteristics of each node are constructed by collecting relevant physical properties, historical data, and field observations to form a feature vector. The initial node connection relationship can be based on expert experience or random initialization to form a basic chemical variable relationship graph.

[0070] Specifically, step S1 includes the following steps: S11: Analyze the chemical process from a graphical perspective and characterize it as , list the relevant variables involved, including controlled variables, manipulated variables and other measurable variables, and use them as nodes in the chemical variable relationship diagram represented .

[0071] S12: The physical properties, historical data, and field observations of each variable are used as node features and represented as feature vectors .

[0072] S13: Initialize connections between nodes based on manual experience or randomly , as the adjacency matrix of the chemical variable relationship graph , forming the original graph of chemical variable relationship .

[0073] In step S2, the constructed original graph data is embedded into the hyperbolic space to generate a hyperbolic embedding vector for each node.

[0074] Specifically, a hyperbolic space graph embedding algorithm is used, which uses a hyperbolic geometry model to embed the original Euclidean graph into a hyperbolic space, and uses the characteristics of hyperbolic geometry to enhance the ability to express complex graph structures. Through specific mathematical transformations, such as the Poincare sphere model, nodes are converted from Euclidean space to hyperbolic space to generate hyperbolic embedding vectors. Further, through linear transformations, the hyperbolic embedding vectors are adjusted to ensure that they have a suitable geometric structure and better reflect the hierarchical relationship between nodes. The following steps are included: S21: Construct the initial graph embedding space in Euclidean space and initialize the node vector in Euclidean space , Indicates that the vector is used to represent a node; S22: Select the Poincare sphere as the hyperbolic space model and use the Poincare embedding exponential mapping formula to transform each node Mapping to corresponding points in Poincare space . Initial vector is converted to a hyperbolic vector : Formula (1) in It is the Möbius addition. are parameters that can be set manually. represents the origin in hyperbolic space, Characterizes the initial vector of the i-th node v, The characterization vector representing the i-th node v in the hyperbolic space H.

[0075] S23: Through a learnable linear transformation layer, the mapped hyperbolic space embedding is ensured to have the appropriate hyperspherical geometry.

[0076] Formula (2) in, is the Möbius multiplication, represents the transformation matrix, Indicates deviation, Represents the transformed representation vector of the i-th node v in the hyperbolic space H.

[0077] As a further preferred embodiment, since the mathematical steps of aggregating neighboring node features in a graph network are inherently complex, and operations such as convolution are defined in Euclidean space, performing these operations directly in Poincare hyperbolic space will incur additional computational overhead. Therefore, before information aggregation, the node feature vectors are remapped back to Euclidean space.

[0078] Formula (3) in It is the Möbius addition. are parameters that can be set manually. represents the origin in hyperbolic space, represents the mapping reference point in the hyperbolic space, The representation vector representing the i-th node v after being mapped back to the Euclidean space τ.

[0079] After aggregation is completed according to the steps of the GAT algorithm, the new features of each node are remapped back to the Poincare hyperbolic space and transformed nonlinearly through the activation function.

[0080] Formula (4) in, represents the transformation matrix, represents a nonlinear activation function, represents the attention coefficient between node i and node j in GAT, represents node i, represents the neighbor node j of node i, represents the set of all neighbor nodes of node i, The representation vector representing the i-th node in the l-th GAT layer in the Euclidean space τ, The representation vector representing the i-th node in the hyperbolic space H in the l+1th GAT layer.

[0081] In step S3, a graph attribute-knowledge tree is constructed, combining the tree structure with the graph structure to retain the high-order information of the graph.

[0082] Here, the graph attribute-knowledge tree is used to represent the high-order information and prior knowledge in the task to supplement the content of the graph representation. In order to enrich the representation of the graph, a graph attribute-knowledge tree is constructed to integrate the high-order information and prior knowledge of the graph. The tree structure classifies the nodes in the graph hierarchically to form a new hierarchical structure, which contains the high-order features and prior knowledge of the nodes. In this way, not only the structural information of the original graph is retained, but also additional semantic information is introduced to enhance the expressive power of the graph. Specifically, it includes the following steps: S31: Extract high-order attributes and hierarchical information in the task and represent them as nodes or branches in the tree; use the graph hierarchical information of the node as the new root node, and the original graph node as the new leaf node. The new branch represents the affiliation of the node at the graph level.

[0083] S32: Constructing graph attribute-knowledge tree, combining tree structure with graph structure to preserve high-order information of graph; secondary root nodes are divided into two levels: structure layer and node layer. The structure layer includes the node order relative to the central node, and the bottom node layer contains the original nodes representing variables in the graph. In this way, the original nodes are divided into different categories according to the high-order features of the graph level, and new nodes and connections are generated in the process to supplement the graph level information of the original graph.

[0084] S33: Due to the excellent scalability of the tree structure, new attribute nodes containing prior knowledge such as "important" and "unimportant" are additionally generated at the structure layer or node layer as a new knowledge layer to flexibly integrate prior knowledge into the newly generated tree to enrich the content of the tree.

[0085] Specifically, a method is adopted to construct a graph attribute-knowledge tree from the perspective of high-order graph features such as the original graph structure and node attributes. Figure 6 A structural diagram of an example of a graph attribute-knowledge tree according to an embodiment of the present application is shown.

[0086] like Figure 6 As shown in the figure, according to the properties and structure of the graph, a knowledge tree is constructed, combining the nodes and edges in the graph with the tree structure to retain the high-order information in the graph. The original nodes are divided into different categories according to the high-order features at the graph level, and new nodes and connections are generated in the process to supplement the graph-level information of the original graph. In addition, due to the excellent scalability of the tree structure, this method additionally generates new attribute nodes containing prior knowledge such as "important" and "unimportant" at the graph or node level as a new knowledge layer to flexibly integrate prior knowledge into the newly generated tree to enrich the content of the tree.

[0087] In step S4, a corpus containing graph structure features and prior knowledge is constructed, and semantic embedding is performed using the word2vec algorithm to integrate prior knowledge.

[0088] Here, the semantic graph construction uses the word2vec algorithm to extract the semantic information of local neighborhood nodes in the graph, convert the graph attribute-knowledge tree into text form, and build a corpus. By training this corpus, vectors that can reflect the semantic information of nodes are generated. These vectors can capture the semantic associations between nodes. The similarity between the node vector and the prior knowledge vector is calculated to generate a weight vector for the subsequent graph reinforcement learning process. Specifically, it includes the following steps: S41: The constructed graph attribute-knowledge tree is converted into textual knowledge; by constructing a graph corpus, which includes words reflecting the structural characteristics of the graph and prior knowledge, including the nodes of the original graph, connection information, hierarchical information in the tree, the actual physical meaning of nodes, edges and node attributes, and learning weights representing prior knowledge.

[0089] S42: Based on the converted textual knowledge, it is digitized into vectors, and the corpus is trained using the word2vec algorithm so that the distribution of word vectors in the semantic space reflects prior knowledge.

[0090] S43: Calculate the word vector for each node Word vectors with prior knowledge The cosine similarity between them is used to further extract the learning focus. It is converted into a weight vector, integrated into the graph level features, and participates in the subsequent graph reinforcement learning training.

[0091] Formula (5) In the formula, The word vector representing a single node, word vectors representing prior knowledge, Represents cosine similarity.

[0092] Here, the text information in the corpus is vectorized, and these terms are mapped to the semantic space in the form of word vectors, forming Figure 7 The vocabulary shown is used to distinguish them. Figure 7 The effect diagram of an example of a corpus of graph structure features and prior knowledge according to an embodiment of the present application is shown. In the process of semantic embedding, a corpus containing graph structure features and prior knowledge and a corresponding vocabulary are constructed. Subsequently, the agent is required to understand the meaning of words such as "node", "level" and "importance", which requires these terms to be distributed in the semantic space according to their original semantics.

[0093] In this method, the training of graph corpus is combined with word2vec algorithm for semantic embedding, using Skipgram model, which is trained by predicting surrounding context words given a central word. During the training process, parameterizable back propagation (BP) neural network is used to learn word vectors.

[0094] Assume a constructed vocabulary where each word has a unique index. For each center word, the goal is to maximize the conditional probability of the context words. Specifically, for a given center word and its context words, the training goal of the corpus is to maximize. The specific method is to do a dot product of the center word vector and the context word vector, and then pass the result to the SoftMax function to obtain the conditional probability. The function is defined as follows: Formula (6) in, It is the central word. is the center word vector, and is the context word vector, and t represents the index interval from the center word.

[0095] Formula (7) in, is the word vector of the context word, t Indicates the upper bound of the index, is the average word vector of the context words.

[0096] However, the computational cost of the SoftMax function increases with the size of the vocabulary, so in actual training, negative sampling is introduced to speed up the training process. The goal is to maximize the probability of the correct context words while minimizing the probability of randomly sampled incorrect context words. Specifically, for each center word and its correct context words, their inner product is maximized, and the inner product between the center word and the incorrect context word is minimized. The objective function of negative sampling is defined as follows: Formula (8) in, Is a central word The word vector of Indicates the central word The word vector of the context word, Indicates the central word The word vector of the wrong context word, Indicates the central word The error context lexicon, represents a non-linear activation function.

[0097] For the entire graph corpus, the loss function can be defined as: Formula (9) in, is the set of all the central words in the corpus, Represents an adjustable coefficient between (0,1).

[0098] After completing corpus learning, the arrangement of the original word vectors in the new embedding space not only reflects the original structure of the graph, but also embodies the guiding role of prior knowledge.

[0099] The word embedding training process is independent of the subsequent graph reinforcement learning training process. The feature vectors learned in this step reflect the high-order information and prior knowledge of the graph and do not participate in the subsequent information aggregation in the graph network, thus avoiding the over-smoothing phenomenon that may occur due to subsequent graph reinforcement learning. Figure 8 A visualization diagram of an example of a priori knowledge weight vector is shown. Figure 8 As shown, the similarities reflecting the importance of each node will form a prior knowledge weight matrix, which is added to the subsequent transformation from node representation to graph representation.

[0100] In step S5, the original graph structure is iteratively optimized according to the reward function of graph reinforcement learning, and the connection relationship between nodes is gradually adjusted to obtain a reconstructed graph for further application.

[0101] Here, the topological and semantic properties of the graph structure are comprehensively considered, so that the reconstructed graph can satisfy the semantic association while retaining the topological characteristics. The reinforcement learning policy network adopts the DDQN (Double Deep Q-Network) algorithm to guide the graph reconstruction by maximizing the reward function of the downstream task.

[0102] The specific steps include: S51: Use the improved GAT and DDQN algorithms as the basic algorithms for graph reinforcement learning; the input of graph reinforcement learning is the original features of each node and the adjacency matrix These features are then converted into hyperbolic vectors through hyperbolic embedding in step S2 and then passed through the GAT layer for feature extraction.

[0103] S52: Representing the Node Hierarchy of a Graph Network Expand to layer-level representation by flattening layers ; Add prior knowledge weights to the layer-level representation , integrate semantic knowledge to make node representation richer; Formula (10) in, is a transformation matrix, represents the flattening operation, represents the concatenation operation of vectors, represents the node-level representation vector of the i-th node, Represents the prior knowledge weight vector of the i-th node.

[0104] S53: The extracted graph-level representation is converted into an action probability vector through the FC layer, which represents the probability of the agent executing the action in the designed action rule library. The graph reconstruction action performed by the agent is reflected in the variable graph as the deletion and modification of the edges between nodes. After an action cycle of the agent, the reconstructed graph is obtained and evaluated based on the performance of the downstream task to train the agent.

[0105] S54: After the training of the intelligent agent is completed, it can be generalized to other chemical processes, autonomously reconstruct the chemical variable relationship diagram, and find out the complex coupling relationship between variables.

[0106] Through the graph reinforcement learning framework, the original graph structure is optimized using the results of hyperbolic embedding and semantic embedding. The improved GAT and DDQN algorithms are used to intelligently adjust the connection relationship between nodes according to the preset reward function. Through iterative optimization, a graph structure that is more in line with the actual chemical process is gradually reconstructed to improve the understanding and prediction capabilities of the chemical process.

[0107] Fig. 9 A schematic diagram of the architecture of an example of graph reinforcement learning according to an embodiment of the present application is shown.

[0108] like Fig. 9 As shown in the figure, the main body of the graph reinforcement learning agent is composed of GAT. Based on the improved graph attention network and the two-layer deep Q-network algorithm, a graph reinforcement learning framework is constructed to perform the graph reconstruction task. The last layer of graph reinforcement learning is the fully connected layer (FC layer), and the extracted graph features are converted into action probability vectors through this layer. The action vector corresponds to a set of heuristic graph reconstruction actions, which are reflected in the modification of the edges of the graph. After the action is completed, a new graph is generated, which is a step in graph reconstruction. After the new graph is generated, it is evaluated using the reward function R, the loss function L is calculated, and the policy network in graph reinforcement learning is updated. The loss of the GAT layer is designed to maximize the sum of the mutual information between all local and global graph features in the global graph: Formula (11) in Represents the image feature representation of the i-th single image, Represents the graph feature representation of the jth step of the entire reconstruction process, represents mutual information calculation.

[0109] Fig.10 A schematic framework diagram of an example of chemical variable relationship diagram reconstruction according to an embodiment of the present application is shown.

[0110] like Fig.10As shown in Figure 1, the graph reconstruction framework consists of two main parts: one is the GAT-FC network as a graph reinforcement learning agent, which outputs actions to reconstruct the variable graph; the other is a graph-recurrent neural network for downstream tasks, whose task performance is used as a reinforcement learning reward. When the downstream tasks of graph reconstruction are different, the reward R of graph reinforcement learning will also be adjusted accordingly. The results of downstream tasks, such as system identification accuracy or fault diagnosis accuracy, serve as the reward R for graph reconstruction.

[0111] In the experimental process combined with the embodiments of the present application, the evaluation index of whether the graph reconstruction is reasonable is the accuracy of using this graph for system identification. The higher the accuracy of identification, the more the reconstructed graph can reflect the complex coupling relationship between process variables. Therefore, before training the graph reinforcement learning agent, the network used for the downstream task is first pre-trained to eliminate the influence of the untrained downstream task network on the graph reinforcement learning process, so that the performance difference in the graph reinforcement learning training process is completely determined by the quality of the graph reconstruction.

[0112] Fig.11 A schematic diagram showing a comparison of experimental simulation effects of an example of the identification result of the system variable 1 according to an embodiment of the present application is shown. Fig.12 A schematic diagram showing a comparison of experimental simulation effects of an example of the identification result of the system variable 2 according to an embodiment of the present application is shown.

[0113] Specifically, Fig.11 The results of downstream system identification tasks using the initial chemical variable relationship graph and the chemical variable relationship graph after graph reconstruction are shown. As can be seen from the figure, the performance of system identification using the reconstructed graph is significantly better than the result using the initial graph. The reconstructed graph is further applied to the system identification of another target control variable (vacuum degree), and the identification results are compared with the results using the initial graph. The comparison results are shown in Fig.12 As shown in the figure, the experimental results show that the result of graph reconstruction has a certain universality in this process. It is not only the best solution driven entirely by the data of a specific downstream task, but also reflects the connection between chemical variables in the actual chemical process to a certain extent.

[0114] In the embodiment of the present application, by integrating graph representation learning, graph reinforcement learning, hyperbolic space embedding and semantic embedding technology, a hybrid graph reconstruction method based on hyperbolic space and semantic graph reinforcement learning is implemented, providing a universal graph reinforcement learning framework to solve the graph reconstruction problem of complex problems, especially suitable for processing large-scale, sparse and high-dimensional data graph structure optimization and reconstruction. Thus, the topological representation characteristics of the graph and the adaptive ability of reinforcement learning can be used to reconstruct a graph structure that conforms to the actual chemical process coupling relationship, provide a reference for industry personnel to understand the corresponding process, and improve the performance of graph tasks.

[0115] It should be noted that this study used publicly available human subject data for retrospective analysis. As the data used were open access and the accompanying license confirmed that no ethical approval was required, no ethical approval was required for this study.

[0116] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of actions combined, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0117] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute any of the above-mentioned graph structure reconstruction methods of the present application.

[0118] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes any one of the above-mentioned graph structure reconstruction methods.

[0119] In some embodiments, an embodiment of the present application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a graph structure reconstruction method.

[0120] Fig.13 is a schematic diagram of the hardware structure of an electronic device for executing a graph structure reconstruction method provided by another embodiment of the present application, such as Fig.13 As shown, the device includes: One or more processors 1310 and memory 1320, Fig.13 A processor 1310 is taken as an example.

[0121] The device for executing the graph structure reconstruction method may further include: an input device 1330 and an output device 1340 .

[0122] The processor 1310, the memory 1320, the input device 1330 and the output device 1340 may be connected via a bus or other means. Fig.13 The example of connecting through bus is taken in the following.

[0123] The memory 1320, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the graph structure reconstruction method in the embodiment of the present application. The processor 1310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 1320, that is, the graph structure reconstruction method of the above method embodiment is implemented.

[0124] The memory 1320 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 1320 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 1320 may optionally include a memory remotely arranged relative to the processor 1310, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0125] The input device 1330 may receive input digital or character information and generate signals related to user settings and function control of the electronic device. The output device 1340 may include a display device such as a display screen.

[0126] The one or more modules are stored in the memory 1320, and when executed by the one or more processors 1310, the graph structure reconstruction method in any of the above method embodiments is executed.

[0127] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.

[0128] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to: (1) Mobile communication equipment: This type of equipment is characterized by having mobile communication functions and its main purpose is to provide voice and data communications. This type of terminal includes: smart phones, multimedia phones, functional phones, and low-end phones.

[0129] (2) Ultra-mobile personal computer devices: These devices belong to the category of personal computers, have computing and processing functions, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, etc.

[0130] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0131] (4) Other onboard electronic devices with data interaction functions, such as on-board devices installed in vehicles.

[0132] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0133] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A graph structure reconstruction method, comprising: Determine the hyperbolic embedding vector corresponding to each graph node in the initial graph structure to be processed, and parse the target prior knowledge that semantically matches each of the hyperbolic embedding vectors; For each of the hyperbolic embedding vectors, extract the graph level representation corresponding to the hyperbolic embedding vector, and fuse the corresponding target prior knowledge to obtain the graph node fusion feature; The fusion features of each of the graph nodes are input into a reinforcement learning model to determine a corresponding target graph reconstruction action, and the edge connections between the graph nodes in the initial graph structure are deleted or modified according to the target graph reconstruction action to obtain a corresponding target reconstructed graph structure.

2. The method according to claim 1, wherein: The step of determining the hyperbolic embedding vector corresponding to each graph node in the initial graph structure to be processed includes: Determine the initial graph embedding space corresponding to the initial graph structure in the Euclidean space; Based on the Poincare sphere model, the node features of each graph node in the initial graph embedding space are mapped to corresponding points in the Poincare space, and a corresponding first hyperbolic embedding vector is obtained through a learnable linear transformation layer.

3. The method according to claim 2, after mapping the node features of each graph node in the initial graph embedding space to the corresponding point in the Poincare space based on the Poincare sphere model and obtaining the corresponding first hyperbolic embedding vector through a learnable linear transformation layer, the method further comprises: Re-mapping each of the first hyperbolic embedding vectors back to the Euclidean space, and performing graph node aggregation through a graph attention algorithm to obtain updated graph nodes; Each updated graph node is remapped back to the Poincare hyperbolic space and nonlinearly transformed through an activation function to obtain the corresponding second hyperbolic embedding vector.

4. The method according to claim 1, wherein: The parsing of the target prior knowledge that each of the hyperbolic embedding vectors is semantically matched includes: For each of the hyperbolic embedding vectors, the semantic similarity between the hyperbolic embedding vector and each prior knowledge weight word vector in the graph corpus is calculated to screen matching target prior knowledge; the prior knowledge weight word vector is a word vector containing learning weights converted from a graph attribute-knowledge tree, and the graph attribute-knowledge tree is constructed based on the fusion of the task attribute tree and the graph structure training samples.

5. The method according to claim 1, wherein: The reinforcement learning model includes a GAT layer and a fully connected layer. The step of inputting the fusion features of each of the graph nodes into a reinforcement learning model to determine a corresponding target graph reconstruction action, and performing deletion and modification operations on the edge connections between the graph nodes in the initial graph structure according to the target graph reconstruction action to obtain a corresponding target reconstructed graph structure includes: Based on the GAT layer, feature extraction is performed on the fusion features of each graph node, and a heuristic graph reconstruction action group is output through the fully connected layer; Sequentially executing the deletion and modification operations of the edge connections corresponding to each of the heuristic graph reconstruction actions in the heuristic graph reconstruction action group to obtain corresponding candidate reconstructed graph structures respectively; Each of the candidate reconstructed graph structures is evaluated according to the downstream task performance index corresponding to each of the candidate reconstructed graph structures to determine a corresponding target reconstructed graph structure.

6. The method according to claim 5, wherein: Determine the loss function of the GAT layer by maximizing the sum of mutual information between the graph feature representations of each candidate reconstructed graph structure and the graph feature representation of the global reconstructed graph structure; The global reconstructed graph structure is constructed by fusing the graph feature representations of all candidate reconstructed graph structures.

7. The method according to any one of claims 1 to 6, wherein: The initial graph structure is a chemical variable relationship graph structure; The graph nodes in the chemical variable relationship graph structure are defined by controlled variables, manipulated variables or measurable variables, and the edge connections in the chemical variable relationship graph structure are defined by the mechanism relationship between variables.

8. A storage medium having a computer program stored thereon, wherein: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

9. An electronic device, comprising: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the method described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images