Unequilibrium graph structure matching analogy reasoning method and system

Abstract

The invention provides an unbalanced graph structure matching analogy reasoning method and system. Performing multi-rule hybrid coding on the original base graph and the original target graph by constructing a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; constructing a robust generalization reasoning model independent of a specific vocabulary library, capturing key information of a graph structure, and utilizing the key information of the graph structure; constructing an unbalanced optimal transmission model; performing optimal transmission analogy reasoning graph matching according to the unbalanced optimal transmission model in combination with the coding information of the original base graph and the original target graph, and processing unmatched abnormal values in analogy graph matching by expanding strict mass conservation constraints; according to the unbalanced optimal transmission model and optimal transmission analogy reasoning graph matching, a graph matching reasoning loss function is constructed, the operation process is continualized, an unbalanced optimal transmission analogy reasoning model is formed, and the sub-graph matching capacity is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of analogical reasoning graph matching for intelligent systems, and more specifically, to an analogical reasoning method and system for unbalanced graph structure matching. Background Art

[0002] Analogical reasoning is an important part of problem solving, where the solution to one problem can be applied to another analogous problem. Structure-Mapping Theory (SMT) is one of the most influential theories, which focuses on mapping the relationship between source and target domains based on their structural similarities rather than surface features. Structural encoding and mapping in SMT can be viewed as an analogy subgraph matching problem, which mainly utilizes the graph structure (edge ​​relationships) while partially considering node semantics. Using interpretable methods to model human analogical reasoning and its downstream tasks has become one of the important issues in cognitive intelligence and a crucial step towards achieving general artificial intelligence. Analogical reasoning is a form of inductive reasoning and is considered as one of the important foundations of intelligent systems in the field of cognitive technology. From basic visual cognition [4] to abstract concept extraction, the process of analogical reasoning can be found in almost every level of simulating human cognition. The basic approach of analogical reasoning can be briefly summarized as mapping information in a source domain to a target domain while maintaining specific structural relationships in the information, including regularities, properties, and even behavioral motivations. Because of the transferability of this structural relationship, analogical reasoning helps to establish relationship mapping between two seemingly unrelated scenarios, thereby bringing about the transfer of knowledge and promoting a deeper understanding based on the relationship level; simulating most of the learning processes of humans is the process of analogical reasoning, and modeling analogical reasoning is naturally an indispensable part of realizing general artificial intelligence; how to explain the process of simulating human analogical reasoning; neural network-based methods are used to simulate the process of human analogical reasoning and modeling has become one of the research hotspots of cognitive intelligence; neural networks are applied to different intelligence tests to detect their analogical reasoning ability, but most of the principles in their black boxes are not explainable, which is obviously inconsistent with the understanding and modeling of simulating human analogical reasoning in cognitive intelligence; using structural mapping theory to model analogical reasoning has many defects in its interpretability, measurement methods and subsequent tasks; problems such as mathematical modeling of cognitive intelligence analogical reasoning have yet to be solved; therefore, it is necessary to propose an unbalanced graph structure matching analogical reasoning method and system to at least partially solve the problems existing in the prior art. Summary of the invention

[0003] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0004] To at least partially solve the above problems, the present invention provides an unbalanced graph structure matching analogy reasoning method, comprising: S100, constructing a hybrid coding structure mapping engine, performing multi-rule hybrid coding on the original base graph and the original target graph, and obtaining coding information of the original base graph and the original target graph; S200, builds a robust generalization reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure; builds an unbalanced optimal transmission model; S300, based on the unbalanced optimal transmission model, combines the original base graph and the original target graph encoding information to perform optimal transmission analogy reasoning graph matching, and handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints; S400, based on the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, constructs a graph matching reasoning loss function, and continuously performs the operation process to form an unbalanced optimal transmission analogy reasoning model to enhance the subgraph matching capability.

[0005] Preferably, S100 includes: S101, selecting semantic coding rules and structural coding rules; constructing a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; S102, performing multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original target graph coding information includes: original target graph semantic coding first coding information and original target graph structural coding second coding information.

[0006] Preferably, S200 includes: S201, building a robust generalization reasoning model that does not rely on a specific vocabulary library to enhance the generalization ability of relations in analogical reasoning; S202, introduces the attention mechanism to strengthen structural relationships and node identities; captures and utilizes key information of the graph structure during analogical reasoning; and constructs an unbalanced optimal transmission model.

[0007] Preferably, S300 includes: S301, performing optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; S302, in the optimal transmission analogy reasoning graph matching process, by extending the strict quality conservation constraint, processing the mismatching outliers in the analogy graph matching.

[0008] Preferably, S400 includes: S401, constructing a graph matching reasoning loss function according to the optimal transmission analogy reasoning graph matching; in the code construction process, a mask is defined according to the common matching matrix, so that the graph matching reasoning loss function operation process is continuous, and an unbalanced optimal transmission analogy reasoning model is constructed; S402, perform DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model; enhance the subgraph matching capability of the training model.

[0009] The present invention provides an unbalanced graph structure matching analogy reasoning system, comprising: The mapping engine hybrid coding module constructs a hybrid coding structure mapping engine, performs multi-rule hybrid coding on the original base graph and the original target graph, and obtains the coding information of the original base graph and the original target graph; The generalized reasoning transmission module builds a robust generalized reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure, and builds an unbalanced optimal transmission model; The analogy reasoning constraint extension module combines the original base graph and the original target graph encoding information according to the non-equilibrium optimal transmission model to perform optimal transmission analogy reasoning graph matching. It handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints. The inference loss graph matching enhancement module builds a graph matching inference loss function based on the unbalanced optimal transmission model and the optimal transmission analogy inference graph matching, and continuously calculates the process to form an unbalanced optimal transmission analogy inference model to enhance the subgraph matching capability. Preferably, the mapping engine hybrid encoding module includes: The hybrid coding structure mapping unit selects semantic coding rules and structural coding rules; and constructs a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; A multi-rule hybrid coding unit performs multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original target graph coding information includes: original target graph semantic coding first coding information and original target graph structural coding second coding information.

[0010] Preferably, the generalized reasoning transmission module includes: Analogy reasoning robust generalization unit, which builds a robust generalization reasoning model that does not rely on a specific vocabulary library and enhances the relational generalization ability in analogy reasoning; The unbalanced optimal transmission model unit introduces the attention mechanism to strengthen the structural relationship and node identity; in the analogical reasoning process, it captures and utilizes the key information of the graph structure; and constructs an unbalanced optimal transmission model.

[0011] Preferably, the analogy reasoning constraint extension module includes: The analogy reasoning graph matching unit performs optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; The mismatch outlier processing unit processes mismatch outliers in analogy graph matching by extending strict quality conservation constraints during the optimal transmission analogy reasoning graph matching process.

[0012] Preferably, the inference loss graph matching enhancement module comprises: The reasoning loss analogy model unit builds a graph matching reasoning loss function based on the optimal transmission analogy reasoning graph matching. During the code construction process, a mask is defined based on the common matching matrix to make the graph matching reasoning loss function calculation process continuous, thus building an unbalanced optimal transmission analogy reasoning model. The network training graph matching enhancement unit performs DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model, and enhances the subgraph matching capability of the training model.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention discloses an unbalanced graph structure matching analogy reasoning method and system. The method comprises the following steps: constructing a hybrid coding structure mapping engine, performing multi-rule hybrid coding on an original base graph and an original target graph, and obtaining coding information of the original base graph and the original target graph; constructing a robust generalized reasoning model that does not depend on a specific vocabulary library, capturing key information of the graph structure and utilizing the key information of the graph structure; constructing an unbalanced optimal transmission model; performing optimal transmission analogy reasoning graph matching according to the unbalanced optimal transmission model combined with the coding information of the original base graph and the original target graph, and processing unmatched outliers in the analogy graph matching by extending strict mass conservation constraints; constructing a graph matching reasoning loss function according to the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, and continuously performing the operation process to form an unbalanced optimal transmission analogy reasoning model, and enhancing the subgraph matching capability; the proposed model is divided into two parts for expansion, the first part is the coding rules and node embedding calculation of the model, and the second part is the mathematical expression and improvement method of the model calculation process; a solution based on the unbalanced optimal transmission theory: a new unbalanced optimal transmission solution is proposed for the problem of unmatched outliers in analogy graph matching. This scheme effectively handles outliers by relaxing the traditional strict mass conservation constraints, thereby improving the accuracy and robustness of analogy graph matching; Application of attention mechanism: In order to enhance the generalization ability of relations in analogy reasoning, an attention mechanism is introduced; This mechanism can mutually reinforce structural relations and node identities, so that the model can better capture and utilize key information in the graph structure during analogy reasoning; Construct a new loss function: According to the characteristics of the model, a corresponding loss function with practical physical significance is proposed, and it is used to train the neural network. It is found that after training, the model has emerged with better subgraph matching ability; Analogy reasoning is a form of inductive reasoning, which is regarded as an important foundation of human intelligence in the field of cognitive technology; It can acquire knowledge in new fields based on previous knowledge and solve problems through limited examples; The present invention proposes a brand-new analogy reasoning computational model. Through experiments on synthetic data and real data, it is found that the model performs better than previous models in analogy reasoning tasks and can extract common structures in different scenarios.

[0014] The unbalanced graph structure matching analogical reasoning method and system described in the present invention, other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by technicians in this field through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1A diagram showing an embodiment of an unbalanced graph structure matching analogy reasoning system according to the present invention; like Figure 1 As shown in the figure, the overall framework diagram of the model, the input original base graph and the original target graph are hash mapped, and the base graph semantic coding, target graph semantic coding, base graph structural coding and target graph structural coding are used. The input directed acyclic graph is processed by the long-term and short-term neural network to form four vectors Lb, Lt, Sb and St, and then the prediction matrix (transition probability matrix) is calculated by the unbalanced optimal transmission to obtain the common structure.

[0016] Figure 2 It is another embodiment diagram of the unbalanced graph structure matching analogy reasoning system of the present invention; like Figure 2 As shown, an original graph structure example of the hybrid coding structure mapping engine uses the structure mapping engine to represent the simulated system structure of the atomic nucleus attracting electrons; wherein, the orbiting rotation indicates the process of the electron orbiting the atomic nucleus, the mass of the atomic nucleus is greater than the mass of the electron, and the nodes of attraction, mass and charge form a directed acyclic graph structure relationship.

[0017] Figure 3 An embodiment diagram of semantic coding of an unbalanced graph structure matching analogy reasoning system and method according to the present invention; like Figure 3 As shown, when considering semantic information, according to hierarchical consistency, all entities are first uniformly encoded as the first entity label; then predicates with different labels are encoded as different semantic labels.

[0018] Figure 4 An embodiment diagram of the structural coding of an unbalanced graph structure matching analogy reasoning system and method according to the present invention; like Figure 4 As shown, the original base graph and the original target graph are encoded again according to the structural information. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments so that those skilled in the art can implement the invention with reference to the instructions. As shown in the figure, the present invention provides an analogy reasoning method for unbalanced graph structure matching, including: S100, constructing a hybrid coding structure mapping engine, performing multi-rule hybrid coding on the original base graph and the original target graph, and obtaining coding information of the original base graph and the original target graph; S200, builds a robust generalization reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure; builds an unbalanced optimal transmission model; S300, based on the unbalanced optimal transmission model, combines the original base graph and the original target graph encoding information to perform optimal transmission analogy reasoning graph matching, and handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints; S400, based on the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, constructs a graph matching reasoning loss function, and continuously performs the operation process to form an unbalanced optimal transmission analogy reasoning model to enhance the subgraph matching capability.

[0020] The principle and effect of the above technical solution are as follows: the present invention provides an unbalanced graph structure matching analogy reasoning method, including: constructing a hybrid coding structure mapping engine, performing multi-rule hybrid coding on the original base graph and the original target graph, and obtaining the encoding information of the original base graph and the original target graph; constructing a robust generalization reasoning model that does not rely on a specific vocabulary library, capturing the key information of the graph structure and utilizing the key information of the graph structure; constructing an unbalanced optimal transmission model; performing optimal transmission analogy reasoning graph matching according to the unbalanced optimal transmission model combined with the encoding information of the original base graph and the original target graph, and processing the unmatched outliers in the analogy graph matching by extending the strict mass conservation constraint; replacing the balanced optimal transmission model OT with the unbalanced optimal transmission model UOT; OT stands for the balanced optimal transmission model optimal transport theory model, and UOT stands for the unbalanced optimal transmission model unbalanced optimal transport theory model; constructing a graph matching reasoning loss function according to the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, and continuously operating the process to form an unbalanced optimal transmission analogy reasoning model to enhance the subgraph matching capability; the overall framework of the model is as follows Figure 1As shown in the figure, the proposed model is divided into two parts. The first part is the encoding rules and node embedding calculation of the model, and the second part is the mathematical expression and improvement method of the model calculation process; Solution based on unbalanced optimal transmission theory: A new unbalanced optimal transmission solution is proposed to solve the problem of mismatched outliers in analogy graph matching. This scheme effectively handles outliers by relaxing the traditional strict mass conservation constraints, thereby improving the accuracy and robustness of analogy graph matching; Application of attention mechanism: In order to enhance the generalization ability of relations in analogy reasoning, an attention mechanism is introduced; This mechanism can mutually reinforce structural relations and node identities, so that the model can better capture and utilize key information in the graph structure during analogy reasoning; Construct a new loss function: According to the characteristics of the model, a corresponding loss function with practical physical significance is proposed, and it is used to train the neural network. It is found that after training, the model has emerged with better subgraph matching ability; Analogy reasoning is a form of inductive reasoning, which is regarded as an important foundation of human intelligence in the field of cognitive technology; It can acquire knowledge in new fields based on previous knowledge and solve problems through limited examples; The present invention proposes a brand-new analogy reasoning computational model. Through experiments on synthetic data and real data, it is found that the model performs better than previous models in analogy reasoning tasks and can extract common structures in different scenarios.

[0021] In one embodiment, S100 includes: S101, selecting semantic coding rules and structural coding rules; constructing a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; S102, performing multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original target graph coding information includes: original target graph semantic coding first coding information and original target graph structural coding second coding information.

[0022] The principle and effect of the above technical solution are as follows: selecting semantic coding rules and structural coding rules; constructing a hybrid coding structure mapping engine according to the semantic coding rules and the structural coding rules; performing multi-rule hybrid coding on the original base graph and the original target graph through the hybrid coding structure mapping engine to obtain the coding information of the original base graph and the original target graph; the coding information of the original target graph includes: the first coding information of the semantic coding of the original target graph and the second coding information of the structural coding of the original target graph; performing multi-rule hybrid coding on the original base graph and the original target graph through the hybrid coding structure mapping engine to obtain the coding information of the original base graph and the original target graph including: the hybrid coding structure mapping engine performs multi-rule hybrid coding on the original base graph and the original target graph according to the two coding rules of semantic coding rules and structural coding, the first coding information of the semantic coding of the original target graph and the second coding information of the structural coding of the original target graph; when searching for a common structure, not only the semantic information of each node is paid attention to, but also the structural information of each node; constructing a hybrid coding structure mapping engine according to the semantic coding rules and the structural coding rules includes: Figure 2 As shown: there are two encoding rules: semantic encoding rule and structural encoding rule. Figure 2 The original target graph in the hybrid coding structure mapping engine instance is as follows: Figure 3 As shown in the figure: when considering semantic information, according to hierarchical consistency, all entities are first uniformly encoded as the first entity label; then predicates with different labels are encoded as different semantic labels; entities are uniformly encoded as the first entity label, including: such as electron (electron) and atomic nucleus (electron) are uniformly encoded as the first entity label of entity (entity); such as greater than (GREATER) is encoded as rp_7, attraction (ATTRACTS) is encoded as rp_9, and rotation around (REVOLVES-AROUND) is encoded as rp_12, retaining the semantic information differences of different nodes; for function nodes, their semantic information is still retained, but structural consistency will be considered when searching for common structures later, such as mass (MASS) is encoded as rf_7, and charge (ELECTRIC) is encoded as rf_3; the semantic encoding of the original base graph and the original target graph is obtained by this rule; the first encoding information of the semantic encoding of the original target graph is obtained; such as Figure 4As shown: the original base graph and the original target graph are encoded again according to the structural information; for entity nodes, the structural position of each node is distinguished, and each entity is encoded as a special structural code; such as electron (electron) is encoded as id_87, and nucleus (electron) is encoded as id_13; for predicate nodes, the number of parameters of the node and whether it is ordered are paid attention to, such as attraction (ATTRACTS), greater than (GREATER), and rotation (REVOLVES-AROUND) The number of parameters is 2, and they are all ordered nodes; the position of the parameters modified by the predicate is not swapped; rp_7, rp_9 and rp_12 multiple nodes are encoded as the same structural code p_2; similarly for functions, charge (ELECTRIC) and mass (MASS) are both unary functions, so they are encoded as the same structural code f_1; obtain the structural codes of the original base graph and the original target graph; obtain the second coding information of the structural code of the original target graph.

[0023] In one embodiment, S200 includes: S201, building a robust generalization reasoning model that does not rely on a specific vocabulary library to enhance the generalization ability of relations in analogical reasoning; S202, introduces the attention mechanism to strengthen structural relationships and node identities; captures and utilizes key information of the graph structure during analogical reasoning; and constructs an unbalanced optimal transmission model.

[0024] The principle and effect of the above technical solution are: construct a robust generalization reasoning model that does not depend on a specific vocabulary library, and enhance the generalization ability of relations in analogical reasoning; introduce an attention mechanism to strengthen structural relations and node identities; in the process of analogical reasoning, capture and utilize key information of the graph structure; construct an unbalanced optimal transmission model; construct a robust generalization model that does not depend on a specific vocabulary library, and enhance the generalization ability of relations in analogical reasoning, including: hashing the encoding of the node, and the value of the node after each hash is not necessarily the same; at the same time, it is also hoped that the differences between different types of nodes can be retained, so that different types of nodes can be distinguished. , the number of parameters and nodes of different order, preset the starting value of the hash table in advance, and use the boundary conditions to distinguish the types of nodes; when the model is actually applied, it is also necessary to input the same sample into the model multiple times, select the best prediction model using specific evaluation criteria, and then compare it with the actual match; obtain the encoding value of the node through this step; for the semantic encoding graph and the structural encoding graph, use two DAG-LSTM network structures with different parameters to encode the embedding information of the original base graph, and the embedding dimension of each node is 32 dimensions; normalize the obtained node embedding; perform distance comparison; The DAG-LSTM selected in the present invention is a tree LSTM network extension. For the calculation of each node embedding, the connected node embeddings are aggregated. For the directed acyclic graph, the topological structure of the graph is used for layer-by-layer calculation. Encoding ,use is the Sigmoid function, The function returns the successor node of this node. They represent the network parameters that need to be learned for various gate types; various gate types include: denote the input gate and output gate respectively, Represents the memory gate, represents the forget gate; Respectively represent the input gate weight vector, output gate weight vector, memory gate weight vector, and forget gate weight vector; , Respectively represent the input gate bias term and the output gate bias term; calculate their embedding ; After the above operations, multiple physical simulation embeddings can be obtained for the original base graph and the original target graph. The physical simulation embedding includes: base graph semantic embedding and base graph structural embedding, original target graph semantic embedding and original target graph structural embedding. The physical simulation embedding methods are respectively expressed as ; Use DAG-LSTM to embed and normalize the information of the directed acyclic graph, and embed the obtained physical simulation into the unbalanced optimal transmission model of the network.

[0025] In one embodiment, S300 includes: S301, performing optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; S302, in the optimal transmission analogy reasoning graph matching process, by extending the strict quality conservation constraint, processing the mismatching outliers in the analogy graph matching.

[0026] The principle and effect of the above technical solution are as follows: according to the unbalanced optimal transmission model, combined with the original base map and the original target map coding information, the analogy reasoning feature algorithm is used to perform optimal transmission analogy reasoning graph matching; in the optimal transmission analogy reasoning graph matching process, the mismatched outliers in the analogy graph matching are processed by extending the strict mass conservation constraint; according to the unbalanced optimal transmission model, combined with the original base map and the original target map coding information, the analogy reasoning feature algorithm is used to perform optimal transmission analogy reasoning graph matching, including: according to the unbalanced optimal transmission model, combined with the original base map and the original target map coding information, an algorithm that conforms to the analogy reasoning characteristics is calculated using the output of the neural network; the distance matrix Use Gromov distance for measurement; in the algorithm, The calculation is divided into two parts. Next, we match Take the distance between two points as an example, Base map Semantic embedding and structural embedding of points, Base map Semantic embedding and structural embedding of points; For semantic embedding, directly use Norm to calculate semantic distance: Its physical meaning is the distance of semantic embedding in the same metric space; Structural embedding calculates the distance between different matches based on the Gromov distance, and chooses to calculate all possible matches and The distance between them and the average of their squares are The distance from the point to all points in the base graph is The average of the distance differences between a point and all points in the original target graph is used to measure whether two points are close enough at the structural level; Distance Matrix The weighted sum of the above two distances yields: in, is a hyperparameter; perform optimal transmission analogy reasoning graph matching; In the optimal transmission analogy reasoning graph matching process, by extending the strict mass conservation constraint, the processing of mismatched outliers in analogy graph matching includes: specifying boundary conditions for algorithm iteration; since in almost all cases it is impossible to know in advance which points are included in the common part of the two images and which points should be ignored, the following settings are made in all iterative steps: That is, all points are set to be evenly distributed; the matching mapping of the common part is set to be much larger than other mappings at the same point, and the transfer probability of the points that should be ignored to all points tends to be average; According to the characteristics of the structure mapping engine, the output of the structure mapping engine focuses on the common structure with parallel connectivity and system, hoping that the transition probability of the matching node can have a high consistency with its neighboring nodes; through the optimization of the graph neural network, the convergence operation is added to the obtained transition probability, that is, The weighted sum of the transition probabilities is Among the transition probabilities, Similarly. Considering the characteristics of the shared structure of the structure mapping engine, the shared structure must extend to the leaf nodes in the lower layer of the graph topology, but there is no requirement for its parent node, so the weight value assigned to the child node must be greater than the weight value of the parent node; when only one layer of convergence is considered, the calculation formula is: in, are the base graph adjacency matrix and the original target graph adjacency matrix respectively, Transpose the base graph adjacency matrix and the original target graph adjacency matrix respectively. is a hyperparameter; however, considering that cousin nodes may appear in the shared structure, it is hoped that more distant nodes can be gathered, but the weight should decrease with the increase of distance; therefore, the matrix index is used to calculate the transition probability matrix after aggregation, and the mathematical expression is as follows: After simplification, it conforms to the above convergence form, where is a hyperparameter; handles unmatched outliers in analogy graph matching; effectively handles outliers, thereby improving the accuracy and robustness of analogy graph matching.

[0027] In one embodiment, S400 includes: S401, constructing a graph matching reasoning loss function according to the optimal transmission analogy reasoning graph matching; in the code construction process, a mask is defined according to the common matching matrix, so that the graph matching reasoning loss function operation process is continuous, and an unbalanced optimal transmission analogy reasoning model is constructed; S402, perform DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model; enhance the subgraph matching capability of the training model.

[0028] The principle and effect of the above technical solution are as follows: according to the optimal transmission analogy reasoning graph matching, a graph matching reasoning loss function is constructed; in the code construction process, a mask is defined according to the common matching matrix to make the graph matching reasoning loss function operation process continuous, and an unbalanced optimal transmission analogy reasoning model is constructed; according to the unbalanced optimal transmission analogy reasoning model, DAG-LSTM neural network training is performed; the subgraph matching capability of the training model is enhanced; Constructing the graph matching inference loss function includes: predicting the transition probability matrix of each transition, generating a sparse matrix of the matching probability between the base graph and the original target graph based on the common matching; the transition probability of the matching node pair is 1, and the transition probability of the remaining node pairs is 0; calculating the cross entropy error loss between the sparse matrix and the transition probability matrix, and learning the neural network with the cross entropy error loss function; according to the attention mechanism, for the matrix rows and columns that are all 0, the cross entropy error loss is learned to be close to the all-zero state; presetting the sum of each row of the matrix and each column of the matrix to be equal; comparing the matching nodes with the unmatched nodes Calculate the loss function separately; calculate the cross entropy error loss for all rows and columns where the matching nodes are located, and then take the average to obtain the mean cross entropy error loss of the matching nodes; calculate the cross entropy error loss for all rows and columns where the unmatched nodes are located, and then take the average to obtain the mean cross entropy error loss of the unmatched nodes; finally, take the weighted sum of the mean cross entropy error loss of the matching nodes and the mean cross entropy error loss of the unmatched nodes as the final loss, and the weight is determined by the ratio of the two parts and the number of nodes to balance the order of magnitude of the two parts of the loss; construct a graph matching inference loss function; The matrix output by the optimal transmission calculation process is a transition probability matrix that predicts each transition. Similarly, a sparse matrix P of the matching probability between the base graph and the original target graph can be generated based on the common matching. The position transfer probability of the matching node pair is 1, and the position transfer probability of the remaining node pairs is 0. The cross-entropy error loss is calculated by combining this matrix with the transition probability matrix output by the model, and the network is learned using this loss function. According to the attention mechanism, for the rows and columns of the matrix that are all 0, the cross-entropy error loss hopes to learn them to a state of all zeros. It is pre-set that the sum of each row of the matrix and each column of the matrix are equal respectively. A loss function suitable for the algorithm is proposed: the loss function is calculated separately for the matching and unmatching nodes. The cross-entropy loss is calculated twice for all rows and columns where the matching nodes are located. is the common matching matrix All non-zero positions in the i-th row are all matching positions. and j column position The cross entropy error loss function is calculated for all rows and columns. Cumulative sum, then add and average; set a total of d positions to be non-zero, and exclude these row and column positions to form a column of size matrix; n and m are the total number of rows and columns of the initial matrix respectively; we hope that the transfer probabilities of these positions are as average as possible during the transfer, which means that the transfer probabilities of these point pairs are not much different in physical sense and cannot be distinguished, which is in contrast to the most transfer with strong preference, so we choose negative entropy As the loss function of this part; the matrix after this part of the matrix passes through the softmax layer is , then the loss function of this part is for Submatrix of The logarithmic product is summed up and the final loss is calculated by taking the weighted sum of the two parts. The weight is determined by the size ratio of the two parts and the number of nodes to balance the order of magnitude of the two parts of loss. During the code operation, in order to solve the problem that the loss function cannot be back-propagated, a mask is defined based on the common matching matrix to make the function calculation process continuous. It is recorded as the matrix ;matrix In accordance with the above loss function calculation process, the rows and columns where the non-zero positions are located are both 1, corresponding to , the rest are 0, corresponding to , the final calculation is as follows: The model is used to train the DAG-LSTM neural network; the subgraph matching ability of the training model is enhanced; after training, the model shows good subgraph matching ability; analogical reasoning is a form of inductive reasoning, which is regarded as an important foundation of human intelligence in the field of cognitive technology; it can acquire knowledge in new fields based on previous knowledge and solve problems through limited examples.

[0029] The present invention provides an unbalanced graph structure matching analogy reasoning system, comprising: The mapping engine hybrid coding module constructs a hybrid coding structure mapping engine, performs multi-rule hybrid coding on the original base graph and the original target graph, and obtains the coding information of the original base graph and the original target graph; The generalized reasoning transmission module builds a robust generalized reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure, and builds an unbalanced optimal transmission model; The analogy reasoning constraint extension module combines the original base graph and the original target graph encoding information according to the non-equilibrium optimal transmission model to perform optimal transmission analogy reasoning graph matching. It handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints. The inference loss graph matching enhancement module builds a graph matching inference loss function based on the unbalanced optimal transmission model and the optimal transmission analogy inference graph matching, and continuously calculates the process to form an unbalanced optimal transmission analogy inference model to enhance the subgraph matching capability. The principle and effect of the above technical solution are as follows: the present invention provides an unbalanced graph structure matching analogy reasoning system, including: a mapping engine hybrid coding module, which constructs a hybrid coding structure mapping engine, performs multi-rule hybrid coding on the original base graph and the original target graph, and obtains the encoding information of the original base graph and the original target graph; a generalized reasoning transmission module, which constructs a robust generalized reasoning model that does not rely on a specific vocabulary library, captures the key information of the graph structure and utilizes the key information of the graph structure; constructs an unbalanced optimal transmission model; an analogy reasoning constraint extension module, which performs optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model combined with the encoding information of the original base graph and the original target graph, and handles the mismatched outliers in the analogy graph matching by extending the strict mass conservation constraint; an inference loss graph matching enhancement module, which constructs a graph matching inference loss function based on the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, and continuously operates the process to form an unbalanced optimal transmission analogy reasoning model to enhance the subgraph matching capability; the overall framework of the model is as follows Figure 1 As shown in the figure, the proposed model is divided into two parts. The first part is the encoding rules and node embedding calculation of the model, and the second part is the mathematical expression and improvement method of the model calculation process; Solution based on unbalanced optimal transmission theory: A new unbalanced optimal transmission solution is proposed to solve the problem of mismatched outliers in analogy graph matching. This scheme effectively handles outliers by relaxing the traditional strict mass conservation constraints, thereby improving the accuracy and robustness of analogy graph matching; Application of attention mechanism: In order to enhance the generalization ability of relations in analogy reasoning, an attention mechanism is introduced; This mechanism can mutually reinforce structural relations and node identities, so that the model can better capture and utilize key information in the graph structure during analogy reasoning; Construct a new loss function: According to the characteristics of the model, a corresponding loss function with practical physical significance is proposed, and it is used to train the neural network. It is found that after training, the model has emerged with better subgraph matching ability; Analogy reasoning is a form of inductive reasoning, which is regarded as an important foundation of human intelligence in the field of cognitive technology; It can acquire knowledge in new fields based on previous knowledge and solve problems through limited examples; The present invention proposes a brand-new analogy reasoning computational model. Through experiments on synthetic data and real data, it is found that the model performs better than previous models in analogy reasoning tasks and can extract common structures in different scenarios.

[0030] In one embodiment, the mapping engine hybrid encoding module includes: The hybrid coding structure mapping unit selects semantic coding rules and structural coding rules; and constructs a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; A multi-rule hybrid coding unit performs multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original target graph coding information includes: original target graph semantic coding first coding information and original target graph structural coding second coding information.

[0031] The principle and effect of the above technical solution are as follows: the mapping engine hybrid coding module includes: a hybrid coding structure mapping unit, which selects semantic coding rules and structural coding rules; constructs a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; a multi-rule hybrid coding unit, which performs multi-rule hybrid coding on the original base graph and the original target graph through the hybrid coding structure mapping engine to obtain the encoding information of the original base graph and the original target graph; the encoding information of the original target graph includes: the first encoding information of the semantic encoding of the original target graph and the second encoding information of the structural encoding of the original target graph; Through the hybrid coding structure mapping engine, the original base graph and the original target graph are subjected to multi-rule hybrid coding to obtain the coding information of the original base graph and the original target graph, including: the hybrid coding structure mapping engine, according to the semantic coding rule and the structural coding two coding rules, respectively, the original target graph semantic coding first coding information, the original target graph structural coding second coding information; when searching for the common structure, not only the semantic information of each node is paid attention to, but also the structural information of each node; according to the semantic coding rule and the structural coding rule, the hybrid coding structure mapping engine is constructed, including: Figure 2 As shown: there are two encoding rules: semantic encoding rule and structural encoding rule. Figure 2 The original target graph in the hybrid coding structure mapping engine instance is as follows: Figure 3 As shown in the figure: when considering semantic information, according to hierarchical consistency, all entities are first uniformly encoded as the first entity label; then predicates with different labels are encoded as different semantic labels; entities are uniformly encoded as the first entity label, including: such as electron (electron) and atomic nucleus (electron) are uniformly encoded as the first entity label of entity (entity); such as greater than (GREATER) is encoded as rp_7, attraction (ATTRACTS) is encoded as rp_9, and rotation around (REVOLVES-AROUND) is encoded as rp_12, retaining the semantic information differences of different nodes; for function nodes, their semantic information is still retained, but structural consistency will be considered when searching for common structures later, such as mass (MASS) is encoded as rf_7, and charge (ELECTRIC) is encoded as rf_3; the semantic encoding of the original base graph and the original target graph is obtained by this rule; the first encoding information of the semantic encoding of the original target graph is obtained; such as Figure 4As shown: the original base graph and the original target graph are encoded again according to the structural information; for entity nodes, the structural position of each node is distinguished, and each entity is encoded as a special structural code; such as electron (electron) is encoded as id_87, and nucleus (electron) is encoded as id_13; for predicate nodes, the number of parameters of the node and whether it is ordered are paid attention to, such as attraction (ATTRACTS), greater than (GREATER), and rotation (REVOLVES-AROUND) The number of parameters is 2, and they are all ordered nodes; the position of the parameters modified by the predicate is not swapped; rp_7, rp_9 and rp_12 multiple nodes are encoded as the same structural code p_2; similarly for functions, charge (ELECTRIC) and mass (MASS) are both unary functions, so they are encoded as the same structural code f_1; obtain the structural codes of the original base graph and the original target graph; obtain the second coding information of the structural code of the original target graph.

[0032] In one embodiment, the generalized reasoning transmission module includes: Analogy reasoning robust generalization unit, which builds a robust generalization reasoning model that does not rely on a specific vocabulary library and enhances the relational generalization ability in analogy reasoning; The unbalanced optimal transmission model unit introduces the attention mechanism to strengthen the structural relationship and node identity; in the analogical reasoning process, it captures and utilizes the key information of the graph structure; and constructs an unbalanced optimal transmission model.

[0033] The principle and effect of the above technical solution are as follows: the generalized reasoning transmission module includes: the analogy reasoning robust generalization unit, which builds a robust generalized reasoning model that does not rely on a specific vocabulary library and enhances the generalization ability of relations in analogy reasoning; the unbalanced optimal transmission model unit, which introduces an attention mechanism to strengthen structural relations and node identities; in the analogy reasoning process, it captures and utilizes key information of the graph structure; and builds an unbalanced optimal transmission model; Constructing a robust generalization model that does not rely on a specific vocabulary library and enhancing the generalization ability of relations in analogical reasoning include: hashing the node encoding, and the value of the node after each hash is not necessarily the same; at the same time, it is also hoped that the difference between different types of nodes can be retained, so the starting value of the hash table is preset in advance for nodes of different types, number of parameters and different order, and the node type is distinguished by boundary conditions; when the model is actually applied, it is also necessary to input the same sample into the model multiple times, select the best prediction model using specific evaluation criteria, and then compare it with the actual match; obtain the node encoding value through this step; for the semantic encoding graph and the structural encoding graph, use two DAG-LSTM network structures with different parameters to encode the original base graph embedding information, and the embedding dimension of each node is 32 dimensions; normalize the obtained node embedding; perform distance comparison; The DAG-LSTM selected in the present invention is a tree LSTM network extension. For the calculation of each node embedding, the connected node embeddings are aggregated. For the directed acyclic graph, the topological structure of the graph is used for layer-by-layer calculation. Encoding ,use is the Sigmoid function, The function returns the successor node of this node. They represent the network parameters that need to be learned for various gate types; various gate types include: denote the input gate and output gate respectively, Represents the memory gate, represents the forget gate; Respectively represent the input gate weight vector, output gate weight vector, memory gate weight vector, and forget gate weight vector; , Respectively represent the input gate bias term and the output gate bias term; calculate their embedding ; After the above operations, multiple physical simulation embeddings can be obtained for the original base graph and the original target graph. The physical simulation embedding includes: base graph semantic embedding and base graph structural embedding, original target graph semantic embedding and original target graph structural embedding. The physical simulation embedding methods are respectively expressed as ; Use DAG-LSTM to embed and normalize the information of the directed acyclic graph, and embed the obtained physical simulation into the unbalanced optimal transmission model of the network.

[0034] In one embodiment, the analogy reasoning constraint expansion module includes: The analogy reasoning graph matching unit performs optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; The mismatch outlier processing unit processes mismatch outliers in analogy graph matching by extending strict quality conservation constraints during the optimal transmission analogy reasoning graph matching process.

[0035] The principle and effect of the above technical solution are as follows: the analogy reasoning constraint expansion module includes: an analogy reasoning graph matching unit, which performs optimal transmission analogy reasoning graph matching according to the unbalanced optimal transmission model, combined with the original base graph and the original target graph coding information, and uses the analogy reasoning feature algorithm; a mismatching outlier processing unit, which processes mismatching outliers in analogy graph matching by extending strict mass conservation constraints during the optimal transmission analogy reasoning graph matching process; the analogy reasoning feature algorithm is used to perform optimal transmission analogy reasoning graph matching according to the unbalanced optimal transmission model, combined with the original base graph and the original target graph coding information; the distance matrix is ​​calculated according to the unbalanced optimal transmission model, combined with the original base graph and the original target graph coding information, and uses the analogy reasoning feature algorithm to calculate the analogy reasoning characteristics; Use Gromov distance for measurement; in the algorithm, The calculation is divided into two parts. Next, we match Take the distance between two points as an example, Base map Semantic embedding and structural embedding of points, Base map Semantic embedding and structural embedding of points; For semantic embedding, directly use Norm to calculate semantic distance: Its physical meaning is the distance of semantic embedding in the same metric space; Structural embedding calculates the distance between different matches based on the Gromov distance, and chooses to calculate all possible matches and The distance between them and the average of their squares are The distance from the point to all points in the base graph is The average of the distance differences between a point and all points in the original target graph is used to measure whether two points are close enough at the structural level; Distance Matrix The weighted sum of the above two distances yields: in, is a hyperparameter; perform optimal transmission analogy reasoning graph matching; In the optimal transmission analogy reasoning graph matching process, by extending the strict mass conservation constraint, the processing of mismatched outliers in analogy graph matching includes: specifying boundary conditions for algorithm iteration; since in almost all cases it is impossible to know in advance which points are included in the common part of the two images and which points should be ignored, the following settings are made in all iterative steps: That is, all points are set to be evenly distributed; the matching mapping of the common part is set to be much larger than other mappings at the same point, and the transfer probability of the points that should be ignored to all points tends to be average; According to the characteristics of the structure mapping engine, the output of the structure mapping engine focuses on the common structure with parallel connectivity and system, hoping that the transition probability of the matching node can have a high consistency with its neighboring nodes; through the optimization of the graph neural network, the convergence operation is added to the obtained transition probability, that is, The weighted sum of the transition probabilities is Among the transition probabilities, Similarly. Considering the characteristics of the shared structure of the structure mapping engine, the shared structure must extend to the leaf nodes in the lower layer of the graph topology, but there is no requirement for its parent node, so the weight value assigned to the child node must be greater than the weight value of the parent node; when only one layer of convergence is considered, the calculation formula is: in, are the base graph adjacency matrix and the original target graph adjacency matrix respectively, Transpose the base graph adjacency matrix and the original target graph adjacency matrix respectively. is a hyperparameter; however, considering that cousin nodes may appear in the shared structure, it is hoped that more distant nodes can be gathered, but the weight should decrease with the increase of distance; therefore, the matrix index is used to calculate the transition probability matrix after aggregation, and the mathematical expression is as follows: After simplification, it conforms to the above convergence form, where is a hyperparameter; handles unmatched outliers in analogy graph matching; effectively handles outliers, thereby improving the accuracy and robustness of analogy graph matching.

[0036] In one embodiment, the inference loss graph matching enhancement module includes: The reasoning loss analogy model unit builds a graph matching reasoning loss function based on the optimal transmission analogy reasoning graph matching. During the code construction process, a mask is defined based on the common matching matrix to make the graph matching reasoning loss function calculation process continuous, thus building an unbalanced optimal transmission analogy reasoning model. The network training graph matching enhancement unit performs DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model, and enhances the subgraph matching capability of the training model.

[0037] The principle and effect of the above technical solution are as follows: the inference loss graph matching enhancement module includes: the inference loss analogy model unit, which constructs a graph matching inference loss function according to the optimal transmission analogy inference graph matching; in the code construction process, a mask is defined according to the common matching matrix to make the graph matching inference loss function operation process continuous, and construct an unbalanced optimal transmission analogy inference model; the network training graph matching enhancement unit performs DAG-LSTM neural network training according to the unbalanced optimal transmission analogy inference model; enhances the subgraph matching ability of the training model; constructs a graph matching inference loss function including: predicting the transition probability matrix of each transfer, generating a sparse matrix of the matching probability of the base graph and the original target graph according to the common matching; the matching node pair position transfer probability is 1, and the remaining node pair position transfer probability is 0; the sparse matrix and the transition probability matrix are used to calculate the cross entropy error loss, The neural network is learned with the cross entropy error loss function; according to the attention mechanism, for the rows and columns of the matrix that are all zero, the cross entropy error loss is learned to be close to the all-zero state; the sum of each row of the matrix and each column of the matrix is ​​pre-set to be equal; the loss function is calculated separately for the matching nodes and the unmatched nodes; the cross entropy error loss is calculated for all rows and columns where the matching nodes are located, and then the average is taken to obtain the mean cross entropy error loss of the matching nodes; the cross entropy error loss is calculated for all rows and columns where the unmatched nodes are located, and then the average is taken to obtain the mean cross entropy error loss of the unmatched nodes; finally, the weighted sum of the mean cross entropy error loss of the matching nodes and the mean cross entropy error loss of the unmatched nodes is taken as the final loss, and the weight is determined by the size ratio of the two parts and the number of nodes to balance the order of magnitude of the two parts of the loss; a graph matching inference loss function is constructed; The matrix output by the optimal transmission calculation process is a transition probability matrix that predicts each transition. Similarly, a sparse matrix P of the matching probability between the base graph and the original target graph can be generated based on the common matching. The position transfer probability of the matching node pair is 1, and the position transfer probability of the remaining node pairs is 0. The cross-entropy error loss is calculated by combining this matrix with the transition probability matrix output by the model, and the network is learned using this loss function. According to the attention mechanism, for the rows and columns of the matrix that are all 0, the cross-entropy error loss hopes to learn them to a state of all zeros. It is pre-set that the sum of each row of the matrix and each column of the matrix are equal respectively. A loss function suitable for the algorithm is proposed: the loss function is calculated separately for the matching and unmatching nodes. The cross-entropy loss is calculated twice for all rows and columns where the matching nodes are located. is the common matching matrix All non-zero positions in the i-th row are all matching positions. and j column position The cross entropy error loss function is calculated for all rows and columns. Cumulative sum, then add and average; set a total of d positions to be non-zero, and exclude these row and column positions to form a column of size matrix; n and m are the total number of rows and columns of the initial matrix respectively; we hope that the transfer probabilities of these positions are as average as possible during the transfer, which means that the transfer probabilities of these point pairs are not much different in physical sense and cannot be distinguished, which is in contrast to the most transfer with strong preference, so we choose negative entropy As the loss function of this part; the matrix after this part of the matrix passes through the softmax layer is , then the loss function of this part is for Submatrix of The logarithmic product is summed up and the final loss is calculated by taking the weighted sum of the two parts. The weight is determined by the size ratio of the two parts and the number of nodes to balance the order of magnitude of the two parts of loss. During the code operation, in order to solve the problem that the loss function cannot be back-propagated, a mask is defined based on the common matching matrix to make the function calculation process continuous. It is recorded as the matrix ;matrix In accordance with the above loss function calculation process, the rows and columns where the non-zero positions are located are both 1, corresponding to , the rest are 0, corresponding to , the final calculation is as follows: The model is used to train the DAG-LSTM neural network; the subgraph matching ability of the training model is enhanced; after training, the model shows good subgraph matching ability; analogical reasoning is a form of inductive reasoning, which is regarded as an important foundation of human intelligence in the field of cognitive technology; it can acquire knowledge in new fields based on previous knowledge and solve problems through limited examples.

[0038] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. An unbalanced graph structure matching analogy reasoning method, characterized in that: include: S100, constructing a hybrid coding structure mapping engine, performing multi-rule hybrid coding on the original base graph and the original target graph, and obtaining coding information of the original base graph and the original target graph; S200, builds a robust generalization reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure; builds an unbalanced optimal transmission model; S300, based on the unbalanced optimal transmission model, combines the original base graph and the original target graph encoding information to perform optimal transmission analogy reasoning graph matching, and handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints; S400, based on the unbalanced optimal transmission model and the optimal transmission analogy reasoning graph matching, constructs a graph matching reasoning loss function, and continuously performs the operation process to form an unbalanced optimal transmission analogy reasoning model to enhance the subgraph matching capability.

2. The unbalanced graph structure matching analogy reasoning method according to claim 1, characterized in that: S100 includes: S101, selecting semantic coding rules and structural coding rules; constructing a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; S102, performing multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original base graph and the original target graph coding information include: the original base graph semantic coding first coding information, the original base graph structural coding second coding information, the original target graph semantic coding first coding information, and the original target graph structural coding second coding information.

3. The unbalanced graph structure matching analogy reasoning method according to claim 1, characterized in that: S200 includes: S201, building a robust generalization reasoning model that does not rely on a specific vocabulary library to enhance the generalization ability of relations in analogical reasoning; S202, introduces the attention mechanism to strengthen structural relationships and node identities; captures and utilizes key information of the graph structure during analogical reasoning; and constructs an unbalanced optimal transmission model.

4. The unbalanced graph structure matching analogy reasoning method according to claim 1, characterized in that: S300 includes: S301, performing optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; S302, in the optimal transmission analogy reasoning graph matching process, by extending the strict quality conservation constraint, processing the mismatching outliers in the analogy graph matching.

5. The unbalanced graph structure matching analogy reasoning method according to claim 1, characterized in that: S400 includes: S401, constructing a graph matching reasoning loss function according to the optimal transmission analogy reasoning graph matching; in the code construction process, a mask is defined according to the common matching matrix, so that the graph matching reasoning loss function operation process is continuous, and an unbalanced optimal transmission analogy reasoning model is constructed; S402, perform DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model; enhance the subgraph matching capability of the training model.

6. An unbalanced graph structure matching analogy reasoning system, characterized in that: include: The mapping engine hybrid coding module constructs a hybrid coding structure mapping engine, performs multi-rule hybrid coding on the original base graph and the original target graph, and obtains the coding information of the original base graph and the original target graph; The generalized reasoning transmission module builds a robust generalized reasoning model that does not rely on a specific vocabulary library, captures and utilizes key information of the graph structure, and builds an unbalanced optimal transmission model; The analogy reasoning constraint extension module combines the original base graph and the original target graph encoding information according to the non-equilibrium optimal transmission model to perform optimal transmission analogy reasoning graph matching. It handles mismatched outliers in analogy graph matching by extending strict quality conservation constraints. The inference loss graph matching enhancement module constructs a graph matching inference loss function based on the unbalanced optimal transmission model and the optimal transmission analogy inference graph matching, and continuously performs the calculation process to form an unbalanced optimal transmission analogy inference model to enhance the subgraph matching capability.

7. The unbalanced graph structure matching analogy reasoning system according to claim 6, characterized in that: Mapping engine hybrid encoding module, including: The hybrid coding structure mapping unit selects semantic coding rules and structural coding rules; and constructs a hybrid coding structure mapping engine according to the semantic coding rules and structural coding rules; A multi-rule hybrid coding unit performs multi-rule hybrid coding on the original base graph and the original target graph through a hybrid coding structure mapping engine to obtain coding information of the original base graph and the original target graph; The original base graph and the original target graph coding information include: the original base graph semantic coding first coding information, the original base graph structural coding second coding information, the original target graph semantic coding first coding information, and the original target graph structural coding second coding information.

8. The unbalanced graph structure matching analogy reasoning system according to claim 6, characterized in that: Generalized reasoning transmission module, including: Analogy reasoning robust generalization unit, which builds a robust generalization reasoning model that does not rely on a specific vocabulary library and enhances the relational generalization ability in analogy reasoning; The unbalanced optimal transmission model unit introduces the attention mechanism to strengthen the structural relationship and node identity; in the analogical reasoning process, it captures and utilizes the key information of the graph structure; and constructs an unbalanced optimal transmission model.

9. The unbalanced graph structure matching analogy reasoning system according to claim 6, characterized in that: Analogy reasoning constraint extension module, including: The analogy reasoning graph matching unit performs optimal transmission analogy reasoning graph matching based on the unbalanced optimal transmission model, combining the original base graph and the original target graph encoding information, and using the analogy reasoning feature algorithm; The mismatch outlier processing unit processes mismatch outliers in analogy graph matching by extending strict quality conservation constraints during the optimal transmission analogy reasoning graph matching process.

10. The unbalanced graph structure matching analogy reasoning system according to claim 6, characterized in that: Inference loss graph matching enhancement module, including: The reasoning loss analogy model unit builds a graph matching reasoning loss function based on the optimal transmission analogy reasoning graph matching. During the code construction process, a mask is defined based on the common matching matrix to make the graph matching reasoning loss function calculation process continuous, thus building an unbalanced optimal transmission analogy reasoning model. The network training graph matching enhancement unit performs DAG-LSTM neural network training based on the unbalanced optimal transmission analogy reasoning model, and enhances the subgraph matching capability of the training model.

Images