A knowledge graph evolution reasoning calculation method based on graph comparison and meta learning

Abstract

The application relates to the technical field of knowledge graph computation, and discloses a knowledge graph evolution reasoning computation method based on graph comparison and meta learning, which comprises the following steps: constructing a heterogeneous graph and generating an entity feature vector by using double-view comparison learning; extracting a meta feature vector containing node degree statistics and graph density, and outputting a local discrimination threshold through a meta threshold network; combining multidimensional similarity and Horn clause rules, and calculating and merging confidence through a gate unit; inputting candidate pairs higher than the local discrimination threshold into a Markov decision model, combining a graph average clustering coefficient improvement amount, a validation set triple link prediction score increment, and deducting a judgment rejection penalty term to calculate an instant reward, arranging a merging order, and updating a topological state; and the application eliminates global static threshold distortion, dynamically balances dense area false merging and sparse area missing detection contradictions, simultaneously blocks error cascade accumulation caused by local greedy merging, and guarantees the global topological structure stability of knowledge graph evolution reasoning.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph computing technology, and in particular relates to a knowledge graph evolutionary reasoning computing method based on graph comparison and meta-learning. Background Technology

[0002] When using specific computational models for knowledge graph evolution and reasoning, graph attention networks and vector space embeddings are typically used to represent entity attributes and relationships between entities. Semantic representations of entity nodes are formed by aggregating features of neighboring nodes, providing a foundation for subsequent knowledge reasoning and topological analysis. As dynamic data is continuously added and the knowledge graph expands, entity name variations and topological connection differences also increase. The graph often contains both densely connected core entities with high connectivity and sparsely connected edge long-tail entities. The topological density varies significantly in different regions. In this case, traditional methods that rely on static alias dictionaries, fixed similarity thresholds, or pure semantic vector matching are difficult to adapt to changes in local topological features and semantic context simultaneously. Fixed discrimination boundaries are also difficult to adjust with the evolution of the graph structure.

[0003] To overcome the limitations of static discrimination boundaries, existing technologies attempt to introduce dynamic adjustment mechanisms to improve entity matching accuracy. For example, Chinese invention patent CN112131395B discloses an iterative knowledge graph entity alignment method based on dynamic thresholds. This method captures candidate pairs and adds them to the training set by using a dynamic threshold that decays uniformly with each iteration. However, the dynamic threshold used in this scheme is essentially just a single time scalar that depends on the number of global iteration steps. Its premise is that all regions of the entire graph have a homogeneous evolution trend, which is out of touch with the objective reality that the topological density of local subgraphs in real knowledge graphs is highly heterogeneous. Due to the lack of perception of local degree distribution and topological density, this global decay threshold is prone to introducing incorrectly merged samples in high-connectivity core regions due to excessive threshold lowering, while it is difficult to adaptively relax in long-tail sparse regions to capture weak connections. In addition, it only relies on unidirectional iterative accumulation and lacks a topological feedback mechanism and sequential decision constraints after merging, which leads to the cascading spread of local discrimination bias during the iteration process.

[0004] Therefore, how to dynamically adjust the discrimination threshold based on the local graph topology density without relying on manual prior rules, limit the high computational cost to the region with blurred boundaries, arrange the entity merging order through global state feedback, and maintain the parameter stability of the model during continuous incremental evolution remains a technical problem that needs to be solved. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of this invention is as follows: A knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning, comprising the following steps: Step S1: Obtain entity nodes, relation types and corresponding triples in the knowledge graph to construct a heterogeneous graph. Construct a structural view by randomly discarding edge samples. Construct a semantic view by masking and perturbing the entity description text. Generate an entity feature vector by using contrastive loss to bring the node representations of the structural view and the semantic view closer together. Step S2: Extract the meta-feature vector of the entity pair to be discriminated. The meta-feature vector contains the logarithmic degree values ​​of the two entity nodes, the mean degree of the entity type, the standard deviation of degree, the degree quartiles, and the global graph density. Input the meta-feature vector into the meta-thresholding network to output the local discrimination threshold. Step S3: Calculate the structural semantic similarity of the entity pairs to be judged based on the entity feature vectors, and construct a basic similarity vector by combining the name similarity and topological overlap of the entity pairs to be judged. Combine the activation intensity vector of the Horn clause rules summarized from the historical judgment database, and calculate the merge confidence of the entity pairs to be judged as candidate pairs through the gating unit. Step S4: Input candidate pairs with a confidence level higher than the local discrimination threshold into the Markov decision process model. Combine the graph average clustering coefficient improvement, the validation set triple link prediction score increment, and deduct the decision rejection penalty to calculate the immediate reward. Use the policy network to arrange the merging order, and complete the topology state update of the local subgraph after confirming the merging.

[0006] Preferably, the generation of entity feature vectors in step S1 includes: step S11, aggregating neighbor node features under relation type in the heterogeneous graph using a relation-aware graph attention encoder, wherein the initial node features are generated by encoding entity names and description text by a language model and fusing entity type embeddings; step S12, constructing a structural view by discarding heterogeneous graph edges using a Bernoulli distribution mask matrix with a parameter of 0.2, and calculating a normalized temperature scale cross-entropy loss to bring the node representations of the two views closer together, and taking the mean of the output vectors of the two views as the final entity feature vector.

[0007] Preferably, step S2, which involves inputting the meta-feature vector into the meta-thresholding network to output a local discrimination threshold, includes: step S21, where the meta-feature vector also includes an entity type embedding vector and the ratio of edges to nodes in the entire graph; the meta-feature vector is input into a meta-thresholding network containing a three-layer multilayer perceptron structure, and a local discrimination threshold in the open interval of 0 to 1 is output via nonlinear mapping; the meta-thresholding network uses a first-order reptile algorithm to iteratively update parameters along the gradient direction that maximizes the entity matching evaluation index on the validation set data.

[0008] Preferably, step S4, which involves inputting candidate pairs with a confidence level higher than the local discrimination threshold into the Markov decision process model, includes: step S41, constructing the queue of candidate pairs to be processed in the graph, the topological state of the entire graph entities, and the global clustering features into a state space, and defining the confirmation of merging and skipping of candidate pairs as the action space; step S42, using a proximal policy optimization algorithm to optimize the policy network parameters, using the merging confidence level to guide the behavior cloning warm-up in the initial training stage, and advancing entity merging according to the output order of the optimized policy network.

[0009] Preferably, step S4 further includes a two-way mutual verification judgment step S43 before arranging the merging order. Step S431: Extract fuzzy candidate pairs with a merging confidence level between 0.4 and 0.75, and input the name, description, and shared relationship triple of the fuzzy candidate pairs into the external language model; Step S432: Use the forward inference chain to complete the same entity judgment, and use the reverse challenge to find contradictions between attributes and spatiotemporal logic; Step S433: Under a set number of sampling times, when the frequency of forward judgments that the same entity reaches a preset judgment threshold and the reverse challenge does not detect a substantial contradiction, the fuzzy candidate pair is judged to be a core-referenced entity, and the core-referenced logic pattern extracted during the inference process is added to the rule base.

[0010] Preferably, after step S43, a global graph partitioning and evidence propagation step S434 is included. Step S4341 involves constructing a weighted graph using verified core-inferred entities, with edge weights generated by weighting based on the frequency of affirmative judgments in the language model, combined confidence, and rule confidence. Step S4342 involves solving for the generalized eigenvalues ​​of the Laplacian matrix of the weighted graph to obtain eigenvectors, and recursively bisecting the graph structure according to the eigenvector signs. Step S4343 involves running a confidence propagation algorithm within the bisected subgraph, constructing a potential function based on node attribute purity and edge weights to update local node messages, and removing nodes with low marginal probabilities after convergence to output high-purity core-inferred clusters.

[0011] Preferably, after completing the topology state update, step S4 further includes a model continuous update step S44. Step S441 involves storing the merged and confirmed positive samples and strong negative samples with similarity higher than 0.85 but determined to be different entities into a scenario memory buffer of a set capacity. Step S442 involves calculating the diagonal elements of the parameter importance Fisher information matrix when the number of new samples in the scenario memory buffer reaches the update threshold, and introducing elastic weights to consolidate the weight shifts of the loss constraint graph attention encoder, gating unit, and meta-threshold network. The joint parameter update is completed by combining the historical samples replayed from the scenario memory buffer.

[0012] Preferably, in step S44, an online meta-learning update step S443 is adopted for the meta-threshold network. In step S443, when a new batch of data is accessed, the current loss gradient is calculated on the validation sample buffer, and the network weights of the meta-threshold network are directly adjusted in a single step to maintain the dynamic response capability of the meta-threshold network to changes in local topological density at each stage of graph evolution.

[0013] Preferably, after completing the topology state update in step S4, the process also includes a lightweight deployment and incremental inference step S45 on the edge. In step S451, a complete architecture including a graph attention encoder and a gating unit is used as the teacher network, and the output of the gating unit and the entity judgment result are used as soft labels to train the dual-tower structure student model. In step S452, when a new entity node is added to the knowledge graph, the first-hop neighborhood features of the new entity node are aggregated on the edge through one-step graph convolution, and the basic confidence is calculated using the dual-tower structure student model. Only the data with confidence in the judgment blind zone is uploaded to the computing cluster to complete the complete inference.

[0014] Compared with existing technologies, the knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning of this invention has the following advantages: 1. In the knowledge graph evolution and reasoning computation of graph comparison and meta-learning, the meta-threshold network dynamically adjusts the local discrimination threshold based on the local connectivity degree of entity nodes and the topological density of the whole graph. Combined with the entity features obtained by dual-view comparison, the discrimination benchmark can be adjusted with the topological density of different regions of the graph. Thus, the probability of false entity merging can be reduced in densely connected regions, and the number of missed topological associations can be reduced in sparsely connected long-tail regions. This avoids the problem that a fixed similarity threshold cannot take into account different graph structures, and enables the knowledge graph to maintain good entity discrimination adaptability during continuous evolution.

[0015] 2. By combining the basic similarity with the judgment criteria provided by the gating unit and the Horn clause rule, the ambiguity caused by relying solely on literal or single similarity can be reduced, and the reliability of entity merging judgment can be improved. On this basis, the Markov decision process model is used to arrange the merging order of candidate entities according to the changes in the graph state, so that the topological changes caused by the previous merging can participate in subsequent decisions, thereby reducing the local greedy bias caused by independent pairwise judgments, suppressing the continuous accumulation of errors along the graph structure, and helping to maintain the stability of the overall topological structure in the evolution of the knowledge graph.

[0016] 3. By focusing complex reasoning on candidate pairs with relatively ambiguous judgment results and using forward reasoning and reverse contradiction checking for bidirectional mutual verification, the impact of generative model judgment fluctuations on entity recognition results can be reduced. At the same time, high-cost reasoning on all candidate entities can be avoided. In the subsequent graph partitioning process, weak connection nodes can be removed by combining multiple verification results and node attribute information, which can improve the purity of the core-referenced entity clustering results. Therefore, while controlling the scope of complex computation calls, good global graph partitioning and entity clustering effects can still be maintained, balancing computational cost and judgment accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart of the dynamic evolution and sequential merging decision-making process of the knowledge graph of this invention; Figure 2 This is the fuzzy candidate pair bidirectional mutual verification and spectral partitioning diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] A computational method for knowledge graph evolutionary reasoning based on graph comparison and meta-learning includes the following steps: Step S1: Obtain entity nodes, relation types and corresponding triples in the knowledge graph to construct a heterogeneous graph. Construct a structural view by randomly discarding edge samples. Construct a semantic view by masking and perturbing the entity description text. Generate an entity feature vector by using contrastive loss to bring the node representations of the structural view and the semantic view closer together. Step S2: Extract the meta-feature vector of the entity pair to be discriminated. The meta-feature vector contains the logarithmic degree values ​​of the two entity nodes, the mean degree of the entity type, the standard deviation of degree, the degree quartiles, and the global graph density. Input the meta-feature vector into the meta-thresholding network to output the local discrimination threshold. Step S3: Calculate the structural semantic similarity of the entity pairs to be judged based on the entity feature vectors, and construct a basic similarity vector by combining the name similarity and topological overlap of the entity pairs to be judged. Combine the activation intensity vector of the Horn clause rules summarized from the historical judgment database, and calculate the merge confidence of the entity pairs to be judged as candidate pairs through the gating unit. Step S4: Input candidate pairs with a confidence level higher than the local discrimination threshold into the Markov decision process model. Combine the graph average clustering coefficient improvement, the validation set triple link prediction score increment, and deduct the decision rejection penalty to calculate the immediate reward. Use the policy network to arrange the merging order, and complete the topology state update of the local subgraph after confirming the merging.

[0020] Preferably, the generation of entity feature vectors in step S1 includes: step S11, aggregating neighbor node features under relation type in the heterogeneous graph using a relation-aware graph attention encoder, wherein the initial node features are generated by encoding entity names and description text by a language model and fusing entity type embeddings; step S12, constructing a structural view by discarding heterogeneous graph edges using a Bernoulli distribution mask matrix with a parameter of 0.2, and calculating a normalized temperature scale cross-entropy loss to bring the node representations of the two views closer together, and taking the mean of the output vectors of the two views as the final entity feature vector.

[0021] Preferably, step S2, which involves inputting the meta-feature vector into the meta-thresholding network to output a local discrimination threshold, includes: step S21, where the meta-feature vector also includes an entity type embedding vector and the ratio of edges to nodes in the entire graph; the meta-feature vector is input into a meta-thresholding network containing a three-layer multilayer perceptron structure, and a local discrimination threshold in the open interval of 0 to 1 is output via nonlinear mapping; the meta-thresholding network uses a first-order reptile algorithm to iteratively update parameters along the gradient direction that maximizes the entity matching evaluation index on the validation set data.

[0022] Preferably, step S4, which involves inputting candidate pairs with a confidence level higher than the local discrimination threshold into the Markov decision process model, includes: step S41, constructing the queue of candidate pairs to be processed in the graph, the topological state of the entire graph entities, and the global clustering features into a state space, and defining the confirmation of merging and skipping of candidate pairs as the action space; step S42, using a proximal policy optimization algorithm to optimize the policy network parameters, using the merging confidence level to guide the behavior cloning warm-up in the initial training stage, and advancing entity merging according to the output order of the optimized policy network.

[0023] Preferably, step S4 further includes a two-way mutual verification judgment step S43 before arranging the merging order. Step S431: Extract fuzzy candidate pairs with a merging confidence level between 0.4 and 0.75, and input the name, description, and shared relationship triple of the fuzzy candidate pairs into the external language model; Step S432: Use the forward inference chain to complete the same entity judgment, and use the reverse challenge to find contradictions between attributes and spatiotemporal logic; Step S433: Under a set number of sampling times, when the frequency of forward judgments that the same entity reaches a preset judgment threshold and the reverse challenge does not detect a substantial contradiction, the fuzzy candidate pair is judged to be a core-referenced entity, and the core-referenced logic pattern extracted during the inference process is added to the rule base.

[0024] Preferably, after step S43, a global graph partitioning and evidence propagation step S434 is included. Step S4341 involves constructing a weighted graph using verified core-inferred entities, with edge weights generated by weighting based on the frequency of affirmative judgments in the language model, combined confidence, and rule confidence. Step S4342 involves solving for the generalized eigenvalues ​​of the Laplacian matrix of the weighted graph to obtain eigenvectors, and recursively bisecting the graph structure according to the eigenvector signs. Step S4343 involves running a confidence propagation algorithm within the bisected subgraph, constructing a potential function based on node attribute purity and edge weights to update local node messages, and removing nodes with low marginal probabilities after convergence to output high-purity core-inferred clusters.

[0025] Preferably, after completing the topology state update, step S4 further includes a model continuous update step S44. Step S441 involves storing the merged and confirmed positive samples and strong negative samples with similarity higher than 0.85 but determined to be different entities into a scenario memory buffer of a set capacity. Step S442 involves calculating the diagonal elements of the parameter importance Fisher information matrix when the number of new samples in the scenario memory buffer reaches the update threshold, and introducing elastic weights to consolidate the weight shifts of the loss constraint graph attention encoder, gating unit, and meta-threshold network. The joint parameter update is completed by combining the historical samples replayed from the scenario memory buffer.

[0026] Preferably, in step S44, an online meta-learning update step S443 is adopted for the meta-threshold network. In step S443, when a new batch of data is accessed, the current loss gradient is calculated on the validation sample buffer, and the network weights of the meta-threshold network are directly adjusted in a single step to maintain the dynamic response capability of the meta-threshold network to changes in local topological density at each stage of graph evolution.

[0027] Preferably, after completing the topology state update in step S4, the process also includes a lightweight deployment and incremental inference step S45 on the edge. In step S451, a complete architecture including a graph attention encoder and a gating unit is used as the teacher network, and the output of the gating unit and the entity judgment result are used as soft labels to train the dual-tower structure student model. In step S452, when a new entity node is added to the knowledge graph, the first-hop neighborhood features of the new entity node are aggregated on the edge through one-step graph convolution, and the basic confidence is calculated using the dual-tower structure student model. Only the data with confidence in the judgment blind zone is uploaded to the computing cluster to complete the complete inference.

[0028] Example 1: In a distributed knowledge graph evolution computing system that continuously accesses massive amounts of heterogeneous data from multiple sources, the graph size expands exponentially. Due to the complexity of entity naming variants, the sparseness of long-tail relationships, and the high connectivity of core entities, traditional models relying on globally fixed similarity thresholds or isolated pairwise calculations suffer from local structural malmerging and topological breaks. This example is deployed on a knowledge graph computing cluster equipped with multi-node parallel heterogeneous accelerator cards. The node memory overhead and communication topology satisfy the constraints of large-scale tensor operations. The input includes a set of entity nodes. Relationship type set and triplet The original heterogeneous diagram Each entity node Includes name text Entity type code and text description sequence To eliminate uncertainty at the initial input, the name text is first... With description text After concatenation, the code is input into the pre-trained language model and the entity type codes are concatenated. Learnable embedding vectors The dimension is generated by mapping through the initial multilayer perceptron. Initial node feature vector ,in, This is a vector concatenation operator; based on this, the system constructs a message passing channel for heterogeneous graphs, where triples exist between entities. Establish with relation types The undirected edges are used, and semantic association edges are added to node pairs where the cosine similarity of term frequency and inverse document frequency between two entity description texts is higher than 0.9. Then, a relation-aware graph attention encoder is used. calculate Layer neighborhood feature aggregation, in the first layer In the feature transformation relationship of a layer, the attention coefficients are based on the learnable vectors specific to that relationship. With projection matrix The computation, feature update is represented as ,in, For entities In relationship Local neighbor set below, For multi-head attention splicing operators, It is a non-linear activation function. For the total number of attention heads, To ensure that the generated feature vectors are invariant to graph topological perturbations, the system constructs a structural view in parallel during the encoding computation process. With semantic view The structural view uses a Bernoulli distribution mask matrix with a parameter of 0.2 to randomly discard heterogeneous graph edges and performs random walk sampling of the subgraph. The semantic view uses a masked language model to perform word-level masking replacement on the entity description sequence and reconstruct the associated edges. The structural view is embedded by outputting it separately via the encoder. With semantic view embedding Normalized temperature-scaled cross-entropy loss is used to approximate the vector representation of the same entity under different views. The loss function is expressed as follows: ,in, For the current training batch set, For the preset scaling temperature coefficient, After the comparison loss converges, the arithmetic mean of the two view vectors is extracted as the final feature vector of the entity. When the computing cluster faces a situation where the local graph structure density is deviated, for the entity pairs to be judged... The system extracts data from concatenated type embedding vectors. Logarithmic degree of two nodes and The meta-feature vector consists of the mean degree, standard deviation of degree, first quartile, third quartile of degree, and the ratio of edges to nodes in the entire graph for nodes of the entity type. The ratio of edges to nodes in the global graph is obtained by the ratio of the number of edges to the number of nodes in the global graph; at the same time, the normalized connection density of the global graph is calculated. Meta-feature vectors are used to characterize the overall connectivity density of the current heterogeneous graph. The dimension is determined by the sum of the dimensions of the actual concatenation of the above components. This meta-feature vector is then input into a meta-thresholding network containing a three-layer feedforward multilayer perceptron architecture. The output of the Sigmoid activation function is mapped to a local discrimination threshold within the open interval of 0 to 1. The meta-thresholding network utilizes a first-order reptile algorithm during the parameter optimization phase to perform single-step parameter fine-tuning on the validation dataset along the gradient direction that maximizes the validation set matching evaluation metric. The update step size is expressed as follows: ,in, The meta-learning rate, For the validation set The comprehensive evaluation function adaptively tightens the discrimination threshold to suppress false merging when the entity node is in a high-connectivity, dense core region, while gradually lowering the threshold in a long-tail, sparse region to improve topology recall capability.

[0029] After obtaining the entity feature vectors and local discrimination thresholds, the system constructs a multi-source evidence fusion channel for the entity pairs to be discriminated, and calculates the name similarity by combining the Jaro-Winkler distance, character overlap rate, and the cosine value of the pre-trained language model vectors. The weighting coefficients satisfy the following conditions: Calculate structural semantic similarity based on graph comparison feature vectors And the topological overlap similarity by fusing neighborhood Jaccard coefficients and iterative SimRank scores. ,in, To pre-determine balanced weights, a basic similarity vector is generated. Simultaneously, Horn clause logic rules are mined from the historical decision database containing confirmed positive examples using differentiable inductive logic programming, and activation intensity vectors representing rule coverage and confidence are extracted from the Horn clause rules. The basic similarity vector and the activation intensity vector of the Horn clause rule are aligned to the same feature dimension via linear projection. , respectively obtained and The input gating unit calculates the adaptive selection weight vector. ,in, and The state representation is obtained by element-wise weighted fusion of the gated mapping parameter matrix and the bias term. The final output is the combined confidence score of the entity pairs to be judged. ,in, and The output layer weights and biases; if the combined confidence level is lower than the corresponding local discrimination threshold... If the current candidate pair is not selected, the merging process will be terminated directly, while candidate pairs that are above the local discrimination threshold will be used as valid candidate inputs for subsequent state transition steps.

[0030] To prevent the cascading accumulation of topological errors caused by isolated pairwise judgments, the system inputs candidate pairs exceeding the local discrimination threshold into the Markov decision process model, constructing the state space from the candidate queue to be processed, the set of topological vectors of entities in the entire graph, and global clustering features. The action space is defined as either confirming the merger and triggering a local subgraph collapse or skipping the candidate pair. The improvement in the average clustering coefficient of the graph after local subgraph folding Validation set triple link prediction score increment The immediate reward will be calculated after deducting the penalty for rejection. ,in, and The preset gain adjustment coefficient, The merge count is used to determine the number of merges rejected by subsequent checks; the policy network is optimized using a near-end policy optimization algorithm. Its objective loss function combines importance sampling ratio truncation, state value loss, and policy entropy regularization. In the early training phase, behavior clone parameters guided by merging confidence are loaded for warm-up. The policy network arranges the merging order of candidate pairs step-by-step according to the dynamic transition sequence of the current global topology state, and updates the local adjacency matrix and feature matrix in situ after each confirmed merge. In the specific operation of behavior clone warm-up, the system pre-maps merging confidence scores above 0.85 from the gating unit output as expert confirmation merge actions, and scores below 0.35 as expert skip actions, thereby automatically constructing a set of demonstration trajectories containing state and expert action pairs on the historical graph evolution slices. During the warm-up training phase, the policy network updates its parameters by minimizing the cross-entropy loss function between its output action probability distribution and the expert discrete actions, enabling the policy network to have basic topology action decision priors before formally interacting with the reinforcement learning environment, accelerating the convergence process of the policy gradient.

[0031] For boundary candidate pairs whose merge confidence falls within the fuzzy range of 0.4 to 0.75, the system triggers a two-way mutual verification judgment. The name, text description, and the first five shared relation triples of the candidate pair are constructed into logical prompts and input into the external language model. A forward inference chain is used to determine whether the two nodes point to the same entity, while a reverse challenge is applied to prompt the language model to check for exclusive contradictions at the temporal, spatial, and attribute logic levels, assuming the two are different. In three independent prompt samplings, if the frequency of forward judgments indicating the same entity reaches at least two times and the reverse challenge confirms the absence of substantial attribute contradictions, the candidate pair is deemed mutually verified, and the coreference logic pattern parsed during the mutual verification process is backfilled into the rule base. Subsequently, the system constructs a weighted undirected graph using the mutually verified candidate pairs, with edge weights calculated based on the language model's affirmative frequency, merge confidence, and rule support. ,in, To verify each other via indicator functions, The number of positive affirmative judgments. To maximize the rule confidence, , and To normalize the weight coefficients, solve for the Laplacian matrix of the weighted graph. Generalized eigenvalue problem ,in, For degree matrix, For the adjacency weight matrix, after extracting the Fiedler eigenvector corresponding to the second smallest eigenvalue, the graph structure is bisected according to the sign of the eigenvector. For subgraphs with more than 2 nodes after bisecting, the Laplacian matrix is ​​constructed in the same way, the Fiedler eigenvector corresponding to the second smallest eigenvalue is solved, and bisecting is performed according to its sign until the number of nodes in the subgraph to be partitioned is less than or equal to 2, completing the recursive bisecting. Then, the confidence propagation algorithm is run inside the bisected subgraph, and a potential function is constructed based on the purity of node attributes and edge weights to propagate edge messages. ,in, and For the binary state label of the node, Let be the boundary potential function. This represents the local prior potential function for the node; the purity of node attributes is determined based on the consistency of attribute values ​​among entities within the current subgraph, and is used to construct the local prior potential function for the node. Edge weights are used to construct the edge potential function. After message passing converges, nodes with low marginal probability are removed, and high-purity co-index clusters are output.

[0032] After completing entity folding and local topology state updates, the system initiates a continuous evolution and model parameter protection mechanism. Newly confirmed merged positive examples and strong negative examples with similarity higher than 0.85 but with different judgments are stored in a fixed-capacity context memory buffer. When the number of new samples reaches a preset threshold, the buffer is activated. At that time, the diagonal elements of the Fisher information matrix are used to calculate the parameter importance of each parameter of the relation-aware graph attention encoder, gating unit, and meta-threshold network corresponding to the historical task. Constructing elastic weights to consolidate losses Constraint parameter displacement, where, Losses due to the current mission For the consolidation coefficient, These are the parameters to be optimized. To solidify historical parameters, joint parameter updates are completed in conjunction with replaying historical samples. For the meta-threshold network, when a new batch of data is introduced, the current loss gradient is directly calculated in the validation sample buffer, and the network weights are individually adaptively fine-tuned to maintain its dynamic sensitivity to changes in local topological density. In addition, a complete system including a relation-aware graph attention encoder and a gating unit is used as the teacher network, and the gating fusion output and decision are used as soft labels to train a dual-tower structure student model. When an external system introduces new entity nodes... At the same time, the first-hop neighbor features are aggregated by one-step graph convolution on the edge side and the basic confidence is calculated using a dual-tower structure student model. Only data with confidence in the decision blind zone is uploaded to the computing cluster. In the end, the convergence of computing resource consumption is achieved while ensuring the stability of the whole graph topology.

[0033] Example 2: This example uses a distributed computing cluster with 8 computing nodes for comparative experiments. Each node is equipped with 64GB of memory and a hardware tensor accelerator. The test dataset uses multi-source heterogeneous graph benchmark data containing 120,000 entity nodes and 450,000 relation triples. 15% entity name variant perturbation sequences and 20% random edge missing noise are superimposed on the original input to form test conditions with both long-tail sparsity and representation degradation. The ground truth labels for the test set are generated using an expert double-blind cross-validation mechanism. 30,000 candidate entity pairs are manually sampled and labeled from the 120,000 nodes, with a ratio of 1:3 between true common positive examples and dissimilar negative examples. All evaluation indicators are statistically analyzed on independent test slices according to the discrete confusion matrix. Subgraphs with node connectivity degrees greater than or equal to 500 are classified into dense regions, and subgraphs with node degrees less than or equal to 2 are classified into sparse regions. Each test is run independently 5 times, and the arithmetic mean is taken as the final evaluation data.

[0034] In the test parameter configuration, the structure view The edge random drop rate is set to 0.2, and the temperature coefficient in the normalized temperature scale cross-entropy loss is... The threshold for updating the context memory buffer is set to 0.5, the bidirectional verification interval for fuzzy candidate pairs is set to 0.4 to 0.75, and the update threshold is set to 0.5. The test window for the random edge drop rate was set to 0.15 to 0.25, with a sample batch size of 500. When the drop rate was below 0.15, the overlap between the structural view and the original topology increased, and the influence of local high-frequency spurious connections on node representation was retained, resulting in a decrease in disambiguation performance. When the drop rate was above 0.25, the structural connectivity of long-tailed sparse nodes decreased, and the topological information in the local subgraph decreased as edges were lost. The experimental calibration used a random edge drop rate of 0.20.

[0035] The test system consisted of one test sample group containing all technical features of the invention, four control sample groups with partially missing features, three control sample groups with parameters exceeding the range, and two extreme perturbation test groups. All groups were run until convergence in the same computing cluster environment. Control sample group 1 maintained a structure-view edge drop rate of 0.20 and a bidirectional mutual verification interval of 0.40 to 0.75, replacing the meta-threshold network with a global static threshold of 0.65, and continuing to use a Markov decision process model in conjunction with a policy network to arrange the merging order. Control sample group 2 set the structure-view edge drop rate to 0.00, replaced the dual-view graph comparison learning with a single topological view encoding, and maintained a bidirectional mutual verification interval. The thresholds remain between 0.40 and 0.75, with the local discrimination thresholds still dynamically generated by the meta-threshold network, and the Markov decision process model is still used to arrange the merging order. Control group 3 uses a structure view edge drop rate of 0.20 and a two-way mutual verification range of 0.40 to 0.75, with the local discrimination thresholds dynamically generated by the meta-threshold network, but the Markov decision process model is removed and replaced with local greedy pair merging. Control group 4 uses a structure view edge drop rate of 0.20, with the local discrimination thresholds dynamically generated by the meta-threshold network, and the merging arrangement of the Markov decision process model is retained, but two-way mutual verification judgment and subsequent global graph partitioning and evidence propagation are not enabled.

[0036] Under the above conditions, the test sample group of this invention uses a structure view edge drop rate of 0.20, a bidirectional mutual verification interval of 0.40 to 0.75, a local discrimination threshold dynamically generated by the meta-threshold network, and a merging arrangement of the Markov decision process model, achieving a disambiguation F1 score of 94.6%, a false merging rate of 2.8% in dense regions, a false negative rate of 3.1% in sparse regions, an average increase in the overall graph clustering coefficient of +0.082, and a heavy computing power usage rate of 0.42%. In the control sample group 1, where only the local discrimination threshold is changed to a global static threshold of 0.65, the disambiguation F1 score is 82.3%, the false merging rate in dense regions is 14.7%, the false negative rate in sparse regions is 18.2%, the average increase in the overall graph clustering coefficient of +0.031, and the heavy computing power usage rate is 0.44%. Compared with this group, the disambiguation F1 score of the test sample group of this invention is improved by 12.3 percentage points. When the other main test conditions of the two groups remain the same, the meta-feature vector composed of node degree statistics and the ratio of edges to nodes in the entire graph is... Input Meta Threshold Network And generate local discrimination thresholds The corresponding false merging rate in dense regions and false negative rate in sparse regions changed from 14.7% and 18.2% to 2.8% and 3.1%, respectively.

[0037] In the comparative test of dual-view graph contrastive learning, the structure view edge drop rate of control group 2 was 0.00, the bidirectional mutual verification interval was 0.40 to 0.75, and the meta-threshold network and Markov decision process model were continued. This group achieved a disambiguation F1 score of 85.1%, a false merging rate of 8.9% in dense regions, a false negative rate of 11.5% in sparse regions, an average clustering coefficient increment of +0.045 across the entire graph, and a heavy computational power usage rate of 0.41%. In the comparative test of merging and arrangement methods... In the test, control group 3 maintained a structural view edge drop rate of 0.20 and a bidirectional mutual verification range of 0.40 to 0.75. Local discrimination thresholds were dynamically generated by the meta-threshold network. After employing local greedy pairwise merging, the disambiguation F1 score was 83.7%, the false merging rate in dense regions was 11.2%, the false negative rate in sparse regions was 9.4%, the average clustering coefficient increment for the entire image was -0.024, and the proportion of heavy computational power usage was 0.43%. The test group of this invention increased the average clustering coefficient of the image by [amount missing]. Integrating link prediction increments into instant rewards , by policy network When merging actions according to state changes, the average clustering coefficient increment for the entire graph is +0.082.

[0038] In the comparative test of two-way mutual verification judgment and global graph partitioning and evidence propagation, control group 4 maintained a structural view edge discard rate of 0.20. Local discrimination thresholds were dynamically generated by the meta-threshold network, and the merging order was arranged using a Markov decision process model. However, the two-way mutual verification interval was not enabled. The disambiguation F1 score of this group was 87.4%, the false merging rate in dense areas was 7.6%, the false detection rate in sparse areas was 8.8%, the average clustering coefficient increment of the whole graph was +0.053, and the proportion of heavy computing power calls was 0.00%. This group directly judged based on the merging confidence output by the gating unit, so no heavy computing power calls were generated corresponding to the two-way mutual verification link.

[0039] The parameter range test continued to use a meta-threshold network to dynamically generate local discrimination thresholds and a Markov decision process model to arrange the merging order. In the out-of-range control group 1, the structure view edge drop rate was reduced to 0.05, and the bidirectional mutual verification interval was maintained between 0.40 and 0.75, resulting in a disambiguation F1 score of 86.4%, a false merging rate in dense regions of 7.2%, a false negative rate in sparse regions of 9.8%, an increase in the average clustering coefficient of the entire graph of +0.049, and a heavy computational power call ratio of 0.42%. In the out-of-range control group 2, the structure view edge drop rate was increased to 0.45, and the bidirectional mutual verification interval was maintained between 0.40 and 0.75. The verification range remains 0.40 to 0.75, yielding a disambiguation F1 score of 83.1%, a false merging rate of 12.4% in dense regions, a false negative rate of 13.6% in sparse regions, an average clustering coefficient increment of +0.028 across the entire graph, and a heavy computing power usage ratio of 0.45%. Compared to the disambiguation F1 score of 94.6% when the edge random drop rate is 0.20, the disambiguation F1 scores corresponding to the two out-of-range values ​​of 0.05 and 0.45 are 86.4% and 83.1%, respectively, corresponding to the aforementioned edge random drop rate test window of 0.15 to 0.25.

[0040] The out-of-range control group 3 maintained a structure-view edge drop-off rate of 0.20, a meta-threshold network, and a Markov decision process model, only expanding the bidirectional mutual verification interval from 0.40-0.75 to 0.20-0.90. This group achieved a disambiguation F1 score of 94.9%, a false merging rate of 2.5% in dense regions, a false negative rate of 2.9% in sparse regions, an average increase in the overall clustering coefficient of +0.084, and a heavy computing power usage rate of 12.80%. The heavy computing power usage rate corresponding to the test sample group of this invention was 0.42%, therefore... The heavy computing power call volume increased by more than 30 times after the interval expansion. The boundary of the bidirectional mutual verification interval was determined by statistically analyzing the posterior error probability distribution of the merged confidence level output by the gating unit. In the historical verification set test, when the merged confidence level was less than 0.40, the posterior probability that the sample was a true core-referenced entity was less than 2%. When the merged confidence level was higher than 0.75, the posterior probability that the sample was a true core-referenced entity was more than 98%. After the candidate pairs in the interval of 0.40 to 0.75 entered the bidirectional mutual verification judgment, the overall judgment error rate converged to less than 3%.

[0041] Under extreme topological perturbation conditions, both test groups adopted a structure view edge drop rate of 0.20 and a bidirectional mutual verification interval of 0.40 to 0.75. Local discrimination thresholds were dynamically generated by the meta-threshold network, and the merging order was arranged using a Markov decision process model. In extreme perturbation test group 1, a 30% topological isomorphic perturbation was applied to the high-density dense subgraph region, resulting in a disambiguation F1 score of 92.8%, a false merging rate of 4.1% in dense areas, a false detection rate of 3.8% in sparse areas, an average clustering coefficient increment of +0.076 across the entire graph, and a heavy computing power usage rate of 0.58%. In extreme perturbation test group 2, a 40% edge loss perturbation was applied to the long-tail sparse region, resulting in a disambiguation F1 score of 91.5%, a false merging rate of 3.6% in dense areas, a false detection rate of 5.2% in sparse areas, an average clustering coefficient increment of +0.071 across the entire graph, and a heavy computing power usage rate of 0.63%.

[0042] Example 3: After continuous access to a distributed knowledge graph computing cluster from multiple heterogeneous sources, the scale and types of graph nodes expand over time, and the node connection degree exhibits a multi-scale unbalanced distribution. The system is configured with a hardware tensor acceleration environment containing multiple computing nodes, and the input includes a set of entity nodes. Relationship type set and relational triples The original heterogeneous graph formed Entity nodes include name text Entity type code and text description sequence When generating initial node features, and After concatenation, the input to the pre-trained language model extracts a 768-dimensional semantic vector, which is then projected onto a 128-dimensional space via a parameter matrix. This vector is then combined with a 64-dimensional entity type learnable embedding vector. The concatenation is input into the initial multilayer perceptron to generate an initial node feature vector with a dimension of 128. Relationship-aware graph attention encoder The network is configured with two layers and four attention heads, each with an output feature dimension of 32. In the [missing information - likely a specific layer or function], the network depth is set to 2 layers and the attention head is set to 4 attention heads. The [missing information - likely a specific function or function] is then set to [missing information - likely a specific function or function]. When aggregating layer features, the projection matrix is ​​based on the specific relationship. With attention vector Calculate attention coefficient The LeakyReLU activation function has a negative slope of 0.2. For entities In relationship Local neighbor set below, This is a vector concatenation operator; the concatenated multi-head outputs are then mapped using a nonlinear activation function to obtain the updated feature vector. In dual-view comparative learning, the structural view Edges in heterogeneous graphs are randomly discarded using a Bernoulli distribution mask matrix with a parameter of 0.2, and random walk subgraph sampling with a maximum path length of 5 is implemented with a restart probability of 0.15; semantic view A masked language model is used to perform word replacement on entity description sequences with a 15% probability, and semantic links with a text cosine similarity higher than 0.9 are reconstructed. The dual-view output vector is then used. and Temperature coefficient Minimize the normalized temperature-scale cross-entropy loss with a batch size of 256. After the loss function converges, the arithmetic mean of the two view vectors is taken as the final entity feature vector. .

[0043] For the entity pair to be determined When generating local discrimination thresholds, the meta-feature vector The input is configured to be 140-dimensional. The embedding vectors of the two entity types are concatenated to form a 128-dimensional feature. The two nodes provide... and For the node sets of each entity's respective type, the mean degree, standard deviation of degree, first quartile, and third quartile are calculated. The ratio of edges to nodes in the entire graph is obtained by taking the ratio of the number of edges to the number of nodes in the entire graph and is used as a meta-feature component. At the same time, the normalized connection density of the global graph is preserved. As another global statistical component, the above components together form a 140-dimensional feature vector. Meta-threshold network A three-layer feedforward multilayer perceptron is used. The first hidden layer maps the 140-dimensional input to 64 dimensions and applies the ReLU activation function. The second hidden layer maps to 32 dimensions and applies the ReLU activation function. The output layer is mapped to a scalar and the local discrimination threshold in the open interval of 0 to 1 is obtained through the Sigmoid function. Meta-threshold network parameters The algorithm is optimized using a first-order reptile model and applied to the validation dataset. The upper layer divides the data into multiple meta-task batches according to the density interval of the subgraph. The inner layer iterates on each batch using a fixed threshold for classification and updates temporary parameters within 5 local steps. The outer layer uses a meta-learning rate. Perform single-step parameter fine-tuning along the gradient direction of maximizing the validation set matching evaluation metric. When entity nodes are located in core high-density subgraph regions with a connection degree greater than 1000, the local discrimination threshold converges to 0.82 to 0.88; when entity nodes are located in long-tailed sparse regions with a connection degree less than or equal to 2, the local discrimination threshold is relaxed to 0.45 to 0.52. The outer iteration of the meta-learning uses continuous smoothing approximate score processing to match the evaluation index: for each sample pair in the validation set, the merged confidence is subtracted from the local discrimination threshold output by the meta-thresholding network, and the difference is mapped to a continuous soft decision probability between 0 and 1 by a logistic smoothing function with an amplification factor of 10. Then, the harmonic mean of continuous soft precision and soft recall is calculated based on the continuous soft decision probability and the true binary label. After the inner task generates local update weights, the outer layer directly performs first-order single-step vector extrapolation update along the gradient direction of the smoothing objective function with the outer layer learning rate.

[0044] In the process of multi-source similarity fusion and rule injection, the basic similarity vector is calculated for the entity pairs to be judged. The name similarity is ,in, , , The structural semantic similarity is Topological overlap similarity is ,in, The SimRank iteration decay coefficient is 0.8, which is achieved when the iteration residual is less than 0.8. When the system stops, it saves historical confirmed positive examples in the historical judgment database. Using differentiable inductive logic, it mines the top 30 Horn clause rules with a confidence level higher than 0.70 and a support frequency greater than or equal to 20, and forms an activation intensity vector for the Horn clause rules. The basic similarity vector and the activation intensity vector of the Horn clause rule are linearly projected onto a unified feature dimension. ,get and Gating unit calculation ,through After obtaining the comprehensive state representation, the combined confidence level is then output. In rule mining, entity type matching, attribute value equality, and multi-hop relationship paths are used as the atomic elements of the rule body, and the equivalent co-reference of two entities is used as the rule head. The rule support is calculated by multiplying the transition probabilities of entities in the historical judgment database. For each pair of entities to be judged, the preconditions of 30 Horn clauses are verified one by one. If the preconditions of the path relationship are met, the corresponding rule confidence is taken as the activation value of that dimension; otherwise, it is taken as 0. The combined confidence is lower than the local discrimination threshold. Candidate pairs that are not selected are discarded directly, while candidate pairs that are above the local discrimination threshold are entered into the subsequent Markov decision queue.

[0045] In the sequential merging decision-making process, the system encodes the candidate queue to be processed, the whole graph node embedding tensor, and the global clustering features into a state space. The confirmation of merging or skipping the current candidate pair will be encoded into a binary action space. When performing subgraph folding, if two merging nodes have parallel edges of the same relation type with the same external neighbor, they are merged into a single edge with a larger attention edge weight. If edges of different relation types exist, the corresponding multiple heterogeneous relation attributes are retained. The entity type embedding vector of the newly generated node is obtained by weighting and summing the type embedding vectors of the two original nodes according to their original degree ratios and normalizing the magnitude. After the policy network outputs confirmation of the merging action, the graph computation engine executes subgraph folding and generates new node feature vectors. And merge the local adjacency lists of the two nodes; based on the improvement in the average clustering coefficient of the folded graph. Validation set triple link prediction average reciprocal ranking increment And the calculation of immediate rewards for rejected penalty items. ,in, , The rejection penalty is determined using the historical cumulative rejection rate within a time sliding window: when the policy network executes the current merging action, it reads the percentage of actions confirmed as erroneous merging by bidirectional mutual verification within the most recent 50 steps before the current decision time; after subsequent bidirectional mutual verification produces the final rejection result for the current action, the penalty signal is converted and attenuated through the asynchronous reward tracing channel and added to the cumulative reward of the corresponding historical state trajectory. Training is performed using a near-end policy optimization algorithm, and policy parameters are pruned. State value loss weight Policy entropy regularization coefficient With a discount factor of 0.95, behavioral clone parameters guided by merged confidence were loaded for warm-up during the first 2000 iterations of training.

[0046] In the time-series scheduling, the sequential merge decision and batch global graph partitioning are divided into two stages. In the first stage, candidate pairs with a merge confidence higher than 0.75 are processed first. The nodes are merged and the local topology is updated one by one according to the action sequence output by the policy network. For candidate pairs with a merge confidence between 0.40 and 0.75, they are written into the mutual verification queue and the current merge action is skipped. In the second stage, when the mutual verification queue reaches 30 pairs or the current batch traversal ends, the batch bidirectional mutual verification judgment and global graph partitioning are performed. After the high-purity co-index clusters are formed, they are uniformly submitted to the main graph update engine.

[0047] For candidate entity pairs with a merge confidence level between 0.40 and 0.75, the system performs a two-way mutual verification judgment. The candidate pair name, text description, and the first five shared relation triples are constructed into a formatted prompt input to the external language model. Three forward inference chain judgments are performed at a sampling temperature of 0.2, and one reverse contradiction check is performed at a sampling temperature of 0.1. The two-way mutual verification judgment passes when at least two positive affirmative judgments are made and no attribute conflict is found in the reverse check. The forward inference prompt includes the entity name, type, text summary, and associated triples. The reverse challenge checks exclusive attributes such as birth year, geographical coordinates, and administrative division under the assumption that the two entities are different. The reverse contradiction parser parses the response text using a preset negative keyword dictionary and attribute conflict extraction template. Merging is rejected when a clear exclusive conflict is found. For candidate pairs that pass the two-way mutual verification judgment, the coreference logic pattern formed by the subject, relational predicate, and object is extracted from the forward inference chain text, converted into a first-order Horn clause format, and the corresponding support frequency is accumulated in the rule base.

[0048] A weighted undirected graph is constructed using candidate pairs determined through bidirectional mutual verification, with edge weights... ,in, , , Before partitioning the spectral graph, a breadth-first search is used to decompose the weighted undirected graph into connected components. Isolated nodes or independent connected components with fewer than or equal to 2 nodes are directly merged based on edge weights. For connected subgraphs with more than or equal to 3 nodes, a Laplacian matrix is ​​constructed for each subgraph. Sum-degree matrix Solve the generalized eigenvalue equations Extract the Fiedler vector corresponding to the second smallest eigenvalue, and divide the node set into two subsets according to the vector sign; for the subgraphs obtained by the division and whose number of nodes is still greater than 2, continue to perform connected component checks, Laplacian matrix construction and Fiedler vector sign division until the number of nodes in the current subgraph to be divided is less than or equal to 2, and complete the recursive binary division of the graph structure.

[0049] The confidence propagation algorithm is run within the subgraph after the binary search, and local messages at nodes are processed according to... Updated, the edge potential function is Local prior potential function of nodes The purity of node attributes is determined based on the consistency of attribute values ​​within the current subgraph. In constructing the prior potential, the proportion of non-empty attribute fields in entity nodes to the total preset number of attribute fields, and the normalized proportion of the number of descriptive text characters within a preset upper limit of 500 characters, are simultaneously calculated. These two factors are weighted by 0.6 and 0.4 respectively, forming a scalar between 0.1 and 0.9. This scalar, combined with the consistency of node attribute values, determines the corresponding prior potential. When the node state value is 1, this weighted scalar is used; when the node state value is 0, 1 minus this weighted scalar is used. Message update residuals are below... Stop iteration and remove edge probabilities. The low edge probability nodes output high-purity co-indexed clusters.

[0050] During continuous evolution and edge deployment, the context memory buffer The capacity is 10,000. Positive samples and strong negative samples with a similarity higher than 0.85 are maintained at 50% each through sampling from a reservoir; new samples are added when the update threshold is reached. Time-triggered parameter consolidation is performed by calculating the diagonal elements of the Fisher information matrix corresponding to the parameter importance of each parameter in the relation-aware graph attention encoder, gating unit, and meta-thresholding network based on the gradient of historical samples. And use elastic weights to consolidate losses. Constraint parameter displacement, where Simultaneously, to update weights in conjunction with historical sample replay, when calculating parameter importance, the relation-aware graph attention encoder and the gating unit calculate the diagonal elements of their respective parameter importance Fisher information matrices based on the expected squared gradient of the log-likelihood of the binary cross-entropy loss for entity classification. The meta-thresholding network uses continuously differentiable soft-threshold cross-entropy as the surrogate loss function, calculating the expected squared gradient of the surrogate loss relative to the network parameters on historical validation samples. When new batches of data are introduced, the meta-thresholding network calculates the current loss gradient on the validation buffer with a learning rate of 0.005 and performs single-step adaptive fine-tuning. Meanwhile, a complete model containing a relation-aware graph attention encoder and a gating unit is used as the teacher network, and the gating unit output generated by temperature coefficient 2 and the entity judgment results are used as soft labels to train a dual-tower structure student model containing a two-layer perceptron. New entity nodes are added when the external system is connected. At the same time, the first-hop neighborhood features are aggregated by one-step graph convolution on the edge, and the basic confidence is calculated by the dual-tower structure student model. Only the data with confidence between 0.35 and 0.80 in the decision blind zone are uploaded to the computing cluster, and the remaining data are directly disambiguated and split on the edge.

[0051] Example 4: This example combines Figures 1 to 2 This paper describes a computational method for knowledge graph evolutionary reasoning based on graph comparison and meta-learning, such as... Figure 1As shown, in step S1, entity nodes, relation types, and corresponding triples in the knowledge graph are obtained to construct a heterogeneous graph. A structural view is constructed by randomly discarding edges, and a semantic view is constructed by reconstructing semantic edges through masking perturbation of the entity description text. Contrastive loss is used to narrow down the node representations of the structural view and the semantic view to generate entity feature vectors. In step S2, meta-feature vectors of the entity pairs to be discriminated are extracted. These meta-feature vectors include the logarithmic degree values ​​of the two entity nodes, the mean degree of the entity type, the standard deviation of the degree, the degree quartiles, and the global graph density. The meta-feature vectors are input into a meta-thresholding network to output a local discrimination threshold. In step S3, based on… The entity feature vector is used to calculate the structural semantic similarity of the entity pairs to be judged, and the basic similarity vector is constructed by combining the name similarity and topological overlap of the entity pairs to be judged. The activation intensity vector of the Horn clause rule summarized from the historical judgment database is combined with the gating unit to calculate the merging confidence of the entity pairs to be judged as candidate pairs. In step S4, the candidate pairs with merging confidence higher than the local discrimination threshold are input into the Markov decision process model. The immediate reward is calculated by combining the graph average clustering coefficient improvement, the validation set triple link prediction score increment and deducting the decision rejection penalty. The merging order is arranged by the policy network, and the topological state of the local subgraph is updated after the merging is confirmed.

[0052] like Figure 2 As shown, fuzzy candidate pairs are extracted and input into an external language model. The same entity is determined using a forward inference chain. The contradiction between attributes and spatiotemporal logic is found through reverse challenge. After being determined as core-referential entities, a weighted graph is constructed. The generalized eigenvalues ​​of the Laplacian matrix are solved, and the graph structure is recursively bisected according to the eigenvector sign. The confidence propagation algorithm is run to update the local messages of the nodes, and finally, a high-purity core-referential cluster is output.

[0053] Example 5: In the initialization phase of deploying the graph comparison and meta-learning inference system to a newly added professional domain subgraph, considering the local edge distribution distortion and entity type embedding bias in the cold start region, the system first collects the initial topological fragment of the target subgraph containing entity nodes and relation triples. Through a sliding sampling window, it statistically analyzes the mean degree, standard deviation of degree, and the ratio of edges to nodes across the entire graph for each relation type. The statistical measures of each dimension are then biased and aligned with the baseline mean to complete the meta-feature vector. Numerical normalization calibration; subsequently, the relation-aware graph attention encoder is fixed during an unsupervised warm-up period. The feedforward projection dimension uses a single batch of 128 samples as validation slices, and a 3-step first-order reptile algorithm is implemented to update the meta-threshold network along the gradient direction that maximizes the matching evaluation index. Local discrimination threshold of output The algorithm converges stably to between 0.80 and 0.85 in the dense core region of the target subgraph and between 0.48 and 0.55 in the long-tailed edge region, thus correcting the discrimination criterion shift caused by the step change in topological density during cross-domain migration.

[0054] During the continuous evolution of the knowledge graph and the access of multi-source incremental nodes, the system initiates a reservoir probability permutation update when the context memory buffer reaches the preset capacity limit of 10,000. The diagonal elements of the parameter importance Fisher information matrix are updated based on the diagonal approximation of the second derivative of the loss from historically merged positive and strong negative examples. And in the elastic weight consolidation loss, the displacement penalty coefficient is included. The value is set to 400 to constrain the deviation of important parameters in the relation-aware graph attention encoder, gating unit, and meta-threshold network. The edge-side decision blind zone interval is jointly calibrated based on the approximation error of the single-hop graph convolution representation and the edge upload communication bandwidth. Since the edge model only aggregates single-hop neighborhood features, its output confidence has a consistency rate of over 99% with the decision of the complete teacher model of the computing cluster in the extremely high and extremely low confidence regions. However, the decision entropy corresponding to the single-hop approximation increases in the range of 0.35 to 0.80. After uploading the data in this interval to the computing cluster, the data traffic uploaded from the edge to the computing cluster is reduced by more than 85%, while retaining the diversion method of processing the decision blind zone data by the computing cluster.

[0055] The complete model containing a relation-aware graph attention encoder and a gating unit in the computing cluster is used as the teacher network, and the output of the gating unit and the entity judgment result are used as soft labels to train the dual-tower structure student model. On the edge, the graph convolution of the student model is restricted to single-hop neighborhood aggregation. When the calculated base confidence falls into the judgment blind zone of 0.35 to 0.80, the corresponding data is uploaded to the computing cluster. Data in other confidence ranges are directly judged and merged locally in the edge hardware register.

[0056] Example 6: During the engineering deployment and system-level parameter solidification of the ultra-large-scale knowledge graph evolutionary reasoning system, the system performs offline calibration and pre-verification on distributed nodes equipped with 8 hardware computing acceleration cards to control the relation-aware graph attention encoder. Meta-threshold network And the numerical fluctuations of the confidence propagation algorithm under extreme graph topological perturbations; firstly, orthogonal initialization is performed on the projection parameters of the heterogeneous graph, and the hidden layer feature dimension of the relation-aware graph attention encoder is set to 128, and the number of attention heads is set to... Attention vector With relation projection matrix The initial parameters follow a normal distribution, the edge drop rate in the structure view is fixed at 0.2, and the connection density is normalized in the global graph. Cross to On multiple sets of benchmark verification subgraphs, for meta-feature vectors The 140-dimensional input data was preprocessed using sliding window statistics for mean and variance normalization; subsequently, the meta-learning iteration of the first-order reptile algorithm was initiated, with the outer meta-learning rate set to [value missing]. The inner layer local optimization step size is set to 5 steps, and the local learning rate is set to 0.05. After training 500 batches of meta-tasks, the L2 norm of the gradient of the comprehensive evaluation function on the validation set is monitored; when the condition is met for 20 consecutive iterations... When the parameters converge and the weights of the meta-threshold network are solidified, and the degree of both nodes is greater than or equal to 500, the local discrimination threshold is determined. The local discrimination threshold converges to 0.82 to 0.88 in dense subgraph regions; when the degree of both nodes is less than or equal to 2, the local discrimination threshold is... It converges to 0.45 to 0.52 in the long-tailed sparse region.

[0057] In online evolutionary reasoning, global graph partitioning, and evidence propagation, numerical boundaries and anomaly degradation strategies are set for solving the generalized eigenvalues ​​of the Laplacian matrix and propagating confidence. The weighted graph Laplacian matrix adopts... Its generalized eigenvalue equation is The solution is obtained using the block-based Lanczos iterative method, with the convergence condition for the iterative residual set as follows: The maximum number of iterations is 100, and the condition number of the Laplace matrix exceeds [a certain threshold]. Degradation warnings are triggered on time and superimposed on the diagonal of the adjacency weight matrix. The regularized perturbation is then applied, and the eigenvalue calculation continues.

[0058] When running the confidence propagation algorithm inside a subgraph, the edge potential function The edge weight scaling factor is set to 1; the node local prior potential function Based on the purity of node attributes, the purity of node attributes is first determined by the consistency of entity attribute values ​​within the current subgraph. Then, it is numerically calibrated in conjunction with the entity description length, ensuring that the value of the local prior potential function of the node is within the range of 0.1 to 0.9, and that local messages... With the norm of the difference between two consecutive rounds of messages being less than As the convergence threshold, the maximum number of update rounds is set to 50. After the convergence condition is met, nodes with an edge probability lower than 0.5 are identified as low edge probability nodes and removed. If the convergence threshold is not reached within 50 rounds, the degradation protection is triggered, and the segmentation of the current subgraph is completed directly according to the threshold of a merge confidence score greater than or equal to 0.85.

[0059] During the incremental model update phase, the context memory buffer Maintain a capacity limit of 10,000 samples; when new samples reach the update threshold. At that time, the diagonal elements of the Fisher information matrix for parameter importance were calculated using 500 replay samples. In the elastic weight consolidation loss, the displacement penalty coefficient is set to 400, and the single-step pruning threshold of the parameter update gradient is set to 1 to limit excessive gradient changes during parameter updates. In the reservoir probability sampling process, sampling weights are maintained hierarchically according to the entity type to which the node belongs, so that the distribution ratio of historical samples of each category is consistent with the actual type distribution of the whole graph. Before triggering parameter consolidation updates, an effectiveness scan is performed on the adjacency subgraph referenced by historical samples in the context memory buffer. If the entity nodes in the historical samples have been folded and merged in the recent evolution, the neighborhood index and feature vector corresponding to the sample are updated synchronously, and the current parameter gradient is calculated using the updated replay sample.

[0060] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A computational method for knowledge graph evolutionary reasoning based on graph comparison and meta-learning, characterized in that, Includes the following steps: Step S1: Obtain entity nodes, relation types and corresponding triples in the knowledge graph to construct a heterogeneous graph. Construct a structural view by randomly discarding edge samples. Construct a semantic view by masking and perturbing the entity description text. Generate an entity feature vector by using contrastive loss to bring the node representations of the structural view and the semantic view closer together. Step S2: Extract the meta-feature vector of the entity pair to be discriminated. The meta-feature vector contains the logarithmic degree values ​​of the two entity nodes, the mean degree of the entity type, the standard deviation of degree, the degree quartiles, and the global graph density. Input the meta-feature vector into the meta-thresholding network to output the local discrimination threshold. Step S3: Calculate the structural semantic similarity of the entity pairs to be judged based on the entity feature vectors, and construct a basic similarity vector by combining the name similarity and topological overlap of the entity pairs to be judged. Combine the activation intensity vector of the Horn clause rules summarized from the historical judgment database, and calculate the merge confidence of the entity pairs to be judged as candidate pairs through the gating unit. Step S4: Input candidate pairs with a confidence level higher than the local discrimination threshold into the Markov decision process model. Combine the graph average clustering coefficient improvement, the validation set triple link prediction score increment, and deduct the decision rejection penalty to calculate the immediate reward. Use the policy network to arrange the merging order, and complete the topology state update of the local subgraph after confirming the merging.

2. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, The generation of entity feature vectors in step S1 includes: Step S11, aggregating neighbor node features under relation type in the heterogeneous graph using a relation-aware graph attention encoder, wherein the initial node features are generated by encoding entity names and description text by a language model and fusing entity type embeddings; Step S12, using a Bernoulli distribution mask matrix with a parameter of 0.2 to discard heterogeneous graph edges to construct a structural view, and calculating a normalized temperature scale cross-entropy loss to bring the node representations of the two views closer together, and taking the mean of the output vectors of the two views as the final entity feature vector.

3. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, Step S2, which inputs the meta-feature vector into the meta-thresholding network to output the local discrimination threshold, includes: Step S21, where the meta-feature vector also includes the entity type embedding vector and the ratio of edges to nodes in the entire graph; the meta-feature vector is input into the meta-thresholding network, which contains a three-layer multilayer perceptron structure, and outputs the local discrimination threshold in the open interval of 0 to 1 through nonlinear mapping; the meta-thresholding network uses a first-order reptile algorithm to perform parameter iterative updates on the validation set data along the gradient direction that maximizes the entity matching evaluation index.

4. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, Step S4 involves inputting candidate pairs with a merging confidence level higher than the local discrimination threshold into the Markov decision process model. This includes: Step S41, constructing the queue of candidate pairs to be processed in the graph, the topological state of the entire graph entities, and the global clustering features into a state space, and defining the confirmation of merging and skipping of candidate pairs as the action space; Step S42, optimizing the policy network parameters using a near-end policy optimization algorithm, using the merging confidence level to guide behavior cloning warm-up in the initial training stage, and advancing entity merging according to the output order of the optimized policy network.

5. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, Step S4, before arranging the merging order, also includes a two-way mutual verification judgment step S43. Step S431: Extract fuzzy candidate pairs with a merging confidence level between 0.4 and 0.75, and input the name, description, and shared relationship triple of the fuzzy candidate pairs into the external language model; Step S432: Use the forward inference chain to complete the same entity judgment, and use the reverse challenge to find contradictions between attributes and spatiotemporal logic; Step S433: Under a set number of sampling times, when the frequency of forward judgments as the same entity reaches a preset judgment threshold and the reverse challenge does not detect a substantial contradiction, the fuzzy candidate pair is judged as a core-referenced entity, and the core-referenced logic pattern extracted during the inference process is added to the rule base.

6. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 5, characterized in that, Step S43 is followed by global graph partitioning and evidence propagation steps S434. Step S4341 involves constructing a weighted graph using verified core-inferred entities, with edge weights generated by weighting based on the frequency of affirmative judgments in the language model, combined confidence, and rule confidence. Step S4342 involves solving for the generalized eigenvalues ​​of the Laplacian matrix of the weighted graph to obtain eigenvectors, and recursively bisecting the graph structure according to the eigenvector signs. Step S4343 involves running a confidence propagation algorithm within the bisected subgraph, constructing a potential function based on node attribute purity and edge weights to update local node messages, and removing nodes with low marginal probabilities after convergence to output high-purity core-inferred clusters.

7. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, Step S4, after completing the topology state update, also includes the model continuous update step S44. Step S441, merged and confirmed positive samples and strong negative samples with similarity higher than 0.85 but determined to be different entities are stored in a scenario memory buffer of a set capacity. Step S442, when the number of new samples in the scenario memory buffer reaches the update threshold, the diagonal elements of the parameter importance Fisher information matrix are calculated, and the weight shifts of the loss constraint graph attention encoder, gating unit and meta-threshold network are introduced to consolidate the weight shifts. The joint parameter update is completed by combining the historical samples replayed from the scenario memory buffer.

8. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 7, characterized in that, In step S44, online meta-learning is used to update the meta-threshold network in step S443. In step S443, when a new batch of data is accessed, the current loss gradient is calculated on the validation sample buffer, and the network weights of the meta-threshold network are directly adjusted in one step to maintain the dynamic response capability of the meta-threshold network to changes in local topological density at each stage of graph evolution.

9. The knowledge graph evolutionary reasoning computation method based on graph comparison and meta-learning according to claim 1, characterized in that, After completing the topology state update in step S4, the process also includes lightweight deployment and incremental inference steps S45. In step S451, a dual-tower structure student model is trained using a complete architecture including a graph attention encoder and a gating unit as the teacher network, and the gating unit output and entity judgment results as soft labels. In step S452, when a new entity node is added to the knowledge graph, the first-hop neighborhood features of the new entity node are aggregated on the edge through one-step graph convolution. The basic confidence is calculated using the dual-tower structure student model, and only data with confidence in the judgment blind zone is uploaded to the computing cluster to complete the complete inference.

Images