A logistics cost intelligent settlement method based on a knowledge graph

Abstract

This invention discloses a smart settlement method for logistics costs based on knowledge graphs, belonging to the field of knowledge graph technology, and includes the following steps: S1, collecting logistics order and rate data, and outputting an indexed association graph; S2, outputting standard entity mapping relationships; S3, constructing a multi-dimensional constraint solution model; S4, outputting a preliminary cost solution; S5, generating statistical feature vectors; S6, calculating node reconstruction errors by performing random masking of features and edges through an improved GraphMAE model, and using a self-supervised mechanism based on neighborhood aggregation, removing abnormal nodes and outputting a cleaned settlement statement; S7, calculating the final settlement price and generating a payment instruction. This invention effectively overcomes the limitations of traditional logistics settlement, such as fuzzy rule matching, poor robustness of anomaly detection, and a single pricing strategy, achieving a technical effect that balances the accuracy of cost calculation, high reliability of data cleaning, and optimal business strategy returns in complex logistics scenarios.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a smart settlement method for logistics costs based on knowledge graphs. Background Technology

[0002] With the extension of the logistics service chain and the diversification of billing scenarios, traditional rule engine systems face severe matching and computing power challenges in real-time billing and anomaly monitoring of massive high-dimensional business data. Existing billing methods based on keyword matching or simple database queries, while capable of handling standardized cost calculations, primarily rely on pre-defined hard-coded logic and static field comparisons. This purely rule-based matching approach ignores the complex topological relationships (such as multi-level transit paths and multi-dimensional rate dependencies) and deep spatiotemporal semantic features implicit in logistics order data. This makes it difficult to automatically identify applicable billing strategies when facing non-standard products or complex changes, thus limiting the accuracy and coverage of cost calculations. Furthermore, classic settlement methods often struggle to effectively quantify the uncertainties and risks in the operational process, and lack robust detection mechanisms for graph structure anomalies during the settlement order cleaning stage. Consequently, when processing settlement requests containing noisy or abnormal operational parameters, the output often deviates from the expected cost results, increasing the enterprise's settlement risk control costs and auditing difficulties.

[0003] Therefore, how to provide a smart settlement method for logistics costs based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention discloses a knowledge graph-based intelligent settlement method for logistics costs. Its core lies in constructing a full-link intelligent settlement mechanism for logistics costs based on topological game theory awareness. By constructing an improved GraphMAE model, it focuses on using a topology-aware mask layer to parse the node topology, combining degree centrality indices with a PageRank algorithm with damping coefficients to calculate node importance distribution, and adjusting Bernoulli sampling parameters accordingly to generate a binary mask matrix. Logical index truncation is then performed on the adjacency matrix and feature matrix of the heterogeneous graph. Simultaneously, a channel semantic decoupling layer and an anomaly measurement layer based on Laplace regularization constraints are introduced to calculate reconstruction errors and quantify anomaly probabilities, thereby outputting a cleaned settlement statement. Based on this, a revenue constraint model is constructed based on the cleaned settlement statement amount and budget reference value. A game theory-based inverse induction algorithm is used to calculate the Nash equilibrium strategy and solve for the final settlement price. This invention effectively overcomes the limitations of traditional logistics settlement methods, such as fuzzy rule matching, poor robustness of anomaly detection, and a single pricing strategy. It achieves a technical effect that balances the accuracy of cost calculation, high reliability of data cleaning, and optimal business strategy returns in complex logistics scenarios.

[0005] According to an embodiment of the present invention, a smart settlement method for logistics costs based on knowledge graphs specifically includes: S1. Collect logistics order and rate data, construct the spatiotemporal-attribute index structure of the billing knowledge graph, establish the binding relationship between index nodes and rule nodes through feature vector mapping, and output the index association graph; S2. Perform neighborhood retrieval based on the index association graph to obtain a set of candidate rules, calculate the semantic matching degree between the feature vector of the candidate rule set and the goods description text, and output the standard entity mapping relationship based on probability estimation. S3. Traverse the billing knowledge graph based on the standard entity mapping relationship, extract the billing constraint rule set, and construct a multi-dimensional constraint solution model that includes decision variables, logical constraint terms, and objective function. S4. Substitute the actual operation parameters into the multidimensional constraint solution model, search for the feasible solution domain that satisfies the logical constraints, calculate the extreme value of the objective function in the feasible solution domain, and output the preliminary cost solution. S5. Based on the preliminary cost solution and actual operation parameters, calculate the joint probability distribution of variables, extract the distribution parameters through iterative optimization, and generate a statistical feature vector containing expectation, variance and confidence interval. S6. Input the statistical feature vector into the improved GraphMAE model to perform random masking of features and edges, calculate the node reconstruction error based on the self-supervised mechanism of neighborhood aggregation to quantify the anomaly probability, remove abnormal nodes and output the cleaning settlement statement. S7. Construct a revenue constraint model based on the amount of the cleaning settlement statement and the budget reference value, use the game theory reverse inducement algorithm to calculate the Nash equilibrium strategy, solve the final settlement price based on the equilibrium strategy and generate payment instructions.

[0006] Optionally, S1 specifically includes: S11. Extract the geographic coordinates and timestamps from the logistics orders, use the geohashing algorithm to map the coordinates to the discrete grid index, and at the same time parse the time constraints of the rate data. Calculate the spatiotemporal correlation based on the intersection of time windows and the overlap of spatial grids, and construct the adjacency matrix of the spatiotemporal-attribute index. S12. Construct the topological adjacency structure of index nodes based on the adjacency matrix. Use the TransE algorithm to perform low-dimensional embedding initialization of nodes to obtain initial node representation vectors. Construct learnable diagonal matrix parameters, map the original feature vectors to the bilinear space for feature interaction transformation, and introduce nonlinear constraints using the ReLU activation function to project the transformed vectors to the low-dimensional manifold subspace. Calculate the cosine similarity matrix between the feature vectors and node embedding vectors in the manifold space, and perform probability normalization on the similarity matrix according to the maximum a posteriori probability criterion. Solve for the node index corresponding to the maximum a posteriori probability to establish the mapping relationship between feature vectors and index nodes. S13. Parse the billing rule predicates corresponding to the index nodes, extract the logical dependencies in the predicates, take the mapping relationship between the feature vectors and the index nodes as input, construct a bipartite graph constraint model from the index nodes to the rule nodes, verify the connectivity between nodes through the constraint satisfaction algorithm, complete the binding and output the index association graph.

[0007] Optionally, S2 specifically includes: S21. Based on the index association graph, perform multi-hop neighborhood retrieval with the index node as the center, aggregate the rule features of the retrieved neighbor nodes, and generate a candidate rule set vector containing topological information. S22. Perform word segmentation and word vectorization operations on the goods description text, and use the self-attention mechanism to extract the deep semantic features of the text and generate text feature vectors. S23. Perform tensor outer product and broadcast operation on the candidate rule set vector and the text feature vector to construct a multi-dimensional feature interaction tensor to represent element-level semantic association; perform normalized dot product operation on the interaction tensor dimension to generate a global similarity matrix; sort the similarity matrix values ​​in descending order and index them according to the preset Top-K truncation strategy, extract the feature subset corresponding to the high score of the matrix, and output the target rule subset. S24. Based on the feature vector distribution in the target rule subset, calculate the posterior probability of entity mapping using maximum a posteriori probability estimation, and output the standard entity mapping relationship with the highest probability.

[0008] Optionally, S3 specifically includes: S31. Perform a directed graph traversal in the billing knowledge graph based on the standard entity mapping relationship to extract the rule logic of the nodes on the path. S32. Based on rule logic, the business logic predicates are transformed into Boolean algebra expressions, which are then mapped to a set of mathematical constraints containing inequalities and equality. S33. Analyze the set of mathematical constraints, define the vector space dimension of the decision variables according to the topological relationship of the parameter index, and determine the boundary conditions by combining the intercept parameters of the inequality constraints. Construct a polyhedral feasible region enclosed by the hyperplane. Map the rate weights and penalty factors to numerical coefficients in the linear programming equation system, establish the algebraic mapping relationship between the objective function scalar and the constraint matrix vector, and construct a multidimensional optimization solution model containing the objective function and constraint conditions.

[0009] Optionally, S4 specifically includes: S41. Map the actual operation parameters to the decision variables of the multidimensional constraint solution model to generate an initial state tensor containing the constraint matrix and the coefficients of the objective function. S42. Construct a Lagrange augmented function based on the initial state tensor, perform a first-order gradient operation on the decision variables with respect to the objective function, and perform iterative updates along the negative half-axis of the gradient. During the iteration process, extract the coefficients of the linear inequalities in the logical constraint terms to construct a constraint matrix. Calculate the projection operator that projects the variables onto the feasible region of the convex set based on the constraint matrix. Use the projection operator to perform orthogonal mapping correction on the updated variable vector to generate an iterative solution sequence that dynamically approximates the extreme points. S43. Based on the iterative solution sequence, calculate the norm distance between adjacent iteration points. When the norm distance is less than the preset convergence threshold or the dual gap approaches zero, determine that the convergence condition is met and extract the current extreme point. The analytical output is the preliminary cost solution.

[0010] Optionally, S5 specifically includes: S51. Construct a joint probability density function with the initial cost solution as the prior mean and the actual operation parameters as covariates, and calculate the partial derivative matrix and log-likelihood gradient of the joint probability density function with respect to the distribution parameters. S52. Construct a Newton-Raphson iterative scheme based on the log-likelihood gradient, calculate the inner product of the inverse of the Hessian matrix and the gradient vector to determine the parameter update direction and step size; perform eigenvalue decomposition of the Hessian matrix during the iteration process to verify positive definiteness, and calculate the L2 norm of the gradient vector until all eigenvalues ​​are positive and the L2 norm is less than the preset convergence threshold, and lock the target joint probability distribution. S53. Based on the analytical expression of the joint probability distribution of the target, the expected value and the second central moment are obtained through integration, and the quantiles of the inverse cumulative distribution function at the preset significance level are calculated to generate a statistical feature vector containing the expected value, variance and confidence interval.

[0011] Optionally, the improved GraphMAE model includes a channel semantic decoupling layer, a topology-aware masking layer, a hierarchical heterogeneous aggregation encoder, a feature reconstruction decoder, and an anomaly measurement layer: The channel semantic decoupling layer is used to project the input statistical feature vector to a high-dimensional latent space, construct a multi-head self-attention mechanism to capture long-range dependencies between feature channels, generate a channel attention weight matrix and perform weighted operations on the original features, and output the semantic decoupling node feature tensor. The topology-aware masking layer is used to parse the topological structure of the semantically decoupled node feature tensor, calculate the degree centrality index of the nodes, and iteratively solve the node importance distribution using the PageRank algorithm with damping coefficient. Based on the node importance distribution, the probability parameters of Bernoulli sampling are adjusted to generate a binary mask matrix. Logical index truncation operations are performed on the adjacency matrix and feature matrix of the heterogeneous graph through the binary mask matrix to construct a masked subgraph view containing only visible nodes, and the corresponding set of masked node indices is extracted. The hierarchical heterogeneous aggregation encoder is used to define the neighborhood sampling window based on the meta-path, receive the masked subgraph view, and use the multi-head attention mechanism to calculate the attention coefficient between visible nodes and heterogeneous neighbors. Through weighted aggregation operation, the neighbor features are mapped into low-dimensional node embedding vectors that fuse high-order structural information. The feature reconstruction decoder is used to parse the index set of the masked nodes, locate and extract the corresponding hidden layer representation from the low-dimensional node embedding vector, perform linear transformation and nonlinear activation operations through a multi-layer fully connected network, and output the predicted feature vector of the masked nodes. The anomaly measurement layer is used to receive the predicted feature vector and the original statistical feature vector, calculate the Euclidean distance to construct the reconstruction loss function, introduce a Laplacian regularization term based on node density to constrain and optimize the loss function, solve and output the node anomaly score.

[0012] Optionally, S7 specifically includes: S71. Construct a revenue function that includes the cleaning settlement amount and the budget reference value, set the budget upper limit and price non-negativity as inequality constraints, and establish a revenue constraint optimization model. S72. Based on the profit-constrained optimization model, calculate the opponent's best response under the current strategy, update and solve the strategy sequence through the fixed-point iteration method until convergence, and lock the Nash equilibrium strategy. S73. Based on the Nash equilibrium strategy, analyze the objective function, calculate the price parameters in the equilibrium state as the final settlement price, and generate payment instructions.

[0013] The beneficial effects of this invention are: This invention effectively solves the challenge of identifying hidden anomalies in logistics cost settlement data by employing an improved GraphMAE model, significantly improving the robustness and accuracy of settlement data cleaning. The invention innovatively introduces a topology-aware masking layer, combining degree centrality indices with a PageRank algorithm with damped coefficients to calculate node importance distribution, and dynamically adjusts the masking probability of Bernoulli sampling based on this distribution. This mechanism can selectively preserve the structural information of high-weight nodes, and by parsing the topological structure of node feature tensors through semantic decoupling, the model can capture more deeply sensitive local neighborhood variations during self-supervised reconstruction. Simultaneously, a channel semantic decoupling layer is introduced to capture long-range dependencies between feature channels, and a Laplace regularization term based on node density is used to constrain and optimize the loss function, effectively enhancing the model's ability to perceive hidden anomaly patterns under complex logistics business models, thereby significantly reducing the false alarm rate and ensuring the purity of settlement data.

[0014] The core of this invention lies in constructing an intelligent settlement mechanism for the entire logistics cost chain based on topological game theory perception. By introducing a game theory-based inverse induction algorithm, it effectively overcomes the limitations of single pricing strategies and the lack of consideration for benefit equilibrium. This invention constructs a revenue constraint model based on the cleaned settlement amount and budget reference value, and uses a game theory-based inverse induction algorithm to calculate the Nash equilibrium strategy, thus solving for the final settlement price. This mechanism transforms the settlement process into a game process of seeking the optimal strategy, deeply exploring the dynamic balance relationship between settlement costs, operational parameters, and budget revenue. It can adaptively find the pricing strategy that maximizes revenue while satisfying logical constraints. The innovation of this mechanism lies in its creative construction of a cross-level mapping from "bottom-level topological structure perception" to "upper-level game strategy decision-making." It directly uses the topological anomaly detection results based on improved GraphMAE as the input constraint of the game model, ensuring that the pricing strategy is based on high-quality data, thereby achieving a deep integration of physical topological constraints and business game decision-making. This not only represents a leap from simple calculation to intelligent decision-making, but also effectively avoids operational risks caused by price fluctuations or rigid strategies, providing scientific and technical support for refined operation and maximizing business benefits in complex logistics scenarios. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a knowledge graph-based intelligent settlement method for logistics costs proposed in this invention. Figure 2 This is a flowchart illustrating the working principle of the improved GraphMAE model for a knowledge graph-based intelligent settlement method for logistics costs proposed in this invention. Detailed Implementation

[0016] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 and Figure 2 A knowledge graph-based intelligent settlement method for logistics costs, specifically including: S1. Collect logistics order and rate data, construct the spatiotemporal-attribute index structure of the billing knowledge graph, establish the binding relationship between index nodes and rule nodes through feature vector mapping, and output the index association graph; S2. Perform neighborhood retrieval based on the index association graph to obtain a set of candidate rules, calculate the semantic matching degree between the feature vector of the candidate rule set and the goods description text, and output the standard entity mapping relationship based on probability estimation. S3. Traverse the billing knowledge graph based on the standard entity mapping relationship, extract the billing constraint rule set, and construct a multi-dimensional constraint solution model that includes decision variables, logical constraint terms, and objective function. S4. Substitute the actual operation parameters into the multidimensional constraint solution model, search for the feasible solution domain that satisfies the logical constraints, calculate the extreme value of the objective function in the feasible solution domain, and output the preliminary cost solution. S5. Based on the preliminary cost solution and actual operation parameters, calculate the joint probability distribution of variables, extract the distribution parameters through iterative optimization, and generate a statistical feature vector containing expectation, variance and confidence interval. S6. Input the statistical feature vector into the improved GraphMAE model to perform random masking of features and edges, calculate the node reconstruction error based on the self-supervised mechanism of neighborhood aggregation to quantify the anomaly probability, remove abnormal nodes and output the cleaning settlement statement. S7. Construct a revenue constraint model based on the amount of the cleaning settlement statement and the budget reference value, use the game theory reverse inducement algorithm to calculate the Nash equilibrium strategy, solve the final settlement price based on the equilibrium strategy and generate payment instructions.

[0018] In this embodiment, S1 specifically includes: S11. Extract the geographic location coordinates and timestamps from logistics orders, map the coordinates to discrete grid indices using the Geohash algorithm, and simultaneously parse the timeliness constraints of rate data. Calculate the spatiotemporal correlation based on the intersection of time windows and spatial grid overlap, and construct an adjacency matrix for the spatiotemporal-attribute index. Specifically, the construction process includes extracting the latitude and longitude coordinate sequences from the order data, using a 6-bit precision Geohash encoding function to convert continuous geographic coordinates into discrete string grid indices, and parsing the effective and ineffective timestamps in the rate rules. Calculate the time overlap ratio based on the order timestamps and rate timeliness windows, and perform a spatial set intersection operation on the order grid index and the rate service area grid index. Use the time overlap ratio and spatial overlap as weight parameters to construct an initial adjacency matrix describing the spatiotemporal correlation strength between nodes. S12. Construct a topological adjacency structure for index nodes based on the adjacency matrix. Use the TransE algorithm to initialize low-dimensional embeddings of nodes to obtain initial node representation vectors. Construct learnable diagonal matrix parameters, map the original feature vectors to a bilinear space for feature interaction transformation, and introduce nonlinear constraints using the ReLU activation function to project the transformed vectors onto a low-dimensional manifold subspace. Calculate the cosine similarity matrix between the feature vectors and node embedding vectors in the manifold space, and perform probability normalization on the similarity matrix according to the maximum a posteriori probability criterion. Solve for the node index corresponding to the maximum a posteriori probability, and establish the mapping relationship between feature vectors and index nodes. The mapping process specifically includes initializing a learnable diagonal matrix with a dimension of 128. The weight matrix is ​​used to perform element-wise multiplication of the original high-dimensional feature vectors with the weight matrix and map them to a bilinear interaction space. The ReLU activation function is used to perform a nonlinear transformation on the interaction feature vectors, filtering negative features and introducing sparsity constraints. The transformed vectors are then mapped to a 64-dimensional manifold subspace through a linear projection layer, while the embedding vectors generated by the TransE algorithm are used as anchors in the manifold space. The dot product of the manifold feature vectors and the anchor embedding vectors is calculated to obtain the cosine similarity matrix. The Softmax function is applied to the similarity matrix to perform row normalization to obtain the posterior probability distribution. The index position corresponding to the probability maxima is locked by solving the Argmax function, thus completing the precise binding of feature vectors and graph nodes. S13. Parse the billing rule predicates corresponding to the index nodes, extract the logical dependencies in the predicates, take the mapping relationship between feature vectors and index nodes as input, construct a bipartite graph constraint model from index nodes to rule nodes, verify the connectivity between nodes through constraint satisfaction algorithm, complete the binding and output the index association graph; the binding process specifically includes parsing the logical predicate expressions in the billing rule text, converting logical AND, logical OR and logical NOT relations into edge constraints of the bipartite graph; using index nodes as the left node set and rule nodes as the right node set, construct potential connection edges between the two sets according to the feature vector mapping results; execute a backtracking algorithm to traverse the bipartite graph structure, verify whether the attribute features of each index node satisfy the predicate constraints of its connection to the rule nodes, perform pruning operations on edges that do not satisfy logical connectivity, retain valid connection edges that satisfy the constraints, thereby generating an index association graph containing a complete logical topology.

[0019] In this embodiment, S2 specifically includes: S21. Based on the index association graph, perform multi-hop neighborhood retrieval with the index node as the center, aggregate the rule features of the retrieved neighbor nodes, and generate a candidate rule set vector containing topological information; the retrieval process specifically includes performing a breadth-first search with the index node as the starting center and a range of 3 hops, traversing the adjacency list structure of the index association graph to collect all neighbor nodes within the multi-hop path; performing an average pooling aggregation operation on the feature vectors of the collected neighbor nodes, and fusing the scattered node features into a candidate rule set vector containing local topological structure information; S22. Perform word segmentation and word vectorization operations on the goods description text, and extract deep semantic features of the text using a self-attention mechanism to generate a text feature vector. The generation process specifically includes using a BERT word segmenter to segment the goods description text, converting the text into a token sequence containing 50 words; mapping each token to a 768-dimensional vector representation through a pre-trained word embedding layer to construct a text input matrix; calculating the attention weights between words in the input matrix using a self-attention mechanism to capture long-distance semantic dependencies; and performing a weighted summation operation on the word vectors based on the attention weights to output a text feature vector that integrates contextual semantic information. S23. Perform tensor outer product and broadcast operations on the candidate rule set vector and the text feature vector to construct a multi-dimensional feature interaction tensor to represent element-level semantic association; perform normalized dot product operation on the interaction tensor dimension to generate a global similarity matrix; sort the similarity matrix values ​​in descending order and index them according to a preset Top-K truncation strategy, extract the feature subsets corresponding to the high scores of the matrix, and output the target rule subset; the construction process specifically includes performing tensor outer product operation on the candidate rule set vector and the text feature vector to generate an interaction tensor; performing L2 normalization operation along the feature dimension of the interaction tensor to calculate the normalized dot product similarity between the rule vector and the text vector to construct a global similarity matrix; according to a preset Top-20 truncation strategy, select the top 20 elements with the largest values ​​in the global similarity matrix, sort them in descending order of value, and extract the original feature vectors corresponding to these 20 high scores, aggregating them into a target rule subset; S24. Based on the feature vector distribution in the target rule subset, the posterior probability of entity mapping is calculated using maximum posterior probability estimation, and the standard entity mapping relationship with the highest probability is output. The calculation process specifically includes statistically analyzing the mean and variance of the feature vectors in the target rule subset, constructing a Gaussian distribution as a class conditional probability density model; combining the preset prior probability, the posterior probability value of the goods description text belonging to each standard entity category is calculated using Bayes' formula; comparing the posterior probability values ​​of all categories, the standard entity category with the highest probability value is selected as the final output result, and the mapping relationship between the goods description and the standard entity is established.

[0020] In this embodiment, S3 specifically includes: S31. Perform a directed graph traversal in the billing knowledge graph based on the standard entity mapping relationship to extract the rule logic of the nodes on the path; the traversal process specifically includes using the node identifier in the standard entity mapping relationship as the starting query anchor point, recursively searching along the directed edges in the billing knowledge graph using a depth-first search algorithm; recording the complete traversal path from the starting node to the leaf node, extracting the rule attribute text carried by all nodes on the path; linearly concatenating the extracted rule attribute text according to the path hierarchy to generate a rule logic sequence containing complete business logic semantics; S32. Based on rule logic, the business logic predicates are transformed into Boolean algebra expressions, and further mapped into a set of mathematical constraints containing inequalities and equality. The mapping process specifically includes using a semantic parser to identify logical connectors and comparison operators in the rule logic sequence; converting the greater than, less than, and equal to comparison operators into algebraic relation symbols, and converting logical AND and logical OR relations into intersection and union relations of constraint conditions; constructing Boolean algebra expressions based on algebraic relation symbols, and mapping the logical truth values ​​in Boolean algebra to the numerical values ​​0 and 1, thereby transforming the business rules into a set of mathematical constraints composed of linear inequalities and linear equality. S33. Analyze the mathematical constraint set, define the vector space dimension of the decision variables based on the topological relationship of the parameter index, and determine the boundary conditions by combining the intercept parameters of the inequality constraints, constructing a polyhedral feasible region enclosed by a hyperplane; map the rate weights and penalty factors to numerical coefficients in the linear programming equation system, establish an algebraic mapping relationship between the objective function scalar and the constraint matrix vector, and construct a multidimensional optimization solution model containing the objective function and constraint conditions; the construction process specifically includes counting the number of independent parameters in the mathematical constraint set, setting the vector space dimension to the total number of parameters; extracting the intercept values ​​from the inequality constraints, using the intercept values ​​as threshold parameters for the boundary conditions, and enclosing a polyhedral feasible region in the spatial coordinate system; reading the rate weight values ​​from the rate configuration table and the penalty factor values ​​from the business rules, filling the weight values ​​into the coefficient positions of the objective function, and filling the penalty factor values ​​into the off-diagonal positions of the constraint matrix; establishing an algebraic mapping relationship between the decision variables, constraint matrix, and objective function scalar through matrix multiplication operations, generating a multidimensional optimization solution model.

[0021] In this embodiment, S4 specifically includes: S41. Map the actual operation parameters to the decision variables of the multidimensional constraint solution model to generate an initial state tensor containing the constraint matrix and objective function coefficients. The mapping process specifically includes reading the weight, volume, and distance values ​​from the actual operation data, filling these values ​​into the decision variable vector according to the predefined parameter index order, and generating the initial variable tensor. At the same time, extract the constraint coefficient matrix and objective function weight vector from the multidimensional constraint solution model configuration file, and align and concatenate the initial variable tensor with the coefficient matrix and weight vector in the tensor space to construct an initial state tensor containing complete model parameters. S42. Construct a Lagrange augmented function based on the initial state tensor, perform a first-order gradient operation on the decision variables with respect to the objective function, and perform iterative updates along the negative half-axis of the gradient. During the iteration process, extract the linear inequality coefficients from the logical constraint terms to construct a constraint matrix, calculate the projection operator that projects the variables onto the feasible region of the convex set based on the constraint matrix, and use the projection operator to perform orthogonal mapping correction on the updated variable vector to generate an iterative solution sequence that dynamically approximates the extreme point. The specific operation process includes constructing a Lagrange augmented function containing equality constraints and inequality constraint penalty terms, calculating the partial derivative vector of the function with respect to the decision variables to obtain the first-order gradient, setting an update step size of 0.01 in the opposite direction of the gradient, and performing iterative update operations on the decision variable vector; extracting the linear inequality coefficient matrix from the logical constraint terms, using the Euclidean projection algorithm to calculate the projection operator that maps the updated variables back to the feasible region; applying the projection operator to perform orthogonal mapping correction on the variable vector, forcibly satisfying the boundary conditions, and generating an iterative solution sequence that gradually approximates the optimal solution. S43. Based on the iterative solution sequence, calculate the norm distance between adjacent iteration points. When the norm distance is less than a preset convergence threshold or the dual gap tends to zero, determine that the convergence condition is met and extract the current extreme point. Analyze the output as a preliminary cost solution. The determination process specifically includes extracting the current iteration point vector and the previous iteration point vector from the iterative solution sequence, calculating the Euclidean distance between the two vectors as the norm distance, and simultaneously calculating the difference between the original problem objective value and the dual problem objective value as the dual gap. Determine whether the norm distance is less than a preset convergence threshold of 0.0001, or whether the dual gap tends to 0. When either condition is met, stop the iteration, extract the value corresponding to the current iteration point as the extreme point, substitute the extreme point into the cost calculation formula, and analyze and output the preliminary cost solution.

[0022] In this embodiment, S5 specifically includes: S51. Construct a joint probability density function with the initial cost solution as the prior mean and the actual operation parameters as covariates, and calculate the partial derivative matrix and log-likelihood gradient of the joint probability density function with respect to the distribution parameters; the construction process specifically includes setting the initial cost solution values ​​as the mean parameters of a Gaussian distribution, taking the actual operation parameters as input covariate matrices, constructing a multivariate joint probability density function describing the correlation of variables; calculating the logarithmic expression of the probability density function, performing partial derivative operations on the mean vector and covariance matrix vector respectively, generating a partial derivative matrix containing the first derivative, and defining the numerical vector in the partial derivative matrix as the log-likelihood gradient; S52. Construct a Newton-Raphson iterative scheme based on the log-likelihood gradient, calculate the inner product of the inverse of the Hessian matrix and the gradient vector to determine the parameter update direction and step size; perform eigenvalue decomposition of the Hessian matrix during the iteration process to verify positive definiteness, and simultaneously calculate the L2 norm of the gradient vector until all eigenvalues ​​are positive and the L2 norm is less than a preset convergence threshold, locking the target joint probability distribution; the iterative process specifically includes calculating the second-order partial derivative of the log-likelihood function to obtain the Hessian matrix, and using a matrix inversion algorithm to obtain the inverse of the Hessian matrix; performing an inner product operation between the inverse matrix and the log-likelihood gradient vector, and using the result value as the step size and direction of parameter updates; performing eigenvalue decomposition on the Hessian matrix, checking whether all eigenvalues ​​are positive to confirm the convexity of the function; simultaneously calculating the L2 norm value of the gradient vector, and stopping the iteration and locking the current distribution parameters as the target joint probability distribution when the L2 norm value is less than the preset convergence threshold of 0.001 and all eigenvalues ​​are positive; S53. Based on the analytical expression of the joint probability distribution of the target, the expected value and second central moment are obtained through integration, and the quantiles of the inverse cumulative distribution function at a preset significance level are calculated to generate a statistical feature vector containing the expected value, variance, and confidence interval. The generation process specifically includes directly extracting the expected feature based on the mean parameter of the joint probability distribution of the target; calculating the diagonal element values ​​of the covariance matrix as the variance feature, i.e., the second central moment; selecting a significance level of 0.05, substituting it into the inverse cumulative distribution function to calculate the function value, and obtaining the quantile values; combining the expected value, variance value, and the boundary values ​​of the 95% confidence interval determined by the quantiles to generate a statistical feature vector containing statistical features.

[0023] In this embodiment, the improved GraphMAE model includes a channel semantic decoupling layer, a topology-aware masking layer, a hierarchical heterogeneous aggregation encoder, a feature reconstruction decoder, and an anomaly measurement layer: The channel semantic decoupling layer projects the input statistical feature vector to a high-dimensional latent space, constructs a multi-head self-attention mechanism to capture long-range dependencies between feature channels, generates a channel attention weight matrix, performs weighted operations on the original features, and outputs a semantic decoupling node feature tensor. Specifically, the operation of the channel semantic decoupling layer includes mapping the input vector to a 512-dimensional high-dimensional latent space through a linear transformation layer, constructing an 8-head multi-head self-attention mechanism to calculate the mutual influence weights between feature channels, normalizing the calculated weight matrix to channel attention coefficients, and performing element-wise multiplication weighted operations with the original input features to highlight key semantic feature channels, suppress background noise, and output a semantic decoupling node feature tensor. The topology-aware masking layer is used to parse the topological structure of the semantically decoupled node feature tensor, calculate the degree centrality index of the nodes, and iteratively solve the node importance distribution using the PageRank algorithm with damping coefficient. Based on the node importance distribution, the probability parameters of Bernoulli sampling are adjusted to generate a binary mask matrix. The binary mask matrix is ​​used to perform logical index truncation operations on the adjacency matrix and feature matrix of the heterogeneous graph, respectively, to construct a masked subgraph view containing only visible nodes, and to extract the corresponding set of masked node indices. The specific operations of the topology-aware masking layer include parsing the graph structure connection relationship corresponding to the tensor, counting the number of neighbors of each node as the degree centrality index, iteratively calculating the importance score of each node using the PageRank algorithm with a damping coefficient of 0.85, normalizing the importance score to the probability distribution parameters of Bernoulli sampling, performing random sampling to generate a binary mask matrix composed of 0 and 1, and using the binary mask matrix to perform index truncation on the feature matrix and adjacency matrix, retaining the nodes with a value of 1 to construct the masked subgraph view, and recording the node indices with a value of 0 as the set of masked node indices. The hierarchical heterogeneous aggregation encoder is used to define a neighborhood sampling window based on the metapath, receive a masked subgraph view, and calculate the attention coefficients between visible nodes and heterogeneous neighbors using a multi-head attention mechanism. It then maps neighbor features into low-dimensional node embedding vectors that incorporate high-order structural information through weighted aggregation operations. Specifically, the operation of the hierarchical heterogeneous aggregation encoder includes defining the hop count window for neighborhood sampling as 2 hops according to a preset metapath pattern; calculating the attention coefficients between visible nodes and heterogeneous neighbor nodes using a 4-head multi-head attention mechanism in the masked subgraph view; performing a weighted summation aggregation operation on the features of neighbor nodes based on the attention coefficients; mapping the aggregated feature vectors to a 128-dimensional low-dimensional space; and outputting a low-dimensional node embedding vector that incorporates high-order topological information. The feature reconstruction decoder is used to parse the index set of the masked nodes, locate and extract the corresponding hidden layer representation from the low-dimensional node embedding vector, perform linear transformation and nonlinear activation operations through a multi-layer fully connected network, and output the predicted feature vector of the masked node. The specific operations of the feature reconstruction decoder include: retrieving the corresponding hidden layer representation from the low-dimensional node embedding vector according to the index set of the masked nodes; constructing a fully connected neural network with 3 hidden layers, each containing 256 neurons; performing linear transformation matrix operations and ReLU nonlinear activation function operations on the hidden layer representation in sequence to restore the dimension to the original feature dimension, and outputting the predicted feature vector of the masked node. The anomaly measurement layer receives the predicted feature vector and the original statistical feature vector, calculates the Euclidean distance to construct a reconstruction loss function, introduces a Laplacian regularization term based on node density to constrain and optimize the loss function, solves and outputs the node anomaly score; the specific operations of the anomaly measurement layer include calculating the Euclidean distance between the predicted feature vector and the original statistical feature vector as the reconstruction error; calculating the local density of the node in the original image as the weight coefficient of the Laplacian regularization term; constructing a loss function containing the reconstruction error and the Laplacian regularization term, where the regularization term coefficient is set to 0.1; solving the loss function and outputting anomaly scores that quantify the degree of node anomaly.

[0024] In this embodiment, S7 specifically includes: To make the steps clearer, more specific, and more actionable, it is recommended to specify the game participants, the details of strategy space construction, and the specific process for generating payment instructions. The following are the rewritten versions of S71-S73: S71. Construct a revenue function that includes the settlement amount and budget reference value, set the budget ceiling and price non-negativity as inequality constraints, and establish a revenue constraint optimization model; the establishment process specifically includes defining the self-revenue function as the difference between the target settlement price and the cost benchmark, and introducing a budget deviation penalty term with a weight coefficient of 0.6; setting the numerical limit that the settlement amount must not exceed the preset budget ceiling as the first inequality constraint, and setting the restriction that the price value must be strictly greater than or equal to 0 as the second inequality constraint; using the self-revenue function as the objective function to be optimized in the model, and in conjunction with the two inequality constraints, construct a self-revenue constraint optimization model describing the cash flow constraints; S72. Based on the payoff constraint optimization model, calculate the opponent's optimal response under the current strategy, update and solve the strategy sequence through a fixed-point iteration method until convergence, and lock the Nash equilibrium strategy. The calculation process specifically includes inputting the self-payoff constraint optimization model, initializing the opponent's strategy set and corresponding payoff model parameters (setting its budget deviation penalty weight coefficient to 0.4 to simulate competitive behavior), simulating the opponent's competitive behavior under the current self-bid, calculating the first derivative condition that maximizes the opponent's own payoff to obtain the optimal response function; using the fixed-point iteration algorithm, using the opponent's optimal response output as the input parameter for the next iteration of the self-strategy, constructing a strategy update sequence; calculating the Euclidean norm distance between adjacent iteration strategy vectors, and determining that the algorithm has converged when the distance value is less than a preset precision threshold of 0.001; extracting the self-bid portion from the converged strategy vector and locking it as the Nash equilibrium strategy. S73. Based on the Nash equilibrium strategy, analyze the objective function, calculate the price parameters under equilibrium state as the final settlement price, and generate a payment instruction. The analysis process specifically includes reading the target price value in the Nash equilibrium strategy vector, substituting it into the revenue constraint optimization model for verification calculation, and confirming that the value maximizes the objective function while satisfying two inequality constraints. After verification, the target price value is encapsulated as the final settlement price field, combined with the order's unique identifier, the payee's account code, and the current system timestamp, and assembled into a payment instruction data packet conforming to the bank payment gateway interface standard to complete the settlement process.

[0025] Example 1: To verify the feasibility of this invention in intelligent logistics settlement, the method was applied to the intelligent freight settlement and risk control system of a large third-party logistics supply chain company (hereinafter referred to as "Company L"). In traditional logistics settlement systems, manual verification of waybills and rate tables or calculations based on simple rule engines are commonly used. These methods not only struggle to handle massive amounts of unstructured cargo descriptions and complex multi-dimensional billing constraints, but also fail to effectively identify hidden anomalies and game-theoretic fraud risks in settlement data, easily leading to errors in cost calculation, fund loss, or settlement disputes. To address these problems, Company L decided to adopt the intelligent logistics cost settlement method based on knowledge graphs proposed in this invention.

[0026] During implementation, Company L first collected historical logistics orders, dynamic rate tables, and budget control data to construct a spatiotemporal attribute index structure for the billing knowledge graph. Through feature vector mapping technology, a binding relationship was established between index nodes and rule nodes, outputting an index association graph, laying the data foundation for subsequent accurate rule matching. Simultaneously, Company L's business experts precisely labeled the cleaned historical data with billing constraints and set budget reference values, serving as the benchmark for model solving and game theory decision-making.

[0027] Company L utilizes an indexed association graph to perform neighborhood retrieval, obtaining a set of candidate rules. By calculating the semantic matching degree between the cargo description text and rule features, it outputs a standard entity mapping relationship. Based on this mapping relationship, the system traverses the billing knowledge graph, extracting logical constraint rules containing multi-dimensional elements such as weight, volume, and transportation distance. It then constructs a multi-dimensional constraint solution model including decision variables, logical constraint terms, and an objective function. Next, the actual operational parameters are substituted into the model, and the feasible solution domain is searched and extreme values ​​are calculated using the Lagrange augmenting function and projection operator algorithm, outputting a preliminary cost solution.

[0028] In the core data risk control and game theory decision-making stage, this invention, based on the preliminary cost solution and actual operational parameters, utilizes Newton-Raphson iterative optimization to extract joint probability distribution parameters, generating a statistical feature vector containing expectation, variance, and confidence interval. This statistical feature vector is input into an improved GraphMAE model. Through channel semantic decoupling and topology-aware masking mechanisms, node reconstruction errors are calculated to quantify anomaly probabilities, successfully eliminating abnormal nodes and outputting a cleansing settlement statement. Based on the cleansing settlement statement amount and budget reference value, a revenue-constrained optimization model is established. Using game theory-based reverse-induction algorithms and fixed-point iterative calculation of the Nash equilibrium strategy, the final settlement price is calculated, and payment instructions are generated, achieving a closed-loop transition from rule calculation to game theory decision-making.

[0029] During implementation, Company L's technical team discovered that, compared to traditional manual settlement and conventional rule engine methods, the method of this invention significantly improves the accuracy, compliance, and risk resistance of logistics settlement. Traditional methods cannot handle complex probability distribution characteristics and lack means to identify hidden data anomalies and strategic fraud. In contrast, the method of this invention effectively achieves accurate cost calculation, automatic cleaning of abnormal data, and game-theoretic optimization of settlement strategies through multi-dimensional constraint solving, graph neural network anomaly detection, and Nash equilibrium game pricing.

[0030] To further verify the actual performance of the method of the present invention, Company L conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1 Performance Comparison of Company L's Intelligent Logistics Settlement Method

[0031] As shown in Table 1, the performance of the intelligent logistics settlement system has been comprehensively improved after applying the method of this invention. The accuracy of cost calculation has increased from 88.5% with the traditional method to 99.2%, and the abnormal settlement order identification rate has significantly increased from 65.0% to 96.5%, significantly enhancing the system's risk control level. The time spent on settlement data cleaning has decreased from 5.5 hours / 10,000 orders to 0.8 hours / 10,000 orders, and the settlement dispute handling cycle has been shortened from 4.5 days to 1.2 days, significantly improving the system's timeliness. In addition, the settlement fund risk exposure has decreased from RMB 250,000 / month to RMB 35,000 / month, and the labor cost for freight settlement has decreased from RMB 1.2 million / year to RMB 450,000 / year, significantly reducing operating costs. Customer reconciliation satisfaction has also significantly improved, from 82.0% to 97.5%.

[0032] Through the method of this invention, Company L has successfully achieved accurate calculation and intelligent risk control of logistics costs, effectively avoiding settlement anomaly risks and strategic fraud, ensuring fund security, significantly improving the automation and intelligence level of logistics settlement business, significantly reducing the workload of financial personnel, enhancing the stability and robustness of the settlement system, and providing strong technical support for the construction of a smart logistics supply chain.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent settlement of logistics costs based on knowledge graphs, characterized in that, Includes the following steps: S1. Collect logistics order and rate data, construct the spatiotemporal-attribute index structure of the billing knowledge graph, establish the binding relationship between index nodes and rule nodes through feature vector mapping, and output the index association graph; S2. Perform neighborhood retrieval based on the index association graph to obtain a set of candidate rules, calculate the semantic matching degree between the feature vector of the candidate rule set and the goods description text, and output the standard entity mapping relationship based on probability estimation. S3. Traverse the billing knowledge graph based on the standard entity mapping relationship, extract the billing constraint rule set, and construct a multi-dimensional constraint solution model that includes decision variables, logical constraint terms, and objective function. S4. Substitute the actual operation parameters into the multidimensional constraint solution model, search for the feasible solution domain that satisfies the logical constraints, calculate the extreme value of the objective function in the feasible solution domain, and output the preliminary cost solution. S5. Based on the preliminary cost solution and actual operation parameters, calculate the joint probability distribution of variables, extract the distribution parameters through iterative optimization, and generate a statistical feature vector containing expectation, variance and confidence interval. S6. Input the statistical feature vector into the improved GraphMAE model to perform random masking of features and edges, calculate the node reconstruction error based on the self-supervised mechanism of neighborhood aggregation to quantify the anomaly probability, remove abnormal nodes and output the cleaning settlement statement. S7. Construct a revenue constraint model based on the amount of the cleaning settlement statement and the budget reference value, use the game theory reverse inducement algorithm to calculate the Nash equilibrium strategy, solve the final settlement price based on the equilibrium strategy and generate payment instructions.

2. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, S1 specifically includes: S11. Extract the geographic coordinates and timestamps from the logistics orders, use the geohashing algorithm to map the coordinates to the discrete grid index, and at the same time parse the time constraints of the rate data. Calculate the spatiotemporal correlation based on the intersection of time windows and the overlap of spatial grids, and construct the adjacency matrix of the spatiotemporal-attribute index. S12. Construct the topological adjacency structure of index nodes based on the adjacency matrix. Use the TransE algorithm to perform low-dimensional embedding initialization of nodes to obtain initial node representation vectors. Construct learnable diagonal matrix parameters, map the original feature vectors to the bilinear space for feature interaction transformation, and introduce nonlinear constraints using the ReLU activation function to project the transformed vectors to the low-dimensional manifold subspace. Calculate the cosine similarity matrix between the feature vectors and node embedding vectors in the manifold space, and perform probability normalization on the similarity matrix according to the maximum a posteriori probability criterion. Solve for the node index corresponding to the maximum a posteriori probability to establish the mapping relationship between feature vectors and index nodes. S13. Parse the billing rule predicates corresponding to the index nodes, extract the logical dependencies in the predicates, take the mapping relationship between the feature vectors and the index nodes as input, construct a bipartite graph constraint model from the index nodes to the rule nodes, verify the connectivity between nodes through the constraint satisfaction algorithm, complete the binding and output the index association graph.

3. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, S2 specifically includes: S21. Based on the index association graph, perform multi-hop neighborhood retrieval with the index node as the center, aggregate the rule features of the retrieved neighbor nodes, and generate a candidate rule set vector containing topological information. S22. Perform word segmentation and word vectorization operations on the goods description text, and use the self-attention mechanism to extract the deep semantic features of the text and generate text feature vectors. S23. Perform tensor outer product and broadcast operation on the candidate rule set vector and the text feature vector to construct a multi-dimensional feature interaction tensor to represent element-level semantic association; perform normalized dot product operation on the interaction tensor dimension to generate a global similarity matrix; sort the similarity matrix values ​​in descending order and index them according to the preset Top-K truncation strategy, extract the feature subset corresponding to the high score of the matrix, and output the target rule subset. S24. Based on the feature vector distribution in the target rule subset, calculate the posterior probability of entity mapping using maximum a posteriori probability estimation, and output the standard entity mapping relationship with the highest probability.

4. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, S3 specifically includes: S31. Perform a directed graph traversal in the billing knowledge graph based on the standard entity mapping relationship to extract the rule logic of the nodes on the path. S32. Based on rule logic, the business logic predicates are transformed into Boolean algebra expressions, which are then mapped to a set of mathematical constraints containing inequalities and equality. S33. Analyze the set of mathematical constraints, define the vector space dimension of the decision variables according to the topological relationship of the parameter index, and determine the boundary conditions by combining the intercept parameters of the inequality constraints. Construct a polyhedral feasible region enclosed by the hyperplane. Map the rate weights and penalty factors to numerical coefficients in the linear programming equation system, establish the algebraic mapping relationship between the objective function scalar and the constraint matrix vector, and construct a multidimensional optimization solution model containing the objective function and constraint conditions.

5. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, S4 specifically includes: S41. Map the actual operation parameters to the decision variables of the multidimensional constraint solution model to generate an initial state tensor containing the constraint matrix and the coefficients of the objective function. S42. Construct a Lagrange augmented function based on the initial state tensor, perform a first-order gradient operation on the decision variables with respect to the objective function, and perform iterative updates along the negative half-axis of the gradient. During the iteration process, extract the coefficients of the linear inequalities in the logical constraint terms to construct a constraint matrix. Calculate the projection operator that projects the variables onto the feasible region of the convex set based on the constraint matrix. Use the projection operator to perform orthogonal mapping correction on the updated variable vector to generate an iterative solution sequence that dynamically approximates the extreme points. S43. Based on the iterative solution sequence, calculate the norm distance between adjacent iteration points. When the norm distance is less than the preset convergence threshold or the dual gap approaches zero, determine that the convergence condition is met and extract the current extreme point. The analytical output is the preliminary cost solution.

6. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, S5 specifically includes: S51. Construct a joint probability density function with the initial cost solution as the prior mean and the actual operation parameters as covariates, and calculate the partial derivative matrix and log-likelihood gradient of the joint probability density function with respect to the distribution parameters. S52. Construct a Newton-Raphson iterative scheme based on the log-likelihood gradient, calculate the inner product of the inverse of the Hessian matrix and the gradient vector to determine the parameter update direction and step size; perform eigenvalue decomposition of the Hessian matrix during the iteration process to verify positive definiteness, and calculate the L2 norm of the gradient vector until all eigenvalues ​​are positive and the L2 norm is less than the preset convergence threshold, and lock the target joint probability distribution. S53. Based on the analytical expression of the joint probability distribution of the target, the expected value and the second central moment are obtained through integration, and the quantiles of the inverse cumulative distribution function at the preset significance level are calculated to generate a statistical feature vector containing the expected value, variance and confidence interval.

7. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, The improved GraphMAE model includes a channel semantic decoupling layer, a topology-aware masking layer, a hierarchical heterogeneous aggregation encoder, a feature reconstruction decoder, and an anomaly measurement layer: The channel semantic decoupling layer is used to project the input statistical feature vector to a high-dimensional latent space, construct a multi-head self-attention mechanism to capture long-range dependencies between feature channels, generate a channel attention weight matrix and perform weighted operations on the original features, and output the semantic decoupling node feature tensor. The topology-aware masking layer is used to parse the topological structure of the semantically decoupled node feature tensor, calculate the degree centrality index of the nodes, and iteratively solve the node importance distribution using the PageRank algorithm with damping coefficient. Based on the node importance distribution, the probability parameters of Bernoulli sampling are adjusted to generate a binary mask matrix. Logical index truncation operations are performed on the adjacency matrix and feature matrix of the heterogeneous graph through the binary mask matrix to construct a masked subgraph view containing only visible nodes, and the corresponding set of masked node indices is extracted. The hierarchical heterogeneous aggregation encoder is used to define the neighborhood sampling window based on the meta-path, receive the masked subgraph view, and use the multi-head attention mechanism to calculate the attention coefficient between visible nodes and heterogeneous neighbors. Through weighted aggregation operation, the neighbor features are mapped into low-dimensional node embedding vectors that fuse high-order structural information. The feature reconstruction decoder is used to parse the index set of the masked nodes, locate and extract the corresponding hidden layer representation from the low-dimensional node embedding vector, perform linear transformation and nonlinear activation operations through a multi-layer fully connected network, and output the predicted feature vector of the masked nodes. The anomaly measurement layer is used to receive the predicted feature vector and the original statistical feature vector, calculate the Euclidean distance to construct the reconstruction loss function, introduce a Laplacian regularization term based on node density to constrain and optimize the loss function, solve and output the node anomaly score.

8. The intelligent settlement method for logistics costs based on knowledge graphs according to claim 1, characterized in that, Specifically, S7 includes: S71. Construct a revenue function that includes the cleaning settlement amount and the budget reference value, set the budget upper limit and price non-negativity as inequality constraints, and establish a revenue constraint optimization model. S72. Based on the profit-constrained optimization model, calculate the opponent's best response under the current strategy, update and solve the strategy sequence through the fixed-point iteration method until convergence, and lock the Nash equilibrium strategy. S73. Based on the Nash equilibrium strategy, analyze the objective function, calculate the price parameters in the equilibrium state as the final settlement price, and generate payment instructions.

Images