AI person and post matching intelligent recommendation method based on graph neural network

Abstract

The invention discloses an AI person and post matching intelligent recommendation method based on a graph neural network, and relates to the technical field of data processing, and the method comprises the following steps: S1, generating a multi-view heterogeneous graph; s2, outputting a community membership matrix and a community-level statistical feature vector; s3, generating an initial feature vector; s4, constructing an improved HetGNN model for non-cold start nodes, and outputting a fusion feature vector; s5, outputting a decoupling feature vector; s6, constructing a reciprocity preference matching mechanism, and generating a corrected matching score; and S7, outputting a final man-post matching result. According to the method, the limitations of single feature representation, poor cold start recommendation effect and insufficient privacy protection in a traditional man-post matching method are overcome, and an efficient, safe and credible solution is provided for intelligent configuration and decision support of human resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an AI person-post matching intelligent recommendation method based on a graph neural network. BACKGROUND

[0002] With the continuous improvement of the intelligent matching demand of human resource allocation, the traditional person-post precise recommendation and intelligent matching method is facing more and more challenges. In modern recruitment management, the precise matching of the skill portrait of job seekers and the demand of posts is not only an important means to improve the efficiency of recruitment and reduce the cost of enterprise recruitment, but also a key link to realize the person-post adaptation of job seekers and optimize the allocation of labor market resources. However, at present, the person-post matching method mostly relies on keyword retrieval, simple rule matching or recommendation algorithm based on collaborative filtering, the process is mechanical and the matching precision is low. Although these traditional methods can meet the basic information docking requirements, due to the lack of intelligent analysis ability of the deep semantic features and complex interaction relationship of users, it is difficult to adapt to the precise matching demand of multi-source heterogeneous data in the recruitment scene, dynamic changes of post demand and personalized recommendation.

[0003] The main limitation of the traditional person-post matching method is the single feature representation and the significant cold start effect. The existing method usually depends on text keywords, static attribute labels or simple click browsing records to calculate the matching degree, and the semantic understanding is shallow and it is difficult to capture the potential adaptation relationship between job seekers and posts. When the person-post interaction data presents sparsity, high dimensionality, multi-modal (such as resume text, skill label, knowledge graph) and dynamic time-varying characteristics, the generalization performance and robustness of the traditional matching method are severely limited. Especially for new job seekers or newly released cold start posts, the traditional single model lacking of historical interaction data is difficult to generate accurate feature representation, resulting in large deviation of recommendation results, low matching success rate, and seriously affecting the user experience and platform activity of recruitment service.

[0004] In addition, the traditional method often ignores the complementarity of multi-view heterogeneous information and the logical implication relationship of external knowledge graph in the person-post matching process, and it is difficult to comprehensively utilize the potential association and privacy constraints between content view, structure view and attribute view. For example, in the recruitment network containing complex social relationship and skill dependence relationship, the traditional homogeneous graph analysis method cannot effectively fuse multi-modal features and privacy protection mechanism, resulting in weak noise resistance of matching model, poor explainability and privacy leakage risk. Even if some methods use general graph neural network technology, they cannot fully exploit the semantic structure information of meta-path, solve the distribution difference of cross-modal data and deal with noise interference under differential privacy, and it is difficult to realize efficient, safe, credible and explainable person-post matching recommendation.

[0005] Therefore, how to provide an AI person-post matching intelligent recommendation method based on a graph neural network is a problem that those skilled in the art urgently need to solve. SUMMARY

[0006] The application provides an AI person-post matching intelligent recommendation method based on a graph neural network. By combining a reciprocity preference matching mechanism with an improved HetGNN model, a deep learning architecture can more accurately calculate and predict the matching score and adaptation trend of job seekers and posts. This method can not only automatically collect text content, interactive topology, and label attributes and other features in multi-source heterogeneous person-post data, construct a multi-view heterogeneous graph, but also combine community discovery and cold start processing strategies, use multi-objective joint modularity optimization and Mahalanobis distance mapping to generate an initial feature vector, and greatly improve the robustness and generalization ability of person-post matching feature representation. By introducing a graph structured entropy path pruning layer, an optimal transport mapping alignment layer, and a cross-modal hash fusion layer, the feature learning process of the improved HetGNN model is optimized, so that the model can adaptively optimize the sampling distribution, eliminate the modal distribution difference, and extract core discriminative features, effectively suppressing the influence of noise on the matching result. At the same time, the bidirectional intention score is jointly constrained and optimized by using a reciprocity loss function, and nodes that do not meet the bidirectional consensus are removed based on the reciprocity adaptation bias, so as to realize accurate person-post matching and intelligent recommendation under the constraint of bidirectional reciprocity. The application overcomes the limitations of traditional person-post matching methods, such as single feature representation, poor cold start recommendation effect, and lack of bidirectional willingness cooperation, and provides an efficient and reliable solution for human resource intelligent allocation and decision support.

[0007] According to the AI person-post matching intelligent recommendation method based on a graph neural network, the following steps are included:

[0008] S1, collect multi-source heterogeneous person-post data, construct content view, structure view and attribute view based on text semantics, interactive topology and label hash, and fuse to generate a multi-view heterogeneous graph;

[0009] S2, perform multi-objective joint modularity optimization based on the heterogeneous graph, weighted aggregate the topology signal and apply spectral clustering constraint, divide the nodes into homogeneous communities, and output the community membership matrix and community-level statistical feature vector;

[0010] S3, identify low-degree cold start nodes according to the community membership matrix, calculate the Mahalanobis distance between the node attribute features and the community centroid, select the feature vector of the target community with the smallest distance for mapping, and generate an initial feature vector;

[0011] S4, for non-cold start nodes, construct an improved HetGNN model, use spectral entropy pruning and multi-view Figure 1Consistency constraint sampling of neighbors, alignment of attribute features with optimal transmission and embedding with knowledge graph, and heterogeneous attention aggregation through meta-path awareness, outputting a fused feature vector;

[0012] S5. Construct a feature decoupling module guided by maximizing mutual information, project the fused feature vector onto the matching latent space, suppress irrelevant noise through mutual information constraints, extract core discriminative features, and output the decoupled feature vector;

[0013] S6. Construct a reciprocal preference matching mechanism. Calculate the two-way intention scores of job seekers and positions based on the decoupled feature vectors. Use the reciprocal loss function to jointly constrain and optimize the two-way scores to generate a corrected matching score.

[0014] S7. Sort the recommended positions in descending order based on the corrected matching score and extract the candidate set. Calculate the reciprocal fit bias of the two-way scores, remove nodes with excessive bias, and output the final person-job matching result.

[0015] Optionally, S1 specifically includes:

[0016] S11. Collect multi-source heterogeneous personnel and job data, perform TF-IDF vectorization mapping on resume text and job description, calculate the cosine similarity of text feature vectors, filter and construct related edges according to similarity threshold, and generate a sparse adjacency matrix of content view.

[0017] S12. Collect multi-source heterogeneous human and work data, parse historical interaction logs to construct a bipartite graph topology, calculate the interaction frequency between nodes and perform normalization processing to determine edge weights, and generate the adjacency matrix of the structure view.

[0018] S13. Collect multi-source heterogeneous personnel and job data, perform MinHash dimensionality reduction mapping on skill tags, estimate the Jaccard similarity coefficient between tag sets based on hash signatures, establish attribute association edges based on similarity coefficients, and generate the adjacency matrix of attribute views.

[0019] S14. Based on node index alignment logic, perform multi-channel feature fusion on the sparse adjacency matrix of the content view, the adjacency matrix of the structure view, and the adjacency matrix of the attribute view to construct a unified noisy multi-view heterogeneous graph.

[0020] Optionally, S2 specifically includes:

[0021] S21. Based on the noisy multi-view heterogeneous graph, traverse the adjacency data of the content view, structure view and attribute view respectively, calculate the node degree matrix of each view, and calculate the normalized Laplace matrix according to the normalized Laplace transform algorithm. Perform eigenvalue decomposition operation on the normalized Laplace matrix and extract the eigenvectors corresponding to the minimum non-zero eigenvalues ​​of a preset number as the spectral domain topological signal features of each view.

[0022] S22. Calculate the ratio of signal energy to noise energy of each view based on the topological signal characteristics in the spectral domain, obtain the signal-to-noise ratio of the view, and calculate the correlation coefficient matrix of the feature vectors between different views as the mutual information between views. Calculate the weight allocation coefficient based on the signal-to-noise ratio of the view and the mutual information between views, and use the weight allocation coefficient to perform weighted linear fusion on the topological signal characteristics of each view to generate a multi-view joint topological signal.

[0023] S23. Construct an objective optimization function containing Rayleigh quotients, and introduce a second-order smoothness constraint regularization term into the objective optimization function. Use multi-view joint topological signals to solve for the generalized feature vector. Under the condition of satisfying spectral clustering constraints, iteratively calculate the maximum value of the objective function until the convergence condition is met, and output the optimal community partitioning result.

[0024] S24. Perform spectral embedding mapping on the feature vectors corresponding to the optimal community division results, calculate the posterior probability distribution of nodes belonging to each community category, generate soft allocation index based on the posterior probability distribution and construct a community membership matrix, aggregate node attribute data according to the index of the community membership matrix, calculate the mean and variance statistics of the attribute data, and generate community-level statistical feature vectors.

[0025] Optionally, S3 specifically includes:

[0026] S31. Analyze the row vector distribution of the community membership matrix, calculate the maximum a posteriori probability value and information entropy of each row, and mark the nodes whose maximum a posteriori probability value is lower than the first preset threshold or whose information entropy is higher than the second preset threshold as the cold start node set.

[0027] S32. For nodes in the cold start node set, calculate the covariance matrix and inverse matrix based on the node attribute features in each community. Use the inverse matrix of the covariance matrix to weight the difference between the node attribute features and the community centroid vector, and calculate the weighted Euclidean norm as the Mahalanobis distance.

[0028] S33. Construct the distance vector of the cold start node relative to all communities, perform a minimum search on the distance vector to locate the index corresponding to the minimum distance, and determine the community corresponding to the minimum distance index as the target community;

[0029] S34. Extract the community-level statistical feature vector of the target community, and project the community-level statistical feature vector onto the attribute space dimension of the cold start node through linear mapping transformation to generate the initial feature vector of the cold start node.

[0030] Optionally, the improved HetGNN model includes a graph-structured entropy path pruning layer, a multi-view path consistency constraint layer, an optimal transport map alignment layer, a cross-modal hash fusion layer, a heterogeneous attention aggregation layer, and a feature output layer.

[0031] The graph structured entropy path pruning layer is used to traverse all second-order meta-paths in the content view, structure view, and attribute view, calculate the Laplacian matrix spectral entropy of the subgraph formed by each meta-path, compare the calculated spectral entropy value with a preset noise threshold, remove meta-paths whose spectral entropy value is greater than or equal to the preset noise threshold, and output a set of candidate meta-paths that meet the preset spectral entropy threshold.

[0032] The multi-view path consistency constraint layer is used to receive a set of candidate meta-paths, perform walk sampling in each view to generate a sequence of neighbor nodes, extract the latent path representation of each view through a shared encoder, construct a contrastive learning loss function to maximize the mutual information of the cross-view path representation of the same node, constrain the consistency of the multi-view sampling distribution, filter nodes according to the consistency distribution, and output a semantically aligned set of neighbor nodes.

[0033] The optimal transport mapping alignment layer is used to extract node attribute features from the neighbor node set and entity embedding vectors from the external recruitment knowledge graph; calculate the Wasserstein distance between the feature distribution and the embedding distribution, and use the Sinkhorn algorithm to solve for the optimal transport plane; project the entity embedding vectors onto the semantic space of the node attribute features through the optimal transport plane, eliminate modal distribution differences, and output the knowledge-enhanced feature tensor after distribution-level alignment.

[0034] The cross-modal hash fusion layer is used to input the knowledge-enhanced feature tensor and the original node attribute features into the deep hash network for binary encoding to generate equal-length binary hash codes; the Hamming distance between the two sets of hash codes is calculated to construct a cross-modal attention weight matrix; the hash codes are weighted, concatenated, and decoded using the attention weight matrix to output a hybrid feature vector that integrates cross-modal semantics.

[0035] The heterogeneous attention aggregation layer is used to receive the mixed feature vector as the node center feature, combine it with the mixed feature vector of the semantically aligned neighbor node set, introduce a meta-path type-aware attention mechanism to calculate the aggregation weight; perform weighted summation and nonlinear transformation based on the aggregation weight, aggregate the high-order structural information of the neighbors, update the node hidden state, and output the aggregated node feature vector.

[0036] The feature output layer is used to input the aggregated node feature vectors into the fully connected layer to perform dimension mapping, perform non-linear mapping through the Tanh activation function, and perform L2 norm normalization on the output vector to generate a fused feature vector with uniform modulus.

[0037] Optionally, the feature decoupling module includes a matching spatial projection layer, a mutual information orthogonal decoupling constraint layer, and a mask feature generation layer:

[0038] The matching space projection layer is used to construct a mapping function containing multiple nonlinear transformations, which maps the input fused feature vector to the matching latent subspace of the Hilbert space, and performs centering and normalization processing on the mapped feature vector to output the latent feature representation tensor.

[0039] The mutual information orthogonal decoupling constraint layer is used to construct an estimator based on the lower bound of mutual information, calculate the mutual information estimate between the latent feature representation tensor and the matching label; at the same time, it constructs the feature covariance matrix and calculates the Frobenius norm as a correlation penalty term, constructs a joint loss function with the goal of maximizing label mutual information and minimizing feature correlation; gradient descent is performed on the joint loss function to calculate the importance weight vector of the feature components.

[0040] The mask feature generation layer is used to construct a binary mask matrix based on the importance weight vector. The binary mask matrix is ​​applied to the latent feature representation tensor through matrix multiplication to suppress feature dimensions with weights below a preset threshold, perform tensor reshaping operation, and output decoupled feature vectors.

[0041] Optionally, the reciprocal preference matching mechanism includes a two-way intention score calculation layer, a reciprocal preference constraint optimization layer, and a corrected score generation layer:

[0042] The bidirectional intention score calculation layer is used to construct a dual-tower neural network. The decoupled feature vectors are input into the job seeker tower and the job tower respectively. The original hidden layer output is calculated by a multilayer perceptron. The original hidden layer output is subjected to a Sigmoid nonlinear transformation to map to the probability space, and the first intention score of the job seeker for the job and the second intention score of the job seeker for the job are output respectively.

[0043] The reciprocal preference constraint optimization layer is used to construct the reciprocal consistency loss function, calculate the mean square error of the first intention score and the second intention score to quantify the two-way preference bias, introduce the Frobenius norm to constrain the alignment of the two-way feature representation, and weightedly fuse the reciprocal consistency loss function with the cross-entropy loss function of the basic recommendation task, and optimize the network parameters through gradient backpropagation.

[0044] The corrected score generation layer is used to recalculate the bidirectional intention score based on the optimized network parameters, construct an attention adjustment factor based on reciprocity strength, perform weighted fusion of the bidirectional intention scores using the reciprocity adjustment factor, and output the corrected matching score.

[0045] Optionally, S7 specifically includes:

[0046] S71. Receive the corrected matching score, sort the recommended positions in descending order according to the corrected matching score, and select the top-K position nodes in the top order to construct the initial candidate set;

[0047] S72. Traverse all nodes in the initial candidate set, extract the first preference score of job seekers for the position and the second preference score of the position for job seekers corresponding to each node, and calculate the absolute value of the difference between the first preference score and the second preference score as the reciprocal fit bias.

[0048] S73. Compare the reciprocal adaptation deviation with the preset deviation threshold. When the reciprocal adaptation deviation is greater than the deviation threshold, mark the node as a one-way intention node and remove it from the initial candidate set. When the reciprocal adaptation deviation is less than or equal to the deviation threshold, retain the node and output the final job matching result.

[0049] The beneficial effects of this invention are:

[0050] This invention effectively addresses the technical challenges of feature alignment and poor cold-start recommendation performance in multi-source heterogeneous job matching data by employing an improved HetGNN model. Traditional job matching methods typically neglect the modal distribution differences between node attribute features and external recruitment knowledge graph entity embeddings, as well as the cold-start problem of new nodes lacking interactive data. The improved HetGNN model calculates the Wasserstein distance between the feature distribution and the embedding distribution through an optimal transport mapping alignment layer and uses the Sinkhorn algorithm to solve for the optimal transport plane, effectively eliminating distribution differences between different modalities and achieving deep fusion of node attributes and the knowledge graph. Simultaneously, by combining multi-objective joint modularity optimization and community centroid feature mapping, the statistical features of the target community are mapped to cold-start nodes based on Mahalanobis distance, effectively utilizing community-level semantic information to generate high-quality initial representations for new nodes. This method overcomes the limitations of traditional methods in heterogeneous feature fusion and new node representation learning, significantly improving the semantic depth of job matching and the recommendation success rate in cold-start scenarios, providing an efficient and accurate solution for intelligent job matching services. Attached Figure Description

[0051] 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:

[0052] Figure 1 This is an overall flowchart of an AI-based intelligent recommendation method for matching people to jobs based on graph neural networks, as proposed in this invention.

[0053] Figure 2 This is a flowchart illustrating the working principle of the improved HetGNN model, which is based on a graph neural network for intelligent recommendation of AI-based human-job matching. Detailed Implementation

[0054] 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.

[0055] refer to Figure 1 and Figure 2 An AI-based intelligent recommendation method for matching people to jobs, based on graph neural networks, includes the following steps:

[0056] S1. Collect multi-source heterogeneous human and job data, construct content view, structure view and attribute view based on text semantics, interaction topology and tag hash, and fuse them to generate a multi-view heterogeneous graph;

[0057] S2. Perform multi-objective joint modularity optimization based on heterogeneous graphs, weighted aggregate topological signals and apply spectral clustering constraints, divide nodes into homogeneous communities, and output the community membership matrix and community-level statistical feature vectors.

[0058] S3. Identify low-degree cold-start nodes based on the community membership matrix, calculate the Mahalanobis distance between the node attribute features and the community centroid, select the feature vector of the target community with the smallest distance for mapping, and generate the initial feature vector.

[0059] S4. Construct an improved HetGNN model for non-cold start nodes, utilizing spectral entropy pruning and multi-view... Figure 1 Consistency constraint sampling of neighbors, alignment of attribute features with optimal transmission and embedding with knowledge graph, and heterogeneous attention aggregation through meta-path awareness, outputting a fused feature vector;

[0060] S5. Construct a feature decoupling module guided by maximizing mutual information, project the fused feature vector onto the matching latent space, suppress irrelevant noise through mutual information constraints, extract core discriminative features, and output the decoupled feature vector;

[0061] S6. Construct a reciprocal preference matching mechanism. Calculate the two-way intention scores of job seekers and positions based on the decoupled feature vectors. Use the reciprocal loss function to jointly constrain and optimize the two-way scores to generate a corrected matching score.

[0062] S7. Sort the recommended positions in descending order based on the corrected matching score and extract the candidate set. Calculate the reciprocal fit bias of the two-way scores, remove nodes with excessive bias, and output the final person-job matching result.

[0063] In this embodiment, S1 specifically includes:

[0064] S11. Collect multi-source heterogeneous job and resume data, and perform TF-IDF vectorization mapping on the resume text and job description. Specifically, this includes parsing the original character sequences from the collected data, performing cleaning operations to remove HTML tags, special symbols, and stop words, using a word segmentation algorithm to divide the cleaned text into word sequences, calculating word frequencies and inverse document frequencies, and multiplying them as weights. The word space is then mapped into sparse text feature vectors containing weight values. Based on the generated sparse text feature vectors, job seeker nodes and job node traverses, calculating the dot product of the resume text feature vector and the job description text feature vector as the weight. The algorithm calculates the L2 norm of two feature vectors and uses their product as the denominator. It then calculates the ratio of the numerator to the denominator to obtain the cosine similarity value. Subsequently, it sets a semantic relevance boundary threshold of 0.3 and compares the calculated cosine similarity value with 0.3. It selects node pairs with similarity values ​​greater than or equal to 0.3 as edges of the graph structure and maps the corresponding similarity values ​​to edge weights to establish connections between nodes. At the same time, it removes node pairs with similarity values ​​lower than 0.3. Finally, it converts the graph structure data containing connections and weights into a sparse matrix form and outputs the sparse adjacency matrix of the content view.

[0065] S12. Collect multi-source heterogeneous user and job data, analyze historical interaction logs to extract the binary association between user nodes and job nodes, and construct a bipartite graph topology; count the interaction frequency between user nodes and job nodes, calculate the maximum and minimum normalized values ​​of the interaction frequency as the initial weights of the bipartite graph edges, and generate the adjacency matrix of the structure view; the process of constructing the bipartite graph topology and determining the edge weights based on historical interaction logs includes: traversing historical interaction log data, identifying the user identifier and job identifier corresponding to each interaction record, and constructing a node set and an edge set; based on the node set and edge set... Initialize the adjacency matrix of the bipartite graph, setting the element values ​​at corresponding positions in the adjacency matrix to the original count values ​​of the interaction frequency; extract all non-zero element values ​​from the adjacency matrix, finding the maximum value of 50 and the minimum value of 1 in the value set; use a linear transformation to map the original count values ​​in the range of 1 to 50 to the range of 0.01 to 1, setting the edge weight to 1 when the interaction frequency is 50 and setting the edge weight to 0.01 when the interaction frequency is 1; backfill the calculated specific edge weights into the adjacency matrix, generating a structured view adjacency matrix containing edge weight information.

[0066] S13. Collect multi-source heterogeneous personnel and job data, and perform MinHash dimensionality reduction mapping on skill tags. Specifically, this includes parsing the skill tag set in the data, constructing a family of hash functions containing 128 different hash functions, taking each skill tag set as input and calculating the minimum hash value through the family of hash functions in turn, generating a fixed-length hash signature of length 128, and then comparing the number of slots with the same value in the hash signatures of different nodes. By calculating the ratio of the number of the same slots to the total number of 128, the Jaccard similarity coefficient between tag sets is estimated. Then, a preset attribute similarity threshold of 0.6 is set, and node pairs with similarity coefficients higher than 0.6 are selected. Attribute connection relationships are established with node pairs as vertices of the graph and similarity coefficients as edge weights. The nodes and attribute connection relationships of the entire graph are converted into matrix form, and the adjacency matrix of the attribute view is output.

[0067] S14. Based on node index alignment logic, perform multi-channel feature fusion on the sparse adjacency matrix of the content view, the adjacency matrix of the structure view, and the adjacency matrix of the attribute view to construct a unified noisy multi-view heterogeneous graph. Specifically, this includes traversing all nodes in all views, assigning a unified index number to each node based on its globally unique identifier, constructing a three-dimensional tensor structure, mapping the three view adjacency matrices after index alignment to three view channels in the unified graph structure, and fusing the node information and edge information of each view channel to generate a unified noisy multi-view heterogeneous graph.

[0068] In this embodiment, S2 specifically includes:

[0069] S21. Based on the noisy multi-view heterogeneous graph, traverse the adjacency data of the content view, structure view, and attribute view respectively, and calculate the node degree matrix of each view. Specifically, for the adjacency matrix of each view, calculate the sum of the weights of the edges connected to each node to generate a diagonal degree matrix. Then, calculate the normalized Laplace matrix according to the normalized Laplace transform algorithm. Construct the normalized Laplace matrix by subtracting the negative half power of the degree matrix from the identity matrix and the product of the adjacency matrix and the negative half power of the degree matrix. Then, perform eigenvalue decomposition operation on the normalized Laplace matrix. Solve the eigenvalues ​​and eigenvectors of the normalized Laplace matrix through matrix decomposition technology. Sort the eigenvalues ​​in ascending order according to their numerical values. Finally, extract the eigenvectors corresponding to the smallest non-zero eigenvalues ​​of a preset number of 32 as the spectral domain topological signal features of each view. According to the preset feature dimension threshold, remove the smallest zero eigenvalue and its corresponding eigenvector, and select the eigenvectors corresponding to the next 32 smallest non-zero eigenvalues. Combine them as the spectral domain topological signal features of each view.

[0070] S22. Calculate the ratio of signal energy to noise energy for each view based on spectral domain topological signal characteristics. Specifically, this includes solving for the squared L2 norm of the feature vectors of each view as the signal energy, estimating the noise energy based on a preset noise model, calculating the ratio of the two as the view signal-to-noise ratio, calculating the correlation coefficient matrix of feature vectors between different views as the mutual information between views, calculating the Pearson correlation coefficient between feature vectors of different views, then calculating the weight allocation coefficient based on the view signal-to-noise ratio and the mutual information between views, constructing a weighted evaluation function of signal-to-noise ratio and mutual information, assigning corresponding weight coefficients to each view based on the evaluation function value, and finally using the weight allocation coefficients to perform weighted linear fusion of the spectral domain topological signal characteristics of each view, multiplying the spectral domain topological signal characteristics of each view with the corresponding weight coefficients and summing them to generate a multi-view joint topological signal.

[0071] S23. Construct a target optimization function containing a Rayleigh quotient. Specifically, this includes defining a Rayleigh quotient function with the ratio of the eigenvector magnitude to the quadratic form of the Laplacian matrix as the objective, introducing a second-order smoothness constraint regularization term into the target optimization function, adding a regularization penalty term based on the trace operation of the Laplacian matrix into the objective function, and then using multi-view joint topological signals to solve for generalized eigenvectors. Using multi-view joint topological signals as input data, the corresponding eigenvectors are obtained by solving the generalized eigenvalue problem. Subsequently, under the condition of satisfying spectral clustering constraints, the maximum value of the objective function is calculated iteratively. An iterative update algorithm is used to repeatedly adjust the eigenvector values ​​to maximize the objective function value. The difference between the objective function values ​​of two adjacent iterations is monitored until the convergence condition is met. The calculation stops when the difference is less than the preset precision threshold of 0.0001 or reaches the maximum number of iterations of 500. The node category label corresponding to the eigenvector at this time is output as the optimal community partitioning result.

[0072] S24. Perform spectral embedding mapping on the feature vectors corresponding to the optimal community division results. Specifically, this includes mapping the optimal community division results to a low-dimensional feature space of dimension 64, then calculating the posterior probability distribution of nodes belonging to each community category, using the distance between the feature vectors and the center of each category to calculate the probability value of a node belonging to each community, then generating a soft assignment index based on the posterior probability distribution and constructing a community membership matrix, arranging the probability values ​​of all nodes belonging to each community in matrix form according to the node index, and finally aggregating node attribute data according to the index of the community membership matrix, aggregating the attribute features of nodes within the same community according to the probability weight of each node in the community membership matrix, calculating the mean and variance statistics of the attribute data, calculating the mathematical expectation and variance of the aggregated attribute features, and generating community-level statistical feature vectors.

[0073] In this embodiment, S3 specifically includes:

[0074] S31. Analyze the row vector distribution of the community membership matrix, calculate the maximum a posteriori probability value and information entropy of each row, and mark the nodes with maximum a posteriori probability values ​​lower than the first preset threshold or information entropy higher than the second preset threshold as the cold start node set. Specifically, this includes reading the community membership matrix, extracting the probability value sequence of each row corresponding to the node belonging to each community category, traversing the probability value sequence, selecting the maximum value as the maximum a posteriori probability value, calculating the information entropy based on the probability value sequence, comparing the maximum a posteriori probability value with the first preset threshold of 0.6 and the information entropy with the second preset threshold of 1.5 respectively, and storing the nodes with maximum a posteriori probability values ​​less than 0.6 or information entropy greater than 1.5 into the cold start node set.

[0075] S32. For nodes in the cold start node set, calculate the covariance matrix and inverse matrix based on the node attribute features in each community. Use the inverse matrix of the covariance matrix to weight the difference between the node attribute features and the community centroid vector. Calculate the weighted Euclidean norm as the Mahalanobis distance. Specifically, this includes aggregating the attribute feature data of nodes in each community, calculating the covariance matrix of the attribute data and solving for the inverse matrix, calculating the difference vector between the attribute feature vector of the cold start node and the community centroid vector, applying the inverse matrix of the covariance matrix as a transformation kernel to the difference vector, and calculating the L2 norm of the weighted vector.

[0076] S33. Construct a distance vector of the cold start node relative to all communities, perform a minimum search on the distance vector to locate the index corresponding to the minimum distance, and determine the community corresponding to the minimum distance index as the target community. Specifically, for the current cold start node, calculate its Mahalanobis distance to each community, arrange all the calculated distance values ​​in community order to form a distance vector, traverse all elements in the distance vector, find the element with the smallest value, obtain the position index of the element in the vector, map the position index to the corresponding community category, and determine the category as the target community.

[0077] S34. Extract the community-level statistical feature vector of the target community, and project the community-level statistical feature vector onto the attribute space dimension of the cold start node through linear mapping transformation to generate the initial feature vector of the cold start node. Specifically, this includes retrieving and obtaining the mean and variance statistical vectors corresponding to the target community from the pre-generated community feature set, constructing a dimension transformation matrix or a fully connected mapping layer, inputting the community-level statistical feature vector into the mapping layer, adjusting the vector dimension to be consistent with the original attribute feature dimension of the cold start node, and performing nonlinear activation and normalization processing on the mapped vector.

[0078] In this embodiment, the improved HetGNN model includes a graph-structured entropy path pruning layer, a multi-view path consistency constraint layer, an optimal transport map alignment layer, a cross-modal hash fusion layer, a heterogeneous attention aggregation layer, and a feature output layer.

[0079] The graph structured entropy path pruning layer is used to traverse all second-order meta-paths in the content view, structure view, and attribute view, and calculate the Laplacian matrix spectral entropy of the subgraph formed by each meta-path. Specifically, this includes constructing the adjacency matrix of the subgraph based on the connection relationship between the first and last nodes of the meta-path, calculating the node degree matrix, and multiplying the negative half power of the degree matrix by the adjacency matrix and the negative half power of the degree matrix to solve for the normalized Laplacian matrix; performing eigenvalue decomposition on the normalized Laplacian matrix to extract all non-zero eigenvalues ​​and remove them. The smallest eigenvalue with a value of 0 is identified. The remaining eigenvalues ​​are normalized so that their sum equals 1 to construct an eigenvalue probability distribution. The logarithm of each element in the eigenvalue probability distribution is calculated. Each probability element is multiplied by its corresponding logarithm, and all products are summed to obtain the spectral entropy value. The calculated spectral entropy value is compared with a preset noise threshold of 3.5. Metapaths with spectral entropy values ​​greater than or equal to 3.5 are removed, while metapaths with spectral entropy values ​​less than 3.5 are retained. A set of candidate metapaths that meet the preset spectral entropy threshold is output.

[0080] The multi-view path consistency constraint layer receives a set of candidate meta-paths and performs walk sampling in each view to generate a sequence of neighboring nodes. Specifically, it randomly initializes a starting node in each view and performs walk sampling with 40 steps to obtain a sequence of neighboring nodes containing 41 nodes. A shared encoder network is constructed, consisting of a single-layer fully connected embedding layer, a single-layer long short-term memory network, and a global average pooling layer. The hidden layer dimension of the long short-term memory network is set to 64, and ReLU is used as the activation function. A Dropout random deactivation mechanism is used to discard neurons at a ratio of 0.2 to prevent overfitting. The neighboring node sequence is then sequentially input into this shared encoder network. In this process, node indices are mapped to a 128-dimensional vector space through a fully connected embedding layer. The temporal dependencies of the sequence are captured through a long short-term memory network. Finally, the temporal features are aggregated through a global average pooling layer, and the low-dimensional latent feature vector of the path with a dimension of 128 is obtained. Then, a contrastive learning loss function is constructed to maximize the mutual information of the path representation of the same node across views. The positive sample similarity of the path representation of the same node under different views is calculated, and a loss function is constructed to maximize this similarity. Finally, nodes are selected based on the consistency distribution. Based on the probability distribution after consistency constraint, the top 10 nodes with the highest probability are selected as high-confidence neighbor nodes to construct a semantically aligned neighbor node set.

[0081] The optimal transport mapping alignment layer is used to extract node attribute features from the neighbor node set and entity embedding vectors from the external recruitment knowledge graph. Specifically, it includes reading the attribute matrix of the graph nodes and the entity vectors of the knowledge graph, calculating the Wasserstein distance between the feature distribution and the embedding distribution (i.e., calculating the Earth's distance between two feature distribution sets), solving for the optimal transport plane using the Sinkhorn algorithm, constructing a cost matrix containing the Euclidean distance between node features and entity vectors, setting an entropy regularization coefficient of 0.1 and performing Sinkhorn iteration, setting the upper limit of the number of iterations to 100 or stopping the iteration when the norm of the difference between two consecutive updated transport matrices is less than a preset precision threshold of 0.001, and outputting the optimal transport matrix that satisfies the entropy regularization constraint; projecting the entity embedding vectors onto the semantic space of the node attribute features through the optimal transport plane, performing matrix multiplication between the optimal transport matrix and the entity embedding vectors of the external recruitment knowledge graph, using the product result to perform a linear transformation on the entity embedding vectors, aligning them to the node attribute space, eliminating modal distribution differences, and outputting the distribution-aligned knowledge-enhanced feature tensor.

[0082] The cross-modal hash fusion layer is used to input the knowledge-enhanced feature tensor and the original node attribute features into a deep hash network for binary encoding. Specifically, it includes constructing a deep hash network containing two fully connected layers and a batch normalization layer; inputting continuous features into the network and using the Tanh activation function to map the values ​​to between -1 and +1; performing symbolic processing on the mapped values, setting values ​​greater than or equal to 0 to 1 and values ​​less than 0 to 0 or -1, generating equal-length binary hash codes; calculating the Hamming distance between the two sets of hash codes to construct a cross-modal attention weight matrix; and performing this calculation on each hash code. A bitwise XOR operation is performed on each bit, and the number of bits with a result of 1 is counted to obtain the bit-level distance. The reciprocal of the bit-level distance is exponentially operated on using the Softmax normalization function to generate the attention weight corresponding to each bit. The hash code is weighted and decoded using the attention weight matrix. The attention weights are multiplied element-wise with the binary hash code to obtain the weighted hash code. The weighted hash code is mapped back to the continuous vector space through a decoding network containing two fully connected layers. The decoding network uses the ReLU activation function for nonlinear transformation and outputs a hybrid feature vector that integrates cross-modal semantics.

[0083] The heterogeneous attention aggregation layer receives a hybrid feature vector as the node's center feature, combines it with the hybrid feature vector of the semantically aligned neighbor node set, and introduces a meta-path type-aware attention mechanism to calculate the aggregation weight. Specifically, it performs a vector concatenation operation between the node's center feature vector and the hybrid feature vector of each neighbor node, and inputs the concatenation result into a single-layer feedforward neural network to calculate the attention score. At the same time, it introduces a learnable semantic vector corresponding to the meta-path type to linearly correct the score. For all neighbors of the same node, it uses the Softmax normalization function to normalize all corrected attention scores, mapping the score values ​​to between 0 and 1 and accumulating them to 1 to generate the final aggregation weight. Based on the aggregation weight, it performs a weighted summation and nonlinear transformation, multiplies the aggregation weight with the corresponding neighbor hybrid feature vector, and accumulates the sum. The accumulated result is input into the ReLU nonlinear activation function for operation, aggregating the high-order structural information of the neighbors, updating the node's hidden state, and using the output of the activation function and the original center feature of the node to perform residual connection and layer normalization to update the node's current feature representation, outputting the aggregated node feature vector.

[0084] The feature output layer is used to input the aggregated node feature vectors into the fully connected layer to perform dimension mapping. Specifically, it includes adjusting the feature vectors to a target dimension of 128 using the fully connected layer, performing non-linear mapping through the Tanh activation function to map the feature values ​​to a preset range of -1 to +1, performing L2 norm normalization on the output vector, calculating the L2 norm of the vector and using this value to scale the vector to generate a fused feature vector with a modulus of 1.

[0085] In this embodiment, the feature decoupling module includes a matching spatial projection layer, a mutual information orthogonal decoupling constraint layer, and a mask feature generation layer:

[0086] The matching space projection layer is used to construct a mapping function containing multiple nonlinear transformations, mapping the input fused feature vector to the matching latent subspace of the Hilbert space, and performing centering and standardization on the mapped feature vector to output a latent feature representation tensor. The process of constructing the mapping function and performing centering and standardization includes: setting three fully connected neural network layers, with 512, 256 and 128 nodes in each layer, respectively, and using the ReLU function as the nonlinear activation function; inputting a fused feature vector of length 2048, which is output as a 128-dimensional feature vector after layer-by-layer linear transformation and nonlinear activation; calculating the mean of the 128-dimensional feature vector in the batch as 0.002 and the standard deviation as 0.155; subtracting the mean of 0.002 from each dimension value of the feature vector and dividing the result by the standard deviation of 0.155 to obtain standardized data with a mean of 0 and a variance of 1, and outputting this standardized data as a latent feature representation tensor.

[0087] The mutual information orthogonal decoupling constraint layer is used to construct an estimator based on the lower bound of mutual information, calculating the estimated mutual information between the latent feature representation tensor and the matched labels; simultaneously, it constructs a feature covariance matrix and calculates the Frobenius norm as a correlation penalty term, constructing a joint loss function with the objective of maximizing label mutual information and minimizing feature correlation; gradient descent is performed on the joint loss function to calculate the importance weight vector of the feature components; the process of constructing the joint loss function and solving for the weight vector includes: constructing a feature covariance matrix of dimension 128 by 128, calculating the sum of the absolute values ​​of the off-diagonal elements of the matrix as the Frobenius norm. Using the us norm and setting the relevance penalty coefficient to 0.1, the relevance penalty term is calculated to be 0.05. A contrastive learning estimator is used to calculate the mutual information estimate between latent features and matching labels, which is 2.35. The mutual information maximization weight coefficient is set to 1.0. The total loss function is constructed by subtracting the relevance penalty term from the mutual information estimate, resulting in a value of 2.30. Gradient descent is performed iteratively 300 times on the total loss function with a learning rate of 0.001 until the loss converges. The output is an importance weight vector corresponding to 128 feature dimensions, where the weight of the 10th dimension is 0.92 and the weight of the 50th dimension is 0.15.

[0088] The mask feature generation layer is used to construct a binary mask matrix based on the importance weight vector. The binary mask matrix is ​​then applied to the latent feature representation tensor through matrix multiplication to suppress feature dimensions with weights below a preset threshold. A tensor reshaping operation is performed, and a decoupled feature vector is output. The process of constructing the mask matrix and suppressing feature dimensions includes: traversing the 128 weight values ​​in the importance weight vector and comparing each weight value with a preset threshold of 0.5; when a weight value is greater than or equal to 0.5, setting the corresponding element in the mask matrix to 1, and when a weight value is less than 0.5, setting the corresponding element to 0, generating a binary mask matrix containing 0 and 1; performing element-wise matrix multiplication on the binary mask matrix and the latent feature representation tensor, setting the tensor dimension values ​​corresponding to 0 elements in the mask matrix to 0, and keeping the tensor dimension values ​​corresponding to 1 elements unchanged; removing the feature dimensions set to 0, rearranging the remaining dimension data, reshaping it into a new tensor structure, and outputting a decoupled feature vector.

[0089] In this embodiment, the reciprocal preference matching mechanism includes a two-way intention score calculation layer, a reciprocal preference constraint optimization layer, and a corrected score generation layer:

[0090] The bidirectional intention score calculation layer is used to construct a dual-tower neural network. Decoupled feature vectors are input into the job seeker tower and the job tower respectively. The original hidden layer output is calculated using a multilayer perceptron. A Sigmoid nonlinear transformation is performed on the original hidden layer output to map it to a probability space, outputting the job seeker's first intention score for the job and the job's second intention score for the job seeker. The process of constructing the dual-tower neural network and calculating the intention scores includes: setting both the job seeker tower and the job tower to contain three fully connected layers, with 256, 128, and 64 neurons in each layer, respectively. The ReLU activation function was chosen. A 128-dimensional decoupled feature vector of job seekers was input into the job seeker pyramid. After forward propagation, the original hidden layer output value was 2.5. The Sigmoid function was applied to map 2.5 to a probability interval of 0 to 1, and the job seeker's first intention score for the position was calculated. Simultaneously, a 128-dimensional decoupled feature vector of the job was input into the job pyramid. After forward propagation, the original hidden layer output value was 1.2. The Sigmoid function was applied to map 1.2 to a probability interval, and the job's second intention score for the job seeker was calculated.

[0091] The reciprocity preference constraint optimization layer is used to construct the reciprocity consistency loss function, calculate the mean squared error between the first intention score and the second intention score to quantify the bidirectional preference bias, introduce Frobenius norm to constrain the alignment of bidirectional feature representations, and weight and fuse the reciprocity consistency loss function with the cross-entropy loss function of the basic recommendation task. The network parameters are optimized through gradient backpropagation. The process of constructing the loss function and optimizing the parameters includes: calculating the difference between the first intention score and the second intention score, and using the square of the difference as the reciprocity consistency loss term; calculating the distance matrix between the 64-dimensional feature vectors output by the job seeker tower and the job tower, and obtaining the Frobenius norm value of the matrix, multiplying the Frobenius norm value by a coefficient of 0.01 to obtain the feature alignment penalty term value; setting the cross-entropy loss term value of the basic recommendation task to 0.3; adding the reciprocity consistency loss term, the feature alignment penalty term, and the cross-entropy loss term to obtain the total loss function value; using the Adam optimizer to perform gradient backpropagation iteration on the total loss function, setting the learning rate to 0.0001, and updating the weight parameters in the network.

[0092] The corrected score generation layer is used to recalculate the bidirectional intention score based on the optimized network parameters, construct an attention adjustment factor based on reciprocity strength, and perform weighted fusion of the bidirectional intention scores using the reciprocity adjustment factor to output a corrected matching score. The process of constructing the adjustment factor and calculating the corrected score includes: re-inputting the feature vector using the updated network parameters to calculate the optimized first intention score and second intention score; adding the first intention score and the second intention score to obtain a sum, dividing by the value of 2 to obtain the reciprocity base value; constructing a Sigmoid nonlinear transformation function, using the reciprocity base value as the function input, and calculating the output value as the reciprocity adjustment factor; calculating the product of the first intention score and the reciprocity adjustment factor, and the product of the second intention score and the reciprocity adjustment factor, adding the two products to obtain the final value, performing normalization processing, and outputting the corrected matching score.

[0093] In this embodiment, S7 specifically includes:

[0094] S71. Receive the corrected matching score, sort the recommended positions in descending order based on the corrected matching score, and select the Top-K position nodes in the top order to construct an initial candidate set. The process of performing the descending sort and constructing the candidate set includes: obtaining the corrected matching score corresponding to all positions to be recommended in the system, using the corrected matching score as the sorting key value, and sorting the recommended positions in descending order; setting the Top-K truncation parameter K to 50, and selecting the top 50 position nodes in the sorting sequence; extracting the position identifier and corresponding corrected matching score of these 50 position nodes, and storing them in the initial candidate set list.

[0095] S72. Traverse all nodes in the initial candidate set, extract the job seeker's first preference score for the position and the position's second preference score for the job seeker corresponding to each node, and calculate the absolute value of the difference between the first preference score and the second preference score as the reciprocal fit bias; the process of extracting scores and calculating bias includes: traversing each of the 50 job nodes in the initial candidate set one by one; for each traversed job node, extracting the job seeker's preference score value and the position's preference score value corresponding to that node from the database; calculating the difference between these two values, and taking the absolute value of the difference result; recording the calculated absolute value as the reciprocal fit bias of that job node.

[0096] S73. Compare the reciprocal adaptation deviation with a preset deviation threshold. When the reciprocal adaptation deviation is greater than the deviation threshold, mark the node as a one-way intention node and remove it from the initial candidate set. When the reciprocal adaptation deviation is less than or equal to the deviation threshold, retain the node and output the final job matching result. The process of comparing deviations and removing nodes includes: setting the preset deviation threshold to 0.20; comparing the reciprocal adaptation deviation value of each job node with 0.20; marking job nodes with a reciprocal adaptation deviation value greater than 0.20 as one-way intention nodes and removing them from the initial candidate set; retaining job nodes with a reciprocal adaptation deviation value less than or equal to 0.20 in the candidate set; after traversal, organize the remaining retained job nodes into the final job matching result and output it.

[0097] Example 1:

[0098] To verify the feasibility of this invention in the field of intelligent human resource matching, the method of this invention was applied to the intelligent job recommendation system of a large comprehensive recruitment platform (hereinafter referred to as "Company G"). In existing recruitment matching systems, content-based TF-IDF text matching or traditional heterogeneous graph neural network algorithms are typically used. These methods not only struggle to solve the problem of view noise interference when fusing multi-source heterogeneous data, but also fail to effectively align the distribution differences between external knowledge graphs and internal node features, resulting in semantic gaps in feature representations and an inability to capture the mutual reciprocity between job seekers and positions. This easily leads to significant deviations between recommendation results and actual hiring intentions. To address these issues, Company G decided to adopt the AI-based intelligent job matching recommendation method based on graph neural networks proposed in this invention.

[0099] During implementation, Company G first collected multi-source heterogeneous job and personnel data from the platform. TF-IDF vectorization was performed on resume text and job descriptions, edge weights were calculated by parsing historical interaction logs, and MinHash dimensionality reduction was applied to skill tags. Sparse adjacency matrices were then constructed for content views, structure views, and attribute views. Based on node index alignment logic, a unified noisy multi-view heterogeneous graph was generated. After normalized Laplace transform, eigenvalue decomposition, and Rayleigh commercial project target optimization, a community membership matrix was output, and community-level statistical feature vectors were generated based on index aggregation.

[0100] Company G employs a graph-structured entropy path pruning layer to calculate the spectral entropy of the Laplacian matrix of each meta-path subgraph, eliminating high-noise meta-paths with spectral entropy values ​​greater than or equal to a preset noise threshold, and selecting a set of candidate meta-paths. A multi-view path consistency constraint layer is used to extract latent path representations through a shared encoder, and a contrastive learning loss function is constructed to maximize cross-view mutual information, outputting a semantically aligned set of neighboring nodes. Subsequently, an optimal transport mapping alignment layer is used to calculate the Wasserstein distance between the feature distribution and the embedding distribution, and the Sinkhorn algorithm is used to solve for the optimal transport plane, projecting the entity embedding vectors of the external recruitment knowledge graph onto the node attribute feature space, eliminating modal distribution differences. In the cross-modal hash fusion layer, the knowledge-enhanced feature tensor and the original node attribute features are input into a deep hash network to generate binary hash codes, and the Hamming distance is calculated to construct an attention weight matrix, outputting a hybrid feature vector.

[0101] In the core matching and generation stage, this invention constructs an improved HetGNN model for non-cold-start nodes, aggregating high-order structural information of neighbors through a heterogeneous attention aggregation layer to output a fused feature vector. Simultaneously, low-degree cold-start nodes are identified based on the community membership matrix, and the Mahalanobis distance between node attribute features and community centroids is calculated. The feature vector of the target community with the smallest distance is selected for mapping, generating an initial feature vector. Subsequently, the system maps the fused feature vector to the matching latent subspace of the Hilbert space through a matching space projection layer, and constructs a joint loss function using a mutual information orthogonal decoupling constraint layer. The importance weight vector of the feature components is calculated, a binarized mask matrix is ​​generated to suppress irrelevant noise, and the core discriminative features are output.

[0102] The system employs a reciprocity preference matching mechanism, utilizing a two-way intention score calculation layer to output the job seeker's first intention score for a position and the position's second intention score for the job seeker. A reciprocity preference constraint optimization layer calculates the mean squared error of the two-way intention scores to construct a reciprocity consistency loss function, which is then weighted and fused with a cross-entropy loss function. Gradient backpropagation optimizes the network parameters. Based on the optimized network parameters, a score generation layer constructs an attention adjustment factor based on reciprocity strength, performs weighted fusion of the two-way intention scores, and outputs a corrected matching score. The recommended positions are sorted in descending order according to the corrected matching score, and the top-K position nodes are selected. The reciprocity fit bias of the two-way scores is calculated, and nodes with a bias greater than a preset threshold are marked as one-way intention nodes and removed, outputting the final job-person matching result.

[0103] During implementation, the technical team at Company G discovered that, compared with traditional keyword matching and conventional graph neural network algorithms, the method of this invention significantly improves the semantic richness and bidirectional adaptability of job matching. Through spectral clustering community discovery, spectral entropy pruning, optimal transmission alignment, feature decoupling, and reciprocal preference constraints, it effectively achieves high-precision, high-security, and bidirectional consensus-based intelligent job matching and recommendation.

[0104] To further verify the actual performance of the method of the present invention, Company G 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:

[0105] Table 1 Performance Comparison of Company G Intelligent Person-Job Matching System

[0106] Indicator Traditional method Method of the present application Lifting amplitude Human-post matching accuracy (%) 78.4 92.6 +14.2% Cold start node matching recall rate (%) 45.2 88.5 +43.3% Recommendation reason generation relevance score 0.72 0.94 +0.22 Cross-modal feature alignment degree (%) 65.8 89.2 +23.4% High-risk post mis-push rate (%) 15.6 2.1 -13.5% Recommendation reason chain generation time-consuming (milliseconds) 320 85 -73.4% User resume delivery conversion rate (%) 12.5 21.8 +9.3% HR resume screening time-consuming (hours / day) 4.5 1.2 -73.3% Platform user retention rate (months%) 68.0 82.0 +14.0% System concurrent processing capacity (QPS) 800 2500 +212.5%

[0107] As shown in Table 1, the performance of the intelligent job matching system was comprehensively improved after applying the method of this invention. The accuracy of job matching increased from 78.4% of the traditional method to 92.6%, and the cold start node matching recall rate increased significantly from 45.2% to 88.5%, significantly solving the problem of matching new users. The relevance score of the recommendation reason generation increased from 0.72 to 0.94, effectively improving user trust and experience. The false recommendation rate for high-risk positions decreased from 15.6% to 2.1%, greatly improving the reliability of recommendations. In addition, the user resume submission conversion rate increased from 12.5% ​​to 21.8%, and the platform user retention rate also improved significantly, from 68.0% to 82.0%.

[0108] Through the method of this invention, Company G has successfully achieved deep semantic alignment and accurate matching between job seeker profiles and job requirements, effectively reducing matching deviations in cold start scenarios, ensuring the activity of the recruitment ecosystem, significantly improving the intelligence and personalization of human resource services, significantly reducing the workload of HR in screening resumes, enhancing the robustness and interpretability of the recommendation system, and providing strong support for the technological upgrade of the intelligent recruitment platform.

[0109] 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. An AI person-post matching intelligent recommendation method based on a graph neural network, characterized in that, The method comprises the following steps: S1, collecting multi-source heterogeneous human-post data, constructing content view, structure view and attribute view based on text semantics, interaction topology and label hash, and fusing to generate multi-view heterogeneous graph; S2, performing multi-objective joint modularity optimization based on the heterogeneous graph, weighting and aggregating the topology signal and applying spectral clustering constraint to divide the nodes into homogeneous communities, and outputting community membership matrix and community-level statistical feature vector; S3, identifying low-degree cold start nodes according to the community membership matrix, calculating the Mahalanobis distance between the node attribute features and the community centroid, selecting the feature vector of the target community with the smallest distance for mapping to generate an initial feature vector; S4, constructing an improved HetGNN model for non-cold start nodes, sampling neighbors using spectral entropy pruning and multi-view consistency constraint, aligning attribute features and knowledge graph embedding through optimal transport, and outputting fusion feature vector through heterogeneous attention aggregation of meta-path perception; S5, constructing a feature decoupling module guided by mutual information maximization, projecting the fusion feature vector into a matching latent space, suppressing irrelevant noise through mutual information constraint, extracting core discriminative features, and outputting decoupled feature vector; S6, constructing a reciprocal preference matching mechanism, calculating the bidirectional intention score of the job seeker and the post based on the decoupled feature vector, and jointly optimizing the bidirectional score using a reciprocal loss function to generate a revised matching score; S7, sorting the recommended posts in descending order according to the revised matching score and selecting a candidate set, calculating the reciprocal adaptation bias of the bidirectional score, removing nodes with bias exceeding the limit, and outputting the final human-post matching result.

2. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The S1 specifically comprises: S11, collecting multi-source heterogeneous human-post data, performing TF-IDF vectorization mapping on resume text and job description, calculating the cosine similarity of the text feature vector, selecting and constructing associated edges according to the similarity threshold, and generating a sparse adjacency matrix of the content view; S12, collecting multi-source heterogeneous human-post data, analyzing historical interaction logs to construct a bipartite graph topology, calculating the interaction frequency between nodes and performing normalization to determine edge weights, and generating an adjacency matrix of the structure view; S13, collecting multi-source heterogeneous human-post data, performing MinHash dimensionality reduction mapping on skill labels, estimating the Jaccard similarity coefficient between label sets based on hash signatures, and establishing attribute association edges according to the similarity coefficient to generate an adjacency matrix of the attribute view; S14, based on node index alignment logic, performing multi-channel feature fusion on the sparse adjacency matrix of the content view, the adjacency matrix of the structure view, and the adjacency matrix of the attribute view to construct a unified noisy multi-view heterogeneous graph.

3. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The S2 specifically comprises: S21, based on the noisy multi-view heterogeneous graph, traversing the adjacency data of the content view, the structure view and the attribute view respectively, calculating the node degree matrix of each view, and calculating the normalized Laplacian matrix according to the normalized Laplacian transformation algorithm, performing feature decomposition operation on the normalized Laplacian matrix, and extracting the feature vector corresponding to the pre-set number of smallest non-zero eigenvalues as the spectral domain topology signal feature of each view; S22, based on the spectral domain topology signal feature, the signal energy and noise energy ratio of each view is calculated, the view signal-to-noise ratio is obtained, and the correlation coefficient matrix of the feature vectors between different views is calculated as the inter-view mutual information; the weight distribution coefficient is calculated according to the view signal-to-noise ratio and the inter-view mutual information; the spectral domain topology signal features of each view are weighted and linearly fused by using the weight distribution coefficient, and a multi-view joint topology signal is generated; S23, a target optimization function containing Rayleigh quotient items is constructed, a second-order smoothness constraint regularization term is introduced into the target optimization function, a generalized eigenvector is solved by using the multi-view joint topology signal, and the maximum value of the target function is calculated by iteration under the condition of meeting the spectral clustering constraint, until the convergence condition is met, and the optimal community partition result is output; S24, the spectral embedding mapping is performed on the feature vector corresponding to the optimal community partition result, the posterior probability distribution of the nodes belonging to each community category is calculated, the soft allocation index is generated according to the posterior probability distribution, and the community membership matrix is constructed; the node attribute data is aggregated according to the index of the community membership matrix, the mean and variance statistics of the attribute data are calculated, and the community-level statistical feature vector is generated.

4. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The S3 specifically comprises: S31, the row vector distribution of the community membership matrix is analyzed, the maximum posterior probability value and the information entropy of each row are calculated, and the node index whose maximum posterior probability value is lower than a first preset threshold or whose information entropy is higher than a second preset threshold is marked as a cold start node set; S32, for the nodes in the cold start node set, the covariance matrix and the inverse matrix are calculated based on the node attribute features in each community, the difference between the node attribute features and the community centroid vector is weighted by using the inverse matrix of the covariance matrix, and the weighted Euclidean norm is calculated as the Mahalanobis distance; S33, a distance vector of the cold start node relative to all communities is constructed, a minimum value search is performed on the distance vector to locate the index corresponding to the minimum distance, and the community corresponding to the minimum distance index is determined as a target community; S34, the community-level statistical feature vector of the target community is extracted, the community-level statistical feature vector is projected to the attribute space dimension of the cold start node through linear mapping transformation, and the initial feature vector of the cold start node is generated.

5. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The improved HetGNN model comprises a graph structured entropy path pruning layer, a multi-view path consistency constraint layer, an optimal transport mapping alignment layer, a cross-modal hash fusion layer, a heterogeneous attention aggregation layer and a feature output layer: The graph structured entropy path pruning layer is used to traverse all second-order meta-paths in the content view, the structure view and the attribute view, calculate the Laplacian matrix spectral entropy of the subgraph formed by each meta-path, compare the calculated spectral entropy value with a preset noise threshold, eliminate the meta-path whose spectral entropy value is greater than or equal to the preset noise threshold, and output a candidate meta-path set meeting the preset spectral entropy threshold; The multi-view path consistency constraint layer is configured to receive a candidate meta-path set, perform walk sampling in each view to generate a neighbor node sequence, extract path latent representations of each view through a shared encoder, construct a contrastive learning loss function to maximize mutual information of path representations of the same node across views, constrain consistency of multi-view sampling distribution, filter nodes according to the consistency distribution, and output a set of neighbor nodes with semantic alignment; The optimal transport mapping alignment layer is configured to extract node attribute features in the set of neighbor nodes and entity embedding vectors of an external recruitment knowledge graph; The Wasserstein distance between the feature distribution and the embedding distribution is calculated, and the Sinkhorn algorithm is used to solve the optimal transport plane; the entity embedding vectors are projected onto the semantic space of the node attribute features through the optimal transport plane, the modal distribution difference is eliminated, and a knowledge-enhanced feature tensor after distribution-level alignment is output; The cross-modal hash fusion layer is configured to input the knowledge-enhanced feature tensor and the original node attribute features into a deep hash network respectively for binary coding, and generate binary hash codes of equal length; the Hamming distance between the two groups of hash codes is calculated to construct a cross-modal attention weight matrix, the hash codes are weighted and spliced using the attention weight matrix, and a hybrid feature vector fused with cross-modal semantics is output; The heterogeneous attention aggregation layer is configured to receive the hybrid feature vector as a node center feature, combine hybrid feature vectors of the set of neighbor nodes with semantic alignment, introduce a meta-path type-aware attention mechanism to calculate aggregation weights; weighted summation and nonlinear transformation are performed according to the aggregation weights, high-order structure information of neighbors is aggregated, the node hidden state is updated, and an aggregated node feature vector is output; The feature output layer is configured to input the aggregated node feature vector into a fully connected layer to perform dimension mapping, perform nonlinear mapping through a Tanh activation function, and perform L2 norm normalization processing on the output vector to generate a fusion feature vector with uniform length.

6. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The feature decoupling module includes a matching space projection layer, a mutual information orthogonal decoupling constraint layer, and a mask feature generation layer: The matching space projection layer is configured to construct a mapping function including multiple nonlinear transformations, map the input fusion feature vector to a matching latent subspace in a Hilbert space, and perform centering and standardization processing on the mapped feature vector to output a latent feature representation tensor; The mutual information orthogonal decoupling constraint layer is configured to construct an estimator based on a lower bound of mutual information, calculate an estimated value of mutual information between the latent feature representation tensor and the matching label; meanwhile, a feature covariance matrix is constructed and a Frobenius norm is calculated as a correlation penalty term, a joint loss function is constructed to maximize label mutual information and minimize feature correlation; the gradient descent is performed on the joint loss function to calculate an importance weight vector of the feature components; The mask feature generation layer is configured to construct a binary mask matrix according to the importance weight vector, apply the binary mask matrix to the latent feature representation tensor through matrix multiplication, suppress feature dimensions with weights below a preset threshold, perform tensor reshaping, and output a decoupled feature vector.

7. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The reciprocity preference matching mechanism comprises a bidirectional intention score calculation layer, a reciprocity preference constraint optimization layer and a revised score generation layer. The bidirectional intention score calculation layer is configured to construct a double-tower neural network, input decoupled feature vectors into a job seeker tower and a job tower respectively, calculate original hidden layer outputs through a multilayer perceptron, perform a Sigmoid nonlinear transformation on the original hidden layer outputs to map to a probability space, and output a first intention score of a job seeker for a job and a second intention score of a job for a job seeker respectively. The reciprocity preference constraint optimization layer is configured to construct a reciprocity consistency loss function, calculate a mean square error of the first intention score and the second intention score to quantify a bidirectional preference deviation, introduce a Frobenius norm to constrain alignment of bidirectional feature representations, weight and fuse the reciprocity consistency loss function and a cross-entropy loss function of a basic recommendation task, and optimize network parameters through gradient backpropagation. The revised score generation layer is configured to recalculate bidirectional intention scores based on the optimized network parameters, construct an attention adjustment factor based on a reciprocity intensity, perform weighted fusion on the bidirectional intention scores through the reciprocity adjustment factor, and output revised matching scores.

8. The AI person-post matching intelligent recommendation method based on a graph neural network according to claim 1, characterized in that, The S7 comprises the following steps: S71, receiving the revised matching scores, performing a descending arrangement operation on the to-be-recommended jobs according to the revised matching scores, selecting top-K job nodes in the front row to construct an initial candidate set; S72, traversing all nodes in the initial candidate set, extracting a first intention score of a job seeker for a job and a second intention score of a job for a job seeker corresponding to each node, and calculating an absolute value of a difference between the first intention score and the second intention score as a reciprocity adaptation deviation; S73, comparing the reciprocity adaptation deviation with a preset deviation threshold, marking the node as a one-way intention node and removing the node from the initial candidate set when the reciprocity adaptation deviation is greater than the deviation threshold, and retaining the node and outputting a final person-job matching result when the reciprocity adaptation deviation is less than or equal to the deviation threshold.

Images