Human-post matching method and system based on fusion graph neural network, and electronic equipment

Abstract

The invention discloses a person and post matching method and system based on a fusion graph neural network, and electronic equipment, and belongs to the technical field of natural language processing and graph neural network crossing. The method comprises the following steps: acquiring position description text data and job seeker resume text data, and preprocessing the position description text data and the job seeker resume text data; performing semantic coding on the preprocessed position description and resume text to generate a high-dimensional text feature vector; constructing graph structure data; performing node information propagation and fusion on the graph structure by using a graph neural network, and updating node features through a multilayer graph convolution operation; a deep learning matching degree calculation module is combined to calculate the comprehensive matching degree of the position and the job seeker; and outputting a matching result of the position and the resume. According to the method, complex position and job seeker information can be efficiently processed, and the precision and efficiency of position description and resume matching are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of natural language processing and graph neural networks. More specifically, it relates to a person - position matching method, system, electronic device and storage medium based on a fused graph neural network. Background Art

[0002] With the rapid development of Internet technology and the continuous expansion of the recruitment market, person - position matching has become a core link in human resource management and recruitment systems. Traditional person - position matching methods mainly rely on keyword matching, rule engines, and simple machine learning models. Although these methods can meet basic needs to a certain extent, they have obvious deficiencies in dealing with complex semantic relationships and deep - level features. For example, the keyword matching method cannot understand the context semantics of the text, resulting in low - precision matching results; the rule engine requires a large number of manually defined rules and is difficult to adapt to diverse job and resume descriptions; while traditional machine learning models are limited by the quality of feature engineering and are difficult to capture the deep - level associations between positions and job seekers.

[0003] In recent years, deep learning technology has made remarkable progress in the fields of natural language processing (NLP) and graph neural networks (GNN), providing new solutions for person - position matching. Transformer learns the semantic representation of text through large - scale corpora, significantly improving the accuracy of text encoding; graph neural networks can effectively model complex relational networks and mine deep - level semantic associations through node information propagation and fusion. However, there are still the following problems in the existing technologies: First, a single text encoding method is difficult to fully capture the multi - dimensional features between positions and job seekers; second, traditional graph neural networks perform limitedly in dealing with heterogeneous information and cannot effectively fuse text features and structural features; finally, existing matching degree calculation methods mostly rely on simple similarity metrics and fail to comprehensively consider local features and global context relationships. Summary of the Invention

[0004] The purpose of the present invention is to provide a person - position matching method, system and electronic device based on a fused graph neural network, so as to solve at least one of the technical problems mentioned in the background art existing in the existing person - position matching methods. Compared with the existing technologies, the present invention can process complex position and job seeker information more efficiently, is applicable to fields such as automated recruitment systems, job search recommendation platforms, and human resource management systems, and has broad application prospects.

[0005] To achieve the above - mentioned purpose, the technical solution provided by the present invention is as follows:

[0006] The first aspect of the present invention provides a person - position matching method based on a fused graph neural network, including:

[0007] Obtain the job description text data and the job seeker's resume text data, and preprocess them to obtain the preprocessed input sequence;

[0008] Use the Transformer model to perform semantic encoding on the preprocessed job description and resume text (the preprocessed input sequence), extract keywords, skill requirements, and experience information, and generate high-dimensional text feature vectors;

[0009] Represent the key information of the job (the preprocessed input sequence data, such as the job title, department, etc.) as a graph structure, where the nodes represent the features of the job or the job seeker, and the edges represent the semantic relationships or associations between the nodes, and construct the graph structure data;

[0010] Use the graph neural network to perform node information propagation and fusion on the graph structure, and update the node features through multi-layer graph convolution operations;

[0011] Based on the output features of the Transformer and the graph neural network, combined with the deep learning matching degree calculation module, calculate the comprehensive matching degree between the job and the job seeker;

[0012] Output the matching result of the job and the resume.

[0013] The present invention combines the advantages of the pre-trained language model and the graph neural network, and through technical means such as semantic encoding, graph structure modeling, node information propagation and fusion, and comprehensive matching degree calculation, significantly improves the accuracy and efficiency of job description and resume matching.

[0014] According to any of the technical solutions described in the first aspect of the present invention, preprocess the job description text data and the job seeker's resume text data, specifically including: performing word segmentation, denoising, and standardization processing on the text data to generate structured input data.

[0015] According to any of the technical solutions described in the first aspect of the present invention, the use of the Transformer model to perform semantic encoding on the preprocessed job description and resume text, extract keywords, skill requirements, and experience information, and generate high-dimensional text feature vectors specifically includes:

[0016] Map the input sequence to a word vector representation through the embedding layer of the Transformer model;

[0017] Use multiple layers of Transformer encoders to perform context semantic encoding on the word vectors to generate high-dimensional text feature vectors; wherein, the self-attention mechanism of the Transformer encoder is calculated by the following improved formula:

[0018]

[0019] In the formula:

[0020] Q, K, and V represent the query matrix, the key matrix, and the value matrix respectively;

[0021] d k is the dimension of the key vector;

[0022] The softmax function is used to calculate the attention weights;

[0023] η(Q, K) is the dynamic weight factor based on the query matrix and the key matrix, and its calculation formula is:

[0024]

[0025] In the above formula:

[0026] represents the square of the Euclidean distance between the query matrix Q and the key matrix K;

[0027] σ is an adjustable scale parameter used to control the speed of weight decay.

[0028] According to any of the technical solutions described in the first aspect of the present invention, when constructing graph-structured data, the edge weights between nodes are calculated by the following formula:

[0029]

[0030] In the formula;

[0031] w uv represents the edge weight between node u and node v;

[0032] sim(h u , h v ) represents the cosine similarity between the feature vectors of node u and node v;

[0033] represents the set of neighbor nodes of node u;

[0034] λ(u, v) is the dynamic adjustment factor based on node features;

[0035] is the additional weight factor based on the node context relationship.

[0036] According to any of the technical solutions described in the first aspect of the present invention, the calculation of the dynamic adjustment factor λ(u, v) based on node features is as follows:

[0037]

[0038] In the formula: α is an adjustable parameter used to control the influence of feature differences on the edge weights;

[0039] and / or an additional weight factor based on the node context relationship is calculated as follows:

[0040]

[0041] In the formula:

[0042] and respectively represent the i-th and j-th components of the feature vectors of nodes u and v;

[0043] deg(u) and deg(v) respectively represent the degrees of nodes u and v.

[0044] Through the introduction of λ(u, v) and , the constructed graph structure can more accurately reflect the complex semantic relationship between positions and job seekers, providing high-quality input data for person-job matching.

[0045] According to any one of the technical solutions described in the first aspect of the present invention, the use of a graph neural network to perform node information propagation and fusion on the graph structure, and update node features through multi-layer graph convolution operations, specifically includes:

[0046] First, perform multi-layer graph convolution operations on the graph structure data, and update node features through the following formula:

[0047]

[0048] In the formula:

[0049] GELU is the Gaussian error linear unit activation function;

[0050] H (l+1) represents the node feature matrix of the (l + 1)-th layer;

[0051] is the adjacency matrix with self-loops added;

[0052] is the degree matrix of;

[0053] W (l) is the learnable weight matrix of the l-th layer;

[0054] Gate(H (l) ,H (0) ) is a gating mechanism for fusing the current layer feature H (l) and the initial layer feature H (0) ;

[0055] β is an adjustable fusion weight parameter for controlling the contribution of the gating mechanism;

[0056] Secondly, regularize the node features to prevent overfitting:

[0057] H = LayerNorm(H (l+1) + Dropout(H (l+1) , p))

[0058] Where: LayerNorm represents the layer normalization operation; Dropout represents the random dropout operation; p is the dropout probability;

[0059] Finally, update the node features according to the results of the regularization process to generate high-quality node feature representations, providing deep semantic information for person-job matching.

[0060] Furthermore, the calculation formula of Gate(H (l) , H (0) ) is:

[0061] Gate(H (l) , H (0) ) = σ(U (l) H (l) + U (0) H (0) ) ⊙ H (0)

[0062] Where: U (l) and U (0) are learnable weight matrices; ⊙ represents element-wise multiplication.

[0063] According to any one of the technical solutions described in the first aspect of the present invention, the comprehensive matching degree between the position and the job seeker is calculated as follows:

[0064]

[0065] Where:

[0066] P represents the set of position features; C represents the set of job seeker features; h pi and h cj respectively represent the feature vectors of the position feature p i and the job seeker feature c j ; α ij is the attention weight of the feature pair (pi, cj); sim(h pi , h cj ) is the cosine similarity between the feature vectors; ψ(h pi , h cj) is the dynamic adjustment factor based on the feature vector; γ is the learnable weight parameter used to balance the contributions of local similarity and global alignment factor; ContextAlign(P, C) is the context-based alignment factor used to capture the global semantic relationship between the position and the job seeker.

[0067] Furthermore, the calculation formula of ψ(h pi , h cj ) is:

[0068]

[0069] In the formula: represents the Euclidean distance between feature vectors; σ is the adjustable scale parameter.

[0070] Furthermore, the calculation formula of ContextAlign(P, C) is:

[0071]

[0072] In the formula: deg(p i ) and deg(c j ) respectively represent the degrees of the position feature p i and the job seeker feature c j .

[0073] By optimizing the calculation formula of the comprehensive matching degree, it is possible to comprehensively consider the local feature similarity and global context relationship, generate a more accurate comprehensive matching degree score, and thus provide a high-quality decision-making basis for the person-job matching.

[0074] According to any one of the technical solutions described in the first aspect of the present invention, the matching result of the output position and resume specifically includes:

[0075] Normalize the matching score generated by the matching degree calculation module and map the matching score to the interval [0, 1];

[0076] Sort the positions and job seekers according to the normalized matching score to generate a recommendation list.

[0077] The second aspect of the present invention provides a person-job matching system based on a fusion graph neural network, including:

[0078] A data acquisition and preprocessing module for acquiring job description text data and job seeker resume text data and preprocessing them;

[0079] A semantic encoding and feature extraction module for semantically encoding the preprocessed job description and resume text using a Transformer model, extracting keywords, skill requirements, and experience information, and generating high-dimensional text feature vectors;

[0080] A graph structure data construction module, which is used to represent the key information of a position as a graph structure, where nodes represent the characteristics of positions or job seekers, and edges represent the semantic relationships or associations between nodes, and construct graph structure data;

[0081] A node information propagation and fusion module, which is used to perform node information propagation and fusion on the graph structure by using a graph neural network, and update the node features through multi-layer graph convolution operations;

[0082] An overall matching degree calculation module, which is used to calculate the overall matching degree between a position and a job seeker based on the output features of a Transformer and a graph neural network, in combination with a deep learning matching degree calculation module;

[0083] An output module, which is used to output the matching result between a position and a resume.

[0084] The third aspect of the present invention provides an electronic device, including a processor and a memory, and a computer program is stored on the memory. When the processor executes the computer program, the person-job matching method described in the first aspect of the present invention can be implemented.

[0085] The fourth aspect of the present invention provides a computer-readable storage medium, and the storage medium stores a computer program, and when the computer program is executed, the person-job matching method described in the first aspect of the present invention can be implemented.

[0086] Adopting the technical solution provided by the present invention, compared with the prior art, the following beneficial effects can be obtained:

[0087] (1) The present invention proposes a person-job matching method based on the fusion of graph neural networks. By fusing the Transformer model and graph neural networks, the matching accuracy of job descriptions and resumes can be effectively improved; among them, the Transformer model is good at processing complex semantics in natural language texts, while graph neural networks can better process structured relational data. The combination of the two makes up for their respective deficiencies and greatly improves the matching effect.

[0088] (2) Based on the output features of the Transformer and graph neural networks, the present invention combines a deep learning matching degree calculation module to calculate the overall matching degree between a position and a job seeker. By optimizing the attention mechanism and context alignment factor, the local feature similarity and global semantic relationship can be comprehensively considered, which is beneficial to generating a more accurate overall matching degree score.

[0089] (3) In the process of constructing and representing the graph structure of the present invention, the calculation formula of the edge weight is further optimized, especially by introducing a dynamic adjustment factor based on node features and an additional weight factor based on node context relationships. As a result, not only can the problem that the traditional cosine similarity may overestimate features that are seemingly similar but actually mismatched be effectively solved, the strong connection of irrelevant nodes can be avoided, the robustness of the model to noise can be enhanced, and the interference of outliers to the graph structure can be prevented. At the same time, the fine-grained similarity of the feature vectors can be combined to supplement the global relationships that cannot be captured by local similarity, alleviate the sparsity problem, and ensure that low-frequency but key nodes can still reasonably participate in the matching.

[0090] (4) The technical framework of the present invention has strong versatility and can be widely applied to automated recruitment systems, intelligent job search recommendation platforms, and various human resource management systems, helping enterprises quickly and accurately screen out suitable candidates. Its application can not only improve recruitment efficiency but also reduce recruitment costs and enhance the job search experience of job seekers. Description of the Drawings

[0091] Figure 1 is a schematic flow chart of the person-job matching method based on the fusion graph neural network according to the embodiment of the present invention;

[0092] Figure 2 is an analysis diagram of the influence of the scale parameter σ on the matching degree score in the embodiment of the present invention. Detailed Embodiments

[0093] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0094] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0095] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0096] In combination withFigure 1 As shown in Figure 1 , the person-job matching method based on the fusion graph neural network according to the embodiment of the present invention includes:

[0097] S1. Obtain the job description text data and the job seeker's resume text data, and preprocess them to obtain the preprocessed input sequence.

[0098] The job description text data includes information such as job responsibilities, skill requirements, work location, and salary range; the job seeker's resume text data includes information such as educational background, work experience, skill description, and project experience. The specific data can be sourced from the human resource management system.

[0099] According to the embodiment of the present invention, the preprocessing of the job description text data and the job seeker's resume text data specifically includes: performing word segmentation, denoising, and normalization on the text data to generate structured input data.

[0100] Specifically, in the embodiment of the present invention, the word segmentation tool NLTK is used to perform word segmentation on the text.

[0101] For example, the job description "Requires proficiency in Python programming and has more than 5 years of relevant experience" will be segmented into ["Requires", "Proficient in", "Python", "Programming", "Has", "5 years", "Above", "Relevant", "Experience"]; the resume text "Proficient in Python programming, has 3 years of relevant experience, and graduated from Tsinghua University." will be segmented into ["Proficient in", "Python", "Programming", "Has", "3 years", "Relevant", "Experience"].

[0102] Then, the segmented text is converted into a fixed-length input sequence, and the length of the input sequence is set to 512 words. If the text length is insufficient, the special symbol "[PAD]" is used for padding; if the text length exceeds 512 words, it is truncated.

[0103] S2. Semantic encoding and feature extraction

[0104] The Transformer model is used to perform semantic encoding on the preprocessed job description and resume text, extract keywords, skill requirements, and experience information, and generate high-dimensional text feature vectors.

[0105] Semantic encoding and feature extraction are one of the core steps of the present invention. According to the embodiment of the present invention, it includes the following steps:

[0106] S21. Map the input sequence into a word vector representation;

[0107] According to an embodiment of the present invention, the input sequence is mapped into a word vector representation through the embedding layer of the Transformer model. Specifically, a pre-trained Transformer model is used to load the vocabulary, and then each word in the input sequence is embedded to generate a word vector sequence. For example, the input sequence ["Need", "Proficient in", "Python", "Programming"] will be mapped into a corresponding 768-dimensional word vector sequence.

[0108] The word "Python" in the job description is mapped into a 768-dimensional word vector:

[0109] H PPython =[0.12, -0.05, 0.23, …, 0.08]

[0110] The word "Python" in the resume text is mapped into a 768-dimensional word vector:

[0111] H CPython =[0.15, -0.02, 0.20, …, 0.10]

[0112] S22. Context semantic encoding: Use multiple layers of Transformer encoders to perform context semantic encoding on the word vectors to generate high-dimensional text feature vectors.

[0113] Preferably, in the embodiment of the present invention, the calculation formula of the self-attention mechanism is:

[0114]

[0115] In the formula:

[0116] Q, K, and V represent the query matrix, key matrix, and value matrix respectively;

[0117] d k is the dimension of the key vector;

[0118] The softmax function is used to calculate the attention weights;

[0119] η(Q, K) is a dynamic weight factor based on the query matrix and key matrix. Further, the calculation formula of η(Q, K) is:

[0120]

[0121] In the above formula:

[0122] represents the square of the Euclidean distance between the query matrix Q and the key matrix K;

[0123] σ is an adjustable scale parameter used to control the speed of weight decay. Specifically, in the embodiments of the present invention, σ ∈ [0.1, 2.0], and σ is determined according to the distance distribution of the feature vectors of the input data, that is, the median of the Euclidean distances of all query-key pairs is taken to adapt σ to the data scale.

[0124] Finally, the output of the self-attention mechanism is input into a multi-layer Transformer encoder to gradually extract the deep semantic features of the text. The output of each layer is used as the input of the next layer, and finally a high-dimensional text feature vector is generated.

[0125] Traditional self-attention mechanisms only calculate similarity through dot products (the transpose of Q*K). On the one hand, it may mask the subtle differences between features; on the other hand, it cannot dynamically adapt to the distribution characteristics of different feature pairs, resulting in the weakening of some key features. By optimizing the calculation formula of the self-attention mechanism, after introducing η(Q, K), the Euclidean distance can be combined to measure differences more comprehensively; at the same time, through the dynamic weight factor, the contribution of similar feature pairs can be highlighted, and the interference of irrelevant feature pairs can be suppressed.

[0126] S23. Feature vector generation: Through the above steps, high-dimensional text feature vectors of the job description and resume text are generated. The job description "Requires proficiency in Python programming and has more than 5 years of relevant experience" and the resume text "Proficient in Python programming and has 3 years of relevant experience" respectively generate 768-dimensional feature vectors.

[0127] S3. Based on the preprocessed input sequence, graph structure construction and representation are performed.

[0128] Graph structure construction and representation are one of the key steps of the present invention, aiming to represent the key information of the job description and resume (information in the preprocessed input sequence, such as: job title, department, etc.) as a graph structure, where nodes represent the features of the job or the job seeker, and edges represent the semantic relationships between nodes.

[0129] Furthermore, in some embodiments, the graph structure construction and representation include:

[0130] S31. Node definition

[0131] In the embodiments of the present invention, node types include job feature nodes and job seeker feature nodes, where:

[0132] Job feature nodes: Include information such as skill requirements, work location, salary range, etc. For example, in the job description "Requires proficiency in Python programming, has more than 5 years of relevant experience, work location in Beijing, salary range 20 - 30K", the skill requirement "Python programming", work location "Beijing", and salary range "20 - 30K" are used as nodes respectively.

[0133] Job seeker feature nodes: Include information such as skills, experience, educational background, etc. For example, in the resume text "Proficient in Python programming, with 3 years of relevant experience, graduated from Tsinghua University", the skill "Python programming", the experience "3 years", and the educational background "Tsinghua University" are used as nodes respectively.

[0134] Among them, the feature vector of each node is generated through semantic encoding and feature extraction steps; the feature vector corresponding to the skill requirement "Python programming" is h Python , and the feature vector corresponding to the work location "Beijing" is h 北京 .

[0135] S32. Edge weight calculation: Calculate the semantic similarity between node feature vectors through cosine similarity to construct edge weights.

[0136] Specifically, for each node pair (u, v), first calculate the cosine similarity between its feature vectors h u and h v :

[0137]

[0138] In the formula:

[0139] h u ·h v represents the dot product of vectors;

[0140] ||h u || and ||h v || represent the magnitudes of vectors.

[0141] Then calculate the edge weight according to the cosine similarity:

[0142]

[0143] In the formula;

[0144] w uv represents the edge weight between node u and node v;

[0145] sim(h u , h v ) represents the cosine similarity between the feature vectors of node u and node v;

[0146] represents the set of neighbor nodes of node u;

[0147] λ(u, v) is a dynamic adjustment factor based on node features;

[0148] is an additional weight factor based on the node context relationship;

[0149] In some embodiments, the dynamic adjustment factor λ(u, v) based on node features is preferably calculated as follows:

[0150]

[0151] where: α is an adjustable parameter used to control the influence of feature differences on the edge weight.

[0152] Traditional cosine similarity may overestimate features that are superficially similar but actually mismatched (e.g., both "Beijing" and "Shanghai" are locations, but they should be distinguished when matching); at the same time, it also enhances the robustness of the model to noise and prevents outliers (such as exaggerated descriptions in a resume) from interfering with the graph structure. Through the introduction and optimized design of the dynamic adjustment factor λ(u, v) based on node features, on the one hand, when the node feature differences are large (e.g., "5 years of experience" vs. "1 year of experience"), λ(u, v) approaches 0, thereby reducing the edge weight and effectively avoiding strong connections between irrelevant nodes; on the other hand, when the node features are similar (e.g., "Python" vs. "Python programming"), λ(u, v) approaches 1, and the contribution of the original similarity can be retained.

[0153] In addition, local similarity (such as cosine similarity) may also ignore the global structure information of nodes (such as the implicit association between "Tsinghua University" and "985 degree"), and there is a sparsity problem. Based on this situation, in some embodiments, an additional weight factor based on the node context relationship is introduced and optimized, so that the importance of nodes in the graph can be directly measured by the node degree (deg). Nodes with a high degree (such as the high-frequency skill "Python") contribute more to the edge weight; at the same time, combining the fine-grained similarity of feature vectors can supplement the global relationships that local similarity cannot capture, effectively alleviating the sparsity problem and ensuring that low-frequency but key nodes (such as niche skills) can still participate in the matching reasonably.

[0154] Specifically, the additional weight factor based on the node context relationship is calculated as follows:

[0155]

[0156] where:

[0157] and represent the i-th and j-th components of the feature vectors of nodes u and v, respectively;

[0158] deg(u) and deg(v) represent the degrees of nodes u and v, respectively.

[0159] For example, the edge weight between the nodes Python Programming (position) and Python Programming (job seeker) is calculated by the formula:

[0160]

[0161] Edge weight between nodes 5 years experience (Job position) and 3 years experience (Job seeker):

[0162]

[0163] S33, graph structure construction: The key information of the job description and resume obtained after preprocessing in S1 is represented as a graph structure, where nodes represent the characteristics of positions or job seekers, and edges represent the semantic relationships between nodes.

[0164] S4. Information propagation and fusion of graph neural networks

[0165] The information propagation and fusion of graph neural networks aims to update node features through multi-layer graph convolution operations, enhance semantic understanding capabilities, explore deep semantic associations between nodes, and generate high-quality node feature representations.

[0166] According to any embodiment of the present invention, the graph neural network information propagation and fusion specifically includes:

[0167] First, perform multi-layer graph convolution operations on the graph structure data and update the node features using the following formula:

[0168]

[0169] Where:

[0170] GELU is the Gaussian error linear unit activation function;

[0171] H (l+1) Represents the node feature matrix of the l+1th layer;

[0172] is the adjacency matrix with self-loops added;

[0173] for The degree matrix of

[0174] W (l) is the learnable weight matrix of the lth layer;

[0175] Gate(H (l) ,H (0) ) is a gating mechanism used to fuse the current layer feature H (l) and the initial layer features H (0) ;

[0176] β is an adjustable fusion weight parameter used to control the contribution of the gating mechanism.

[0177] It should be noted that traditional graph convolution operations usually use the ReLU function. ReLU completely truncates negative values, which may lead to information loss; GELU can retain some negative value information through a smooth transition and is more suitable for dealing with complex patterns in text and graph data. At the same time, in deep graph networks, ReLU may cause neurons to "die" (gradient is 0), while the continuous differentiability of GELU alleviates the above gradient vanishing problem. In addition, GELU can better fuse node features with the context information of their neighbors, especially suitable for dealing with the long-tail distribution in job-resume matching (such as low-frequency skills or rare experiences).

[0178] More preferably, the calculation formula of Gate(H (l) ,H (0) ) is:

[0179] Gate(H (l) ,H (0) ) = sigmoid (l) H (l) + U (0) H (0) ) ⊙ H (0)

[0180] In the formula, U (l) and U (0) are learnable weight matrices, and ⊙ represents element-wise multiplication.

[0181] In traditional GNNs, after multiple convolutions, all node features may converge (for example, "Python programming" and "Java programming" become indistinguishable); at the same time, directly stacking graph convolution layers will cause the original input features (such as skill keywords) to gradually decay. However, through the optimization of the gating mechanism in the present invention, the initial feature differences can be retained, and the semantic distinguishability of nodes can be maintained, that is, different nodes can adopt different feature fusion strategies. The gating is adaptively adjusted through the learnable parameters U (l) and U (0) . At the same time, the gating explicitly transmits the initial features, which can avoid the deep network from "forgetting" important matching conditions (such as hard skill requirements) and explicitly retain the key initial features.

[0182] Secondly, the node features are regularized to prevent overfitting:

[0183] H = LayerNorm(H (l+1) + Dropout(H (l+1) , p))

[0184] In the formula: LayerNorm represents the layer normalization operation; Dropout represents the random dropout operation; p is the dropout probability.

[0185] Finally, update the node features according to the results after regularization processing.

[0186] Specifically, in the embodiments of the present invention, first initialize, input the graph structure data into the graph neural network, and initialize the node feature matrix H (0) Specifically as follows:

[0187]

[0188] Then perform the first-layer graph convolution operation on the graph structure data to update the node feature matrix H (1) :

[0189]

[0190] Among them, W (0) is the learnable weight matrix of the first layer.

[0191] Perform the second-layer graph convolution operation on the graph structure data to update the node feature matrix H (2) :

[0192]

[0193] Among them, W (1) is the learnable weight matrix of the second layer.

[0194] Perform the third-layer graph convolution operation on the graph structure data to update the node feature matrix H (3) :

[0195]

[0196] Among them, W (1) is the learnable weight matrix of the third layer.

[0197] Through the three-layer graph convolution operation, the dimension of the node feature matrix H (3) remains unchanged, but its semantic representation ability is significantly improved. In the embodiments of the present invention, the final node feature matrix H (3) is as follows:

[0198]

[0199] S5. Matching degree calculation and comprehensive evaluation

[0200] In some embodiments, based on the output features of the Transformer and the graph neural network, combined with the deep learning matching degree calculation module, calculate the comprehensive matching degree between the position and the job seeker.

[0201] Furthermore, its steps include:

[0202] S51. Calculate the attention weights of each feature pair based on the attention mechanism

[0203] For the job feature set P and the job seeker feature set C, calculate the attention weight of each feature pair (p i , c j ). The specific formula is:

[0204]

[0205] In the formula:

[0206] α ij represents the attention weight of the feature pair (p i , c j ), and is used to measure the correlation between the job feature p i and the job seeker feature c j ;

[0207] h pi and h cj respectively represent the feature vectors of the job feature p i and the job seeker feature c j ;

[0208] sim(h pi , h cj ) is the cosine similarity between feature vectors, and the calculation formula is:

[0209]

[0210] Through the attention mechanism, the weights of different feature pairs can be dynamically adjusted to highlight the contributions of key features, thereby improving the accuracy of match degree calculation.

[0211] S52. Comprehensive match degree calculation

[0212] Based on the attention weights and feature similarities, calculate the comprehensive match degree between the job and the job seeker. The specific formula is:

[0213]

[0214]

[0215] In the formula:

[0216] P represents the job feature set;

[0217] C represents the job seeker feature set;

[0218] h pi and h cjrespectively represent the job feature p i and the job seeker feature c j of the feature vectors;

[0219] α ij is the attention weight of the feature pair (p i , c j );

[0220] sim(h pi , h cj ) is the cosine similarity between the feature vectors;

[0221] ψ(h pi , h cj ) is the dynamic adjustment factor based on the feature vectors;

[0222] γ is a learnable weight parameter used to balance the contributions of local similarity and global alignment factor;

[0223] ContextAlign(P, C) is the context-based alignment factor used to capture the global semantic relationship between the job and the job seeker.

[0224] Among them, the calculation formula of ψ(h pi , h cj ) is:

[0225]

[0226] In the formula:

[0227] represents the Euclidean distance between the feature vectors;

[0228] σ is an adjustable scale parameter.

[0229] Exaggerated descriptions in the resume (such as writing "familiar" as "proficient") may lead to inflated similarity, while ψ can effectively suppress unreasonable matches through distance penalty. In addition, in comprehensive matching, the importance of different features (such as skills, experience, education) is different, and ψ allows the model to flexibly adapt to the strictness requirements of different features through σ.

[0230] Furthermore, the calculation formula of ContextAlign(P, C) is:

[0231]

[0232] In the formula:

[0233] deg(p i ) and deg(c j ) respectively represent the job feature p i and the job seeker feature c jDegree.

[0234] Traditional methods only accumulate local feature similarities and may ignore the rationality of overall matching. ContextAlign, from a global perspective, can avoid overgeneralization and also consider the synergistic effects of all features. For example, it can judge the differences between the combinations of "Python + SQL" and "Java + NoSQL". Therefore, it can also avoid the one-sidedness of local matching.

[0235] In summary, through the optimized design of the comprehensive matching degree calculation formula, it is possible to comprehensively consider local feature similarities and global context relationships, generate a more accurate comprehensive matching degree score, and provide a high-quality decision-making basis for the job-person matching.

[0236] As Figure 2 shown, it demonstrates the influence of different scale parameters σ on the matching degree. Experiments show that when the value of σ is 0.3, the matching degree reaches the peak value of 0.5, indicating that the balance effect between the dynamic adjustment factor and the cosine similarity is the best under this parameter. This result verifies the sensitivity and robustness of the method of the present invention to parameters and supports optimizing the matching effect by adjusting the value of σ.

[0237] S6, Generation and Output of Matching Results

[0238] According to the matching scores generated by the matching degree calculation module, sort the positions and job seekers, generate a recommended list, and provide a reliable basis for recruitment decisions.

[0239] Furthermore, the specific operations are as follows:

[0240] Normalize the matching scores to ensure the fairness and comparability of the matching results:

[0241]

[0242] In the formula:

[0243] MatchScore(P, C) is the original matching score;

[0244] min(MatchScore) and max(MatchScore) are respectively the minimum and maximum values of all matching scores;

[0245] ζ(P, C) is the dynamic adjustment factor based on the characteristics of the position and the job seeker.

[0246] More preferably, the calculation formula of ζ(P, C) is as follows:

[0247]

[0248] In the formula:

[0249] sim(hpi , h cj ) is the cosine similarity between feature vectors;

[0250] deg(p i ) and deg(c j ) respectively represent the degrees of the job feature p i and the job seeker feature c j ;

[0251] n and m respectively represent the numbers of job features and job seeker features.

[0252] Through normalization, the matching scores are mapped to the interval [0, 1], so as to ensure the comparability of the matching scores between different jobs and job seekers. According to the normalized matching scores, the jobs and job seekers are sorted to generate a recommendation list. The recommendation list includes the job name, the matching score, and the key skill matching situation.

[0253] Traditional [min, max] normalization treats all features equally and ignores the particularity of key conditions; moreover, the fixed normalization interval cannot adapt to the matching strictness requirements of different job types; in the embodiments of the present invention, the introduction of a dynamic adjustment factor can respond to the matching quality of different feature combinations in real time, highlighting the decisive role of core requirements, and making the score distribution more in line with the actual recruitment decision logic.

[0254] In the embodiments of the present invention, after calculating the matching scores of each job and each job seeker through S5, the normalized matching scores are then calculated:

[0255]

[0256] Similarly, the normalized values of other matching scores are calculated.

[0257] Finally, the recommendation list is output as shown in the following table:

[0258]

[0259] The embodiments of the present invention also provide a job-person matching system based on a fusion graph neural network, including:

[0260] A data acquisition and preprocessing module, configured to acquire job description text data and job seeker resume text data, and preprocess them;

[0261] A semantic encoding and feature extraction module, configured to perform semantic encoding on the preprocessed job description and resume text by using a Transformer model, extract keywords, skill requirements, and experience information, and generate high-dimensional text feature vectors;

[0262] A graph structure data construction module, which is used to represent the key information of a position as a graph structure, where nodes represent the characteristics of positions or job seekers, and edges represent the semantic relationships or associations between nodes, and construct graph structure data;

[0263] A node information propagation and fusion module, which is used to perform node information propagation and fusion on the graph structure by using a graph neural network, and update node features through multi-layer graph convolution operations;

[0264] A comprehensive matching degree calculation module, which is used to calculate the comprehensive matching degree between a position and a job seeker based on the output features of a Transformer and a graph neural network, in combination with a deep learning matching degree calculation module;

[0265] An output module, which is used to output the matching result between a position and a resume.

[0266] It should be noted that the specific working processes and working principles of the above functional modules in the embodiments of the present invention are the same as those described in the previous person-position matching method, and will not be elaborated here.

[0267] The embodiments of the present invention also provide an electronic device, including a processor and a memory, and a computer program is stored on the memory. When the processor executes the computer program, the person-position matching method of the present invention can be implemented.

[0268] The embodiments of the present invention also provide a computer-readable storage medium, and the storage medium stores a computer program, and when the computer program is executed, the person-position matching method of the present invention can be implemented.

Claims

1. A person-job matching method based on a fused graph neural network, characterized in that, Including: Obtain the job description text data and the job seeker's resume text data, and preprocess them to obtain the preprocessed input sequence; Use the Transformer model to perform semantic encoding on the preprocessed input sequence, extract keywords, skill requirements, and experience information, and generate a high-dimensional text feature vector; Represent the key information of the job as a graph structure, where the nodes represent the features of the job or the job seeker, and the edges represent the semantic relationships or associations between the nodes, and construct the graph structure data; Use the graph neural network to perform node information propagation and fusion on the graph structure, and update the node features through multiple graph convolution operations; Based on the output features of the Transformer and the graph neural network, combined with the deep learning matching degree calculation module, calculate the comprehensive matching degree between the job and the job seeker; Output the matching result of the job and the resume.

2. The person-job matching method according to claim 1, wherein Preprocess the job description text data and the job seeker's resume text data, specifically including: performing word segmentation, denoising, and standardization processing on the text data to generate structured input data.

3. The person-job matching method according to claim 1, wherein The step of using the Transformer model to perform semantic encoding on the preprocessed job description and resume text, extract keywords, skill requirements, and experience information, and generate a high-dimensional text feature vector specifically includes: Map the input sequence to a word vector representation through the embedding layer of the Transformer model; Use multiple layers of Transformer encoders to perform context semantic encoding on the word vectors to generate a high-dimensional text feature vector; among them, the self-attention mechanism of the Transformer encoder is calculated by the following improved formula: In the formula: Q, K, and V represent the query matrix, key matrix, and value matrix respectively; d k is the dimension of the key vector; The softmax function is used to calculate the attention weights; η(Q, K) is a dynamic weight factor based on the query matrix and the key matrix, and its calculation formula is: In the above formula: Represents the square of the Euclidean distance between the query matrix Q and the key matrix K; σ is an adjustable scale parameter used to control the speed of weight decay.

4. The person-job matching method according to claim 1, characterized in that When constructing the graph structure data, calculate the edge weights between nodes through the following formula: In the formula; w uv represents the edge weight between node u and node v; sim(h u ,h v ) represents the cosine similarity between the feature vectors of node u and node v; Denote the set of neighbor nodes of node u; λ(u, v) is a dynamic adjustment factor based on node features; is an additional weight factor based on the node context relationship.

5. The person-job matching method according to claim 4, characterized in that The calculation of the dynamic adjustment factor λ(u, v) based on node features is as follows: In the formula: α is an adjustable parameter used to control the influence of feature differences on the edge weights; and / or an additional weight factor based on the node context relationship is calculated as follows: In the formula: and represent the i-th and j-th components of the feature vectors of nodes u and v, respectively; deg(u) and deg(v) represent the degrees of nodes u and v respectively.

6. The person-job matching method according to any one of claims 1-5, characterized in that, The step of using the graph neural network to perform node information propagation and fusion on the graph structure, and update the node features through multiple graph convolution operations specifically includes: First, perform multiple graph convolution operations on the graph structure data, and update the node features through the following formula: In the formula: GELU is the Gaussian error linear unit activation function; H (l+1) represents the node feature matrix of the (l + 1)-th layer; is the adjacency matrix with self-loops added; is degree matrix of; W (l) is the learnable weight matrix for the l-th layer; Gate(H (l) ,H (0) ) is a gating mechanism for fusing the features H of the current layer (l) and the features H of the initial layer (0) ; β is an adjustable fusion weight parameter used to control the contribution of the gating mechanism; Secondly, perform regularization processing on the node features to prevent overfitting: H = LayerNorm(H (l+1) + Dropout(H (l+1) , p)) In the formula: LayerNorm represents the layer normalization operation; Dropout represents the random dropout operation; p is the dropout probability. Finally, update the node features according to the results of the regularization processing.

7. The person-job matching method according to claim 6, wherein The calculation of the comprehensive matching degree between the job and the job seeker is as follows: In the formula: Let \(P\) denote the set of job characteristics; \(C\) denote the set of job seeker characteristics; \(h\) pi and \(h\) cj denote the feature vectors of job characteristic \(p\) i and job seeker characteristic \(c\) j respectively; \(\alpha\) ij is the attention weight of the feature pair \((p_i, c_j)\); \(\text{sim}(h\) pi , h\) cj ) is the cosine similarity between feature vectors; \(\psi(h\) pi , h\) cj ) is the dynamic adjustment factor based on feature vectors; \(\gamma\) is a learnable weight parameter used to balance the contributions of local similarity and global alignment factor; \(\text{ContextAlign}(P, C)\) is the context-based alignment factor used to capture the global semantic relationship between the job and the job seeker.

8. The person-job matching method according to claim 7, wherein The matching results of the output position and the resume specifically include: Normalize the matching scores generated by the matching degree calculation module and map the matching scores to the interval [0, 1]; Sort the positions and job seekers according to the normalized matching scores to generate a recommendation list.

9. A person-job matching system based on a fusion graph neural network, characterized in that, It includes: A data acquisition and preprocessing module, which is used to acquire the job description text data and the job seeker's resume text data, and preprocess them to obtain the preprocessed input sequence; A semantic encoding and feature extraction module, which is used to perform semantic encoding on the preprocessed input sequence by using a Transformer model, extract keywords, skill requirements and experience information, and generate a high-dimensional text feature vector; A graph structure data construction module, which is used to represent the key information of the position as a graph structure, where the nodes represent the features of the position or the job seeker, and the edges represent the semantic relationship or relevance between the nodes, and construct the graph structure data; A node information propagation and fusion module, which is used to use a graph neural network to perform node information propagation and fusion on the graph structure, and update the node features through multi-layer graph convolution operations; A comprehensive matching degree calculation module, which is used to calculate the comprehensive matching degree of the position and the job seeker based on the output features of the Transformer and the graph neural network, in combination with the deep learning matching degree calculation module; An output module, which is used to output the matching results of the position and the resume.

10. An electronic device, comprising a processor and a memory, where a computer program is stored on the memory, characterized in that, When the processor executes the computer program, it can implement the person-job matching method described in any one of claims 1-8.

Images