A Method and System for Constructing Heterogeneous Graph Convolutional Neural Network Models for Information Recommendation

Abstract

This invention provides a method and system for constructing a heterogeneous graph convolutional neural network model for information recommendation. The method includes the following steps: acquiring resume information from multiple early job seekers and constructing a heterogeneous resume graph based on all the resume information; extracting a feature matrix from the heterogeneous resume graph; acquiring the association attributes between the resume information and constructing an initial adjacency matrix based on the association attributes; constructing multiple meta-paths with each resume information as a node based on the educational and skill information in the resume information; aggregating all the meta-paths and the initial adjacency matrix into a target adjacency matrix; using the feature matrix and the target adjacency matrix as model input, and constructing a heterogeneous graph convolutional neural network model using a graph convolutional neural network framework. The heterogeneous graph convolutional neural network model constructed by this invention can save early job seekers time and effort in the process of screening recruitment information.

Description

Technical Field

[0001] This invention belongs to the field of Internet technology, specifically relating to a method and system for constructing a heterogeneous graph convolutional neural network model for information recommendation. Background Technology

[0002] In recent years, with the development of artificial intelligence-related technologies, online recruitment platforms have emerged, aiming to provide services for potential job seekers and employers. Given the large number of job vacancies and job seekers, recruiters often struggle to find suitable candidates in a timely manner. Generally, recruiters focus on two key aspects of a candidate's resume in the early stages: educational background and skills. Educational background reflects a candidate's consistent effort, while skills reflect their ability to learn new things. Traditional recruitment processes rely heavily on recruiters' subjective experience, making it difficult to quantitatively and comprehensively evaluate the information from a large number of resumes. Therefore, many recruitment platforms utilize machine learning methods to build neural network models, enabling rapid person-job matching and pushing suitable resumes to recruiters, thus achieving intelligent human resource management.

[0003] However, recruitment is a two-way selection process that should meet the needs of both job seekers and recruiters. Recommending suitable positions to job seekers is also a crucial task for recruitment platforms. However, existing recruitment platforms heavily rely on human features and expert knowledge in their machine learning methods, leading to high costs, difficulty in updating, and a high risk of errors. Furthermore, existing recruitment platforms using machine learning models recommend job postings based on job seekers' work experience. Since early job seekers often lack relevant work experience, existing machine learning models cannot be directly applied to early-stage job seekers such as recent graduates. Faced with a massive amount of job postings, early-stage job seekers have to spend considerable time and effort searching and filtering to find relatively suitable positions. Summary of the Invention

[0004] This invention provides a method and system for constructing a heterogeneous graph convolutional neural network model for information recommendation, in order to solve the problem that machine learning models cannot be applied to early job seekers, which leads to early job seekers having to spend a lot of time and energy screening massive amounts of recruitment information in order to find a relatively suitable job.

[0005] In a first aspect, the present invention provides a method for constructing a heterogeneous graph convolutional neural network model for information recommendation, the method comprising the following steps:

[0006] Obtain resume information from multiple early job seekers and construct a heterogeneous resume graph based on all the resume information;

[0007] Extract the feature matrix from the heterogeneous resume graph;

[0008] Obtain the association attributes between the resume information, and construct an initial adjacency matrix based on the association attributes;

[0009] Construct multiple meta-paths with each resume information as a node based on the education and skills information in the resume information;

[0010] Aggregate all the meta-paths and the initial adjacency matrix into a target adjacency matrix;

[0011] The feature matrix and the target adjacency matrix are used as model inputs, and a heterogeneous graph convolutional neural network model is constructed using a graph convolutional neural network framework.

[0012] Optionally, the heterogeneous graph convolutional neural network model further includes two hidden layers and one output layer.

[0013] Optionally, the functional expression of the neural network layer in the heterogeneous graph convolutional neural network model is:

[0014] f(H (l) ,B)=RELU(BH (l) W (l) )

[0015] In the formula: H (0) =X,H (l) =Z, X is the feature matrix, l is the number of layers in the neural network, Z is the graph-level output, W (l) Let be the weights of the neural network layer, and RELU(·) be the non-linear activation function ReLU, and RELU(x) = max(0,x).

[0016] Optionally, the output layer is a softmax layer, and the expression of the output layer is:

[0017]

[0018] Optionally, the loss function in the heterogeneous graph convolutional neural network model is defined as the negative log-likelihood loss, and the expression of the loss function is:

[0019]

[0020] Optionally, the method further includes the following steps:

[0021] The heterogeneous resume graph is set to self-loop by adding the preset identity matrix to the target adjacency matrix.

[0022] Optionally, the method further includes the following steps:

[0023] The target adjacency matrix is ​​normalized to a symmetric adjacency matrix;

[0024] The neural network layers in the heterogeneous graph convolutional neural network model are updated based on the symmetric adjacency matrix.

[0025] Optionally, the formula for calculating the symmetric adjacency matrix is:

[0026]

[0027] In the formula: B is the target adjacency matrix. Let be the symmetric adjacency matrix. Let be the diagonal node degree matrix of the symmetric adjacency matrix.

[0028] Optionally, the expression for updating the neural network layer is as follows:

[0029]

[0030] In the formula: H (0) =X, where X is the feature matrix, l is the number of layers in the neural network, and W (l) σ represents the weights of the neural network layer, and σ represents the output layer of the heterogeneous graph convolutional neural network model.

[0031] In a second aspect, the present invention also provides a heterogeneous graph convolutional neural network model construction system for information recommendation, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0032] The beneficial effects of this invention are:

[0033] The method for constructing a heterogeneous graph convolutional neural network model for information recommendation in this invention includes the following steps: acquiring resume information from multiple early job seekers and constructing a heterogeneous resume graph based on all the resume information; extracting a feature matrix from the heterogeneous resume graph; acquiring the association attributes between the multiple resume information and constructing an initial adjacency matrix based on the association attributes; constructing multiple meta-paths with each resume information as a node based on the education and skills information in the resume information; aggregating all the meta-paths and the initial adjacency matrix into a target adjacency matrix; using the feature matrix and the target adjacency matrix as model input, and constructing a heterogeneous graph convolutional neural network model using a graph convolutional neural network framework. The heterogeneous graph convolutional neural network model can mine education and skills-related information from the resumes of early job seekers, thereby accurately locating the category of early job seekers. This allows for the recommendation of the most suitable recruitment information based on the category of early job seekers, saving early job seekers time and effort in the recruitment information screening process. Simultaneously, it accelerates the process for recruiters to screen suitable early job seekers from a massive number of resumes. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for constructing a heterogeneous graph convolutional neural network model for information recommendation in one embodiment of this application.

[0035] Figure 2 This is a schematic diagram illustrating the relationship between the dropout parameter and the accuracy of the heterogeneous graph convolutional neural network model in one embodiment of this application.

[0036] Figure 3 This is a schematic diagram illustrating the relationship between the parameter learning rate and the accuracy of the heterogeneous graph convolutional neural network model in one embodiment of this application.

[0037] Figure 4 This is a schematic diagram illustrating the accuracy of adding different meta-paths to the adjacency matrix in one embodiment of this application. Detailed Implementation

[0038] This invention discloses a method for constructing a heterogeneous graph convolutional neural network model for information recommendation.

[0039] In one embodiment, reference is made to Figure 1 The method for constructing a heterogeneous graph convolutional neural network model for information recommendation specifically includes the following steps:

[0040] S101. Obtain resume information from multiple early job seekers and construct a heterogeneous resume graph based on all resume information.

[0041] Early job seekers refer to recent graduates and other job seekers with no work experience but possessing certain educational backgrounds, skills, and learning abilities. Resumes of early job seekers can be obtained through recruitment platforms. Heterogeneous resume graphs are a type of heterogeneous graph, defined as follows: in ε represents the set of nodes, and ε represents the set of edges. The network pattern of a heterogeneous graph is represented as follows: The mapping function from node set to node type set is expressed as follows: The mapping of a link set to a link type set is represented as follows: For each node v∈V, it is associated with a typical node type. Furthermore, each link, e∈E, is associated with a typical link type. The sets of node types and link types are respectively represented as and in

[0042] In this embodiment, a heterogeneous resume graph is defined as G = (A G ,L G AG It is a combination of vertices of different types, L G It is a collection of various links, where l G ∈L G , a G ∈A G Assume that each node is connected to itself, that is, for any a G Satisfy (a) G ,a G )∈A G The feature matrix is ​​defined as follows: Where m is the dimension of the feature vector, and n is the number of nodes. Each row It is the eigenvector of a.

[0043] S102. Extract the feature matrix from the heterogeneous resume graph.

[0044] In the heterogeneous resume graph, each resume entry represents a node, and each node possesses node features. The node features are defined as an N×M dimensional feature matrix X, where N is the number of nodes and M is the number of node features.

[0045] S103. Obtain the association attributes between resume information and construct an initial adjacency matrix based on the association attributes.

[0046] In this embodiment, the association attribute refers to the relationship between users on the recruitment platform. This association attribute includes a list of other users displayed on the target user's resume information page on the recruitment platform. That is, when viewing the target user's resume information page, users in this list will also be seen. The resume information of these other users is similar to that of the target user. The association attribute between two resumes can be considered as a link between them, allowing the construction of an initial adjacency matrix B based on this attribute. The degree matrix D of the initial adjacency matrix B is also defined, and its expression is D. pp =∑ p B pq The degree matrix is ​​a commonly used concept in graph theory, typically used to describe the degree statistics of a graph. In an undirected graph, the degree of each node refers to the number of nodes directly adjacent to that node, and the degree matrix is ​​a diagonal matrix where each element on the diagonal represents the degree of the corresponding node. Due to the existence of self-loops, the diagonal elements of the initial adjacency matrix B are all assigned 1.

[0047] In one implementation, the self-loop of the heterogeneous resume graph is established by adding a preset identity matrix to the target adjacency matrix. Specifically, in the adjacency matrix representation, each node and its neighboring nodes have a weight representing the connection strength between them. However, if the connection between a node and itself is lost, this can be compensated for by adding a weight representing the self-loop. This self-loop weight can be implemented by adding an identity matrix to the adjacency matrix.

[0048] The specific steps are as follows: add the adjacency matrix to the identity matrix to obtain a weighted adjacency matrix, and then use the weighted adjacency matrix as input to a self-looping neural network layer, thus introducing self-looping connections into the network. Since the identity matrix represents the connection between a node and itself, adjusting the adjacency matrix by adding it to the adjacency matrix achieves the purpose of setting up self-loops.

[0049] S104. Construct multiple meta-paths with each resume information as a node based on the education and skills information in the resume.

[0050] Metapath defines the nodes. and Relationship set Among the symbols Metapaths represent composition operators on relations. They extend the concept of link types in heterogeneous graph networks and carry different semantics depending on their composition.

[0051] In this embodiment, Res = {r1, r2, ..., r n} represents the resume information of n early job seekers. Each individual resume can be described by a set of attributes, denoted as Res = {Edu, Sk}, where Edu = {e1, e2, ..., e...} t} represents the educational information of early job seekers. The smaller the t-value, the earlier the educational experience of the early job seeker. Therefore, e t This represents the most recent educational experience of an early job seeker. Educational information can be further subdivided, specifically represented as Edu = {Deg, Unr, Maj}, where Deg represents the degree information in the resume, Unr represents the school's world ranking information in the resume, and Maj represents the major information in the resume. Sk represents the early job seeker's skills information, defined as Sk = {s1, s2, ..., s...} j}

[0052] Because the overall resume information is too complex and scattered, it can be processed into categorized data, represented as Job={T1,T2,...,T...}. oA resume can be described using a set of attributes, represented as Res = {Deg, Unr, Maj, Sk}. The relationship between two resumes is represented as... Where r * This refers to the attribute information displayed on the resume.

[0053] In this embodiment, different meta-paths capture the semantic relationships between resume information from different perspectives. For example, the meta-path "RDR" indicates that the early job seekers corresponding to two resumes have the same educational level. The meta-path "RSR" indicates that the two resumes have similar skill information. The meta-path "RSR" indicates that the two resumes have the same school world ranking information. The meta-path "RMR" indicates that the two resumes have the same professional information.

[0054] S105. Aggregate all metapaths and the initial adjacency matrix into the target adjacency matrix.

[0055] After constructing the initial adjacency matrix and different meta-paths among the resume information, they can be aggregated into a new target adjacency matrix. In this embodiment, the following steps can be taken:

[0056] Constructing the metapath matrix: Metapath sampling is performed on the graph to obtain several types of metapaths. Then, a metapath matrix is ​​constructed, where each column represents a metapath and each row represents a node. In the metapath matrix, each element indicates whether a metapath has passed through that node, using 0 / 1 encoding.

[0057] Obtaining the node context vector: Multiply the metapath matrix by the adjacency matrix to obtain the node context vector matrix. In the node context vector matrix, each row represents a node, and each column represents the node context vector corresponding to the metapath.

[0058] Metapath aggregation: For each node, its context vector is aggregated into a single vector according to a certain method. Common aggregation methods include summing all context vectors, averaging, and using attention mechanisms.

[0059] To obtain a new adjacency matrix: combine the vectors obtained after aggregating each node in a certain way to get a new node representation vector as a row (or column) of the new adjacency matrix.

[0060] S106. Using the feature matrix and the target adjacency matrix as model inputs, construct a heterogeneous graph convolutional neural network model using a graph convolutional neural network framework.

[0061] The heterogeneous graph convolutional neural network (HPN) model is a semi-supervised HPN model for node classification. Graph convolutional neural networks aggregate feature information from a node's first-order neighbors. Therefore, the heterogeneous graph HPN model can capture nearest-neighbor information through two convolutional layers. The heterogeneous graph HPN model operates directly on the graph and derives node embedding vectors based on their neighborhood attributes. Stacking multiple graph HPN layers aggregates information about larger neighborhoods.

[0062] In one embodiment, the heterogeneous graph convolutional neural network model constructed in step S106 includes one input layer, two hidden layers, and one output layer. The feature matrix and the target adjacency matrix serve as the input layer, and the node embeddings of the two hidden layers are the same size as the label set. In this embodiment, each neural network layer in the heterogeneous graph convolutional neural network model can be represented as a nonlinear function, and the function expression of the neural network layer is:

[0063] f(H (l) ,B)=RELU(BH (l) W (l) )

[0064] In the formula: H (0) =X,H (l) =Z, where X is the feature matrix, l is the number of layers in the neural network, Z is the graph-level output, and W (l) Let X be the weights of the neural network layer. RELU(·) is the non-linear activation function ReLU, and RELU(x) = max(0,x).

[0065] In this embodiment, the output layer is a softmax layer, and the expression for the output layer is:

[0066]

[0067] After inputting the heterogeneous graph into a heterogeneous graph convolutional neural network model, node embedding vectors under different metapaths are obtained. Then, all node embedding vectors are aggregated to obtain the final embedding vector. Finally, the final embedding vector is input into a softmax layer to obtain the graph-level output Z, expressed as:

[0068]

[0069] Where X is the characteristic matrix. and All are weights of the hidden layer with feature mapping, and σ represents the softmax layer.

[0070] In one embodiment, since the heterogeneous graph convolutional neural network model constructed in this invention is used to solve multi-class classification problems, the loss function is defined as negative log-likelihood loss, and the expression of the loss function is:

[0071]

[0072] Where y i This represents the logarithm C of the probability values ​​corresponding to only the true class within the categories. For example, if C = 2, then... It follows a probability distribution and sums to 1.

[0073] Heterogeneous graph convolutional neural network (HNN) models learn the mapping relationship from input to output through training. During model training, the backpropagation algorithm calculates the error and distributes it to the connection weights between each neuron. Then, gradient descent is used to update the weights and biases, thereby continuously optimizing the model's performance. Updating neural network layers refers to updating the connection weights and biases between neurons within the neural network, allowing the network to gradually adapt to the training data, thus improving its performance and accuracy. In the field of deep learning, updating neural network layers is typically an iterative process that requires continuous adjustment of the neural network structure and parameter settings based on actual data and task requirements to achieve optimal results.

[0074] In one implementation, the neural network layers in a heterogeneous graph convolutional neural network model can be updated via the following steps:

[0075] Normalize the target adjacency matrix into a symmetric adjacency matrix.

[0076] Updating neural network layers in a heterogeneous graph convolutional neural network model based on symmetric adjacency matrices.

[0077] In this embodiment, the formula for calculating the symmetric adjacency matrix is:

[0078]

[0079] In the formula: B is the target adjacency matrix. It is a symmetric adjacency matrix. It is the diagonal node degree matrix of a symmetric adjacency matrix.

[0080] The expression for updating the neural network layer is as follows:

[0081]

[0082] In the formula: H (0) =X, where X is the feature matrix, l is the number of layers in the neural network, and W (l) σ represents the weights of the neural network layer, and σ is the output layer of the heterogeneous graph convolutional neural network model.

[0083] In one implementation, the feasibility of the heterogeneous graph convolutional neural network model in this invention is verified based on a real dataset and eight different baseline models. The eight baseline models are Graph Convolutional Network (GCN), Graph Attention Network (GAT), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Adaboost (AB), Random Forest (RF), Logistic Regression (LR), and Naive Bayes Classifier (NBC).

[0084] GCN is a semi-supervised learning method for graph-structured data, and it's a more efficient version of convolutional neural networks. It's cleverly designed to extract features from graph data, which can then be used for node classification. Compared to GCN, GAT employs an attention mechanism to achieve more efficient neighbor aggregation. GAT does not require complex matrix operations or prior knowledge of the graph structure. During convolution, GAT assigns different importances to different nodes within the same neighborhood by stacking self-attention layers.

[0085] SVM handles classification tasks by identifying a separating boundary called the classification plane, which maximizes the distance between the boundary and the nearest data point on each side, with the aim of efficiently dividing data points into separate classes. KNN is a well-established theory with a simple concept, and can be used for both classification and regression. The idea behind KNN is that if most of the K nearest neighbors (i.e., nearest neighbors in the feature space) of a sample belong to a certain class in the feature space, then the sample also belongs to that class.

[0086] A / B testing is a method that involves training multiple simple classifiers using training data and then merging them into a powerful classifier. Random Forest (RF) is a classifier that uses multiple decision trees and handles multi-class problems well. The output class of an RF is determined individually by each decision tree in the forest, and the final predicted class is determined by the class chosen most frequently.

[0087] Logistic Regression (LR) can solve multiple classification tasks and is used to represent the probability of events occurring. LR is easy to deploy, requires minimal computational power, provides fast results, and uses minimal storage resources during classification. National Classification (NBC) is one of the most widely used classification algorithms based on Bayes' theorem. The essence of Bayesian classification algorithms is to calculate conditional probabilities.

[0088] To evaluate the classification performance of all models, the following five representative evaluation metrics were selected: accuracy, precision, Macro-F1 score, and AUC. Accuracy measures the proportion of correctly predicted labels by the model. It is calculated by correctly predicting labels out of a total number of instances. The specific formula for calculating accuracy is as follows:

[0089]

[0090] Precision refers to the proportion of items identified as positive by the model that are actually positive. It can be calculated by dividing the number of correctly predicted positive labels by the total number of positive predictions. Higher precision indicates a lower false positive rate. The specific formula for calculating precision is as follows:

[0091]

[0092] The Macro-F1 score combines precision and recall under the concept of the harmonic mean, finding the optimal trade-off between these two metrics. The specific calculation formula is as follows:

[0093]

[0094]

[0095]

[0096] TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

[0097] AUC represents the probability, indicating the ability of a classification algorithm to distinguish between positive and negative samples. A higher AUC value suggests that the algorithm more effectively prioritizes positive samples over negative samples, indicating better classification performance. The specific formula for calculating AUC is as follows:

[0098]

[0099] Where P, N, and M represent the number of positive samples, negative samples, and classification categories, respectively.

[0100] In this implementation, experiments were conducted using PyTorch Geometric 1.11.0. The learning rate was initially set to 0.01, and dropout to 0.2. The model was then trained for over 900 generations with decaying weights set to 0.0005. To generate the same model initialization learnable parameters each time, a random number seed of 42 was set. In this case, the hidden heterogeneous graph convolutional neural network model had 16 layers. A two-step meta-path was used for the heterogeneous graph convolutional neural network model, with a total of 4 meta-paths.

[0101] Table 1 Performance data of the heterogeneous graph convolutional neural network model and the baseline model

[0102] accuracy Accuracy Macro-F1 score AUC Heterogeneous graph convolutional neural network model 0.8591 0.9011 0.8218 0.8475 GCN 0.7350 0.4905 0.5173 0.6911 GAT 0.6644 0.6532 0.6189 0.6299 SVM 0.5745 0.5209 0.5401 0.5495 KNN 0.5532 0.1829 0.1879 0.5345 AB 0.5390 0.4683 0.5003 0.5382 RF 0.5887 0.1951 0.2105 0.5553 LR 0.6312 0.3206 0.2714 0.5829 NBC 0.0922 0.0768 0.0974 0.5643

[0103] The experimental results are shown in Table 1. The heterogeneous graph convolutional neural network (HBM) model performed best, significantly outperforming all baseline models. The HBM model without meta-paths performed worse than the one with meta-paths. This indicates that adding meta-paths to the heterogeneous resume graph preserves the syntactic and semantic relationships between resume information, providing additional information in large external resume datasets. The significant superiority of the HBM model over the baseline models demonstrates that it can effectively combine the features of nodes and their neighbors to analyze resume information. NBC, relying on the assumption of independent sample attributes, achieved the worst results on the experimental dataset, implying that sample attributes are correlated. GAT performed worse than the HBM model, indicating that GAT's unsupervised resume embeddings are not very discriminative in job classification.

[0104] Heterogeneous graph convolutional neural network (HBM) models are an improved form of GCN model. As a special form of Laplacian smoothing, the new features of a node are calculated as a weighted average of the node itself and its second-order neighbors. Therefore, HBM models combine the features of the node itself and its neighbors in the heterogeneous graph to analyze nodes, resulting in good performance in node classification. Furthermore, by adding meta-paths as features, heterogeneous graphs can capture node-node relationships and global node-meta-path relationships. Therefore, HBM models are more interpretable.

[0105] In one implementation, the robustness of the heterogeneous graph convolutional neural network model was further evaluated through accuracy testing experiments. (See also...) Figure 2 , Figure 2 The test accuracy is shown under different parameter settings. As dropout increases, the test accuracy gradually decreases, but there is a brief increase in average accuracy at 0.5. This indicates that dropout at 0.5 randomly generates relatively more network structures.

[0106] Reference Figure 3 ,exist Figure 3 The paper describes the classification performance of heterogeneous graph convolutional neural network models with different learning rates. It can be observed that the test accuracy reaches the optimal solution and is relatively robust at a learning rate of 0.01. Furthermore, excessively high learning rates do not improve classification performance and may lead to oscillations.

[0107] Reference Figure 4 , Figure 4The values ​​represent the accuracy of adding different meta-paths to the adjacency matrix. Variable A on the X-axis represents using the association attribute as the adjacency matrix alone. Furthermore, variables AD, AM, AU, AS, AMS, and ADMSU on the X-axis represent adding meta-paths RDR, RMR, RUR, RSR, RMR and RSR, RDR, RMR, RUR, and RSR to the association attribute, respectively. The heterogeneous graph convolutional neural network model achieves optimal performance when the adjacency matrix adds all four meta-paths to the association attribute. When the adjacency matrix is ​​a combination of the association attribute and the meta-path RDR, the model performance decreases compared to not adding any. However, by adding the meta-path RDR as AMSU to the adjacency matrix, the accuracy of the heterogeneous graph convolutional neural network model is significantly improved. These results suggest that early job seekers should not only focus on academic qualifications when selecting jobs but also comprehensively evaluate their educational background and skill level.

[0108] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0109] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for constructing a heterogeneous graph convolutional neural network model for information recommendation, characterized in that, Includes the following steps: Obtain resume information from multiple early job seekers and construct a heterogeneous resume graph based on all the resume information; Extract the feature matrix from the heterogeneous resume graph; Obtain the association attributes between the resume information, and construct an initial adjacency matrix based on the association attributes; Construct multiple meta-paths with each resume information as a node based on the education and skills information in the resume information; Aggregate all the meta-paths and the initial adjacency matrix into a target adjacency matrix; The feature matrix and the target adjacency matrix are used as model inputs, and a heterogeneous graph convolutional neural network model is constructed using a graph convolutional neural network framework.

2. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 1, characterized in that, The heterogeneous graph convolutional neural network model also includes two hidden layers and one output layer.

3. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 2, characterized in that, The functional expression of the neural network layer in the heterogeneous graph convolutional neural network model is: f(H (l) ,B)=RELU(BH (l) W (l) ) In the formula: H (0) =X,H (l) =Z, X is the feature matrix, l is the number of layers in the neural network, Z is the graph-level output, W (l) Let be the weights of the neural network layer, and RELU(·) be the non-linear activation function ReLU, and RELU(x) = max(0,x).

4. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 2, characterized in that, The output layer is a softmax layer, and the expression for the output layer is:

5. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 1, characterized in that, The loss function in the heterogeneous graph convolutional neural network model is defined as the negative log-likelihood loss, and the expression of the loss function is:

6. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 1, characterized in that, The method further includes the following steps: The heterogeneous resume graph is set to self-loop by adding the preset identity matrix to the target adjacency matrix.

7. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 1, characterized in that, The method further includes the following steps: The target adjacency matrix is ​​normalized to a symmetric adjacency matrix; The neural network layers in the heterogeneous graph convolutional neural network model are updated based on the symmetric adjacency matrix.

8. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 7, characterized in that, The formula for calculating the symmetric adjacency matrix is ​​as follows: In the formula: B is the target adjacency matrix. Let be the symmetric adjacency matrix. Let be the diagonal node degree matrix of the symmetric adjacency matrix.

9. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 8, characterized in that, The expression for updating the neural network layer is as follows: In the formula: H (0) =X, where X is the feature matrix, l is the number of layers in the neural network, and W (l) σ represents the weights of the neural network layer, and σ represents the output layer of the heterogeneous graph convolutional neural network model.

10. A system for constructing a heterogeneous graph convolutional neural network model for information recommendation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 9.

Images