Intelligent recruitment system and method for matching positions with candidates based on heterogeneous graph neural network

Abstract

This invention discloses a method and system for matching people and jobs in an intelligent recruitment system based on heterogeneous graph neural networks. The method includes: preprocessing collected user and job data; constructing a heterogeneous graph based on the preprocessed data; negative sampling of the heterogeneous graph along the "user-job" edge and setting meta-paths; learning the nodes in the heterogeneous graph based on the heterogeneous graph neural network model to obtain user node embeddings and job node embeddings; inputting the user node embeddings and job node embeddings into a scoring model to obtain predicted scores between users and jobs; training the heterogeneous graph neural network model and the scoring model using a cross-entropy loss function; inputting users and preset jobs into the trained model to obtain the optimal Top-K job recommendations. This invention incorporates auxiliary information about users and jobs, making the feature embedding of users and jobs more accurate, the recommendation results more personalized and accurate, and allowing for flexible graph construction based on collected information.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for matching people to jobs in an intelligent recruitment system based on heterogeneous graph neural networks. Background Technology

[0002] The job-person matching problem in intelligent recruitment systems is essentially a recommendation system problem. Based on knowledge of recommendation systems, the following are some common techniques used to achieve job-person matching in intelligent recruitment systems:

[0003] ① Collaborative Filtering-Based Intelligent Recruitment System Job Matching Method: The core idea of ​​this method is "birds of a feather flock together." This method can be further divided into user-based collaborative filtering, item-based collaborative filtering, and model-based collaborative filtering. The basic idea of ​​user-based collaborative filtering is to calculate the most similar users to the target user based on the user's historical records on the recruitment platform (cosine similarity, Pearson similarity, Jaccard similarity), and then recommend jobs to the target user based on these similar users. The basic idea of ​​item-based collaborative filtering is to find the group of jobs most similar to the target job based on the similarity between jobs on the platform (such as job description, required skills, etc.), and then recommend jobs similar to the user's historical positions. Model-based collaborative filtering: The most typical algorithm is matrix factorization. First, the user-job rating matrix is ​​obtained. Through matrix factorization and multiplication, matrix completion is achieved, and jobs with higher ratings after completion are recommended to the user.

[0004] ② Content-Based Intelligent Recruitment System for Job Matching: This method collects information such as job seekers' resumes, skills, and experience, as well as job descriptions and requirements, from recruitment platforms. Keywords, phrases, and sentences are extracted as features from the text of both job seekers and job postings. Features can be represented using methods such as the bag-of-words model and TF-IDF (term frequency-inverse document frequency). Similarity metrics (such as cosine similarity and Euclidean distance) are used to calculate the similarity score between job seekers and jobs. The similarity score measures the degree of match between the job seeker and the job; a higher score indicates a higher degree of match.

[0005] ③ Hybrid recommendation methods combine collaborative filtering and content-based methods. This is currently the most commonly used method in recommendation systems. The simplest hybrid recommendation method involves obtaining recommendation sequences from both collaborative filtering and content-based methods, and then combining them according to a certain principle to obtain the final recommendation result.

[0006] However, these methods fail to effectively uncover the complex relationships between users and jobs. Graph neural networks (GNNs), by modeling users and jobs as a graph structure, utilize node embeddings and GNN algorithms to learn the features of users and items, aiming to capture the connections between them more comprehensively and accurately. However, most GNNs only contain a bipartite graph of users and jobs, lacking the personal attributes of users and the attributes of jobs. Recruitment systems, on the other hand, can collect auxiliary information about users and jobs. The collection and analysis of this auxiliary information helps recruitment systems better understand the needs of users and jobs, and perform intelligent matching based on algorithms. By comprehensively considering the personal characteristics of users and the requirements of jobs, recruitment systems can provide job recommendations that better meet user expectations, improving recruitment effectiveness and satisfaction. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] In view of the above-mentioned problems, the present invention is proposed.

[0009] A first aspect of this invention provides a method for matching job candidates to positions in an intelligent recruitment system based on a heterogeneous graph neural network, comprising: preprocessing collected user data and job data; constructing a heterogeneous graph based on the preprocessed data; negative sampling of the heterogeneous graph along the edge "user-job", dividing the graph into a training set, a validation set, and a test set according to a ratio, and setting a meta-path; learning nodes in the heterogeneous graph based on a heterogeneous graph neural network model to obtain user node embeddings and job node embeddings; inputting the user node embeddings and job node embeddings into a scoring model to obtain a predicted score between the user and the job; training the heterogeneous graph neural network model and the scoring model using a cross-entropy loss function; inputting a user and a preset job into the trained model to obtain the optimal Top-K job recommendations.

[0010] As a preferred embodiment of the job matching method for an intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the construction of the heterogeneous graph includes:

[0011] The collected user data includes the user's name, age, education, major, university, and work experience. The collected job data includes all positions the user has held. The preset job data includes all previously held positions collected, denoted as S. v ;

[0012] The collected user data and job data are then processed in a structured manner;

[0013] Map the strings in the structured data to numerical values;

[0014] Based on the processed data, construct a heterogeneous graph, and design the node types and names, as well as the edge types and names of the heterogeneous graph.

[0015] As a preferred embodiment of the job matching method for an intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the method includes: negative sampling of the heterogeneous graph with respect to the edge "user-job";

[0016] The "user-position" edges of the generated heterogeneous graph are represented using triples;

[0017] The generated heterogeneous graph is negatively sampled with respect to the "user-job" edge, that is, a certain number of negative edges are randomly constructed, positive edges and negative edges are merged, the order of the triples is shuffled, and the training set, validation set and test set are divided according to a certain ratio;

[0018] Remove the edges of the validation set and the test set from the heterogeneous graph to obtain the processed heterogeneous graph, and save the processed heterogeneous graph, the training set, the validation set, and the test set.

[0019] Set the meta-paths required when training heterogeneous graph neural networks.

[0020] As a preferred embodiment of the human-job matching method for an intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the acquisition of user node embedding and job node embedding includes,

[0021] Initialize the embeddings of different types of nodes in the heterogeneous graph G = (V, ε), where V represents the set of nodes, ε represents the set of edges, and different node types correspond to different dimensions. The initial embedding of node e is... Node w is initialized as embedded Node j is initialized as embedded

[0022] Using projection matrix P r This projects different types of neighboring nodes of the target node onto the same dimension of the target node.

[0023] The attention weights of the target node and its other neighboring nodes are calculated using the following formula:

[0024]

[0025] Where N1(e) represents all one-hop neighbor nodes of node e;

[0026] The one-hop neighbor aggregation of the target node is performed according to the calculated attention weights, as follows:

[0027] Two nodes can be connected via the meta-path. The attention weights of neighboring nodes based on the meta-path can be calculated using the following formula:

[0028]

[0029] Where, N ρ (e) represents the set of all neighboring nodes of node e on the metapath ρ;

[0030] The attention mechanism is used to aggregate information about neighboring nodes based on the meta-path ρ, as follows:

[0031] The attention score for each metapath is calculated using the following formula:

[0032]

[0033] Where q represents the semantic attention vector, and W represents the weight matrix. This represents the aggregated information of the neighbor nodes of a certain meta-path, where b represents the deviation vector and N represents the number of meta-paths.

[0034] After obtaining the attention score for each metapath, the attention score for each metapath is normalized using the softmax function, and expressed as:

[0035]

[0036] Where, γ ρj Represents the metapath ρ j Attention weights;

[0037] The learned weights are used as coefficients to fuse semantically specific embeddings to obtain the final embedding of node e based on the meta-path, represented as:

[0038] Initialize and embed the node e. What I learned and what I learned Aggregated into node e, the final embedding of node e is obtained, represented as:

[0039]

[0040] Where W1 and W2 represent weight matrices, and b1 and b2 represent bias vectors;

[0041] After updating the heterogeneous graph node embedding, the user node embedding u is obtained. t embedding v with job node t .

[0042] As a preferred embodiment of the job matching method for an intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the acquisition of the predicted score between the user and the job includes:

[0043] Embed the user node in u t embedding v with job node t The scoring model is used to obtain the user's predicted score for the job, expressed by the formula:

[0044]

[0045] in, denoted by f, which represents the user's predicted score for the job position.

[0046] As a preferred embodiment of the job matching method for an intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the step of training the heterogeneous graph neural network model and the scoring model using the cross-entropy loss function includes:

[0047] The cross-entropy loss function L loss The calculations include,

[0048]

[0049] Among them, Y t This represents the node pairs representing all users and roles;

[0050] Update all weight parameters of the heterogeneous graph neural network model and the scoring model, and continuously correct the weight parameters during the iteration process in order to obtain the minimum loss function, and save the trained heterogeneous graph neural network model and the scoring model.

[0051] As a preferred embodiment of the job matching method for the intelligent recruitment system based on heterogeneous graph neural networks described in this invention, the acquisition of the optimal Top-K job recommendations includes:

[0052] Calculate the target user u and all preset positions v i ,i∈S v The predicted score between them is first obtained by using the heterogeneous graph neural network model to obtain the user node embedding u. t and job node embedding Then substitute it into the trained scoring model;

[0053] The calculation of the trained scoring model includes,

[0054]

[0055] All predicted scores are sorted to obtain the Top-K job postings for recommendation.

[0056] A second aspect of this invention provides an intelligent recruitment system for matching job candidates with candidates based on heterogeneous graph neural networks, comprising:

[0057] The heterogeneous graph construction unit is used to preprocess the collected user data and job data, construct a heterogeneous graph based on the preprocessed data, perform negative sampling on the heterogeneous graph about the edge "user-job", divide the training set, validation set and test set according to the ratio, and set the meta path.

[0058] The scoring prediction unit is used to learn the nodes in the heterogeneous graph based on the heterogeneous graph neural network model to obtain user node embeddings and job node embeddings. The user node embeddings and job node embeddings are input into the scoring model to obtain the predicted score between the user and the job.

[0059] The job recommendation unit is used to train the heterogeneous graph neural network model and the scoring model using the cross-entropy loss function. By inputting a user and a preset job into the trained model, the best Top-K job recommendations are obtained.

[0060] A third aspect of the present invention provides an apparatus, the apparatus comprising,

[0061] processor;

[0062] Memory used to store processor-executable instructions;

[0063] The processor is configured to invoke instructions stored in the memory to execute the method described in any embodiment of the present invention.

[0064] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions, including:

[0065] When the computer program instructions are executed by the processor, they implement the method as described in any embodiment of the present invention.

[0066] The beneficial effects of this invention are as follows: The intelligent recruitment system for matching people and jobs based on heterogeneous graph neural networks provided by this invention constructs a heterogeneous graph based on user and job data, and considers the semantic information between users and recommended content, thereby improving the utilization rate of person and job data. It effectively alleviates the inability of traditional recommendation systems to effectively mine the complex relationship between users and jobs. At the same time, it incorporates auxiliary information of users and jobs, making the feature embedding of users and jobs more accurate, and making the recommendation results more personalized and accurate. Attached Figure Description

[0067] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0068] Figure 1 The present invention provides a method for matching job candidates to positions in an intelligent recruitment system based on heterogeneous graph neural networks, and a flowchart of the overall system.

[0069] Figure 2 The present invention provides a flowchart of the pre-recruitment process for a job matching method based on heterogeneous graph neural networks in an intelligent recruitment system.

[0070] Figure 3 The present invention provides a method for matching job candidates to positions in an intelligent recruitment system based on heterogeneous graph neural networks, and a heterogeneous graph of the system.

[0071] Figure 4 This invention provides a meta-path example diagram of the intelligent recruitment system for matching job candidates with jobs based on heterogeneous graph neural networks.

[0072] Figure 5 The diagram shows a detailed illustration of the heterogeneous graph neural network model of the intelligent recruitment system for matching people to jobs based on heterogeneous graph neural networks, as provided by this invention. Detailed Implementation

[0073] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0074] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0075] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0076] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0077] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0078] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0079] Example 1

[0080] Reference Figures 1-5 As one embodiment of the present invention, a method for matching job candidates to positions in an intelligent recruitment system based on heterogeneous graph neural networks is provided, comprising:

[0081] S1: Preprocess the collected user and job data, and construct a heterogeneous graph based on the preprocessed data. It should be noted that:

[0082] The collected user data includes the user's name, age, education, major, university, and work experience. The collected job data includes all positions the user has held. The preset job data includes all previously held positions collected, denoted as S. v ;

[0083] The collected user data and job data are structured, and the strings in the structured data are mapped to numerical values. Then, a heterogeneous graph is constructed based on the above-processed data, and the types and names of nodes and edges of the heterogeneous graph are designed.

[0084] It should be noted that structured processing refers to organizing, organizing, and transforming data to give it a specific structure and format for better analysis, storage, and use. User information or job postings are usually collected by crawling websites; such data cannot be used directly and requires structured processing before it can be used. Common structured processing methods include the following:

[0085] ① Data cleaning: Data cleaning refers to detecting and correcting errors, missing values, duplicates, or inconsistencies in data. This includes operations such as deleting duplicate data, filling in missing values, and repairing erroneous data to ensure the accuracy and consistency of the data.

[0086] ② Data Conversion: Data conversion includes operations that transform data from one representation to another. For example, converting unstructured data to structured data, or changing the format of data from one data source to another. This can involve data format conversion, data type conversion, unit conversion, etc.

[0087] ③ Data Integration: Data integration is the process of merging and combining data from different sources. During data integration, it is necessary to address issues such as inconsistent data schemas, key mismatches, and data redundancy to ensure that the integrated data has a unified structure.

[0088] ④ Data Standardization: Data standardization refers to mapping data to a certain standard range or format to facilitate comparison and analysis. Common data standardization methods include normalization (scaling data to the 0-1 range) and standardization (conforming to a normal distribution).

[0089] ⑤ Data Aggregation: Data aggregation is the process of summarizing and statistically analyzing data. By aggregating data, useful information and indicators can be extracted from a large amount of details to assist in decision-making and analysis.

[0090] S2: Negatively sample the heterogeneous graph along the "user-job" edge, divide it into training, validation, and test sets according to the ratio, and set meta-paths. It should be noted that:

[0091] The "user-position" edges of the generated heterogeneous graph are represented using triples;

[0092] Negative sampling of the "user-job" edges is performed on the generated heterogeneous graph. That is, a certain number of negative edges are randomly constructed, positive and negative edges are merged, the order of triples is shuffled, and the training set, validation set and test set are divided according to a certain ratio.

[0093] Remove the edges from the validation and test sets from the heterogeneous graph to obtain the processed heterogeneous graph, and save the processed heterogeneous graph, training set, validation set, and test set.

[0094] Set the meta-paths required when training heterogeneous graph neural networks;

[0095] It should be noted that a heterogeneous graph can be represented as G = (V, ε), where V is the set of nodes and ε is the set of edges. The heterogeneous graph is also associated with a node type mapping function ξ: V → A and an edge type mapping function ψ: ε → R, where A and R represent predefined sets of node and edge types, respectively, where |A| + |R| > 2. In this invention, the designed heterogeneous graph is as follows: Figure 3 As shown, the heterogeneous graph has 5 node types: user, degree, industry, position, and major. There are 10 edge types: um(user-major): user's major, mu(major-user): the major studied by the user, ud(user-degree): the user's degree level, du(degree-user): the user corresponding to that degree level, up(user-position): the user's position, pu(position-user): the user suitable for that position, iu(industry-user): the user in that department, ui(user-industry): the department the user belongs to, pi(position-industry): the department corresponding to the position, and ip(industry-position): the position in that department.

[0096] It should be noted that training a link prediction model involves comparing the scores between two connected nodes with the scores between any two pairs of nodes. For example, given an edge connecting u and v, a good model aims for a higher score between u and v than between u and a node v′ sampled from an arbitrary noise distribution; this approach is called negative sampling. Specifically, negative sampling of the up (user-position) edge is performed as follows: the edge up is represented as a triple (user_id, position_id, label). If the user has held a position in the user's job, the label is 1, and the triple is (user_id, position_id, 1). Some fake edges for users and irrelevant positions are randomly generated; these edges are labeled 0, and the triple is (user_id, position_id, 0).

[0097] S3: As Figure 5 As shown, a heterogeneous graph neural network model is used to learn the nodes in the heterogeneous graph, resulting in user node embeddings and job node embeddings. It should be noted that:

[0098] S3.1: Initialize the embeddings of different types of nodes in the heterogeneous graph G = (V, ε), where V represents the set of nodes and ε represents the set of edges. Different node types correspond to different dimensions. The initial embedding of node e is... Node w is initialized as embedded Node j is initialized as embedded

[0099] S3.2: Using the projection matrix P r This projects different types of neighboring nodes of the target node onto the same dimension of the target node.

[0100] S3.3: Calculate the attention weights of the target node and its other neighboring nodes, expressed by the formula:

[0101]

[0102] Where N1(e) represents all one-hop neighbor nodes of node e;

[0103] S3.4: Aggregate the one-hop neighbors of the target node according to the calculated attention weights, as follows:

[0104] S3.5: Two nodes can be connected via meta-paths. The attention weights of neighboring nodes based on meta-paths are calculated using the following formula:

[0105]

[0106] Where, N ρ(e) represents the set of all neighboring nodes of node e on the metapath ρ;

[0107] S3.6: An attention mechanism is used to aggregate information about neighboring nodes based on the meta-path ρ, represented as:

[0108] S3.7: Calculate the attention score for each metapath, expressed by the formula:

[0109]

[0110] Where q represents the semantic attention vector, and W represents the weight matrix. This represents the aggregated information of the neighbor nodes of a certain meta-path, where b represents the deviation vector and N represents the number of meta-paths.

[0111] S3.8: After obtaining the attention score for each metapath, the attention score for each metapath is normalized using the softmax function, as follows:

[0112]

[0113] Where, γ ρj Represents the metapath ρ j Attention weights;

[0114] S3.9: Using the learned weights as coefficients, the semantically specific embeddings are fused to obtain the final embedding of node e based on the meta-path, represented as:

[0115] S3.10: Initialize and embed node e What I learned and what I learned Aggregated into node e, the final embedding of node e is obtained, represented as:

[0116]

[0117] Where W1 and W2 represent weight matrices, and b1 and b2 represent bias vectors;

[0118] S3.11: After updating the heterogeneous graph node embeddings, the user node embedding u is obtained. t embedding v with job node t ;

[0119] It should be noted that the metapath ρ is defined as (abbreviated as A1A2…A) l+1 It describes objects A1 and A. l+1 A composite relationship between them: R = R1οR2ο···οR l, where ο represents a relational compound operator. For example Figure 4 As shown, two users can be connected through different meta-paths. If they are connected through the meta-path UMU (User-Major-User), it means that the two users are in the same major. If they are connected through the meta-path UPU (User-Position-User), it means that the two users have held the same position.

[0120] S4: Embed user nodes and job node embeddings into the scoring model to obtain predicted scores between users and jobs. It should be noted that:

[0121] Embed user nodes in u t embedding v with job node t The scoring model is used to obtain the user's predicted score for the job, expressed by the formula:

[0122]

[0123] in, denoted by f, which represents the user's predicted score for the job position.

[0124] S5: Train the heterogeneous graph neural network model and the scoring model using the cross-entropy loss function. Input a user and a preset job title into the trained model to obtain the optimal Top-K job recommendations. It should be noted that:

[0125] Cross-entropy loss function L loss The calculations include,

[0126]

[0127] Among them, Y t This represents the node pairs representing all users and roles;

[0128] Update all weight parameters of the heterogeneous graph neural network model and the scoring model, and continuously adjust the weight parameters during the iteration process in order to obtain the minimum loss function, and save the trained heterogeneous graph neural network model and the scoring model;

[0129] Furthermore, obtaining the best Top-K job recommendations includes,

[0130] Calculate the target user u and all preset positions v i ,i∈S v The predicted scores between them are first obtained by using a heterogeneous graph neural network model to obtain the user node embedding u. t and job node embedding Then substitute it into the trained scoring model;

[0131] Specifically, the calculation of the trained scoring model includes,

[0132]

[0133] All predicted scores are sorted to obtain the Top-K job postings for recommendation.

[0134] It should be noted that the intelligent recruitment system for matching people and jobs based on heterogeneous graph neural networks provided by this invention constructs a heterogeneous graph based on user and job data, while also considering the semantic information between users and recommended content, thereby improving the utilization rate of person and job data and effectively alleviating the inability of traditional recommendation systems to effectively mine the complex relationship between users and jobs. At the same time, the addition of auxiliary information for users and jobs makes the feature embedding of users and jobs more accurate, and makes the recommendation results more personalized and accurate.

[0135] The second aspect disclosed in this invention,

[0136] Provides an intelligent recruitment system and job matching system based on heterogeneous graph neural networks, including:

[0137] The heterogeneous graph construction unit is used to preprocess the collected user data and job data, construct a heterogeneous graph based on the preprocessed data, perform negative sampling on the heterogeneous graph about the edge "user-job", divide the training set, validation set and test set according to the ratio, and set the meta path;

[0138] The scoring prediction unit is used to learn the nodes in the heterogeneous graph based on the heterogeneous graph neural network model to obtain the user node embedding and the job node embedding. The user node embedding and the job node embedding are input into the scoring model to obtain the predicted score between the user and the job.

[0139] The job recommendation unit is used to train a heterogeneous graph neural network model and a scoring model using the cross-entropy loss function. By inputting a user and a preset job into the trained model, the best Top-K job recommendations are obtained.

[0140] The third aspect disclosed in this invention,

[0141] A device is provided, comprising:

[0142] processor;

[0143] Memory used to store processor-executable instructions;

[0144] The processor is configured to invoke instructions stored in memory to execute any of the aforementioned methods.

[0145] The fourth aspect disclosed in this invention,

[0146] A computer-readable storage medium is provided, having stored thereon computer program instructions, including:

[0147] A method by which computer program instructions are executed by a processor to implement any of the foregoing.

[0148] The present invention may be a method, apparatus, system and / or computer program product, and the computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the present invention.

[0149] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0150] Example 2

[0151] This embodiment differs from the first embodiment in that it provides a method for matching job candidates to positions in an intelligent recruitment system based on heterogeneous graph neural networks, as well as verification tests of the system, to verify and explain the technical effects used in this method.

[0152] This embodiment uses a sample of a smart resume parsing platform as an example to recommend job positions to a logged-in user of the platform. The specific operation steps are as follows:

[0153] Step 1: Obtain user and job information from the intelligent recruitment system, including user name, age, education, major, university, and work experience; job data includes all positions the user has held; preset jobs include all previously held positions; perform structured processing and string mapping on the collected data, which is collected from web pages and is often represented in JSON format. One example is as follows: {"age":"34","degree":Bachelor's degree,"gender":"Male","id":"558d2bcd1dc76a2178189ce6","major":"Computer Science and Technology","workExperienceList":[{"department":null,"end_date":"2015-02","industry":"IT Services (System / Data)"}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} ≠ " " """"""" "" ... {"Department":null,"end_date":"2007-01","industry":"IT Services (System / Data / Maintenance)","position_name":"System Engineer","salary":3,"size":4,"start_date":"2004-08","type":"-"},{"department":null,"end_date":"2004-07","industry":"IT Services (System / Data / Maintenance)","position_name":"Technical Support","salary":5,"size":4,"start_date":"2003-08","type":"-"}]};

[0154] First, select the key information needed to build the heterogeneous graph from this JSON data, such as the user's ID, degree, industry, position, and major, as shown in Table 1.

[0155] Table 1: Extracting key information.

[0156] id 558d2bcd1dc76a2178189ce6 degree Undergraduate industry IT services (systems / data / maintenance) position Technical support, system engineers major Computer Science and Technology

[0157] Then, through data conversion, the JSON format is converted into the CSV or DAT format required to construct the heterogeneous graph. The converted information is shown in Tables 2 to 5.

[0158] Table 2: user-major.csv after data conversion.

[0159] user major 558d2bb91dc76a2178188c3e e-commerce 558d2bb91dc76a2178188c40 Computer Science and Technology 558d2bba1dc76a2178188c6f Computer Information Management 558d2bbc1dc76a2178188d89 Computer software 558d2bbc1dc76a2178188d9e Information and Computing Science

[0160] Table 3: Industry-position.csv after data conversion.

[0161] industry position Computer hardware Operations Manager Internet / E-commerce Operations Specialist Real Estate / Construction / Building Materials / Engineering Technical Support Internet / E-commerce Information Technology Specialist IT services (systems / data / maintenance) Software testing

[0162] Table 4: user-industry.csv after data conversion.

[0163] user industry 558d2bb91dc76a2178188c3e Computer hardware 558d2bb91dc76a2178188c40 Internet / E-commerce 558d2bba1dc76a2178188c6f Real Estate / Construction / Building Materials / Engineering 558d2bbc1dc76a2178188d89 Internet / E-commerce 558d2bbc1dc76a2178188d9e IT services (systems / data / maintenance)

[0164] Table 5: User-degree.csv after data conversion.

[0165] user degree 558d2bb91dc76a2178188c3e Undergraduate 558d2bb91dc76a2178188c40 Undergraduate 558d2bba1dc76a2178188c6f Undergraduate 558d2bbc1dc76a2178188d89 Undergraduate 558d2bbc1dc76a2178188d9e master

[0166] Table 6: User-position.csv after data conversion.

[0167] user position 558d2bb91dc76a2178188c3e Operations Manager 558d2bb91dc76a2178188c40 Operations Specialist 558d2bba1dc76a2178188c6f Technical Support 558d2bbc1dc76a2178188d89 Information Technology Specialist 558d2bbc1dc76a2178188d9e Software testing

[0168] Then create user dictionaries, job dictionaries, major dictionaries, etc., and map these strings into numerical forms that can be processed by machine learning and deep learning through word encoding. The mapping results are shown in Tables 7 to 11. Use the mapped CSV or DAT data to construct a heterogeneous graph.

[0169] Table 7: Mapped user-major.csv

[0170] user major 457 34 315 56 142 245 16 67 268 432

[0171] Table 8: Mapped industry-position.csv

[0172]

[0173]

[0174] Table 9: Mapped user-industry.csv

[0175] user industry 457 34 315 56 142 234 16 214 268 214

[0176] Table 10: Mapped user-degree.csv

[0177] user degree 457 1 315 1 142 1 16 1 268 2

[0178] Table 11: Mapped user-position.csv

[0179] user position 457 498 315 234 142 255 16 382 268 454

[0180] Step 2: Designing a suitable heterogeneous graph is also crucial. This graph should ideally contain as much information as possible related to users and job roles, while minimizing noisy information to avoid reducing accuracy and wasting computational resources. A suitable graph could be designed as follows: Figure 3 The heterogeneous graph shown has five types of nodes: user, education level, job title, department, and major. These five types of nodes can summarize the key information users need when looking for a job. There are ten types of edges: user-job title, job title-user, user-education level, education level-user, user-department, department-user, user-major, major-user, department-job title, and job title-department. These ten types of edges can summarize the potential relationships between these nodes. To further explore the semantic relationships between these nodes, it is necessary to set some meta-paths. These meta-paths need to be designed manually based on experience. For example, designing a meta-path like "user-job title-user" can reveal the semantics that these two users have previously held the same job title.

[0181] Step 3: As a job recommendation model, the task of this model is to predict the probability of the user-job edge in the interval (0, 1) as accurately as possible. After constructing the heterogeneous graph and setting the meta-path, the existing user-job edges are labeled as label 1. These edges are called positive edges. In order to make the model more accurate, the user-job edges are negatively sampled to randomly generate some non-existent false edges, labeled as 0. These positive and negative edges are mixed, shuffled, and divided into training set, validation set and test set according to the set ratio.

[0182] Step 4: Remove the edges from the validation and test sets in the original heterogeneous graph, keeping only the edges from the training set, resulting in a processed heterogeneous graph. Feed the processed heterogeneous graph and the set meta-paths into the TAHIN (Three-level Attention Aggregation Heterogeneous Information Network) model for training. Details of this model are as follows... Figure 5As shown; first, initialize nodes of various types; when performing embedding calculations on the target node, first collect the set of one-hop neighbors of the target node. One-hop neighbors contain nodes of various types, and project them onto the dimension of the target node using a projection matrix for subsequent calculations; then, using an attention mechanism, first calculate the attention weight of each one-hop neighbor, and then perform weighted aggregation to obtain the embedding of the target node's relational neighbor aggregation; on the set meta-path, find the target node's meta-path-based neighbors, and similarly use the attention mechanism to first calculate the attention weight of the meta-path-based neighbors, and then perform weighted aggregation to obtain the node-level attention aggregation based on the meta-path; next, calculate the attention weight of each meta-path, use the learned weight as coefficients, and fuse these semantically specific embeddings to obtain the final embedding of the target node based on the meta-path; finally, aggregate the initial embedding, the embedding of the relational neighbor aggregation, and the embedding based on the meta-path. This aggregation can use the simplest vector concatenation operation;

[0183] Step 5: After obtaining the user embedding and job embedding, use the user embedding and job embedding to predict the user's score for the job. This score is calculated by a scoring model, which is designed as a deep neural network. The input is the interaction matrix multiplied by the user embedding and job embedding, and the output is the score of the user-job node pair. The score is activated by the sigmoid function to obtain the final predicted score.

[0184] Step 6: Train the model using the training set, calculate the cross-entropy loss function using the predicted score and the labels added in step 3, reset the previous gradients to zero, and then calculate the gradients through backpropagation to update the model parameters. Continuously adjust the weight parameters to obtain the minimum loss function. Calculate the validation loss and validation accuracy using the validation set, and save the best model. Test the test set with the best model to obtain the test set loss, accuracy, and F1 score.

[0185] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for matching job candidates to positions in an intelligent recruitment system based on heterogeneous graph neural networks, characterized in that: include: The collected user data and job data are preprocessed, and a heterogeneous graph is constructed based on the preprocessed data. Negative sampling is performed on the heterogeneous graph with respect to the edge "user-position", and the training set, validation set and test set are divided according to the ratio, and meta-paths are set; The nodes in the heterogeneous graph are learned based on the heterogeneous graph neural network model to obtain user node embeddings and job node embeddings. The user node embedding and the job node embedding are input into the scoring model to obtain the predicted score between the user and the job. The heterogeneous graph neural network model and the scoring model are trained using the cross-entropy loss function. A user and a preset job are input into the trained model to obtain the best Top-K job recommendations. The acquisition of user node embedding and job node embedding includes, Initialize the heterogeneous graph Embedding of different types of nodes, among which, Represents a set of nodes. This represents a set of edges, where different node types correspond to different dimensions. e Initialize embedding as ,node w Initialize embedding as ,node j Initialize embedding as ; Using projection matrix This projects different types of neighboring nodes of the target node onto the same dimension of the target node. The attention weights of the target node and its other neighboring nodes are calculated using the following formula: , in, Represents a node e All one-hop neighbor nodes; The one-hop neighbor aggregation of the target node is performed according to the calculated attention weights, as follows: ; Two nodes can be connected via the meta-path. The attention weights of neighboring nodes based on the meta-path can be calculated using the following formula: , in, Metapath upper node e The set of all neighboring nodes; Aggregation based on the meta-path using an attention mechanism The information of the neighboring nodes is represented as follows: ; The attention score for each metapath is calculated using the following formula: , in, Represents a semantic-level attention vector. Represents the weight matrix. This represents the aggregated information of neighboring nodes for a given meta-path. Represents the deviation vector. Indicates the number of metapaths; After obtaining the attention score for each metapath, the attention score for each metapath is normalized using the softmax function, as follows: , in, Metapath Attention weights; The learned weights are used as coefficients to fuse semantically specific embeddings to obtain nodes. e The final embedding based on the meta-path is represented as: ; The node e Initialize embedding What I learned and what I learned Aggregate to node e In, obtain the node e The final embedding is represented as: , in, and Represents the weight matrix. and Represents the deviation vector; After updating the heterogeneous graph node embedding, the user node embedding is obtained. and job node embedding .

2. The job matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in claim 1, characterized in that: The construction of the heterogeneous graph includes, The collected user data includes the user's name, age, education, major, university, and work experience. The collected job data includes all positions the user has held. The preset job data includes all previously held positions collected, denoted as... ; The collected user data and job data are then processed in a structured manner; Map the strings in the structured data to numerical values; Based on the processed data, construct a heterogeneous graph, and design the node types and names, as well as the edge types and names of the heterogeneous graph.

3. The job matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in claim 2, characterized in that: Negative sampling of the heterogeneous graph with respect to the "user-position" edge includes, The "user-position" edges of the generated heterogeneous graph are represented using triples; The generated heterogeneous graph is negatively sampled with respect to the "user-job" edge, that is, a certain number of negative edges are randomly constructed, positive edges and negative edges are merged, the order of the triples is shuffled, and the training set, validation set and test set are divided according to a certain ratio; Remove the edges of the validation set and the test set from the heterogeneous graph to obtain the processed heterogeneous graph, and save the processed heterogeneous graph, the training set, the validation set, and the test set. Set the meta-paths required when training heterogeneous graph neural networks.

4. The job matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in claim 3, characterized in that: The acquisition of the predicted score between the user and the job includes, Embed the user node and job node embedding The scoring model is used to obtain the user's predicted score for the job, expressed by the formula: , in, This represents the user's predicted score for the job position. This represents the scoring model.

5. The job matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in claim 4, characterized in that: The step of training the heterogeneous graph neural network model and the scoring model using the cross-entropy loss function includes... The cross-entropy loss function The calculations include, , in, This represents the node pairs representing all users and roles; Update all weight parameters of the heterogeneous graph neural network model and the scoring model, and continuously correct the weight parameters during the iteration process in order to obtain the minimum loss function, and save the trained heterogeneous graph neural network model and the scoring model.

6. The job matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in claim 5, characterized in that: The acquisition of the best Top-K job recommendations includes, Calculate target users and all preset positions The predicted scores between them are first obtained by using the heterogeneous graph neural network model to obtain the user node embeddings. and job node embedding Then substitute it into the trained scoring model; The calculation of the trained scoring model includes, , All predicted scores are sorted to obtain the Top-K job postings for recommendation.

7. A job-person matching system for an intelligent recruitment system based on heterogeneous graph neural networks, implementing the job-person matching method for an intelligent recruitment system based on heterogeneous graph neural networks as described in any one of claims 1 to 6, characterized in that, include: The heterogeneous graph construction unit is used to preprocess the collected user data and job data, construct a heterogeneous graph based on the preprocessed data, perform negative sampling on the heterogeneous graph about the edge "user-job", divide the training set, validation set and test set according to the ratio, and set the meta path. The scoring prediction unit is used to learn the nodes in the heterogeneous graph based on the heterogeneous graph neural network model to obtain user node embeddings and job node embeddings. The user node embeddings and job node embeddings are input into the scoring model to obtain the predicted score between the user and the job. The job recommendation unit is used to train the heterogeneous graph neural network model and the scoring model using the cross-entropy loss function. By inputting a user and a preset job into the trained model, the best Top-K job recommendations are obtained.

8. A device, characterized in that, The device includes, processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method described in any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images