A bidirectional employment recommendation system and method based on a graph neural network and a collaborative attention mechanism

Abstract

The application belongs to the field of recommendation system, and discloses a bidirectional employment recommendation system and method based on a graph neural network and a collaborative attention mechanism, aiming at the limitations of the existing employment recommendation method, such as the insufficient distinction of the contribution of different work experience of candidates to the demand of a specific position, the lack of consideration of the experience of recruiters and the like, the application makes the following innovations: (1) a double-view heterogeneous graph is constructed, the selection preferences of job seekers and recruiters are respectively modeled as active and passive selection nodes, so as to capture the bidirectional interaction relationship; (2) a layered ability perception module is designed, a multi-layer attention mechanism is used to analyze the matching degree between the position requirements and the experience of candidates, and the understanding of fine-grained features is enhanced; (3) a collaborative attention mechanism is introduced, the historical recruitment experience of recruiters is mapped to the current scene, and the adaptability and accuracy of the recommendation are optimized. The application can more comprehensively model the bidirectional selection behavior in the recruitment process, and significantly improve the accuracy of the employment recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a two-way employment recommendation system and method based on graph neural networks and a collaborative attention mechanism. Background Art

[0002] With the rapid development of the digital age, career matching technology has emerged. This recommendation mechanism, based on intelligent algorithms, processes massive amounts of job search data to accurately match positions with talent. The current talent market is facing the challenge of information overload, and relying solely on manual recruitment screening is time-consuming and difficult to achieve optimal matches. To address this challenge, intelligent career recommendation platforms have emerged. They can deeply analyze applicants' professional skills, career aspirations, and work experience, and compare these with job requirements across multiple dimensions, thereby creating more valuable matching recommendations for both parties.

[0003] This type of recommendation algorithm primarily relies on two dimensions of information to achieve precise recommendations: first, analyzing a job seeker's past application behavior, using collaborative filtering techniques to identify potential matches; and second, parsing the text of resumes and job descriptions, using natural language processing to extract key information points. Notably, the new generation of recommendation systems goes a step further, employing a fusion model that organically combines user interaction data with text features. This not only improves the accuracy of recommendations but also better adapts to the career development needs of different groups. This innovative matching approach is reshaping the traditional recruitment and job search model.

[0004] While existing recommendation systems have made significant progress in talent matching, they still face several key challenges. When assessing applicants' qualifications, these systems struggle to accurately measure the actual value of different work experiences to specific job requirements. They overly rely on certain quantitative metrics and overlook the variability in experience quality. Furthermore, existing algorithms have significant shortcomings when processing recruitment needs. They fail to fully integrate the recruiter's experience and struggle to effectively assess the exceptional skills that recruiters value in past successful hires. These limitations can lead to discrepancies between recommendation results and a company's actual hiring needs. Summary of the Invention

[0005] In order to overcome the above-mentioned limitations of the prior art in the contribution of the matching system to the specific position requirements and the lack of consideration of the recruiter's experience, the purpose of the present application is to provide a bidirectional employment recommendation system and method based on graph neural network and collaborative attention mechanism, which proposes hierarchical ability perception, uses multi-layer attention mechanism to deeply analyze the corresponding relationship between job requirements and candidate experience, and significantly improves the understanding ability of the system to fine-grained features; at the same time, the present application introduces a collaborative attention mechanism to map the recruiter's historical recruitment experience to the current recruitment scene, thereby effectively evaluating the outstanding skills that the recruiter values in the past successful recruitment.

[0006] In order to achieve the above-mentioned purpose, the technical solution adopted by the present application is:

[0007] A bidirectional employment recommendation system based on graph neural network and collaborative attention mechanism, comprising:

[0008] A node initialization module: the node initialization module initializes each node by combining preference embedding and text embedding, and further extracts text embedding through hierarchical ability perception function to obtain different contribution degrees of candidate work experience to specific position requirements, and finally obtains initial representation

[0009] A dual-view heterogeneous graph construction module: after the node initialization module initializes the nodes, the dual-view heterogeneous graph construction module divides the candidate and position nodes into active selection nodes and passive selection nodes respectively by creating a dual-view heterogeneous graph, and establishes edges according to bidirectional interaction, thereby modeling the bidirectional selection behavior in recruitment in detail;

[0010] A dual-view preference learning module: the dual-view preference learning module uses graph convolution network to propagate node preference information, updates the node representation constructed by the dual-view heterogeneous graph construction module, captures the multi-level interaction features between candidates and positions, and obtains the final passive node representation and the preliminary active node representation of both parties;

[0011] A recruitment experience learning module: the recruitment experience learning module introduces a collaborative attention mechanism to integrate historical recruitment experience into the current scene, optimizes the preliminary active node representation updated by the dual-view preference learning module, and further optimizes the matching accuracy and recommendation effect of the position and the candidate, i.e. the final active node representation of both parties;

[0012] A prediction and model optimization module: in the prediction stage, the bidirectional selection intention between the candidate and the position is calculated, and the final node representation of both parties obtained by the recruitment experience learning module is calculated by the inner product formula to obtain the selection intention of the candidate to the position r i→k And the selection intention of the position to the candidate s k→i , and the matching score is calculated by combining the two In the model optimization stage, an extended BPR four-tuple loss function is adopted to optimize the partial order relationship between the matching score and the non-matching score, so as to improve the accuracy of matching prediction.

[0013] The node initialization module performs the following operations:

[0014] Step 1.1: Preference embedding, in the initial stage, a random generated way is adopted, so that the model gradually adjusts the embedding in the training process, so that it is closer to the real preference pattern in the data; Specifically, a lookup table operation is used to generate a preference embedding for each node in the dual-view heterogeneous graph;

[0015] Step 1.2: Text embedding, using a BiLSTM model to model the word level in the job description and resume to obtain job requirement text representation and candidate ability representation After obtaining the word-level representation, a hierarchical ability perception function is used to further extract higher-level representations to distinguish the contribution of different work experience of the candidate to the specific job requirements;

[0016] Step 1.3: The preference embedding representation and the text embedding representation in steps 1.1-1.2 are fused as the initial representation of node n

[0017] The hierarchical ability perception performs the following operations:

[0018] Step 2.1: Perception of single ability of job requirements, first based on the job requirement text representation extracted by the BiLSTM model Attention calculation is performed to evaluate the importance of different semantic fragments in the job description, and the weight α l,t is obtained; then the weighted sum after softmax normalization processing is obtained, and the comprehensive representation of the job requirement is obtained The key semantic information of each ability requirement in the job requirement is effectively extracted;

[0019] Step 2.2: Perception and aggregation of multi-ability of job requirements, input the ability requirement representation of all jobs into the BiLSTM model to model the order information of the ability requirements and generate the hidden state representation of the job Then, the importance of each ability requirement is calculated by adding an attention layer, and the weight β t is obtained by softmax normalization, and finally the hidden states of different requirements are weighted and summed to obtain the overall requirement representation g J of the job. This step generates a comprehensive requirement representation of the job by assigning different weights to each requirement;

[0020] Step 2.3: For the single ability perception of the candidate’s experience, first, the candidate’s ability representation extracted based on the BiLSTM model Combined with specific job requirements k A single ability-aware representation of the candidate's experience is used to calculate the weight of each word in the candidate's experience using a collaborative attention mechanism. Next, the attention weight γ is obtained through the softmax function l′,k,t′ , to measure the contribution of each word to the job requirement; finally, sum the weighted word representations to obtain the candidate's contribution to the job requirement j k A single experience This allows you to filter out key information in your resume that is relevant to specific job requirements;

[0021] Step 2.4: Aggregate the multi-ability perception of the candidate's experience. First, the representation of each candidate's experience Perform mean pooling to generate semantic representations for each experience Next, the BiLSTM model is used to concatenate these semantic representations to generate the hidden state of the candidate experience sequence Then, the hidden state of each experience is calculated Importance score And obtain the attention weight δ through the softmax function t , finally, the comprehensive representation g of the candidate is obtained based on the weighted summation R , comprehensively integrate all the candidate's experience information to form a holistic representation of their abilities.

[0022] The dual-view heterogeneous graph construction module performs the following operations: each candidate and position node is divided into an active selection node and a passive selection node to represent the selection preferences of both parties. Specifically, when a candidate applies for a position but is not accepted, an edge is established between the candidate's active selection node and the position's passive selection node; when an employer contacts a candidate but is rejected, the position's active selection node is connected to the candidate's passive selection node; if the two parties reach an interview agreement, the active selection nodes and passive selection nodes of both parties are connected.

[0023] The dual-view preference learning module performs the following operations:

[0024] After obtaining the initial representation of node n Finally, the lightweight graph convolutional network LightGCN is used to learn node representation. In each layer, the representation of node n is obtained by matching the set M n and interaction set A n The preference propagation update is expressed as Finally, the node representations of all layers are averaged to obtain the final representation z for each node n n .

[0025] The recruitment experience learning module performs the following operations:

[0026] Step 3.1: Enhance the modeling of candidate ability by establishing a position-candidate relationship, map the position embedding vector in the historical recruitment record to the vector space of the current candidate, and use the collaborative attention mechanism to identify the skills that the recruiter pays more attention to. First, use the soft attention mechanism to map the active node representation of the position to the passive representation space of the candidate, and calculate the attention weight α j to obtain the weighted mapping representation V R of the position, then evaluate the matching degree of the position and the mapping representation V R of the position through the collaborative attention mechanism, and finally obtain the experience representation e J of the active node of the position by weighting the attention weight and the position node representation;

[0027] Step 3.2: Model the recruiter's preference by establishing a candidate-position relationship, map the candidate embedding in the historical successful recruitment record to the vector space of the current position, and thus obtain the recruiter's preference for skills. First, use the soft attention mechanism to map the active node representation of the candidate to the passive representation space of the position, and calculate the attention weight γ R to generate the mapping representation V J of the candidate, then evaluate the matching degree of the candidate and the mapping representation V J of the candidate through the collaborative attention mechanism, and finally obtain the experience representation e R of the active node of the candidate by weighting the attention weight and the candidate node representation.

[0028] The prediction and model optimization module performs the following operations:

[0029] Step 4.1: After learning the node representation, calculate the bidirectional selection preference: the intention r i of the candidate c k to select the position j i→k and the intention s k→i of the position j k to select the candidate c i , specifically, for the candidate c i and the position j k , use the inner product to calculate the scores of the two intentions: and wherein and are the experience representations of the active nodes of the candidate c i and the position j k , respectively, and are the positions j k and the candidates c i respectively. Finally, we predict the matching score by integrating the two intent scores:

[0030] Step 4.2: In the model optimization phase, use the BPR loss function based on the expansion to introduce a quadruple loss function for optimizing the matching problem in the recruitment scenario, the formula is as follows:

[0031]

[0032] where the set D={(i,k,i′,k')∣(i,k)∈M,(i,k')∈M - ,(i',k)∈M -} represents the training data, M and M - represent the matching and unmatched sets respectively, then the unmatched scores and are averaged, and the matching score predicted in step 4.1 is compared using the sigmoid function σ(·) to optimize the partial order relation.

[0033] A bidirectional employment recommendation method based on graph neural network and collaborative attention mechanism, comprising the following steps:

[0034] Step one: first, according to the interactive data of job seekers and employers and their respective text information, the node initialization operation is carried out, wherein the node initialization includes preference embedding and text embedding two parts, in the preference embedding part, the embedding is initialized by random generation, and the preference embedding e n is obtained for each node n in the double view heterogeneous graph through lookup table operation.

[0035] In the text embedding part, in order to model the word level representation in the position description and resume, BiLSTM model is used to calculate the word level representation of each word in the position description and resume, to obtain the position requirement text representation and the candidate ability representation to capture the semantic information of position and resume;

[0036] Step two: after obtaining the word level representation of position and resume, the hierarchical ability perception function is used to further extract higher level representation to distinguish the different contribution of different work experience to specific position requirements. Through the layer ability perception function, first, the single ability perception of position requirements is carried out, and the position requirement text representation ​Attention is calculated to evaluate the importance of different semantic segments in the job description and the overall representation of the job requirement is obtained by weighted summation Key semantic information in the job requirement is extracted, followed by multi-capability perception aggregation of the job requirement, and the capability requirement representation of all jobs (such as ) is input into the BiLSTM model to model the sequential information of the capability requirement and generate the hidden state of the capability requirement sequence Then the importance weight β of each capability requirement is learned through the attention mechanism t Finally, the overall requirement representation g of the job is generated by weighted summation J , which realizes the comprehensive representation of different capability requirements;

[0037] In the single capability perception of the candidate's experience, the collaborative attention mechanism is used to evaluate the contribution of each word in the candidate's experience to the specific job capability requirement, and the candidate capability representation extracted by the BiLSTM model Combining the single capability perception representation of the specific job requirement j k , the word weight of the candidate's experience is calculated Then, the attention weight γ is obtained by the softmax function l′,k,t′ to measure the contribution of each word to the job requirement; and the single experience representation of the candidate for the job requirement j k is generated by weighted summation Subsequently, in the multi-capability perception aggregation step, the representation of each experience of the candidate is mean-pooled to generate the semantic representation of each experience and these representations are concatenated using the BiLSTM model to generate the hidden state of the experience sequence Subsequently, the importance score of the hidden state of each experience is calculated and the attention weight δ is obtained by the softmax function t Finally, the importance weight of each experience is assigned through the attention mechanism, and the comprehensive representation g of the candidate is obtained by weighted summation R , which comprehensively integrates the relevant information in the candidate's resume to form the overall capability representation;

[0038] Step three: the preference embedding representation obtained in step one and step two is fused with the text embedding representation after hierarchical capability perception, which is used as the initial representation of node n

[0039] Step four: after the initialization of the node representation in step three, a dual-view heterogeneous graph is constructed, and each candidate and position node is set as an active and passive selection node to represent the active and passive selection preference, specifically, when the candidate applies for the position and is not accepted, an edge is established between the active node of the candidate and the passive node of the position; when the employer contacts the candidate but is rejected, the active node of the position and the passive node of the candidate are connected; if both parties reach an interview agreement, the active and passive nodes of both parties are connected;

[0040] Step five: after obtaining the initial representation of the nodes in the dual-view heterogeneous graph in step four , a light graph convolutional network LightGCN is used to learn the node representation, in the graph convolution operation of each layer, the representation of node n is updated based on the preference propagation of the matching set M n and the interaction set A n , wherein the hyperparameter ω is introduced to adjust the influence of different interaction relationships, finally, the representation z n of each node n is obtained by averaging the node representation of the L+1 layer, wherein the node representation includes the position active node representation , the candidate active node representation , the position passive node representation and the candidate passive node representation

[0041] Step six: the past recruitment experience of the recruiter is used to guide the current recruitment matching decision, first, the position-candidate relationship is established to enhance the modeling of the candidate's ability, the position embedding vector in the historical recruitment record is mapped to the vector space of the current candidate, so as to identify the skills that the recruiter pays more attention to, the specific operation includes: the position active node representation in the historical recruitment record obtained in step five is mapped to the candidate passive representation vector space through a soft attention mechanism, and the attention weight a j is calculated, then the mapping representation V R of the position is obtained by weighting the position active node representation with the weight, and the matching degree of the position and V R is evaluated through a collaborative attention mechanism, finally, the experience representation e J of the position active node is generated to provide a basis for the recruitment decision;

[0042] The candidate-position relationship is established, and the skill preference of the recruiter is modeled by introducing the historical recruitment record, the specific operation includes: the candidate embedding in the historical successful recruitment record is mapped to the vector space of the current position to obtain the preference of the recruiter for specific skills, first, the active node representation of the candidate is mapped to the passive representation vector space of the position through a soft attention mechanism, and the attention weight g R, then, the candidate mapping representation V is generated by weighting J and use the co-attention mechanism to evaluate the candidate and V J Finally, the attention weight is used to generate the empirical representation e of the candidate active node R , thereby helping recruiters identify candidate traits that match their preferences;

[0043] Step 7: Calculate the two-way choice preference: Candidate c i Select Position k Intention i→k and position j k Select Candidate c i Intentions k→i , specifically, for candidate c i and position j k , using the inner product to calculate these two intent scores: and In the formula and Candidate C i and position j k The empirical representation of the active node obtained from step 6 is and Position j k and candidate c i From the passive node representation obtained in step 5, we finally predict the matching score by integrating these two intent scores:

[0044] Step 8: Expand the BPR loss function and introduce the quadruple loss function to optimize the matching problem in the recruitment scenario. The formula is as follows:

[0045]

[0046] Among them, the set D = {(i,k,i',k')|(i,k)∈M,(i,k')∈M - ,(i',k)∈M -} represents training data, M and M - Represent the matched and unmatched sets respectively, and then score the unmatched and Average and use the sigmoid function σ(·) to compare the matching score predicted in step 7 This optimizes the partial order relationship.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1.The application constructs a dual-view heterogeneous graph, divides the position and candidate nodes into active and passive selection nodes respectively, and fully captures the selection preferences of the two-way interaction in the recruitment process. This dual-view structure can accurately model the interaction between the two parties in the recruitment process, thereby better reflecting the matching characteristics between positions and candidates and improving the matching accuracy of the recommendation system.

[0049] 2.The application introduces a hierarchical ability perception module, which gradually analyzes the fine-grained features of position requirements and candidate experience through a multi-level attention mechanism, and can distinguish the contribution of different experiences of the candidate to the requirements of each ability of the position. This hierarchical perception effectively improves the detailed understanding of the candidate's ability by the recommendation model, thereby more accurately matching the position requirements and the candidate's ability.

[0050] 3.The application uses a recruitment experience learning module to embed historical recruitment data into the current recruitment scenario using a collaborative attention mechanism to help the system identify the recruiter's preference for specific skills. This mechanism enables the recruiter's past experience to be effectively utilized in the recommendation process, thereby effectively evaluating the outstanding skills that the recruiter valued in past successful recruitment.

[0051] In summary, the application improves the accuracy between positions and candidates, the accuracy of position requirements and candidate abilities, and the effectiveness of evaluating outstanding skills that the recruiter valued in past successful recruitment. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The application is a framework diagram of a bidirectional employment recommendation system and method based on a graph neural network and a collaborative attention mechanism. DETAILED DESCRIPTION

[0053] The application will be described in detail below with reference to the accompanying drawings.

[0054] As shown in the drawings, a bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism includes: Figure 1 Node initialization module: combine preference embedding and text embedding to initialize each node, and further extract text embedding through hierarchical ability perception function to obtain different contribution of candidate work experience to specific position requirements, and finally obtain initial representation

[0055] Dual-view heterogeneous graph construction module: after the node initialization module initializes the nodes, it divides the candidate and position nodes into active and passive selection nodes by creating a dual-view heterogeneous graph, and establishes edges according to the two-way interaction, thereby modeling the two-way selection behavior in recruitment in detail.

[0056]

[0057] ​The dual-view preference learning module: uses a graph convolution network to propagate node preference information, updates the node representation constructed by the dual-view heterogeneous graph construction module through dual-view interaction, captures multi-level interaction features between candidates and positions, and obtains the final passive node representation and the preliminary active node representation of both parties;

[0058] The recruitment experience learning module: introduces a collaborative attention mechanism, integrates historical recruitment experience into the current scene, optimizes the preliminary active node representation updated by the dual-view preference learning module, and further optimizes the matching accuracy and recommendation effect of positions and candidates, i.e., the final active node representation of both parties;

[0059] The prediction and model optimization module: in the prediction stage, the bidirectional selection intention between the candidate and the position is calculated, and the final node representation of both parties obtained by the recruitment experience learning module is used to obtain the selection intention of the candidate to the position r i→k and the selection intention of the position to the candidate s k→i by the inner product formula, and the matching score is calculated by combining the two. In the model optimization stage, an extended BPR quadruple loss function is used to optimize the partial order relationship between the matching score and the unmatched score, and the accuracy of the matching prediction is improved.

[0060] The node initialization module performs the following operations:

[0061] Step 1.1: preference embedding, in the initial stage, a random generation method is used to gradually adjust the embedding in the training process, so that it is closer to the real preference pattern in the data; specifically, in order to initialize the node representation, for each node n in the dual-view heterogeneous graph, a preference embedding is generated for each node in the dual-view heterogeneous graph by a lookup table operation where d E represents the dimension of the preference embedding;

[0062] Step 1.2: text embedding, in order to model the word-level representation in the position description and the resume, a BiLSTM model is used to calculate the word-level representation of each word in the lth position requirement and the l' th candidate experience in the resume, where m l and n l′ are the number of words in the position description and the resume, respectively, to obtain the position requirement text representation and the candidate ability representation The position requirement text representation and the candidate ability representation are the d l th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l th word in the candidate experience r' O th word in the jth position requirement and the d l thThe semantic representation is obtained after obtaining the word-level representation, and higher-level representation is further extracted through hierarchical ability perception function to distinguish the contribution of different work experience of the candidate to the specific job requirements. Specifically, for the job description, each skill requirement corresponds to a specific work requirement, and the sum of these requirements summarizes the overall requirements of the job. Similarly, each experience in the candidate's resume collectively constitutes the hierarchical structure of the overall ability of the candidate. In addition, since the position and the resume usually have a clear time sequence, this feature helps to extract deep information.

[0063] Step 1.3: Fusion of preference embedding representation and text embedding representation in step 1.1-1.2 as initial representation of node n

[0064] The hierarchical ability perception performs the following operations:

[0065] Step 2.1: Perception of single ability of job requirements, first extract job requirement text representation based on the foregoing BiLSTM model Attention calculation is performed to evaluate the importance of different semantic fragments in the job description. The calculation process is as follows:

[0066]

[0067] The softmax function is used to normalize to obtain the attention weight α l,t :

[0068]

[0069] Then, the comprehensive representation of the job requirement is obtained by weighted summation after softmax normalization The key semantic information of each ability requirement in the job requirement is effectively extracted:

[0070]

[0071] Step 2.2: Perception and aggregation of multi-ability of job requirements, input the ability requirement representation (such as ) of all jobs into the BiLSTM model to model the order information of the ability requirement and generate the hidden state representation of the job

[0072]

[0073] Then, the importance of each ability requirement is calculated by adding an attention layer, and the calculation process is as follows:

[0074]

[0075] And the attention score is normalized by softmax to get the weight β of each capability requirement t :

[0076]

[0077] Finally, the hidden states of different requirements are weighted and summed to get the overall requirement representation g of the position J :

[0078]

[0079] This step generates a comprehensive requirement representation of the position by assigning different weights to each capability requirement;

[0080] Step 2.3: For the single capability perception of the candidate's experience, a new capability perception attention mechanism is designed to measure the contribution of each word in the candidate's experience to the specific capability requirement. First, based on the candidate's capability representation extracted by the BiLSTM model Combine the single capability perception representation of the specific position requirement j k , use the collaborative attention mechanism to calculate the weight of each word in the candidate's experience

[0081]

[0082] where, is the representation of the kth position requirement;

[0083] Then, the attention weight γ is obtained by the softmax function l′,k,t′ to measure the contribution of the candidate's capability representation to the kth position requirement:

[0084]

[0085] Finally, the weighted word representation is summed to get the single experience representation of the candidate for the position requirement j k Thus, the key information related to the specific position requirement in the resume is screened out:

[0086]

[0087] Step 2.4: Aggregate the multi-capability perception of the candidate's experience. First, the representation of each segment of the candidate's experience is averaged and pooled to generate the semantic representation of the lth experience

[0088] ​​

[0089] Then, considering the time relationship between these semantic vectors, we concatenate these semantic representations using a BiLSTM model to generate the hidden state of the candidate's experience sequence

[0090]

[0091] Subsequently, the hidden state of each experience is calculated The importance score of the experience

[0092]

[0093] And get the attention weight δ by the softmax function t :

[0094]

[0095] Finally, the comprehensive representation g of the candidate is obtained by weighted summation based on the weight R , which comprehensively integrates all experience information of the candidate to form the overall ability representation:

[0096]

[0097] The dual-view heterogeneous graph construction module performs the following operations: in order to model the bidirectional selection behavior in the recruitment scene, each candidate and position node is divided into an active selection node and a passive selection node to represent the selection preferences of both parties. Specifically, when a candidate applies for a position but is not accepted, an edge is established between the candidate's active selection node and the position's passive selection node; when an employer contacts a candidate but is rejected, the active selection node of the position is connected to the passive selection node of the candidate; if both parties reach an interview agreement, the active selection nodes and passive selection nodes of both parties are connected. This bidirectional structure can more comprehensively capture the bidirectional interaction characteristics in recruitment, which helps to improve the accuracy of matching.

[0098] The dual-view preference learning module performs the following operations:

[0099] After obtaining the initial representation of node n , we use a lightweight graph convolutional network LightGCN to learn the node representation. In the graph convolution operation at the l-th layer, for each node n, we consider the preference propagation from two different interaction sets: the matching set M n related to node n and the interaction set A n . Formally, we use a light propagation mechanism Lightgcn to update the node representation:

[0100]

[0101] Among them, z u and z v They represent the representation u of the neighbor nodes with matching relationship and the representation v of the neighbor nodes with unidirectional interaction relationship, respectively. n and N v The neighbor sets representing node n, u, and v represent neighbor nodes with matching relationships, respectively. A hyperparameter ω is introduced here to balance the two types of propagation. Finally, we average the node representations of the L+1 layer to obtain the final representation of each node n:

[0102]

[0103] Where L represents the number of graph convolutional layers.

[0104] The recruitment experience learning module uses the recruiter's past recruitment experience to guide current recruitment decisions during recruitment matching. Specifically, it performs the following operations:

[0105] Step 3.1: Enhance the modeling of candidate capabilities by establishing a position-candidate relationship. Use the collaborative attention mechanism to map the position embedding vectors in the historical recruitment records to the vector space of the current candidate to identify the skills that the recruiter pays more attention to. First, use the soft attention mechanism to represent the active node of the position. Mapping to the candidate passive representation space:

[0106]

[0107] Then use the attention weight α j and position active node representation The weighted position mapping representation V R , then, the mapping representation V between positions is evaluated through the collaborative attention mechanism R The matching degree is calculated as follows:

[0108]

[0109] Then use the softmax function to convert the attention score f R Convert to attention weight β, and use attention weight β and position active node representation Weighted to get the experience representation of the active node of the position e J ;

[0110] Step 3.2: By establishing a candidate-position relationship to model the recruiter's preferences, the candidate embeddings in the historical successful recruitment records are mapped to the vector space of the current position to obtain the recruiter's preference for skills. First, the active node representation of the candidate is represented by the soft attention mechanism. Map to the position passive representation space and calculate the attention weight γ R :

[0111]

[0112] Then use the attention weight γ R and candidate active nodes are represented by Weighted candidate mapping representation V J , then, the candidate-to-candidate mapping representation V is evaluated through the collaborative attention mechanism J The matching degree is calculated as follows:

[0113]

[0114] Then use the softmax function to convert the attention score f J Convert to attention weight δ and use attention weight δ with candidate active node representation Weighted to get the experience representation e of the candidate active node R .

[0115] The prediction and model optimization module performs the following operations:

[0116] Step 4.1: After learning the node representation, calculate the two-way selection preference: candidate c i Select Position k Intention i→k and position j k Select Candidate c i Intentions k→i , specifically, for candidate c i and position j k , using the inner product to calculate these two intent scores: and In the formula and Candidate C i and position j k The experience of the active node shows that and Position j k and candidate c i The passive node representation of , ultimately, we predict the matching score by integrating these two intent scores:

[0117] Step 4.2: In the model optimization stage, the BPR loss function is extended and the quadruple loss function is introduced to optimize the matching problem in the recruitment scenario. The formula is as follows:

[0118]

[0119] where D = {(i, k, i', k') | (i, k) e M, (i', k') e M - , (i', k) e M -} denotes the training data, M and M - denote the matched and unmatched sets, respectively, then the unmatched scores and are averaged, and the matched scores predicted in step 4.1 are compared using the sigmoid function σ(·) to optimize the partial order relation.

[0120] A bidirectional employment recommendation method based on a graph neural network and a collaborative attention mechanism, comprising the following steps:

[0121] Step one: first, according to the interactive data of job seekers and employers and the respective text information, the node initialization operation is carried out, wherein the node initialization includes two parts of preference embedding and text embedding, in the preference embedding part, the embedding is initialized in a random generation manner, and the preference embedding e n is obtained for each node n in the double-view heterogeneous graph through a lookup table operation.

[0122] In the text embedding part, in order to model the word-level representation in the position description and the resume, a BiLSTM model is used to calculate the word-level representation of each word in the position description and the resume, to obtain the position requirement text representation and the candidate ability representation to capture the semantic information of the position and the resume.

[0123] Step two: after obtaining the word-level representation of the position and the resume, a hierarchical ability perception module is used to further extract higher-level representations to distinguish the different contributions of different work experiences to specific position requirements. Through the layer ability perception function, first, the single ability perception of the position requirement is carried out, and the position requirement text representation extracted in step one is input into the BiLSTM model to calculate the attention, evaluate the importance of different semantic fragments in the position description, and obtain the comprehensive representation of the position requirement through weighted summation to extract the key semantic information in the position requirement, then the multi-ability perception aggregation of the position requirement is carried out, and all the ability requirement representations (such as ) of the position are input into the BiLSTM model to model the order information of the ability requirement, to generate the hidden state of the ability requirement sequence Subsequently, the importance weight β t of each ability requirement is learned through the attention mechanism, and finally the overall requirement representation g J, to achieve a comprehensive representation of different capability requirements;

[0124] In the single ability perception of the candidate experience, the collaborative attention mechanism is used to evaluate the contribution of each word in the candidate experience to the ability requirements of a specific position, and the candidate ability representation extracted by the BiLSTM model is obtained. Combined with specific job requirements k A single ability perception representation, calculating the word weight of the candidate's experience Next, the attention weight γ is obtained through the softmax function l′,k,t′ , to measure the contribution of each word to the job requirement; and generate the candidate's weighted sum for the job requirement j k A single experience Then, in the multi-ability-aware aggregation step, the representation of each candidate’s experience Perform mean pooling to generate semantic representations of each experience And use the BiLSTM model to concatenate these representations to generate the hidden state of the candidate experience sequence Then, the hidden state of each experience is calculated Importance score And obtain the attention weight δ through the softmax function t Finally, the importance weight is assigned to each experience through the attention mechanism, and the weighted sum is used to obtain the candidate's comprehensive representation g R , comprehensively integrate relevant information in the candidate's resume to form a holistic representation of capabilities.

[0125] Step 3: Combine the preference embedding representation obtained in steps 1 and 2 with the text embedding representation after layered ability perception as the initial representation of node n

[0126] Step 4: After completing the initial node representation in step 3, to model the two-way selection behavior in the recruitment scenario, a dual-view heterogeneous graph is constructed. Each candidate and position node is set as an active and passive selection node to represent active and passive selection preferences, respectively. Specifically, when a candidate's application for a position is not accepted, an edge is established between the candidate's active node and the position's passive node. When an employer contacts a candidate but is rejected, the position's active node is connected to the candidate's passive node. If both parties reach an interview agreement, the active and passive nodes of both parties are connected.

[0127] Step 5: Obtain the initial representation of the nodes in the dual-view heterogeneous graph in step 4 Finally, a lightweight graph convolutional network LightGCN is used to learn node representation. In the graph convolution operation of each layer, the representation of node n is based on the matching set M n and interaction set A nThe preference propagation is updated, in which the hyperparameter ω is introduced to adjust the influence of different interaction relationships. Finally, the representation z of each node n is obtained by averaging the node representations of the L+1 layer. n , where the node representation includes the position active node representation Active node representation of the candidate Position passive node representation and candidate passive nodes are represented by

[0128] Step 6: In the recruitment experience learning module, the recruiter's past recruitment experience is used to guide the current recruitment matching decision. First, by establishing a position-candidate relationship, the modeling of candidate capabilities is enhanced, and the position embedding vector in the historical recruitment record is mapped to the vector space of the current candidate, thereby identifying the skills that the recruiter pays more attention to. The specific operations include: using the soft attention mechanism to represent the active node of the position in the historical recruitment record obtained in step 5. Map to the candidate passive representation vector space and calculate the attention weight α j Then, the weight is added to the active node representation of the position to obtain the mapping representation V of the position R and evaluate positions and V through the collaborative attention mechanism R The matching degree of the position is finally generated, and the empirical representation of the active node is generated. J , providing a basis for recruitment decisions;

[0129] Establish candidate-position relationship and model the recruiter's skill preference by introducing historical recruitment records. Specifically, the candidate embedding in the historical successful recruitment records is mapped to the vector space of the current position to obtain the recruiter's preference for specific skills. First, the soft attention mechanism is used to represent the active node of the candidate. Map to the passive representation vector space of the position and calculate the attention weight γ R , then, the candidate mapping representation V is generated by weighting J and use the co-attention mechanism to evaluate the candidate and V J Finally, the attention weight is used to generate the empirical representation e of the candidate active node R , thereby helping recruiters identify candidate traits that match their preferences;

[0130] Step 7: After learning the node representation, calculate the two-way selection preference: candidate c i Select Position k Intention i→k and position j k Select Candidate c i Intentions k→i , specifically, for candidate ci and position j k The two intention scores are calculated using inner product: and where and are the candidate c i and position j k The active node representation from step six, and are the position j k and candidate c i The passive node representation from step five, Finally, the matching score is predicted by integrating the two intention scores:

[0131] Step eight: In the model optimization phase, the BPR loss function is extended based on the four-tuple loss function, which is used to optimize the matching problem in the recruitment scenario, and the formula is as follows:

[0132]

[0133] Where the set D = {(i, k, i', k') | (i, k) e M, (i, k') e M - , (i', k) e M -} represents the training data, M and M - represent the matching and non-matching sets respectively, then the non-matching scores and are averaged, and the matching score predicted in step seven is compared using the sigmoid function σ(·) to optimize the partial order relationship.

[0134] In order to illustrate the effectiveness of the bidirectional employment recommendation system and method based on graph neural network and collaborative attention mechanism proposed in the embodiment, a large number of experiments were carried out on the Xi'an University of Electronic Science and Technology graduate employment data set (referred to as UEMD). The data set covers multi-dimensional information of students during their school period, including academic performance, scientific research achievements, social practice, language ability, skill certification and employment signing, aiming to comprehensively evaluate the comprehensive ability of students. Compared with the relatively novel 8 methods, the effectiveness of the bidirectional employment recommendation system and method based on graph neural network and collaborative attention mechanism is verified. The indicators used in the experiment are Recall@5, Precision@5, NDCG@5 and MRR@5, and the experimental results are shown in Table 1. The experimental results show that the performance of the bidirectional employment recommendation system and method based on graph neural network and collaborative attention mechanism is better than that of other baselines in most indicators, i.e. DPGCA in Table 1, which proves again that the method proposed in the embodiment is effective.

[0135] Table 1

[0136]

[0137]

[0138] The models compared in this embodiment have been published in top journals or conferences in the field of artificial intelligence, including SIGIR, WWW, RecSys, TMIS, and UAI, etc. The compared models are: BPRMF (Bayesian Personalized Ranking from Implicit Feedback), NCF (Neural Collaborative Filtering), LightGCN (Simplifying and Powering Graph Convolution Network for Recommendation), LFRR (Latent Factor Reciprocal Recommender), PJFNN (Person-Job Fit Neural Network), BPJFNN (Basic Person-Job Fit Neural Network), APJFNN (Ability-aware Person-Job Fit Neural Network), and DPGNN (Dual-Perspective Graph Neural Network).

[0139] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any content that does not deviate from the technical solution of the present application should be included in the protection scope of the present application.

Claims

1. A bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism, characterized in that, Comprise: The node initialization module: combine the preference embedding and the text embedding to initialize each node, and further extract the text embedding through the hierarchical ability perception function to obtain the different contribution degrees of the candidate's work experience to the specific position requirements, and finally obtain the initial representation The specific execution operation is as follows: Step 1.1: preference embedding, in the initial stage, a randomly generated way is adopted, so that the model adjusts the embedding gradually in the training process, so that it is closer to the real preference mode in the data; Specifically, the preference embedding of each node in the dual-view heterogeneous graph is generated by lookup table operation; Step 1.2: Text Embedding, using BiLSTM model to model the word level in job description and resume to get job requirement text representation and candidate ability representation After obtaining the word-level representation, further extract higher-level representation through hierarchical ability perception function to distinguish the contribution of different work experience of candidates to the specific job requirements; Step 1.3: Fusing the preference embedding representation and the text embedding representation in Step 1.1-Step 1.2 as the initial representation of node n In step 1.2, the hierarchical capability perception is executed as follows: Step 2.1: Perception of single ability of job requirements, first based on the job requirement text representation extracted by BiLSTM model Attention calculation is performed to evaluate the importance of different semantic segments in the job description, and the weight α is obtained l,t ; then the weighted sum is obtained after normalization processing by softmax The key semantic information of each ability requirement in the job requirement is effectively extracted; Step 2.2: Perceive the multi-capability of the job requirement, input the capability requirement representation of all jobs into the BiLSTM model to model the sequential information of the capability requirement and generate the hidden state representation of the job Then, the importance of each capability requirement is calculated by adding an attention layer, and the weight β is obtained by softmax normalization t Finally, the hidden states of different requirements are weighted and summed to obtain the overall requirement representation g of the job J This step generates a comprehensive requirement representation of the job by assigning different weights to each capability requirement; Step 2.3: For the single ability perception of the candidate experience, first, the candidate ability representation extracted based on the BiLSTM model combined with the specific job requirement j k The single ability perception representation is used to calculate the weight of each word in the candidate experience using the collaborative attention mechanism Then, the attention weight γ is obtained by the softmax function l′,k,t′ to measure the contribution of each word to the job requirement; finally, the weighted word representation is summed up to obtain the single experience representation of the candidate for the job requirement j k Thus, the key information related to the specific job requirement in the resume is screened out;​ Step 2.4: Aggregate the multi-ability perception of the candidate's experience. First, the representation of each candidate's experience Perform mean pooling to generate semantic representations for each experience Next, the BiLSTM model is used to concatenate these semantic representations to generate the hidden state of the candidate experience sequence Then, the hidden state of each experience is calculated Importance score And obtain the attention weight δ through the softmax function t , finally, the comprehensive representation g of the candidate is obtained based on the weighted summation R , comprehensively integrate all the candidate's experience information to form a holistic representation of ability; The dual-view heterogeneous graph construction module: after the node initialization module initializes the nodes, the candidate and position nodes are divided into active selection nodes and passive selection nodes respectively by creating a dual-view heterogeneous graph, and edges are established according to the bidirectional interaction, so as to model the bidirectional selection behavior in recruitment in detail; The dual-view preference learning module: using graph convolution network to propagate node preference information, updating the node representation constructed by the dual-view heterogeneous graph construction module, capturing the multi-level interaction features between candidates and positions, and obtaining the final passive node representation and preliminary active node representation of both parties; The recruitment experience learning module: introducing a collaborative attention mechanism, integrating historical recruitment experience into the current scene, optimizing the preliminary active node representation updated by the dual-view preference learning module, so as to further optimize the matching accuracy and recommendation effect of positions and candidates, that is, the final active node representation of both parties; Prediction and model optimization module: in the prediction stage, the bidirectional selection intention between the candidate and the position is calculated, and the final node representation of both parties obtained by the recruitment experience learning module is optimized to obtain the selection intention of the candidate to the position r i→k and the selection intention of the position to the candidate s k→i , and the matching score is calculated by synthesizing the two In the model optimization stage, the extended BPR four-tuple loss function is used to optimize the partial order relationship between the matching score and the unmatched score, and the accuracy of the matching prediction is improved.

2. The bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism according to claim 1, characterized in that, The dual-view heterogeneous graph construction module executes the following operations: each candidate and position node is divided into active selection node and passive selection node to represent the selection preference of both parties, specifically, when the candidate applies for the position but is not accepted, an edge is established between the active selection node of the candidate and the passive selection node of the position; When the employer contacts the candidate but is refused, connect the active selection node of the position with the passive selection node of the candidate; If both parties reach an interview agreement, connect the active and passive nodes of both parties.

3. The bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism according to claim 1, characterized in that, The dual-view preference learning module executes the following operations: After obtaining the initial representation of node n Finally, the lightweight graph convolutional network LightGCN is used to learn node representation. In each layer, the representation of node n is obtained by matching the set M n and interaction set A n The preference propagation update is expressed as Finally, the node representations of all layers are averaged to obtain the final representation z for each node n n .

4. The bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism according to claim 1, characterized in that, The recruitment experience learning module executes the following operations: Step 3.1: Enhance the modeling of candidate ability by establishing a position-candidate relationship, use a collaborative attention mechanism to map the position embedding vector in the historical recruitment record to the vector space of the current candidate to identify the skills that the recruiter pays more attention to, first map the active node representation of the position to the passive representation space of the candidate through a soft attention mechanism, and calculate the attention weight a Map the position's mapping representation V $ to the passive representation space of the candidate, and calculate the attention weight a R , then evaluate the matching degree of the position and the mapping representation V R of the position through a collaborative attention mechanism, and finally weight the attention weight with the position node representation to obtain the experience representation e J of the position active node; Step 3.2: To model the recruiter's preference by establishing candidate-job relationship, the candidate embedding in historical successful hiring records is mapped to the vector space of the current job, thus obtaining the recruiter's preference on skills. First, the active node representation of the candidate is mapped to the passive representation space of the job using a soft attention mechanism, and the attention weight γ is calculated to generate the mapped representation V of the candidate R weighted J . Then, the matching degree of the candidate and the mapped representation V of the candidate is evaluated by a co-attention mechanism, and finally the experience representation e of the candidate active node is obtained by weighting the attention weight and the candidate node representation J . R ​ 5. The bidirectional employment recommendation system based on a graph neural network and a collaborative attention mechanism according to claim 1, characterized in that, The prediction and model optimization module executes the following operations: Step 4.1: After learning the node representation, calculate the two-way selection preference: candidate c i Select Position k Intention i→k and position j k Select Candidate c i Intentions k→i , specifically, for candidate c i and position j k , using the inner product to calculate these two intent scores: and In the formula and Candidate C i and position j k The experience of the active node shows that and Position j k and candidate c i The passive node representation of , ultimately, we predict the matching score by integrating these two intent scores: Step 4.2: in the model optimization stage, use BPR loss function for extension, introduce four tuple loss function, used to optimize the matching problem in the recruitment scene, the formula is as follows: where D = {(i, k, i', k') | (i, k) G M, (i, k') G M D} denotes the training data, M and M D} denote the matched and unmatched sets, respectively, then the unmatched scores D and are averaged and compared to the matched scores predicted in step 4.1 using the sigmoid function σ(·) to optimize the partial order relation.​ 6. A bidirectional employment recommendation method based on a graph neural network and a collaborative attention mechanism, characterized in that, Comprise the following steps: Step one: first, according to the interactive data of the job seeker and the recruiter and the respective text information, the initialization operation of the node is carried out, wherein the node initialization includes two parts of preference embedding and text embedding, in the preference embedding part, the embedding is initialized in a random generation manner, and the preference embedding e of each node n in the double view heterogeneous graph is obtained through lookup table operation n ; In the text embedding part, in order to model the word-level representation in the job description and the resume, the BiLSTM model is used to calculate the word-level representation of each word in the job description and the resume, and the job requirement text representation is obtained and the candidate ability representation to capture the semantic information of the job and the resume; Step 2: After obtaining the word-level representation of the position and resume, we use the hierarchical ability perception function to further extract higher-level representations to distinguish the different contributions of different work experiences to specific position requirements. Through the hierarchical ability perception function, we first perform single ability perception of the position requirements, and then use the BiLSTM model to extract the position requirement text representation in step 1. Perform attention calculation to evaluate the importance of different semantic segments in the job description and obtain a comprehensive representation of the job requirements through weighted summation Extract key semantic information from job requirements, then perform multi-ability perception aggregation of job requirements, input the capability requirements of all positions into the BiLSTM model to model the sequential information of capability requirements, and generate the hidden state of the capability requirement sequence The importance weight β of each capability requirement is then learned through the attention mechanism t , and finally the weighted summation generates the overall demand representation g of the position J , to achieve a comprehensive representation of different capability requirements; In the single ability perception of the candidate's experience, the contribution of each word in the candidate's experience to the specific position ability requirement is evaluated by using the collaborative attention mechanism, and the candidate ability representation extracted by the BiLSTM model Combined with the single ability perception representation of the specific position requirement j k The word weight of the candidate's experience is calculated Then, the attention weight γ is obtained by the softmax function l′,k,t′ to measure the contribution of each word to the position requirement; and weighted summation generates the single experience representation of the candidate for the position requirement j k Subsequently, in the multi-ability perception aggregation step, the representation of each experience of the candidate is Mean-pooled to generate the semantic representation of each experience And use the BiLSTM model to concatenate these representations to generate the hidden state of the candidate's experience sequence Subsequently, the importance score of the hidden state of each experience is calculated And the attention weight δ is obtained by the softmax function t Finally, the importance weight of each experience is assigned by the attention mechanism, and the weighted summation of the comprehensive representation g of the candidate R integrates the relevant information in the candidate's resume to form the overall ability representation;​ Step three: fuse the preference embedding representation obtained in step one, step two with the text embedding representation after hierarchical capability perception as the initial representation of node n Step four: after completing the initialization of the node representation in step three, build a dual-view heterogeneous graph, set each candidate and position node as active and passive selection nodes to represent the active and passive selection preference respectively, specifically, when the candidate applies for the position but is not accepted, an edge is established between the active node of the candidate and the passive node of the position; When the employer contacts the candidate but is refused, connect the active node of the position with the passive node of the candidate; If both parties reach an interview agreement, connect the active and passive nodes of both parties; Step five: obtaining initial representations of nodes in the dual-view heterogeneous graph in step four After that, a lightweight graph convolutional network LightGCN is used to learn the node representations, in the graph convolution operation at each layer, the representation of node n is updated based on the preference propagation of the matching set M n and the interaction set A n , where the hyperparameter ω is introduced to adjust the influence of different interaction relationships, finally, the representation z n of each node n is obtained by averaging the node representations of the L+1 layer, where the node representation includes the active node representation of the position the active node representation of the candidate the passive node representation of the position and the passive node representation of the candidate Step six: using the recruiter's past recruitment experience to guide the current recruitment matching decision, first, by establishing the position-candidate relationship, enhancing the modeling of the candidate's ability, embedding the position in the historical recruitment record into the vector space of the current candidate, so as to identify the skills that the recruiter pays more attention to, the specific operation includes: through the soft attention mechanism, the position active node representation in the historical recruitment record obtained in step five is mapped to the vector space of the candidate passive representation, and the attention weight α is calculated Mapping to candidate passive representation vector space and calculating attention weight α $ Then, the mapping representation V of the position is obtained by weighting the weight and the position active node representation R And the matching degree of the position and V R Is evaluated through the collaborative attention mechanism, and finally the experience representation e of the position active node is generated J , which provides the basis for recruitment decision. establishing a candidate-job relationship, modeling the skill preferences of the recruiter by introducing historical recruitment records, including: obtaining the recruiter's preference for a specific skill by mapping the candidate's embedding in the historical successful recruitment record to the vector space of the current job, first, using a soft attention mechanism to map the active node representation of the candidate to the passive representation vector space of the position, and calculating the attention weight γ R , then, generating the mapping representation V J of the candidate by weighting, and evaluating the matching degree of the candidate and V J using the collaborative attention mechanism, finally, generating the experience representation e R of the candidate's active node using the attention weight, so as to help the recruiter identify the preferred candidate characteristics; Step 7: Calculate the two-way choice preference: Candidate c i Select Position k Intention i→k and position j k Select Candidate c i Intentions k→i , specifically, for candidate c i and position j k , using the inner product to calculate these two intent scores: and In the formula and Candidate C i and position j k The empirical representation of the active node obtained from step 6 is and Position j k and candidate c i From the passive node representation obtained in step 5, we finally predict the matching score by integrating these two intent scores: Step eight: use BPR loss function for extension, introduce four tuple loss function, used to optimize the matching problem in the recruitment scene, the formula is as follows: where the set D = {(i, k, i', k') | (i, k) e M, (i, k') e M D , (i', k) e M D} represents the training data, M and M D represent the matched and unmatched sets, respectively, then the unmatched scores and are averaged and the matching scores predicted by step seven are compared using the sigmoid function σ(·) to optimize the partial order relation. Thus, the partial order relation is optimized.

Images