Disclosure of Invention
In order to solve the problems, the invention aims to provide a bidirectional recommendation system and a bidirectional recommendation method which can support recommendation requirements in an online recruitment market, simultaneously meet consideration requirements for independent preference of both job seekers and recruiters and realize high-accuracy prediction of reciprocal matching between users.
The invention provides a bidirectional recommendation system for online recruitment, which comprises: the system comprises an input module, a job seeker feature extraction module, a recruiter feature extraction module, a feature distribution module, a job seeker multi-stage behavior coding module, a recruiter multi-stage behavior coding module and an output module; wherein, the liquid crystal display device comprises a liquid crystal display device,
the input module converts original features to obtain input features, and inputs the input features to the job seeker feature extraction module and the recruiter feature extraction module respectively, wherein the original features are features related to job seekers 'job seekers and recruiters' recruitments;
the job seeker feature extraction module and the recruiter feature extraction module respectively extract job seeker feature information and recruiter feature information from the input features;
the feature distribution module performs feature aggregation on the extracted feature information of the job seeker and the recruiter, and distributes the aggregated feature information to the job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module;
the job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module respectively process the aggregated characteristic information to generate a plurality of job seeker behavior probability values and a plurality of recruiter behavior probability values;
and the output module calculates and outputs a bidirectional matching result of the job seeker and the recruiter according to the plurality of job seeker behavior probability values and the plurality of recruiter behavior probability values.
As a further improvement of the present invention, the original features include job seeker attribute information, job seeker preference information, recruiter attribute information, recruiter preference information, information on interactions of the job seeker with the recruiter, and environmental information; all original features are classified into category type features and numerical type features.
As a further refinement of the invention, the input module converts the original feature into a one-hot encoded vector and takes the one-hot encoded vector as the input feature.
As a further refinement of the present invention, the converting the original feature into the one-hot encoded vector by the input module includes:
the input module builds a dictionary of all the category type features, and obtains the index of the dictionary corresponding to each category type feature through mapping;
the input module performs barrel division processing on all the numerical type features, and obtains indexes of barrels to which each numerical type feature belongs after mapping;
the input module converts the indexed category type features and numerical type features into independent heat coding vectors.
As a further improvement of the invention, the job seeker feature extraction module and the recruiter feature extraction module have the same structure and operate independently, both comprise a feature embedding layer and a feature interaction layer,
the feature embedding layer outputs a learnable dense vector corresponding to an input feature, and inputs the obtained learnable dense vector to the feature interaction layer;
and the feature interaction layer performs feature interaction on the same input feature to obtain a plurality of feature fusion vectors after interaction, and outputs the plurality of feature fusion vectors to the feature distribution module.
As a further improvement of the invention, the feature interaction layer adopts a mixed expert model, and independent nonlinear feature interaction is carried out on the same input feature by a plurality of experts, so as to obtain a plurality of feature fusion vectors after interaction.
As a further refinement of the present invention, the feature distribution module comprises a feature aggregation layer and a feature adaptive selector,
the feature aggregation layer performs feature aggregation on the extracted feature information of the job seeker and the recruiter to obtain aggregated feature information, and inputs the aggregated feature information to the feature adaptive selector;
the characteristic self-adaptive selector self-adaptively learns weights through a neural network and distributes the aggregated characteristic information to the job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module.
As a further improvement of the invention, the job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module have the same structure and operate independently;
the job seeker multi-stage behavior coding module comprises a job seeker multi-behavior coding hidden layer, wherein the job seeker multi-behavior coding hidden layer independently codes a plurality of behaviors of the job seeker according to characteristic information of the job seeker, and generates a corresponding behavior probability value for each behavior of the job seeker;
the recruiter multi-stage behavior coding module comprises a recruiter multi-behavior coding hidden layer, wherein the recruiter multi-behavior coding hidden layer independently codes a plurality of behaviors of the recruiter according to the recruiter characteristic information, and generates a corresponding behavior probability value for each behavior of the recruiter.
As a further improvement of the present invention, the action probability value corresponding to the previous action and the action probability value corresponding to the current action are multiplied by a conditional probability formula to obtain a first conditional probability value, where the first conditional probability value is used for being input into a classification loss function corresponding to the current action to calculate so as to output a classification result corresponding to the current action, the first conditional probability value and the action probability value corresponding to the next action are multiplied by the conditional probability formula to obtain a second conditional probability value, and the second conditional probability value is used for the next multiplication until the output module outputs a final classification result, where the previous action, the current action and the next action are one of a plurality of job seekers and a plurality of recruiters, and the final classification result represents a bidirectional matching result of the job seekers and the recruiters.
The invention also provides a bidirectional recommendation method for online recruitment, which comprises the following steps:
feature input, namely converting original features into input features, wherein the original features are features related to job hunting behaviors of job seekers and recruitment behaviors of recruiters;
extracting features, namely respectively extracting the feature information of the job seeker and the feature information of the recruiter according to the input features;
feature distribution, namely performing feature aggregation on the feature information of the job seeker and the feature information of the recruiter, and distributing the aggregated feature information;
performing multi-stage behavior coding, namely performing coding processing on the aggregated characteristic information, and generating a plurality of job seeker behavior probability values and a plurality of recruiter behavior probability values;
and outputting results, namely calculating and outputting a two-way matching result of the job seeker and the recruiter according to the plurality of job seeker behavior probability values and the plurality of recruiter behavior probability values.
As a further improvement of the present invention, the original features include job seeker attribute information, job seeker preference information, recruiter attribute information, recruiter preference information, information on interactions of the job seeker with the recruiter, and environmental information;
all original features are classified into category type features and numerical type features.
As a further refinement of the invention, the original feature is converted into a one-hot encoded vector and the one-hot encoded vector is used as the input feature.
As a further improvement of the present invention, converting the original feature into the input feature includes:
dictionary construction is carried out on all the category type features in the original features, and indexes of the dictionary corresponding to each category type feature are obtained through mapping;
barrel division processing is carried out on all the numerical type features in the original features, and indexes of barrels to which each numerical type feature belongs are obtained after mapping;
and converting the indexed category type features and the indexed numerical type features into single-hot coding vectors.
As a further improvement of the present invention, extracting the job seeker feature information and the recruiter feature information, respectively, according to the input features includes:
and acquiring a learnable dense vector corresponding to the input feature, and performing feature interaction on the same input feature to obtain a plurality of feature fusion vectors after interaction.
As a further improvement of the invention, a mixed expert model is adopted for carrying out feature interaction, and a plurality of experts are used for carrying out independent nonlinear feature interaction on the same input feature, so as to obtain a plurality of interacted feature fusion vectors.
As a further improvement of the invention, the feature distribution comprises feature aggregation and feature self-adaptive selection distribution, the extracted feature information of the job seeker and the feature information of the recruiter are subjected to feature aggregation, self-adaptive learning weights are adopted through a neural network, and the aggregated feature information is distributed to respectively carry out multi-behavior coding of the job seeker and multi-behavior coding of the recruiter;
as a further improvement of the present invention, the multi-phase behavior code comprises a job seeker multi-phase behavior code and a recruiter multi-phase behavior code, the job seeker multi-phase behavior code and the recruiter multi-phase behavior code being independent of each other;
the job seeker multi-stage behavior coding independently codes a plurality of behaviors of the job seeker according to the characteristic information of the job seeker, and generates a corresponding behavior probability value for each behavior of the job seeker;
the recruiter multi-behavior code independently codes a plurality of behaviors of the recruiter according to the recruiter characteristic information, and generates a corresponding behavior probability value for each behavior of the recruiter.
As a further improvement of the present invention, the calculating and outputting the bidirectional matching result between the job seeker and the recruiter according to the job seeker behavior probability value and the recruiter behavior probability value includes:
the method comprises the steps that a behavior probability value corresponding to the previous behavior and a behavior probability value corresponding to the current behavior are subjected to multiplication operation through a conditional probability formula to obtain a first conditional probability value, the first conditional probability value is used for being input into a classification loss function corresponding to the current behavior to be calculated so as to output a classification result corresponding to the current behavior, the first conditional probability value and a behavior probability value corresponding to the next behavior are subjected to multiplication operation through the conditional probability formula to obtain a second conditional probability value, the second conditional probability value is used for the next multiplication operation until the output module outputs a final classification result, wherein the previous behavior, the current behavior and the next behavior are one of a plurality of job seekers and a plurality of recruiters, and the final classification result represents a bidirectional matching result of the job seekers and recruiters.
The beneficial effects of the invention are as follows: the characteristic extraction mode that the job seeker and the recruiter are mutually independent is adopted, two input characteristic forms are different, independent characteristic information is extracted by preference of independent user groups, the independent characteristic information is distributed to the independent behavior coding models through the independent user coding models and the self-adaptive learning weights, so that the respective behaviors of the job seeker and the recruiter are reasonable, the characteristic information of the job seeker and the recruiter is fully utilized, the requirement of simultaneously watching independent preferences of the job seeker and the recruiter is met, the problem that a common recommendation system only pays attention to high unidirectional matching and low bidirectional matching caused by unilateral preference is avoided, and the accuracy and efficiency of mutual matching between the job seeker and the recruiter are improved.
After the preference probabilities of the job seeker and the recruiter are changed, the expression of the reciprocity property in the reciprocity recommendation is realized in a product mode through a conditional probability formula and by maximally utilizing the coding information of the multi-stage behaviors of the job seeker and the recruiter; and because of the existence of multi-stage behaviors, the output is a multi-dimensional vector, a multi-task model is adopted at the same time, and the front-end behavior of the matching behavior is designed as an auxiliary task to strengthen the learning of the matching prediction main task, so that the bidirectional recommendation behavior is more reasonable and accurate, and the high-accuracy prediction of reciprocal matching between the job seeker and the recruiter is improved.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by one of ordinary skill in the art without inventive effort, based on the embodiments of the present invention are intended to be within the scope of the present disclosure.
It should be noted that, if directional indications (such as up, down, left, right, front, and rear … …) are included in the embodiments of the present invention, the directional indications are merely used to explain the relative positional relationship, movement conditions, etc. between the components in a specific posture (as shown in the drawings), and if the specific posture is changed, the directional indications are correspondingly changed.
In addition, in the description of the present invention, the terminology used is for the purpose of illustration only and is not intended to limit the scope of the present disclosure. The terms "comprises" and/or "comprising" are used to specify the presence of elements, steps, operations, and/or components, but do not preclude the presence or addition of one or more other elements, steps, operations, and/or components. The terms "first," "second," and the like may be used for describing various elements, do not represent a sequence, and are not intended to limit the elements. Furthermore, in the description of the present invention, unless otherwise indicated, the meaning of "a plurality" is two or more. These terms are only used to distinguish one element from another element. These and/or other aspects will become apparent to those of ordinary skill in the art from a review of the following drawings and description of embodiments of the invention. The drawings are intended to depict embodiments of the disclosure for purposes of illustration only. Those skilled in the art will readily recognize from the following description that alternative embodiments of the illustrated structures and methods of the present invention may be employed without departing from the principles of the present disclosure.
The embodiment of the invention provides a bidirectional recommendation system for online recruitment, as shown in fig. 1, the system comprises: the system comprises an input module, a job seeker feature extraction module, a recruiter feature extraction module, a feature distribution module, a job seeker multi-stage behavior coding module, a recruiter multi-stage behavior coding module and an output module; wherein, the liquid crystal display device comprises a liquid crystal display device,
the input module converts the original characteristics to obtain input characteristics, and inputs the input characteristics to the job seeker characteristic extraction module and the recruiter characteristic extraction module respectively; the original features are features related to the job seeker's job hunting behavior and the recruiter's recruitment behavior;
the recruiter characteristic extraction module and the recruiter characteristic extraction module respectively extract the recruiter characteristic information and the recruiter characteristic information from the input characteristics, the extraction of the recruiter characteristic and the extraction of the recruiter characteristic are independent of each other and do not affect each other, the extraction of the recruiter characteristic information is extracted by the recruiter characteristic extraction module, and the extraction of the recruiter characteristic information is extracted by the recruiter characteristic extraction module;
the feature distribution module is used for carrying out feature aggregation on the extracted feature information of the job seeker and the feature information of the recruiter, and distributing the aggregated feature information to the multi-stage behavior coding module of the job seeker and the multi-stage behavior coding module of the recruiter;
the job seeker multi-stage behavior coding module processes the aggregated characteristic information to generate a plurality of job seeker behavior probability values; the recruiter multi-stage behavior coding module processes the aggregated characteristic information to generate a plurality of recruiter behavior probability values, wherein each behavior probability value represents the occurrence probability of the current behavior predicted under the condition that the previous behavior occurs, the current behavior can be the job seeker behavior or the recruiter behavior, and the previous behavior can be the job seeker behavior or the recruiter behavior.
And the output module calculates and outputs a bidirectional matching result of the job seeker and the recruiter according to the plurality of job seeker behavior probability values and the plurality of recruiter behavior probability values.
In an alternative embodiment, the input module processes the original features, where the original features include job seeker attribute information, job seeker preference information, recruiter attribute information, recruiter preference information, information about interactions between job seekers and recruiters, and environmental information, and all the original features can be classified into category type features and numerical type features.
In an alternative embodiment, the input module converts the original feature into a single thermal encoding vector, and inputs the single thermal encoding vector as an input feature into the job seeker feature extraction module and the recruiter feature extraction module respectively, so that the job seeker feature extraction module and the recruiter feature extraction module can perform feature extraction.
The input module converting the original features into the one-hot encoded vector includes: the input module builds a dictionary of all the category type features, and obtains the index of the dictionary corresponding to each category type feature through mapping; the input module performs barrel dividing processing on all the numerical type features, wherein the barrel dividing processing can adopt equal-frequency barrel dividing, the barrel dividing quantity is respectively set according to different features, and indexes of barrels to which each numerical type feature belongs are obtained after mapping; the input module converts the indexed category type features and the indexed numerical value features into independent heat coding vectors, and the converted independent heat coding vectors are respectively input to the job seeker feature extraction module and the recruiter feature extraction module for processing.
An alternative embodiment, as shown in fig. 2, the job seeker feature extraction module and the recruiter feature extraction module are identical in structure and operate independently, each comprising a feature embedding layer and a feature interaction layer,
the feature embedding layer outputs a learnable dense vector corresponding to the input features, and inputs the obtained learnable dense vector to the feature interaction layer; the feature embedding layer may be designed as n key-value dictionaries, where keys are the corresponding n input features and value is a unified dimension learnable dense vector. And the independent heat coding vector provided by the input module returns a learning dense vector corresponding to the value in the feature embedding layer through key inquiry and is input to the downstream feature interaction layer.
The feature interaction layer performs feature interaction on the same input feature to obtain a plurality of feature fusion vectors after interaction, and outputs the plurality of feature fusion vectors to the feature distribution module, wherein the plurality of feature fusion vectors are used for representing the feature information of the job seeker and the feature information of the recruiter.
Furthermore, in order to realize that the job seeker and the recruiter can share the characteristic interaction codes to the downstream job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module for matching by self-adaptive learning weights, a characteristic interaction layer adopts a mixed expert model, and independent nonlinear characteristic interaction is carried out on the same input characteristic through a plurality of experts, so that a plurality of characteristic fusion vectors after interaction are obtained.
The feature interaction layer adopts a mode of MoE, a MoE structure is shown in fig. 3, feature extraction of job seekers and recruiters is adaptively and independently learned through a plurality of expert models, and downstream subtasks are adaptively and weight-distributed to learn through a plurality of door model structures.
In an alternative embodiment, the feature distribution module includes a feature aggregation layer and a feature adaptive selector, where the feature aggregation layer performs feature aggregation on the extracted feature information of the job seeker and the feature information of the recruiter to obtain aggregated feature information, and inputs the aggregated feature information to the feature adaptive selector; the characteristic self-adaptive selector self-adaptively learns weights through a neural network and distributes the aggregated characteristic information to the job seeker multi-stage behavior coding module and the recruiter multi-stage behavior coding module.
An alternative embodiment, as shown in fig. 2, has the same structure and operates independently of the recruiter multi-stage behavior encoding module;
the job seeker multi-stage behavior coding module comprises a job seeker multi-behavior coding hidden layer, wherein the job seeker multi-behavior coding hidden layer independently codes a plurality of behaviors of the job seeker according to characteristic information of the job seeker, n independent multi-behavior coding structures are arranged for n behaviors of each job seeker, weights are adaptively learned, and corresponding behavior probability values are generated for each behavior of the job seeker;
the recruiter multi-stage behavior coding module comprises a recruiter multi-behavior coding hidden layer, m independent multi-behavior coding structures are arranged for m behaviors of each recruiter, weights are adaptively learned, the recruiter multi-behavior coding hidden layer independently codes the behaviors of the recruiter according to the characteristic information of the recruiter, and corresponding behavior probability values are generated for each behavior of the recruiter.
The job seeker multi-stage behavior coding module comprises a processing procedure of a job seeker multi-behavior coding hidden layer, for example, the following modes are adopted: matching feature fusion vectors from a job seeker feature coding module input at the upstream, weighting and fusing, entering n three-layer feedforward neural networks using a ReLU as an activation function, and finally outputting probabilities representing the occurrence of the corresponding n behaviors after the occurrence of the previous behaviors respectively through a Sigmoid activation function. Correspondingly, the recruiter multi-stage behavior encoding module includes a recruiting rule multi-behavior encoding hidden layer processing process, for example, adopting the following manner: matching feature fusion vectors from a recruiter feature coding module input at the upstream, weighting and fusing, entering m three-layer feedforward neural networks using a ReLU as an activation function, and finally outputting probabilities representing the occurrence of corresponding m behaviors after the occurrence of the previous behaviors respectively through a Sigmoid activation function.
For example, as shown in fig. 2, for n actions of the job applicant, the job applicant multi-action coding hidden layer outputs n action probability values respectively, wherein the 1 st action probability value corresponds to a probability that the job applicant action P1 occurs after a previous action of the action P1 occurs, the 2 nd action probability value corresponds to a probability that the job applicant action P2 occurs after a previous action of the action P2 occurs, … …, the i-th action probability value corresponds to a probability that the job applicant action Pi occurs after a previous action of the action Pi occurs, and the n-th action probability value corresponds to a probability that the job applicant action Pn occurs after a previous action of the action Pn, wherein i represents an action number of the job applicant. Correspondingly, for m behaviors of the recruiter, the recruiter multi-behavior coding hidden layer outputs m behavior probability values respectively, wherein the 1 st behavior probability value corresponds to the probability that the recruiter behavior P1 occurs after the occurrence of the previous behavior of the behavior P1, the 2 nd behavior probability value corresponds to the probability that the recruiter behavior P2 occurs after the occurrence of the previous behavior of the behavior P2, … …, the i-th behavior probability value corresponds to the probability that the recruiter behavior Pi occurs after the occurrence of the previous behavior of the behavior Pi, and the n-th behavior probability value corresponds to the probability that the recruiter behavior Pn occurs after the occurrence of the previous behavior of the behavior Pn, wherein i represents the behavior number of the recruiter.
In an alternative implementation manner, the behavior probability value corresponding to the previous behavior and the behavior probability value corresponding to the current behavior are multiplied by a conditional probability formula to obtain a first conditional probability value, the first conditional probability value is used for being input into a classification loss function corresponding to the current behavior to be calculated so as to output a classification result corresponding to the current behavior, the first conditional probability value and the behavior probability value corresponding to the next behavior are multiplied by the conditional probability formula to obtain a second conditional probability value, and the second conditional probability value is used for the next multiplication until the output module outputs a final classification result, wherein the previous behavior, the current behavior and the next behavior are one of a plurality of job seekers and a plurality of recruiters, and the final classification result represents a bidirectional matching result of the job seeker and the recruiter.
The output module multiplies the behavior probability value of the job seeker and the behavior probability value of the recruiter through a conditional probability formula, inputs the result into a two-class loss function corresponding to each behavior for calculation, and finally outputs a two-way matching result of the job seeker and the recruiter. Through multiplication between two user behaviors, a primary task expression and a secondary task expression are provided that are capable of interpreting the seeker-recruiter reciprocity nature.
For example, as shown in fig. 2, the output module is configured to:
inputting the behavior probability value corresponding to the behavior P1 of the job seeker into the classification Loss function corresponding to the behavior P1, namely P1 classification Loss, to calculate;
multiplying the action probability value corresponding to the job seeker action P1 and the action probability value corresponding to the job seeker action P2 by a conditional probability formula to obtain a conditional probability value P 1 Inputting the two kinds of class Loss functions corresponding to the job seeker behaviors P2, namely P2 two kinds of classes Loss, and calculating;
the conditional probability value p 1 The action probability value corresponding to the job seeker action P3 is multiplied by a conditional probability formula to obtain a conditional probability value P 2 Inputting the two kinds of class Loss functions corresponding to the job seeker behaviors P3, namely P3 two kinds of class Loss, and calculating;
……,
conditional probability value p i-2 The action probability value corresponding to the job seeker action Pi is calculated through a conditional probability formula to obtain a conditional probability value p i-1 Inputting the two types of Loss functions corresponding to job seeker behaviors Pi, namely Pi two types of Loss functions, and calculating;
conditional probability value p i-1 The behavior probability value corresponding to the job seeker behavior Pi+1 is calculated through a conditional probability formula to obtain a conditional probability value p i Inputting the two types of Loss functions corresponding to the job seeker behavior Pi+1, namely Pi+1 types of Loss functions, and calculating;
……,
conditional probability value p n-2 The action probability value corresponding to the job seeker action Pn is calculated through a conditional probability formula to obtain a conditional probability value p n-1 And input into Pn classification Loss function corresponding to the job seeker behavior Pn, namely Pn classification Loss for calculation.
Conditional probability value p n-1 Calculating the behavior probability value corresponding to the recruiter behavior R1 through a conditional probability formula to obtain a conditional probability value R 1 Inputting the recruiter behavior R1 into R1 class Loss function corresponding to the recruiter behavior R1 to calculate;
conditional probability value r 1 Calculating the behavior probability value corresponding to the recruiter behavior R2 through a conditional probability formula to obtain a conditional probability value R 2 And input into two classification loss functions R2 corresponding to recruiter behavior R2Calculating in the classification Loss;
……,
conditional probability value r j-1 The behavior probability value corresponding to the recruiter behavior Rj is calculated through a conditional probability formula to obtain a conditional probability value r j Inputting the recruiter behavior Rj into a two-class Loss function corresponding to the recruiter behavior Rj, namely Rj two-class Loss for calculation;
……,
conditional probability value r m-1 The behavior probability value corresponding to the recruiter behavior Rm is calculated by a conditional probability formula to obtain a conditional probability value r m And inputting the result into a binary class Loss function (Rm) corresponding to the recruiter behavior (Rm) to calculate, so as to obtain a final matching result, wherein the matching result is a two-way matching result of the job seeker and the recruiter.
In an alternative implementation, a specific flow of an online recruitment scenario applied by the system is shown in fig. 4, a job seeker sends a request recommendation service to a recommendation system provided by any embodiment of the invention, the system provides a recommendation list for the job seeker, the job seeker browses the recommendation list provided by the system, the job seeker browses positions or applies positions based on the recommendation list, meanwhile, an online recruitment platform records a plurality of multi-stage behavior data of the job seeker browsing the recommendation list and applying positions, the job seeker of the application positions forms a candidate list, the recruiter browses the candidate list, and meanwhile, the online recruitment platform records the plurality of multi-stage behavior data of the application positions. The online recruitment platform imports the recorded recruiter and recruiter behavior data into the multitasking model for model training, and applies the model training result to the bidirectional recommendation system, so that the matching accuracy of the system is further improved.
The embodiment of the invention discloses a bidirectional recommendation method for online recruitment, which comprises the following steps of:
the method comprises the steps of inputting characteristics, namely converting original characteristics into input characteristics, wherein the original characteristics are characteristics related to job hunting behaviors of a job seeker and recruitment behaviors of a recruiter, and the original characteristics are characteristics related to the job hunting behaviors of the job seeker and the recruitment behaviors of the recruiter;
extracting features, namely respectively extracting the feature information of the job seeker and the feature information of the recruiter according to the input features;
feature distribution, namely performing feature aggregation on the feature information of the job seeker and the feature information of the recruiter, and distributing the aggregated feature information;
performing multi-stage behavior coding, namely performing coding processing on the aggregated characteristic information, and generating a plurality of job seeker behavior probability values and a plurality of recruiter behavior probability values; each behavior probability value represents a predicted occurrence probability of the phase under the condition that the previous phase occurred.
And outputting results, namely calculating and outputting a two-way matching result of the job seeker and the recruiter according to the plurality of job seeker behavior probability values and the plurality of recruiter behavior probability values.
In an alternative embodiment, the original features include job seeker attribute information, job seeker preference information, recruiter attribute information, recruiter preference information, job seeker and recruiter interaction information, and environmental information, and all of the original features may be categorized into category type features and numerical type features. Dictionary construction is carried out on all the category type features, and indexes of the dictionary corresponding to each category type feature are obtained through mapping; and (3) carrying out barrel division processing on all the numerical type features, mapping to obtain indexes of barrels to which each numerical type feature belongs, converting the indexed type features and the indexed numerical type features into independent heat coding vectors, and extracting the unique heat coding vectors.
In an alternative implementation manner, feature extraction of job seekers and feature extraction of recruiters are independently carried out feature embedding vectorization and feature interaction, a mixed expert model is adopted for the feature interaction, independent nonlinear feature interaction is carried out on the same input feature through a plurality of experts, a plurality of feature fusion vectors after interaction are obtained, feature information is distributed to a multi-behavior coding hidden layer to carry out multi-stage behavior coding through a neural network self-adaptive learning weight based on the feature fusion vectors, a plurality of behavior probability values corresponding to the behaviors are generated, and each behavior probability value represents the prediction occurrence probability of the stage under the condition that the stage occurs in the previous stage. Multiplying a plurality of behavior probability values representing the occurrence probability of the behavior prediction through a conditional probability formula, and inputting the multiple behavior probability values into a classification loss function corresponding to each behavior for calculation; through multiplication operation between two user behaviors, reciprocal property expression in the form of product of job seeker/recruiter behaviors is established.
That is, it will be understood by those skilled in the art that all or part of the steps in implementing the methods of the embodiments described above may be implemented by a program stored in a storage medium, where the program includes several instructions for causing a device (which may be a single-chip microcomputer, a chip or the like) or a processor (processor) to perform all or part of the steps in the methods of the embodiments described herein. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.
Furthermore, one of ordinary skill in the art will appreciate that while some embodiments described herein include some features but not others included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention and form different embodiments. For example, in the claims, any of the claimed embodiments may be used in any combination.
It will be understood by those skilled in the art that while the invention has been described with reference to exemplary embodiments, various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the invention. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the invention without departing from the essential scope thereof. Therefore, it is intended that the invention not be limited to the particular embodiment disclosed, but that the invention will include all embodiments falling within the scope of the appended claims.