CN113742563B

CN113742563B - Work prediction model establishment and work recommendation method, device, equipment and medium

Info

Publication number: CN113742563B
Application number: CN202010469752.8A
Authority: CN
Inventors: 王超; 祝恒书; 马超; 张敬帅
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2020-05-28
Filing date: 2020-05-28
Publication date: 2023-08-18
Anticipated expiration: 2040-05-28
Also published as: CN113742563A

Abstract

The application discloses a method, a device, equipment and a medium for establishing and recommending work prediction models, and relates to the technical field of artificial intelligence. The specific implementation scheme of the method for establishing the work prediction model is as follows: determining sample user resume feature representation and target work of sample user jump grooves according to sample user resume information; inputting the sample user resume feature representation into a time sequence neural network in an original model to obtain a user work experience representation, inputting the user work experience representation into a collaborative neural network in the original model to obtain a user work feature representation, inputting the candidate work feature representation and the user work feature representation into a prediction neural network in the original model to predict the probability of the sample user jumping to the candidate work; and training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model. To improve the accuracy of work recommendations.

Description

Work prediction model establishment and work recommendation method, device, equipment and medium

Technical Field

The application relates to the technical field of computers, in particular to a work prediction model building method and a work recommendation method using an artificial intelligence technology.

Background

In the fast-paced environment of the current society, the working mobility is greatly increased compared with the past, more and more working opportunities are presented to people, and job seekers tend to easily fall into mass job information on the Internet and are difficult to select. Therefore, a work recommendation system (Job Recommender System) for recommending proper work for job seekers is gradually rising. At present, the existing work recommendation system generally performs preliminary work screening on the basis of the geographical position, salary level, working time and other conditions selected by the job seeker, and then sorts the screened works according to the working heat of the current period and recommends the sorted works to the job seeker. The existing work recommendation system cannot recommend based on personalized requirements of job seekers, and recommendation results are not accurate enough.

Disclosure of Invention

Provided are a method, a device, equipment and a medium for establishing a work prediction model and recommending works.

According to an aspect of the present disclosure, there is provided a method for establishing a work prediction model, the method including:

determining sample user resume feature representation and target work of sample user jump grooves according to sample user resume information;

Inputting the sample user resume feature representation into a time sequence neural network in an original model to obtain a user work experience representation, inputting the user work experience representation into a collaborative neural network in the original model to obtain a user work feature representation, inputting the candidate work feature representation and the user work feature representation into a prediction neural network in the original model to predict the probability of the sample user jumping to the candidate work;

and training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

According to a second aspect, there is provided a work recommendation method implemented using a work prediction model established by the method of any embodiment of the present application, the method comprising:

determining a target user resume feature representation according to the target user resume information;

inputting the resume feature representation of the target user into a work prediction model to obtain the probability of the target user jumping to the candidate work;

and recommending the work for the target user according to the probability of the user jump candidate work.

According to a third aspect, there is provided an apparatus for building a work prediction model, the apparatus comprising:

the sample preprocessing module is used for determining sample user resume feature representation and target work of sample user jump grooves according to sample user resume information;

The model training module is used for inputting the sample user resume feature representation into the time sequence neural network in the original model to obtain user work experience representation, inputting the user work experience representation into the collaborative neural network in the original model to obtain user work feature representation, inputting the candidate work feature representation and the user work feature representation into the prediction neural network in the original model to predict the probability of the sample user jumping to the candidate work; and training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

According to a fourth aspect, there is provided a work recommendation device implemented using a work prediction model established by the method of any embodiment of the present application, the device comprising:

the user resume processing module is used for determining the feature representation of the target user resume according to the target user resume information;

the work prediction module is used for inputting the resume feature representation of the target user into the work prediction model to obtain the probability of the target user jumping to the candidate work;

and the work recommending module is used for recommending works for the target user according to the probability of the user jump candidate works.

According to a fifth aspect, there is provided an electronic device comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of building a work prediction model or the method of recommending work according to any of the embodiments of the present application.

According to a sixth aspect, there is provided a non-transitory computer readable storage medium storing computer instructions. The computer instructions are for causing a computer to perform the method of building a work prediction model or the method of recommending work according to any of the embodiments of the present application.

The technical scheme of the embodiment of the application solves the problem that the conventional work recommendation system cannot conduct work recommendation based on personalized demands of job seekers, and improves the accuracy of work recommendation.

It should be understood that the description in this section is not intended to identify key or critical features of the embodiments of the disclosure, nor is it intended to be used to limit the scope of the disclosure. Other features of the present disclosure will become apparent from the following specification.

Drawings

The drawings are included to provide a better understanding of the present application and are not to be construed as limiting the application. Wherein:

FIG. 1A is a flowchart of a method for building a work prediction model according to an embodiment of the present application;

FIG. 1B is a schematic diagram of the structure of an original model provided according to an embodiment of the present application;

FIG. 2 is a schematic diagram of another original model provided in accordance with an embodiment of the present application;

FIG. 3A is a flowchart of another method for building a work prediction model according to an embodiment of the present application;

fig. 3B is a schematic diagram of an operating principle of a time-series neural network of an original model according to an embodiment of the present application;

FIG. 4A is a flowchart of another method for building a work prediction model according to an embodiment of the present application;

FIG. 4B is a schematic diagram of the working principle of a predictive neural network of an original model according to an embodiment of the application;

FIG. 5A is a flow chart of a work recommendation method provided in accordance with an embodiment of the present application;

FIG. 5B is a schematic diagram of a work prediction model provided according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a device for creating a work prediction model according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a work recommendation device according to an embodiment of the present application;

Fig. 8 is a block diagram of an electronic device for implementing a method for establishing a work prediction model or a method for recommending works according to an embodiment of the present application.

Detailed Description

Exemplary embodiments of the present application will now be described with reference to the accompanying drawings, in which various details of the embodiments of the present application are included to facilitate understanding, and are to be considered merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

FIG. 1A is a flowchart of a method for building a work prediction model according to an embodiment of the present application; FIG. 1B is a schematic diagram of the structure of an original model provided according to an embodiment of the present application; the embodiment is suitable for the situation of constructing and training a neural network model capable of executing work prediction tasks. This embodiment may be performed by means of a construction of a work prediction model configured in an electronic device, which may be implemented in software and/or hardware. As shown in fig. 1A-1B, the method includes:

s101, determining sample user resume feature representation and target work of sample user jump grooves according to sample user resume information.

The sample user resume information may be text information contained in a user resume as a sample, and optionally, the sample user resume information may include personal attribute information and work experience information in the user resume. For example, the personal attribute information may include the sex, age, self-evaluation, etc. of the user, and the work experience information may include the work duration and work attribute information of each work of the user, such as the work attribute information may include the size, type, address, profile, etc. of the work unit. The user resume feature representation may be a user resume feature characterized in the form of numbers or letters after encoding sample user resume information of the text class. The user profile representation may be in the form of a vector or matrix or the like. The target work of the sample user jump groove can be determined according to the work experience of the sample user resume, and specifically, the later work in the work experience can be sequentially used as the target work of the sample user jump groove corresponding to the former work. For example, assuming that there are three jobs in the sample user resume, the second job may be the target job of the sample user jump corresponding to the first job; the third job may be the target job of the sample user jump corresponding to the second job. The target job of the sample user jump may include, but is not limited to: sample user time to jump to the target job, company and post of the target job, etc.

Optionally, in the embodiment of the present application, the text sample user resume information may be encoded into a numeric or alphabetical sample user resume feature representation according to a preset encoding rule, and the second to last work in the work experience of the sample user resume is used as the target work of the sample user jump slot corresponding to the previous work respectively. Optionally, the preset encoding rule may be that numerical value class information in the resume information of the sample user, such as age, working age of the sample user or registration duration of each company that the sample user has spent, is directly used as the sample resume feature representation. And converting the word class information in the user resume information into standard numerical class vectors or sample user resume characteristic representation in a matrix form by adopting word segmentation processing technology and word vector coding technology (such as word2vec technology).

Optionally, because the collected data in the sample user resume information is generally messy, the embodiment of the application can preprocess the sample user resume information in advance, delete the abnormal data (such as abnormal symbols) in the sample user resume information, and execute the operation of determining the sample user resume feature representation and the target work of the sample user jump groove in the step on the residual sample user resume information so as to improve the accuracy of the determined sample user resume feature representation and the target work.

S102, inputting the sample user resume feature representation into a time sequence neural network in an original model to obtain a user work experience representation, inputting the user work experience representation into a collaborative neural network in the original model to obtain a user work feature representation, inputting the feature representation of the candidate work and the user work feature representation into a prediction neural network in the original model to predict the probability of the sample user jumping to the candidate work.

The original model may be a work prediction model which is already constructed but not trained. Optionally, as shown in fig. 1B, the original model 1 in the embodiment of the present application includes three parts including a time-series neural network 10, a cooperative neural network 11, and a prediction neural network 12. The input of the original model 1 is the input of the time-series neural network 10, the output of the time-series neural network 10 is connected with the input of the cooperative neural network 11, the output of the cooperative neural network 11 is connected with the input of the prediction neural network 12, and the output of the prediction neural network 12 is the output of the original model 1.

Optionally, in the embodiment of the present application, the user work experience representation may be a hidden variable feature corresponding to each work of the sample user extracted from the sample user resume feature; the user work characteristic representation may be an implicit variable characteristic of the user further extracted from the implicit variable characteristic corresponding to each work. Wherein, among a large number of sample users, the user work feature representations of similar sample users are also similar. The feature representation of a candidate job may be a hidden variable feature of the candidate job determined for each candidate job, wherein the feature representations of candidate jobs that are similar to candidate jobs (e.g., similar companies and/or similar posts) are also similar among a vast number of candidate jobs.

Optionally, the embodiment of the present application may be that the sample user resume feature determined in S101 is input into the time-series neural network 10 of the original model 1, the time-series neural network 10 analyzes the input sample user resume feature representation, and outputs a user experience representation, where the user experience representation may include a sub-work experience representation corresponding to each work included in the work experience of the sample user resume. The user work experience representation output by the time sequence neural network 10 is further input into the collaborative neural network 11, the collaborative neural network 11 performs collaborative sensing on the sub-work experience representation corresponding to each work, the user work feature representation corresponding to each work is output and further input into the prediction neural network 12, the prediction neural network 12 performs analysis by combining the user work feature representation corresponding to each work and the feature representation of each candidate work, and the probability of jump from the work to each candidate work is output.

Optionally, the probability of the sample user jumping to the candidate job predicted by the embodiment of the present application may include, but is not limited to: sample user time to skip to candidate job, candidate company to skip, candidate post to skip, etc.

And S103, training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

The prediction result is a probability that the prediction neural network in the original model predicts, for each work in the sample user resume, a jump from the work to each candidate work in S102.

Alternatively, this step may be to predict, for each sample user, the probability of each work being predicted from the work jump to each candidate work, except for the last work, and the target work that the sample user actually jumps from the work jump (i.e., the next work to the work) as a set of training data. Based on multiple sets of training data of each sample user, the original model is continuously trained by adopting a gradient descent method, and network parameters of a time sequence neural network, a cooperative neural network and a prediction neural network in the original model are continuously updated and optimized until the model converges, so that a working prediction model is obtained.

Optionally, after the original model is trained for a preset duration or for a preset number of times, the method and the device can use test data to test the prediction accuracy of the trained original model, and if the accuracy of the trained original model is required to be preset, the trained original model is the working prediction model.

Optionally, the work prediction model trained by the method can be embedded into a work recommendation system and used for carrying out online prediction on the probability value of the jump of the target user (such as a job seeker) to each candidate work. Specifically, the target user can upload the resume thereof to the work recommendation system, the work recommendation system can firstly determine the resume characteristic representation of the target user according to the resume characteristic of the target user, then input the resume characteristic representation of the target user into the work prediction model to obtain the probability of each candidate work of the target user predicted by the work prediction model from the current work jump slot thereof, and further select one or more candidate works with higher jump slot probability as the jump slot work to be recommended to the user.

According to the technical scheme, an original model comprising a time sequence neural network, a cooperative sensing network and a prediction neural network is constructed, a user resume feature representation determined based on sample user resume information is input to the time sequence neural network of the original model, user work experience representation is obtained and input to the cooperative neural network, user work feature representation is obtained and input to the prediction neural network, probability of a sample user jumping to candidate works is predicted, and then the original model is trained by combining with target works of actual jumping of the sample user, so that a work prediction model is obtained. Because the work prediction model of the embodiment of the application is trained based on resume information of a large number of users with different samples, when the work prediction model carries out the follow-up prediction on a jump work line aiming at a certain user, the work prediction model can cooperate with the selection of other users similar to the target user work characteristic representation in the jump process to determine the probability of the target user jump to each candidate work. The accuracy of the prediction result is greatly improved while the work recommendation based on the personalized requirements of the user can be realized.

Fig. 2 is a schematic structural diagram of another original model provided according to an embodiment of the present application. The embodiment further optimizes the structure of the constructed original model based on the embodiment, and provides a specific description of the internal structures of the time sequence neural network and the collaborative neural network of the original model. Specifically, the time sequence neural network of the original model comprises a plurality of sub-time networks, the cooperative neural network of the original model comprises a plurality of sub-sensing networks, and the sub-time networks in the time sequence neural network are in one-to-one correspondence with the sub-sensing networks in the cooperative neural network; the first output end of the sub-timing network is connected with the next sub-timing network; the second output end of the sub-time sequence network is connected with the corresponding sub-sensing network. Optionally, the sub-timing network is a long-short-period memory network LSTM; the sub-sensing network is a multi-layer sensing machine MLP. Illustratively, as shown in FIG. 2, the time-series neural network 10 of the original model 1 includes T sub-time-series networks 101, namely LSTM-1 through LSTM-T. The collaborative neural network 11 in the original model 1 includes T sub-perceptron networks 111, i.e., MLP-1 through MLP-T. And LSTM-1 corresponds to MLP-1, LSTM-2 corresponds to MLP-2, … LSTM-T corresponds to MLP-T. For each LSTM, its first output A is connected to the pass input C of the next LSTM; the second output B of the LSTM is connected to the input of its corresponding MLP. Each LSTM has, in addition to the transfer input C, a parameter input D for inputting a sample user profile representation corresponding to each job.

The method for introducing time sequence modeling in the embodiment of the application constructs an original model comprising a plurality of sub-time networks, can completely analyze the whole work experience link of the user, and excavates the work change process information of the user. Meanwhile, each sub-time sequence network corresponds to one sub-sensing network, the user work characteristic representation can be accurately extracted, and as the user work characteristics of similar users are similar, the embodiment of the application can assist in providing suggestions for professional development of the users by utilizing the selection information of the similar users on different positions, so that more accurate work prediction results are obtained.

Further, in the embodiment of the present application, the collaborative neural network of the original model includes a plurality of sub-sensing networks, and for each sub-sensing network, at least one of the first sub-sensing unit, the second sub-sensing unit, and the third sub-sensing unit is included; the input ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the second output ends of the corresponding sub-timing network, and the output ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the input ends of the prediction neural network; the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are respectively used for outputting user time feature representation, user company feature representation and user post feature representation in the user work feature representation. Alternatively, for each sub-sensing unit, it may be a set of multi-layer perceptron MLPs.

For example, as shown in fig. 2, taking the sub-sensing network MLP-1 as an example, three groups of sub-sensing units, namely, a first sub-sensing unit corresponding to a slash box, a second sub-sensing unit corresponding to a vertical line box and a third sub-sensing unit corresponding to a horizontal line box are included. Alternatively, the three groups of sub-sensing units may be formed by three groups of multi-layer sensing machines, each group of multi-layer sensing machines including an input layer, at least one hidden layer and an output layer. The input layers of the sub-sensing units are the same and are connected with the second output end B of the sub-timing network, and the output layer of each sub-sensing unit is connected with the input end of the prediction neural network. Wherein the first sub-sensing unit of the sub-sensing network MLP-1 is used for outputting the user time feature representation x in the user work feature representation through multi-layer sensing _i1 (i.e., user time feature representation corresponding to work 1 of the ith user); the second sub-perception unit is used for outputting the user company characteristic representation u in the user work characteristic representation through multi-layer perception _i1 (i.e., user company feature representation corresponding to work 1 of the ith user); third sub-perception unit 1113 is configured to output a user position feature representation w in the user work feature representation through multi-layer perception _i1 (i.e., the user post feature representation corresponding to the 1 st job of the i-th user).

It should be noted that, in the embodiment of the present application, each sub-sensing network may include at least one of the three sub-sensing units, which is not limited to this embodiment. The user work characteristics output by different sub-sensing units are used for predicting different jump operation categories, for example, the user time characteristics output by the first sub-sensing unit are used for predicting jump time, the user company characteristics output by the second sub-sensing unit are used for predicting jump company, and the user position characteristics output by the third sub-sensing unit are used for measuring jump position. The embodiment of the application has the advantages that the multidimensional user work characteristic representation can be extracted through the perception network, so that the more accurate prediction of the jump groove work for the user from the multidimensional degree is realized.

Optionally, for each sub-sensing network or each sub-sensing unit, the specific working principle may be: user work experience representation of user i's T-th work received by each sub-sense network or sub-sense unit from sub-time network transmissionAfter this, the ++can be determined by the multilayer perceptron inside it, i.e. at least one hidden layer >The corresponding user work characteristic representation. Specifically, each hidden layer may have two parameters, namely a weight coefficient and a paranoid vector. For the t-th hidden layer, it is assumed that its weight coefficient is W _t And a paranoid vector of b _t Then it can be according to formula g _t ＝f _t (W _t g _t-1 +b _t )，t∈[2,n-1]Determining the perception result g of the t-th hidden layer _t . Wherein n is the network layer number of the sub-sensing network or the sub-sensing unit, g _t-1 F, which is the result of the perception of the t-1 th hidden layer _t () Is the activation function of the t-th hidden layer. Optionally, for different hidden layers, the activation function may be selected according to requirements, for example, a Sigmoid activation function is adopted for the first n-1 hidden layers, and a tanh activation function is adopted for the last hidden layer.

Optionally, in the embodiment of the present application, the model parameters are shared by different sub-aware networks in the collaborative neural network. The model parameters may include, but are not limited to, weight coefficients and bias vectors of the network. Specifically, the collaborative neural network comprises a plurality of sub-sensing networks, and the sub-sensing networks share the weight coefficient and the bias vector of the same multi-layer sensing mechanism. Optionally, when each sub-sensing network comprises a plurality of sub-sensing units, the same type of sub-sensing units share the same weight coefficient and bias vector. The benefits of such an arrangement are seen in the first aspect in that, since the process of obtaining a representation of user operational characteristics by collaborative neural network parsing is independent of time, and similar representations of user operational characteristics represent similar users, even though they are not within the same time period, the corresponding users are identical, so that the sub-aware network processing using the same model parameters is required for the representations of user operational experiences output by different sub-time networks. On the other hand, the parameters of the model shared by different sub-sensing networks can greatly reduce the parameters needed to be learned by the cooperative neural network, so that the robustness of the model is enhanced, and the risk of over-fitting is reduced.

Fig. 3A is a flowchart of another method for creating a work prediction model according to an embodiment of the present application. Fig. 3B is a schematic diagram of an operation principle of a time-series neural network of an original model according to an embodiment of the present application. The embodiment is further optimized based on the embodiment, and specific situations of determining the sample user resume feature representation according to the sample user resume information and how to input the sample user resume feature representation into the original model are presented. As shown in fig. 3A-3B, the method includes:

s301, determining sample user attribute information, work attribute information of at least one work of a sample user and the work duration of the work according to the sample user resume information.

The sample user attribute information may be related information characterizing a static attribute of the sample user, and may include, but is not limited to, age, gender, age, personal assessment, and the like of the sample user, for example. The job attribute information may be information about each job in the sample user's job experience, and may include, for example, but not limited to, the size, type, job positions, annual outflow rates of company employees, and company introduction text, among others.

Optionally, the embodiment of the application can classify the sample user resume information, and screen out the information of age, sex, age, personal evaluation and the like representing the static attribute information of the user from the sample user resume information as the sample user attribute information. Because the working experience in the sample resume information generally comprises at least one work, the embodiment can be to screen out information such as company scale, type, job positions, annual outflow and inflow rate of company staff, company introduction text and the like representing each work from the sample user resume as the working attribute information, and screen out the working time of each work of the user.

S302, determining the static attribute characteristic representation of the sample resume according to the sample user attribute information.

Optionally, as the sample user attribute information includes two attribute information of text and numeric, the step may be to directly represent the corresponding numeric value of the sample user attribute information of numeric, such as age, service age and expected wages of the sample user, in the form of vector or matrix, as a sample resume static attribute feature representation; for sample user attribute information of characters, such as gender, personal evaluation of character content and the like, firstly, word segmentation technology is adopted to carry out word segmentation processing, then word vector coding technology (such as word2vec technology) is adopted to encode word groups after word segmentation into numerical values, and the numerical values are expressed in a vector or matrix form to be used as sample resume static attribute characteristic expression.

Optionally, because abnormal data, such as abnormal symbols, may exist in the sample user attribute information determined in S301, this step may be to clean the sample user attribute information determined in S301, delete the abnormal data contained therein, and then execute the operation of determining the static attribute feature representation of the sample resume in this step.

S303, determining the dynamic attribute characteristic representation of the sample resume according to the work attribute information of at least one work of the sample user.

Optionally, for the work attribute information of each work of the sample user, similar to the sample user attribute information, the work attribute information includes two attribute information of text type and numerical value type, so the step can be similar to S302, and for the numerical value type information in the work attribute information of each work, such as company scale, annual outflow and inflow rate of company staff, company creation year and the like, the corresponding numerical value is directly expressed in a vector or matrix form and is used as a sample resume dynamic attribute feature representation; for the working attribute information of the characters, such as formula type, address, company introduction text and the like, firstly, word segmentation technology is adopted to carry out word segmentation processing, then word vector coding technology (such as word2vec technology) is adopted to encode word groups after word segmentation into numerical values, and the numerical values are expressed in a vector or matrix form and serve as sample resume dynamic attribute characteristic expression.

Optionally, similarly, abnormal data, such as abnormal symbols, may also exist in the working attribute information, so this step may adopt a method similar to the above-mentioned S302, and the working attribute information of each job determined in S301 is cleaned first, and after the abnormal data included in the working attribute information is deleted, the operation of determining the dynamic attribute feature representation of the sample resume in this step is executed.

S304, the working time length of at least one work of the sample user is encoded, and the sample time encoding representation is determined.

Optionally, the working time length of each work of the sample user (such as one year, two years, etc.) belongs to a one-dimensional discrete time variable, and in order to deeply utilize the deep resolution capability of the neural network, the embodiment of the application needs to encode the working time length of each work of the sample user and convert the working time length into a time hidden variable in a high-dimensional continuous space, namely, a sample time code representation. In the process of encoding the working time length, the characteristic of the original time value needs to be maintained, for example, the sample time encoded representation after the working time length of 3 years and 4 years is encoded, and the euclidean distance corresponding to the high-dimensional space is closer than the euclidean distance corresponding to the sample time encoded representation after the working time length of 1 year and 6 years.

Optionally, in the embodiment of the present application, two possible implementation manners of a preset algorithm or a model learning method may be adopted to encode the working duration to obtain the sample time code representation. Specific:

in the first embodiment, the working time length is encoded by adopting a preset algorithm, and when the sample time encoding representation is determined, the one-dimensional discrete working time length can be encoded into a high-dimensional continuous sample time encoding representation by analyzing time information according to a preset formula as shown in the following formulas (1) - (2).

Wherein D is _iT The working time of the T-th work of the user i is the working time of the T-th work of the user i;and->Respectively represent D _iT Sample time encoded values in the 2j and 2j+1 dimensions, r, are the dimension sizes of the sample time encoded representation.

For example, assuming that the one-dimensional discrete operating time length is encoded into a 5-dimensional continuous time sample encoded representation, then r is 5, the sample time encoded values corresponding to 1, 3, 5 dimensions can be calculated using equation (1); and (3) calculating sample time coding values corresponding to 2 and 4 dimensions by adopting a formula (2), and taking the calculated sample time coding values of 5 dimensions as a coded 5-dimensional continuous sample time coding representation. The advantage of using this embodiment to determine the sample time encoded representation is that no more calculation is needed to learn the encoding parameters, the encoding process is simple and easy.

In the second embodiment, the model learning method is adopted to encode the working time length, when the sample time coding representation is determined, the encoding process can be fused in the learning training process of the whole original model, specifically, a time coding representation is established for each working time length and paired with the working time length, and then in the subsequent process of training the original model, the encoding parameters between the working time length and the paired time coding representation are continuously updated. The advantage of using this embodiment is that the determined sample time code represents a time code that more closely conforms to the real data distribution.

S305, taking the sample resume static attribute feature representation, the sample resume dynamic attribute feature representation and the sample time code representation as sample user resume feature representations.

Optionally, the embodiment of the present application uses the static attribute feature representation of the sample resume determined in S302, the dynamic attribute feature representation of the sample resume determined in S303, and the sample time code representation determined in S304 as the feature representation of the sample user resume.

S306, determining target work of the jump groove of the sample user according to the resume information of the sample user.

S307, inputting the sample user resume feature representation into the time sequence neural network in the original model to obtain the user work experience representation.

Optionally, in this embodiment, the initial transfer parameter may be obtained according to a sample resume static attribute feature representation; taking the initial transfer parameter as the transfer input of the first sub-time sequence network of the time sequence neural network in the original model; and sequentially taking the dynamic attribute characteristic representation and the sample time coding representation of each working sample resume and the static attribute characteristic representation of the sample resume as the parameter input of the sub-time network in the time sequence neural network. Specifically, as shown in fig. 3B, the static attribute features of the sample resume in the sample user resume features may be represented by F _is Mapping to obtain initial transfer parametersAlternatively, it may be that F _is Input to the data mapping processing layer, which can convert F _is Mapping to obtain->Then will->As a delivery input to the first sub-timing network LSTM-1. Then sequentially representing the dynamic characteristics of the sample resume of each work F _iT The working time D of the work _iT Sample time encoded representation d of (c) _iT And sample resume static attribute feature representation F _is Added hidden variable +.>As a sub-sequenceParameter inputs to the network LSTM. For example, it may be F that works the first time _i1 The working time D of the work _i1 Sample time encoded representation d of (c) _i1 And sample resume static attribute feature representation F _is Added hidden variable +.>As a parameter input for LSTM-1; second work F to second work _i2 The working time D of the work _i2 Sample time encoded representation d of (c) _i2 And sample resume static attribute feature representation F _is Added hidden variable +.>As the parameter input of LSTM-2, and so on, the T-th work F _iT The working time D of the work _iT Sample time encoded representation d of (c) _iT And sample resume static attribute feature representation F _is Added hidden variableAs parameter input of LSTM-T, i.e. to input D of LSTM-T.

Optionally, for each sub-timing network in the timing neural network, the input includes two parts, one part being the transfer input, i.e. the transfer parameter of the last sub-timing network outputAnother part is the parameter input +>The output of the method also comprises two parts, wherein one part is the transmission parameter of the current time sequence network output +.>Another part is the user work experience representation of the current sub-time network output +.>Optionally, the transfer input +.>Includes->Is a piece of information related to the information. When the sub-timing network is an LSTM network, it can be expressed as +. >To determine the output of the current sub-timing network>Andwherein (1)>Sub-aware network for outputting to the corresponding sub-time sequence network of the current sub-time sequence>And the transmission parameter is used as a transmission parameter of the current sub-time network and is output to the next sub-time network.

It should be noted that, in the embodiment of the present application, the process of obtaining the initial transfer parameters according to the sample resume static attribute feature representation, and the process of representing the sample resume dynamic attribute feature representation and the sample time code representation of each job, i.e. in fig. 3BAnd d _i1 -d _iT The exact process can be performed with the data processing layer in the original model, so long as the sample resume static attribute features are represented by F _is Sample resume dynamic feature representation F for each job _iT The working time length D of the work _iT Input sourceAnd (5) starting the model. The initial transmission parameters and the sample time code representation of each working time duration may be determined before the original model is input, and the initial transmission parameters and the sample time code representation may be directly input to the original model, which is not limited in this embodiment.

S308, inputting the user work experience representation into the collaborative neural network in the original model to obtain the user work characteristic representation.

S309, inputting the characteristic representation of the candidate work and the user work characteristic representation into a prediction neural network in the original model, and predicting the probability of the sample user jumping to the candidate work.

And 310, training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

According to the technical scheme, according to the sample user resume information, the sample resume static attribute feature representation, the sample resume dynamic attribute feature representation and the sample time coding representation are determined and used as the sample user resume feature representation, the dimension division of the sample user resume feature is finer, the sample user resume feature representation is used for training the obtained work prediction model, the sample resume static attribute feature representation representing the user attribute is considered, the sample resume dynamic attribute feature representation representing the user work experience is considered, and the accuracy of the work model prediction result obtained through training is greatly improved.

Fig. 4A is a flowchart of another method for creating a work prediction model according to an embodiment of the present application. Fig. 4B is a schematic diagram of an operating principle of a predictive neural network of an original model according to an embodiment of the application. The embodiment further optimizes the process of predicting the probability of the sample user jumping to the candidate work based on the embodiment, and specifically includes the following steps:

S401, determining sample user resume feature representation and target work of sample user jump grooves according to sample user resume information.

S402, inputting the sample user resume feature representation into a time sequence neural network in an original model to obtain a user work experience representation.

S403, inputting the user work experience representation into the collaborative neural network in the original model to obtain the user work characteristic representation.

S404, inputting the characteristic representation of the candidate work, the jump time and the user work characteristic representation into a prediction neural network in the original model, and predicting the probability of the sample user jump to the candidate work.

Optionally, in the embodiment of the application, the slot-skipping time is introduced in the process of predicting the candidate working probability of the sample user slot-skipping, and the slot-skipping time can be predicted by the prediction neural network of the original model or can be input by the user according to the self requirement.

Optionally, in an embodiment of the present application, the prediction neural network includes a skip company prediction sub-network and/or a skip post prediction sub-network. The S403 derived user work feature representation may include a user company feature representation and/or a user post feature representation. When the user work characteristic representation comprises a user company characteristic representation, the prediction neural network comprises a jump groove company prediction sub-network, and the prediction neural network can predict the probability of the jump groove of the user to each candidate company; when the user work characteristic representation comprises a user position characteristic representation, the prediction neural network comprises a jump position prediction sub-network, and the prediction neural network can predict the probability of the jump of the user to each candidate position. Specifically, the step may include at least one of the following two cases:

In the first case, the feature representation of the candidate company in the feature representation of the candidate work, the jump time and the user company feature representation in the user work feature representation are input into a jump company prediction sub-network to predict the probability of the sample user jump to the candidate company.

Specifically, as shown in FIG. 4B, in this case the predicted neural network 12 of the original model 1 includes a slot-jumping company predicted subnetwork 122, at which time the slot-jumping time D may be first set _iT+1 Coding into a code representation d of the jump time _iT+1 (specific encoding procedure the above embodiment has been described and will not be described here in detail), and then input into a layer of fully connected network according to formula e _iT+1 ＝Wd _iT+1 +b new mapped vector e _iT+1 Wherein W and b are the weight coefficients and bias vectors of the fully connected network. User company feature representation u _iT And e _iT+1 Adding to obtain final hidden variable of dynamic user companyFinally, predicting the probability of the sample user to jump to the candidate company according to the following formula (3) by using a normalized index (softmax) function.

Wherein C is _ij,T+1 For the predicted probability of sample user i jumping to candidate company j,transpose of the feature representation for candidate company j, +.>Hidden variables for dynamic user companies.

And in the second case, inputting the feature representation of the candidate post in the feature representation of the candidate work, the jump time and the user post feature representation in the user work feature representation into a jump post prediction sub-network to predict the probability of the sample user jump to the candidate post.

Specifically, as shown in FIG. 4B, in this case, the predicted neural network 12 of the original model 1 includes a skip-slot-position prediction sub-network 123, where the skip time D may be first determined _iT+1 Coding into a code representation d of the jump time _iT+1 (specific encoding procedure the above embodiment has been described and will not be described here in detail), and then input into a layer of fully connected network according to formula e' _iT+1 ＝W′d _iT+1 +b 'gets the mapped new vector e' _iT+1 Wherein, W 'and b' are the weight coefficient and the bias vector of the fully connected network. The user post feature is then represented as w _iT And e' _iT+1 Adding to obtain the final productDynamic user position hidden variableFinally, predicting the probability of the sample user to jump to the candidate post according to the following formula (4) by using a normalized index (softmax) function. />

Wherein p is _ik,T+1 For the predicted probability of sample user i jumping to candidate position k,transpose of the feature representation for candidate position k, +.>Hidden variables for dynamic user companies.

S405, training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

Optionally, in the embodiment of the present application, the method for determining the feature representation of the candidate company in the feature representation of the candidate job may be to encode attribute information of each candidate company, such as company name, address, phone profile, etc., into a vector or matrix representation, and then input the vector or matrix representation into a multi-layer perceptual network for extracting the feature representation of each candidate company to obtain the feature representation of each candidate company. Because the post information will not change with time, the method for determining the feature representation of the candidate posts according to the embodiment of the application may be to set a hidden variable feature for each candidate post in advance as the feature representation of each candidate post. The characteristic representation of the same post of the company may be the same or different, and the present embodiment is not limited thereto.

According to the technical scheme, when the probability of the sample user jumping to the candidate work is predicted, the predicted neural network for constructing the trained original model introduces a jumping time factor, and the probability of the sample user jumping to the candidate work is predicted by combining the characteristic representation of the candidate work, the jumping time and the user work characteristic representation. The corresponding dynamic work recommendation result can be given according to the dynamically-changed jump time, so that the recommended work for the user is more accurate.

Optionally, if the two situations predict the slot-jumping post and the slot-jumping time required by the slot-jumping company is predicted by the prediction neural network itself, the prediction neural network further includes: a slot time prediction sub-network; the user time feature representation also needs to be included in the user work feature representation obtained in S403. And the step of inputting the user time characteristic representation in the user work characteristic representation into the jump time prediction sub-network to predict and obtain the jump time. Specifically, as shown in FIG. 4B, the predicted neural network 12 of the original model 1 in this case includes a dead-time predicted self-network 121, where x can be expressed for the user time feature _iT Final optimal slot time prediction by a layer of fully connected neural network, optionally where the spliced layer of fully connected neural network is equivalent to the pair x _iT Logistic regression was performed. Compared with the prior art, the working prediction model trained by the embodiment of the application can also realize accurate prediction of the user slot-skipping time on the basis of predicting the slot-skipping company and the slot-skipping post, the prediction result is more abundant, and the slot-skipping time provides guarantee for the follow-up accurate prediction of the slot-skipping company and the slot-skipping post.

FIG. 5A is a flow chart of a work recommendation method provided in accordance with an embodiment of the present application; FIG. 5B is a schematic diagram of a work prediction model provided according to an embodiment of the present application; the embodiment is suitable for the situation that the work prediction model established in any embodiment is deployed into a work recommendation system to recommend work for a target user. The embodiment can be executed by an electronic device where the work recommendation system is located, wherein the electronic device is configured with a work prediction model and a work recommendation device, and the device can be implemented by software and/or hardware. As shown in fig. 5A-5B, the method includes:

s501, determining the feature representation of the resume of the target user according to the resume information of the target user.

Optionally, the target user in the embodiment of the present application may be a job seeker user. The method comprises the steps of obtaining resume information of a target user, and then determining resume feature representation of the target user according to the resume information of the target user and a similar method of determining the resume feature representation of the sample user according to resume information of the sample user when a work prediction model is trained. Optionally, the specific execution process may be determining, according to the resume information of the target user, attribute information of the target user, work attribute information of at least one work of the target user, and a work duration of the work; determining a target resume static attribute characteristic representation according to the target user attribute information; determining target resume dynamic attribute characteristic representation according to the work attribute information of at least one work of the target user; encoding the working time length of at least one work of the target user, and determining a target time encoding representation; and taking the target resume static attribute characteristic representation, the target resume dynamic attribute characteristic representation and the target time code representation as target user resume characteristic representations. It should be noted that, the process is similar to the above embodiment in that the determining of the characteristic representation of the sample user resume and the target task of the jump slot of the sample user according to the sample user resume information, and only the sample user resume in the above embodiment is required to be changed into the target user resume, and the specific determining process is the same, and details are not repeated here. In the process of extracting the pre-characteristics of the resume information of the target user, the dimension division of the resume characteristics of the extracted target user is finer, the predicted jump operation is represented by the resume characteristics of the target user, the target resume static attribute characteristic representation representing the user attribute is considered, and the target resume dynamic attribute characteristic representation representing the user work experience is considered, so that the accuracy of the prediction result of the work prediction model is greatly improved.

The target user resume feature representation determined in this step includes a target resume static attribute feature representation F, assuming that the work experience record of the resume of the target user has two works _is The dynamic attribute feature of the target resume of the first work represents F _i1 Working of first partDuration of operation D _i1 Corresponding target time encoded representation d _i1 The dynamic attribute feature of the target resume of the second job represents F _i2 Working time D of second work _i2 Corresponding target time encoded representation d _i2 。

S502, inputting the resume feature representation of the target user into the work prediction model to obtain the probability of the target user jumping to the candidate work.

The optional target user resume feature determined in S501 includes: the method comprises the steps of target resume static attribute feature representation, target resume dynamic attribute feature representation of each work in target user work experience and target time coding representation. When the work prediction model is trained, a method similar to the target user resume feature representation is input into the original model to be trained, so that the target user resume feature is input into the trained work prediction model, the work prediction model analyzes the input target user feature representation according to an algorithm during training, and the probability of target user jump to each candidate work is determined.

Specifically, the initial transfer parameters can be obtained according to the static attribute characteristic representation of the target resume; taking the initial transfer parameter as the transfer input of the first sub-time sequence network of the time sequence neural network in the work prediction model; and sequentially taking the dynamic attribute characteristic representation and the target time code representation of the target resume and the static attribute characteristic representation of the target resume of each work as the parameter input of the sub-time network in the time sequence neural network. The procedure is similar to the procedure of inputting the sample user resume feature representation into the time-series neural network in the original model described in the above embodiment, and the description is omitted here. Each sub-timing network analyzes the transmission input and the parameter input to obtain two outputs, one is a user work experience representation, the user work experience representation needs to be transmitted to a sub-sensing network corresponding to the sub-timing network in the cooperative neural network, and the other is a transmission parameter which is used as the parameter input of the next sub-timing network. Each sub-sensor network in the collaborative neural network analyzes the received user work experience representation to determine a user work feature representation, which optionally includes: at least one of a user time feature representation, a user company feature representation, and a user post feature representation. When predicting the next slot-skipping operation for the target user, the embodiment of the application can take the user operation characteristic representation output by the sub-sensing network corresponding to the last sub-time network in time sequence arrangement as the input of the prediction neural network, so that the prediction network predicts the probability of the target user from the current operation slot-skipping to each candidate operation according to the input user operation characteristic representation corresponding to the current operation. Optionally, the prediction sub-network for predicting the slot time in the neural network predicts the slot time according to the user time feature representation. The slot company prediction sub-network predicts the probability of a target user to jump to each candidate company according to the feature representation of the candidate company in the feature representation of the candidate work, the slot time and the feature representation of the user company in the feature representation of the user work, and the slot post prediction sub-network predicts the probability of the target user to jump to each candidate post according to the feature representation of the candidate post in the feature representation of the candidate work, the slot time and the feature representation of the user post in the feature representation of the user work.

For example, when the working experience of the resume of the target user records two works, the step can enable the target resume static attribute feature representation F in the resume feature representation of the target user to be represented _is Mapped initial transfer parametersAs a pass-through input to LSTM-1 in FIG. 5B, the first work of the target user resume feature representation is represented by dynamic attribute feature representation F _i1 Working time D of first work _i1 Corresponding target time encoded representation d _i1 And the target resume static attribute feature represents F _is As a parameter input of LSTM-1, the transfer parameter of LSTM-1 output is likewise +.>As a delivery input to LSTM-2, target profile for the second job in the target user profileCalendar dynamic attribute feature representation F _i2 Working time D of second work _i2 Corresponding target time encoded representation d _i2 And the target resume static attribute feature represents F _is As a parameter input to LSTM-2. At this time, LSTM-1 and LSTM-2 will process according to the algorithm when training LSTM of the work prediction model according to the above embodiment based on the input parameters, and output the user work experience representation corresponding to the first work>User work experience representation corresponding to the second work +. >The embodiment of the application can be that the user work experience corresponding to the second work is represented as +.>Inputting to MLP-2 corresponding to LSTM-2, wherein LSTM-1 may not represent user work experience corresponding to the first work ∈>The transmission parameters to MLP-1 are only output +.>Inputs to LSTM-2.MLP-2 will process according to the algorithm when training MLP in the work prediction model to obtain user time feature expression x _i2 User company feature representation u _i2 And user post feature representation w _i2 Is input to the predictive neural network 52, and the jump time predictive sub-network 521 in the predictive neural network 52 is then based on x _i2 The next jump time Di3 is predicted. The jump-slot company predictive sub-network 522 will predict the characteristic representation v of the candidate company from the characteristic representations of the candidate job _j2 、D _i3 And u _i2 The probability of the target user i jumping from the second work (i.e., the current work) to each candidate company j is predicted. The jump post prediction sub-network 523 will rootFeature representation q of candidate positions in feature representation of candidate work _k 、D _i3 And w _i2 The probability of predicting the target user i to jump from the second work (i.e., the current work) to each candidate post k is predicted.

It should be noted that, in the embodiment of the present application, when predicting the probability of the user's jump to the candidate work, the jump time may be predicted by the jump time prediction sub-network 521 in the prediction neural network 52, or may be input according to the actual requirement of the target user, which is not limited in this embodiment.

S503, recommending work for the target user according to the probability of the user jump candidate work.

Optionally, when recommending the work for the target user according to the probability of the user's jump candidate work, the step may be to select one or more of the candidate works with higher probability as the work which is finally recommended for the target user.

According to the technical scheme provided by the embodiment of the application, the working prediction model which is constructed and trained by adopting any one of the embodiments is adopted. Determining a target user resume feature representation according to the target user resume information, inputting the target user resume feature representation into a work prediction model, and obtaining the probability of target user jump to each candidate work, and determining the recommended work for the target user according to the probability of jump to the candidate work. Because the work prediction model of the embodiment of the application is trained based on resume information of a large number of users with different samples, when the work prediction model carries out the follow-up prediction on a jump work line aiming at a certain user, the work prediction model can cooperate with the selection of other users similar to the target user work characteristic representation in the jump process to determine the probability of the target user jump to each candidate work. The accuracy of the prediction result is greatly improved while the work recommendation based on the personalized requirements of the user can be realized.

Fig. 6 is a schematic structural diagram of a device for establishing a working prediction model according to an embodiment of the present application. The method and the device are suitable for constructing and training the neural network model capable of executing the work prediction task, and can be used for realizing the method for constructing the work prediction model in any embodiment of the application.

The apparatus 600 specifically includes the following:

the sample preprocessing module 601 is configured to determine a sample user resume feature representation and a target job of a sample user jump according to sample user resume information;

the model training module 602 is configured to input a sample user resume feature representation to a time-sequential neural network in the original model to obtain a user work experience representation, input the user work experience representation to a collaborative neural network in the original model to obtain a user work feature representation, and input a candidate work feature representation and a user work feature representation to a prediction neural network in the original model to predict a probability of a sample user jumping to a candidate work; and training the original model according to the target work and the prediction result of the jump groove of the sample user to obtain a work prediction model.

Further, the sub-time sequence network in the time sequence neural network corresponds to the sub-sensing network in the cooperative neural network one by one;

the first output end of the sub-timing network is connected with the next sub-timing network;

the second output end of the sub-time sequence network is connected with the corresponding sub-sensing network.

Further, the sub-sensing network comprises at least one of a first sub-sensing unit, a second sub-sensing unit and a third sub-sensing unit; the input ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the second output ends of the corresponding sub-timing network, and the output ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the input ends of the prediction neural network;

the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are respectively used for outputting user time feature representation, user company feature representation and user post feature representation in the user work feature representation.

Further, different sub-aware networks in the collaborative neural network share model parameters.

Further, the sub-time network is a long-short-period memory network LSTM; the sub-sensing network is a multi-layer sensing machine MLP.

Further, the sample preprocessing module 601 includes:

the information determining unit is used for determining sample user attribute information, work attribute information of at least one work of the sample user and the work duration of the work according to the sample user resume information;

the static characteristic determining unit is used for determining a sample resume static attribute characteristic representation according to the sample user attribute information;

the dynamic characteristic determining unit is used for determining a sample resume dynamic attribute characteristic representation according to the work attribute information of at least one work of the sample user;

the time coding unit is used for coding the working time length of at least one work of the sample user and determining sample time coding representation;

and the characteristic integration unit is used for taking the sample resume static attribute characteristic representation, the sample resume dynamic attribute characteristic representation and the sample time coding representation as sample user resume characteristic representations.

Further, the model training module 602 includes a first data input unit for inputting a sample user resume feature representation into the time-sequential neural network in the original model; the first data input unit is specifically configured to:

obtaining initial transfer parameters according to the sample resume static attribute characteristic representation;

Taking the initial transfer parameter as the transfer input of the first sub-time sequence network of the time sequence neural network in the original model;

and sequentially taking the dynamic attribute characteristic representation and the sample time coding representation of each working sample resume and the static attribute characteristic representation of the sample resume as the parameter input of the sub-time network in the time sequence neural network.

Further, the model training module 602 comprises a second data input unit for inputting the feature representation of the candidate work and the user work feature representation into a predictive neural network in the original model, predicting the probability of the sample user jumping to the candidate work; the second data input unit is specifically configured to: the feature representation of the candidate work, the jump time and the user work feature representation are input into a prediction neural network in the original model, and the probability of the sample user jump to the candidate work is predicted.

Further, the prediction neural network comprises a jump groove company prediction sub-network and/or a jump groove post prediction sub-network;

correspondingly, the second data input unit is specifically configured to:

inputting the feature representation of the candidate company in the feature representation of the candidate work, the jump time and the user company feature representation in the user work feature representation into a jump company prediction sub-network to predict the probability of the sample user jump to the candidate company; and/or the number of the groups of groups,

And inputting the feature representation of the candidate post in the feature representation of the candidate work, the jump time and the user post feature representation in the user work feature representation into a jump post prediction sub-network to predict the probability of the sample user jump to the candidate post.

Further, the prediction neural network further includes: a slot time prediction sub-network; accordingly, model training module 602 further includes:

and the third data input unit is used for inputting the user time characteristic representation in the user work characteristic representation into the jump time prediction sub-network to predict and obtain the jump time.

Fig. 7 is a schematic structural diagram of a work recommendation device according to an embodiment of the present application. The embodiment is suitable for the situation that the work prediction model established in any embodiment is deployed into a work recommendation system to recommend work for a target user. The device can realize the work recommendation method of any embodiment of the application. The apparatus 700 specifically includes the following:

the user resume processing module 701 is configured to determine a target user resume feature representation according to the target user resume information;

the work prediction module 702 is configured to input the target user resume feature representation into a work prediction model, so as to obtain a probability of the target user jumping to a candidate work;

The job recommendation module 703 is configured to recommend a job to the target user according to the probability of the user's jump candidate job.

Further, the user resume processing module 701 is specifically configured to:

Determining target user attribute information, work attribute information of at least one work of a target user and the work time length of the work according to the target user resume information;

determining a target resume static attribute characteristic representation according to the target user attribute information;

determining target resume dynamic attribute characteristic representation according to the work attribute information of at least one work of the target user;

encoding the working time length of at least one work of the target user, and determining a target time encoding representation;

and taking the target resume static attribute characteristic representation, the target resume dynamic attribute characteristic representation and the target time code representation as target user resume characteristic representations.

Further, the work prediction module 702 is specifically configured to:

obtaining initial transfer parameters according to the static attribute characteristic representation of the target resume;

taking the initial transfer parameter as the transfer input of the first sub-time sequence network of the time sequence neural network in the work prediction model;

and sequentially taking the dynamic attribute characteristic representation and the target time code representation of the target resume and the static attribute characteristic representation of the target resume of each work as the parameter input of the sub-time network in the time sequence neural network.

According to an embodiment of the present application, the present application also provides an electronic device and a readable storage medium.

As shown in fig. 8, there is a block diagram of an electronic device of a work prediction model establishment method or a work recommendation method according to an embodiment of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the applications described and/or claimed herein.

As shown in fig. 8, the electronic device includes: one or more processors 801, memory 802, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions executing within the electronic device, including instructions stored in or on memory to display graphical information of the GUI on an external input/output device, such as a display device coupled to the interface. In other embodiments, multiple processors and/or multiple buses may be used, if desired, along with multiple memories and multiple memories. Also, multiple electronic devices may be connected, each providing a portion of the necessary operations (e.g., as a server array, a set of blade servers, or a multiprocessor system). One processor 801 is illustrated in fig. 8.

Memory 802 is a non-transitory computer readable storage medium provided by the present application. The memory stores instructions executable by the at least one processor to cause the at least one processor to execute the method for establishing the work prediction model or the method for recommending the work provided by the application. The non-transitory computer-readable storage medium of the present application stores computer instructions for causing a computer to execute the method of establishing the work prediction model or the work recommendation method provided by the present application.

The memory 802 is used as a non-transitory computer readable storage medium, and may be used to store a non-transitory software program, a non-transitory computer executable program, and a module, such as program instructions/modules corresponding to a method for creating a work prediction model or a method for recommending work in an embodiment of the present application (e.g., the sample preprocessing module 601 and the model training module 602 shown in fig. 6, or the user profile processing module 701, the work prediction module 702, and the work recommendation module 703 shown in fig. 7). The processor 801 executes various functional applications of the server and data processing, that is, implements the method of establishing a work prediction model or the method of recommending works in the above-described method embodiment, by executing non-transitory software programs, instructions, and modules stored in the memory 802.

Memory 802 may include a storage program area that may store an operating system, at least one application program required for functionality, and a storage data area; the storage data area may store data created according to the use of the electronic device of the work prediction model establishment method or the work recommendation method, and the like. In addition, memory 802 may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 802 may optionally include memory remotely located with respect to processor 801, which may be connected via a network to the electronic device of the work prediction model creation method or work recommendation method. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The electronic device of the work prediction model establishing method or the work recommending method may further include: an input device 803 and an output device 804. The processor 801, memory 802, input devices 803, and output devices 804 may be connected by a bus or other means, for example in fig. 8.

The input device 803 may receive input numeric or character information and generate key signal inputs related to user settings and function control of the electronic device of the work prediction model creation method or the work recommendation method, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, a pointer stick, one or more mouse buttons, a track ball, a joystick, etc. The output device 804 may include a display apparatus, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD), a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device may be a touch screen.

Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASIC (application specific integrated circuit), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs, the one or more computer programs may be executed and/or interpreted on a programmable system including at least one programmable processor, which may be a special purpose or general-purpose programmable processor, that may receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device.

These computing programs (also referred to as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and pointing device (e.g., a mouse or trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic input, speech input, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a background component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such background, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the internet.

The computer system may include a client and a server. The client and server are typically remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

According to the technical scheme of the embodiment of the application, an original model comprising a time sequence neural network, a collaborative perception network and a prediction neural network is constructed, a user resume feature representation determined based on sample user resume information is input to the time sequence neural network of the original model, a user work experience representation is obtained and input to the collaborative neural network, a user work feature representation is obtained and input to the prediction neural network, the probability of a sample user jumping to a candidate work is predicted, and then the original model is trained by combining with a target work of a sample user actual jumping to obtain a work prediction model. Because the work prediction model of the embodiment of the application is trained based on resume information of a large number of users with different samples, when the work prediction model carries out the follow-up prediction on a jump work line aiming at a certain user, the work prediction model can cooperate with the selection of other users similar to the target user work characteristic representation in the jump process to determine the probability of the target user jump to each candidate work. The accuracy of the prediction result is greatly improved while the work recommendation based on the personalized requirements of the user can be realized.

It should be appreciated that various forms of the flows shown above may be used to reorder, add, or delete steps. For example, the steps described in the present application may be performed in parallel, sequentially, or in a different order, provided that the desired results of the disclosed embodiments are achieved, and are not limited herein.

The above embodiments do not limit the scope of the present application. It will be apparent to those skilled in the art that various modifications, combinations, sub-combinations and alternatives are possible, depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the scope of the present application.

Claims

1. A work recommendation method implemented using a work prediction model, the method comprising:

inputting the target user resume feature representation into the work prediction model to obtain the probability of target user jump to candidate works;

recommending the work for the target user according to the probability of the candidate work of the user jump;

Wherein, the work prediction model is established by the following modes:

inputting the sample user resume feature representation into a time sequence neural network in an original model to obtain a user work experience representation, inputting the user work experience representation into a collaborative neural network in the original model to obtain a user work feature representation, inputting a candidate work feature representation and the user work feature representation into a prediction neural network in the original model, and predicting the probability of the sample user jumping to the candidate work;

training the original model according to the target work and the prediction result of the sample user jump, so as to obtain a work prediction model;

wherein determining a sample user resume feature representation according to the sample user resume information comprises:

according to the sample user resume information, determining sample user attribute information, working attribute information of at least one work of a sample user and working time of the work;

determining a sample resume static attribute characteristic representation according to the sample user attribute information;

Determining a sample resume dynamic attribute characteristic representation according to the work attribute information of at least one work of the sample user;

encoding the working time length of at least one work of the sample user, and determining a sample time encoding representation;

and taking the sample resume static attribute characteristic representation, the sample resume dynamic attribute characteristic representation and the sample time coding representation as the sample user resume characteristic representation.

2. The method of claim 1, wherein determining a target user resume feature representation from target user resume information comprises:

determining a target resume dynamic attribute characteristic representation according to the work attribute information of at least one work of the target user;

and taking the target resume static attribute characteristic representation, the target resume dynamic attribute characteristic representation and the target time code representation as the target user resume characteristic representation.

3. The method of claim 2, wherein inputting the user resume feature representation into the work prediction model comprises:

obtaining initial transfer parameters according to the target resume static attribute characteristic representation;

taking the initial transfer parameter as a transfer input of a first sub-time sequence network of a time sequence neural network in the work prediction model;

and sequentially inputting the dynamic attribute characteristic representation and the target time coding representation of each working object resume and the static attribute characteristic representation of the target resume as parameters of a sub-time network in the time sequence neural network.

4. The method of claim 1, wherein sub-timing networks in the timing neural network are in one-to-one correspondence with sub-sensing networks in the collaborative neural network;

and a second output end of the sub-timing network is connected with a corresponding sub-sensing network.

5. The method of claim 4, wherein the sub-sensing network comprises at least one of a first sub-sensing unit, a second sub-sensing unit, and a third sub-sensing unit; the input ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the second output ends of the corresponding sub-time network, and the output ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the input ends of the prediction neural network;

The first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are respectively used for outputting user time characteristic representation, user company characteristic representation and user post characteristic representation in the user work characteristic representation.

6. The method of claim 4, wherein different sub-aware networks in the collaborative neural network share model parameters.

7. The method of claim 4, wherein the sub-timing network is a long and short memory network LSTM; the sub-sensing network is a multi-layer sensing machine MLP.

8. The method of claim 1, wherein inputting the sample user resume feature representation into a temporal neural network in an original model comprises:

taking the initial transfer parameter as a transfer input of a first sub-timing network of a timing neural network in the original model;

and sequentially inputting the dynamic attribute characteristic representation and the sample time coding representation of each working sample resume and the static attribute characteristic representation of the sample resume as parameters of a sub-time network in the time sequence neural network.

9. The method of claim 1, wherein inputting a feature representation of a candidate work and the user work feature representation into a predictive neural network in the original model predicts a probability of the sample user jumping to the candidate work, comprising:

inputting a feature representation of a candidate job, a slot-jump time, and the user job feature representation into a predictive neural network in the original model, predicting a probability of the sample user jumping to the candidate job.

10. The method of claim 9, wherein the predictive neural network comprises a slot company predictive sub-network and/or a slot post predictive sub-network;

correspondingly, inputting the feature representation of the candidate work, the jump time and the user work feature representation into a prediction neural network in the original model, predicting the probability of the sample user jump to the candidate work, comprising:

inputting a feature representation of a candidate company in the feature representation of the candidate work, a jump time and a user company feature representation in the user work feature representation into the jump company prediction sub-network to predict the probability of the sample user jump to the candidate company; and/or the number of the groups of groups,

And inputting the feature representation of the candidate post in the feature representation of the candidate work, the jump time and the user post feature representation in the user work feature representation into the jump post prediction sub-network to predict the probability of the sample user jump to the candidate post.

11. The method of any of claims 9-10, wherein the predictive neural network further comprises: a slot time prediction sub-network;

correspondingly, inputting the feature representation of the candidate work, the jump time and the user work feature representation into the prediction neural network in the original model, and before predicting the probability of the sample user jump to the candidate work, further comprising:

and inputting the user time characteristic representation in the user work characteristic representation into the jump time prediction sub-network to predict and obtain the jump time.

12. A work recommendation device implemented using a work prediction model, the device comprising:

the work prediction module is used for inputting the target user resume feature representation into the work prediction model to obtain the probability of target user jump to candidate works;

The work recommending module is used for recommending works for the target user according to the probability of the candidate works of the user jump groove;

wherein, the work prediction model is established by the following modes:

13. The apparatus of claim 12, wherein sub-timing networks in the timing neural network are in one-to-one correspondence with sub-sensing networks in the collaborative neural network;

14. The apparatus of claim 13, wherein the sub-sensing network comprises at least one of a first sub-sensing unit, a second sub-sensing unit, and a third sub-sensing unit; the input ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the second output ends of the corresponding sub-time network, and the output ends of the first sub-sensing unit, the second sub-sensing unit and the third sub-sensing unit are connected with the input ends of the prediction neural network;

15. The apparatus of claim 13, wherein different sub-aware networks in the collaborative neural network share model parameters.

16. The apparatus of claim 13, wherein the sub-timing network is a long-short-term memory network LSTM; the sub-sensing network is a multi-layer sensing machine MLP.

17. The apparatus of claim 12, wherein inputting the sample user resume feature representation into a temporal neural network in an original model comprises:

18. The apparatus of claim 12, wherein inputting a feature representation of a candidate work and the user work feature representation into a predictive neural network in the original model predicts a probability of the sample user jumping to the candidate work, comprising:

19. The apparatus of claim 18, wherein the predictive neural network comprises a slot company predictive sub-network and/or a slot post predictive sub-network;

20. The apparatus of any of claims 18-19, wherein the predictive neural network further comprises: a slot time prediction sub-network;

21. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the work recommendation method of any one of claims 1-11.

22. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the work recommendation method of any one of claims 1-11.