Method for generating job expectation transfer model and method for constructing training sample

Abstract

The invention discloses a construction method of a training sample and a generation method and device of a job hunting expectation transfer model. The construction method comprises the steps of obtaining first historical delivery position data of a job hunter in a first time period; the first historical delivery position data comprises position information of a plurality of delivered positions; sequentially sorting the first historical delivery position data according to a sequence from back to front from the current time to obtain first delivery sequence data; segmenting the first delivery sequence data into multiple groups of positive sample sets according to a preset segmentation strategy; each group of positive sample set comprises a target position and n delivered positions, and the target position is the position with the delivery time closest to the current time in each group of positive sample set; randomly selecting a plurality of positions from the delivery sequence data to obtain pre-selected positions; replacing a target position in each group of positive sample set with a pre-selected position to obtain a plurality of groups of corresponding negative sample sets; and obtaining a training sample set according to the positive sample set and the negative sample set.

Description

Technical Field

[0001] The present application relates to the technical field of computers and the Internet, and in particular to a method for constructing a training sample, a method for generating a job search expectation transfer model, a device, an electronic device, and a storage medium. Background Art

[0002] In the field of recruitment, candidates' expectations will change to varying degrees during the entire job search cycle, such as writing resumes, searching and screening positions, submitting resumes, and interviews. For example, candidates initially expect and submit a relatively high salary and a position they like. As the submission feedback and interview results are not ideal, the initial expectations will gradually decrease, and the salary can be reduced, the working city can be changed, the position can be changed, etc. In order to improve the accuracy of recommended positions, it is necessary to adjust the recommendation model in real time. At present, due to the limited log data that can be recorded by the recruitment system, most of the existing practices are based on the candidate's behavior, such as searching for keywords, browsing exposed positions, clicking on positions, collecting positions, submitting positions, etc. to construct positive and negative samples. For example, for a series of behaviors after the candidate enters the search keyword, the position that the candidate finally submits is the positive sample of the search, and the exposed but not submitted and / or not clicked positions are negative samples. These positive and negative samples reflect the candidate's selection tendency to a certain extent and can be used as training samples for the recommendation model. However, the sample diversity determined in the above manner is poor, and is quite different from the data distribution actually processed by the model. It does not accurately capture and understand the information about the expected changes and shifts of candidates, and therefore the training effect of the recommendation model is difficult to be satisfactory. Summary of the invention

[0003] The purpose of this application is to provide a method for constructing a training sample, a method for generating a job search expectation transfer model, a device, an electronic device and a storage medium.

[0004] The present application provides a method for constructing a training sample, including:

[0005] Acquire the first historical job posting data of the job seeker in the first time period; wherein the first historical job posting data includes job information of multiple jobs that have been posted;

[0006] The first historical posting position data is sorted in order from the last time to the last time to obtain first posting sequence data; wherein the first posting sequence data includes position information of multiple posted positions sorted in order from the last time to the last time;

[0007] The first delivery sequence data is divided into multiple groups of positive sample sets according to a preset segmentation strategy; each group of positive sample sets includes a target position and n delivered positions, and the target position is the position in each group of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2;

[0008] Randomly selecting a plurality of positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions;

[0009] The pre-selected positions are used to replace the target positions in each set of positive samples to obtain corresponding sets of negative samples.

[0010] According to the positive sample set and the negative sample set, a training sample set is obtained.

[0011] Furthermore, the step of segmenting the first delivery sequence data into a plurality of positive sample sets according to a preset segmentation strategy includes:

[0012] Using a set sliding window to sequentially intercept the first delivery sequence data to obtain multiple groups of sequence samples;

[0013] The positions corresponding to the sequences in each group of sequence samples are marked as positive samples, and the positions corresponding to the first sequence number in each group of sequence samples are taken as target positions.

[0014] Furthermore, the step of segmenting the first delivery sequence data into a plurality of positive sample sets according to a preset segmentation strategy includes:

[0015] If the number of samples in the current sequence of samples does not meet the number of samples set in the sliding window, "0" is filled after the sequence of the current sequence of samples to make the number of samples in the current sequence of samples meet the number of samples set in the sliding window.

[0016] Furthermore, the position information includes one or more of the following influencing factors: position ID, function, industry, working city, salary range, and the time difference between the nth position and the target position.

[0017] The embodiment of the present application provides a method for generating a job-seeking expectation transfer model, and a training sample set constructed based on the method includes:

[0018] Constructing a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and the vector data of multiple conversion sequences output by the transformer structure are concatenated as the input of the multi-layer perceptron;

[0019] The constructed model to be trained is trained using the training sample set to obtain a trained job search expectation transfer model; wherein,

[0020] The job-seeking expectation transfer model is used to predict the current job seeker's delivery sequence data and the relevant scores of multiple positions to be predicted.

[0021] Furthermore, the step of using the training sample set to train the constructed model to be trained to obtain a trained job search expectation transfer model further includes:

[0022] Extract various influencing factors and position information of job information in each sample set;

[0023] Map each influencing factor in the job information into a first-class eigenvector of the corresponding dimension;

[0024] Mapping the position information into a second type of feature vector of corresponding dimension;

[0025] Concatenate each first-category eigenvector and the second-category eigenvector to obtain concatenated first-category vector data;

[0026] Input the concatenated first-category vector data into a transformer structure, and train the transformer structure to output vector data of multiple transformation sequences;

[0027] splicing the vector data of the plurality of conversion sequences to obtain the spliced ​​second-category vector data;

[0028] The concatenated second-category vector data is input into a multilayer perceptron, and the multilayer perceptron is trained to output a first correlation score between the target position and the multiple positions that have been applied for.

[0029] Furthermore, mapping the position information into a second type of feature vector of a corresponding dimension includes:

[0030] At the t-th position, the i-th vector PE can be calculated by the following formula:

[0031]

[0032] Where: d is the vector dimension; the value of k is (vector dimension d / 2-1);

[0033] When i=2k, the vector is at an even position; when i=2k+1, the vector is at an odd position.

[0034] Furthermore, the step of using the training sample set to train the constructed model to be trained to obtain a trained job search expectation transfer model further includes:

[0035] Input the relevant scores into the cross entropy loss function and calculate the loss value;

[0036] According to the calculated loss value, the optimization algorithm is used to calculate the update gradient, and the parameters of the job search expectation transfer model and the underlying first-class eigenvector and second-class eigenvector data are updated through reverse conduction; wherein the loss function is shown in the following conditional formula:

[0037]

[0038] In the formula, f(x i ) is the first correlation score; when the training sample is a positive sample, y i is 1; when the training sample is a negative sample, y i is 0; m is the total number of sample groups in the sample set.

[0039] The present application embodiment provides a training sample construction device, including:

[0040] A first acquisition module is used to acquire first historical job posting data of a job seeker within a first time period; wherein the first historical job posting data includes job information of multiple jobs that have been applied for;

[0041] A first sequence module, which is used to sort the first historical posting position data in order from the current time to the previous time, so as to obtain first posting sequence data; wherein the first posting sequence data includes position information of multiple posted positions sorted in order from the current time to the previous time;

[0042] A segmentation module, which is used to segment the first delivery sequence data into multiple sets of positive sample sets according to a preset segmentation strategy; wherein each set of positive sample sets includes a target position and n delivered positions, and the target position is the position in each set of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2;

[0043] A selection module, which is used to randomly select multiple positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions;

[0044] A replacement module, which is used to replace the target position in each set of positive samples with the pre-selected position to obtain a corresponding set of negative samples;

[0045] The training sample module is used to obtain a training sample set based on a positive sample set and a negative sample set.

[0046] The embodiment of the present application provides a device for generating a job-seeking expectation transfer model, and the training sample set constructed based on the method described includes:

[0047] A construction module is used to construct a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and vector data of multiple conversion sequences output by the transformer structure are concatenated as input of the multi-layer perceptron;

[0048] A training module is used to train the constructed model to be trained using the training sample set to obtain a trained job search expectation transfer model; wherein,

[0049] The job-seeking expectation transfer model is used to predict the current job seeker's delivery sequence data and the relevant scores of multiple positions to be predicted.

[0050] Furthermore, the training module includes:

[0051] An extraction unit, which is used to extract various influencing factors and position information of the position information in each set of sample sets;

[0052] A first mapping unit, which is used to map each influencing factor in the position information into a first type of feature vector of a corresponding dimension;

[0053] A second mapping unit, configured to map the position information into a second type of feature vector of a corresponding dimension;

[0054] A first concatenation unit, which is used to concatenate each first-category feature vector and the second-category feature vector to obtain concatenated first-category vector data;

[0055] A first training unit, which is used to input the concatenated first-category vector data into a transformer structure, train the transformer structure, and output vector data of a plurality of transformation sequences;

[0056] A second splicing unit, which is used to splice the vector data of multiple conversion sequences to obtain spliced ​​second-category vector data;

[0057] The second training unit is used to input the concatenated second-category vector data into a multilayer perceptron, and train the multilayer perceptron to output a first correlation score between the target position and the multiple positions that have been applied for.

[0058] An embodiment of the present application provides an electronic device, which includes a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the steps of the method described above are implemented.

[0059] An embodiment of the present application provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the steps of the method described above are implemented.

[0060] An embodiment of the present application provides a computer program product, wherein the computer program product includes computer program instructions, and when the computer program instructions are executed by a processor, the steps of the method described above are implemented.

[0061] The above technical solution of the present application has the following beneficial technical effects:

[0062] In the embodiments of the present application, in the recruitment field, there is a significant connection between the expectation transfer of job seekers and the time interval. The expectation transfer of job seekers can be reflected in the change of any element in the job information such as job ID, function, industry, work city, salary range and welfare benefits. By providing positive samples and negative samples to train the model to be trained, the model to be trained can learn the feature changes in the training samples, and then obtain a trained job search expectation transfer model. The job search expectation transfer model can predict the expectation changes of the current job seeker, which is conducive to improving the accuracy of job information recommendation and improving the matching degree between job seekers and jobs to be recruited; therefore, time interval information is introduced into the training sample sequence information so that the model can better learn and characterize the influence of time factors on expectation transfer, and the introduction of position information can characterize the order of the job seekers submitting job sequences, so that high-quality training samples can be obtained, thereby achieving the purpose of optimizing the model training effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following briefly introduces the drawings in the embodiments of the present application.

[0064] Figure 1 It is a schematic diagram of the system architecture of an embodiment of the present application.

[0065] Figure 2 It is a flowchart of a method for constructing a training sample in an embodiment of the present application.

[0066] Figure 3 It is a flowchart of a method for generating a job-seeking expectation transfer model according to an embodiment of the present application.

[0067] Figure 4 It is a flowchart of a method for recommending job information according to an embodiment of the present application.

[0068] Figure 5 It is a schematic diagram of the processing process of the sample sequence of an embodiment of the present application.

[0069] Figure 6 It is a schematic diagram of the training process of the job-seeking expectation transfer model according to an embodiment of the present application.

[0070] Figure 7 It is a structural block diagram of a training sample construction device according to an embodiment of the present application.

[0071] Figure 8 It is a structural block diagram of a device for generating a job-seeking expectation transfer model according to an embodiment of the present application.

[0072] Fig. 9 It is a structural block diagram of a device for recommending job information according to an embodiment of the present application.

[0073] Fig.10 It is a schematic diagram of an electronic device used to implement the method for constructing training samples, the method for generating a job-seeking expectation transfer model, and the method for recommending job information in the embodiments of the present application. DETAILED DESCRIPTION

[0074] The principles and spirit of the present application will be described below with reference to several exemplary embodiments. It should be understood that the purpose of providing these embodiments is to make the principles and spirit of the present application clearer and more thorough, so that those skilled in the art can better understand and implement the principles and spirit of the present application. The exemplary embodiments provided herein are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments herein, all other embodiments obtained by ordinary technicians of the art without creative work are within the scope of protection of this application.

[0075] Embodiments of the present application relate to terminal devices and / or servers. Those skilled in the art will appreciate that the embodiments of the present application may be implemented as a system, apparatus, device, method, computer-readable storage medium, or computer program product. Therefore, the present disclosure may be specifically implemented in at least one of the following forms: complete hardware, complete software, or a combination of hardware and software. According to the embodiments of the present application, the present application seeks protection for a method for constructing a training sample and a device thereof, a method for generating a job search expectation transfer model and a device thereof, a method for recommending job information and a device thereof, an electronic device, a computer-readable storage medium, and a computer program product. Figure 1 A schematic diagram of a system architecture of an embodiment of the present application is shown. Figure 1As shown, the system includes a terminal device 102 and a server 104. Among them, the terminal device 102 may include at least one of the following: a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart TV, various wearable devices, an augmented reality AR device, a virtual reality VR device, etc. A client can be installed on the terminal device 102. For example, the client can be a client that specifically performs a specific function (such as an application app), or a client with multiple application applets (different functions) embedded in it, or a client logged in through a browser. The user can operate on the terminal device 102. For example, the user can open the client installed on the terminal device 102 and input instructions through the client operation, or the user can open the browser installed on the terminal device 102 and input instructions through the browser operation. After the terminal device 102 receives the instruction input by the user, the request information containing the instruction is sent to the server 104. After receiving the request information, the server 104 performs corresponding processing and then returns the processing result information to the terminal device 102. The user instruction is completed through a series of data processing and information interaction.

[0076] In this document, terms such as first, second, third, etc. are only used to distinguish one entity (or operation) from another entity (or operation), but not to require or imply any order or relationship between these entities (or operations).

[0077] The following is a brief description of the concepts and technical terms that may be involved in the embodiments of the present application.

[0078] Function: The positions applied for by job seekers can be classified into first-level, second-level and third-level, such as R&D / algorithm engineer / recommended algorithm engineer, among which recommended algorithm engineer is a third-level classification, and the third-level classification can be defined as "function"; R&D is a first-level classification, and the first-level classification can be defined as "industry".

[0079] Figure 2 A flow chart of a method for constructing a training sample according to an embodiment of the present application is shown, and the method comprises the following steps:

[0080] S110: Acquire the first historical job application data of the job seeker within a first time period; wherein the first historical job application data includes job information of multiple applied positions.

[0081] Specifically, the first time period can be set to a predetermined cycle, such as a job search life cycle, that is, the entire process from the job seeker starting to look for a job to the job interview, which has a time span; when the job seeker browses the job description of the job to be recruited posted by the recruiter on the terminal device 102, the job description may include job information such as job ID, function, industry, work city, salary range and benefits, and the job seeker can click the submit resume button for the position that meets his / her job search intention to send his / her resume to the recruiter; therefore, by obtaining the resume sent by each job seeker on the recruitment platform after browsing different job descriptions on the terminal device 102, the job information of multiple positions that the job seeker has submitted can be obtained; for example, setting the first time period to two months can obtain the corresponding historical job submission data of different job seekers.

[0082] S120: Sort the first historical delivered position data in order from the back to the current time to obtain first delivery sequence data; wherein the first delivery sequence data includes location information of multiple delivered positions sorted in order from the back to the current time in terms of delivery time.

[0083] Specifically, the job information of each job seeker who has applied for multiple positions can be sorted in order from the time of application to the current time to obtain the first application sequence data; for example, if a job seeker has applied for 100 positions in two months, the position applied for last can be identified as number "1", and the position applied for earliest can be identified as number "100" in order from the time of application to the current time, and the same applies to other positions; wherein the position information corresponds to the position identification serial numbers of the multiple positions applied for.

[0084] S130: Split the first delivery sequence data into multiple groups of positive sample sets according to a preset segmentation strategy; wherein each group of positive sample sets includes a target position and n delivered positions, and the target position is the position in each group of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2.

[0085] Specifically, each group of positive sample sets can be set as a part of the first delivery sequence data, that is, the first delivery sequence data is a long sequence set, each group of positive sample sets is a short sequence set in the long sequence set, and each group of positive sample sets may include a target position and n delivered positions. The target position is the position in each group of positive sample sets whose delivery time is closest to the current time, and the n delivered positions are other positions whose delivery time of the job seeker is slightly later than the delivery time of the target position and are sorted in order from the back to the front from the current time.

[0086] S140: Randomly select multiple positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions.

[0087] S150: Replace the target position in each set of positive samples with the pre-selected position to obtain multiple sets of corresponding negative samples.

[0088] Specifically, the target position in each positive sample set is replaced by a randomly selected position, and the sequence of the other n submitted positions remains unchanged to obtain the corresponding negative sample set; the number of negative sample sets and positive sample sets can be generated in a 1:10 manner. By providing positive and negative samples to train the model to be trained, the model to be trained is, for example, a neural network model, which can learn to distinguish which features are associated with the category of interest and which are not, which helps to improve the model's ability to distinguish and improve the robustness of the model.

[0089] S160: Obtain a training sample set according to the positive sample set and the negative sample set. Specifically, the positive sample set and the negative sample set may be shuffled to generate a random sequence to obtain the training sample set.

[0090] In the embodiments of the present application, in the recruitment field, there is a significant connection between the expectation transfer of job seekers and the time interval. The expectation transfer of job seekers can be reflected in the change of any element in the job information such as job ID, function, industry, work city, salary range and welfare benefits. By providing positive samples and negative samples to train the model to be trained, the model to be trained can learn the feature changes in the training samples, and then obtain a trained job search expectation transfer model. The job search expectation transfer model can predict the expectation changes of the current job seeker, which is conducive to improving the accuracy of job information recommendation and improving the matching degree between job seekers and jobs to be recruited; therefore, time interval information is introduced into the training sample sequence information so that the model can better learn and characterize the influence of time factors on expectation transfer, and the introduction of position information can characterize the order of the job seekers submitting job sequences, so that high-quality training samples can be obtained, thereby achieving the purpose of optimizing the model training effect.

[0091] In some embodiments, step S130, dividing the first delivery sequence data into multiple sets of positive sample sets according to a preset division strategy, may include the following specific steps:

[0092] S131: intercepting the first delivery sequence data in sequence using a set sliding window to obtain multiple groups of sequence samples;

[0093] S132: Mark the position corresponding to the sequence in each group of sequence samples as a positive sample, and take the position corresponding to the first sequence number in each group of sequence samples as the target position.

[0094] Specifically, the sliding window may include multiple delivery sequence numbers, that is, a certain number of samples can be set in the sliding window. Since the job information of each job seeker who has delivered multiple positions is sorted in order from the back to the front from the current time, the set sliding window is used to intercept the first delivery sequence data in sequence to obtain multiple groups of sequence samples that can be marked as positive samples. The target position is the last delivered position in each group of sequence samples. In this way, time interval information is introduced into the training samples so that the model can better learn and characterize the impact of time factors on expected transfer.

[0095] In some embodiments, step S130, dividing the first delivery sequence data into multiple sets of positive sample sets according to a preset segmentation strategy, may further include the following specific steps:

[0096] S133: If the number of samples in the current sequence of samples does not meet the number of samples set in the sliding window, fill "0" after the sequence of the current sequence of samples to make the number of samples in the current sequence of samples meet the number of samples set in the sliding window.

[0097] In an exemplary embodiment, the job information includes one or more of the following influencing factors: job ID, function, industry, work city, salary range, and the time difference between the nth job and the target job. In the job search process, the job ID, function, industry, work city, salary range, etc. of the job seeker may change over time. Introducing these influencing factors into the training samples can characterize the trajectory changes of the job seeker's expected migration.

[0098] The present application embodiment provides a method for generating a job-seeking expectation transfer model, based on a training sample set constructed by the method, such as Figure 3 As shown, the following specific steps may be included:

[0099] S210: constructing a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and vector data of multiple conversion sequences output by the transformer structure are concatenated and used as input of the multi-layer perceptron;

[0100] S220: Using the training sample set to train the constructed model to be trained to obtain a trained job-seeking expectation transfer model; wherein the job-seeking expectation transfer model is used to predict the current job seeker's submission sequence data and the relevant scores of multiple positions to be predicted.

[0101] Specifically, in the field of deep learning, the Transformer structure is a model architecture for processing sequence data. The Transformer structure may include multi-head self-attention, positional encoding, feed forward neural network, residual connections and layer normalization, etc.; among them, the use of multi-head self-attention can characterize and learn the relationship between position sequences, and can better capture the multiple relationships and subtle differences between the elements of each position sequence; by adding position encoding to provide position information in the sequence, and performing parallel calculations on each position through a fully connected feedforward neural network, residual connections are used between Transformer layers, and layer normalization is applied before or after each sublayer to achieve training of deeper network structures; the vector data of multiple transformation sequences output by the transformer structure are concatenated as the input of the multilayer perceptron MLP; the multilayer perceptron MLP is a feedforward neural network model that can It includes: an input layer, one or more hidden layers and an output layer. After learning and training by the multi-layer perceptron MLP, the relevant scores of the target position and other positions in each group of training samples can be output according to the activation function. The trained job search expectation transfer model can predict the current job seeker's delivery sequence data and the relevant scores of multiple positions to be predicted. The corresponding predicted relevant scores of each position to be predicted are sorted from large to small, so as to obtain recommended position information, that is, according to the relevant scores, it can be determined what position the job seeker is most likely to apply for after the expectation transfer; in this way, the information on the change and transfer of job seekers' expectations can be accurately captured and understood, which is of great help for the optimization of person-job matching and job recommendation. In the embodiments of the present application, in the recruitment field, there is a significant connection between job seekers' expectation transfer and time intervals. Job seekers' expectation transfer can be reflected in changes in any element of job information such as job ID, function, industry, work city, salary range and benefits. The trained job expectation transfer model can learn feature changes in training samples. In this way, time interval information is introduced into the training sample sequence information so that the model can better learn and characterize the impact of time factors on expectation transfer. The introduction of position information can characterize the order of job seekers' job submission sequences. The job expectation transfer model can predict the current job seekers' expectation changes, which is conducive to improving the accuracy of job information recommendations and improving the matching degree between job seekers and positions to be recruited.

[0102] In some embodiments, step S220, using the training sample set to train the constructed model to be trained to obtain a trained job search expectation transfer model, may include the following specific steps:

[0103] S221: extracting various influencing factors and position information of the position information in each sample set;

[0104] S222: Mapping each influencing factor in the position information into a first-class feature vector of a corresponding dimension;

[0105] S223: Mapping the position information into a second type of feature vector of a corresponding dimension;

[0106] S224: concatenating each first-category feature vector with the second-category feature vector to obtain concatenated first-category vector data;

[0107] S225: inputting the concatenated first-category vector data into a transformer structure, training the transformer structure, and outputting vector data of a plurality of transformation sequences;

[0108] S226: splicing the vector data of the plurality of conversion sequences to obtain spliced ​​second-category vector data;

[0109] S227: Input the concatenated second-category vector data into a multi-layer perceptron, and train the multi-layer perceptron to output a first correlation score between the target position and the multiple positions that have been applied for.

[0110] Specifically, before entering the Transformer structure, the underlying embedding is linearly mapped to map each vector into the same vector space.

[0111] In some embodiments, step S223, mapping the position information into a second type of feature vector of a corresponding dimension, includes:

[0112] At the t-th position, the i-th vector PE can be calculated by the following formula:

[0113]

[0114] Where: d is the vector dimension; the value of k is (vector dimension d / 2-1);

[0115] When i=2k, the vector is at an even position; when i=2k+1, the vector is at an odd position.

[0116] Specifically, k is the vector dimension number. For example, the vector dimension d = 512 can be grouped in pairs, for a total of 256 groups, k = 0 to 255, 2k corresponds to 0, 2, 4...510, that is, the even positions; 2k+1 corresponds to 1, 3, 5...511, that is, the odd positions.

[0117] In some embodiments, step S220, wherein the training sample set is used to train the constructed model to be trained to obtain a trained job search expectation transfer model, further comprises the following specific steps:

[0118] S228: Input the relevant score into the cross entropy loss function to calculate the loss value;

[0119] S229: According to the calculated loss value, an optimization algorithm is used to calculate the update gradient, and reverse conduction is performed to update the parameters of the job search expectation transfer model and the underlying first-class eigenvector and second-class eigenvector data; wherein the loss function is shown in the following conditional formula:

[0120]

[0121] In the formula, f(x i ) is the first correlation score; when the training sample is a positive sample, y i is 1; when the training sample is a negative sample, y i is 0; m is the total number of sample groups in the sample set.

[0122] The present application embodiment provides a method for recommending job information, based on a job search expectation transfer model generated by the method, such as Figure 4 As shown, the following specific steps may be included:

[0123] S310: Acquire the second historical job posting data of the current job seeker in the second time period; wherein the second historical job posting data includes job information of multiple jobs that have been posted;

[0124] S320: Sort the second historical posting position data in order from the last to the last time to obtain second posting sequence data; wherein the second posting sequence data includes position information of multiple posted positions sorted in order from the closest time to the current time;

[0125] S330: Obtain multiple positions to be predicted;

[0126] S340: Inputting the plurality of positions to be predicted and the second delivery sequence data into the trained job-seeking expectation transfer model to obtain a prediction-related score for each position to be predicted;

[0127] S350: Obtaining recommended job information based on the prediction-related score of each job to be predicted.

[0128] Specifically, the second time period can be set to be the same as the first time period, and both can be set to a predetermined period, such as a two-month period. By obtaining the resume sent by the current job seeker on the recruitment platform after browsing different job descriptions on the terminal device 102, the job information of multiple positions that the job seeker has submitted can be obtained. The job information of multiple positions submitted by the current job seeker is sorted in order from the back to the front from the current time, and the second submission sequence data can be obtained; for example, if the current job seeker has submitted 50 positions in two months, the job that the job seeker has submitted last can be identified as the serial number "1", and according to the order from the back to the front from the current time, the job that the job seeker has submitted the earliest can be identified as the serial number "50", and the other job identifications are similar; wherein, the position information corresponds to the serial number of the position identification of the multiple positions that have been submitted. Input multiple positions to be predicted and the second delivery sequence data into the trained job-seeking expectation transfer model to obtain the prediction related score of each position to be predicted; sort the corresponding prediction related scores of each position to be predicted from large to small, so as to obtain the recommended position information, that is, according to the related scores, the position that the job seeker is most likely to deliver after the expectation transfer can be judged; in this way, the information of the job seeker's expectation change and transfer can be accurately captured and understood, which is very helpful for the optimization of person-job matching and position recommendation. In the embodiment of the present application, in the recruitment field, there is a significant connection between the job seeker's expectation transfer and the time interval. The job seeker's expectation transfer can be reflected in the change of any element in the position information such as position ID, function, industry, working city, salary range and welfare benefits. The trained job-seeking expectation transfer model can learn the feature changes in the training samples, so that the time interval information is introduced into the training sample sequence information so that the model can better learn and characterize the influence of time factors on expectation transfer, and the introduction of position information can characterize the order of the job seeker's delivery position sequence. The job-seeking expectation transfer model can predict the current job seeker's expectation change, which is conducive to improving the accuracy of job information recommendation and improving the matching degree between job seekers and positions to be recruited.

[0129] In some embodiments, step S350, obtaining recommended job information according to the prediction-related score of each job to be predicted, includes the following specific steps:

[0130] The prediction relevance scores of each position to be predicted are sorted in descending order, and the position to be predicted corresponding to the maximum prediction relevance score is taken as the recommended position information.

[0131] In some embodiments, step S310, obtaining a plurality of positions to be predicted, includes the following specific steps:

[0132] Obtain the positions clicked and / or viewed by the current job seeker, and use the positions clicked and / or viewed most recently as the positions to be predicted; or,

[0133] Obtain the job positions posted by the recruiter, and use the nearest job positions to be predicted; or,

[0134] The historical application positions of the current job seeker are recalled, and the positions with high application rates are selected as the positions to be predicted.

[0135] The above describes the implementation methods and advantages of the embodiments of the present application through multiple embodiments. Figure 5 and Figure 6 , describes in detail the specific processing process of the embodiment of the present application.

[0136] 1. Construction process of training samples:

[0137] S11. Get the historical job posting data within two months, extract the posting sequence of each job seeker, and sort them by the time of posting resumes, with the ones closest to the current time at the front, that is, the serial number of the position for which the resume was last submitted is jd1:

[0138] time i ≥time j

[0139] Among them, jd1 is the identification number of the position to which the job seeker last submitted his resume; id n The identification number of the position to which the job seeker first submitted his resume;

[0140] S12. The splitting logic of the sample sequence can be referenced Figure 5 For example, the sequence size in the sample in the sliding window is 51, with a total of 51 positions, including one target position (that is, the position to which the job seeker submitted his resume last is the target position, which is at the top of the sample sequence) and 50 other positions for which resumes have been submitted; for sample sequences with a job seeker's submission sequence < 51, fill the sequence with 0 to make it 51. For job seekers with a submission sequence > 51, split the sequence into multiple sample sequences according to the sliding window of 51, so that 1-51, 2-52, ..., n-50–n can be split in turn, for a total of n-50 sample sequences. For example, if a job seeker submitted resumes to 100 positions in two months, that is, when n = 100, 50 sample sequences can be split.

[0141] S13. Positive sample construction: Each sample sequence containing 51 positions generated in step S12 is marked as a positive sample.

[0142] S14. Negative sample construction: For each sample sequence generated in step S12, the first position is the target value; for example, 10 positions are randomly selected from all positions (excluding the target position) and the first position is replaced. This will generate 10 new sample sequences, which are negative sample sequences.

[0143] S15. Shuffle the positive sample sequence and the negative sample sequence to generate a final sample set.

[0144] S16. Select 20% of the final sample set as test samples and 80% as training samples.

[0145] 2. Training process of job search expectation transfer model:

[0146] S21. Feature extraction and construction of the underlying Embedding layer: position information, location information and their vector mapping.

[0147] a) Position information: Set the position information to a 128-dimensional vector. The calculation formula for the t-th position and the i-th vector PE is as follows:

[0148]

[0149] Where: d is the vector dimension; the value of k is (vector dimension d / 2-1);

[0150] When i=2k, the vector is at an even position; when i=2k+1, the vector is at an odd position.

[0151] b) (3-level classification): Function ID feature mapping is 64-dimensional Embedding;

[0152] c) Time span: the difference in delivery time between position n and the target position, in days or hours, mapped to a 64-dimensional Embedding. For example, if the gap is 5 days, it can be mapped to a 64-dimensional Embedding.

[0153] d) Position ID: Position ID feature map is 64-dimensional Embedding;

[0154] e) Industry (Level 1 function): Industry ID feature mapping is 64-dimensional Embedding;

[0155] f) Working city: The city ID feature can be mapped to a 64-dimensional Embedding;

[0156] g) Salary range: The salary ID feature is mapped to a 64-dimensional Embedding. For example, if the salary range is 6000-7999, it can be marked as 5 and mapped to a 64-dimensional Embedding.

[0157] S22. Concat: Perform concat operation on the location information and position information, and concatenate the above 7 vectors into a large vector Embedding_concat_1. The dimension of the concatenated vector is a 512-dimensional vector.

[0158] S23. Input Embedding_concat_1 as input data into the Transformer structure.

[0159] S24. Build Transformer structure:

[0160] a)Multi-Head Self-Attention

[0161] i. The conversion formula is as follows:

[0162]

[0163] MultiHead(Q,K,V)=Concat(head1,...head n )W o

[0164] Among them head i =Attention(QW i Q , KW i k , VW i v )

[0165] Among them: the number of heads is 8 dimensions, and each head vector is 64 dimensions;

[0166] ii. Assign Embedding_concat_1 to QKV, i.e., Q = Embedding_concat_1, K = Embedding_concat_1, V = Embedding_concat_1;

[0167] Perform a linear mapping on each QKV, map it to the same vector space, and split it into 8 heads;

[0168]

[0169] Perform Attention(Q, K, V) on QKV, then concat the result of Attention(Q, K, V), and finally perform a FC (linear mapping of the fully connected layer), recorded as Embedding_concat_2'

[0170] Concat(head1,...head n )W o

[0171] b)Add&Norm

[0172] Embedding_2=Embedding_concat_1+Embedding_concat_2'

[0173] Normalize, Embedding_3 = LN(Embedding_2)

[0174]

[0175] Normalization is performed at the layer level, where the mean μ and variance σ are the mean and variance of all feature vector dimensions under the layer;

[0176] The role of Add is residual learning.

[0177] c) Feed Forward (Feed Forward Network Layer)

[0178] i. Two-layer FC, the dimension of the hidden layer vector is 512

[0179] Embedding_4=FC2(FC1(Embedding_3))

[0180] FC(X)=W*X+B

[0181] In the formula, W is the weight, X is the input vector, and B is the bias;

[0182] d)Add&Norm

[0183] Same as the first Add&Norm logic above

[0184] Add:Embedding_5=Embedding_3+Embedding_4

[0185] Norm:Embedding_5=LN(Embedding_5)

[0186] S25. Concat: Concatenate the 51 embedding vectors output by the transformer layer together, denoted as Embedding_concat_6, as the input of the MLP.

[0187] S26. Building MLP (Multi-layer Perceptron)

[0188] a) 4-layer FC, 3 hidden layer vector dimensions are 512, 512, 256 respectively, the output layer is 1-dimensional, and sigmoid is used as the activation function;

[0189] b)Embedding_7=FC4(FC3(FC2(FC1(Embedding_concat_6))))

[0190] FC(X)=W*X+B

[0191] predict=sigmoid(Embedding_7)

[0192] The conditional form of the activation function is as follows:

[0193]

[0194] The correlation score can be calculated by the above formula, and the value of the correlation score is between 0 and 1.

[0195] S27. Construct loss function

[0196] The loss function adopts the cross entropy loss function. The relevant score result calculated by the activation function sigmoid in step S26 is input into the cross entropy loss function loss to calculate the loss value. The specific formula is as follows:

[0197]

[0198] Where f(x i ) is the correlation score calculated in step S26. When the training sample is a positive sample, y i is 1; when the training sample is a negative sample, y i is 0; m is the total number of sample groups in the sample set.

[0199] S28. Model parameter update (learning)

[0200] Based on the loss value calculated in step S27, the adagrad optimization algorithm is used to calculate the updated gradient, and the model parameters and the underlying embedding vector are updated by reverse conduction. Steps S27-S28 are repeated until the predetermined number of training rounds is reached or other stopping conditions are met, the training stops, and the trained job search expectation transfer model is obtained.

[0201] 3. Reasoning process of recommended job information:

[0202] S31. For a job seeker, multiple positions that the job seeker has applied for in the past 60 days can be extracted, and the application time of the positions can be sorted in order from the back to the current time to obtain a sequence of applied positions (the order is the same as the order in step S11), and the top 50 positions are intercepted (if there are less than 50 positions, they are filled with 0 at the end) as the prediction sequence.

[0203] S32. Get the positions to be predicted. For example, there are three positions to be predicted, jd1, jb2, and jb3. Input the three positions to be predicted and the prediction sequence into the trained job-seeking expectation transfer model respectively to obtain three corresponding correlation scores. Compare them according to the sizes of the correlation scores, and take the position to be predicted corresponding to the maximum value of the correlation score as the recommended position information for the job seeker, that is, the position that the job seeker is most likely to apply for after the expectation transfer.

[0204] The following is an example of the shift in salary expectations during the job search life cycle:

[0205] There is a group of job seekers who have high expectations at the beginning. As the feedback from their resumes and interview results are not ideal, their expected salary gradually decreases. The following is the actual data of job seekers' job applications from the beginning to the end in order:

[0206] 6000-7999,6000-7999,8000-9999,6000-7999,4500-5999,4500-5999,6000-7999,4500-5999,4500-5999,3000-4499,4500-5999,4500-5999 (the entire submission sequence is longer than this, only part is extracted here for illustration). It is not difficult to see from the above data that the positions submitted at the beginning are concentrated in 6000-7999 yuan or even higher, and the positions submitted later are concentrated in 4500-5999 yuan or even lower. Therefore, the position information of 6000-7999 yuan can be used as the first position to be predicted, and the position information of 4500-5999 yuan can be used as the second position to be predicted.

[0207] Sorting in order from the last to the last time, the delivery sequence data can be obtained as follows:

[0208] 4500-5999,4500-5999,3000-4499,4500-5999,4500-5999,6000-7999,4500-5999,4500-5999,6000-7999,8000-9999,6000-7999,6000-7999;

[0209] The predicted positions 1 and 2 are respectively input into the trained job application expectation transfer model with the above delivery sequence data, and the output correlation scores are approximately 0.65 and 0.89, respectively. It can be determined that positions with a salary of 4500-5999 yuan are more suitable for job seekers, thereby realizing a prediction of the job seeker's expected salary transfer. In this embodiment, taking salary as an example, the job seeker's expected functions, industries, and working cities all have different transfers. The algorithm can reflect the job seeker's expected transfer by integrating these influencing factors as a whole through feature fusion.

[0210] The embodiment of the present application proposes an algorithm based on the Transformer mechanism to describe the expected transfer of job seekers. The algorithm is applied to the advertisement ranking model, and the display-to-investment ratio (resume delivery volume / exposure volume) can be increased by 5.7%. The algorithm is applied to the click-through rate and conversion rate model, and can be used for click-through rate and conversion rate estimation.

[0211] Corresponding to the method embodiment of the present application, the present application also provides a training sample construction device, such as Figure 7 As shown, including:

[0212] A first acquisition module 710 is used to acquire first historical job posting data of a job seeker within a first time period; wherein the first historical job posting data includes job information of multiple jobs that have been posted;

[0213] A first sequence module 720 is used to sort the first historical posting position data in order from the last time to the last time to obtain first posting sequence data; wherein the first posting sequence data includes position information of multiple posted positions sorted in order from the last time to the last time;

[0214] A segmentation module 730 is used to segment the first delivery sequence data into multiple sets of positive sample sets according to a preset segmentation strategy; each set of positive sample sets includes a target position and n delivered positions, and the target position is the position in each set of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2;

[0215] A selection module 740 is used to randomly select a plurality of positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions;

[0216] A replacement module 750, which is used to replace the target position in each set of positive samples with the pre-selected position to obtain a corresponding plurality of sets of negative samples;

[0217] The training sample module 760 is used to obtain a training sample set according to the positive sample set and the negative sample set.

[0218] The embodiment of the present application provides a device for generating a job-seeking expectation transfer model, based on a training sample set constructed by the method described, such as Figure 8 As shown, including:

[0219] A construction module 810 is used to construct a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and vector data of multiple conversion sequences output by the transformer structure are concatenated and used as input of the multi-layer perceptron;

[0220] The training module 820 is used to train the constructed model to be trained using the training sample set to obtain a trained job-seeking expectation transfer model; wherein the job-seeking expectation transfer model is used to predict the current job seeker's submission sequence data and the relevant scores of multiple positions to be predicted.

[0221] In some embodiments, the training module 820 includes:

[0222] An extraction unit, which is used to extract various influencing factors and position information of the position information in each set of sample sets;

[0223] A first mapping unit, which is used to map each influencing factor in the position information into a first type of feature vector of a corresponding dimension;

[0224] A second mapping unit, configured to map the position information into a second type of feature vector of a corresponding dimension;

[0225] A first concatenation unit, which is used to concatenate each first-category feature vector and the second-category feature vector to obtain concatenated first-category vector data;

[0226] A first training unit, which is used to input the concatenated first-category vector data into a transformer structure, train the transformer structure, and output vector data of a plurality of transformation sequences;

[0227] A second splicing unit, which is used to splice the vector data of multiple conversion sequences to obtain spliced ​​second-category vector data;

[0228] The second training unit is used to input the concatenated second-category vector data into a multilayer perceptron, and train the multilayer perceptron to output a first correlation score between the target position and the multiple positions that have been applied for.

[0229] The embodiment of the present application provides a device for recommending job information, based on the job search expectation transfer model generated by the method, such as Fig. 9 As shown, including:

[0230] The second acquisition module 910 is used to acquire the second historical job posting data of the current job seeker within the second time period; wherein the second historical job posting data includes job information of multiple jobs that have been posted;

[0231] A second sequence module 920, which is used to sort the second historical posting position data in order from the last to the last time to obtain second posting sequence data; wherein the second posting sequence data includes position information of multiple posted positions sorted in order from the closest to the current time;

[0232] A third acquisition module 930 is used to acquire a plurality of positions to be predicted;

[0233] An input module 940 is used to input a plurality of positions to be predicted and the second delivery sequence data into the trained job-seeking expectation transfer model to obtain a prediction correlation score for each position to be predicted;

[0234] The recommendation module 950 is used to obtain recommended job information according to the prediction-related score of each job to be predicted.

[0235] The electronic device in the embodiment of the present application may be a user terminal device, a server, other computing devices, or a cloud server. Fig.10 A schematic diagram of the hardware structure of an electronic device according to an embodiment of the present application is shown. The electronic device may include a processor 601 and a memory 602 storing computer program instructions. When the processor 601 executes the computer program instructions, the process or function of any of the above-mentioned embodiments of the method is implemented.

[0236] Specifically, the processor 601 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application. The memory 602 may include a large-capacity memory for data or instructions. For example, the memory 602 may be at least one of the following: a hard disk drive (HDD), a read-only memory (ROM), a random access memory (RAM), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a tape, a universal serial bus (USB) drive or other physical / tangible memory storage device. For another example, the memory 602 may include a removable or non-removable (or fixed) medium. For another example, the memory 602 may be inside or outside the integrated gateway disaster recovery device. The memory 602 may be a non-volatile solid-state memory. In other words, the memory 602 usually includes a tangible (non-transitory) computer-readable storage medium (such as a memory device) encoded with computer-executable instructions, and when the software is executed (such as executed by one or more processors), the operation described in the method of the embodiment of the present application can be performed. The processor 601 implements the process or function of any method in the above embodiments by reading and executing the computer program instructions stored in the memory 602.

[0237] In one example, Fig.10 The electronic device shown may also include a communication interface 603 and a bus 610. Among them, the processor 601, the memory 602, and the communication interface 603 are connected through the bus 610 and complete the communication between each other. The communication interface 603 is mainly used to realize the communication between each module, device, unit and / or device in the embodiment of the present application. The bus 610 includes hardware, software or both, and can couple the components of the online data traffic billing device to each other. For example, the bus may include at least one of the following: an accelerated graphics port (AGP) or other graphics bus, an enhanced industrial standard architecture (EISA) bus, a front-end bus (FSB), a hypertransport (HT) interconnect, an industrial standard architecture (ISA) bus, an infinite bandwidth interconnect, a low pin count (LPC) bus, a memory bus, a micro channel architecture (MCA) bus, a peripheral component interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a serial advanced technology attachment (SATA) bus, a video electronics standard association local (VLB) bus or other suitable buses. The bus 610 may include one or more buses. Although the embodiments of the present application describe or illustrate a specific bus, the embodiments of the present application may consider any suitable bus or interconnection method.

[0238] In combination with the method in the above embodiments, an embodiment of the present application also provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the process or function of any method in the above embodiments is implemented.

[0239] In addition, an embodiment of the present application further provides a computer program product, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the process or function of any one of the methods in the above embodiments is implemented.

[0240] The above exemplarily describes the flowcharts and / or block diagrams of the methods, devices, systems and computer program products of the embodiments of the present application, and describes the relevant various aspects. It should be understood that each box or combination thereof in the flowchart and / or block diagram can be implemented by computer program instructions, or by dedicated hardware that performs a specified function or action, or by a combination of dedicated hardware and computer instructions. For example, these computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to form a machine that enables these instructions executed by such a processor to enable the implementation of the functions / actions specified in each box or combination thereof in the flowchart and / or block diagram. Such a processor can be a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit.

[0241] The functional blocks shown in the structured block diagram of the embodiment of the present application can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc.; when implemented in software, it is a program or code segment used to perform the required task. The program or code segment can be stored in a memory, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0242] It should be noted that the present application is not limited to the specific configurations and processes described above or shown in the figures. The above is only a specific implementation mode of the present application. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the described system, device, module or unit can refer to the corresponding process in the method embodiment without further description. It should be understood that the scope of protection of the present application is not limited to this. Any technician familiar with the technical field can think of various equivalent modifications or substitutions within the technical scope disclosed in the present application, and these modifications or substitutions should be included in the scope of protection of the present application.

Claims

1. A method for constructing a training sample, characterized in that: include: Acquire the first historical job posting data of the job seeker in the first time period; wherein the first historical job posting data includes job information of multiple jobs that have been posted; The first historical posting position data is sorted in order from the last to the last time to obtain first posting sequence data; wherein the first posting sequence data includes position information of multiple posted positions sorted in order from the last to the last time; The first delivery sequence data is divided into multiple groups of positive sample sets according to a preset segmentation strategy; wherein each group of positive sample sets includes a target position and n delivered positions, and the target position is the position in each group of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2; Randomly selecting a plurality of positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions; The pre-selected positions are used to replace the target positions in each set of positive samples to obtain corresponding sets of negative samples. According to the positive sample set and the negative sample set, a training sample set is obtained.

2. The method according to claim 1, characterized in that The step of dividing the first delivery sequence data into a plurality of positive sample sets according to a preset division strategy includes: Using a set sliding window to sequentially intercept the first delivery sequence data to obtain multiple groups of sequence samples; The positions corresponding to the sequences in each group of sequence samples are marked as positive samples, and the positions corresponding to the first sequence number in each group of sequence samples are taken as target positions.

3. The method according to claim 1, characterized in that The method of dividing the first delivery sequence data into a plurality of positive sample sets according to a preset division strategy further includes: If the number of samples in the current sequence of samples does not meet the number of samples set in the sliding window, "0" is filled after the sequence of the current sequence of samples to make the number of samples in the current sequence of samples meet the number of samples set in the sliding window.

4. The method according to claim 1, characterized in that: The job information includes one or more of the following influencing factors: job ID, function, industry, work city, salary range, and the time difference between the nth job and the target job.

5. A method for generating a job search expectation transfer model, characterized in that: The training sample set constructed based on the method according to any one of claims 1 to 4 comprises: Constructing a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and the vector data of multiple conversion sequences output by the transformer structure are concatenated as the input of the multi-layer perceptron; The constructed model to be trained is trained using the training sample set to obtain a trained job search expectation transfer model; wherein, The job-seeking expectation transfer model is used to predict the current job seeker's delivery sequence data and the relevant scores of multiple positions to be predicted.

6. The method according to claim 5, characterized in that The step of using the training sample set to train the constructed model to be trained to obtain a trained job search expectation transfer model includes: Extract various influencing factors and position information of job information in each sample set; Map each influencing factor in the job information into a first-class eigenvector of the corresponding dimension; Mapping the position information into a second type of feature vector of corresponding dimension; Concatenate each first-category eigenvector and the second-category eigenvector to obtain concatenated first-category vector data; Input the concatenated first-category vector data into a transformer structure, and train the transformer structure to output vector data of multiple transformation sequences; splicing the vector data of the plurality of conversion sequences to obtain the spliced ​​second-category vector data; The concatenated second-category vector data is input into a multilayer perceptron, and the multilayer perceptron is trained to output a first correlation score between the target position and the multiple positions that have been applied for.

7. The method according to claim 6, characterized in that Mapping the position information into a second type of feature vector of a corresponding dimension includes: At the t-th position, the i-th vector PE can be calculated by the following formula: Where: d is the vector dimension; the value of k is (vector dimension d / 2-1); When i=2k, the vector is at an even position; when i=2k+1, the vector is at an odd position.

8. The method according to claim 5, characterized in that The step of using the training sample set to train the constructed model to be trained to obtain a trained job search expectation transfer model further includes: Input the relevant scores into the cross entropy loss function and calculate the loss value; According to the calculated loss value, the optimization algorithm is used to calculate the update gradient, and the parameters of the job search expectation transfer model and the underlying first-class eigenvector and second-class eigenvector data are updated through reverse conduction; wherein the loss function is shown in the following conditional formula: In the formula, f(x i ) is the first correlation score; when the training sample is a positive sample, y i is 1; when the training sample is a negative sample, y i is 0; m is the total number of sample groups in the sample set.

9. A training sample construction device, characterized in that: include: A first acquisition module is used to acquire first historical job posting data of a job seeker within a first time period; wherein the first historical job posting data includes job information of multiple jobs that have been applied for; A first sequence module, which is used to sort the first historical posting position data in order from the current time to the previous time, so as to obtain first posting sequence data; wherein the first posting sequence data includes position information of multiple posted positions sorted in order from the current time to the previous time; A segmentation module, which is used to segment the first delivery sequence data into multiple sets of positive sample sets according to a preset segmentation strategy; wherein each set of positive sample sets includes a target position and n delivered positions, and the target position is the position in each set of positive sample sets whose delivery time is closest to the current time; n is greater than or equal to 2; A selection module, which is used to randomly select multiple positions from the delivery sequence data to obtain pre-selected positions; wherein the pre-selected positions are non-target positions; A replacement module, which is used to replace the target position in each set of positive samples with the pre-selected position to obtain a corresponding set of negative samples; The training sample module is used to obtain a training sample set based on a positive sample set and a negative sample set.

10. A device for generating a job-seeking expectation transfer model, characterized in that: The training sample set constructed based on the method according to any one of claims 1 to 4 comprises: A construction module is used to construct a model to be trained; wherein the model to be trained includes a transformer structure and a multi-layer perceptron, and vector data of multiple conversion sequences output by the transformer structure are concatenated as input of the multi-layer perceptron; A training module is used to train the constructed model to be trained using the training sample set to obtain a trained job search expectation transfer model; wherein, The job-seeking expectation transfer model is used to predict the current job seeker's delivery sequence data and the relevant scores of multiple positions to be predicted.

11. The generating device according to claim 10, characterized in that: The training module includes: An extraction unit, which is used to extract various influencing factors and position information of the position information in each set of sample sets; A first mapping unit, which is used to map each influencing factor in the position information into a first type of feature vector of a corresponding dimension; A second mapping unit, configured to map the position information into a second type of feature vector of a corresponding dimension; A first concatenation unit, which is used to concatenate each first-category feature vector and the second-category feature vector to obtain concatenated first-category vector data; A first training unit, which is used to input the concatenated first-category vector data into a transformer structure, train the transformer structure, and output vector data of a plurality of transformation sequences; A second splicing unit, which is used to splice the vector data of multiple conversion sequences to obtain spliced ​​second-category vector data; The second training unit is used to input the concatenated second-category vector data into a multilayer perceptron, and train the multilayer perceptron to output a first correlation score between the target position and the multiple positions that have been applied for.

12. An electronic device, characterized in that: The electronic device comprises: a processor and a memory storing computer program instructions; when the electronic device executes the computer program instructions, the method according to any one of claims 1 to 8 is implemented.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images