Personalized employment information accurate pushing method based on deep learning

Abstract

The invention discloses a personalized employment information accurate pushing method based on deep learning, and relates to the technical field of artificial intelligence data processing, and the method comprises the steps: executing post screening based on user information and post information, extracting user baseline information through a behavior sequence, and generating a post set; performing feature coding on the post set and the user baseline information to generate a post feature vector and a user feature vector, and inputting the post feature vector and the user feature vector into a deep learning model to generate a personalized employment information push sequence; and based on the newly added behavior event in the personalized employment information push sequence, calculating a deep learning model prediction deviation and a convergence threshold, and when the deep learning model prediction deviation is greater than or equal to the convergence threshold, executing deep learning model self-learning updating, and generating an updated personalized employment information push sequence. According to the invention, through prediction deviation calculation and self-learning updating, dynamic optimization of the post pushing threshold is realized, and the real-time performance and accuracy of personalized employment information pushing are enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, and in particular to a method for accurately pushing personalized employment information based on deep learning. Background Technology

[0002] With the rapid development of online recruitment platforms and intelligent human resource services, personalized job recommendations have become an important way to connect job seekers and employers. Conventional job recommendation methods primarily rely on user-provided resume information, historical application records, and job database content, using keyword matching, similarity calculation, or rule-based filtering to establish a correspondence between jobs and users. In recent years, with the growth of data scale and the improvement of computing power, some methods have incorporated statistical modeling and shallow machine learning algorithms to uncover the potential correlation between user interests and job characteristics, thereby improving the targeting and accuracy of job recommendations to a certain extent. These methods have been widely used in recruitment platforms and talent management systems, forming a relatively mature basic technical system.

[0003] However, traditional job recommendation methods still have limitations in complex data environments. On the one hand, the matching relationship between jobs and users is mostly based on static feature extraction, which makes it difficult to reflect the dynamic characteristics of job-seeking behavior over time and to fully capture the phased changes in user interests. On the other hand, recommendation models lack adaptive learning and updating mechanisms during long-term operation. When user behavior patterns or job requirement structures change, the recommendation results are difficult to optimize in real time, thus limiting the timeliness and personalization of job recommendations. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for accurately pushing personalized employment information based on deep learning, which solves the problems of not being able to dynamically depict changes in user behavior and lacking an adaptive learning and update mechanism in the job recommendation process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for accurately pushing personalized employment information based on deep learning, which includes collecting user information and job information, performing structured processing, and generating behavioral sequences; Job selection is performed based on user information and job information, and user baseline information is extracted through behavioral sequences to generate a set of jobs; Feature encoding is performed on the job set and user baseline information to generate job feature vectors and user feature vectors, which are then input into a deep learning model to generate a personalized employment information push sequence. Based on the new behavioral events in the personalized employment information push sequence, the prediction bias and convergence threshold of the deep learning model are calculated. When the prediction bias of the deep learning model is greater than or equal to the convergence threshold, the deep learning model is self-learned and updated to generate an updated personalized employment information push sequence.

[0007] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the user information includes resume information and behavioral information; The resume information includes educational background, work experience, professional skills, qualification certificates, desired work location, expected salary and available start date; The behavioral information includes the user's browsing history, favorites history, application history and interview history on the employment platform; The job information includes the job title, job responsibilities, job requirements, salary range, job location, required qualifications, and start date requirements.

[0008] As a preferred embodiment of the deep learning-based personalized employment information accurate push method described in this invention, the structured processing includes unifying the field format of resume information and job information, removing redundant items and filling in missing fields, and establishing a one-to-one correspondence between resume information and job information.

[0009] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the generation of the behavior sequence includes sorting the browsing records, collection records, application records and interview records in the behavior information according to the time order, and integrating all the behavior records of the same user into a behavior sequence according to the user identifier.

[0010] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the job screening based on user information and job information includes: matching and screening based on the corresponding fields in the structured resume information and job information; comparing educational background and academic qualifications, work experience and years of work experience, professional skills and required qualifications, qualification certificates and required qualifications, expected work location and job location, expected salary and salary range, and available time and required time to work; retaining job records that meet the matching and screening conditions; and summarizing them to form a pre-screened job set. The pre-screened job list is screened based on job title and job responsibilities. When the job title belongs to the same occupational category as the job category in the work experience section, and the job responsibilities description matches the skill tags in the occupational skills section, the job record is retained, and a job matching set is generated.

[0011] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the step of extracting user baseline information through behavior sequence and generating job set includes: analyzing the event type, event time and event object in the behavior sequence, counting the number of times browsing records, collection records, application records and interview records appear in each job category, and recording the number of events corresponding to each job category as the behavior frequency field; The job records in the job matching set are sorted according to the value of the behavior frequency field, with job categories with higher behavior frequency field values ​​appearing first. Extract user baseline information by combining resume information and behavioral sequences; The user baseline information includes user location, user salary, user qualifications, and user attendance. Based on the sorted job matching set, a further filtering is performed according to the user's baseline information. If any four or more of the job location, salary range, required qualifications, start date requirements, job responsibilities, and job category match the user's baseline information, the job record is retained. The job records retained after the further filtering are then summarized to form a job set.

[0012] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the generation of job feature vectors and user feature vectors includes performing feature encoding on job records and user baseline information in the job set, performing semantic transformation on text fields, and performing normalization transformation on numerical fields. The job feature vector is formed by combining the job title, job responsibilities, job requirements, job location, salary range, required qualifications, and start date requirements in that order. The user feature vector is formed by concatenating the user's location, salary, qualifications, and attendance in that order.

[0013] As a preferred embodiment of the personalized employment information push method based on deep learning described in this invention, the step of generating a personalized employment information push sequence by inputting the deep learning model includes inputting the job feature vector and the user feature vector into the job channel and the user channel of the deep learning model respectively, performing nonlinear mapping, and generating the job hidden layer vector and the user hidden layer vector. In the fusion layer, the inner product operation and normalization are performed on the hidden vectors of job positions and users to calculate the feature similarity scores between job positions and users. The feature similarity score is input into the activation function to calculate the matching probability. Based on the average and standard deviation of the matching probabilities in the job set, the job push threshold is calculated. When the job matching probability is greater than or equal to the job push threshold, the job record is retained and sorted from high to low matching probability to generate a personalized employment information push sequence.

[0014] As a preferred embodiment of the personalized employment information accurate push method based on deep learning described in this invention, the calculation of the prediction bias and convergence threshold of the deep learning model includes: adding job records that generate new behavioral events in the personalized employment information push sequence to the positive sample set, and adding job records that do not generate new behavioral events to the negative sample set. The absolute difference between the average positive sample matching probability and the average negative sample matching probability is calculated to obtain the prediction bias of the deep learning model. Calculate the average matching probability during the learning phase based on the distribution of matching probabilities of all job records in the positive and negative sample sets. Centered on the average matching probability during the learning phase, the variance of the matching probability of job records in the positive and negative sample sets is calculated to obtain the standard deviation of the matching probability, which is used as the convergence threshold. When the prediction bias of a deep learning model is greater than or equal to the convergence threshold, the deep learning model is determined to have not reached convergence.

[0015] As a preferred embodiment of the personalized employment information push method based on deep learning described in this invention, the generation of the updated personalized employment information push sequence includes, when the deep learning model has not reached convergence, performing gradient descent updates on the weight matrix of the job channel, the weight matrix of the user channel, the bias vector of the job channel, and the bias vector of the user channel based on the positive sample set and the negative sample set. After the update is completed, the job matching probability and job push threshold are recalculated, and an updated personalized employment information push sequence is generated.

[0016] The beneficial effects of this invention are as follows: by using the dual-channel nonlinear mapping of a deep learning model, the fusion expression of job features and user features is realized, thereby improving the matching accuracy between jobs and users; by calculating prediction deviation and updating through self-learning, the dynamic optimization of job push threshold is realized, thereby enhancing the real-time performance and accuracy of personalized employment information push. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for accurately pushing personalized employment information based on deep learning.

[0019] Figure 2This is a flowchart for job screening and user baseline information extraction.

[0020] Figure 3 A flowchart for generating personalized push sequences.

[0021] Figure 4 A flowchart for self-learning updates of deep learning models. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for accurately pushing personalized employment information based on deep learning, including the following steps: S1. Collect user information and job information, perform structured processing, and generate behavior sequences.

[0026] Furthermore, user information is collected, including resume information and behavioral information.

[0027] Resume information includes educational background, work experience, professional skills, qualification certificates, expected work location, expected salary, and available start date; behavioral information includes the user's browsing history, favorites history, application history, and interview history on the employment platform, and records the time of each behavioral information; a user identifier is set for each user to match resume information with behavioral information; all user information is collected with the user's consent and used for legitimate purposes.

[0028] We collect job information, including job title, job responsibilities, job requirements, salary range, job location, required qualifications, and start date requirements; we assign a job identifier to each job posting to match job information with behavioral information; the job information comes from recruitment platform databases or enterprise recruitment interfaces and is used only for legitimate data analysis and employment information matching.

[0029] Furthermore, the collected resume and job information are processed in a structured manner.

[0030] Standardize the field format of resume information and job information, remove redundant items and fill in missing fields, and establish a one-to-one correspondence between resume information and job information.

[0031] Redundancy removal includes deleting resume and job information that is completely identical or submitted repeatedly, such as resumes uploaded multiple times by the same user or content for the same job posting repeatedly on different channels.

[0032] Missing field completion includes filling in missing expected salaries based on users' historical job search records, or filling in missing job responsibilities based on the company's job database; the priority of completion is determined by time order, with the most recent valid record taking priority, followed by the latest complete record in the platform's database.

[0033] In the resume information, education experience is uniformly organized into education level names, such as doctoral, master's, bachelor's, associate's, and high school; work experience is organized into years of work and job category; professional skills are organized into a set of skill tags, with each skill tag corresponding to a standard skill term; qualification certificates are organized into certificate names and level names; desired work location is uniformly organized into city names; expected salary is uniformly organized into monthly salary ranges; and available time is uniformly organized into the number of days.

[0034] In the job postings, job titles are standardized to job category names, such as Mechanical Engineer, Data Analyst, and Software Development Engineer; key responsibility phrases are extracted for job duties, each recorded in a verb-object structure, such as responsible for equipment maintenance, production line management, and data interface development; job requirements are extracted as independent sub-items such as education requirements, years of work experience requirements, and professional direction requirements; salary ranges are standardized to monthly salary intervals; job locations are standardized to city-level expressions; required qualifications are compiled into a list of certificate level requirements or a list of skill requirements; and start date requirements are standardized to a number of days.

[0035] Establish a one-to-one correspondence between resume information and job information, matching educational background with the academic qualifications required in the job requirements, work experience with the years of service required in the job requirements, professional skills with the required qualifications, professional certificates with the required qualifications, expected work location with the job location, expected salary with the salary range, available start date with the required start date, job category in work experience with job title, and skill tags in professional skills with key responsibility phrases in job responsibilities.

[0036] Furthermore, the browsing history, favorites history, submission history, and interview history in the behavioral information are sorted in chronological order, and all behavioral records of the same user are integrated into a continuous behavioral sequence based on the user identifier. During the integration process, a time index is generated for each behavioral record, the time of the event is recorded in ISO date and time format, and the sequence number is used to mark the order of events.

[0037] The behavior sequence is set with three fields: event type, event time, and event object. The event object records the value of the job identifier, which is used to perform field-level mapping between the behavior sequence and the job information. The event types include browsing, favorites, application, and interview.

[0038] The frequency distribution of different event types in the time dimension is statistically analyzed in the form of numerical values ​​or labels in the behavioral sequence to characterize the user's job search activity and job interest trends in a continuous time period.

[0039] S2. Perform job screening based on user information and job information, extract user baseline information through behavior sequence, and generate a set of jobs.

[0040] Furthermore, based on the structured resume information, job information, and behavioral sequences, job screening and user baseline extraction are performed. The specific steps are as follows: S201. Match and filter based on the corresponding fields in the resume information and job information.

[0041] The resume information is compared with the educational background in the job posting, retaining job records where the educational level meets the job requirements; the resume information is compared with the work experience in the job posting, retaining job records where the years of work experience meet the job requirements; the resume information is compared with the job posting's required qualifications, retaining job records where the skill sets overlap; the resume information is compared with the job posting's required qualifications, retaining job records where the certificate level meets the job requirements; the resume information is compared with the job posting's location, retaining job records where the geographical location is the same or adjacent cities; the resume information is compared with the job posting's salary range, retaining job records where the salary ranges overlap; and the resume information is compared with the job posting's start date, retaining job records where the start date requirement is no earlier than the user's start date.

[0042] After completing the matching and filtering, all job records retained by the matching and filtering are summarized to form a pre-screened job set.

[0043] S202. Screen the pre-screened job list based on job title and job responsibilities.

[0044] The job title is compared with the work experience in the resume information by occupational category. The job record is retained when the occupational category of the job title matches the job category in the work experience. The job responsibilities are compared with the professional skills in the resume information by responsibility keywords. The job record is retained when the job responsibility description contains skill tags or work task words from the professional skills.

[0045] After screening job titles and job responsibilities, job records that meet both the job title matching criteria and the job responsibility matching criteria are merged, duplicate job records are removed, and a job matching set is generated.

[0046] S203. Adjust the sorting and priority of the job matching set through behavior sequence.

[0047] By analyzing the event type, event time, and event object in the behavioral sequence, the frequency of occurrence of browsing history, collection history, application history, and interview history in each job category is counted.

[0048] The job records in the job matching set are sorted according to the frequency of occurrence; job categories that occur more frequently are ranked first.

[0049] The sorting results record the number of events corresponding to each job category as a behavior frequency field, which reflects the level of user interaction activity in different job categories.

[0050] S204. Extract user baseline information from resume information and behavioral sequences, including user location, user salary, user qualifications, and user attendance.

[0051] The user's location is determined by the expected work location in the resume information. The location is determined by matching the event objects in the behavior sequence with the job identifiers in the job information, reading the job location from the job records, and counting the cities that appear most frequently.

[0052] User compensation is determined by the expected salary range in the resume information. The salary range is read from the job records by matching the event objects in the behavior sequence with the job identifiers in the job information. When the behavior frequency field reflects that the user has submitted many events in a certain job category, the salary range corresponding to the job category is used first. If there are multiple salary records, the average of the salary ranges of the job records corresponding to the three most recent submissions is taken as the user's compensation.

[0053] User qualifications are determined by the basic abilities of the qualification certificates in the resume information. The required qualification conditions are read from the job record based on the correspondence between the event objects in the behavior sequence and the job identifiers in the job information. When the certificate or skill name in the resume information is consistent with the job information requirements, the highest level corresponding item is selected as the user qualification according to the level order.

[0054] A user's availability for work is calculated by comparing the available time in their resume with the time difference between the most recent interview record in their behavioral sequence. The result is in days and represents the user's actual available time range.

[0055] S205. Based on the sorted job matching set, perform further filtering according to the user baseline information.

[0056] For job records in the job matching set, conditional comparison is performed based on user baseline information.

[0057] The comparison criteria include whether the job location is the same as or in a neighboring city as the user's location, whether the salary range overlaps with the user's salary range, whether the required qualifications are no higher than the user's qualifications, whether the start date requirement is no earlier than the user's start date, whether the job responsibilities include the professional skills tags in the resume information, and whether the job title's occupational category is consistent with the work experience category in the resume information.

[0058] If a job record meets any four or more of the conditions, the job record will be retained.

[0059] The remaining job records after further filtering are compiled into a job set.

[0060] S3. Perform feature encoding on the job set and user baseline information to generate job feature vectors and user feature vectors, and input them into a deep learning model to generate a personalized employment information push sequence.

[0061] Furthermore, feature encoding is performed on the job records and user baseline information in the job set, and a vectorized approach is used to achieve semantic transformation of text fields and normalization transformation of numerical fields.

[0062] Text fields include job title, job responsibilities, job requirements, required qualifications, and user qualifications. The semantic content of text fields is processed using word vectorization. Words in each field are represented by word frequency statistics to create word vectors, and corresponding semantic vectors are obtained through vector averaging. All semantic vectors are normalized by unit length to maintain consistency in magnitude.

[0063] Numerical fields include salary range, job location, user location, arrival time requirement, user salary, and user arrival time. The normalization rule for numerical fields is a scaling method based on the minimum and maximum value intervals, meaning the scaling interval is determined by the minimum and maximum values ​​of the field's historical samples, mapping all numerical field values ​​to... The range is defined by taking the critical value beyond the boundary.

[0064] The job feature vector is formed by combining the following information in the order of job title, job responsibilities, job requirements, job location, salary range, required qualifications, and start date requirements.

[0065] The user feature vector is formed by concatenating the user's location, salary, qualifications, and attendance in that order.

[0066] Furthermore, the job feature vector and user feature vector are input into the deep learning model.

[0067] Deep learning models belong to the existing neural network architecture and adopt a dual-channel structure. They perform nonlinear mapping on job feature vectors and user feature vectors through job channel and user channel respectively, so as to realize parallel learning and fusion of job features and user features.

[0068] During the input phase of the deep learning model, the job feature vector and user feature vector are input into the job feature extraction function and user feature extraction function, respectively, as follows: ; ; in, This represents the hidden layer vector of a job position, i.e., the hidden layer representation of job features. This represents the user's hidden layer vector, which is the hidden layer representation of the user's features. This represents the weight matrix for job channels, used to control the relative importance of each field in the job characteristics. This represents the weight matrix for the user channel, used to control the influence ratio of each field in the user baseline information. Indicates time index The following job feature vectors Indicates time index The user feature vector below, This represents a non-linear activation function, exemplified by ReLU. This represents the bias vector for the job channel. This represents the bias vector of the user channel.

[0069] The bias vector of the job channel is used to correct the overall mean of the job feature vector. It consists of the average of all numerical fields in the job set, including salary range, job location, and start date requirements.

[0070] The bias vector of the user channel is used to correct the overall mean of the user feature vector. It consists of the average value of the user baseline information and the numerical fields in the behavioral sequence, including user salary, user location and user attendance.

[0071] In the deep learning model fusion layer, the inner product operation and normalization are performed on the hidden vectors of job positions and users to obtain the feature similarity score between job positions and users, represented as: ; in, Indicates time index The similarity score between job positions and user characteristics. , The vector magnitude representing the job characteristics, The vector magnitude representing user characteristics.

[0072] The higher the feature similarity score between a job and a user, the closer the job features are to the user features.

[0073] In the output layer of the deep learning model, the sigmoid activation function is used to convert the feature similarity scores between job postings and users into matching probabilities, expressed as: ; in, Indicates time index The matching probability is used to reflect the likelihood that the job posting will be accepted by users. This represents the natural constant, used to control the smooth curve variation of the Sigmoid function.

[0074] Furthermore, the matching probability of all job records in the job set is calculated, and the job push threshold is dynamically obtained based on the matching probability distribution of all job records, expressed as: ; in, Indicates the threshold for job postings. This represents the average probability of matching all positions in the job set. It represents the standard deviation of the matching probability in the job set.

[0075] The job recommendation threshold is used to distinguish between highly relevant jobs and ordinary jobs. When the job matching probability is greater than or equal to the job recommendation threshold, the job is considered to be highly matched with the user, and the job record is retained.

[0076] All eligible job records are sorted from highest to lowest job matching probability to generate a personalized employment information push sequence.

[0077] S4. Based on the new behavioral events in the personalized employment information push sequence, calculate the prediction bias and convergence threshold of the deep learning model. When the prediction bias of the deep learning model is greater than or equal to the convergence threshold, perform self-learning update of the deep learning model to generate the updated personalized employment information push sequence.

[0078] Furthermore, after the personalized employment information push sequence is generated, new behavioral events of users within the personalized employment information push sequence are monitored.

[0079] The newly added behavioral events include browsing history, favorites history, application history, and interview history. Each newly added behavioral event includes fields for event type, event time, and event object.

[0080] The event object field records the job identifier, which is used to map the newly added behavior event to the job set.

[0081] When a user generates a new behavioral event on a job record in the personalized employment information push sequence, the new behavioral event is appended to the end of the behavioral sequence to maintain the temporal continuity of the behavioral sequence and is updated accordingly with the user identifier.

[0082] In time index Next, job records that generate new behavioral events in the personalized employment information push sequence will be incorporated into the positive sample set, and the same time index will be used to index the job records. The job records that did not generate any new behavioral events were merged into the negative sample set.

[0083] The absolute difference between the average positive sample matching probability and the average negative sample matching probability is calculated to obtain the prediction bias of the deep learning model, which is expressed as: ; in, Indicates time index Prediction bias of deep learning models Indicates time index The following positive sample set, Indicates the number of positive samples. Indicates time index Next The matching probability corresponding to each positive sample job record. Indicates time index The negative sample set below, Indicates the number of negative samples. Indicates time index Next The matching probability corresponding to each negative sample job record.

[0084] By calculating the time index The standard deviation of the matching probability of all job records in the positive sample set and the negative sample set is used to obtain the convergence threshold, which is expressed as: ; ; in, Indicates time index The average matching probability during the learning phase. Indicates time index The convergence threshold below.

[0085] When the prediction bias of the deep learning model is greater than or equal to the convergence threshold, the deep learning model self-learning update is performed.

[0086] When the prediction bias of the deep learning model is less than the convergence threshold, it indicates that the deep learning model has reached a stable learning state. At this time, the existing deep learning model parameters are not updated, and the current personalized employment information push sequence remains unchanged to ensure the stability and convergence of the push results.

[0087] The deep learning parameters include the weight matrix of the job channel, the weight matrix of the user channel, the bias vector of the job channel, and the bias vector of the user channel.

[0088] Furthermore, during the self-learning update process of the deep learning model, gradient descent updates are performed on the weight matrices of the job channel, the weight matrix of the user channel, the bias vector of the job channel, and the bias vector of the user channel using positive and negative samples.

[0089] Gradient descent updates use a learning rate decay method to gradually reduce the step size of deep learning model parameter updates in order to prevent overfitting.

[0090] After the update is completed, the job matching probability and job push threshold are recalculated, and a new personalized employment information push sequence is generated.

[0091] Within the time index where no update is triggered, only new behavioral events are recorded to improve the behavioral sequence; no adjustments are made to the deep learning model parameters, job matching probability, or job push threshold.

[0092] By repeatedly executing deep learning model self-learning updates, job matching probability calculations, job push threshold calculations, and personalized employment information push sequence generation within multiple consecutive time indices, a closed-loop adaptive push optimization mechanism is formed, ensuring that the personalized employment information push sequence continuously conforms to the changing characteristics of user behavior.

[0093] In summary, this invention achieves the fusion expression of job features and user features through the dual-channel nonlinear mapping of a deep learning model, thereby improving the matching accuracy between jobs and users; and dynamically optimizes the job push threshold through prediction deviation calculation and self-learning update, thereby enhancing the real-time performance and accuracy of personalized employment information push.

[0094] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for accurately pushing personalized employment information based on deep learning, characterized in that: include, Collect user information and job information, process them in a structured manner, and generate behavioral sequences; Job selection is performed based on user information and job information, and user baseline information is extracted through behavioral sequences to generate a set of jobs; Feature encoding is performed on the job set and user baseline information to generate job feature vectors and user feature vectors, which are then input into a deep learning model to generate a personalized employment information push sequence. Based on the new behavioral events in the personalized employment information push sequence, the prediction bias and convergence threshold of the deep learning model are calculated. When the prediction bias of the deep learning model is greater than or equal to the convergence threshold, the deep learning model is self-learned and updated to generate an updated personalized employment information push sequence.

2. The method for accurately pushing personalized employment information based on deep learning as described in claim 1, characterized in that: The user information includes resume information and behavioral information; The resume information includes educational background, work experience, professional skills, qualification certificates, desired work location, expected salary and available start date; The behavioral information includes the user's browsing history, favorites history, application history and interview history on the employment platform; The job information includes the job title, job responsibilities, job requirements, salary range, job location, required qualifications, and start date requirements.

3. The method for accurately pushing personalized employment information based on deep learning as described in claim 2, characterized in that: The structured processing includes standardizing field formats, removing redundant items, and filling in missing fields for resume and job information, and establishing a one-to-one correspondence between resume and job information.

4. The method for accurately pushing personalized employment information based on deep learning as described in claim 3, characterized in that: The generation of the behavior sequence includes sorting the browsing history, collection history, submission history and interview history in the behavior information according to time order, and integrating all the behavior records of the same user into a behavior sequence according to the user identifier.

5. The method for accurately pushing personalized employment information based on deep learning as described in claim 4, characterized in that: The job screening based on user information and job information includes matching and screening based on the structured resume information and corresponding fields in the job information. This involves comparing educational background and degree requirements, work experience and years of service, professional skills and required qualifications, qualification certificates and required qualifications, expected work location and job location, expected salary and salary range, and available start time and start time requirements. Job records that meet the matching and screening conditions are retained and summarized to form a pre-screened set of jobs. The pre-screened job list is screened based on job title and job responsibilities. When the job title belongs to the same occupational category as the job category in the work experience section, and the job responsibilities description matches the skill tags in the occupational skills section, the job record is retained, and a job matching set is generated.

6. The method for accurately pushing personalized employment information based on deep learning as described in claim 5, characterized in that: The step of extracting user baseline information through behavior sequences and generating a set of job positions includes analyzing the event type, event time, and event object in the behavior sequences, counting the number of times browsing records, collection records, application records, and interview records appear in each job category, and recording the number of events corresponding to each job category as the behavior frequency field. The job records in the job matching set are sorted according to the value of the behavior frequency field, with job categories with higher behavior frequency field values ​​appearing first. Extract user baseline information by combining resume information and behavioral sequences; The user baseline information includes user location, user salary, user qualifications, and user attendance. Based on the sorted job matching set, a further filtering is performed according to the user's baseline information. If any four or more of the job location, salary range, required qualifications, start date requirements, job responsibilities, and job category match the user's baseline information, the job record is retained. The job records retained after the further filtering are then summarized to form a job set.

7. The method for accurately pushing personalized employment information based on deep learning as described in claim 6, characterized in that: The generation of job feature vectors and user feature vectors includes performing feature encoding on job records and user baseline information in the job set, performing semantic transformation on text fields, and performing normalization transformation on numerical fields. The job feature vector is formed by combining the job title, job responsibilities, job requirements, job location, salary range, required qualifications, and start date requirements in that order. The user feature vector is formed by concatenating the user's location, salary, qualifications, and attendance in that order.

8. The method for accurately pushing personalized employment information based on deep learning as described in claim 7, characterized in that: The process of generating a personalized employment information push sequence by inputting the deep learning model includes inputting the job feature vector and the user feature vector into the job channel and user channel of the deep learning model, respectively, performing nonlinear mapping, and generating the job hidden layer vector and the user hidden layer vector. In the fusion layer, the inner product operation and normalization are performed on the hidden vectors of job positions and users to calculate the feature similarity scores between job positions and users. The feature similarity score is input into the activation function to calculate the matching probability. Based on the average and standard deviation of the matching probabilities in the job set, the job push threshold is calculated. When the job matching probability is greater than or equal to the job push threshold, the job record is retained and sorted from high to low matching probability to generate a personalized employment information push sequence.

9. The method for accurately pushing personalized employment information based on deep learning as described in claim 8, characterized in that: The calculation of prediction bias and convergence threshold of the deep learning model includes: adding job records that generate new behavioral events in the personalized employment information push sequence to the positive sample set, and adding job records that do not generate new behavioral events to the negative sample set. The absolute difference between the average positive sample matching probability and the average negative sample matching probability is calculated to obtain the prediction bias of the deep learning model. Calculate the average matching probability during the learning phase based on the distribution of matching probabilities of all job records in the positive and negative sample sets. Centered on the average matching probability during the learning phase, the variance of the matching probability of job records in the positive and negative sample sets is calculated to obtain the standard deviation of the matching probability, which is used as the convergence threshold. When the prediction bias of a deep learning model is greater than or equal to the convergence threshold, the deep learning model is determined to have not reached convergence.

10. The method for accurately pushing personalized employment information based on deep learning as described in claim 9, characterized in that: The generated updated personalized employment information push sequence includes, when the deep learning model has not reached convergence, performing gradient descent updates on the weight matrix of the job channel, the weight matrix of the user channel, the bias vector of the job channel, and the bias vector of the user channel based on the positive sample set and the negative sample set. After the update is completed, the job matching probability and job push threshold are recalculated, and an updated personalized employment information push sequence is generated.

Images