Job application recommendation method based on multi-modal depth semantics and dynamic intention modeling

Abstract

The invention provides a job recommendation method based on multi-modal depth semantics and dynamic intention modeling, and relates to the technical field of Internet, the method comprises the following steps: collecting and preprocessing multi-modal data to obtain preprocessed multi-modal data, the multi-modal data comprises user resume data, user behavior sequence data, post description text data, company portrait data, post popularity data and industry trend data; based on the user resume data and the user behavior sequence data, modeling a user dynamic intention to obtain a user intention vector; and based on the post description text data, the company portrait data, the post popularity data and the industry trend data, performing deep semantic representation on posts to obtain post vectors. According to the invention, the accuracy of user intention recognition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to a job recommendation method based on multi-modal deep semantic and dynamic intention modeling. BACKGROUND

[0002] According to the 2024 Internet recruitment industry report, the user daily average number of browsed positions is more than 50, but the effective matching (HR initiates communication after the user delivers) rate is insufficient; under the traditional keyword search and list browsing mode, the user needs to manually filter more than 30 positions to find one opportunity that meets the basic expectations, and the single effective position filtering time is averagely up to 45 minutes, and the job seeking efficiency is seriously low. There are two major problems in traditional recommendation algorithms: cold start problem: the matching accuracy of new users or new positions is low, and the algorithm relying on collaborative filtering cannot handle the scene without similar behaviors; insufficient semantic understanding: only through keyword matching or shallow text features (such as position title segmentation), it is impossible to identify the implicit needs of the position (such as the implicit team management experience of independently responsible projects) and the potential ability of the user (such as the implicit high concurrency processing ability of participating in e-commerce system development), resulting in a high rate of mismatch between people and positions.

[0003] The existing system is based on single time granularity behavior to infer user intention, and cannot distinguish between short-term urgent needs (such as high-frequency clicking of urgent position by users who quit their jobs within 7 days), medium-term preferences (such as collecting technical positions in large factories within 30 days), and long-term career planning (such as focusing on technical management positions within 3 months). When the intention drifts, the intention recognition accuracy is low, and the recommendation precision drops sharply. Only relying on the text of the position description for representation ignores the value of cross-modal data: company portrait; position popularity; industry trends, resulting in a representation error of more than 45% for newly released positions. There is no effective feedback after the user delivers, and most users cannot know the reason why the resume is not passed; the recommended results have no explanation basis, and the user cannot judge why the position is recommended, resulting in low acceptance of the recommendation, and the monthly retention rate of the platform is insufficient; the user interaction data is not fed back to the model in real time, and the model iteration cycle is as long as 15 days, and the recommendation effect cannot be dynamically optimized. These problems lead to long job seeking cycle, low user satisfaction, and low platform conversion rate, etc. industry pain points. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a job recommendation method based on multi-modal deep semantic and dynamic intention modeling, which improves the accuracy of user intention recognition.

[0005] To solve the above technical problems, the technical solution of the present application is as follows: In a first aspect, a job recommendation method based on multi-modal deep semantic and dynamic intention modeling, the method comprising: The multi-modal data is collected and preprocessed to obtain preprocessed multi-modal data, and the multi-modal data includes user resume data, user behavior sequence data, job description text data, company portrait data, job heat data and industry trend data. Based on the user resume data and the user behavior sequence data, the user dynamic intention is modeled to obtain a user intention vector; and based on the job description text data, the company portrait data, the job heat data and the industry trend data, the job is deeply semantically represented to obtain a job vector. According to the user intention vector and the job vector, candidate jobs are recalled to obtain recalled candidate jobs; according to the recalled candidate jobs, a user intention alignment score is calculated, a second-stage precision sorting is performed on the candidate jobs sorted in the first stage to calculate a competitiveness prediction score, and finally a recommended job list is obtained based on a comprehensive ranking score; Based on the interaction feedback data of the user on the recommended job list, the user dynamic intention modeling and the job semantic representation model are incrementally updated.

[0006] Further, the multi-modal data is collected and preprocessed to obtain preprocessed multi-modal data, and the multi-modal data includes user resume data, user behavior sequence data, job description text data, company portrait data, job heat data and industry trend data, including: The original multi-modal data is collected from the recruitment platform database, and the original multi-modal data includes user resume data, user behavior sequence data, job description text data, company portrait data, job heat data and industry trend data; The collected original multi-modal data is cleaned to obtain cleaned data, and the cleaned data is standardized to obtain standardized data; The standardized data is feature extracted to obtain preprocessed multi-modal data.

[0007] Further, based on the user resume data and the user behavior sequence data, the user dynamic intention is modeled to obtain a user intention vector, including: The user resume data is extracted from the preprocessed multi-modal data as a long-term portrait, and the user behavior sequence data is extracted, and the user behavior sequence data is split into short-term behavior sequence and medium-term behavior sequence according to a preset time window; The short-term behavior sequence is time-encoded by a Transformer encoder to obtain a short-term intention vector, the medium-term behavior sequence is time-encoded by a GRU network to obtain a medium-term intention vector, and the long-term portrait is feature mapped by a full connection layer to obtain a long-term intention vector; Based on the short-term intention vector, the medium-term intention vector and the long-term intention vector, dynamic weight calculation and weighted fusion are performed by a multi-granularity attention layer to obtain an initial user intention vector; Intent drift is detected by comparing the user's real-time behavior data with the initial user intent vector. When intent drift is detected, the weights of the multi-granularity attention layer are recalculated and adjusted to obtain the final user intent vector.

[0008] Furthermore, based on job description text data, company profile data, job popularity data, and industry trend data, deep semantic representation of jobs is performed to obtain job vectors, including: The job description text data is semantically encoded using a BERT pre-trained model to obtain text semantic vectors, and the company profile data is feature encoded, including embedding discrete features and normalizing continuous features to obtain company vectors. The job popularity data is time-series encoded, and the popularity change trend is captured by an LSTM network to obtain a popularity vector. The industry trend data is standardized and feature mapping is performed through a fully connected layer to obtain a trend vector. Using text semantic vectors as query vectors and company vectors, popularity vectors, and trend vectors as key vectors, cross-modal semantic alignment is performed through a cross-attention mechanism to obtain aligned text semantic vectors, company vectors, popularity vectors, and trend vectors. The aligned text semantic vector, company vector, popularity vector, and trend vector are concatenated, then fused through residual connections and fully connected layers, and finally processed by Dropout to obtain the final job vector.

[0009] Furthermore, based on the user intent vector and job vector, candidate jobs are recalled to obtain recalled candidate jobs; based on the recalled candidate jobs, a user intent-job alignment score is calculated, including: Based on user intent vectors and job vectors, a vector retrieval tool is used to calculate similarity and quickly retrieve a set of candidate jobs from the job database. Based on the retrieved candidate job list, rules are applied to filter the candidate job list to exclude industries that the user has blocked or jobs with salaries lower than expected, resulting in a filtered candidate job list. Based on the filtered candidate job list, for each job in the filtered candidate job list, the semantic similarity between the user intent vector and the job vector is calculated, and keyword matching features between the user and the job are extracted. Based on semantic similarity and keyword matching features, a weighted fusion method is used to calculate the alignment score between user intent and job position.

[0010] Furthermore, the candidate positions ranked in the first stage undergo a second stage of ranking to calculate the predicted competitiveness score. Finally, a recommended list of positions is obtained based on the comprehensive ranking score, including: Based on the candidate positions with qualified alignment scores, user competitiveness features and position competitiveness features are extracted. Based on the user competitiveness features and position competitiveness features, the user competitiveness score and position competitiveness score are obtained by weighted calculation. The user's competitiveness score and the job competitiveness score are weighted and combined to obtain the predicted competitiveness score; The score for alignment between intent and job position is weighted and combined with the score for predicted competitiveness to calculate the overall ranking score for each candidate job position. All candidate positions are ranked based on their comprehensive ranking scores to obtain the final list of recommended positions.

[0011] Furthermore, based on user interaction feedback data regarding the recommended job list, the user dynamic intent model and job semantic representation model are incrementally updated, including: Collect user interaction feedback data on the recommended job list. The interaction feedback data includes clicks, favorites, applications, and ignores. Based on the interaction feedback data, obtain training data with positive and negative sample labels. Incremental fine-tuning of the multi-granularity temporal intent encoder based on training data updates the parameters of user dynamic intent modeling. Incremental fine-tuning of the cross-modal job semantic enhancement model based on training data updates the parameters of job semantic representation.

[0012] Secondly, a job recommendation system based on multimodal deep semantics and dynamic intent modeling includes: The acquisition module is used to collect and preprocess multimodal data to obtain preprocessed multimodal data, which includes user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data. Based on user resume data and user behavior sequence data, the module models user dynamic intent to obtain user intent vectors. Based on job description text data, company profile data, job popularity data, and industry trend data, the module performs deep semantic representation of jobs to obtain job vectors. The processing module is used to recall candidate positions based on user intent vectors and job vectors to obtain the recalled candidate positions; based on the recalled candidate positions, it calculates the alignment score between user intent and job position; and performs a second-stage fine ranking of the candidate positions ranked in the first stage to calculate the competitiveness prediction score. Finally, it obtains a recommended job list based on the comprehensive ranking score; and incrementally updates the user dynamic intent model and job semantic representation model based on user interaction feedback data of the recommended job list.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: By employing multimodal data collection and preprocessing of user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data, combined with multi-granular temporal intent modeling, deep semantic representation of jobs, and two-stage fine ranking, this approach recalls candidate jobs based on vector similarity and calculates intent alignment. It also combines user competitiveness and job competitiveness to calculate a competitiveness prediction score and a feedback closed-loop incremental update model. This overcomes the limitations of traditional job recommendation methods, such as static matching of user needs, inability to dynamically track changes in user intent, one-sided semantic representation of jobs, lack of consideration for user and job competitiveness leading to low recommendation feasibility, and lack of continuous optimization mechanisms. This approach achieves accurate characterization of users' dynamic job search intent and comprehensive job features, improves the adaptability and practical feasibility of recommended jobs, and continuously adapts to user needs and market dynamics through feedback iteration. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a job recommendation method based on multimodal deep semantics and dynamic intent modeling, provided by an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a job recommendation system based on multimodal deep semantics and dynamic intent modeling, provided by an embodiment of the present invention.

[0018] Figure 3 This is a flowchart of the job recommendation algorithm.

[0019] Figure 4 A flowchart illustrating the user's interaction process from browsing to delivery.

[0020] Figure 5 The user interaction flowchart from the access sequence. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] like Figure 1 As shown, embodiments of the present invention propose a job recommendation method based on multimodal deep semantics and dynamic intent modeling, the method comprising the following steps: Step 1: Collect multimodal data and preprocess it to obtain preprocessed multimodal data. The multimodal data includes user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data. Step 2: Based on user resume data and user behavior sequence data, model the user's dynamic intent to obtain the user intent vector; based on job description text data, company profile data, job popularity data and industry trend data, perform deep semantic representation of the job to obtain the job vector. Step 3: Based on the user intent vector and job vector, recall candidate jobs to obtain the recalled candidate jobs; based on the recalled candidate jobs, calculate the user intent and job alignment score, and conduct a second stage of fine ranking of the candidate jobs ranked in the first stage to calculate the competitiveness prediction score. Finally, obtain the recommended job list based on the comprehensive ranking score. Step 4: Based on user interaction feedback data of the recommended job list, incrementally update the user dynamic intent model and the job semantic representation model.

[0023] In this embodiment of the invention, by employing multimodal data acquisition and preprocessing, a multi-granular temporal intent encoder to model user dynamic intent, and a cross-modal job semantic enhancement model to achieve deep semantic representation of jobs, and a two-stage dynamic fine ranking framework to integrate candidate job retrieval and two-stage fine ranking with incremental model updates based on user interaction feedback, this invention overcomes the problems of low job search efficiency due to information overload, poor matching accuracy, single user intent modeling, job representation relying solely on text dimensions, slow model iteration due to lack of feedback loop, and low user trust in existing job recommendation technologies. As a result, it achieves improvements in user intent recognition accuracy from 58% to 83%, new user / new job matching accuracy from 28% to 68%, recommended job CTR from 12% to 15.6%, HR response rate after application from 35% to 49%, model iteration cycle from 15 days to 7 days, platform monthly retention rate from 35% to 60%, and average user job search cycle from 60 days to 45 days. Simultaneously, multi-dimensional optimization enhances user recommendation trust and the intelligent auxiliary value of the entire job search process.

[0024] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Collect raw multimodal data from the recruitment platform database. This raw multimodal data includes user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data. Specifically, this involves systematically collecting multi-dimensional raw multimodal data from the recruitment platform database to ensure that the data covers key dimensions missing in traditional technologies. The specific collection content includes: collecting user resume data, covering user skill tags, years of work experience, educational background, past project experience, expected salary, target work city, professional qualification certificates, etc. This information is used to identify users' potential abilities and compensate for the shortcomings of traditional technologies in uncovering users' hidden strengths; collecting user behavior sequence data, recording user behaviors such as clicking on jobs, saving jobs, applying for jobs, and the duration of time spent browsing jobs on the platform, while also associating the timestamps corresponding to each behavior to provide time for distinguishing between short-term urgent needs, medium-term preferences, and long-term career plans. Dimensional data support addresses the limitations of traditional intent modeling's single-time granularity; it collects job description text data to fully acquire textual information such as job responsibilities, skill requirements, experience requirements, job content, team composition, and promotion paths, providing a foundation for deep semantic understanding and overcoming the limitations of traditional methods that rely solely on shallow textual features; it collects company profile data, including information such as the company's industry, size, salary range, financing stage, corporate culture, office location, and employee reviews, supplementing the missing company attribute dimensions in traditional job representation; it collects job popularity data, statistically analyzing dynamic data such as the number of applicants in the past seven days after a job posting, the HR response rate within 24 hours, the number of days the job was posted, and the ratio of applicants to available positions, addressing the problem of traditional technologies ignoring job competition; and it collects industry trend data, obtaining data on the annual demand growth rate of each job's field, the number of talent supply gaps, the impact of industry policies, and the impact of technological development on job demand.

[0025] Step 1.2 involves cleaning the collected raw multimodal data to obtain cleaned data. The cleaned data is then standardized to obtain standardized data. Specifically, this includes: First, handling missing values ​​during data cleaning. This involves distinguishing between continuous and discrete features. Continuous features include numerical features that are continuously changing, such as years of work experience and salary range. Discrete features include categorical features, such as education level and industry type. For missing values ​​in continuous features, the KNN imputation method is used. For each sample with a missing value, the dataset is searched for the multiple samples most similar to that feature, and the average value of these similar samples for that feature is calculated. This average value is then used to imput the missing value. For missing values ​​in discrete features, the frequency of all non-missing values ​​for that feature is counted, and the most frequent value is selected as the mode. This mode is then used to imput the missing value. Next, outlier detection is performed. For numerical features, the IQR criterion is used. The upper and lower quartiles of the feature data are calculated, and the difference between them is the interquartile range. The upper quartile is then increased by 1.5. The upper limit is set at 1 / 4 interquartile range, and the lower limit is set at 1.5 times the lower interquartile range. Values ​​exceeding these limits are considered outliers and removed from the data. For text data, invalid characters are filtered according to preset rules, including identifying and removing special symbols, garbled characters, invisible characters, and other meaningless content. Only normal text, numbers, and necessary punctuation are retained, ensuring the text information is valid. After data cleaning, format standardization is performed. For date data, regardless of whether the original format uses slashes, periods, or other symbols, it is uniformly converted to year, month, and day format, with the year represented by four digits and the month and day each represented by two digits. Dates of all formats are uniformly adjusted to a standard format similar to October 20, 2025. For salary data, it is uniformly converted to a monthly salary range in yuan. If the original data is an annual salary, it is divided by twelve to convert to a monthly salary. If it is a daily or hourly salary, it is converted to a monthly salary based on the number of working days or hours per month. The salary range is clearly marked, and the unit is uniformly set to yuan. All salary data is presented consistently. After processing, standardized data is obtained.

[0026] Step 1.3 involves feature extraction from standardized data to obtain preprocessed multimodal data. Specifically, feature extraction of standardized data requires processing according to data type. For textual data, including job descriptions and position descriptions in resumes, the TF-IDF algorithm is used to extract keyword features. First, the text is segmented into words, breaking down continuous text into independent words. Then, the frequency of each word in a single text is counted (term frequency). Next, the number of documents in which each word appears across all texts is calculated, yielding the inverse document frequency (IVF). The IVF is inversely proportional to the prevalence of a word in a document; the more prevalent a word is, the lower its IVF. The term frequency is multiplied by the IVF to obtain the TF-IDF value for each word. A higher value indicates greater importance of the word in the current text. Words with higher TF-IDF scores are selected as keyword features for textual data. For numerical data, including work experience, job title, and salary, the Z-index is used... The score standardization algorithm converts data into zero-mean, unit-variance features. First, the mean of the feature across all samples is calculated, then the standard deviation is calculated. Next, for each sample's value, the mean is subtracted, and the difference is divided by the standard deviation. This process ensures the transformed feature has a mean of zero and a standard deviation of one, resulting in a zero-mean, unit-variance feature. For time-series data, including user behavior sequences and job popularity data, a sliding window approach is used to extract time-series features. The sliding window's time range is set, for example, 30 days per window. The windows slide sequentially, each covering 30 consecutive days of data. Within each window, relevant time-series indicators are statistically analyzed, such as the number of times users clicked on a job posting, the duration of browsing a job posting, or the frequency of changes in job popularity within 30 days. These statistical results are used as features for the time-series data. After feature extraction processing of textual, numerical, and time-series data, preprocessed multimodal data is obtained.

[0027] In this embodiment of the invention, the use of raw multimodal data collected from a recruitment platform database, including user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data, followed by cleaning the raw data to remove invalid / redundant information, standardizing the format to unify the data format, and finally extracting features from the standardized data, overcomes the limitations of traditional job recommendation technologies, such as single data sources, poor raw data quality, inconsistent data formats leading to inefficient modeling, and insufficient data feature extraction failing to provide effective input for accurate recommendations. This results in high-quality, multi-dimensional, and standardized feature data for subsequent multi-granularity temporal intent encoder user dynamic intent modeling and cross-modal job semantic enhancement model job deep semantic representation.

[0028] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Extract user resume data from the preprocessed multimodal data to create a long-term user profile, and extract user behavior sequence data. This user behavior sequence data is then divided into short-term and medium-term behavior sequences according to a preset time window. Specifically, this includes: extracting user resume data from the preprocessed multimodal data to create a long-term user profile. The extracted content includes fixed attribute information such as the user's core skills, years of work experience, educational background, past project experience, expected salary range, target work city, and professional qualification certificates. This information reflects the user's long-term stable career foundation and career planning direction. Simultaneously, user behavior sequence data is extracted, covering the user's job click records, job collection records, job application records, and the duration of browsing each job on the platform. Each behavior record is associated with a corresponding timestamp. Then, the user behavior sequence data is divided according to a preset time window. User behavior sequences within the last 7 days are classified as short-term behavior sequences to capture users' short-term urgent needs, and user behavior sequences within the last 30 days are classified as medium-term behavior sequences to capture users' medium-term job-seeking preferences. This division method solves the problem that traditional systems, relying solely on a single time granularity, cannot distinguish job-seeking needs across different periods.

[0029] Step 2.2 involves temporally encoding the short-term behavior sequence using a Transformer encoder to obtain a short-term intent vector, temporally encoding the medium-term behavior sequence using a GRU network to obtain a medium-term intent vector, and performing feature mapping on the long-term profile using a fully connected layer to obtain a long-term intent vector. Specifically, this includes temporally encoding the split short-term behavior sequence using a Transformer encoder. Because there are obvious temporal correlations in the short-term behavior sequence—for example, if a user clicks on multiple urgent job postings consecutively within the past 7 days—the Transformer encoder can effectively capture the underlying logic of this continuous behavior, transforming the short-term behavior sequence into a short-term intent vector that reflects the user's short-term needs. The mid-term behavioral sequences are encoded temporally using a GRU network. These sequences primarily reflect users' relatively stable job preferences, such as users consistently saving technical positions at major companies within the past 30 days. The GRU network can process this stable sequence data, uncovering patterns in preferences and outputting a mid-term intent vector that reflects the user's mid-term preferences. The extracted long-term user profiles are mapped using fully connected layers. These profiles contain both discrete and continuous features. Discrete features include education level and types of professional qualifications, while continuous features include years of work experience and expected salary. The fully connected layers can unify these different types of features into a structured feature vector, ultimately outputting a long-term intent vector that reflects the user's long-term career planning.

[0030] Step 2.3: Based on short-term, medium-term, and long-term intent vectors, dynamic weight calculation and weighted fusion are performed through a multi-granularity attention layer to obtain the initial user intent vector. Specifically, this includes: In the multi-granularity attention layer, firstly, analyzing the user's recent behavioral activity and statistically analyzing the frequency of user behavior in short-term and medium-term behavioral sequences. For example, users who quit their jobs have a much higher frequency of job clicks and applications in the past 7 days than in the past 30 days, and employed users who are changing jobs have a more prominent focus on long-term planning related behaviors for technical management positions in the past 3 months. Dynamic weights are calculated for each intent vector based on behavioral frequency. For users with high short-term behavior frequency, the weight of the short-term intent vector is increased; for users with obvious long-term career planning characteristics, the weight of the long-term intent vector is increased, ensuring that the weight allocation aligns with the user's current job-seeking status. The short-term, medium-term, and long-term intent vectors are then weighted and fused according to the calculated dynamic weights. By superimposing vectors, different dimensions of intent information are integrated to ultimately generate an initial user intent vector that comprehensively reflects the user's short-, medium-, and long-term job-seeking needs, solving the problem that traditional intent modeling is singular and cannot fully cover user needs.

[0031] Step 2.4: Detect intent drift based on the difference between real-time user behavior data and the initial user intent vector. When intent drift is detected, the weights of the multi-granularity attention layer are recalculated and adjusted to obtain the final user intent vector. This specifically includes: collecting the latest user behavior data in real-time, including information such as the job types clicked by the user in real-time, the job attributes saved in real-time, and the job requirements applied for in real-time; comparing the job characteristics corresponding to these real-time behavior data with the job characteristics associated with the initial user intent vector to determine the degree of difference. For example, if the initial user intent vector corresponds to the characteristics of a Python development job, while the user continuously clicks on... When a user clicks on a Java development job posting three or more times, it is considered an intent drift. Upon detecting intent drift, the weights of the multi-granularity attention layer are recalculated. At this point, the weight of the short-term intent vector is increased because real-time behavior falls under the category of short-term behavior and can more accurately reflect the user's current intent after drift. Based on the recalculated weights, the short-term, medium-term, and long-term intent vectors are then weighted and fused again. The adjusted vector can better match the user's job search needs after drifting, ultimately obtaining a final user intent vector that accurately reflects the user's current true intent. This solves the problems of low recognition accuracy and sharp drop in recommendation precision when intent drift occurs in traditional systems.

[0032] In this embodiment of the invention, user resume data is extracted from preprocessed multimodal data to form a long-term profile. Simultaneously, user behavior sequence data is extracted and split into short-term and medium-term behavior sequences according to a preset time window. The short-term behavior sequence is temporally encoded or feature-mapped using a Transformer encoder, the medium-term behavior sequence using a GRU network, and the long-term profile using a fully connected layer to obtain short-term, medium-term, and long-term intent vectors. Then, a multi-granularity attention layer is used to dynamically calculate and weight the three types of intent vectors to obtain an initial user intent vector. Finally, intent drift is detected based on the difference between real-time user behavior data and the initial user intent vector, and the weights of the multi-granularity attention layer are readjusted. Therefore, this invention overcomes the technical problems of existing job recommendation technologies where user intent modeling relies solely on a single time-granularity behavior, failing to distinguish between short-term urgent needs, medium-term preferences, and long-term career plans, and where intent recognition accuracy is below 60% during intent drift, making it difficult to accurately capture dynamic user job-seeking intentions. This improves the accuracy of user intent recognition, effectively enhances the initial intent matching effect for new users, and achieves accurate differentiation and dynamic tracking of users' short, medium, and long-term job-seeking needs, providing high-quality intent evidence for accurate matching of job positions and users.

[0033] In a preferred embodiment of the present invention, step 2 above may include: Step 2.5 involves semantically encoding the job description text data using a BERT pre-trained model to obtain a text semantic vector, and performing feature encoding on the company profile data, including embedding discrete features and normalizing continuous features to obtain a company vector. Specifically, for the job description text data, the complete job description content is input into the BERT pre-trained model. This model, through adaptive learning of recruitment texts, can deeply mine the implicit requirements in the job description. For example, it can identify the implicit team management experience requirement from a description requiring independent responsibility for the entire project development process, thereby transforming the job description text into text semantics that accurately reflects the semantic information of the job. Vector technology addresses the problem that traditional techniques, relying solely on shallow textual features, cannot deeply understand job requirements. For company profile data, it first distinguishes between discrete and continuous features. Discrete features include the company's industry, funding stage, and type of enterprise; these features are embedded to transform non-numerical categorical information into a computable vector form. Continuous features include company size, salary range, and average annual revenue; these features are normalized to eliminate the impact of different data scales on subsequent calculations. Finally, the processed discrete and continuous feature vectors are integrated to obtain a company vector that comprehensively reflects the company's attributes, overcoming the deficiency of traditional job representations that ignore the company profile dimension.

[0034] Step 2.6 involves temporal encoding of job posting popularity data, using an LSTM network to capture trends in popularity and obtain a popularity vector. Industry trend data is standardized, and feature mapping is performed through a fully connected layer to obtain a trend vector. Specifically, this includes collecting real-time dynamic data from the past 7 days after a job posting, and inputting this chronologically ordered popularity data into an LSTM network. The LSTM network can capture trends in popularity data over time, such as identifying patterns like a surge in applications followed by stabilization in the first 3 days after a job posting, or an anomaly like a continuous decline in HR response rates. This allows for the dynamic encoding of job posting popularity data. This approach transforms job information into a trend vector, addressing the problem of traditional technologies focusing only on static job information and ignoring the dynamic value of job popularity. For industry trend data, it collects data such as the annual demand growth rate of the job's field, the number of talent supply gaps, and the impact of industry technology development on job skills. This data is first standardized to unify the value range, and then input into a fully connected layer. Through the feature mapping function of the fully connected layer, the industry trend information is transformed into a trend vector that reflects the future development potential of the job, avoiding the problem of traditional technologies failing to prioritize high-potential jobs due to ignoring industry trends.

[0035] Step 2.7: Using the text semantic vector as the query vector and the company vector, popularity vector, and trend vector as key-value vectors, cross-modal semantic alignment is performed through a cross-attention mechanism to obtain the aligned text semantic vector, company vector, popularity vector, and trend vector. Specifically, this includes: using the text semantic vector as the query vector, which carries the core semantic requirements information of the job; using the company vector, popularity vector, and trend vector as key-value vectors, which respectively carry information on company attributes, job competition dynamics, and industry development potential; and using the cross-attention mechanism to calculate the correlation between the query vector and each key-value vector, for example, the text semantic vector... The semantic information of flat management in the semantic vector is associated with the attribute information of 100-500 employees in the company vector. The demand for urgently recruiting core developers in the text semantic vector is associated with the dynamic information of more than 200 people applying in the past 3 days in the popularity vector. At the same time, the requirement to master AI training skills in the text semantic vector is associated with the industry information of 120% annual growth in demand for AI in the trend vector. Through this association calculation, the text semantics and cross-modal features are accurately aligned, avoiding the problem of multimodal data existing in isolation and unable to play a synergistic role in traditional technologies, so that the information of each dimension of the position forms an organic whole.

[0036] Step 2.8 involves concatenating the aligned text semantic vector, company vector, popularity vector, and trend vector, then fusing them through residual connections and fully connected layers, and applying Dropout to obtain the final job vector. Specifically, this involves: first, concatenating the aligned text semantic vector, company vector, popularity vector, and trend vector according to feature dimensions, integrating four core information categories: job semantic requirements, company attributes, popularity dynamics, and industry trends, forming a more comprehensive initial fused vector; then, inputting the initial fused vector into a fully connected layer with residual connections. Residual connections effectively alleviate the gradient vanishing problem in deep network training, ensuring that key information is not lost during the fusion process. The fully connected layer further integrates the concatenated multi-dimensional features through non-linear transformations, forming a more representative job feature vector; finally, applying Dropout to the integrated feature vector randomly masks some feature nodes, preventing the model from over-relying on a single feature dimension and causing overfitting. The final output is a job vector that comprehensively and accurately represents job information, solving the problems of missing dimensions in traditional technical job representation and over 45% error in newly released job representation, laying the foundation for accurate matching of user intent.

[0037] In this embodiment of the invention, semantic vectors are obtained by semantically encoding job description text data using a BERT pre-trained model; company profile data undergoes embedding processing on discrete features and normalization processing on continuous features to obtain company vectors; job popularity data is temporally encoded using an LSTM network to capture popularity trends and obtain popularity vectors; industry trend data is standardized and then feature-mapped through a fully connected layer to obtain trend vectors; cross-modal semantic alignment is achieved using the text semantic vector as the query vector, the company vector, the popularity vector, and the trend vector as key-value vectors through a cross-attention mechanism; and finally, the aligned vectors are... The technique of concatenating class vectors and obtaining the final job vector through residual connections, fully connected layer fusion, and Dropout processing overcomes the technical problems of existing job recommendation technologies, which rely solely on job description text for job representation, ignore cross-modal data such as company profiles, job popularity, and industry trends, and suffer from shallow semantic understanding that makes it impossible to accurately uncover implicit job requirements, resulting in a job representation error of over 45%. This technique expands the job representation dimension from a single textual modality to four cross-modal types, reduces the new job matching error from 45% to 10%, improves the correlation with industry trends by 60%, and more accurately identifies implicit job requirements.

[0038] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Based on the user intent vector and job vector, a vector retrieval tool is used to calculate similarity and quickly retrieve a set of candidate jobs from the job database. Specifically, this involves: first, calling the vector retrieval tool, using the user intent vector and the job vectors of all jobs in the database as input data; then, calculating the similarity between the user intent vector and each job vector using a similarity calculation method that accurately measures the degree of vector association to ensure that jobs initially related to the user's intent are selected. Considering the inefficiency of users manually filtering a large number of jobs in traditional keyword search and list browsing modes, the vector retrieval tool needs to have rapid retrieval capabilities, able to quickly filter out the top-ranked jobs with high similarity from a job database containing over 500,000 jobs, obtaining a set of candidate jobs. Typically, around 100 jobs initially matching the user's intent are selected, significantly reducing the filtering scope and solving the problem of low job search efficiency caused by information overload.

[0039] Step 3.2: Based on the retrieved candidate job set, apply rule filtering to the candidate job set to exclude industries blocked by the user or jobs with salaries lower than expected, resulting in a filtered candidate job list. Specifically, this includes: first, obtaining the personalized filtering conditions pre-set by the user within the platform, mainly including industries blocked by the user and the user's expected minimum salary; for the candidate job set, checking the industry attributes and salary level of each candidate job one by one; if a candidate job belongs to an industry within the user's blocked industry list, the job is directly excluded from the candidate job set; if the minimum salary range of a candidate job is lower than the user's expected minimum salary, the job is also excluded. Through these two types of rule filtering, jobs that clearly do not meet the user's basic expectations are removed, preventing the user from encountering irrelevant jobs, and obtaining a filtered candidate job list, further narrowing the job selection range, reducing the time cost for users to manually identify invalid jobs, and solving the problem in the traditional model where users need to filter a large number of jobs to find opportunities that meet their basic expectations.

[0040] Step 3.3: Based on the filtered candidate job list, for each job in the filtered candidate job list, calculate the semantic similarity between the user intent vector and the job vector, and extract keyword matching features between the user and the job. Specifically, for each job in the filtered candidate job list, perform two core calculation and extraction tasks. The first is to calculate the semantic similarity by performing an association analysis between the job vector and the user intent vector. Through calculation methods that can capture deep semantic associations, explore the degree of matching between the two at the level of implicit needs and potential capabilities. For example, identify the matching relationship between the user's potential high-concurrency processing capabilities in e-commerce system development and the implicit needs of the job to deal with high-concurrency business scenarios, and obtain the semantic similarity result between the job and the user intent. The second step is to extract keyword matching features. First, keywords such as core skills, work experience, and education requirements are extracted from the user's long-term profile. Then, corresponding job requirement keywords are extracted from the job description text. The two types of keywords are compared one by one, and the number of matching keywords is counted. At the same time, core keywords and auxiliary keywords are distinguished to clarify the matching situation of different types of keywords. This forms the keyword matching features between the job and the user, solving the problem that traditional algorithms rely only on shallow keyword matching or text features, resulting in insufficient semantic understanding.

[0041] Step 3.4: Based on semantic similarity and keyword matching features, a weighted fusion calculation of the user intent-job alignment score is performed. Specifically, this involves: first, determining the weight allocation of semantic similarity and keyword matching features in the alignment score calculation. Considering that semantic similarity reflects the depth of matching and keyword matching features are the basic matching basis, the weight of semantic similarity is set to a higher proportion, and the weight of keyword matching features is set to a lower proportion. The weight of semantic similarity is 0.6, and the weight of keyword matching features is 0.4. Next, the semantic similarity results and keyword matching features are converted into quantitative scores of 0-100 points respectively. The semantic similarity score is converted proportionally based on the previously calculated similarity results, and the keyword matching feature score is converted comprehensively based on the number and type of matched keywords. Finally, the two types of quantitative scores are weighted and summed according to the set weights to obtain the intent-job alignment score for each candidate job. This score comprehensively reflects the degree of matching between the job and the user's intent, providing a quantitative basis for further screening of precise jobs and reducing the problem of a job mismatch rate exceeding 40% in traditional algorithms.

[0042] In this embodiment of the invention, by employing a method based on user intent vectors and job vectors, calculating similarity using vector retrieval tools to quickly retrieve a set of candidate jobs from the job database, applying rules to filter and exclude jobs in industries blocked by the user and jobs with salaries lower than expected, calculating the semantic similarity between the user intent vector and job vector for each filtered candidate job, extracting keyword matching features between the user and the job, and finally calculating the alignment score between user intent and job through weighted fusion, this invention overcomes the efficiency problems of traditional job recommendation, such as information overload leading to time-consuming manual job screening and requiring screening more than 30 jobs to find one that meets basic expectations, as well as the technical problems of insufficient semantic understanding and mismatched matching accuracy due to relying solely on shallow keyword matching or list browsing. This achieves the goal of quickly filtering candidate jobs that initially match user needs from a massive job database, reducing interference from invalid jobs, shortening user job screening time, and improving the initial alignment accuracy between intent and job by combining semantic similarity and keyword matching features, laying a high-quality candidate foundation for the second stage of fine ranking and helping to improve the effective matching rate of recommended jobs.

[0043] In a preferred embodiment of the present invention, step 3 above may include: Step 3.5: Based on the candidate positions with qualified alignment scores, extract user competitiveness features and position competitiveness features. Based on these features, calculate the user competitiveness score and position competitiveness score using weighted averages. Specifically, when extracting user competitiveness features, focus on the match between the implicit requirements of the position and the user's potential abilities. Extract three core features: First, skill gap features: compare the core skills required for the position (e.g., Python, TensorFlow) with the skills the user already possesses, count the number of missing core skills, and distinguish between core skills (required for the position) and auxiliary skills (preferred for the position). Missing core skills have a greater impact on competitiveness. Second, experience matching features: compare the user's actual years of work experience with the required years for the position. If the user's years of experience exceed the position requirement by 3 years or more, they are considered to have sufficient experience; otherwise, the matching degree is calculated proportionally. Third, expected matching features: compare the user's expected salary with the salary range of the position. If the user's expected salary is within the range, they are considered a perfect match; if it is below the lower limit or above the upper limit, the matching degree is calculated based on the degree of deviation. When extracting features related to job competitiveness, two core features are extracted based on job popularity data and actual competition: First, the competition coefficient, calculated as the ratio of applicants to available positions in the past 7 days; a higher ratio indicates more intense competition. Second, the HR response rate, calculated as the percentage of applicants whose applications were responded to within 24 hours in the past 7 days; a lower response rate suggests potentially lower efficiency in subsequent communication. When calculating a user's competitiveness score, weights are assigned based on feature importance, with skill gap features having the highest weight (e.g., 40%), experience matching features second (e.g., 30%), and expected matching features the lowest weight (e.g., 30%). Each feature is converted into a quantifiable value of 0-100 based on its matching degree, and then summed according to weights to obtain the user's competitiveness score. When calculating the job competitiveness score, the competition coefficient and HR response rate each have a 50% weight. The competition coefficient is converted into a quantifiable value of 0-100 based on the rule that a lower ratio results in a higher score, and the HR response rate is converted into a quantifiable value of 0-100 based on weights, and then summed according to weights to obtain the job competitiveness score.

[0044] Step 3.6 involves weighted fusion of the user's competitiveness score and the job's competitiveness score to obtain a predicted competitiveness score. Specifically, this involves clarifying the equal importance of the user's competitiveness score and the job's competitiveness score in competitiveness prediction. User competitiveness determines whether a user is competent for the job, while job competitiveness determines whether a user can stand out in the competition. Both are crucial for improving the HR response rate after application (traditionally only 35%). The user's competitiveness score and the job's competitiveness score are each set within the 0-100 range, and then weighted and fused in a 5:5 ratio: first, the user's competitiveness score is multiplied by 0.5, then the job's competitiveness score is multiplied by 0.5, and finally, the two results are added together to obtain the predicted competitiveness score for each candidate job. This score directly reflects the user's competitive advantage in the position, avoiding the problem of no response after application caused by traditional recommendations that only focus on semantic matching and ignore actual competitive feasibility.

[0045] Step 3.7 involves weighting the intent-job alignment score and the competitiveness prediction score to calculate the overall ranking score for each candidate job. This includes: first, clarifying the weighting logic of the intent-job alignment score and the competitiveness prediction score: the alignment score is the foundation of the relevance between the job and user needs, while the competitiveness prediction score is crucial for effective communication after application. Therefore, the alignment score is weighted at 60%, and the competitiveness prediction score at 40%. The two scores are then uniformly quantified and calibrated to ensure they fall within the 0-100 score range: the alignment score uses the calculation results from Step 3.4, and the competitiveness prediction score uses the calculation results from Step 3.6. Finally, a weighted combination calculation is performed according to the set weights: the intent-job alignment score is multiplied by 0.6, the competitiveness prediction score is multiplied by 0.4, and the two products are summed to obtain the overall ranking score for each candidate job. This score covers both the demand matching dimension and the competitive feasibility dimension, effectively solving the core problem of traditional job recommendations where the mismatch rate exceeds 40% and the effective matching rate is less than 10%.

[0046] Step 3.8: Sort all candidate positions based on the comprehensive ranking score to obtain the final recommended position list. Specifically, this includes: sorting all candidate positions in descending order of their comprehensive ranking scores, prioritizing the retention of positions with higher scores; considering the pain point of long processing times for users, controlling the final number of recommended positions to within 30 to avoid information overload; if positions with the same comprehensive ranking score are found during the sorting process, further sorting is performed based on industry trend data, with positions having higher demand growth rates ranked higher to align with users' long-term career planning needs; after sorting, the final recommended position list is generated, with each position in the list including the core components of its comprehensive ranking score, providing a foundation for the interpretability module; and enabling users to quickly focus on highly relevant and competitive positions, reducing manual screening time and improving job search efficiency.

[0047] In this embodiment of the invention, for candidate positions with qualified alignment scores, user competitiveness features and position competitiveness features are first extracted. These are then weighted to obtain user competitiveness scores and position competitiveness scores, respectively. The two scores are then weighted and fused to obtain a predicted competitiveness score. Subsequently, the intention-position alignment score and the predicted competitiveness score are weighted and combined according to preset weights to calculate the comprehensive ranking score for each candidate position. Finally, all candidate positions are ranked based on the comprehensive ranking score to obtain the final recommended position list. This overcomes the technical problem of traditional job recommendation technologies that only focus on the semantic matching of users and positions while ignoring the correlation between users' actual competitiveness and position competitiveness. This results in high click-through rates but low application success rates and a mismatch rate exceeding 40%. Furthermore, users need to manually assess their suitability from a large number of recommended positions, exacerbating the inefficiency of job seeking under information overload. This invention achieves a 28% increase in the click-to-application conversion rate from 12% to 15.6%, an increase in the HR response rate after application from 35% to 49%, and a 28% increase in the user application success rate. It also reduces the time users spend identifying invalid positions, allowing them to quickly focus on highly suitable and competitive positions.

[0048] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Collect user interaction feedback data on the recommended job list. This data includes clicks, favorites, submissions, and ignore actions. Based on this data, training data with positive and negative labels is obtained. Specifically, this involves: first, collecting interaction feedback data by recording all user actions after viewing the recommended job list in real time. This includes clicks, favorites, submissions, and ignore actions for individual recommended jobs. During recording, each action data point is associated with the corresponding user identity information, job information, and the timestamp of the action. User identity information is used to locate the specific user's intent model, job information is used to associate with the corresponding job vector, and the timestamp is used to distinguish user feedback at different times, providing a time dimension reference for subsequent incremental model updates. Then, based on the collected interaction feedback data... The system employs positive and negative sample labeling to clarify the correspondence between feedback behaviors and sample labels: user clicks, favorites, and applications are classified as positive feedback because these behaviors directly reflect the user's interest in or approval of the position, aligning with their job search needs, and are thus labeled as positive samples. User ignores are classified as negative feedback because this behavior indicates that the user believes the position does not meet their needs, and are thus labeled as negative samples. Finally, the labeling results are integrated with relevant feature data, binding each feedback behavior with positive or negative labels to the user's intent vector at the time of the behavior and the corresponding position's job vector, resulting in a complete training dataset. Simultaneously, all labeled training data is organized chronologically to ensure that subsequent incremental fine-tuning is based on the latest user feedback, addressing the problem of user interaction data not being collected and utilized in real time in traditional systems.

[0049] Step 4.2 involves incrementally fine-tuning the multi-granularity temporal intent encoder based on the training data to update the parameters of user dynamic intent modeling, and incrementally fine-tuning the cross-modal job semantic enhancement model based on the training data to update the parameters of job semantic representation. Specifically, this includes: To address the pain point of traditional models having iteration cycles as long as 15 days, incremental fine-tuning is achieved by selecting the training data collected and labeled within the previous day as the incremental training set every morning at midnight. This avoids the time consumption caused by training with the full dataset, significantly shortening the model iteration cycle. For incremental fine-tuning of the multi-granularity temporal intent encoder, the focus is on adjusting the weight parameters of the multi-granularity attention layer and the network parameters of the temporal coding layer. The adjustment of the multi-granularity attention layer weight parameters needs to be combined with the changes in intent reflected by the user's latest feedback behavior. For example, if the user has recently made multiple clicks on job postings that are significantly different from their original intent, incremental training will be used to address this. The attention weights for short-term behavior sequences, medium-term behavior sequences, and long-term profiles are adjusted to make the model focus more on capturing the latest short-term needs of users. The adjustment of the temporal coding layer network parameters focuses on optimizing the Transformer encoder's ability to capture the temporal correlation of short-term behavior and the GRU network's ability to handle the stability of medium-term preferences. This ensures that the adjusted model can more accurately distinguish users' short-term urgent needs, medium-term preferences, and long-term career plans, thereby improving the accuracy of intent recognition. The incremental fine-tuning of the cross-modal job semantic enhancement model focuses on optimizing the parameters of the single-modal coding layer and the cross-modal attention alignment layer.

[0050] When adjusting the parameters of the single-modal encoding layer, the BERT pre-trained model's ability to mine implicit needs in job description texts and the encoding accuracy of discrete features of company profiles in the embedding layer are optimized based on user feedback on jobs with different company profiles, job popularity, and industry trend features. This optimizes the alignment effect between text semantic vectors and company vectors, popularity vectors, and trend vectors, allowing job vectors to more comprehensively integrate the value of multimodal data and reduce the representation error of newly released jobs. During fine-tuning, the model parameters are updated gradually with small batches of data to avoid drastic parameter fluctuations affecting the overall stability of the model. After each fine-tuning, internal validation is conducted to ensure that the model's performance in intent recognition and job representation does not significantly decline. At the same time, the click-through conversion rate of recommended jobs and the HR response rate after application are continuously tracked. The fine-tuned model can continuously optimize the recommendation effect, improve user recommendation acceptance and platform monthly retention rate.

[0051] In this embodiment of the invention, by collecting user interaction feedback data such as clicks, favorites, submissions, and ignores on the recommended job list, and labeling clicks, favorites, and submissions as positive samples and ignores as negative samples to obtain training data with positive and negative labels, and then incrementally fine-tuning the multi-granularity temporal intent encoder and the cross-modal job semantic enhancement model based on this training data, and updating the relevant parameters of user dynamic intent modeling and job semantic representation, this invention overcomes the shortcomings of existing job recommendation technologies, such as the lack of feedback loop, the failure of user interaction data to feed back to the model in real time leading to a model iteration cycle of up to 15 days, the inability to dynamically optimize the recommendation effect, and the lack of effective feedback support for the recommendation results resulting in a user recommendation acceptance rate of less than 50% and a platform monthly retention rate of less than 35%. This invention shortens the model iteration cycle from 15 days to 7 days, dynamically optimizes the recommendation effect, makes user dynamic intent modeling more accurate, and makes job semantic representation more aligned with user needs, thereby helping to improve user trust and acceptance of the recommendation results, and ultimately increasing the platform monthly retention rate from 35% to 60%.

[0052] like Figure 2 As shown, embodiments of the present invention also provide a job recommendation system based on multimodal deep semantics and dynamic intent modeling, including: The acquisition module is used to collect and preprocess multimodal data to obtain preprocessed multimodal data, which includes user resume data, user behavior sequence data, job description text data, company profile data, job popularity data, and industry trend data. Based on user resume data and user behavior sequence data, the module models user dynamic intent to obtain user intent vectors. Based on job description text data, company profile data, job popularity data, and industry trend data, the module performs deep semantic representation of jobs to obtain job vectors. The processing module is used to recall candidate positions based on user intent vectors and job vectors to obtain the recalled candidate positions; based on the recalled candidate positions, it calculates the alignment score between user intent and job position; and performs a second-stage fine ranking of the candidate positions ranked in the first stage to calculate the competitiveness prediction score. Finally, it obtains a recommended job list based on the comprehensive ranking score; and incrementally updates the user dynamic intent model and job semantic representation model based on user interaction feedback data of the recommended job list.

[0053] In this embodiment, the MG-TIE structure consists of four layers: an input layer, a temporal coding layer, a multi-granularity attention layer, and an intent drift correction layer, wherein: The input layer is used to receive three types of data: Short-term behavioral sequence S: User click-favorite-submit behavior in the past 7 days, sorted by timestamp, each behavior includes job ID-behavior type-stay duration; Mid-term preference sequence M, user's job type favorites, industry focus, and salary filtering records over the past 30 days; Long-term profile P consists of fixed attributes in the user's resume, such as skill tags, years of work experience, education, expected salary, or city.

[0054] 2. Timing coding layer: Short-term sequence S: Using a Transformer encoder, the temporal correlation of short-term behaviors is captured, such as clicking 3 consecutive 'urgently hiring Python' posts, which implies a short-term need for quick onboarding. The output vector is VS. Intermediate sequence M: A GRU network is used to handle the stability of intermediate preferences, and the output vector is VM; Long-term profile P: Through a fully connected layer + BatchNorm, discrete and continuous features are transformed into vectors VP.

[0055] 3. Multi-granularity attention layer: Introducing behavioral activity weights: The attention weights are dynamically adjusted based on the frequency of users' recent behaviors. For example, the short-term behavioral activity of users who quit their jobs without a backup plan is 0.8, and the weight αS is 0.6; the short-term activity of users who are currently employed and changing jobs is 0.3, and the weight αS is 0.2. The formula for calculating weights is: αS = frequency of short-term behavior or total frequency of behavior × 0.7 + 0.3, αM = frequency of medium-term behavior or total frequency of behavior × 0.5 + 0.2, αP = 1 - αS - αM, and the sum of the weights is 1. Attention fusion: VI = αS×VS + αM×VM + αP×VP, outputting the user's dynamic intent vector VI.

[0056] 4. Intent Drift Correction Layer: If a user clicks on a job posting that differs from the current intent VI by more than 60% three times in a row (e.g., if the original intent was Python development but the user clicks on Java development), a weight recalculation is triggered to correct the intent vector in real time.

[0057] 4.2.2 Innovation Point 2: Cross-modal Job Semantic Enhancement Model—Solving the Challenge of Multimodal Job Representation Design Background: To address the shortcomings of traditional job representation that relies solely on text and ignores cross-modal data, this design integrates multimodal data including job text, company profile, job popularity, and industry trends to achieve deep semantic representation of jobs, especially improving the representation accuracy of new jobs.

[0058] In this embodiment, the CM-SRM structure consists of four layers: an input layer, a single-modal coding layer, a cross-modal attention alignment layer, and a multimodal fusion layer. 1. Input layer: Receives four types of cross-modal data: Job Description (T): Company Profile C: Industry, Size, Salary Range, Funding Stage; Job Popularity H: Number of applications in the past 7 days, HR response rate, and number of days the job posting has been published; Industry Trends (Tend): Demand growth rate in the field of the job and talent supply gap.

[0059] 2. Single-mode coding layer: Text T: Using the BERT-base pre-trained model, the [CLS] vector is taken as the text semantic vector VT; Company Profile C: Discrete features are input into the Embedding layer after One-Hot encoding, and continuous features are input into the fully connected layer after Min-Max normalization, and then fused into the company vector VC; Job popularity H: The number of applications or job openings and the HR response rate are normalized, then input into the LSTM layer, and the output is a popularity vector VH. Industry Trend Tend: Demand growth rate and talent gap, after being standardized, are input into the fully connected layer and output as trend vector VTend.

[0060] 3. Cross-modal attention alignment layer: Using text vector VT as the query and VC, VH, and VTend as the key or value, the alignment of text semantics and cross-modal features is achieved through a cross-attention mechanism; Output the aligned vectors VTattn, VCattn, VHattn, and VTendattn.

[0061] 4. Multimodal fusion layer: Vector fusion is performed using Concat + residual join: VJconcat=Concat; b) After passing through two fully connected layers and Dropout, the final job vector VJ is output.

[0062] 4. New Job Representation Optimization: For new jobs with no historical behavior, the representation error of new jobs is reduced by more than 35% by adjusting the average vector of similar jobs in the industry plus the weight of the company profile. For example, Java developers and Python developers in the same company share 50% of the company characteristics.

[0063] In this embodiment, a two-stage dynamic ranking framework addresses the challenge of balancing recommendation accuracy and conversion rate, specifically including: 1. Recall Layer: Based on the user intent vector VI and the job vector VJ, FAISS vector retrieval with rule filtering is used to exclude industries blocked by users or jobs with salaries lower than expected, and 100 candidate jobs are selected from a job database of 500,000 jobs.

[0064] 2. First-stage fine-tuning: Intent-job alignment model: Input: User intent vector VI, job vector VJ, user-job keyword matching features, such as the number of matches between user skills containing Python and job requirements containing Python; Calculation logic: ① Semantic similarity: cosine similarity between VI and VJ; ② Keyword matching score: number of matched skill / experience keywords / total number of job keywords; Output: Alignment score S1, filter candidate positions with S1 ≥ 60.

[0065] 3. Second-stage refined ranking: Competitiveness prediction enhancement model: Input features: User competitiveness features: skills gap score, experience matching score, expectation matching score; Job competitiveness features: competition coefficient, HR response rate; Calculation logic: S2 = 0.5 × user competitiveness score + 0.5 × job competitiveness score; Output: final ranking score S = 0.6 × S1 + 0.4 × S2.

[0066] 4. Explainability Generation Module: Score Decomposition: S1 is decomposed into skill matching score, experience matching score, and expectation matching score, and S2 is decomposed into user competitiveness score and job competitiveness score; Natural Language Generation: Explanatory text is obtained based on the decomposed scores.

[0067] The recommendation system of this invention is deployed on a distributed server cluster, with the following specific hardware configuration: 1. Compute nodes: 2 Intel Xeon Gold 6330 CPUs, 4 NVIDIA A100 GPUs; 2. Storage node: 1 distributed storage server, supporting real-time data read and write; 3. Retrieval Node: 1 dedicated FAISS acceleration server, supporting IVFPQ indexing, with a retrieval latency of <100ms; 4. Network environment: 10Gbps Ethernet, supporting low-latency data transmission between nodes.

[0068] In embodiments of the present invention, the software environment of the present invention includes: 1. Operating System: As the underlying operating environment, Ubuntu 22.04LTS is the mainstream choice for servers.

[0069] 2. Development languages: Python 3.9 is used for the core logic development of the system and is adapted to the entire job recommendation process; Java 17 is used to focus on front-end interface development, serving as an interaction bridge between users and the system.

[0070] 3. Deep learning frameworks: PyTorch 2.0 focuses on model training and real-time inference of user dynamic intent and job semantic representation, while TensorFlow 2.14 specializes in BERT pre-training for the recruitment field.

[0071] 4. Search tools: FAISS 1.7.4 is a core tool for candidate job retrieval, which efficiently matches user intent vectors with job vectors. Elasticsearch 8.10 supplements this with vector retrieval, which is responsible for accurate text matching and rule filtering.

[0072] 5. Data processing tools: Pandas 2.1.0 is used to process small-scale data and support rapid validation during the model development stage; Spark 3.4.0 is used to process large-scale data and support industrial preprocessing of full data.

[0073] 6. Service Deployment: Docker 25.0.0, which enables environment isolation and standardization of system modules, and achieves automated management and high availability of container clusters.

[0074] The innovative algorithm parameters of this invention are set as follows: MG-TIE module: short-term window of 7 days, medium-term window of 30 days; 6-layer Transformer encoder, 4 attention heads; 256-dimensional GRU hidden layer; 512-dimensional intent vector. CM-SRM module: BERT-base, i.e., fine-tuned learning rate 2e-5; number of cross-attention heads 2; dimensionality of the fused job vector 512; Dropout probability 0.2; TS-DRF module: recall candidate count from 100 to fine-tune 30; S1 weight 0.6, S2 weight 0.4; after fine-tuning, interpretable text generation adopts T5-small model.

[0075] The model training and validation in the embodiments of the present invention include: Dataset: Real data from a comprehensive recruitment platform, including 100,000 users, 20,000 new users, 500,000 job postings, and 10 million user-job posting interaction records, divided into training set, validation set, and test set in a 7:2:1 ratio; Pre-training: The BERT model was fine-tuned using job description text from the recruitment domain; End-to-end training: MG-TIE, CM-SRM, and TS-DRF are trained jointly, with the loss function being cross-entropy loss and L2 regularization; Online iteration: Incremental fine-tuning is performed every day at midnight based on the user interaction data of the previous day, and the updated model is verified through A / B testing; Verification results: The comparison of core metrics on the test set is shown in Table 1 below: Table 1 Comparison of core metrics on the test set The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A job recommendation method based on multi-modal deep semantic and dynamic intent modeling, characterized in that, The method comprises: Collecting multi-modal data and preprocessing to obtain preprocessed multi-modal data, the multi-modal data including user resume data, user behavior sequence data, job description text data, company profile data, job heat data and industry trend data; Based on the user resume data and the user behavior sequence data, the user dynamic intention is modeled to obtain a user intention vector; based on the job description text data, the company profile data, the job heat data and the industry trend data, the job is deeply semantically represented to obtain a job vector; According to the user intention vector and the job vector, candidate jobs are recalled to obtain recalled candidate jobs; according to the recalled candidate jobs, the user intention alignment score is calculated; the second stage of the candidate jobs is refined to calculate the competitiveness prediction score, and finally the recommended job list is obtained based on the comprehensive ranking score; Based on the interactive feedback data of the user on the recommended job list, the user dynamic intention modeling and the job semantic representation model are incrementally updated.

2. The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 1, characterized in that, Collecting multi-modal data and preprocessing to obtain preprocessed multi-modal data, the multi-modal data including user resume data, user behavior sequence data, job description text data, company profile data, job heat data and industry trend data, comprising: Collecting original multi-modal data from a recruitment platform database, the original multi-modal data including user resume data, user behavior sequence data, job description text data, company profile data, job heat data and industry trend data; The collected original multi-modal data is cleaned to obtain cleaned data, and the cleaned data is standardized to obtain standardized data; the standardized data is feature extracted to obtain preprocessed multi-modal data. 3.The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 2, characterized in that, Based on the user resume data and the user behavior sequence data, the user dynamic intention is modeled to obtain a user intention vector, comprising: Extracting user resume data from the preprocessed multi-modal data as a long-term profile, and extracting user behavior sequence data, and splitting the user behavior sequence data into short-term behavior sequence and medium-term behavior sequence according to a preset time window; The short-term behavior sequence is time-encoded by a Transformer encoder to obtain a short-term intention vector, the medium-term behavior sequence is time-encoded by a GRU network to obtain a medium-term intention vector, and the long-term profile is feature mapped by a fully connected layer to obtain a long-term intention vector; Based on the short-term intention vector, the medium-term intention vector and the long-term intention vector, dynamic weight calculation and weighted fusion are performed through a multi-granularity attention layer to obtain an initial user intention vector; According to the difference between the user real-time behavior data and the initial user intention vector, the intention drift is detected, and when the intention drift is detected, the weights of the multi-granularity attention layer are recalculated and adjusted to obtain a final user intention vector. 4.The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 3, characterized in that, Based on the job description text data, the company profile data, the job heat data and the industry trend data, the job is deeply semantically represented to obtain a job vector, comprising: The BERT pre-training model is used for semantic coding of the job description text data to obtain a text semantic vector, and feature coding is performed on the company portrait data, including embedding processing of discrete features and normalization processing of continuous features, to obtain a company vector; The job heat data is time-series coded, the LSTM network is used to capture the heat change trend to obtain a heat vector, and the industry trend data is standardized and mapped through a fully connected layer to obtain a trend vector; The text semantic vector is used as a query vector, and the company vector, heat vector and trend vector are used as key value vectors, and cross attention mechanism is used for cross-modal semantic alignment to obtain aligned text semantic vector, company vector, heat vector and trend vector; The aligned text semantic vector, company vector, heat vector and trend vector are spliced, then fused through residual connection and fully connected layer, and processed by Dropout to obtain the final job vector. 5.The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 4, characterized in that, According to the user intention vector and the job vector, the candidate job recall is performed to obtain the recalled candidate jobs; According to the recalled candidate jobs, the user intention and job alignment score is calculated, including: Based on the user intention vector and the job vector, the vector retrieval tool is used for similarity calculation, and the candidate job set is quickly retrieved from the job library; Based on the retrieved candidate job set, rule filtering is applied to the candidate job set to exclude jobs in the user shielded industry or with a salary lower than the expected salary to obtain a filtered candidate job list; Based on the filtered candidate job list, the semantic similarity between the user intention vector and the job vector is calculated for each job in the filtered candidate job list, and the keyword matching features between the user and the job are extracted; Based on the semantic similarity and the keyword matching features, the user intention and job alignment score is calculated through weighted fusion. 6.The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 5, characterized in that, The first-stage refined candidate jobs are used for second-stage refinement to calculate the competitiveness prediction score, and finally the recommended job list is obtained based on the comprehensive ranking score, including: Based on the candidate jobs with qualified alignment scores, user competitiveness features and job competitiveness features are extracted, and based on the user competitiveness features and job competitiveness features, user competitiveness scores and job competitiveness scores are calculated through weighting respectively; The user competitiveness score and the job competitiveness score are weighted and fused to obtain the competitiveness prediction score; The intention and job alignment score and the competitiveness prediction score are weighted and combined to calculate the comprehensive ranking score of each candidate job; Based on the comprehensive ranking score, all candidate jobs are sorted to obtain the final recommended job list.

7. The job recommendation method based on multi-modal deep semantic and dynamic intent modeling according to claim 6, characterized in that, Based on the user interaction feedback data on the recommended job list, the user dynamic intention model and the job semantic representation model are incrementally updated, including: Collecting user interaction feedback data on the recommended job list, the interaction feedback data including clicking, collecting, delivering, ignoring behavior, based on the interaction feedback data, obtaining training data with positive and negative sample labels; Incrementally fine-tune the multi-granularity time sequence intention encoder based on the training data, update the parameters of the user dynamic intention modeling, and incrementally fine-tune the cross-modal post semantic enhancement model based on the training data, update the parameters of the post semantic representation.

8. A job recommendation system based on multi-modal deep semantic and dynamic intent modeling, the system implements the method as claimed in any one of claims 1 to 7, characterized in that, The method comprises the steps of: An acquisition module is configured to collect and preprocess multi-modal data to obtain preprocessed multi-modal data, wherein the multi-modal data comprises user resume data, user behavior sequence data, post description text data, company profile data, post heat data, and industry trend data; user dynamic intention is modeled based on the user resume data and the user behavior sequence data to obtain a user intention vector; and a post is deeply semantically represented based on the post description text data, the company profile data, the post heat data, and the industry trend data to obtain a post vector; A processing module is configured to perform candidate post recall based on the user intention vector and the post vector to obtain recalled candidate posts; calculate a user intention and post alignment score based on the recalled candidate posts; perform second-stage precision ranking on the candidate posts ranked in the first stage to calculate a competitiveness prediction score; and finally obtain a recommended post list based on a comprehensive ranking score; and incrementally update the user dynamic intention modeling and the post semantic representation model based on user interaction feedback data on the recommended post list.

9. A computing device, comprising: The method comprises the steps of: One or more processors; A storage device is configured to store one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images