Employment post recommendation method, system, device and storage medium based on big data analysis

Abstract

The application provides an employment post recommendation method and system based on big data analysis, a device and a storage medium. Firstly, talent demand data, basic characteristics of job seekers and trajectory data of job seekers in the talent market in a target administrative region are obtained. Then, multi-dimensional features of talent demand are extracted to construct a demand feature set, and combined with the basic characteristics and the demand feature set, industry inclination data of the job seekers is obtained by using a support vector machine model. Then, the inclination data is adapted and screened based on the trajectory data to obtain the target industry corresponding to the job seekers. The K-nearest neighbor algorithm is used to calculate the similarity between the job seeker characteristics and the post characteristics corresponding to the target industry, and all target posts with a similarity value greater than a preset threshold are sorted to generate a post recommendation list for the job seekers. The method can effectively solve the problem of insufficient post matching accuracy in the regional talent market, and further realize more accurate and adaptive regional employment recommendation.

Description

Technical Field

[0001] This application relates to the technical field of job recommendation, and in particular to a method, system, device, and storage medium for job recommendation based on big data analysis. Background Technology

[0002] In various scenarios such as regional job fairs, campus recruitment events, and local employment service platforms, the accuracy of job recommendations directly impacts job search efficiency and talent allocation effectiveness. Furthermore, with the continuous expansion of employment market data and the increasing demand for personalized services, achieving highly adaptable job recommendations has significant practical implications and promising application prospects.

[0003] Currently, most common job recommendation technologies are based on keyword matching between static text features in job seekers' resumes and job descriptions, or on building collaborative filtering models using historical application and browsing behavior data, and then providing generalized recommendations to individuals by analyzing group behavior patterns.

[0004] However, when these methods target regional talent markets with clear geographical limitations and industry clustering characteristics, they often fail to effectively integrate the dynamic supply and demand situation of the industry within the region, and they also rarely consider the real-time behavioral feedback of job seekers in actual recruitment scenarios. This leads to the recommendation results being easily detached from the actual structure of the employment market, with significant limitations in industry relevance and regional adaptability, thus affecting the final effect of job matching. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, device, and storage medium for job recommendation based on big data analysis, in order to solve the problem of insufficient accuracy in job matching in the prior art.

[0006] To address the aforementioned technical problems, firstly, this application provides a job recommendation method based on big data analysis, comprising:

[0007] Acquire talent demand data for various industries within the target administrative region, basic characteristics of job seekers, and data on job seekers' movements in the talent market;

[0008] Multidimensional feature extraction is performed on the talent demand data to obtain a demand feature set;

[0009] The basic features and the set of demand features are input into a classification model built on support vector machines to obtain the preference data of each job seeker.

[0010] Based on the trajectory data, the tendency data is adapted and filtered to obtain the target industry corresponding to the job seeker;

[0011] The K-nearest neighbor algorithm is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry. Based on the similarity calculation results, all target jobs with similarity values ​​greater than a preset threshold are sorted to generate a job recommendation list for job seekers.

[0012] Optionally, the step of inputting the basic features and the set of demand features into a classification model built on a support vector machine to obtain the preference data of each job seeker includes:

[0013] Based on trajectory data within a historical preset period, the frequency of job seekers' visits to different industries and the duration of their stay are determined in order to generate a preference vector for job seekers.

[0014] The preference vector is used to weight the demand feature set by importance to obtain the modulation feature set. All features in the modulation feature set are then fused with the basic features to obtain composite features.

[0015] Support vector machines are used to perform boundary learning and partitioning of the composite features in high-dimensional space to generate classification results and corresponding initial confidence scores.

[0016] Based on the inter-industry relationships within the target administrative region, an inter-industry relationship diagram is constructed, and the classification results are amplified on the relationship diagram to enhance the inter-industry relationships.

[0017] Based on the correlation degree, the initial confidence degree is propagated and updated under the constraints of the preset graph structure to generate the final confidence degree;

[0018] The classification results are integrated with the final confidence level to form the propensity data.

[0019] Optionally, the step of constructing an inter-industry association diagram based on the inter-industry relationships within the target administrative region, and then applying influence diffusion to the classification results on the association diagram to enhance the inter-industry correlation, includes:

[0020] Based on the upstream and downstream dependencies between industries within the target administrative region, a first graph edge is constructed; based on the degree of overlap between core skills, a second graph edge is constructed; and based on the clustering density of the geographical locations of each industry, a third graph edge is constructed.

[0021] By merging the edges of the first graph, the second graph, and the third graph, an inter-industry association graph is generated.

[0022] Based on the classification results, the nodes in the association graph are initialized to obtain an initial state vector;

[0023] A graph attention network is used to propagate and gating the initial state vector on the association graph to obtain the target state vector, and the target state vector is decoded into a numerical form of association degree.

[0024] Optionally, the step of using the K-nearest neighbor algorithm to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry includes:

[0025] By integrating the basic features, the preference vector corresponding to the trajectory data within the historical preset period, and the final confidence level, a core vector for representing job seeker characteristics is generated. By integrating the text features of the job and the correlation between various industries, a context vector for representing job characteristics is generated.

[0026] A graph-augmented metric learning method is adopted, which uses the inter-industry relationship graph as a constraint to learn a metric function;

[0027] The core vector and the context vector are mapped to a unified metric space using the metric function, and the similarity between the core vector and the context vector is calculated using the K-nearest neighbor algorithm in the metric space.

[0028] Optionally, the graph-augmented metric learning method, using an inter-industry association graph as a constraint, learns a metric function, including:

[0029] Extract the graph topology structure of the inter-industry relationship graph, and use the graph topology structure as graph topology constraints;

[0030] By using a graph attention mechanism, neighborhood information is aggregated from the context vector based on the graph topological constraints to generate a perceptual feature set;

[0031] A contrastive learning strategy is employed to construct positive and negative sample pairs in the perceptual feature set based on the graph topological constraints.

[0032] The similarity reward between the positive sample pairs and the similarity penalty between the negative sample pairs are calculated to form the structural contrast loss;

[0033] The structural contrast loss is used as a regularization term, and the regularization term is fused with the main loss function of metric learning to form a hybrid loss function;

[0034] The optimization aims to minimize the hybrid loss function. The internal parameters of the preset function are updated, and the function corresponding to the updated parameters is used as the metric function.

[0035] Optionally, the step of adapting and filtering the tendency data based on the trajectory data to obtain the target industry corresponding to the job seeker includes:

[0036] The preference vector corresponding to the trajectory data is fused with the classification results and final confidence scores in the tendency data to obtain the comprehensive suitability of each candidate industry;

[0037] Based on the overall fit, the candidate industry with the highest fit value is selected as the target industry from the candidate industries.

[0038] Optionally, the step of extracting multi-dimensional features from the talent demand data to obtain a demand feature set includes:

[0039] Multi-dimensional features are extracted from the job information in the talent demand data to obtain multi-dimensional features for each industry. The multi-dimensional features include: a first feature corresponding to the number of jobs, a second feature corresponding to skill requirements, a third feature corresponding to salary levels, a fourth feature corresponding to growth trends, and a fifth feature corresponding to talent supply and demand surplus.

[0040] The multi-dimensional features are standardized, and the standardized features are then structured and organized according to industry type to obtain a set of demand features.

[0041] Secondly, this application provides a job recommendation system based on big data analysis, including:

[0042] The acquisition module is used to acquire talent demand data for various industries within the target administrative region, basic characteristics of job seekers, and trajectory data of job seekers in the talent market.

[0043] The extraction module is used to extract multi-dimensional features from the talent demand data to obtain a demand feature set;

[0044] The analysis module is used to input the basic features and the set of demand features into a classification model built on a support vector machine to obtain the preference data of each job seeker;

[0045] The filtering module is used to adapt and filter the tendency data based on the trajectory data to obtain the target industry corresponding to the job seeker.

[0046] The calculation module is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry using the K-nearest neighbor algorithm, and to sort all target jobs with similarity values ​​greater than a preset threshold based on the similarity calculation results, thereby generating a job recommendation list for job seekers.

[0047] Thirdly, this application provides an electronic device, comprising:

[0048] Memory, used to store computer programs;

[0049] A processor is configured to execute the computer program to implement the steps of the job recommendation method based on big data analysis as described in the first aspect above.

[0050] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the job recommendation method based on big data analysis as described in the first aspect above.

[0051] The job recommendation method based on big data analysis provided in this application has the following beneficial effects:

[0052] This application acquires talent demand data, basic characteristics of job seekers, and their trajectory data in the talent market within the target administrative region. This provides a multi-dimensional data foundation integrating macro-market data, individual static attributes, and dynamic behaviors for subsequent analysis. Next, multi-dimensional feature extraction is performed on the talent demand data to construct a demand feature set, transforming regional industry supply and demand into a structured and quantifiable feature expression. Then, the basic characteristics of job seekers and the demand feature set are input into a support vector machine model for processing, combining individual abilities with regional market job requirements to initially determine quantifiable industry preferences of job seekers. Furthermore, based on trajectory data, these industry preferences are adapted and filtered, and the initial preferences can be verified and corrected by incorporating the job seeker's real-time behavioral preferences, thereby determining target industries that better match their true intentions. Finally, the K-nearest neighbor algorithm is used to perform refined similarity calculation and ranking of positions within the target industry. This enables multi-dimensional and accurate matching of job seeker characteristics with specific job requirements within the determined industry direction, ultimately generating a highly personalized job recommendation list, effectively improving the recommendation success rate and user satisfaction in regional employment scenarios.

[0053] Furthermore, this application generates preference vectors reflecting job seekers' interests based on historical trajectory data, and uses these vectors to weight the regional demand feature set based on importance. This integrates individual behavioral preferences into market feature analysis, thereby highlighting industry information that is more relevant to them. Subsequently, the weighted market features and individual basic features are fused to form composite features, and support vector machines are used for high-dimensional classification. This allows for the generation of more reliable initial industry classifications and confidence levels from a composite perspective. Furthermore, by constructing a graph structure that reflects the real connections between regional industries, and updating the initial classification results through influence diffusion and confidence propagation on this graph, the inherent connections between industries can be used to collaboratively verify and optimize the preliminary judgment. This ensures that the final generated preference data simultaneously includes the classification results and the confidence levels calibrated by the graph structure, thus providing a more reliable and comprehensive decision-making basis for subsequent screening and matching steps. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating a job recommendation method based on big data analysis provided in this application embodiment;

[0056] Figure 2 A schematic diagram illustrating a specific implementation of a job recommendation method based on big data analysis provided in this application embodiment;

[0057] Figure 3 A schematic diagram of the structure of a job recommendation system based on big data analysis provided in an embodiment of this application;

[0058] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0059] The matching bias that often occurs in existing job recommendation systems in regional talent markets is due to the fact that most existing methods rely on static resume texts or historical application records for analysis. This makes it difficult to reflect real-time changes in industry demand within the region, and it also ignores the preference signals conveyed by job seekers' actual behavior at recruitment sites. This disconnect between data and application scenarios means that the recommendation results often differ from the actual industrial structure of the region and the immediate intentions of individuals, ultimately resulting in limited accuracy in matching people with jobs.

[0060] To address this issue, this application proposes a job recommendation method based on big data analysis. This method first simultaneously collects regional industry demand data, job seeker basic information, and their actual activity data in the talent market. Then, through feature extraction and fusion analysis, it combines regional supply and demand with individual dynamic behavior to form a comprehensive judgment of job seeker preferences. Finally, it performs refined matching and ranking at the job level within the target industry. This method organically integrates macro-level regional industry characteristics, individual static ability attributes, and dynamic behavioral preferences, ensuring that job recommendation results not only conform to the actual structure of the employment market but also align with job seekers' individual intentions and real-time behavior. This effectively solves the problem of insufficient matching accuracy in regional scenarios using traditional methods, thus improving the effectiveness of employment services.

[0061] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] Example 1

[0063] The core of this application is to provide a job recommendation method based on big data analysis, and a flowchart of one specific implementation method is shown below. Figure 1 As shown, the method includes:

[0064] S101. Obtain talent demand data for various industries within the target administrative region, basic characteristics of job seekers, and trajectory data of job seekers in the talent market;

[0065] Among them, the target administrative region refers to the specific geographical area set for job recommendations, such as a city or district; the talent demand data mainly reflects the recruitment demand information such as the number of job postings, skill requirements, and salary ranges of employers in various industries within the region; the basic characteristics refer to static personal information such as educational background, major, and years of work experience; the trajectory data refers to the physical data of job seekers' movement trajectory at the talent market event site. This data is collected through positioning devices deployed in booths and passageways, including information such as the order in which job seekers visit different industry booths, the duration of their stay, and their movement paths, which is used to objectively reflect their real-time industry attention and interest preferences;

[0066] In step S101, in order to achieve precise regional recommendations, the raw data of talent demand in various industries in the recent period are first obtained from the government employment service platform, mainstream recruitment websites and recruitment interfaces of key enterprises in the target administrative region. At the same time, basic features are extracted from the electronic resumes or registration information filled in by job seekers.

[0067] For example, when holding offline job fairs, acoustic or Bluetooth positioning devices are deployed in various industry booths and main passages of the venue to collect the identification signals worn by job seekers in real time. The signals change as the job seekers move and record the location coordinates of each job seeker. Then, by matching the coordinates with the booth area, trajectory data containing timestamps, booth industry type and dwell time is finally generated, and thus a trajectory data sequence is constructed.

[0068] S102. Perform multi-dimensional feature extraction on the talent demand data to obtain a demand feature set.

[0069] Among them, multi-dimensional features refer to the quantitative description of the employment market situation of an industry in a target region from different perspectives.

[0070] In a specific implementation scheme, step S102 involves extracting multi-dimensional features from the talent demand data to obtain a demand feature set. The talent demand data is important data on job postings in the market and serves as fundamental information for analysis, including:

[0071] Step 1021: Extract multi-dimensional features from the job information in the talent demand data to obtain multi-dimensional features for each industry. The multi-dimensional features include: a first feature corresponding to the number of jobs, a second feature corresponding to skill requirements, a third feature corresponding to salary levels, a fourth feature corresponding to growth trends, and a fifth feature corresponding to talent supply and demand surplus.

[0072] The first feature reflects the industry's recruitment scale; the second feature, through analyzing high-frequency keywords in job descriptions, characterizes the industry's core competency requirements for talent; the third feature reflects the salary attractiveness of the industry's positions; the fourth feature refers to the degree of change in the number of positions currently compared to the same period in history, used to determine whether the industry is in a state of expansion or contraction; the fifth feature is a composite indicator, which can be calculated from the total number of job positions in the industry and the number of job seekers, used to intuitively reflect whether the industry in the region faces a "difficulty in finding jobs" or a "difficulty in recruiting." Among these, talent demand data is important data and basic information for analyzing job information in the market, and has significant research value. It can further obtain a demand feature set, and provide a technical foundation for subsequent classification models and analyses.

[0073] In step 1021, multi-dimensional feature extraction can be performed simultaneously through multiple channels. For example, the first channel categorizes talent demand data by industry and counts the total number of job postings for each industry as the first feature. Simultaneously, the second channel performs word segmentation and keyword analysis on all job description texts to extract high-frequency words representing skill requirements, such as "Python" and "project management," and their frequency of occurrence, thereby constructing the second feature. The third channel calculates the median or average of the salary range for the industry's job postings to obtain the third feature.

[0074] Similarly, the fourth channel retrieves the number of job openings in the industry over a past period and compares it with the current number to calculate the growth rate, which serves as the fourth feature. The fifth channel obtains the number of job seekers in the industry within the target region from talent supply data and compares it with the total job supply to obtain the fifth feature. The talent supply and demand surplus is the ratio of the total job supply to the number of job seekers. The total job supply refers to the deduplicated count of all unique job openings in the corresponding industry within the target region that are marked as "open".

[0075] Step 1022: Standardize the multi-dimensional features and organize the standardized features in a structured manner according to industry type to obtain the demand feature set.

[0076] In step 1022, since the units and dimensions of the above five features are different, this embodiment of the invention can use methods such as min-max standardization or Z-score standardization to transform the values ​​of all features to the same scale. After standardization, the five standardized feature values ​​corresponding to each industry are combined into an ordered feature vector with industry as the basic unit. Finally, the feature vectors of all industries are integrated together to form a "region-industry feature table" with a clear structure that can be directly read and analyzed by the algorithm, i.e., the demand feature set.

[0077] This invention transforms raw, disorganized recruitment information into a unified and comparable set of structured data indicators, thereby converting regional, qualitative industry employment trends into quantitative, machine-understandable feature expressions, thus laying a solid data foundation for subsequent intelligent analysis of regional talent market conditions.

[0078] S103. Input the basic features and the set of demand features into a classification model built on support vector machine to obtain the preference data of each job seeker.

[0079] Support Vector Machine (SVM) is a machine learning model whose core idea is to find an optimal decision boundary in a high-dimensional feature space to classify data into different categories.

[0080] In one specific implementation, such as Figure 2 As shown, step S103 includes:

[0081] Step 1031: Based on trajectory data within a historical preset period, determine the frequency of job seekers' visits to different industries and the duration of their stay, in order to generate a preference vector for job seekers.

[0082] Here, the preference vector refers to a list of values, where each position corresponds to a specific industry, and the value at that position represents the degree of interest a job seeker has in that industry based on their historical behavior.

[0083] In step 1031, a preset historical period is first selected, such as the period of the most recent large-scale job fair or within the past month. Then, all trajectory data of job seekers within this period are analyzed, and for each industry, the total number of times job seekers visited the industry's booths and the cumulative dwell time at each booth are counted. Next, the frequency and duration are normalized, and the two normalized values ​​are added according to preset weights. Finally, an interest score between 0 and 1 is calculated for each industry. The interest scores of all industries are arranged in a fixed order to form the job seeker's preference vector. For example, the higher the value at the i-th position in the vector, the higher the job seeker's attention to the i-th industry. This preference vector is a numerical vector based on user behavior history, quantifying their interest in various industries. It transforms users' explicit behaviors (visits, dwell times) in the talent market into implicit interest signals. Of course, in this embodiment, this step can also obtain the job seeker's preference vector through the user's click-through behavior and dwell time on the online job search platform under this system architecture.

[0084] For example, the target area contains three industries: Internet, Finance, and Manufacturing, corresponding to the 1st, 2nd, and 3rd positions of the vector, respectively. Statistics show that job seekers visited the Internet booth 3 times for a total of 60 minutes at the most recent job fair, visited the Finance booth once for 15 minutes, and did not visit the Manufacturing booth. The frequencies are then normalized: Internet = 3 / (3+1+0) = 0.75, Finance = 0.25, and Manufacturing = 0. The durations are then normalized: Internet = 60 / (60+15+0) = 0.8, Finance = 0.2, and Manufacturing = 0. Assuming the frequency and duration weights are each 0.5, the final preference vector is calculated as: [0.775, 0.225, 0].

[0085] Step 1032: Use the preference vector to perform importance weighting on the demand feature set to obtain the modulation feature set, and fuse all features in the modulation feature set with the basic features to obtain composite features.

[0086] The modulated feature set refers to the new feature set obtained by weighting and adjusting the features of each industry using a preference vector based on the original demand feature set. In other words, the modulated feature set is obtained by multiplying each industry feature in the demand feature set element by the interest score of that industry in the job seeker's preference vector. The composite feature refers to the comprehensive feature vector formed by combining the "basic features" representing individual abilities with the "modulated features" reflecting the weighted market conditions.

[0087] In step 1032, for the multi-dimensional feature vector of each industry in the demand feature set, the feature value of each dimension is multiplied by the interest score of that industry in the preference vector. This enhances the features of industries with high interest and weakens the features of industries with low interest, thus forming a modulated feature set that is more in line with the job seeker's perspective. Then, the job seeker's basic feature vector is concatenated with the feature vectors of all industries in the modulated feature set in turn to form a longer composite feature vector that integrates personal attributes and personalized market insights.

[0088] For example, the demand feature vectors for the internet, finance, and manufacturing industries are [100, 0.8, 20000], [50, 0.6, 18000], and [80, 0.5, 15000], respectively, representing the number of job openings, the average skill matching degree, and the average salary (i.e., a three-dimensional dataset). Then, the preference vector [0.775, 0.225, 0] is used for modulation to obtain the following modulated features: the internet feature is [77.5, 0.62, 15500]; the finance feature becomes [11.25, 0.135, 4050]; and the manufacturing feature becomes [0, 0, 0]. Assuming the job seeker's basic features are [Master's degree, Computer Science, 3 years] = [1, 5, 3], concatenating all the above values ​​yields the final composite feature: [1, 5, 3, 77.5, 0.62, 15500, 11.25, 0.135, 4050, 0, 0, 0].

[0089] Step 1033: Using a support vector machine, perform boundary learning and partitioning processing on the composite features in a high-dimensional space to generate classification results and corresponding initial confidence scores.

[0090] The classification result refers to the industry category that the job seeker is most likely to belong to, as predicted by the model; the initial confidence score is a preliminary probability score given by the model regarding the reliability of the classification result.

[0091] For example, the structure of a support vector machine aims to find an optimal decision hyperplane to maximize the margin between samples of different classes. Its core mathematical form can be expressed as solving the following optimization problem: Its constraints are ,in, Let w be the feature vector of the i-th training sample, w be the normal vector of the hyperplane, and b be the bias term. It is a penalty parameter used to balance model complexity and classification error. These are slack variables, allowing for a certain degree of misclassification. It is the class label of the i-th training sample. It is a nonlinear mapping function that maps the original features to a high-dimensional space. This application selects the radial basis kernel function: ,in, It is the feature vector of the j-th training sample. It is a sample and In the high-dimensional feature space, similarity is expressed as o, which is a kernel parameter that controls the range of influence of a single sample on the decision boundary.

[0092] The support vector machine is pre-trained. The training process includes: First, preparing training samples. The features of each sample are formed by concatenating the "basic features" of a historical job seeker with the feature vector of the target industry in the corresponding "demand feature set". The label of the sample is the industry category in which the job seeker was finally employed or applied for. Then, a radial basis function kernel function is used, and grid search and cross-validation are used to determine the optimal penalty coefficient and kernel parameters. Finally, the sequential minimum optimization algorithm is used to solve for the model parameters to complete the model training.

[0093] In step 1033, the composite feature vector is input into a pre-trained support vector machine classification model. This model has learned a large number of samples of "composite features and industry categories" during the training phase, thereby determining the optimal decision boundary for classifying different industry categories. When a new composite feature is input, the model calculates the relative position of the feature to the decision boundary and determines its industry category, thereby outputting the category label as the classification result. At the same time, the model calculates an initial confidence score between 0 and 1 based on factors such as the distance from the feature point to the decision boundary. The higher the score, the more confident the model is in the classification judgment.

[0094] For example, a Support Vector Machine (SVM) model categorizes industries into "Internet," "Finance," and "Manufacturing," etc. After inputting the composite feature vector from the example above, the model calculates that the feature is closest to the decision boundary for "Internet," thus outputting the classification result as "Internet." Simultaneously, the model calculates that the functional margin from this feature point to the "Internet" decision boundary is relatively large, so after mapping using the Sigmoid function, an initial confidence score of 0.85 is obtained. The above process maps to a probability value between 0 and 1 using the Sigmoid function, serving as the initial confidence score. The Sigmoid function mapping is a commonly used technique and will not be elaborated further.

[0095] Step 1034: Based on the inter-industry relationships within the target administrative region, construct an inter-industry relationship graph, and apply influence diffusion to the classification results on the relationship graph to enhance the inter-industry relationships.

[0096] Among them, the association graph refers to a network structure that uses nodes and edges to represent entities and their relationships. The nodes in the association graph represent industries, and the edges represent the associations between industries. Influence diffusion refers to a computational process that simulates the spread of information or effects on a network. The degree of association is a numerical value used to quantify the additional influence intensity that an industry node gains due to its association with other nodes after the diffusion process is completed.

[0097] This association graph is used to model the multidimensional relationships between industries within a target region. The graph data structure is G=(V,E,W). V: Set of nodes, each node representing an industry. E: Set of edges, representing relationships between industries. W: Edge weight matrix, w... jk ∈W represents the correlation strength between industry j and industry k.

[0098] Step 1034 involves constructing an inter-industry correlation graph based on the inter-industry relationships within the target administrative region, and then applying influence diffusion to the classification results on the correlation graph to enhance the correlation between industries. This specifically includes the following steps:

[0099] Step a1: Based on the upstream and downstream dependencies between industries within the target administrative region, construct the first graph edges; based on the degree of overlap between core skills, construct the second graph edges; and based on the clustering density of the geographical locations of the industries, construct the third graph edges.

[0100] In step a1, based on the determination that the products of industry A are the raw materials of industry B based on regional industrial economic data, a first graph edge representing upstream and downstream dependence is constructed; based on the analysis of job descriptions in the skill dictionary, the Jaccard similarity of the skill sets required by different industries is calculated, and if it exceeds the threshold, a second graph edge representing skill affinity is constructed; based on the enterprise registration location information, the overlap of the distribution of enterprises in different industries in each block is calculated, and if it exceeds the threshold, a third graph edge representing geographical clustering is constructed.

[0101] Specifically, in order to construct a relationship graph with clear weights, it is necessary to quantify the "upstream and downstream dependencies", the "overlap between core skills" and the "cluster density of geographical locations";

[0102] For upstream and downstream dependencies, a direct consumption coefficient is used to quantify them. This coefficient reflects the product flow relationship between industries; the larger the coefficient, the stronger the supply effect of the upstream industry on the downstream industry. Then, based on regional input-output table data, the direct consumption coefficient of industry i on industry j is calculated. Its formula is: ,in, This represents the consumption of product i by industry j in the production of industry j. Let j be the total output of industry j. Greater than the preset threshold If the value is 0.05, then a directed edge is established from upstream industry i to downstream industry j;

[0103] To quantify the overlap between core skills, the cosine similarity of the feature vectors corresponding to the core skills is used. This cosine similarity reflects the similarity of talent skill requirements across industries. Then, based on the skill requirement feature vectors of each industry position, the cosine similarity between the feature vectors corresponding to industry i and industry j is calculated. If the value is greater than the preset threshold If the value is 0.2, then an undirected edge is established between the two industry nodes;

[0104] For the clustering density of geographical locations, the correlation coefficient of the spatial density distribution obtained by kernel density estimation is used to quantify it. This coefficient reflects the degree of clustering and co-occurrence of industries in geographical space. Then, based on the latitude and longitude information of the company's registered address, the spatial density distribution of each industry is calculated using the kernel density estimation method, and then the Pearson correlation coefficient between the density distributions of different industries is calculated. ,like Greater than the preset threshold For example, 0.3;

[0105] Finally, the overall weight of each edge in the association graph is determined by the above three pieces of information according to preset coefficients. The weighted fusion yielded the following, These coefficients can be optimized and determined through effect experiments on historical data. Specifically, different weight combinations can be tried on historical data through grid search, with the goal of maximizing the offline evaluation index of recommendation effect, and the optimal weight combination can be selected. The typical initial value is [0.4, 0.3, 0.3], which means that the three relationships have similar importance in industry relationships, but the upstream and downstream dependence relationship has a slightly higher weight.

[0106] For example, if the target region has three industries: Internet (I), Finance (F), and Manufacturing (M), and analysis shows that the Internet is an upstream supplier of Manufacturing, then a directed edge is constructed from the Internet node to the Manufacturing node with a weight of 0.8. Subsequently, since Finance and the Internet have 30% overlap in core skills such as data analysis and risk control models, exceeding the threshold of 20%, an undirected edge is constructed connecting the Finance node and the Internet node with a weight of 0.5. Then, since Manufacturing and Finance have 25% overlap in enterprise cooperation or business in the field of supply chain finance, exceeding the threshold of 20%, an undirected edge is constructed connecting the Manufacturing node and the Finance node with a weight of 0.4.

[0107] Step a2: Merge the edges of the first graph, the second graph, and the third graph to generate an inter-industry association graph.

[0108] For example, all the constructed edges can be merged to generate a mixed directed and undirected industry association graph containing three nodes and three edges.

[0109] Step a3: Based on the classification results, initialize the state of the nodes in the association graph to obtain the initial state vector.

[0110] In step a3, the state value of the node corresponding to the industry that the classification result hits is set to 1, and the state value of the other nodes is set to 0, thus forming an initial state vector; for example, if the classification result is "Internet", then the corresponding node is I, and the initial state vector is: I=1, F=0, M=0, that is, [1, 0, 0].

[0111] Step a4: Using a graph attention network, the initial state vector is propagated and gated in the association graph to obtain the target state vector, and the target state vector is decoded into a numerical form of association degree.

[0112] The graph attention network employs a two-layer structure. The first layer is the graph attention layer, which is responsible for aggregating neighborhood information and updating node features. The second layer is the output layer, which maps the updated node features to a correlation scalar value.

[0113] The network with the above structure is trained as follows: The first layer maps node features to a 64-dimensional hidden layer using four attention heads and LeakyReLU activation, and concatenates the outputs of the four heads to form a 256-dimensional intermediate feature; the second layer uses a single attention head to map the 256-dimensional intermediate feature to a 1-dimensional output, and its training task is semi-supervised node regression, using some known industry association strengths as labels, and its loss function is mean squared error; then it is trained using an Adam optimizer with a learning rate of 0.005 and a weight decay of 5e-4.

[0114] In step a4, a graph attention network is used to perform multi-round iterative message propagation on the association graph. In each round, each node aggregates information from all its neighboring nodes, and the aggregation weight is dynamically calculated by the attention mechanism, paying particular attention to neighbors connected by strong association edges. The aggregated information is then combined with the node's state from the previous round through a gating fusion unit to generate a new state. This process is repeated until the node's state changes tend to stabilize, resulting in the final target state vector. Decoding the state value of each node in this vector yields the association degree obtained by the industry due to network association.

[0115] For example, the initial state vector [1, 0, 0] and the association graph are input into a two-layer graph attention network. After computation in the first layer, the node state is updated to [0.8, 0.4, 0.5]. Then, it is further propagated and merged in the second layer, and finally the target state vector converges to [0.9, 0.5, 0.6]. Finally, this vector is decoded to obtain the association degree of each industry: Internet 0.9, Finance 0.5, Manufacturing 0.6.

[0116] Step 1035: Based on the correlation degree, propagate and update the initial confidence degree under the constraints of the preset graph structure to generate the final confidence degree.

[0117] It should be noted that the three industries (I, F, M) in the case study represent V. The three edges constructed in the case study (I→M, IF, MF) represent E. The weight assigned to each edge in the case study (0.8, 0.5, 0.4) represents W. Then, starting from the initial state [1, 0, 0], after multiple rounds of message propagation and aggregation by the Graph Attention Network (GAT), the final result is [0.9, 0.5, 0.6]. However, the values ​​obtained after decoding the target state vector in the above case study (Internet 0.9, Finance 0.5, Manufacturing 0.6) are precisely the correlation degree that "quantifies the additional influence strength gained by the industry due to the association".

[0118] Among them, graph structure constraint refers to a rule that uses the topological structure of the association graph to guide numerical updates. In this embodiment, the content of graph structure constraint is not specifically limited and can be set according to the actual situation. The final confidence score refers to the reliability score obtained after the initial confidence score has been propagated and fused under the graph structure constraint and calibrated by network information.

[0119] The first graph edge uses upstream and downstream dependencies to quantify industrial linkages. The second graph edge uses the overlap of core skills to quantify talent / technology linkages. The third graph edge uses the overlap of geographical distribution to quantify spatial clustering linkages. In the above scheme, the "edge set (E)" is a mathematical set that contains all the edges in the graph. It is a holistic concept.

[0120] The "first, second, and third graph edges" in the subsequent solution refer to three different methods or data sources for constructing the edge set (E). This does not mean that the final graph has only three edges, but rather that the edge definition can originate from three relational dimensions. The final fused relational graph's edge set E is composed of all edges constructed using these three methods.

[0121] Using Case 1 as an example: The target region comprises three industries: Internet (I), Finance (F), and Manufacturing (M). The "First Graph Edge" (upstream / downstream) method constructs the edge I→M. The "Second Graph Edge" (Skills) method constructs the edge IF. Through business cooperation analysis (which can be viewed as another type of association), the edge MF is constructed. Ultimately, the edge set E = {I→M, IF, MF}. Here, the "First Graph Edge Method" contributes one edge, and the "Second Graph Edge Method" contributes the other.

[0122] In step 1035, the initial confidence score and the correlation between industries are used as input. Then, a confidence score propagation rule based on the correlation graph is defined. For example, the confidence score of a node is diffused to its neighboring nodes with high correlation scores in a certain proportion. Then, through multiple rounds of iterative updates, the confidence score of each industry node is no longer isolated, but integrates its own initial value and the support or weakening information from other nodes in the correlation network. When the propagation process reaches a stable state, the confidence score held by each node is the final confidence score. The stable state means that the difference between two adjacent values ​​is less than a preset threshold. The threshold is set to balance the accuracy requirements and computational efficiency: when the vector change is less than this value, the confidence score distribution can be considered to have reached a stable state, and further iteration has little impact on the result. This empirical value is widely used in the field of numerical optimization, which can effectively ensure the sufficiency of convergence while avoiding unnecessary computational overhead.

[0123] For example, the propagation rule is defined as follows: each node allocates 20% of its own confidence to its neighbors according to the correlation ratio. In the first round, the initial confidence of the Internet is 0.85, so it allocates 0.17; the financial allocation is 0.17 × (0.5 / (0.5+0.6)) = 0.077; and the manufacturing allocation is 0.17 × (0.6 / (0.5+0.6)) = 0.093. In the second round, the initial confidence of the financial allocation is 0.077, so it allocates 0.015; and the Internet allocation is 0.015 × ( 0.9 / (0.9+0.6))=0.009, manufacturing allocation 0.015×(0.6 / (0.9+0.6))=0.006, the initial confidence level of manufacturing is 0.093, allocation is 0.019, internet allocation is 0.019×(0.9 / (0.9+0.5))=0.012, financial allocation is 0.019×(0.5 / (0.9+0.5))=0.007, and after several rounds of iteration and stabilization, the final confidence level of internet nodes may converge to 0.82.

[0124] The final confidence level mentioned above can also be calculated as follows: Initialization: Let the confidence level of node i be in the t-th iteration. .make =Initial confidence score (the initial confidence score of the node corresponding to the classification result is the SVM output, and the confidence score of other nodes is 0). Define the propagation matrix T: construct a transition probability matrix based on the edge weights of the association graph G (the association score can be used as the edge weight or calculated from it). ,in This represents the proportion of confidence propagation from node i to node j, satisfying... For example, weighting can be done by relevance: in It is the edge weight (association degree) between nodes i and j.

[0125] Iterative updates: Where α∈[0,1] is the damping factor, representing the degree of trust in neighbor information; This represents the confidence level aggregated from all neighbors j. The damping factor is set based on the experience and practice of the classic PageRank algorithm. This value can ensure that the confidence level is effectively diffused in the network while avoiding oscillation or non-convergence of the results due to over-propagation, and achieve a good balance between the "inheritance" and "stability" of information propagation.

[0126] Convergence criterion: When all nodes satisfy The iteration stops when ε (where ε is a preset threshold). This represents the stability confidence level of node i, which serves as its industry confidence level. To generate trend data: all industries and their corresponding stability confidence levels are sorted in descending order of confidence level, forming a trend data list.

[0127] Step 1036: Integrate the classification results with the final confidence level to form the tendency data.

[0128] In step 1036, the classification result representing the most likely direction and the final confidence level representing the reliability of that direction are encapsulated into a structured data object. This object clarifies the job seeker's core industry preference and provides a quantitative confidence assessment of this preference. For example, the final generated preference data is: {Classification result: "Internet", final confidence level: 0.82}.

[0129] This invention quantifies the discrete behavioral trajectory data of job seekers into a structured, computable representation of their interests and preferences, providing a weighted basis from a personal perspective for subsequent integration of market information. Then, personal behavioral preferences are used as a "lens" to examine regional market data, and static personal attributes are integrated with dynamically weighted market dynamics to generate a unified feature representation that combines personalization and contextual information, thus laying the foundation for accurate classification. Finally, mature classification algorithms are used to automatically discriminate the complex features after integration, resulting in a preliminary, quantitative judgment of industry preferences and its reliability estimate.

[0130] The classification results are then validated and disseminated across the entire regional industry association network, moving beyond single-node judgments. This allows other industries closely related to the core industry to also receive a degree of association assessment, thereby capturing potential cross-industry employment opportunities. Subsequently, the industry association network is used to smooth and enhance the confidence level of the initial model, ensuring that the confidence assessment is not only based on single-point features but also considers the group information of the industry ecosystem. This results in a more robust and reliable measure that better reflects real-world industry relationships. Finally, structured trend data is produced, providing not only directional conclusions about "what it is" but also deterministic measurements of "how reliable it is," offering more complete and robust support for subsequent screening and matching decisions.

[0131] S104. Based on the trajectory data, the tendency data is adapted and filtered to obtain the target industry corresponding to the job seeker.

[0132] In one specific implementation, step S104 includes:

[0133] Step 1041: Fuse the preference vector corresponding to the trajectory data with the classification results and final confidence scores in the tendency data to obtain the comprehensive fit of each candidate industry.

[0134] Among them, the overall fit is a comprehensive evaluation index used to quantify the overall matching degree between a candidate industry and a job seeker. It integrates the following three aspects of information: the job seeker's behavioral preferences, the industry tendency of personal ability and market matching as analyzed by the model, and its goal is to find industry directions that simultaneously satisfy personal "interest", "ability" and "opportunity".

[0135] In step 1041, for each candidate industry in the candidate industry set, the interest score of that industry is extracted from the preference vector, the industry is confirmed to be a classification result from the tendency data, and its corresponding final confidence score is extracted. Then, a single score is calculated according to the preset fusion rule, which is the comprehensive fit of the candidate industry. For example, the fusion rule can refer to a weighted summation method, where the weights are set based on training with historical data. Specifically, a logistic regression model can be used, with interest and final confidence scores as features and whether the job seeker selects the industry as a label. The feature weights obtained from the training are the weights corresponding to interest and final confidence scores, respectively.

[0136] For example, if the propensity data indicates a classification result of "Internet" with a final confidence score of 0.82, and the preference vector shows job seekers' interest in "Internet," "Finance," and "Manufacturing" at 0.775, 0.225, and 0 respectively, then the candidate industry is "Internet." The fusion rule is then set as: Overall Fit = 0.4 × Interest Score + 0.6 × Final Confidence Score. For the "Internet" industry, its Overall Fit = 0.4 × 0.775 + 0.6 × 0.82 = 0.802. This example is only provided to facilitate understanding of the fusion process for a single candidate industry; however, in practical applications, the candidate industry set typically contains multiple candidate industries.

[0137] Step 1042: Based on the comprehensive fit, select the candidate industry with the highest fit value from the candidate industries as the target industry.

[0138] In step 1042, the overall suitability of all candidate industries is compared, and the industry with the highest value is directly selected as the final target industry, which will be used in the subsequent job-level matching process.

[0139] The embodiments of the present invention ensure that the selected target industry is not only based on theoretical analysis but also verified by the actual behavior of job seekers by integrating and adjudicating objective behavioral preferences and subjective inclination judgments, thereby improving the reliability and personalization of industry direction selection.

[0140] S105. Using the K-nearest neighbor algorithm, the similarity between the job characteristics and job seeker characteristics corresponding to the target industry is calculated. Based on the similarity calculation results, all target jobs with similarity values ​​greater than a preset threshold are sorted to generate a job recommendation list for job seekers.

[0141] In one specific implementation, step S105 includes:

[0142] Step 1051: Integrate the basic features, the preference vector corresponding to the trajectory data within the historical preset period, and the final confidence level to generate a core vector representing the job seeker's characteristics. Integrate the text features of the job and the correlation between various industries to generate a context vector representing the job's characteristics.

[0143] Among them, the core vector refers to the feature vector used to comprehensively represent the job seeker's personal status. This vector integrates static abilities, dynamic interests, and network-optimized tendency credibility. The text features are obtained by natural language processing of recruitment information corresponding to the target industry. Specifically, key information items can be extracted from texts such as job descriptions and job requirements and quantitatively represented. For example, bag-of-words models or word vector techniques can be used to transform unstructured job description text into structured feature vectors. The context vector refers to the feature vector used to represent the job and its industry background. This vector not only contains the descriptive information of the job itself, but also incorporates the correlation strength information of the industry to which the job belongs in the entire regional industry network, providing a richer context for the job.

[0144] In step 1051, the job seeker's basic characteristics, such as education and major codes, are concatenated or weighted and fused with the preference vector extracted from the trajectory data and the final confidence level obtained from the tendency data to form a high-dimensional unified vector, which is used to comprehensively encode the job seeker's personal profile.

[0145] Simultaneously, for each position within the target industry, a corresponding context vector is constructed. This vector consists of two parts: one part is the job description text, such as job requirements and responsibilities, after feature extraction, such as text features obtained after word vectorization; the other part is the dynamic correlation degree of the industry to which the position belongs, calculated in step 1034 through influence diffusion based on the current job seekers' preference data. Subsequently, these two parts of information are merged so that the features of each position not only reflect its own requirements, but also carry the position information of its industry in the regional ecosystem.

[0146] For example, a job seeker's basic feature vector is [1, 5, 3], the interest in the target industry "Internet" in the preference vector is 0.775, and the final confidence level is 0.82, resulting in the core vector [1, 5, 3, 0.775, 0.82]. Subsequently, for a "Java Development Engineer" position within the "Internet" industry, its text feature vector is [0, 0.9, 0.2, ...]. Combined with the job seeker's dynamic correlation degree in the Internet industry calculated in step 1034 based on the Internet category, for example, 0.9, the context vector of the position is [0, 0.9, 0.2, ..., 0.9].

[0147] Step 1052: Using a graph-augmented metric learning method, a metric function is learned by using the inter-industry relationship graph as a constraint.

[0148] Here, the metric function refers to a mathematical function used to calculate the "distance" or "similarity" between two vectors in a high-dimensional space. In this embodiment of the invention, the expression of the metric function is not specifically limited and can be set according to actual conditions. The graph-enhanced metric learning method refers to transforming the topological structure of the associated graph into a neighborhood constraint on the feature representation by introducing a graph attention mechanism, and constructing training sample pairs based on this constraint using a contrastive learning strategy, thereby generating a structural contrastive loss that can quantify the consistency of the graph structure. Finally, this loss is used as a regularization term and fused with the main loss to jointly drive the parameter optimization of the metric function, so that the learned metric function can not only distinguish individual feature differences, but also encode industry-related knowledge into the distance metric in the metric space.

[0149] For example, the core of graph-augmented metric learning methods is to obtain a metric function, which is typically implemented by a neural network to map the input feature vector to a metric space. In this application, the neural network uses two fully connected layers. The first layer maps the input vector to 128 dimensions and uses the ReLU activation function; the second layer maps to 64 dimensions and also uses the ReLU activation function; finally, the output is to a 32-dimensional metric space. The parameters θ of the network include the weight matrix and bias vector of the two fully connected layers.

[0150] The training process for the network with the above structure is as follows: First, construct training sample pairs, which include job pairs from strongly related industries as positive sample pairs and job pairs from weakly related or unrelated industries as negative sample pairs; second, define a metric function to map the context vector of the job to a metric space; then, design a hybrid loss function for optimization; finally, use the gradient descent algorithm to optimize the parameter θ, minimize the hybrid loss function, and thus obtain the trained metric function.

[0151] Step 1052 employs a graph-augmented metric learning method, using the inter-industry relationship graph as constraints, to learn and output a metric function, specifically including the following steps:

[0152] Step b1: Extract the graph topology of the inter-industry association graph and use the graph topology as graph topology constraints.

[0153] In step b1, the graph topology, namely the connection relationships and weights of nodes and edges, is extracted from the constructed industry association graph as constraints for subsequent learning.

[0154] Step b2: Using a graph attention mechanism, the context vector is aggregated with neighborhood information based on the graph topological constraints to generate a perceptual feature set.

[0155] In step b2, the context vector of each job is used as the initial feature of its industry node. Under the constraints of the graph topology, each node aggregates the feature information of its neighboring nodes. The aggregation weight is dynamically calculated by the attention mechanism based on the edge weight and node features. After aggregation, the context vector of each job is updated to a "perceptual feature" that integrates its industry neighborhood information.

[0156] Step b3: Using a contrastive learning strategy, construct positive and negative sample pairs in the perceptual feature set according to the graph topological constraints.

[0157] In step b3, a contrastive learning strategy is used to construct training sample pairs. This strategy can define what is "similar" and "dissimilar" based on graph topological constraints. That is, two positions belonging to strongly related industries are considered positive sample pairs, while two positions belonging to weakly related or unrelated industries are constructed as negative sample pairs.

[0158] Step b4: Calculate the similarity reward between the positive sample pairs and the similarity penalty between the negative sample pairs to form the structural contrast loss.

[0159] In step b4, the similarity between the two perceptual feature vectors of each positive sample pair is calculated. This similarity can be cosine similarity, and it is a value between -1 and 1, with the value closer to 1 indicating greater similarity. Since ideally, all positive sample pairs should have the highest possible similarity, close to 1, the actual similarity value of each positive sample pair is compared to the target value of 1. The difference reflects the degree to which the positive sample pair is "not similar enough," and this difference is included in the loss and is the part that needs to be minimized. This is essentially a "penalty" for failing to achieve high similarity. Although the initial intention is to reward high similarity, within the framework of the loss function, it is usually achieved by penalizing low similarity.

[0160] Meanwhile, the similarity of each negative sample pair is calculated. Since it is ideal to want negative sample pairs to be as dissimilar as possible, i.e., the similarity value should be as low as possible, close to 0 or negative, the actual similarity value of each negative sample pair is compared with the target value of 0. The difference reflects the degree to which the negative sample pair is "not distant enough". This difference is also included in the loss as another penalty that needs to be minimized.

[0161] Finally, the penalties for all positive sample pairs due to "insufficient similarity" and the penalties for all negative sample pairs due to "insufficient distance" are weighted and summed according to preset weights to form a single scalar value, namely the structural contrast loss. This loss value measures the extent to which the current perceived features violate the similarity relationship defined by the graph topological constraints. That is, the smaller the loss value, the more similar the positive sample pairs are and the less similar the negative sample pairs are, meaning that the feature representation is more consistent with the industry association structure.

[0162] For example, based on graph topological constraints, two positive sample pairs are constructed: positive pair 1 and positive pair 2; and two negative sample pairs: negative pair 1 and negative pair 2. Then, the cosine similarity of each pair of perceptual features is calculated as follows: similarity of positive pair 1 = 0.85, similarity of positive pair 2 = 0.75; similarity of negative pair 1 = 0.40, similarity of negative pair 2 = 0.30. Subsequently, the target similarity of the positive sample pairs is set to 1, and the target similarity of the negative sample pairs is set to 0, with preset weight coefficients: the weight of the penalty term for the positive sample pairs is λ1 = 0.7, and the weight of the penalty term for the negative sample pairs is λ2 = 0.3. These specific weight values ​​are determined as follows: during model training, λ1 and λ2 are treated as hyperparameters, and a grid search method is used on the validation set to optimize the overall recommendation performance of the model.

[0163] Then, for each positive sample pair, the penalty for "not being similar enough" is calculated as follows: penalty for positive pair 1 = 1 - 0.85 = 0.15; penalty for positive pair 2 = 1 - 0.75 = 0.25; for each negative sample pair, the penalty for "not being distant enough" is calculated as follows: penalty for negative pair 1 = 0.40 - 0 = 0.40; penalty for negative pair 2 = 0.30 - 0 = 0.30.

[0164] Next, all penalty terms are weighted and summed according to preset weights to calculate the structural contrast loss: Structural Contrast Loss = (λ1 × Penalty for positive pair 1) + (λ1 × Penalty for positive pair 2) + (λ2 × Penalty for negative pair 1) + (λ2 × Penalty for negative pair 2) = (0.7 × 0.15) + (0.7 × 0.25) + (0.3 × 0.40) + (0.3 × 0.30) = 0.49. This loss value of 0.49 comprehensively reflects the degree of deviation between the current perceptual features and the graph topological constraints. Further optimization to reduce this loss will drive the model to adjust its parameters, making the similarity of positive sample pairs approach 1 and the similarity of negative sample pairs approach 0, thereby achieving alignment between the feature representation and the industry-related structure.

[0165] Step b5: Use the structural contrast loss as a regularization term, and fuse the regularization term with the main loss function of metric learning to form a hybrid loss function.

[0166] In step b5, the aforementioned structural contrast loss is used as a regularization term and weighted and fused with the main loss function of metric learning itself. For example, the main loss function... ,in, Let represent the main loss function, where, for a given anchor job feature , . Select one job position from the same industry (or strongly related industry) as a positive sample. Select one job from its different industries (or weakly related industries) as a negative sample. Together they form a training triplet; This represents the summation of all N constructed triplet samples in a training batch. This represents the feature vector of the anchor sample in the i-th triplet. Indicates and Feature vectors of positive samples belonging to the same category or that should be similar. Indicates and Feature vectors of negative samples belonging to different categories or that should be dissimilar L2 norm is used to calculate the Euclidean distance between two vectors. , Indicates boundary value hyperparameters, such as >0 is used to enforce the minimum distance difference between positive and negative sample pairs. This represents the hinge function, ensuring it is zero when the distance difference already meets the requirements, avoiding unnecessary optimization, thus forming a hybrid loss function that simultaneously optimizes feature discriminative power and graph structure consistency. For example, ,in, This represents the mixed loss function, which is the final objective function that needs to be minimized. This represents the hyperparameters of the weighting coefficients, such as... ≥0 is used to control the proportion of structural contrast loss in the total loss. This represents the structural contrast loss.

[0167] Step b6: Optimize with the goal of minimizing the hybrid loss function, update the internal parameters of the preset function, and use the function corresponding to the updated parameters as the metric function.

[0168] In step b6, with the goal of minimizing this mixed loss function, a pre-defined metric function model, such as the internal parameters of a projection function based on a neural network, is iteratively updated using optimization algorithms such as gradient descent. The function with the new parameters obtained after the training process converges is the learned "graph-augmented metric function," for example, z = f(x; Θ), where z represents the output vector, i.e., the mapping representation of the input x in the metric space, and Θ represents the parameter set of the metric function. This is achieved by minimizing the mixed loss function. The optimized result is that x represents the input feature vector, which can be either the job seeker's core vector or the job context vector. This represents the metric function, which maps the input features to the optimized metric space.

[0169] Step 1053: Using the metric function, map the core vector and the context vector to a unified metric space, and in the metric space, use the K-nearest neighbor algorithm to calculate the similarity between the core vector and the context vector.

[0170] In step 1053, a metric function is used to map the job seeker's core vector and the context vectors of all positions in the target industry into a unified, optimized metric space. In this space, the distance between vectors directly reflects the overall matching degree after graph knowledge enhancement.

[0171] Subsequently, the K-nearest neighbor algorithm is used to calculate the similarity between the job seeker core vector and each job vector in the metric space. For example, the inverse or cosine similarity of Euclidean distance is used. Then, all similarity values ​​are compared with a preset threshold to filter out all jobs with similarity higher than the threshold as candidate target jobs. The preset threshold is determined by calculating the similarity of all "job seeker and job" pairs on the historical dataset and calculating the recall rate under different thresholds. The similarity value that makes the recall rate reach a predetermined level, such as 85%, is selected as the preset threshold.

[0172] Finally, these candidate job positions are sorted in descending order according to their similarity scores with job seekers, generating an ordered job recommendation list as the final output of this method.

[0173] The embodiments of the present invention achieve a precise conversion from macro-industry screening to micro-job matching, significantly improving the fit between the recommendation results and the regional market context and the individual's overall situation.

[0174] Example 2

[0175] Figure 3 This is a schematic diagram illustrating a specific implementation of a job recommendation system based on big data analysis, as provided in Embodiment 2 of the present invention. (Refer to...) Figure 3 The system may include:

[0176] The acquisition module 31 is used to acquire talent demand data of various industries within the target administrative region, basic characteristics of job seekers, and trajectory data of job seekers in the talent market.

[0177] Extraction module 32 is used to extract multi-dimensional features from the talent demand data to obtain a demand feature set.

[0178] Analysis module 33 is used to input the basic features and the set of demand features into a classification model built on support vector machine to obtain the preference data of each job seeker.

[0179] The filtering module 34 is used to adapt and filter the tendency data based on the trajectory data to obtain the target industry corresponding to the job seeker.

[0180] The calculation module 35 is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry using the K-nearest neighbor algorithm, and to sort all target jobs with similarity values ​​greater than a preset threshold based on the similarity calculation results, thereby generating a job recommendation list for job seekers.

[0181] The job recommendation system based on big data analysis in this embodiment of the invention is used to implement the aforementioned job recommendation method based on big data analysis. Therefore, the specific implementation of the job recommendation system based on big data analysis can be found in the embodiment section of the job recommendation method based on big data analysis above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0182] like Figure 4 As shown, this embodiment of the invention also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described big data analysis-based job recommendation methods.

[0183] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for recommending job positions based on big data analysis.

[0184] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0185] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the job recommendation method based on big data analysis.

[0186] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0187] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this solution includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which the embodiments of this solution pertain. The processor performs the various methods and processes described above. For example, the method embodiments of this solution can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).

[0188] The logic and / or steps represented in the flowchart or otherwise described herein may be specifically implemented in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0189] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A job recommendation method based on big data analysis, characterized in that, include: Acquire talent demand data for various industries within the target administrative region, basic characteristics of job seekers, and data on job seekers' movements in the talent market; Multidimensional feature extraction is performed on the talent demand data to obtain a demand feature set; The basic features and the demand feature set are input into a classification model built on a support vector machine to obtain the preference data of each job seeker, including: determining the frequency of visits and dwell time of job seekers to different industries based on trajectory data within a preset historical period to generate a preference vector of job seekers; using the preference vector to weight the demand feature set by importance to obtain a modulated feature set; fusing all features in the modulated feature set with the basic features to obtain a composite feature; using a support vector machine to perform boundary learning and partitioning processing in high-dimensional space on the composite feature to generate classification results and corresponding initial confidence scores; constructing an industry association graph based on the inter-industry relationships within the target administrative region, and spreading the influence of the classification results on the association graph to enhance the correlation between industries; propagating and updating the initial confidence score according to the correlation score under the constraints of a preset graph structure to generate a final confidence score; integrating the classification results and the final confidence score to form preference data; Based on the trajectory data, the tendency data is adapted and filtered to obtain the target industry corresponding to the job seeker; The K-nearest neighbor algorithm is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry. Based on the similarity calculation results, all target jobs with similarity values ​​greater than a preset threshold are sorted to generate a job recommendation list for job seekers.

2. The job recommendation method based on big data analysis according to claim 1, characterized in that, The process of constructing an inter-industry correlation diagram based on the inter-industry relationships within the target administrative region, and then applying influence diffusion to the classification results on the correlation diagram to enhance the correlation between industries, includes: Based on the upstream and downstream dependencies between industries within the target administrative region, a first graph edge is constructed; based on the degree of overlap between core skills, a second graph edge is constructed; and based on the clustering density of the geographical locations of each industry, a third graph edge is constructed. By merging the edges of the first graph, the second graph, and the third graph, an inter-industry association graph is generated. Based on the classification results, the nodes in the association graph are initialized to obtain an initial state vector. A graph attention network is used to propagate and gating the initial state vector on the association graph to obtain the target state vector, and the target state vector is decoded into a numerical form of association degree.

3. The job recommendation method based on big data analysis according to claim 1, characterized in that, The K-nearest neighbor algorithm is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry, including: By integrating the basic features, the preference vector corresponding to the trajectory data within the historical preset period, and the final confidence level, a core vector for representing job seeker characteristics is generated. By integrating the text features of the job and the correlation between various industries, a context vector for representing job characteristics is generated. A graph-augmented metric learning method is adopted, which uses the inter-industry relationship graph as a constraint to learn a metric function; The core vector and the context vector are mapped to a unified metric space using the metric function, and the similarity between the core vector and the context vector is calculated using the K-nearest neighbor algorithm in the metric space.

4. The job recommendation method based on big data analysis according to claim 3, characterized in that, The graph-augmented metric learning method, using an inter-industry association graph as a constraint, learns a metric function, including: Extract the graph topology structure of the inter-industry relationship graph, and use the graph topology structure as graph topology constraints; By using a graph attention mechanism, neighborhood information is aggregated from the context vector based on the graph topological constraints to generate a perceptual feature set. A contrastive learning strategy is employed to construct positive and negative sample pairs in the perceptual feature set based on the graph topological constraints. The similarity reward between the positive sample pairs and the similarity penalty between the negative sample pairs are calculated to form the structural contrast loss; The structural contrast loss is used as a regularization term, and the regularization term is fused with the main loss function of metric learning to form a hybrid loss function; The optimization aims to minimize the hybrid loss function. The internal parameters of the preset function are updated, and the function corresponding to the updated parameters is used as the metric function.

5. The job recommendation method based on big data analysis according to claim 1, characterized in that, The process of adapting and filtering the tendency data based on the trajectory data to obtain the target industry corresponding to the job seeker includes: The preference vector corresponding to the trajectory data is fused with the classification results and final confidence scores in the tendency data to obtain the comprehensive suitability of each candidate industry; Based on the overall fit, the candidate industry with the highest fit value is selected as the target industry from the candidate industries.

6. The job recommendation method based on big data analysis according to claim 1, characterized in that, The process of extracting multidimensional features from the talent demand data yields a demand feature set, including: Multi-dimensional features are extracted from the job information in the talent demand data to obtain multi-dimensional features for each industry. The multi-dimensional features include: a first feature corresponding to the number of jobs, a second feature corresponding to skill requirements, a third feature corresponding to salary levels, a fourth feature corresponding to growth trends, and a fifth feature corresponding to talent supply and demand surplus. The multi-dimensional features are standardized, and the standardized features are then structured and organized according to industry type to obtain a set of demand features.

7. A job recommendation system based on big data analysis, characterized in that, The process is implemented using the job recommendation method based on big data analysis according to any one of claims 1-6, including: The acquisition module is used to acquire talent demand data for various industries within the target administrative region, basic characteristics of job seekers, and trajectory data of job seekers in the talent market. The extraction module is used to extract multi-dimensional features from the talent demand data to obtain a demand feature set; The analysis module is used to input the basic features and the set of demand features into a classification model built on a support vector machine to obtain the preference data of each job seeker; The filtering module is used to adapt and filter the tendency data based on the trajectory data to obtain the target industry corresponding to the job seeker. The calculation module is used to calculate the similarity between the job characteristics and job seeker characteristics corresponding to the target industry using the K-nearest neighbor algorithm, and to sort all target jobs with similarity values ​​greater than a preset threshold based on the similarity calculation results, thereby generating a job recommendation list for job seekers.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the job recommendation method based on big data analysis as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the job recommendation method based on big data analysis as described in any one of claims 1 to 6.

Images