A method for establishing an intelligent job matching model and matching method based on deep reinforcement learning

Abstract

The present invention provides a method for establishing an intelligent job matching model based on deep reinforcement learning and a matching method, which belongs to the technical field of job matching based on computer data processing. Based on the constructed original data set, the present invention performs deep reinforcement learning training on the newly designed job matching model, and finally outputs the matching score between the job seeker's resume and the company's recruitment position; and makes corresponding job matching recommendations. The model includes a feature extraction module and a prediction output module. The feature extraction module performs three-dimensional feature extraction and fusion to obtain multi-granularity attention feature representation. The prediction output module adopts a deep Q network to model the job matching problem as a Markov decision process. The present invention can comprehensively capture the multi-dimensional features of job seekers and positions, greatly improving the richness and accuracy of feature representation. By designing a reasonable state space, action space and reward function, the model can continuously learn and optimize matching strategies in interaction.

Description

Technical Field

[0001] The present invention belongs to the technical field of job matching based on computer data processing, and in particular, relates to a method for establishing an intelligent job matching model based on deep reinforcement learning and a matching method. Background Art

[0002] With the rapid development of social economy, enterprises' demand for talents is becoming increasingly complex, and job seekers are also facing the challenge of how to choose suitable positions. Traditional job matching methods usually rely on manual screening and rule-based systems, which are difficult to cope with the complex matching of diverse job requirements and personal abilities. In recent years, deep reinforcement learning technology has made significant progress in the field of intelligent decision-making and recommendation systems, providing new solutions for job matching. By introducing deep reinforcement learning technology, dynamic learning and optimized matching of complex job requirements and job seeker characteristics can be achieved, thereby improving the accuracy and efficiency of matching. However, the existing technology has the following defects:

[0003] 1. Low matching accuracy: Traditional keyword- and rule-based matching systems have difficulty accurately understanding the deep matching relationship between job requirements and job seekers’ abilities;

[0004] 2. Lack of dynamic adjustment capabilities: Existing methods cannot be updated and optimized in real time based on changes in job requirements or job seekers’ feedback, resulting in delayed matching results;

[0005] 3. Lack of personalization: Many systems have difficulty in making personalized recommendations based on personal traits, career development preferences and other factors, and the matching results are relatively simple;

[0006] 4. Limited data utilization: Existing technologies are often unable to fully utilize multi-source data (such as career history, skill evaluation, etc.) for in-depth analysis, resulting in unsatisfactory matching results. Summary of the invention

[0007] In view of the above problems, the first aspect of the present invention proposes a method for establishing an intelligent job matching model based on deep reinforcement learning, which includes the following process:

[0008] S1, obtain job seekers’ resume information and company recruitment position information data to construct the original data set;

[0009] S2, preprocessing the original data set to obtain the model standard training data set;

[0010] S3, building an intelligent job matching model, including a feature extraction module and a prediction output module;

[0011] The feature extraction module includes a comprehensive feature extraction unit and a feature fusion unit; the comprehensive feature extraction unit includes a contextual semantic feature extraction network, a low-level semantic feature extraction network and a high-level semantic feature extraction network; the low-level semantic feature extraction network is used to extract feature keywords of job seeker resume information and company job recruitment information; the high-level semantic feature extraction network is used to obtain job seeker position change feature vectors and job seeker historical work content feature vectors, and perform vector splicing to obtain high-level features of job seeker resumes, and at the same time obtain high-level features of company recruitment positions; the feature fusion unit is used to fuse the extracted features to obtain multi-granularity attention feature representation;

[0012] The prediction output module uses a deep Q network to model the job matching problem as a Markov decision process, where the Q value is the expected return of recommending a job given the resume information; the final output is the matching score between the job seeker's resume and the company's recruitment position;

[0013] S4, conduct deep reinforcement learning training on the job matching model to obtain the final model.

[0014] Preferably, the preprocessing of the original data set in S2 includes data cleaning and data conversion;

[0015] The data cleaning first uses regular expressions to remove HTML tags and special characters, and unifies the date format to "YYYY-MM-DD". For missing values, different strategies are used: numerical missing values ​​are filled with mean or median, categorical missing values ​​are filled with mode, and text missing values ​​are filled with "Unknown".

[0016] The data conversion is to convert the unstructured text represented by the job seeker's resume information and the company's recruitment position matching information into a numerical representation that can be understood by a machine; wherein, the jieba library is used for Chinese word segmentation, and the NLTK library is used for English word segmentation.

[0017] Preferably, the preprocessing of the original data set in S2 also includes data label processing, including similarity calculation, comprehensive matching score calculation and label generation;

[0018] The similarity calculation: First, use TF-IDF to convert the text into a vector, and calculate the cosine similarity to get the overall text similarity ; Secondly, convert resume skills and job requirement skills into word vectors, calculate the average cosine similarity to get the skill matching degree ; Then, the empirical matching degree is calculated through numerical comparison and scoring function Finally, design the education level mapping, compare the difference between the resume education level and the job requirements to get the education level matching degree ;

[0019] The comprehensive matching score calculation: First, define the comprehensive matching score :

[0020] ;

[0021] in is the weight; then based on the score , the matching degree is divided into four levels: For mismatch, For partial matching, For a good match, is an excellent match; the confidence is calculated at the same time:

[0022] ;

[0023] Where threshold is the lower limit threshold of the matching degree;

[0024] The label generation: based on the comprehensive matching score calculation result, a label including the resume ID, position ID, matching score and confidence level is generated for each pair of job seeker resume information and company recruitment position information.

[0025] Preferably, the specific processing process of the low-level semantic feature extraction network is:

[0026] First, extract the low-level semantic features of the job seeker's resume , as follows:

[0027] Work experience keywords: extracted using TF-IDF method Work experience keywords in;

[0028] Educational qualification coding: One-hot coding is used to encode the educational qualification in the applicant’s resume;

[0029] Certificate keywords: extracted using TF-IDF method The certificate keywords in;

[0030] Skill keywords: extracted using the TF-IDF method Skill keywords in;

[0031] Input work experience keywords, education codes, certificate keywords, and skill keywords into BERT to obtain low-level features of job seekers’ resumes ;

[0032] Secondly, extract the low-level semantic features of the company's job recruitment information , as follows:

[0033] Educational requirements coding: Use one-hot coding to encode the educational requirements in the company's recruitment information;

[0034] Essential Skills Keywords: Extract using TF-IDF method Essential skills keywords in

[0035] Job requirement keywords: Extracted using TF-IDF method Keywords of job requirements in the article;

[0036] Job description keywords: Extracted using TF-IDF method Keywords in the job description;

[0037] Input the educational requirement codes, required skills keywords, job requirement keywords, and job description keywords into BERT to obtain the low-level features of the company's recruitment positions. ;

[0038] in The word segmentation sequence for the job seeker resume information text, The word segmentation sequence for the company's recruitment position information text.

[0039] Preferably, the specific processing process of the high-level semantic feature extraction network is:

[0040] First, use the pre-trained Word2Vec model to and Each word in is converted into a word vector, and the word vector set of job seeker resumes is , the word vector set of company job recruitment information is ; Then calculate the average of all word vectors to get the vector representation of the entire job seeker resume and the vector representation of the company's job positions :

[0041] ;

[0042] ;

[0043] in is the number of words in the job seeker's resume, For job seekers resume The vector representation of words, For the company's job recruitment information The vector representation of words, The number of words in the company's job postings; The word segmentation sequence for the job seeker resume information text, Segment word sequences for company recruitment job information text;

[0044] Then obtain the job applicant's position change feature vector: use a bidirectional LSTM network to capture the position sequence information in the job applicant's resume;

[0045] Obtaining the feature vector of the applicant’s historical work content: Extracting using the Transformer model Semantic features of job seekers’ historical work content in the;

[0046] Get the semantic information feature vector of the company's recruitment position requirements: Use the Transformer model to embed the company's recruitment position word vector Perform feature extraction to obtain the semantic information feature vector of the company's recruitment position requirements;

[0047] Finally, the job seeker's position change feature vector and the job seeker's historical work content feature vector are concatenated to obtain the job seeker's resume high-level features. , the semantic information feature vector of the company's recruitment position requirements is the high-level feature of the company's recruitment position .

[0048] Preferably, the feature fusion unit includes a multi-level feature fusion network and a feature attention mechanism layer;

[0049] The multi-level feature fusion network includes a local feature extraction module, a global feature extraction module, and a feature fusion module;

[0050] Use the local feature extraction module to extract the word vector set of job seekers' resumes The local context features of the local feature extraction module adopts 1D-CNN, as follows:

[0051] ;

[0052] in For the The vector representation of words, is the window size, is the local feature of the word, and the output of the local feature extraction module is ;

[0053] Extracted using the global feature extraction module The global time series features of the global feature extraction module use bidirectional GRU (BIGRU), as follows:

[0054] ;

[0055] in Output of BIGRU The output of the global feature extraction module is ;

[0056] The feature fusion module is used to combine local features, global features and contextual semantic features of job seekers’ resumes. Perform attention fusion and obtain the attention weight vector :

[0057] ;

[0058] in is the learnable weight matrix, is the bias term;

[0059] Then, the interactive attention between the resume features of the job seeker and the job features of the recruiting company is calculated through the feature attention mechanism layer :

[0060] ;

[0061] in is the job seeker resume feature matrix, ;

[0062] Matrix of job characteristics for company recruitment, ;

[0063] is the learnable weight matrix;

[0064] Attention is calculated at different granularity levels, including word level, sentence level, and paragraph level, and finally the multi-granularity attention feature results are fused to obtain interactive features. :

[0065] ;

[0066] in For the The weight of each granularity level, is the corresponding interactive attention result.

[0067] Preferably, the prediction output module uses DQN as the core network, including an input layer, a feature extraction layer, an attention layer, a feature concatenation layer, a deep Q network layer and an output layer, as follows:

[0068] Input layer: receiving state representation ; yes The interaction feature vector at time, yes A collection of positions waiting to be matched at any time;

[0069] Feature extraction layer: interactive features , use the fully connected layer to perform deep feature extraction and obtain deep interactive features :

[0070] ;

[0071] in, is the weight matrix, is the bias term;

[0072] For each position , usage and interaction features The fully connected layer with shared weights obtains job features as follows:

[0073] ;

[0074] in, It means a certain position The hidden layer output is and is the corresponding weight matrix and bias term;

[0075] Attention layer: Use multi-head self-attention mechanism to focus on job features The fusion is as follows:

[0076] ;

[0077] in,

[0078] Feature concatenation layer: deep interactive features and post-integration job characteristics Perform splicing as follows:

[0079] ;

[0080] Deep Q network layer: the fused features Enter three fully connected layers in sequence:

[0081] ;

[0082] ;

[0083] ;

[0084] Each layer is processed by the ReLU activation function, and finally the implicit features in the matching process are learned;

[0085] Output layer: The output layer corresponds to the Q value of each possible action:

[0086] ;

[0087] in , , and are the weights of the first fully connected layer, the second fully connected layer, the third fully connected layer, and the output layer, respectively. , , and are the biases of the first fully connected layer, the second fully connected layer, the third fully connected layer, and the output layer, respectively. , and They are the hidden layer output features of the first fully connected layer, the second fully connected layer, and the third fully connected layer respectively.

[0088] Preferably, the Q value is normalized to the range of 0 to 1, and the formula is as follows:

[0089] ;

[0090] in and are the minimum and maximum Q values ​​of all resume-position pairs, It is the Q value between the resume and the position;

[0091] By normalization, all matching scores It falls between 0 and 1, where 0 represents the lowest match and 1 represents the highest match;

[0092] At the same time, set a matching threshold , only when the matching degree between the position and the job seeker exceeds the threshold will it be recommended to the job seeker.

[0093] Preferably, a strategy including deep reinforcement learning is used during training to establish a state space and an action space, and determine the state transition process of the environment;

[0094] State Space :

[0095] ;

[0096] in yes The state space at time, yes The feature vector of the interaction between job seekers’ resumes and company positions at any given moment, yes A collection of positions waiting to be matched at any time;

[0097] Action Space :

[0098] ;

[0099] in yes The action space at the moment Reject all positions. Indicates selection positions;

[0100] Reward Function :

[0101] ;

[0102] in Function evaluates the quality of the match, is the rejection penalty coefficient, is the indicator function;

[0103] State transition probability :

[0104] ;

[0105] in represents the transition probability of the job set, represents the evolution probability of candidate characteristics;

[0106] The optimization goal is to achieve the best semantic similarity score between job seekers and available positions. , the reward function is:

[0107] ;

[0108] in, Parameters for weighing the semantic similarity score and the optimal semantic similarity score.

[0109] The second aspect of the present invention provides an intelligent job matching method based on deep reinforcement learning, which deploys and uses the intelligent job matching model established by the establishment method described in the first aspect, and includes the following processes:

[0110] S1, obtain the resume information of job seekers;

[0111] S2, cleaning and transforming the data to obtain standard model input data;

[0112] S3, inputs the standard input data into the intelligent job matching model to generate a job matching score;

[0113] S4, based on the setting of matching thresholds or the ranking of candidate positions, outputs matching company positions or top-ranked position recommendations.

[0114] Compared with the prior art, the present invention has the following beneficial effects:

[0115] (1) Improve matching accuracy:

[0116] Traditional methods often only focus on simple keyword matching or single-dimensional features, which makes it difficult to fully capture the deep connection between job seekers and positions. The present invention proposes a multi-level, multi-dimensional feature extraction method. It not only extracts low-level features (such as work experience, education, certificates, etc.), but also extracts the contextual semantic features of the text through the BERT model, and extracts keyword features in combination with methods such as TF-IDF. This comprehensive feature extraction method can more comprehensively characterize the attributes of job seekers and positions, providing a rich information basis for subsequent matching. Through the feature extraction module, the multi-dimensional features of job seekers and positions can be fully captured, greatly improving the richness and accuracy of feature representation, thereby significantly improving the accuracy of job matching;

[0117] (2) Realize dynamic optimization:

[0118] Traditional job matching methods usually use static rule matching or simple machine learning models, which are difficult to adapt to the dynamically changing job market environment. The present invention introduces deep reinforcement learning technology to model the job matching problem as a Markov decision process. By designing a reasonable state space, action space, and reward function, the model can continuously learn and optimize the matching strategy during interaction. This method can better capture the complex relationship between job seekers and jobs, and can dynamically adjust the matching strategy as the market changes. This dynamic optimization capability ensures that the matching results always maintain a high degree of consistency with market demand, greatly improving the timeliness and accuracy of recommendations;

[0119] (3) Reduce reliance on labeled data:

[0120] Traditional model training methods often rely on a large amount of labeled data, but high-quality job-matching labeled data is difficult to obtain. The present invention combines supervised learning and reinforcement learning. First, the initial model is obtained through pre-training, and then the policy gradient method of reinforcement learning is used for further optimization. This method can not only make full use of limited labeled data, but also continuously improve model performance through interaction with the environment. It effectively solves the problem of lack of high-quality labeled data. By combining supervised learning and reinforcement learning, the system can continuously improve performance through continuous learning based on limited labeled data;

[0121] (4) Provide explainable matching results:

[0122] Traditional matching scores are often oversimplified and difficult to reflect the multiple dimensions of job matching. The present invention designs a comprehensive matching scoring mechanism that takes into account multiple aspects such as text similarity, skill matching, experience matching, and academic matching. By introducing confidence calculation, the reliability of the matching results can also be quantitatively evaluated. This multi-dimensional, explainable scoring mechanism can provide job seekers and companies with more comprehensive and reliable matching results. The multi-dimensional matching scoring mechanism not only improves the accuracy of matching, but also enhances the interpretability of the results. Job seekers and companies can clearly understand the composition of the matching score, thereby having a deeper understanding of the matching results. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] In order to more clearly illustrate the technical solutions of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, what is described below is only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0124] Figure 1 The figure is a flowchart of the overall process of the intelligent job matching method of the present invention.

[0125] Figure 2 Schematic diagram of the network structure of the feature extraction module of the present invention.

[0126] Figure 3 Schematic diagram of the deep Q network DQN network structure of the present invention.

[0127] Figure 4 Schematic diagram of the Q value update process during model training. DETAILED DESCRIPTION

[0128] The present invention proposes a method for establishing an intelligent job matching model based on deep reinforcement learning and a matching method, wherein the job matching model used in the job matching method is trained using a training data set. The overall process is as follows: Figure 1 shown.

[0129] 1. Dataset Collection and Production

[0130] High-quality, high-precision, and diverse data are the basis for building an intelligent job matching model. This stage aims to collect, process, and label data used to train functional job matching models in order to build efficient job matching models and improve recommendation accuracy.

[0131] 1. Obtaining original job seeker resume information and company recruitment position information:

[0132] To obtain the original resume information of job seekers and company recruitment information, we use web crawler technology to obtain information from public recruitment websites (51job, Zhaopin.com), internal talent pools of enterprises, and campus recruitment platforms;

[0133] The crawler framework uses Python's Scrapy framework, which is a powerful asynchronous web crawler framework suitable for large-scale data crawling. The URL management in the crawler design uses Scrapy's start_urls and Rule classes to define the starting URL and link extraction rules to achieve incremental crawling, record the crawled URLs, and avoid repeated crawling. The request construction uses Scrapy's Request class to construct HTTP requests and set request headers such as User-Agent and Referer to simulate real browser behavior to implement IP proxy pools and request delays to avoid being intercepted by anti-crawling measures; page parsing uses Scrapy's Selector and XPath / CSS selectors to extract page information. For dynamically loaded content, Selenium WebDriver is used to render it before crawling; the Item class is designed to define the data structure to ensure that the extracted data fields are unified, and the Item Pipeline is used to clean and format the data for data extraction; at the same time, Scrapy-Redis can be used to implement distributed crawlers to improve crawling efficiency, and Redis can be used as a URL queue and temporary data storage;

[0134] Adopt dynamic User-Agent switching and use Tor network or VPN service to hide the real IP. You can also add a verification code recognition module to handle the verification code challenge of the website to implement anti-crawler strategy response; use MongoDB to store original resume data, support flexible document structure, and use data sharding and index optimization to improve the reading and writing performance of large-scale data; use Scrapy's built-in statistics collector to monitor the crawler running status to implement exception handling and log recording, timely discover and solve problems in the crawling process, and set scheduled tasks to regularly update the resume database;

[0135] When performing crawling operations, we will abide by the website rules, carefully read and comply with the robots.txt file of the target website, and control the crawling frequency to avoid burdening the target website.

[0136] 2. Data preprocessing:

[0137] The original collected data of job seekers’ resume information and company recruitment position information may have three problems: privacy leakage, format errors, and invalid information. Therefore, comprehensive data preprocessing is required to ensure the data quality for subsequent model training. The main steps include:

[0138] (1) Data cleaning

[0139] This step aims to improve data quality and ensure the accuracy of subsequent processing. First, regular expressions are used to remove HTML tags and special characters, and the date format is unified as "YYYY-MM-DD". For missing values, different strategies are adopted: numerical missing values are filled with the mean or median, categorical missing values are filled with the mode, and text missing values are filled with "Unknown". This processing can maintain the integrity of the data without introducing too much bias;

[0140] (2) Data Transformation

[0141] First, convert the unstructured text represented by the job seeker's resume information and the company's recruitment position matching information into a numerical representation that can be understood by machines. Use the jieba library for Chinese word segmentation and the NLTK library for English word segmentation. Then, remove stop words, such as "的", "了", "the", and "a". Finally, obtain the word segmentation sequences of the job seeker's resume information text and the word segmentation sequences of the company's recruitment position information text ;

[0142] (3) Data Labeling

[0143] Similarity calculation: Adopt multiple methods to calculate the similarity between the job seeker's resume information and the company's recruitment position information. First, use TF-IDF to convert the text into vectors and calculate the cosine similarity to obtain the overall text similarity . Secondly, convert the resume skills and the required skills of the position into word vectors and calculate the average cosine similarity to obtain the skill matching degree . Then, calculate the experience matching degree through numerical comparison and scoring functions . Finally, design a mapping of academic degree levels and compare the difference between the resume academic degree and the required academic degree of the position to obtain the academic degree matching degree . This multi-dimensional similarity calculation can comprehensively evaluate the matching situation;

[0144] Comprehensive matching score calculation: Define the comprehensive matching degree score:

[0145] ;

[0146] where is the weight. The initial weight can be set according to domain knowledge and can be optimized later through machine learning algorithms (such as linear regression). Based on the score , divide the matching degree into four levels: is non-matching, is partial matching, is good matching, is excellent matching. At the same time, calculate the confidence level:

[0147] ;

[0148] Where threshold is the lower limit threshold of the matching degree;

[0149] Label generation and quality control: Based on the calculation results, a label containing resume_id (resume id), job_id (job id), (match score) and confidence. To ensure the quality of labels, the following strategies are adopted: First, a small number of high-quality datasets with manual annotations are used as validation sets to regularly evaluate the accuracy of automatic annotations. Second, samples with low confidence are manually reviewed. Finally, the matching process is continuously improved by adjusting the matching algorithm, optimizing weights, and introducing more complex machine learning models (such as random forests or gradient boosting trees). This continuous optimization method can ensure the reliability and accuracy of the annotation results. This method realizes the automated matching and annotation of resumes and job descriptions, which provides a basis for subsequent machine learning model training. This process takes into account matching factors in multiple dimensions, and gives the degree of match and confidence through comprehensive scoring, making the annotation results more reliable and explainable;

[0150] (4) Dataset creation

[0151] After pre-processing and cleaning, the processed corpus is divided into three independent data sets: training set, validation set and test set in a ratio of 8:1:1. The training set is used for iterative learning of model parameters, the validation set is used for indicator monitoring during model training, and the test set is used for the final evaluation of the model's generalization ability. Among them, the validation set and test set are randomly sampled and constructed by the data set partitioning tool, and are independent of the training set data;

[0152] The final input data generated is divided into two parts: job seeker resume data and company job recruitment information data ; Output target is the matching score between job seekers and company recruitment information ,The higher the matching score, the more the job seeker’s personal information fits the job requirements;

[0153] Through a rigorous data collection, cleaning, enhancement, and segmentation process, we finally obtained a high-quality, semantically rich, and sufficient training dataset, denoted as . This ensures the efficiency and accuracy of subsequent semantic feature representation learning and self-supervised training, laying a solid foundation for achieving excellent matching recommendation results.

[0154] 2. Build a job matching model

[0155] 1. Feature extraction module

[0156] The feature extraction module includes a comprehensive feature extraction unit and a feature fusion unit; the comprehensive feature extraction unit includes a contextual semantic feature extraction network, a low-level semantic feature extraction network and a high-level semantic feature extraction network; the contextual semantic feature extraction network is used to extract contextual semantic features of job seekers’ resume information data and company recruitment position information; the low-level semantic feature extraction network is used to extract feature keywords of job seekers’ resume information and company recruitment position information; the high-level semantic feature extraction network is used to obtain the job seekers’ job transition feature vectors and the job seekers’ historical work content feature vectors, and perform vector splicing to obtain the job seekers’ resume high-level features, and at the same time obtain the company’s recruitment position requirement semantic information feature vector as the company’s recruitment position high-level features; the feature fusion unit is used to fuse contextual semantic features, low-level semantic features and high-level semantic features, and fully explore the relationship between different features to obtain multi-granular attention feature representations of the interactive information between resumes and positions. Its overall network structure is as follows: Figure 2 shown.

[0157] (1) Comprehensive feature extraction unit

[0158] Extract key features from job seekers' resume information and company recruitment information, including contextual semantic features, low-level semantic features and high-level semantic features, which cover multiple dimensions required for evaluating matching. Finally, obtain the word segmentation sequence of job seekers' resume information and the word segmentation sequence of the company's recruitment position information text .

[0159] Contextual semantic feature extraction:

[0160] The BERT model is used to extract contextual semantic features of job seeker resume information data and company recruitment position information, as follows:

[0161] ;

[0162] ;

[0163] in is the contextual semantic feature of the job seeker's resume information, Contextual semantic features of the company's recruitment job information;

[0164] The context semantic feature extraction method is relatively conventional and will not be explained in detail here.

[0165] Low-level semantic feature extraction:

[0166] First, extract the low-level semantic features of the job seeker's resume , as follows:

[0167] Work experience keywords: extracted using TF-IDF method Work experience keywords in;

[0168] Educational qualification coding: One-hot coding is used to encode the educational qualification in the applicant’s resume;

[0169] Certificate keywords: extracted using TF-IDF method The certificate keywords in;

[0170] Skill keywords: extracted using the TF-IDF method Skill keywords in;

[0171] Input work experience keywords, education codes, certificate keywords, and skill keywords into BERT to obtain low-level features of job seekers’ resumes ;

[0172] Secondly, we construct the low-level semantic features of the company's job recruitment information , as follows:

[0173] Educational requirements coding: Use one-hot coding to encode the educational requirements in the company's recruitment information;

[0174] Essential Skills Keywords: Extract using TF-IDF method Essential skills keywords in

[0175] Job requirement keywords: Extracted using TF-IDF method Keywords of job requirements in the article;

[0176] Job description keywords: Extracted using TF-IDF method Keywords in the job description;

[0177] Input the educational requirement codes, required skills keywords, job requirement keywords, and job description keywords into BERT to obtain the low-level features of the company's recruitment positions. .

[0178] High-level semantic feature extraction:

[0179] 1) First, use the pre-trained Word2Vec model to and Each word in is converted into a word vector, and the word vector set of job seeker resumes is , the word vector set of company job recruitment information is ; Then calculate the average of all word vectors to get the vector representation of the entire job seeker resume and the vector representation of the company's job positions :

[0180] ;

[0181] ;

[0182] in is the number of words in the job seeker's resume, For job seekers resume The vector representation of words, For the company's job recruitment information The vector representation of words, is the number of words in the company's job recruitment information. This method takes into account both word order and semantic information;

[0183] 2) Feature vector of job applicant’s position transition: Use a bidirectional LSTM network to capture the position sequence information in the job applicant’s resume;

[0184] ;

[0185] in for The position word vector at the moment, is a hidden state. The job seeker's work experience is time-dependent, so the LSTM network is used to extract the characteristics of the job seeker's work experience and obtain the job seeker's position change feature vector;

[0186] 3) Feature vector of job seeker’s historical work content: extracted using Transformer model Semantic features of job seekers' historical work content in .

[0187] ;

[0188] Where Q, K, and V are query, key, and value matrices, respectively. The attention mechanism in the transformer is used to focus on important information in the job seeker's historical work content and ignore unimportant text content.

[0189] The semantic information feature vector of the company's recruitment position requirements: Use the Transformer model to embed the company's recruitment position word vector Perform feature extraction to obtain the semantic information feature vector of the company's recruitment job requirements. Use the attention mechanism in the transformer to focus on the important information in the company's recruitment job requirements and ignore unimportant text content.

[0190] Concatenate the job applicant's job change feature vector and the job applicant's historical work content feature vector to obtain the job applicant's resume high-level features , the semantic information feature vector of the company's recruitment position requirements is the high-level feature of the company's recruitment position .

[0191] (2) Feature fusion unit

[0192] The present invention adopts the innovative "Hierarchical Attention Fusion Network" (HAFN) to fuse three levels of features: contextual semantic features, low-level semantic features and high-level semantic features, specifically including a multi-level feature fusion network and a feature attention mechanism layer.

[0193] 1) Multi-level feature fusion network:

[0194] The multi-level feature fusion network specifically includes a local feature extraction module, a global feature extraction module, and a feature fusion module:

[0195] Use the local feature extraction module to extract the word vector set of job seekers' resumes The local context features of the local feature extraction module adopts 1D-CNN, as follows:

[0196] ;

[0197] in For the The vector representation of words, is the window size, is the local feature of the word, and the output of the local feature extraction module is .

[0198] Extracted using the global feature extraction module The global time series features of the global feature extraction module use bidirectional GRU (BIGRU), as follows:

[0199] ;

[0200] in Output of BIGRU The output of the global feature extraction module is .

[0201] The feature fusion module is used to combine local features, global features and contextual semantic features of job seekers’ resumes. Perform attention fusion and obtain the attention weight vector :

[0202] ;

[0203] in is the learnable weight matrix, is the bias term.

[0204] 2) Feature attention mechanism layer:

[0205] Introducing the "Multi-granularity Interactive Attention, MIA" mechanism to calculate the interactive attention between the resume features of job seekers and the job features of recruiting companies

[0206] ;

[0207] in is the job seeker resume feature matrix, ;

[0208] Matrix of job characteristics for company recruitment, ;

[0209] is the learnable weight matrix

[0210] Calculate attention at different granularity levels (word level, sentence level, paragraph level), and finally fuse the multi-granularity attention feature results:

[0211] ;

[0212] in For the The weight of each granularity level, is the corresponding attention result.

[0213] This innovative feature extraction and fusion method can fully utilize the multi-level features of resumes and job information and capture the complex interactions between them, thereby providing a richer and more expressive feature representation for subsequent matching models. This method not only considers local and global information, but also introduces multi-granular interactions, which is expected to significantly improve the accuracy and interpretability of matching.

[0214] 2. Prediction output module

[0215] This module uses deep reinforcement learning methods to model the job matching problem as a Markov decision process (MDP), continuously optimizes the matching strategy through interaction with the environment, and finally outputs accurate matching prediction results.

[0216] (1) Environmental modeling

[0217] The job matching problem is modeled as an MDP, which is defined as follows:

[0218] Defines the state space of the agent , Action Space , reward function and state transition probability , through the interaction between the agent and the environment, the matching strategy is continuously optimized and the matching prediction results are output; the state space Indicates the matching status and action space between resume and position Represents the position and reward function selected by the agent at a certain action step Indicates the reward and state transition probability obtained by the agent when choosing this position represents the transition probability of the job set;

[0219] State Space : , in yes The state space at time, yes The feature vector of job seekers at a given moment, yes A collection of positions waiting to be matched at any time;

[0220] Action Space : , in yes The action space at the moment Reject all positions. Indicates selection positions;

[0221] Reward Function : ;

[0222] in Function evaluates the quality of the match, is the rejection penalty coefficient, is the indicator function;

[0223] State transition probability :

[0224] in represents the transition probability of the job set, represents the evolution probability of candidate characteristics;

[0225] Discount Factor : Used to balance immediate rewards and long-term benefits, usually set to 0.9~0.99.

[0226] (2) Deep Q Network (DQN) Design

[0227] DQN is used as the core network, which includes input layer, feature extraction layer, attention layer, feature concatenation layer, deep Q network layer and output layer. The structure is as follows: Figure 3 As shown, the details are as follows:

[0228] Input layer: receiving state representation ;

[0229] Feature extraction layer:

[0230] Interaction Features , use the fully connected layer for deep feature extraction:

[0231] ;

[0232] in, is the weight matrix, is the bias term;

[0233] For each position , usage and interaction features The fully connected layer with shared weights obtains job features as follows:

[0234] ;

[0235] in, It means a certain position The hidden layer output of the features, and is the corresponding weight matrix and bias term;

[0236] Attention layer: Use multi-head self-attention mechanism to focus on job features The fusion is as follows:

[0237] ;

[0238] in:

[0239] The core of attention is: ;

[0240] Feature concatenation layer: concatenates interactive features and post-integration job characteristics Perform splicing as follows:

[0241] ;

[0242] Deep Q-Network Layer: Fusion Features Pass through three fully connected layers in sequence:

[0243] ;

[0244] ;

[0245] ;

[0246] Each layer is processed by the ReLU activation function, and finally the implicit features in the matching process are learned;

[0247] Output layer: The output layer corresponds to the Q value of each possible action:

[0248] ;

[0249] in and are the parameters to be learned.

[0250] Q value explanation:

[0251] Q Value It represents the expected long-term return of taking an action (recommending a job) in a given state (job seeker's resume). A higher Q value indicates that the job is more likely to match the job seeker;

[0252] state :Resume characteristics of job seekers;

[0253] action : Recommend a specific position;

[0254] Q value: reflects the expected return of recommending a job given a resume. The higher the Q value, the more the job matches the job seeker, so a higher Q value means that the job is more suitable for the job seeker.

[0255] Q value conversion:

[0256] In order to facilitate the interpretation of the Q value as the actual matching score, it is normalized to the range of 0 to 1, and the formula is as follows:

[0257] ;

[0258] in and are the minimum and maximum Q values ​​of all (resume, position) pairs, It is the Q value between the resume and the position;

[0259] By normalization, all matching scores Will fall between 0 and 1.

[0260] 3. Model Training

[0261] In the job matching method, a self-supervised deep reinforcement learning method is used to optimize the matching strategy to achieve the optimal job matching result.

[0262] In deep reinforcement learning, in order to obtain optimal results, it is necessary to establish a state space and action space and determine the state transition process of the environment.

[0263] State Space : , in yes The state space at time, yes The feature vector of the interaction between job seekers’ resumes and company positions at any given moment, yes A collection of positions ready to be matched.

[0264] Action Space : , in yes The action space at the moment Reject all positions. Indicates selection Positions.

[0265] Design a reward function to guide the behavior of reinforcement learning ad recommendation. Reward function It should be consistent with the optimization goal. In this invention, the optimization goal is to achieve the best semantic similarity score between job seekers and available positions. , so the reward function is:

[0266] ;

[0267] in, Parameters for weighing the semantic similarity score and the optimal semantic similarity score.

[0268] The Q-value function is used to learn the action-value function to guide the selection of actions. The job matching model is trained based on the Q-value function, and the loss function is defined by each experience tuple. The expected value of the sample is calculated as is the next state. The loss function is:

[0269] ;

[0270] in, Indicates the expected operator, is the discount factor, is the parameter of the target network, which is used to calculate the target Q value, that is, the expected cumulative reward in the future. are the parameters of the main network, used to evaluate the Q value of taking different actions in a given state, For the status The following actions are taken.

[0271] Based on this loss function, the job matching is trained. The specific steps are:

[0272] 1) Initialize the experience tuple , parameters of the main network and target network parameters And the state space ;

[0273] 2) Select actions based on the current Q network ;

[0274] 3) Perform actions in the environment , observe the next state , get instant rewards ;

[0275] 4) Store the experience tuple;

[0276] 5) Sample batches of data from the experience tuple and use the loss function defined above Calculate the loss of the batch data;

[0277] 6) Use stochastic gradient descent to update the Q value network parameters and perform state Updates;

[0278] 7) Repeat steps 2)-6). When the total reward of the model does not increase significantly, it proves that the model has converged and the training ends.

[0279] The Q value update process during model training is as follows Figure 4 shown.

[0280] 4. Job matching based on model output scores

[0281] Use the set threshold to screen the matches between job seekers' resumes and all company recruitment positions with matching scores greater than the threshold, and obtain a set of job seekers' resumes and company recruitment positions matching sets, which are the company recruitment positions recommended to the job seekers; or sort the match scores between job seekers' resumes and all company recruitment positions from large to small, and select the top-ranked company recruitment positions as recommendations to job seekers.

[0282] 1. Ranking of candidate positions

[0283] In actual application, the matching score between each job seeker and all available positions is calculated, and then the positions are sorted according to the matching score. Positions with higher matching scores are ranked higher, indicating that these positions are more compatible with the job seeker.

[0284] 2. Matching threshold setting

[0285] In order to control the quality of recommended positions, set a matching threshold . Only when the matching degree between the position and the job seeker exceeds this threshold will it be recommended to the job seeker.

[0286] 5. Model deployment

[0287] 1. Model packaging and optimization:

[0288] Serialize the trained deep reinforcement learning model and save it in a deployable format (such as ONNX);

[0289] Prune and quantize the model to reduce the model size and improve the inference speed;

[0290] Use tools such as TensorRT to optimize the model and further improve reasoning performance.

[0291] 2. Deployment environment preparation:

[0292] Choose a suitable deployment platform, such as a cloud server or edge device;

[0293] Configure necessary hardware resources, such as GPU accelerator cards;

[0294] Install the required dependent libraries and runtime environment, such as CUDA, cuDNN, etc.

[0295] 3. Reasoning process implementation:

[0296] Real-time data collection, and three main steps: data preprocessing, feature extraction and model prediction;

[0297] In the preprocessing stage, the collected resumes and job information are converted into a format acceptable to the model;

[0298] Use feature extraction module to extract semantic features of text;

[0299] The extracted features are input into the deep reinforcement learning model to generate a job matching score;

[0300] Based on the setting of matching thresholds or the ranking of candidate positions, matching company positions or top-ranked position recommendations are output.

[0301] 4. Load balancing and expansion:

[0302] Use a load balancer (such as Nginx) to distribute requests and improve the system's concurrent processing capabilities;

[0303] Implement horizontal expansion architecture and dynamically increase service nodes based on demand.

[0304] 5. Monitoring and logging system:

[0305] Deploy monitoring tools such as Prometheus to monitor model performance and system resource usage in real time;

[0306] Implement a logging system to track model predictions and potential anomalies.

[0307] 6. Safety measures:

[0308] Implement data encryption transmission to protect user privacy information;

[0309] Deploy a Web Application Firewall (WAF) to prevent malicious attacks;

[0310] Implement access control and authentication mechanisms to ensure secure use of the API.

[0311] 7. Continuous Integration and Deployment (CI / CD):

[0312] Build a CI / CD pipeline to automate the testing, building, and deployment of models;

[0313] Configure a rollback mechanism to quickly restore to a stable version when problems occur in the new version.

[0314] 8.A / B Testing System:

[0315] Implement an A / B testing framework to compare the performance of different versions of the model;

[0316] Dynamically adjust the model version of the online service based on the test results.

[0317] 9. Data Feedback Loop

[0318] Design data collection mechanisms to capture user interactions and matching result feedback;

[0319] Implement a regular model retraining process to continuously optimize model performance using newly collected data.

[0320] 10. Documentation and Training:

[0321] Write detailed API documentation and instructions;

[0322] Provide system maintenance and troubleshooting training to the operation and maintenance team.

[0323] The present invention uses the data set published by Zhilian Zhaopin in Ali Tianchi for experiments. The comparison models of the method of the present invention are multi-scale convolutional attention network, sequence-to-sequence matching network, reinforcement learning strategy network and graph neural network. Data preprocessing is performed before the job seeker resume information and company recruitment information are input into the model. The processed data set contains 2716 resumes, 13326 recruitment information and 26337 interaction records. The accuracy of the model of the present invention is 92.3%, which is slightly higher than the traditional reinforcement learning strategy network, and far exceeds the multi-scale convolutional attention network, sequence-to-sequence matching network and graph neural network.

[0324] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0325] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for establishing an intelligent job matching model based on deep reinforcement learning, characterized in that: The process includes: S1, obtain job seekers’ resume information and company recruitment position information data to construct the original data set; S2, preprocessing the original data set to obtain the model standard training data set; S3, building an intelligent job matching model, including a feature extraction module and a prediction output module; The feature extraction module includes a comprehensive feature extraction unit and a feature fusion unit; the comprehensive feature extraction unit includes a contextual semantic feature extraction network, a low-level semantic feature extraction network and a high-level semantic feature extraction network; The low-level semantic feature extraction network is used to extract the characteristic keywords of the job seeker's resume information and the company's job recruitment information; the high-level semantic feature extraction network is used to obtain the job seeker's position change feature vector and the job seeker's historical work content feature vector, and perform vector splicing to obtain the high-level features of the job seeker's resume, and at the same time obtain the high-level features of the company's recruitment positions; the feature fusion unit is used to fuse the extracted features to obtain a multi-granularity attention feature representation; The feature fusion unit includes a multi-level feature fusion network and a feature attention mechanism layer; The multi-level feature fusion network includes a local feature extraction module, a global feature extraction module, and a feature fusion module; Use the local feature extraction module to extract the word vector set of job seekers' resumes The local context features of the local feature extraction module adopts 1D-CNN, as follows: ； in For the The vector representation of words, is the window size, is the local feature of the word, and the output of the local feature extraction module is ; Extracted using the global feature extraction module The global time series features of the global feature extraction module use bidirectional GRU, namely BIGRU, as follows: ； in Output of BIGRU The output of the global feature extraction module is ; The feature fusion module is used to combine local features, global features and contextual semantic features of job seekers’ resumes. Perform attention fusion and obtain the attention weight vector : ； in is the learnable weight matrix, is the bias term; Then, the interactive attention between the resume features of the job seeker and the job features of the recruiting company is calculated through the feature attention mechanism layer : ； in is the job seeker resume feature matrix, ; Matrix of job characteristics for company recruitment, ; is a learnable weight matrix; The low-level semantic features of job seekers’ resumes are obtained by inputting work experience keywords, education codes, certificate keywords, and skill keywords into BERT; The low-level semantic features of the company's job recruitment information are obtained by inputting the education requirement codes, required skills keywords, job requirement keywords, and job description keywords into BERT; The high-level features of the job seeker's resume are obtained by concatenating the job seeker's position change feature vector and the job seeker's historical work content feature vector; It is the high-level features of the company's recruitment positions, that is, the feature vector of semantic information required by the company's recruitment positions; Attention is calculated at different granularity levels, including word level, sentence level, and paragraph level, and finally the multi-granularity attention feature results are fused to obtain interactive features. : ； in For the The weight of each granularity level, is the corresponding interactive attention result; The prediction output module uses a deep Q network to model the job matching problem as a Markov decision process, where the Q value is the expected return of recommending a job given the resume information; the final output is the matching score between the job seeker's resume and the company's recruitment position; S4, conduct deep reinforcement learning training on the job matching model to obtain the final model.

2. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The preprocessing of the original data set in S2 includes data cleaning and data conversion; The data cleaning first uses regular expressions to remove HTML tags and special characters, and unifies the date format to "YYYY-MM-DD"; For missing values, different strategies are used: numeric missing values ​​are filled with mean or median, categorical missing values ​​are filled with mode, and text missing values ​​are filled with "Unknown"; The data conversion is to convert the unstructured text represented by the job seeker's resume information and the company's recruitment position matching information into a numerical representation that can be understood by a machine; wherein, the jieba library is used for Chinese word segmentation, and the NLTK library is used for English word segmentation.

3. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The preprocessing of the original data set in S2 also includes data label processing, including similarity calculation, comprehensive matching score calculation and label generation; The similarity calculation: First, use TF-IDF to convert the text into a vector, and calculate the cosine similarity to get the overall text similarity ; Secondly, convert resume skills and job requirement skills into word vectors and calculate the average cosine similarity to get the skill matching degree. ; Then, the empirical matching degree is calculated through numerical comparison and scoring function Finally, design the education level mapping, compare the difference between the resume education level and the job requirements to get the education level matching degree ; The comprehensive matching score calculation: First, define the comprehensive matching score : ； in is the weight; then based on the score , the matching degree is divided into four levels: For mismatch, For partial matching, For a good match, is an excellent match; the confidence is calculated at the same time: ； Where threshold is the lower limit threshold of the matching degree; The label generation: based on the comprehensive matching score calculation result, a label including the resume ID, position ID, matching score and confidence level is generated for each pair of job seeker resume information and company recruitment position information.

4. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The specific processing process of the low-level semantic feature extraction network is as follows: First, extract the low-level semantic features of the job seeker's resume , as follows: Work experience keywords: extracted using TF-IDF method Work experience keywords in; Educational qualification coding: One-hot coding is used to encode the educational qualification in the applicant’s resume; Certificate keywords: extracted using TF-IDF method The certificate keywords in; Skill keywords: extracted using the TF-IDF method Skill keywords in; Input work experience keywords, education codes, certificate keywords, and skill keywords into BERT to obtain low-level features of job seekers’ resumes ; Secondly, extract the low-level semantic features of the company's job recruitment information , as follows: Educational requirements coding: Use one-hot coding to encode the educational requirements in the company's recruitment information; Essential Skills Keywords: Extract using TF-IDF method Essential skills keywords in Job requirement keywords: Extracted using TF-IDF method Keywords of job requirements in the article; Job description keywords: Extracted using TF-IDF method Keywords in the job description; Input the educational requirement codes, required skills keywords, job requirement keywords, and job description keywords into BERT to obtain the low-level features of the company's recruitment positions. ; in The word segmentation sequence for the job seeker resume information text, Segment word sequence for company recruitment job information text.

5. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The specific processing process of the high-level semantic feature extraction network is as follows: First, use the pre-trained Word2Vec model to and Each word in is converted into a word vector, and the word vector set of job seeker resumes is , the word vector set of company job recruitment information is ; Then calculate the average of all word vectors to get the vector representation of the entire job seeker resume and the vector representation of the company's job positions : ； ； in is the number of words in the job seeker's resume, For job seekers resume The vector representation of words, For the company's job recruitment information The vector representation of words, The number of words in the company's job postings; The word segmentation sequence for the job seeker resume information text, Segment word sequences for company recruitment job information text; Then obtain the job applicant's position change feature vector: use a bidirectional LSTM network to capture the position sequence information in the job applicant's resume; Obtaining the feature vector of the applicant’s historical work content: Extracting using the Transformer model Semantic features of job seekers’ historical work content in the; Get the semantic information feature vector of the company's recruitment position requirements: Use the Transformer model to embed the company's recruitment position word vector Perform feature extraction to obtain the semantic information feature vector of the company's recruitment position requirements; Finally, the job seeker's position change feature vector and the job seeker's historical work content feature vector are concatenated to obtain the job seeker's resume high-level features. , the semantic information feature vector of the company's recruitment position requirements is the high-level feature of the company's recruitment position .

6. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The prediction output module uses DQN as the core network, including input layer, feature extraction layer, attention layer, feature concatenation layer, deep Q network layer and output layer, as follows: Input layer: receiving state representation ; yes The interaction feature vector at time, yes A collection of positions waiting to be matched at any time; Feature extraction layer: interactive features , use the fully connected layer to perform deep feature extraction and obtain deep interactive features : ； in, is the weight matrix, is the bias term; For each position , usage and interaction features The fully connected layer with shared weights obtains job features as follows: ； in, It means a certain position The hidden layer output is and is the corresponding weight matrix and bias term; Attention layer: Use multi-head self-attention mechanism to focus on job features The fusion is as follows: ； in, ; Feature concatenation layer: deep interactive features and post-integration job characteristics Perform splicing as follows: ； Deep Q network layer: the fused features Enter three fully connected layers in sequence: ； ； ； Each layer is processed by the ReLU activation function, and finally the implicit features in the matching process are learned; Output layer: The output layer corresponds to the Q value of each possible action: ； in , , and are the weights of the first fully connected layer, the second fully connected layer, the third fully connected layer, and the output layer, respectively. , , and are the biases of the first fully connected layer, the second fully connected layer, the third fully connected layer, and the output layer, respectively. , and They are the hidden layer output features of the first fully connected layer, the second fully connected layer, and the third fully connected layer respectively.

7. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: The Q value is normalized to the range of 0 to 1, and the formula is as follows: ； in and are the minimum and maximum Q values ​​of all resume-position pairs, It is the Q value between the resume and the position; By normalization, all matching scores It falls between 0 and 1, where 0 represents the lowest match and 1 represents the highest match; At the same time, set a matching threshold , only when the matching degree between the position and the job seeker exceeds the threshold will it be recommended to the job seeker.

8. The method for establishing an intelligent job matching model based on deep reinforcement learning according to claim 1, characterized in that: During training, strategies including deep reinforcement learning are used to establish a state space and action space, and determine the state transition process of the environment; State Space : ； in yes The state space at time, yes The feature vector of the interaction between job seekers’ resumes and company positions at any given moment, yes A collection of positions waiting to be matched at any time; Action Space : ； in yes The action space at the moment Reject all positions. Indicates selection positions; Reward Function : ； in Function evaluates the quality of the match, is the rejection penalty coefficient, is the indicator function; State transition probability : ； in represents the transition probability of the job set, represents the evolution probability of candidate characteristics; The optimization goal is to achieve the best semantic similarity score between job seekers and available positions. , the reward function is: ； in, Parameters for weighing the semantic similarity score and the optimal semantic similarity score.

9. An intelligent job matching method based on deep reinforcement learning, characterized in that: Deploy and use the intelligent job matching model established by the establishment method according to any one of claims 1 to 8, and include the following processes: S1, obtain the resume information of job seekers; S2, cleaning and transforming the data to obtain standard model input data; S3, inputs the standard input data into the intelligent job matching model to generate a job matching score; S4, based on the setting of matching thresholds or the ranking of candidate positions, outputs matching company positions or top-ranked position recommendations.

Images