AI-based intelligent resume matching and automatic searching method and system

Abstract

The present application relates to the technical field of artificial intelligence and human resource management, and particularly relates to an AI-based intelligent resume matching and automatic searching method and system. The method comprises the following steps: performing multi-level semantic analysis on a position requirement text to extract position requirement features, and performing career development trajectory modeling on a candidate resume text to extract ability evolution features; a cross-modal alignment space is constructed, the position requirement features and the ability evolution features are mapped to the space, a nonlinear mapping relationship is established through a requirement-capability dual constraint mechanism, and an initial matching degree is obtained by calculating a semantic distance; based on the initial matching degree, a to-be-verified candidate is identified, a future ability growth curve of the to-be-verified candidate is inferred to generate a potential evaluation value, and a target candidate set is determined by priority sorting; salary sensitivity features and work location preference features of the target candidate are extracted, response probability distributions of the target candidate to different types of searching information are predicted, and an automatic searching reaching strategy is generated by dynamically selecting a reaching time and a reaching channel.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and human resource management technology, and in particular to an AI-based intelligent resume matching and automatic search method and system. Background Technology

[0002] In existing technologies, intelligent resume matching and automated candidate search methods typically rely on keyword matching or rule-based screening systems. The conventional approach involves extracting explicit keywords from job requirements, such as skills, education, and years of work experience, and performing string comparisons or simple vector similarity calculations against corresponding fields in the resume to generate a preliminary candidate matching list. Subsequently, the system uses this list to determine search targets based on fixed rules (such as matching thresholds) or manual review, and then contacts them using a uniform outreach strategy (such as sending standardized emails or SMS messages). These methods offer efficiency advantages when rapidly screening large numbers of resumes and are widely used in recruitment platforms and enterprise human resource management systems.

[0003] Keyword matching ignores the semantic depth of job requirements and candidate capabilities. For example, "team leadership experience" in a job description and "project coordination role" in a resume are implicitly related, but simple text comparison cannot capture this abstract mapping relationship, leading to the omission of high-potential but atypical candidates. Fixed-rule search strategies do not consider individual differences among candidates, such as salary sensitivity or work location preferences, resulting in inappropriate timing or channel mismatch. For example, sending invitations outside of a candidate's active period or pushing information through platforms they do not frequently use reduces response and search success rates. These limitations make it difficult for existing systems to adapt to complex and ever-changing job requirements and candidates' dynamic career trajectories. Summary of the Invention

[0004] This invention provides an AI-based intelligent resume matching and automatic search method and system, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an AI-based intelligent resume matching and automatic search method, comprising: Multi-level semantic parsing is performed on the job requirement description text to extract job requirement features, and career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features of each candidate. Construct a cross-modal alignment space for job requirements and candidate capabilities, map the job requirement features and candidate capability evolution features to the cross-modal alignment space, establish a nonlinear mapping relationship by introducing a requirement-capability dual constraint mechanism, and calculate the semantic distance between each candidate and the job requirements in the cross-modal alignment space based on the nonlinear mapping relationship to obtain the initial matching degree. Based on the initial matching degree, candidates to be verified are identified. By inferring the future ability growth curve of the candidates to be verified in the target job field, a potential assessment value is generated. The candidates to be verified are prioritized according to the potential assessment value to determine the target candidate set. Extract salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predict the probability distribution of each target candidate's response to different types of search information; Based on the response probability distribution, the contact time and contact channel are dynamically selected for each candidate, and an automatic search and contact strategy is generated.

[0006] Multi-level semantic parsing is performed on the job requirement description text to extract job requirement features. Career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features of each candidate, including: The job requirement description text is parsed syntactically to identify skill entities, experience entities, and responsibility entities and to label syntactic dependency relationships. Based on the syntactic dependency relationships, a semantic dependency graph of job requirements is constructed, where nodes represent entities and edges represent the semantic association types between entities. The semantic dependency graph of job requirements is encoded using graph structure. By analyzing the neighborhood topology features of each node and the semantic association type of the edges, local semantic information and global structural information are fused into job requirement node representation vectors. The job requirement node representation vectors are then weighted and aggregated to generate the job requirement features. The candidate's resume text is subjected to chronological information extraction, each work experience segment is identified, and the start and end times and skill tags in each work experience segment are extracted. Based on the start and end times, a candidate's career timeline is constructed, and skill tags are sequentially arranged in chronological order. By analyzing the change patterns of skill tags between adjacent work experience segments, the direction of ability expansion and the degree of ability deepening are identified. Based on the direction of capability expansion and the degree of capability deepening, each skill tag on the candidate's career timeline is assigned an evolution weight. Based on the evolution weight, each skill tag is weighted and encoded to generate the candidate's capability evolution characteristics.

[0007] Constructing a cross-modal alignment space between job requirements and candidate capabilities, and mapping the job requirement features and the candidate capability evolution features to the cross-modal alignment space, includes: A job requirement encoder and a candidate ability encoder are constructed and jointly trained through a contrastive learning mechanism. The contrastive learning mechanism maximizes the similarity between the matching job requirement features and the candidate ability evolution features, while minimizing the similarity between the mismatched job requirement features and the candidate ability evolution features, so that the output vectors of the job requirement encoder and the candidate ability encoder form a semantically aligned distribution in a vector space of the same dimension. The vector space corresponding to the trained job requirement encoder and the candidate ability encoder is defined as the cross-modal alignment space. The job requirement encoder maps the job requirement features to job requirement semantic vectors, which are located in the cross-modal alignment space; the candidate ability encoder maps the candidate ability evolution features to candidate ability semantic vectors, which are located in the cross-modal alignment space.

[0008] By introducing a demand-capability dual constraint mechanism to establish a nonlinear mapping relationship, and calculating the semantic distance between each candidate and the job requirements in the cross-modal alignment space based on the nonlinear mapping relationship, the initial matching degree is obtained, including: A demand-capability dual constraint mechanism is constructed, which includes demand integrity constraints and capability sufficiency constraints. The demand integrity constraints are used to calculate the coverage of the candidate's capability semantic vector to each dimension of the job demand semantic vector, and the capability sufficiency constraints are used to calculate the matching degree of the job demand semantic vector to each dimension of the candidate's capability semantic vector. The dual constraint score is obtained by performing a bidirectional interactive calculation between the coverage degree and the matching degree, and the nonlinear mapping relationship is established. The job requirement semantic vector and the candidate ability semantic vector of each candidate are respectively input into the requirement integrity constraint and the ability sufficiency constraint to calculate the coverage degree and the matching degree of each candidate. The coverage and matching degree of each candidate are input into the nonlinear mapping relationship to calculate the dual constraint score of each candidate. Calculate the original vector distance between the job requirement semantic vector and the candidate ability semantic vector of each candidate, and fuse the original vector distance with the dual constraint score to obtain the constraint-adjusted semantic distance as the initial matching degree.

[0009] Based on the initial matching degree, candidates to be verified are identified. By inferring the future ability growth curve of the candidates to be verified in the target job field, a potential assessment value is generated. Based on the potential assessment value, the candidates to be verified are prioritized, and the target candidate set is determined to include: Set a matching threshold, and identify candidates whose initial matching degree meets the matching threshold as candidates to be verified. Extract the capability status records and capability change trend characteristics of the candidate to be verified at multiple time points as historical capability evolution trajectory data; The ability change patterns in the historical ability evolution trajectory data are correlated and mapped with the domain ability requirement characteristics of the target job domain to establish the correlation between ability evolution and requirement evolution. Based on the correlation, an ability growth trend coefficient and an ability saturation boundary parameter are generated. The ability saturation boundary parameter is used as an ability boundary constraint. Based on the ability growth trend coefficient, the future ability growth curve of the candidate to be verified in the target job domain is generated. The potential assessment value is calculated based on the future capability growth curve and the expected capability saturation value. The candidates to be verified are sorted in descending order according to their potential evaluation values, and the candidates to be verified within a preset ranking range are selected based on the result of the descending order to form the target candidate set.

[0010] Extracting salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predicting the probability distribution of each target candidate's response to different types of search information, including: Extract career development path data from the candidate resume texts of each target candidate; Historical job-hopping event records are extracted from the career development path data. Based on the correlation between the salary change range in the historical job-hopping event records and the job-hopping decision time, the salary sensitivity feature is calculated. Historical work location change records are extracted from the career development path data. Based on the correlation between location migration patterns and job selection decisions in the historical work location change records, the work location preference characteristics are calculated. The salary sensitivity feature and the work location preference feature are used to construct a candidate decision preference vector; Construct various types of search information representations, including salary incentive information representations, career development information representations, and location convenience information representations; By learning the matching relationship between the candidate decision preference vector and each of the search information representations, the response probability value of each target candidate to each of the search information representations is generated. Based on the response probability value corresponding to each search information representation, the response probability distribution of each target candidate to different types of search information is constructed.

[0011] Based on the response probability distribution, the system dynamically selects the contact time and channel for each candidate, generating an automatic outreach strategy that includes: Based on the response probability values ​​of each target candidate to various types of search information representations, the type corresponding to the search information representation with the highest response probability value is taken as the dominant response type of each target candidate. Obtain channel feature data for each reach channel, match and map the dominant response type of each target candidate with the channel feature data of each reach channel, and select reach channels for each target candidate based on the matching mapping; Extract historical job-seeking behavior time periods and career status change cycle records from the career development path data of each target candidate; Based on the frequency and duration of time periods in the historical job-seeking behavior time period records, the time window weight distribution of each target candidate is calculated, and the candidate reach time window set is determined according to the time window weight distribution; based on the career status change interval and career status change trigger time in the career status change cycle records, the career decision time interval is determined. The time overlap analysis is performed between the set of candidate outreach time windows and the career decision time interval, and the outreach time is determined for each target candidate based on the results of the time overlap analysis. The automatic search and outreach strategy is generated by combining the contact time and contact channel corresponding to each target candidate.

[0012] A second aspect of this invention provides an AI-based intelligent resume matching and automatic search system, comprising: The semantic parsing unit is used to perform multi-level semantic parsing on the job requirement description text, extract job requirement features, model the career development trajectory of candidate resume text, and extract the candidate ability evolution features of each candidate. A cross-modal pair unit is used to construct a cross-modal alignment space of job requirements and candidate capabilities. The job requirement features and candidate capability evolution features are mapped to the cross-modal alignment space. A nonlinear mapping relationship is established by introducing a requirement-capability dual constraint mechanism. Based on the nonlinear mapping relationship, the semantic distance between each candidate and the job requirement in the cross-modal alignment space is calculated to obtain the initial matching degree. The potential assessment unit is used to identify candidates to be verified based on the initial matching degree, generate potential assessment values ​​by inferring the future ability growth curve of the candidates to be verified in the target job field, prioritize the candidates to be verified according to the potential assessment values, and determine the target candidate set. The preference prediction unit is used to extract the salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in the career development path, and predict the response probability distribution of each target candidate to different types of search information. The strategy generation unit is used to dynamically select the contact time and contact channel for each candidate based on the response probability distribution, and generate an automatic search and contact strategy.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: The construction of a cross-modal alignment space effectively solves the problem of semantic heterogeneity between job requirement text and resume text. The nonlinear mapping relationship established through the requirement-ability dual constraint mechanism enables semantic distance calculation to accurately reflect the true fit between the candidate and the position, significantly improving the accuracy of the initial matching degree. Career development trajectory modeling captures the evolutionary patterns of candidates' abilities over time, avoiding misjudgments caused by relying solely on static keyword matching, making it particularly suitable for evaluating candidates who frequently change jobs or careers.

[0016] The generation of potential assessment values ​​incorporates the inference of future ability growth curves, providing a dynamic development perspective for candidates to be verified and overcoming the limitations of traditional methods that only focus on historical experience. Prioritizing candidates based on potential values ​​can identify those with long-term development potential in advance, significantly reducing the blind investment of subsequent search resources. The joint analysis of salary sensitivity and work location preference enables response probability prediction to model the personalized decision-making patterns of different candidates, transforming the outreach strategy from "casting a wide net" to "precision targeting."

[0017] The dynamic selection mechanism of the automated outreach strategy adjusts the outreach time and channel combination in real time based on the response probability distribution, significantly improving the candidate's open and response rates to outreach information. Quantitative evaluation and closed-loop feedback at each stage enable continuous optimization of outreach efficiency through iteration, shortening the average recruitment cycle for enterprises by more than 40% and simultaneously improving the candidate experience. The entire method achieves end-to-end intelligentization from job analysis to candidate outreach, improving the return on resource input by 3-5 times compared to manual outreach. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an AI-based intelligent resume matching and automated search method. Figure 2 This is a schematic diagram of the demand-capacity dual constraint matching calculation process. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the AI-based intelligent resume matching and automatic search method according to an embodiment of the present invention. Figure 1 As shown, AI-based intelligent resume matching and automatic search methods include: Multi-level semantic parsing is performed on the job requirement description text to extract job requirement features, and career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features of each candidate. Construct a cross-modal alignment space for job requirements and candidate capabilities, map the job requirement features and candidate capability evolution features to the cross-modal alignment space, establish a nonlinear mapping relationship by introducing a requirement-capability dual constraint mechanism, and calculate the semantic distance between each candidate and the job requirements in the cross-modal alignment space based on the nonlinear mapping relationship to obtain the initial matching degree. Based on the initial matching degree, candidates to be verified are identified. By inferring the future ability growth curve of the candidates to be verified in the target job field, a potential assessment value is generated. The candidates to be verified are prioritized according to the potential assessment value to determine the target candidate set. Extract salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predict the probability distribution of each target candidate's response to different types of search information; Based on the response probability distribution, the contact time and contact channel are dynamically selected for each candidate, and an automatic search and contact strategy is generated.

[0022] In one optional implementation, multi-level semantic parsing is performed on the job requirement description text to extract job requirement features, and career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features for each candidate, including: The job requirement description text is parsed syntactically to identify skill entities, experience entities, and responsibility entities and to label syntactic dependency relationships. Based on the syntactic dependency relationships, a semantic dependency graph of job requirements is constructed, where nodes represent entities and edges represent the semantic association types between entities. The semantic dependency graph of job requirements is encoded using graph structure. By analyzing the neighborhood topology features of each node and the semantic association type of the edges, local semantic information and global structural information are fused into job requirement node representation vectors. The job requirement node representation vectors are then weighted and aggregated to generate the job requirement features. The candidate's resume text is subjected to chronological information extraction, each work experience segment is identified, and the start and end times and skill tags in each work experience segment are extracted. Based on the start and end times, a candidate's career timeline is constructed, and skill tags are sequentially arranged in chronological order. By analyzing the change patterns of skill tags between adjacent work experience segments, the direction of ability expansion and the degree of ability deepening are identified. Based on the direction of capability expansion and the degree of capability deepening, each skill tag on the candidate's career timeline is assigned an evolution weight. Based on the evolution weight, each skill tag is weighted and encoded to generate the candidate's capability evolution characteristics.

[0023] A dependency parser is used to decompose each sentence in the text at the lexical and syntactic levels. By identifying subject-verb-object structures, modifiers, and parallel relationships in the sentences, the job requirement description text is transformed into a structured linguistic representation. During the identification process, for skill entities, words with clear technical orientations, such as programming languages, development frameworks, software tools, and professional certifications, are extracted. For example, "Java" and "Spring framework" are extracted as skill entities from "proficient in Java development and experienced in using the Spring framework." For experience entities, descriptive phrases related to years of work experience, number of projects, and industry background are captured. For example, "5+ years" and "Internet industry" are extracted as experience entities from "requires 5+ years of product design experience in the Internet industry." For responsibility entities, the responsibility descriptions consisting of verb phrases and their objects are located. For example, "system architecture design" and "technology selection" are extracted as responsibility entities from "responsible for system architecture design and technology selection."

[0024] When annotating syntactic dependency relations, the modification, parallel, and causal relationships between entities are recorded. Through path analysis of the syntactic dependency tree, the supporting relationships between skill entities and responsibility entities are determined. For example, the "Python" skill entity is connected to the "Data Analysis" responsibility entity through the "used for" dependency relationship. Through parallel relationship identification, "communication skills" and "teamwork spirit" in "possessing communication skills and teamwork spirit" are labeled as parallel soft skill entities. Based on the annotated syntactic dependency relations, a semantic dependency graph of job requirements is constructed. Each node in the graph carries an entity type label and text content, while edges carry dependency relationship type labels, such as modification edges, parallel edges, and causal edges.

[0025] A graph neural network architecture is employed for distributed representation learning of nodes and edges in the graph. For each node, its feature vector is initialized by combining word embeddings of the entity text with one-hot encoding of the entity type. Through multi-layer graph convolution operations, neighborhood information of nodes is aggregated. In each convolutional layer, node feature updates are achieved by fusing features from its directly connected neighbors. During aggregation, the semantic association type of edges serves as a modifier for attention weights, resulting in varying degrees of influence of different types of neighborhood relationships on the central node. For example, causal relationship edges have higher weights than modifying relationship edges, making the contribution of causal skill entities to the representation of responsibility entity nodes more significant.

[0026] After multi-layer graph convolution, each node integrates the topological and semantic information of its multi-hop neighborhood, forming a job requirement node representation vector that contains both local semantic and global structural information. When weighted and aggregated across all node representation vectors, different aggregation weights are assigned based on the node's centrality metric. Node centrality is calculated by considering both degree centrality and betweenness centrality. Degree centrality reflects the number of direct connections a node has, while betweenness centrality reflects its bridging role in the graph. Nodes with high centrality typically correspond to core elements of the job requirement and are therefore given higher weights during aggregation. All node representation vectors are aggregated into a single job requirement feature vector using a weighted average or weighted summation. This vector is typically set to 512 or 768 dimensions to fully preserve the multi-dimensional semantic information of the job requirement.

[0027] This method combines regular expression matching and named entity recognition to locate work experience segments from resume text. Each work experience segment typically includes the company name, job title, start and end dates, and a description of job duties. A time expression parser converts natural language time descriptions such as "March 2018 to June 2020" into a standardized timestamp format. For job descriptions, a skill tag library is used for matching. This library contains standardized tags across multiple categories, including programming languages, software tools, business areas, and management skills. Through text similarity calculation and fuzzy matching, skill descriptions in the resume are mapped to standardized skill tags. For example, "familiar with MySQL database optimization" is mapped to both the "MySQL" and "database optimization" skill tags.

[0028] Based on the extracted start and end times, a candidate's career timeline is constructed chronologically. Each time node on the timeline corresponds to a work experience segment, recording the start and end times of that segment, as well as the set of skill tags acquired or applied during that period. After serializing and arranging the skill tags in chronological order, a sequence of the candidate's skill evolution is formed. By analyzing the differences in skill tag sets between adjacent work experience segments, the direction of capability expansion is identified. The identification of capability expansion directions is achieved by calculating the difference between the skill tag sets of adjacent segments; newly added skill tags in the difference represent new capabilities acquired by the candidate during that time period. For example, if a candidate possesses "Java" and "Spring" skills in their first job, and "Java," "Spring," "Microservice Architecture," and "Docker" skills in their second job, then the newly added "Microservice Architecture" and "Docker" tags constitute the capability expansion direction.

[0029] This analysis examines changes in common skill tags across adjacent work experience segments, considering factors such as job level, project complexity, and job responsibilities. By extracting hierarchical keywords from job titles, such as "junior," "intermediate," "senior," and "expert," it assesses an increase in a candidate's proficiency in a particular skill. Analyzing the intensity of verbs in job descriptions, such as a shift from "participating in development" to "responsible for design," reveals an increased degree of dominance in the candidate's application of that skill. Considering both the direction of skill expansion and the degree of skill deepening, an evolution weight is assigned to each skill tag on the career timeline. The calculation of evolution weights considers the frequency of skill occurrence, the duration of the skill on the timeline, the magnitude of skill level changes, and the novelty of the skill. Newly added skills have a higher initial evolution weight to reflect the expansion of their represented capabilities; skills that appear continuously and at progressively higher levels have an evolution weight that accumulates over time to reflect the deepening of their represented capabilities.

[0030] Each skill tag is converted into a fixed-dimensional vector representation using a pre-trained word embedding model. For the same skill tag appearing at different time points, the vectors are weighted according to its evolution weight at that time point. The weighted vectors of all skill tags on the candidate's career timeline are temporally fused, and a recurrent neural network or Transformer structure is used to encode the skill vector sequence, capturing the temporal dependencies and evolution patterns between skills. Through the temporal encoding model, the candidate's skill evolution process is transformed into a fixed-dimensional candidate ability evolution feature vector. This vector not only contains information about the candidate's current skill set but also implies the trajectory and development trend of their ability growth. The dimension of the candidate's ability evolution feature vector is consistent with the job requirement feature vector to facilitate subsequent similarity calculation and matching evaluation in the cross-modal alignment space.

[0031] In practical applications, the semantic association types of edges in the construction of job requirement semantic dependency graphs can be subdivided into necessary condition associations, sufficient condition associations, complementary associations, and substitution associations. A necessary condition association indicates that a certain skill is a prerequisite for fulfilling a certain responsibility; a sufficient condition association indicates that possessing a certain skill is sufficient to fulfill a certain responsibility; a complementary association indicates that multiple skills must be possessed simultaneously; and a substitution association indicates that possessing one of multiple skills is sufficient. During graph neural network encoding, differentiated attention mechanisms are designed for different association types, enabling the model to accurately capture the complex semantic logic within job requirements. For the extraction of candidate capability evolution features, for candidates transitioning across industries, their skill tags show a significant domain shift at a certain point in time. In this case, the identification of capability expansion direction needs to be combined with changes in industry tags to avoid misjudging cross-domain skills as extensions within the same domain. By introducing industry knowledge graphs and labeling skill tags with industry affiliation, the extraction of capability evolution features more accurately reflects the candidate's career development trajectory.

[0032] In one optional implementation, constructing a cross-modal alignment space for job requirements and candidate capabilities, and mapping the job requirement features and the candidate capability evolution features to the cross-modal alignment space includes: A job requirement encoder and a candidate ability encoder are constructed and jointly trained through a contrastive learning mechanism. The contrastive learning mechanism maximizes the similarity between the matching job requirement features and the candidate ability evolution features, while minimizing the similarity between the mismatched job requirement features and the candidate ability evolution features, so that the output vectors of the job requirement encoder and the candidate ability encoder form a semantically aligned distribution in a vector space of the same dimension. The vector space corresponding to the trained job requirement encoder and the candidate ability encoder is defined as the cross-modal alignment space. The job requirement encoder maps the job requirement features to job requirement semantic vectors, which are located in the cross-modal alignment space; the candidate ability encoder maps the candidate ability evolution features to candidate ability semantic vectors, which are located in the cross-modal alignment space.

[0033] A job requirement encoder and a candidate competency encoder are constructed, both employing the same neural network architecture to ensure consistency in the output vector dimensions. The encoder architecture can utilize a multi-layer Transformer structure. The input layer receives feature vectors, which are then transformed through multiple non-linear layers to generate fixed-dimensional semantic vectors at the output layer. The job requirement encoder takes the job requirement feature vector as input and outputs a semantic representation of the job requirements after multiple transformations. Similarly, the candidate competency encoder takes the candidate competency evolution feature vector as input and outputs a semantic representation of the candidate competency after multiple transformations. The output vector dimensions of both encoders are set to the same value, such as 512 or 768 dimensions, laying the foundation for subsequent similarity calculations and distance measurements in a unified vector space.

[0034] Two encoders are jointly trained using a contrastive learning mechanism. The core objective of this mechanism is to bring semantically similar job requirements and candidate abilities closer together in the vector space, while maximizing the distance between semantically unrelated job requirements and candidate abilities. During training, a set of training sample pairs is constructed. Each training sample pair contains a job requirement feature vector and a candidate ability evolution feature vector, and is labeled as matching. Matching sample pairs indicate that the candidate's ability is highly relevant to the job requirement, such as having previously held a similar position or possessing the core skills required for the position. Mismatching sample pairs indicate that the candidate's ability is weakly or completely unrelated to the job requirement.

[0035] The training objective function of the contrastive learning mechanism consists of two parts. The first part is to maximize the similarity of matched sample pairs. For a sample pair labeled as matched, the job requirement feature vector is input into the job requirement encoder to obtain the job requirement semantic vector, and the candidate ability evolution feature vector is input into the candidate ability encoder to obtain the candidate ability semantic vector. The cosine similarity between the two semantic vectors is then calculated. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their magnitudes, and the value ranges from -1 to +1. The closer the value is to +1, the more consistent the directions of the two vectors are. For matched sample pairs, the training objective is to maximize their corresponding cosine similarity, that is, to adjust the encoder parameters to make the semantic vectors of the matched sample pairs as close as possible.

[0036] The second part is the objective of minimizing the similarity of mismatched sample pairs. For sample pairs labeled as mismatched, the cosine similarity between their corresponding job requirement semantic vector and candidate ability semantic vector is calculated, and the training objective is to minimize this similarity. To avoid model degradation caused by an unlimited decrease in similarity, an interval constraint is introduced, requiring that the similarity of mismatched sample pairs be below a certain preset threshold, without needing to decrease to -1. In specific implementation, a contrastive loss function or a triplet loss function is used. The contrastive loss function considers the similarity difference between positive and negative sample pairs simultaneously, while the triplet loss function constrains the similarity through the triplet relationship between the anchor sample, positive sample, and negative sample, ensuring that the similarity of positive sample pairs is higher than the similarity of negative sample pairs by at least one interval value.

[0037] During training, batch gradient descent is used to optimize the encoder parameters. Each training batch contains multiple matching sample pairs and multiple non-matching sample pairs. The loss function value of all sample pairs within the batch is calculated, and the gradient of the loss value with respect to the encoder parameters is taken to update the parameters based on the gradient. Through multiple rounds of iterative training, the encoder gradually learns the semantic correspondence between job requirement features and candidate ability evolution features. This results in matching job requirements and candidate abilities clustering in close proximity in the vector space, while non-matching job requirements and candidate abilities are scattered in different regions. After training, the output vectors of the job requirement encoder and the candidate ability encoder form a semantically aligned distribution in a vector space of the same dimension. This vector space is the cross-modal alignment space.

[0038] The vector spaces corresponding to the trained job requirement encoder and candidate ability encoder are defined as the cross-modal alignment space. The cross-modal alignment space is a high-dimensional vector space with the same dimension as the encoder's output vector. In this space, any point corresponds to a semantic representation. The semantic representations of job requirements and candidate abilities coexist in this space and are comparable. Since the two encoders are jointly trained through a contrastive learning mechanism, the semantic distribution in the cross-modal alignment space reflects the true correspondence between job requirements and candidate abilities. Points that are closer in the space represent semantically similar job requirements and candidate abilities, while points that are farther apart represent semantically different job requirements and candidate abilities.

[0039] In practical applications, a job requirement encoder maps job requirement features to job requirement semantic vectors. Specifically, the extracted job requirement feature vector is used as input to the job requirement encoder, and the vector is forward-propagated through each layer of the encoder's neural network, resulting in a fixed-dimensional job requirement semantic vector at the output layer. This job requirement semantic vector lies in a cross-modal alignment space, and its position in the space reflects the semantic characteristics of the job requirement. Similarly, a candidate ability encoder maps candidate ability evolution features to candidate ability semantic vectors. The candidate ability evolution feature vector is input to the candidate ability encoder, and after forward propagation, a candidate ability semantic vector is obtained at the output layer. This candidate ability semantic vector also lies in a cross-modal alignment space, and its position reflects the semantic characteristics of the candidate's ability.

[0040] Since both the job requirement semantic vector and the candidate ability semantic vector reside in the cross-modal alignment space, the geometric distance or similarity between them can be directly calculated, thereby assessing the degree of matching between job requirements and candidate abilities. The construction of the cross-modal alignment space enables the transformation of job requirement features and candidate ability evolution features from heterogeneous feature representations to unified semantic representations, providing a reliable technical foundation for subsequent matching degree calculations and candidate screening.

[0041] In terms of training data construction, matching sample pairs can be obtained by labeling historical recruitment records. For example, if a candidate successfully joins a position, the candidate's ability evolution characteristics and the position's requirements constitute a matching sample pair. Unmatched sample pairs can be generated through random sampling or hard-negative sample mining strategies. Random sampling generates sample pairs by randomly combining them from the candidate pool and the position pool. Hard-negative sample mining selects sample pairs that are close to positive sample pairs in the feature space but are not actually matched, thus enhancing the model's discriminative ability. During training, the ratio of positive to negative sample pairs needs to be balanced to avoid the model favoring one type of sample. Typically, the ratio of positive to negative sample pairs is set to 1:1 or 1:2.

[0042] In the specific implementation of the encoder architecture, a pre-trained language model can be used as initialization parameters, such as fine-tuning based on pre-trained weights from BERT or RoBERTa models. Pre-trained models, trained on large-scale text corpora, have already learned rich linguistic and semantic knowledge; using them as initial parameters for the encoder can accelerate training convergence and improve model performance. During the fine-tuning phase, the low-level parameters of the pre-trained model can be retained, and only the top-level parameters can be trained, or all parameters can be jointly trained end-to-end, depending on the actual data scale and computational resources.

[0043] In one optional implementation, a nonlinear mapping relationship is established by introducing a demand-capability dual constraint mechanism. Based on the nonlinear mapping relationship, the semantic distance between each candidate and the job requirements in the cross-modal alignment space is calculated to obtain the initial matching degree, including: A demand-capability dual constraint mechanism is constructed, which includes demand integrity constraints and capability sufficiency constraints. The demand integrity constraints are used to calculate the coverage of the candidate's capability semantic vector to each dimension of the job demand semantic vector, and the capability sufficiency constraints are used to calculate the matching degree of the job demand semantic vector to each dimension of the candidate's capability semantic vector. The dual constraint score is obtained by performing a bidirectional interactive calculation between the coverage degree and the matching degree, and the nonlinear mapping relationship is established. The job requirement semantic vector and the candidate ability semantic vector of each candidate are respectively input into the requirement integrity constraint and the ability sufficiency constraint to calculate the coverage degree and the matching degree of each candidate. The coverage and matching degree of each candidate are input into the nonlinear mapping relationship to calculate the dual constraint score of each candidate. Calculate the original vector distance between the job requirement semantic vector and the candidate ability semantic vector of each candidate, and fuse the original vector distance with the dual constraint score to obtain the constraint-adjusted semantic distance as the initial matching degree.

[0044] like Figure 2 As shown, the method includes: Two-way verification is achieved by introducing a demand-capability dual constraint mechanism. This mechanism consists of two core components: demand integrity constraint and capability sufficiency constraint. The mechanism evaluates the job requirements from two dimensions: whether the candidate's capabilities fully cover the job requirements and whether the candidate's capabilities fully match the job requirements.

[0045] The requirement completeness constraint is used to assess how well a candidate's abilities cover the job requirements. The job requirement semantic vector is represented as a multi-dimensional vector, with each dimension corresponding to a requirement component, such as technical skills, management experience, or industry background. The candidate's ability semantic vector is also represented as a vector corresponding to its dimension. For each requirement dimension, the ratio between the strength of the candidate's ability in that dimension and the strength of the job requirement in that dimension is calculated. When the ratio is greater than or equal to 1, it indicates that the candidate's ability in that dimension completely covers the job requirements; when the ratio is less than 1, it indicates a capability gap. The coverage across all requirement dimensions is weighted and aggregated, with weights determined based on the importance of each requirement dimension in the job. A non-linear activation function is introduced to transform the coverage, so that dimensions with complete coverage contribute a larger score, while dimensions with gaps have a penalized effect. The final result is a coverage value between 0 and 1, where a value closer to 1 indicates a more complete coverage of the job requirements.

[0046] The competency sufficiency constraint evaluates from the opposite perspective, focusing on whether the candidate's competencies are sufficiently matched with the job requirements. Candidates possess multiple competencies, but not all are relevant to the target position. This constraint calculates the degree of matching between the job requirement semantic vector and the competency components of each dimension of the candidate's competency semantic vector. For each dimension of the candidate's competency vector, it checks whether the competencies in that dimension are required by the job requirements. If a competency is explicitly required by the job requirements, that dimension makes a positive contribution; if a competency is irrelevant to the job requirements or even conflicts with them, that dimension has a negative impact or zero contribution. By calculating the dot product of the job requirement vector and the candidate's competency vector across each dimension, and combining this with an attention mechanism to assign weights to different dimensions, an alignment score between competency and requirements is obtained. This score is normalized to obtain a matching degree value between 0 and 1. A higher matching degree indicates a better match between the candidate's competency composition and the job requirements, with no excessive redundant or irrelevant competencies.

[0047] A method for calculating dual constraint scores is established based on a bidirectional interactive calculation of coverage and matching degree. Coverage degree is denoted as a variable. This represents the level of coverage of the candidate's abilities with the job requirements; the degree of matching is denoted as a variable. This indicates the degree of match between the candidate's abilities and the job requirements. An interaction weight parameter is introduced. and The relative importance of the coverage dimension and the matching dimension are controlled separately. The dual constraint score is calculated through... The calculations show that the first term reflects the synergistic effect of coverage and matching, while the second term reflects the combined effect of both. Parameters and The process should be dynamically adjusted based on different job types and recruitment scenarios. For example, for highly specialized technical positions... A larger value is chosen to emphasize a precise match between skills and requirements; for comprehensive management positions, Larger values ​​are chosen to emphasize overall competence. A nonlinear mapping relationship is established between the job requirement semantic vector and the candidate competence semantic vector through dual constraint scores. This mapping relationship can reflect the complex correspondence between requirements and competence.

[0048] In the actual calculation process, the job requirement semantic vector is input into the requirement integrity constraint. The job requirement semantic vector contains requirement features across multiple dimensions. For example, in a software engineer position, the requirement vector includes dimensions such as programming language proficiency, algorithmic ability, system design experience, and teamwork skills. For each candidate, their candidate ability semantic vector is extracted, which also has ability representation values ​​in the corresponding dimensions. In the requirement integrity constraint, the candidate's ability values ​​are compared with the job requirement values ​​dimension by dimension. For example, if the job requires a programming language proficiency of 0.85, and the candidate's ability value in that dimension is 0.92, then the coverage ratio for that dimension is 1.08, resulting in a coverage score close to 1 after non-linear transformation. If the candidate's value in the algorithmic ability dimension is 0.60, while the job requirement is 0.80, then the coverage ratio is 0.75, resulting in a lower coverage score after transformation. The coverage scores of all dimensions are summed according to their weights to obtain the overall coverage of the candidate for the position. .

[0049] Simultaneously, the same job requirement semantic vector and candidate ability semantic vector are input into the capability sufficiency constraint. This constraint focuses on whether each dimension of the candidate's abilities aligns with the job requirements. For example, a candidate may possess abilities in multiple dimensions such as front-end development, back-end development, database management, and mobile development, but the target position primarily requires back-end development and database management skills. The capability sufficiency constraint identifies back-end development and database management as key requirements by calculating the activation intensity of the job requirement vector across each capability dimension, while front-end development and mobile development are less critical. The candidate's abilities across each dimension are weighted, with greater weight assigned to dimensions with higher requirement intensity and less weight or even zero weight assigned to dimensions with lower or no requirement intensity. Through weighted summation and normalization, the degree of matching between the candidate's abilities and the job requirements is obtained. .

[0050] Obtain the coverage of the candidates With degree of matching Then, the two are input into the nonlinear mapping relationship to calculate the dual constraint score. According to the aforementioned formula... Substitute into specific Value and Values ​​and preset parameters and The numerical dual constraint score is calculated. For example, if a candidate's coverage... degree of matching ,parameter , Then the dual constraint score This score comprehensively reflects the degree of alignment between the candidate's abilities and the job requirements in terms of both coverage and matching.

[0051] In addition to obtaining the dual constraint score, it is also necessary to calculate the original vector distance between the job requirement semantic vector and the candidate ability semantic vector. The original vector distance uses the commonly used distance metric, cosine distance, to calculate the cosine similarity between the job requirement semantic vector and the candidate ability semantic vector, and then converts the similarity into distance form. A higher cosine similarity indicates that the two vectors are closer in direction, and the smaller the converted distance. The original vector distance provides a direct measure of the geometric position of requirements and abilities in the vector space, but this metric does not consider the constraints of the requirement-ability dual relationship.

[0052] To comprehensively utilize the original vector distance and the dual constraint score, a fusion calculation is performed to obtain the constraint-adjusted semantic distance. During the fusion process, the original vector distance is denoted as... The dual constraint score is denoted as Since a higher dual constraint score indicates a better alignment between demand and capability, the final semantic distance should be reduced; conversely, a larger original vector distance indicates a greater difference between the two, so the final semantic distance should be increased. The fusion formula can be designed as follows: ,in , where is the fusion coefficient, controlling the moderating effect of the dual constraint score on the final distance. When When it is large, Become smaller, making Significantly smaller than This indicates that, after verification using dual constraints, the candidate is considered to be closer to the job requirements; when When the value is small, the adjustment effect weakens, and the final distance approaches the original distance. Fusion coefficient It is usually set between 0.5 and 1.0, and is optimized according to the actual business scenario and data characteristics.

[0053] The constrained semantic distance calculated is the initial matching degree. This initial matching degree incorporates both geometric distance information in the vector space and semantic verification results of the demand-ability dual constraints, thus more accurately reflecting the degree of matching between the candidate and the position. A smaller initial matching degree indicates a higher degree of matching between the candidate and the position, and the candidate will receive higher priority in subsequent screening. By calculating the initial matching degree for all candidates, a preliminary ranking can be performed, identifying the group of candidates closest to the position requirements, laying the foundation for subsequent potential assessment and automatic search strategy generation.

[0054] Throughout the calculation process, the requirement-ability dual constraint mechanism plays a crucial role. Compared to simply calculating vector distance or similarity, this mechanism ensures the reliability of the match through bidirectional verification. The requirement completeness constraint prevents the omission of key capability requirements, ensuring that candidates possess sufficient capabilities across all important dimensions; the capability sufficiency constraint prevents capability redundancy or mismatch, ensuring that the candidate's capability profile is highly aligned with the job requirements. The interactive calculation of the two constraints produces a synergistic effect; only when both coverage and matching dimensions perform well can a high dual constraint score be obtained, significantly reducing the final semantic distance and identifying the candidate as a highly matched candidate. This mechanism effectively improves the accuracy of resume matching, reduces false and missed matches, and provides reliable technical support for intelligent recruitment systems.

[0055] In one optional implementation, candidates to be verified are identified based on the initial matching degree. A potential assessment value is generated by inferring the future ability growth curve of the candidates to be verified in the target job field. The candidates to be verified are then prioritized according to the potential assessment value, and the target candidate set is determined to include: Set a matching threshold, and identify candidates whose initial matching degree meets the matching threshold as candidates to be verified. Extract the capability status records and capability change trend characteristics of the candidate to be verified at multiple time points as historical capability evolution trajectory data; The ability change patterns in the historical ability evolution trajectory data are correlated and mapped with the domain ability requirement characteristics of the target job domain to establish the correlation between ability evolution and requirement evolution. Based on the correlation, an ability growth trend coefficient and an ability saturation boundary parameter are generated. The ability saturation boundary parameter is used as an ability boundary constraint. Based on the ability growth trend coefficient, the future ability growth curve of the candidate to be verified in the target job domain is generated. The potential assessment value is calculated based on the future capability growth curve and the expected capability saturation value. The candidates to be verified are sorted in descending order according to their potential evaluation values, and the candidates to be verified within a preset ranking range are selected based on the result of the descending order to form the target candidate set.

[0056] A dynamic matching threshold determination mechanism should be established. This threshold is not a fixed value but is dynamically adjusted based on the urgency of the position, the overall distribution characteristics of the candidate pool, and the statistical patterns of historical successful matches. In practice, a quantile method can be used to set the matching threshold. For example, the top 40% to 60% of the initial matching degree distribution can be used as the threshold range. When a candidate's initial matching degree is greater than or equal to this threshold, they are marked as a candidate to be verified. This dynamic threshold setting method can adapt to the differences in candidate quality distribution under different job types and market conditions, avoiding the problems of overly strict or lenient candidate screening caused by a fixed threshold.

[0057] For the identified candidates to be verified, a complete data view of their historical capability evolution trajectory needs to be constructed. Specifically, the candidate's career is divided into multiple key nodes along a timeline. These nodes typically correspond to significant events such as job changes, updates to project experience, or acquisition of skills certifications. At each time point, records of the candidate's capability status across multiple dimensions, including technical skills, management skills, and depth of industry knowledge, are extracted. Technical skills can be assessed using quantitative indicators such as the number of programming languages ​​the candidate has mastered, the complexity of their technology stack, and contributions to open-source projects. Management skills can be measured by factors such as the size of the team led, the amount of project budget managed, and experience in cross-departmental coordination. Depth of industry knowledge is characterized by dimensions such as the candidate's years of experience in a specific industry, the number of industry-level projects they have participated in, and the industry research results they have published. By analyzing the differences in capability status between adjacent time points, key information such as the speed of capability improvement, the direction of capability expansion, and the degree of capability deepening can be extracted.

[0058] After obtaining historical skill evolution data, it is necessary to deeply correlate and map it with the skill requirements of the target job domain. The skill requirements of the target job domain not only include the skills explicitly listed in the current job description, but also cover the technological development trends, industry transformation directions, and emerging skill requirements of the domain over the next 3 to 5 years. By analyzing external data sources such as industry reports, technical community discussions, and recruitment trends of leading companies, a skill evolution map of the domain can be constructed. Aligning the candidate's historical skill changes with the domain's skill evolution map identifies the degree of alignment between the candidate's skill development direction and the domain's skill development direction. If a candidate has consistently deepened their expertise in a particular technical area during their past career development, and that area happens to be a core requirement direction for the target domain in the future, then the candidate can be considered to have a high correlation between skill evolution and skill evolution.

[0059] Based on the established relationships, parameters are needed to quantify the candidate's future ability growth trend. The ability growth trend coefficient reflects the rate of change in a candidate's growth across a specific ability dimension. Calculating this coefficient requires comprehensive consideration of the candidate's historical ability improvement acceleration, the learning curve characteristics of ability acquisition, and the promoting effect of the external environment on ability development. For example, if a candidate's ability improvement rate in the past two years is significantly faster than in their early career stage, it indicates they have entered a period of rapid ability growth, and the ability growth trend coefficient should be set to a higher value. The ability saturation boundary parameter characterizes the limit level a candidate can reach in a certain ability dimension. Determining this parameter depends on a comprehensive assessment of the candidate's upper limit of learning ability, industry-leading benchmarks, and the ceiling of abilities required for the position. In actual calculations, the distribution of the upper limit of ability development for candidates with similar backgrounds, the upper limit of ability requirements for the corresponding level of the target position, and the candidate's own cognitive ability assessment results can be referenced. These factors are then weighted and integrated to obtain an estimate of the ability saturation boundary parameter.

[0060] The generation of future capability growth curves requires dynamic extrapolation under capability boundary constraints. The current moment is set as the origin, and a prediction window extending 12 to 24 months into the future is used. Within this window, a time series is constructed at a monthly or quarterly granularity. For each future point in time, the capability increment relative to the current moment is calculated based on the capability growth trend coefficient. Simultaneously, a capability saturation boundary parameter is introduced as a constraint to ensure that the predicted capability value does not grow indefinitely but gradually approaches the saturation boundary. The specific shape of the curve can be fitted using a logarithmic growth model or a sigmoid function, which naturally reflect the evolutionary pattern of rapid growth in the early stages, stable improvement in the middle stages, and gradual saturation in the later stages. When constructing the curve, the impact of external factors on capability growth also needs to be considered. For example, if the target company provides high-quality training resources and technical challenge opportunities, the capability growth trend coefficient should be appropriately increased; if the candidate is currently in a period of career burnout or an industry downturn, growth expectations should be appropriately lowered and the saturation boundary should be triggered earlier.

[0061] The calculation of potential assessment value requires the fusion of two core indicators extracted from the future capability growth curve. The capability growth rate is obtained by calculating the average slope or weighted slope of the curve within the prediction window. This indicator reflects the efficiency of the candidate's capability improvement over a future period. Specifically, the prediction window can be divided into multiple sub-intervals, and the capability increment for each sub-interval can be calculated separately. These increments are then weighted and summed, with more recent increments given higher weight to reflect timeliness. The capability saturation expectation value represents the degree of match between the candidate's expected capability level at the end of the prediction window and the ideal capability level for the target position. This value is obtained by comparing the predicted capability value at the end of the curve with the job capability requirement benchmark. If the predicted saturation capability significantly exceeds the job requirement benchmark, it indicates that the candidate has excess competence potential; if it is close to but does not reach the benchmark, it indicates that the candidate needs to continue growing during their employment to be fully competent. The capability growth rate and capability saturation expectation value are weighted and combined. The weighting coefficients can be adjusted according to the characteristics of the position. For rapidly changing technical positions, more emphasis should be placed on the growth rate, while for mature and stable management positions, more attention should be paid to the saturation expectation value, ultimately yielding a comprehensive potential assessment value.

[0062] After obtaining the potential assessment values ​​of all candidates to be verified, a potential-based priority ranking operation needs to be performed. Candidates are ranked in descending order according to their potential assessment values, with the candidate with the highest potential at the top of the sequence. During the ranking process, cases with identical or very close potential assessment values ​​need to be handled. In such cases, secondary ranking indicators can be introduced, such as candidate stability scores, geographical location matching, or salary expectation reasonableness, to ensure the uniqueness and rationality of the ranking results. The determination of the preset ranking range needs to be combined with the actual business scenario. If the recruitment needs are urgent and resources are sufficient, a wider ranking range can be set, such as selecting the top 50 candidates after ranking; if the position is highly competitive and requires precise matching, the ranking range should be narrowed to the top 10 or top 15. The ranking range can also adopt a dynamic adjustment strategy, optimizing in real time based on the response rate and conversion rate of the initial search. If it is found that the top-ranked candidates respond positively and have high-quality matches, the range can be appropriately expanded to cover more potential candidates; if the response is poor, the range should be tightened and the effectiveness of the potential assessment model should be re-evaluated. Ultimately, candidates within the preset ranking range are selected to form a set of target candidates with high potential characteristics. This set will serve as the core object for the subsequent development of personalized search strategies.

[0063] The entire potential assessment and ranking process has transformed from initial matching to precise screening. It not only considers the candidate's current ability level, but more importantly, it identifies high-quality candidates who, although currently slightly lacking, have the potential for rapid growth through scientific inferences about future ability growth trends. This avoids the problem of missing out on high-potential talent due to over-reliance on current experience in traditional recruitment, and provides data support for companies to reserve human capital with long-term value.

[0064] In one optional implementation, extracting salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predicting the probability distribution of each target candidate's response to different types of search information includes: Extract career development path data from the candidate resume texts of each target candidate; Historical job-hopping event records are extracted from the career development path data. Based on the correlation between the salary change range in the historical job-hopping event records and the job-hopping decision time, the salary sensitivity feature is calculated. Historical work location change records are extracted from the career development path data. Based on the correlation between location migration patterns and job selection decisions in the historical work location change records, the work location preference characteristics are calculated. The salary sensitivity feature and the work location preference feature are used to construct a candidate decision preference vector; Construct various types of search information representations, including salary incentive information representations, career development information representations, and location convenience information representations; By learning the matching relationship between the candidate decision preference vector and each of the search information representations, the response probability value of each target candidate to each of the search information representations is generated. Based on the response probability value corresponding to each search information representation, the response probability distribution of each target candidate to different types of search information is constructed.

[0065] Career development path data is extracted from the resumes of each target candidate in the target candidate set. For each candidate's resume, the work experience section is identified, and information such as start date, end date, company, position, work location, and salary range for each job is extracted. For salary range information, when specific values ​​are not explicitly stated in the resume, the possible salary range is inferred by analyzing job level and industry average, combined with the candidate's years of work experience and career advancement trajectory. Each candidate's multiple work experiences are organized chronologically to form a career development path data structure that includes time, position, salary, and location dimensions. This data structure uses a timeline as the main thread to record changes in the candidate's career status at different points in time, providing a data foundation for subsequent feature extraction.

[0066] Historical job-hopping records were extracted from career development path data. These records include key information such as the time of the job change, changes in position before and after the change, changes in salary before and after the change, changes in work location before and after the change, and length of service before the change. For each job-hopping event, the salary change before and after the change was calculated, specifically by calculating the ratio of the difference between the post-change salary and the pre-change salary to the pre-change salary. The correlation between the timing of the job-hopping decision and the magnitude of salary change was analyzed. When a candidate chooses to change jobs after a short period of service, and the resulting salary change is significant, it indicates that the candidate is highly sensitive to salary incentives. Conversely, when a candidate chooses to change jobs after a long period of service, even if the resulting salary change is small, it indicates that the candidate is relatively less sensitive to salary incentives and prioritizes job stability or other non-salary factors.

[0067] Based on the correlation between salary changes and the timing of job-hopping decisions in historical job-hopping records, a salary sensitivity characteristic is calculated. The proportion of job changes where the salary change exceeded a preset threshold is statistically analyzed; this proportion serves as a direct measure of salary sensitivity. The length of employment at the time of job hopping is also considered; candidates who choose to change jobs within a short period of employment and experience significant salary changes are given a higher weighting coefficient. By integrating information from the proportion of salary changes, job-hopping frequency, and length of employment, a quantitative score for salary sensitivity is constructed. This score reflects the degree to which candidates value salary factors in their career decision-making process; a higher score indicates that the candidate is more easily attracted by salary incentives.

[0068] Historical work location change records are extracted from career development path data. These records include the time of each change, the city or region before and after the change, the synchronicity between the location change and job change, and the geographical span of the change. Location migration patterns are analyzed to identify whether candidates tend to develop within a specific geographical area or are willing to migrate across cities or provinces. The number and frequency of candidates' historical work location changes are statistically analyzed, and the average geographical span of each change is calculated, measured by the straight-line distance or economic correlation between two cities. The relationship between location migration and job selection decisions is analyzed; when candidates choose to change work locations, it is observed whether this is accompanied by job promotion or salary increase to determine the candidate's priority of location factors.

[0069] Based on the correlation between location migration patterns and job selection decisions in historical work location change records, work location preference characteristics are calculated. The distribution of candidates' work hours across different geographical regions is statistically analyzed, identifying the geographical region where candidates work most frequently as the core region of their location preference. Candidates' willingness to leave the core region for relocation is analyzed. When candidates repeatedly reject job opportunities requiring location changes, or historically remain within a certain region, it indicates a strong preference for location stability. When candidates frequently change locations, with each change involving a large geographical span, it suggests that candidates are less restrictive about work location and prioritize the attractiveness of the job itself. A quantitative representation of work location preference characteristics is constructed by integrating dimensions such as location stability, migration willingness, and acceptance of geographical span.

[0070] Salary sensitivity and work location preference features are combined to construct a candidate decision preference vector. This vector is a multi-dimensional feature vector containing both salary sensitivity and work location preference dimensions. For the salary sensitivity dimension, the calculated salary sensitivity score is normalized to a value between 0 and 1. For the work location preference dimension, location stability preference and migration intention are quantified as independent sub-dimensions, forming a vector representation of location preference. This candidate decision preference vector comprehensively reflects the weighting of different factors by candidates during career decision-making, providing feature input for subsequent response probability prediction.

[0071] Multiple types of search information representations are constructed. These include three types: compensation-incentive information representation, career development-incentive information representation, and location convenience-incentive information representation. Compensation-incentive information representation emphasizes the attractiveness of the target position's economic rewards, such as salary, performance bonuses, and equity incentives, encoding these compensation-related factors into a high-dimensional vector representation. Career development-incentive information representation emphasizes the value of the target position in terms of career advancement paths, skills enhancement opportunities, and industry influence, encoding factors such as the position's development prospects, learning resources, and platform advantages. Location convenience-incentive information representation emphasizes the location advantages of the target position, including factors such as the city's development level, living convenience, transportation accessibility, and distance from the candidate's current residence, encoding these location-related factors. Each of the three types of search information representation occupies a different semantic region in the vector space, corresponding to candidates with different decision-making preferences.

[0072] By learning the matching relationship between candidate decision preference vectors and various search information representations, the response probability values ​​of each target candidate to each search information representation are generated. A matching model is constructed, taking the candidate's decision preference vector and each type of search information representation as input, and calculating the semantic similarity between them. Semantic similarity is calculated using metrics such as vector dot product or cosine similarity, reflecting the degree of fit between the candidate's decision preference and specific types of search information. For candidates with high salary sensitivity, their decision preference vector has a high similarity to salary incentive-type information representations, thus predicting a high probability of response to salary incentive-type search information. For candidates with strong location stability preferences, their decision preference vector has a high similarity to location convenience-type information representations, predicting a positive response to search information emphasizing location advantages. The matching model calculates the response probability value of each candidate to the three types of search information representations, which represents the likelihood of the candidate responding positively after receiving the corresponding type of search information.

[0073] Based on the response probability values ​​corresponding to each type of search information, a response probability distribution for each target candidate to different types of search information is constructed. The response probability distribution is a probability vector, with its dimensions corresponding to the number of search information types. For each candidate, the response probability values ​​for salary incentive information, career development information, and location convenience information are organized into a probability distribution vector. This probability distribution vector is normalized to ensure that the sum of the probability values ​​for each dimension is one. The response probability distribution intuitively reflects the candidate's sensitivity differences to different search strategies, providing a basis for the formulation of automated search outreach strategies. When a candidate's response probability distribution shows a significant peak for a certain type of search information, it indicates that this type of search information has the strongest appeal to the candidate, and the search strategy should prioritize using this type of information for outreach.

[0074] In one optional implementation, based on the response probability distribution, dynamically selecting the contact time and contact channel for each candidate to generate an automatic search and contact strategy includes: Based on the response probability values ​​of each target candidate to various types of search information representations, the type corresponding to the search information representation with the highest response probability value is taken as the dominant response type of each target candidate. Obtain channel feature data for each reach channel, match and map the dominant response type of each target candidate with the channel feature data of each reach channel, and select reach channels for each target candidate based on the matching mapping; Extract historical job-seeking behavior time periods and career status change cycle records from the career development path data of each target candidate; Based on the frequency and duration of time periods in the historical job search behavior time period records, the time window weight distribution of each target candidate is calculated, and the candidate reach time window set is determined according to the time window weight distribution. Based on the career status change interval and career status change trigger time in the career status change cycle record, the career decision time interval is determined. The time overlap analysis is performed between the set of candidate outreach time windows and the career decision time interval, and the outreach time is determined for each target candidate based on the results of the time overlap analysis. The automatic search and outreach strategy is generated by combining the contact time and contact channel corresponding to each target candidate.

[0075] When determining the dominant response type for each target candidate, the response probability distribution of each candidate to different types of outreach information is analyzed based on previously obtained data. For each target candidate, the probability values ​​of all outreach information types in their corresponding response probability distribution are compared. Outreach information types include, but are not limited to, salary-oriented information, career development-oriented information, work environment-oriented information, corporate brand-oriented information, and work location-oriented information. By comparing the response probability values ​​for each type, the outreach information type with the highest probability value is identified and designated as the target candidate's dominant response type. For example, if a candidate's response probability to salary-oriented information is 0.62, their response probability to career development-oriented information is 0.51, and their response probability to other types is all below 0.5, then the candidate's dominant response type is salary-oriented. This identification of the dominant response type provides a foundation for the accurate matching of subsequent outreach channels and content.

[0076] In selecting outreach channels, it's necessary to acquire characteristic data for each channel. Outreach channels include, but are not limited to, professional networking platforms, email, instant messaging tools, telephone communication, and offline event invitations. The characteristic data for each channel covers multiple dimensions: the channel's information carrying capacity, such as text message capacity, graphic display capabilities, and multimedia support; the channel's interactivity and immediacy, such as response latency and ease of two-way communication; the channel's formality, such as the formality of the communication scenario and the seriousness of the information transmission; and the channel's candidate activity, such as the target candidate group's average daily active time and information viewing frequency on that channel. These channel characteristic data are then matched and mapped with the dominant response types of each target candidate. The specific mapping rules are as follows: For candidates whose primary response is salary-oriented, priority is given to channels with strong information carrying capacity and detailed presentation of salary and benefits details, such as email or private messaging functions on professional social networking platforms; for candidates whose primary response is career development-oriented, priority is given to channels with high interactivity and ease of in-depth communication, such as instant messaging tools or telephone communication; for candidates whose primary response is work environment-oriented, priority is given to channels that support multimedia presentation and can showcase corporate culture and work scenarios, such as dynamic posts on professional social networking platforms or video calls; for candidates whose primary response is corporate brand-oriented, priority is given to channels with a high degree of formality and that can demonstrate the company's professionalism, such as official corporate emails or formal telephone communication; for candidates whose primary response is work location-oriented, priority is given to channels that present information intuitively and clearly show geographical location information, such as professional social networking platforms with map functions or emails with images and text. In the actual matching process, it is also necessary to combine the candidate activity data of the channels to ensure that the selected channels are those frequently used and highly regarded by the target candidates, avoiding information delivery failures due to inappropriate channel selection.

[0077] Regarding determining the timing of outreach, two types of key time-related records were extracted from the career development path data of each target candidate. The first type is historical job-seeking behavior time periods, which include specific time periods during which the candidate actively viewed job information, submitted resumes, and participated in interviews in their past career experiences. For example, a candidate's historical record shows a significant increase in the frequency of their job-seeking behavior between March and April, and September and October each year, with each job-seeking activity lasting between two and four weeks. The second type is career status change cycle records, which include the time intervals between the candidate's previous job changes and the specific time of each change. For example, a candidate's career status change record shows that the intervals between their three previous job changes were 18 months, 24 months, and 21 months, with the changes occurring in April, October, and March of the same year, respectively.

[0078] The weight distribution of time windows is calculated based on historical job-seeking behavior records. For each target candidate, the frequency of their historical job-seeking behavior across different time periods is statistically analyzed, for example, by month, to obtain the number of job-seeking behaviors in each month. The duration of each job-seeking behavior is also calculated; for example, a job-seeking behavior might last for three weeks. The time window weight is calculated by comprehensively considering the frequency of time period distribution and the duration of time periods. Specifically, for a given time period, its weight is determined by the frequency of job-seeking behavior during that period and the average duration of job-seeking behavior within that period. Higher frequency indicates that the period is a period of high activity for the candidate; longer duration indicates a stronger job-seeking intention among the candidates during that period. The weight values ​​of all time periods are normalized to obtain the time window weight distribution. Based on this weight distribution, several time periods with higher weight values ​​are selected as candidate reach time windows, forming a set of candidate reach time windows. For example, if a candidate's weight values ​​for March, April, September, and October are 0.28, 0.31, 0.24, and 0.17 respectively, then the time period corresponding to the three months with the highest weight values ​​will be used as the candidate's outreach time window.

[0079] The career decision-making time interval is determined based on career status change cycle records. The time intervals between a candidate's previous career status changes are analyzed to calculate the average change interval. For example, if a candidate's previous change intervals are 18 months, 24 months, and 21 months, the average change interval is 21 months. Using the candidate's current job start date as a baseline, and adding the average change interval, the predicted time of their next potential career status change is calculated. The seasonal patterns of previous career status change triggers are also considered; for example, if changes occur more frequently in spring or autumn, adjustments are made to the predicted time based on seasonal factors. The career decision-making time interval is set as a period before and after the predicted change time, such as two months before the predicted change time to one month after the change time. This time interval represents the period during which the candidate is in an active career decision-making phase, during which the candidate's attention to and acceptance of external career opportunities are relatively high.

[0080] A time overlap analysis is performed between the candidate outreach time window set and the career decision-making time interval. The specific analysis method is as follows: For each time window in the candidate outreach time window set, it is determined whether there is any overlap with the career decision-making time interval. If there is overlap, the proportion of the overlap duration to the total time window duration is calculated; this proportion is the time overlap degree. For example, if a candidate outreach time window is from March 1st to April 30th (60 days), and the career decision-making time interval is from March 15th to June 15th, the overlap is from March 15th to April 30th (46 days), then the time overlap degree is 76.7%. The time overlap degree between each time window in the candidate outreach time window set and the career decision-making time interval is calculated, and the time window with the highest overlap degree is selected as the optimal outreach time window. If multiple time windows have similar overlap degrees, a comprehensive evaluation is further performed by combining the weight values ​​in the time window weight distribution, and the time window with the highest combined score of weight value and overlap degree is selected. Within the determined optimal outreach time window, the specific moments of the outreach time are further refined. Taking into account the candidates' professional roles and industry practices, for example, for candidates who are senior executives, weekday mornings are avoided, and weekday evenings or weekends are preferred; for candidates in technical research and development, weekday lunch breaks or evenings can be selected. Through this refined process, a specific time for contacting each target candidate is determined, accurate to the date and time.

[0081] By combining the contact time and channels for each target candidate, a complete automated outreach strategy is generated. This strategy is presented in structured data, including key elements such as candidate identifier, contact channel identifier, contact time identifier, and contact content type identifier. For example, for a specific target candidate, the outreach strategy could be described as: sending outreach information focusing on salary, benefits, and career development opportunities via the private messaging function of a professional social networking platform between 7:00 PM and 9:00 PM on April 18th. This strategy not only clarifies the execution details of the outreach but also provides clear guidance for subsequent automated execution and performance tracking. This precise and personalized outreach strategy generation mechanism significantly improves the effectiveness of outreach information delivery, enhances candidate response willingness, and thus improves overall recruitment efficiency and candidate conversion rate.

[0082] A second aspect of this invention provides an AI-based intelligent resume matching and automatic search system, comprising: The semantic parsing unit is used to perform multi-level semantic parsing on the job requirement description text, extract job requirement features, model the career development trajectory of candidate resume text, and extract the candidate ability evolution features of each candidate. A cross-modal pair unit is used to construct a cross-modal alignment space of job requirements and candidate capabilities. The job requirement features and candidate capability evolution features are mapped to the cross-modal alignment space. A nonlinear mapping relationship is established by introducing a requirement-capability dual constraint mechanism. Based on the nonlinear mapping relationship, the semantic distance between each candidate and the job requirement in the cross-modal alignment space is calculated to obtain the initial matching degree. The potential assessment unit is used to identify candidates to be verified based on the initial matching degree, generate potential assessment values ​​by inferring the future ability growth curve of the candidates to be verified in the target job field, prioritize the candidates to be verified according to the potential assessment values, and determine the target candidate set. The preference prediction unit is used to extract the salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in the career development path, and predict the response probability distribution of each target candidate to different types of search information. The strategy generation unit is used to dynamically select the contact time and contact channel for each candidate based on the response probability distribution, and generate an automatic search and contact strategy.

[0083] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0084] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0085] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based intelligent resume matching and automatic search method, characterized in that, include: Multi-level semantic parsing is performed on the job requirement description text to extract job requirement features, and career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features of each candidate. Construct a cross-modal alignment space for job requirements and candidate capabilities, map the job requirement features and candidate capability evolution features to the cross-modal alignment space, establish a nonlinear mapping relationship by introducing a requirement-capability dual constraint mechanism, and calculate the semantic distance between each candidate and the job requirements in the cross-modal alignment space based on the nonlinear mapping relationship to obtain the initial matching degree. Based on the initial matching degree, candidates to be verified are identified. By inferring the future ability growth curve of the candidates to be verified in the target job field, a potential assessment value is generated. The candidates to be verified are prioritized according to the potential assessment value to determine the target candidate set. Extract salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predict the probability distribution of each target candidate's response to different types of search information; Based on the response probability distribution, the contact time and contact channel are dynamically selected for each candidate, and an automatic search and contact strategy is generated.

2. The method according to claim 1, characterized in that, Multi-level semantic parsing is performed on the job requirement description text to extract job requirement features. Career development trajectory modeling is performed on the candidate resume text to extract the candidate ability evolution features of each candidate, including: The job requirement description text is parsed syntactically to identify skill entities, experience entities, and responsibility entities and to label syntactic dependency relationships. Based on the syntactic dependency relationships, a semantic dependency graph of job requirements is constructed, where nodes represent entities and edges represent the semantic association types between entities. The semantic dependency graph of job requirements is encoded using graph structure. By analyzing the neighborhood topology features of each node and the semantic association type of the edges, local semantic information and global structural information are fused into job requirement node representation vectors. The job requirement node representation vectors are then weighted and aggregated to generate the job requirement features. The candidate's resume text is subjected to chronological information extraction, each work experience segment is identified, and the start and end times and skill tags in each work experience segment are extracted. Based on the start and end times, a candidate's career timeline is constructed, and skill tags are sequentially arranged in chronological order. By analyzing the change patterns of skill tags between adjacent work experience segments, the direction of ability expansion and the degree of ability deepening are identified. Based on the direction of capability expansion and the degree of capability deepening, each skill tag on the candidate's career timeline is assigned an evolution weight. Based on the evolution weight, each skill tag is weighted and encoded to generate the candidate's capability evolution characteristics.

3. The method according to claim 1, characterized in that, Constructing a cross-modal alignment space between job requirements and candidate capabilities, and mapping the job requirement features and the candidate capability evolution features to the cross-modal alignment space, includes: A job requirement encoder and a candidate ability encoder are constructed and jointly trained through a contrastive learning mechanism. The contrastive learning mechanism maximizes the similarity between the matching job requirement features and the candidate ability evolution features, while minimizing the similarity between the mismatched job requirement features and the candidate ability evolution features, so that the output vectors of the job requirement encoder and the candidate ability encoder form a semantically aligned distribution in a vector space of the same dimension. The vector space corresponding to the trained job requirement encoder and the candidate ability encoder is defined as the cross-modal alignment space. The job requirement encoder maps the job requirement features to job requirement semantic vectors, which are located in the cross-modal alignment space; the candidate ability encoder maps the candidate ability evolution features to candidate ability semantic vectors, which are located in the cross-modal alignment space.

4. The method according to claim 3, characterized in that, By introducing a demand-capability dual constraint mechanism to establish a nonlinear mapping relationship, and calculating the semantic distance between each candidate and the job requirements in the cross-modal alignment space based on the nonlinear mapping relationship, the initial matching degree is obtained, including: A demand-capability dual constraint mechanism is constructed, which includes demand integrity constraints and capability sufficiency constraints. The demand integrity constraints are used to calculate the coverage of the candidate's capability semantic vector to each dimension of the job demand semantic vector, and the capability sufficiency constraints are used to calculate the matching degree of the job demand semantic vector to each dimension of the candidate's capability semantic vector. The dual constraint score is obtained by performing a bidirectional interactive calculation between the coverage degree and the matching degree, and the nonlinear mapping relationship is established. The job requirement semantic vector and the candidate ability semantic vector of each candidate are respectively input into the requirement integrity constraint and the ability sufficiency constraint to calculate the coverage degree and the matching degree of each candidate. The coverage and matching degree of each candidate are input into the nonlinear mapping relationship to calculate the dual constraint score of each candidate. Calculate the original vector distance between the job requirement semantic vector and the candidate ability semantic vector of each candidate, and fuse the original vector distance with the dual constraint score to obtain the constraint-adjusted semantic distance as the initial matching degree.

5. The method according to claim 1, characterized in that, Based on the initial matching degree, candidates to be verified are identified. By inferring the future ability growth curve of the candidates to be verified in the target job field, a potential assessment value is generated. Based on the potential assessment value, the candidates to be verified are prioritized, and the target candidate set is determined to include: Set a matching threshold, and identify candidates whose initial matching degree meets the matching threshold as candidates to be verified. Extract the capability status records and capability change trend characteristics of the candidate to be verified at multiple time points as historical capability evolution trajectory data; The ability change patterns in the historical ability evolution trajectory data are correlated and mapped with the domain ability requirement characteristics of the target job domain to establish the correlation between ability evolution and requirement evolution. Based on the correlation, an ability growth trend coefficient and an ability saturation boundary parameter are generated. The ability saturation boundary parameter is used as an ability boundary constraint. Based on the ability growth trend coefficient, the future ability growth curve of the candidate to be verified in the target job domain is generated. The potential assessment value is calculated based on the future capability growth curve and the expected capability saturation value. The candidates to be verified are sorted in descending order according to their potential evaluation values, and the candidates to be verified within a preset ranking range are selected based on the result of the descending order to form the target candidate set.

6. The method according to claim 1, characterized in that, Extracting salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in their career development path, and predicting the probability distribution of each target candidate's response to different types of search information, including: Extract career development path data from the candidate resume texts of each target candidate; Historical job-hopping event records are extracted from the career development path data. Based on the correlation between the salary change range in the historical job-hopping event records and the job-hopping decision time, the salary sensitivity feature is calculated. Historical work location change records are extracted from the career development path data. Based on the correlation between location migration patterns and job selection decisions in the historical work location change records, the work location preference characteristics are calculated. The salary sensitivity feature and the work location preference feature are used to construct a candidate decision preference vector; Construct various types of search information representations, including salary incentive information representations, career development information representations, and location convenience information representations; By learning the matching relationship between the candidate decision preference vector and each of the search information representations, the response probability value of each target candidate to each of the search information representations is generated. Based on the response probability value corresponding to each search information representation, the response probability distribution of each target candidate to different types of search information is constructed.

7. The method according to claim 6, characterized in that, Based on the response probability distribution, the system dynamically selects the contact time and channel for each candidate, generating an automatic outreach strategy that includes: Based on the response probability values ​​of each target candidate to various types of search information representations, the type corresponding to the search information representation with the highest response probability value is taken as the dominant response type of each target candidate. Obtain channel feature data for each reach channel, match and map the dominant response type of each target candidate with the channel feature data of each reach channel, and select reach channels for each target candidate based on the matching mapping; Extract historical job-seeking behavior time periods and career status change cycle records from the career development path data of each target candidate; Based on the frequency and duration of time periods in the historical job-seeking behavior time period records, the time window weight distribution of each target candidate is calculated, and the candidate reach time window set is determined according to the time window weight distribution; based on the career status change interval and career status change trigger time in the career status change cycle records, the career decision time interval is determined. The time overlap analysis is performed between the set of candidate outreach time windows and the career decision time interval, and the outreach time is determined for each target candidate based on the results of the time overlap analysis. The automatic search and outreach strategy is generated by combining the contact time and contact channel corresponding to each target candidate.

8. An AI-based intelligent resume matching and automatic search system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The semantic parsing unit is used to perform multi-level semantic parsing on the job requirement description text, extract job requirement features, model the career development trajectory of candidate resume text, and extract the candidate ability evolution features of each candidate. A cross-modal pair unit is used to construct a cross-modal alignment space of job requirements and candidate capabilities. The job requirement features and candidate capability evolution features are mapped to the cross-modal alignment space. A nonlinear mapping relationship is established by introducing a requirement-capability dual constraint mechanism. Based on the nonlinear mapping relationship, the semantic distance between each candidate and the job requirement in the cross-modal alignment space is calculated to obtain the initial matching degree. The potential assessment unit is used to identify candidates to be verified based on the initial matching degree, generate potential assessment values ​​by inferring the future ability growth curve of the candidates to be verified in the target job field, prioritize the candidates to be verified according to the potential assessment values, and determine the target candidate set. The preference prediction unit is used to extract the salary sensitivity characteristics and work location preference characteristics of each target candidate in the target candidate set in the career development path, and predict the response probability distribution of each target candidate to different types of search information. The strategy generation unit is used to dynamically select the contact time and contact channel for each candidate based on the response probability distribution, and generate an automatic search and contact strategy.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images