An ai-driven talent intelligent matching and recommendation system oriented to post demand

Abstract

The application discloses a post-demand-oriented AI-driven talent intelligent matching and recommendation system, comprising: a contract template configuration module that collects real-time post demand texts and dynamically constructs a post skill tree to obtain a target post skill tree graph; a compliance data chaining module that diffuses the key node reachable domain of the target post skill tree graph and matches multiple primary screening talent vectors in a preset talent vector database; a salary calculation module that performs multidimensional skill attenuation mapping on the multiple primary screening talent vectors based on the target post skill tree graph to obtain multiple attenuation talent vectors, calculates the depth weighted distance between each attenuation talent vector and the root node of the target post skill tree graph, and obtains multiple depth weighted distance values; and a risk control audit module that takes the multiple depth weighted distance values as path constraints, performs cooperative path backtracking on the multiple primary screening talents, generates multiple cooperative recommendation paths, and outputs a recommended target talent list.

Description

Technical Field

[0001] This invention relates to the field of intelligent talent matching technology, specifically to an AI-driven intelligent talent matching and recommendation system oriented towards job requirements. Background Technology

[0002] In the field of talent recruitment and job matching, existing automated talent recommendation methods mostly rely on resume keyword matching or static skill tag libraries. When extracting required skills from job postings published by companies in real time, only shallow semantic analysis is usually performed, treating skills as flat, independent tags, failing to reflect the hierarchical structure, inclusion relationships, and semantic diffusion connections between skills. This makes it difficult to identify candidates who, while not directly possessing the explicit skills required for the position, have closely related or extended skills, resulting in the omission of potentially suitable candidates.

[0003] Existing methods for quantifying the gap between talent and job requirements generally employ geometric metrics such as cosine similarity or Euclidean distance, treating all dimensions of the talent vector equally and ignoring the differences in the level and importance of different skills within job requirements. Furthermore, the proficiency of skills acquired by individuals diminishes over time, and conventional matching mechanisms lack the ability to dynamically map this skill decay, leading to assessment results that deviate from the actual usable skill level of the talent.

[0004] Talent recommendations are typically output on a per-candidate basis, without considering collaborative relationships between candidates. Even if the system can provide multiple high-scoring initial candidates, it cannot determine whether there are reasonable collaborative paths or transferable relationships among these candidates. The recommendation results lack support for practical recruitment scenarios such as team building and talent relevance. Job seekers are often not only concerned with the skill fit of individual candidates, but also expect to understand which relevant talents can be sourced for a particular position, forming a reliable talent supply chain through collaboration or transferable relationships between positions.

[0005] Therefore, it is necessary to address how to dynamically parse the job skill structure with hierarchical expansion capabilities under the real-time job requirement text-driven approach, incorporate the skill association diffusion range into the initial screening and matching, and integrate the skill level depth and decay characteristics when assessing talent skill gaps, and use this gap to explore the collaborative paths between talents, thereby providing more relevant and hierarchical talent recommendations. Summary of the Invention

[0006] This paper presents an AI-driven intelligent talent matching and recommendation system that is oriented towards job requirements. It can dynamically construct a hierarchical skill tree from real-time job requirement text and perform reachability domain diffusion matching. At the same time, it introduces multi-dimensional decay mapping and depth-weighted distance in skill assessment, and uses this distance as a constraint to perform collaborative path backtracking, so as to achieve high-precision talent recommendation with collaborative relationship awareness.

[0007] The objective of this invention can be achieved through the following technical solutions: This invention discloses an AI-driven intelligent talent matching and recommendation system oriented towards job requirements. The system achieves high-precision talent recommendation by constructing a dynamic skill tree, diffusion matching, decay mapping, and collaborative path backtracking.

[0008] The system includes a contract template configuration module, a performance data on-chain module, a payroll calculation module, and a risk control and audit module. The contract template configuration module collects real-time job requirement text from target companies and dynamically constructs a job skill tree from this text, obtaining a target job skill tree graph. Through dependency parsing, it extracts core phrases containing job action words and job skill object words, using job action words as root nodes and job skill object words as first-level child nodes. Each job skill object word undergoes word embedding expansion to obtain second-level skill extension words. The initial connection weights between the root node and first-level child nodes are determined based on co-occurrence frequency, and the extended connection weights between first-level and second-level child nodes are determined based on semantic distance. The nodes and their connection weights are combined to form the target job skill tree graph. This graph accurately reflects the skills required for the job and their logical hierarchy, providing a structured foundation for subsequent matching.

[0009] The on-chain module for performance data performs key node reachability domain diffusion on the target job skill tree graph, matching multiple initial talent vectors from a pre-set talent vector database. As a preferred technical solution, all first-level child nodes are extracted from the skill tree graph as key skill nodes. Using each key skill node as the starting point for diffusion, a random walk diffusion is performed according to the connection weights to obtain the set of reachable skill nodes for each key skill node. A union operation is then performed to obtain the reachability domain of the key node. Each skill node in the reachability domain is converted into a query vector, and a dot product similarity calculation is performed with the talent vectors in the pre-set talent vector database. Talent vectors that meet the similarity threshold are used as initial talent vectors. This diffusion matching method can expand the range of key skills required for the job, improving the recall rate and coverage accuracy of the initial talent screening. Preferably, the maximum step size of the random walk diffusion is determined by the maximum depth of the skill tree graph, and the stopping probability of each walk is inversely proportional to the connection weight of the current node, ensuring that the diffusion process is guided by the importance of skills and avoiding over-diffusion that introduces irrelevant skills. When calculating dot product similarity, multiple query vectors can be constructed into a query matrix, and talent vectors can be constructed into a talent matrix. The similarity matrix can be obtained at once through matrix multiplication. After extracting candidate talent vectors that exceed the threshold in each row, the union is performed to remove duplicates, which effectively improves the computational efficiency under a large-scale talent pool.

[0010] The salary calculation module, based on the target job skill tree graph, performs multi-dimensional skill decay mapping on multiple initially screened talent vectors to obtain multiple decayed talent vectors. It then calculates the depth-weighted distance between each decayed talent vector and the root node, obtaining multiple depth-weighted distance values. Preferably, multiple decay functions corresponding to the number of depth levels in the skill tree graph are extracted from a preset skill decay function library, with each decay function corresponding to one depth level. The initially screened talent vectors are split into multiple skill sub-vectors according to the skill dimension. Based on the depth level of each skill sub-vector in the skill tree graph, the corresponding decay function is selected for decay calculation. The decayed skill sub-vectors are then concatenated to form the decayed talent vector. The decay function can be an exponential decay function or a linear decay function, with the depth level as the independent variable. This decay mapping simulates the decay effect of talent skills over time or technological iteration, making talent representation more closely reflect actual job requirements. In calculating the depth-weighted distance, a decay coefficient is extracted from each decaying skill sub-vector, and the level contribution value is calculated in conjunction with the corresponding depth level. Using the unit vector corresponding to the root node as a reference vector, the cosine of the angle between each decaying skill sub-vector and the reference vector is calculated. The level contribution values ​​are then summed by the weighted cosine values ​​and divided by the sum of the decay coefficients to obtain the depth-weighted distance value. This distance value comprehensively reflects the degree of deviation between the talent's skills and the core requirements of the job at the skill level, as well as the closeness of the skill correlation, providing a quantitative basis for subsequent accurate screening and ranking.

[0011] The risk control and auditing module uses multiple depth-weighted distance values ​​as path constraints to perform collaborative path backtracking on multiple initially screened talents, generating multiple collaborative recommendation paths. The talent corresponding to the endpoint of each collaborative recommendation path is selected as the target talent for recommendation, and a list of target talents is output. Specifically, the first and second initially screened talents are randomly selected from the initially screened talents, and the difference between their depth-weighted distance values ​​is used as the first path constraint threshold. Starting from the first initially screened talent, a breadth-first search is performed in a preset talent relationship graph to search for all talent nodes whose inter-node migration cost is less than the first path constraint threshold, obtaining a first candidate path set. Then, starting from the second initially screened talent, a reverse path backtracking is performed in the first candidate path set to obtain collaborative recommendation paths that can reach the second initially screened talent. The preset talent relationship graph is constructed with talents as nodes and historical collaboration relationships or skill similarity between talents as edge weights. Through collaborative path backtracking, while considering the matching degree between individuals and positions, the collaborative relationships and complementary skills between talents can be explored, improving the synergy of recommendation results and the adaptability of team building. When outputting the list of recommended target talents, the total migration cost is calculated for each collaborative recommendation path, and the endpoint of the path with the minimum total migration cost is selected as the first recommended target talent. After removing the talents already used from the initial screening, the above process is repeated to obtain the second recommended target talent and subsequent recommended talents. The recommended target talent list is generated in the order obtained to ensure the diversity and reasonableness of the recommendation results.

[0012] As a technical solution of the present invention, the system further includes a historical data acquisition module, a historical distance calculation module, a regression model training module, and a depth distance output module. The historical data acquisition module collects historical job requirement texts and historical successful recruitment data of the target enterprise to construct a historical job skills tree graph. The historical distance calculation module uses the historical depth-weighted distance value between the historical successful recruitment vector and the root node of the historical job skills tree graph as a label. The regression model training module uses historical decaying talent vectors as input and historical depth-weighted distance values ​​as labels to iteratively train the depth-weighted distance regression model to obtain a trained model. The depth distance output module inputs multiple currently obtained decaying talent vectors into the model and directly outputs the depth-weighted distance value. By introducing a regression model, the distance calculation can be adaptively calibrated using the enterprise's historical successful recruitment data, making the depth-weighted distance more accurately reflect the enterprise's actual hiring preferences and talent quality assessment standards, thereby improving the reliability and implementability of the recommendations.

[0013] The beneficial effects of this invention are: When extracting core phrases of job requirements from real-time job requirement text, dependency parsing is used to obtain verb-object pairs containing job action words and job skill object words. The job action words are used as the root node of the skill tree graph, and the job skill object words are used as first-level child nodes. Second-level skill extension words are generated through word embedding. The connection weights between first-level child nodes and the root node, and between second-level child nodes and first-level child nodes, are determined based on co-occurrence frequency and semantic distance, forming the target job skill tree graph. This dynamic construction method transforms the flat skill list into a hierarchical and weighted tree structure, explicitly expressing the primary and secondary, and derivative relationships between skills. Based on this, the target job skill tree graph undergoes key node reachability domain diffusion. Using first-level child nodes as key skill nodes, random walk diffusion is performed along the connection weights. The maximum step size of the diffusion is determined by the maximum depth of the graph, and the stopping probability is inversely proportional to the connection weight. This yields the set of reachable skill nodes for each key skill node, and the union of these sets forms the key node reachability domain. The skill nodes in the reachability domain are converted into query vectors and their dot product similarity is calculated with a pre-set talent vector database to filter out multiple initial talent vectors. By expanding the search scope, the search range is extended from a few isolated key skill terms to a set of skills that are semantically and structurally related to them, effectively capturing talents with derivative and adjacent skills and reducing omissions caused by the limitations of job description terms.

[0014] When performing multi-dimensional skill attenuation mapping on the initial talent vectors, a corresponding attenuation function is selected from a pre-set skill attenuation function library based on the number of depth levels in the target job skill tree graph. After each initial talent vector is split by skill dimension, the corresponding attenuation function is called to perform attenuation calculations based on the depth level of each skill sub-vector in the skill tree, obtaining attenuated skill sub-vectors and concatenating them into an attenuated talent vector. The attenuation functions in the pre-set skill attenuation function library are exponential or linear attenuation functions with depth level as the independent variable, applying a greater attenuation to skills at deeper levels to simulate the weakening of skill proficiency over time. When calculating the depth-weighted distance between each attenuated talent vector and the root node, attenuation coefficients are extracted from the attenuated skill sub-vectors and combined with the depth level to calculate the level contribution value. Then, the cosine value of the angle between each attenuated skill sub-vector and the root node unit vector is calculated. The product of the level contribution value and the cosine value is summed and divided by the sum of the attenuation coefficients to obtain the depth-weighted distance value. This distance metric embeds hierarchical importance and skill decay status, which, compared to conventional vector distance, better reflects the structural gap between the actual set of skills available to talents and the core requirements of the job, thus improving the accuracy of ranking and screening.

[0015] When performing collaborative path backtracking using multiple depth-weighted distance values ​​as path constraints, the difference in depth-weighted distance values ​​between two initially screened talents is randomly selected as the first path constraint threshold. Starting from one of the initially screened talents, a breadth-first search is performed in a pre-defined talent relationship graph to find all talent nodes whose inter-node migration cost is less than the threshold, obtaining a first candidate path set. Then, a reverse path backtracking is performed with the other initially screened talent as the endpoint, and the path that can reach the endpoint is used as the collaborative recommendation path. The pre-defined talent relationship graph is constructed with talents as nodes and historical collaboration relationships or skill similarity as edge weights, with migration cost defined by the edge weights. The resulting collaborative recommendation path is constrained by skill gaps and relies on a real talent relationship network. The recommended talents are not only close to the job's skill tree but also have connectable collaborative relationships with the identified reference talents. Repeatedly selecting different pairs of initially screened talents to perform this process outputs a list containing multiple recommended target talents, extending the recommendation results from single-point matching to the interconnected links between talents. This provides recruiters with talent combinations with internal collaboration potential, enhancing the reference value of the recommendations in actual recruitment decisions. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of the structure of an AI-driven intelligent talent matching and recommendation system; Figure 2 It is a flowchart for constructing the skill tree map for the target position; Figure 3 It is a flowchart of the initial talent screening process based on the diffusion of reachable domains of skill tree graphs; Figure 4 This is a flowchart of the training and inference process for a deep-weighted distance regression model. Detailed Implementation

[0018] 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.

[0019] See Figure 1This invention provides an AI-driven intelligent talent matching and recommendation system oriented towards job requirements, including a contract template configuration module, a performance data upload module, a salary calculation module, and a risk control and audit module. The contract template configuration module collects real-time job requirement text from target companies and dynamically constructs a job skill tree from this text to obtain a target job skill tree graph. The performance data upload module expands the reachability domains of key nodes in the target job skill tree graph and matches multiple initial screening talent vectors from a pre-set talent vector database. These initial screening talent vectors correspond to multiple initial screening talents. The salary calculation module, based on the target job skill tree graph, performs multi-dimensional skill attenuation mapping on the multiple initial screening talent vectors to obtain multiple attenuated talent vectors. It then calculates the depth-weighted distance between each attenuated talent vector and the root node of the target job skill tree graph to obtain multiple depth-weighted distance values. The risk control and audit module uses these depth-weighted distance values ​​as path constraints to perform collaborative path backtracking on the multiple initial screening talents, generating multiple collaborative recommendation paths. The talent corresponding to the endpoint of each collaborative recommendation path is selected as the recommended target talent, and a list of recommended target talents is output.

[0020] The contract template configuration module dynamically constructs a job skills tree from real-time job requirement texts, obtaining a target job skills tree map. For specific implementation details, please refer to [link / reference needed]. Figure 2 The real-time job requirement text is segmented using a hybrid segmentation method combining dictionary and statistical approaches. This method divides the text into multiple word units, and a sequence labeling model is used to tag each word unit with its corresponding part-of-speech tag. The part-of-speech tag set includes verb tags, noun tags, and adjective tags. A dependency parsing tree is constructed based on these word units and their part-of-speech tags. A neural network-based dependency parser takes a sequence of word units and a sequence of part-of-speech tags as input and outputs the dependency relations between word units. These relations include verb-object relations, attributive-head relations, and adverbial-head relations, and all dependency relations form the dependency parsing tree. All dependency arcs with verb-object relations are extracted from the dependency parsing tree. Each verb-object relation arc contains a core word and a dependent word. The core word is tagged with a verb, and the dependent word is tagged with a noun. The core word in each verb-object relation arc is used as the job action word, and the dependent word in each verb-object relation arc is used as the job skill object word. A job action word and a job skill object word are combined to form a core phrase of job requirements. Each core phrase of job requirements contains a job action word and a job skill object word, representing the target company's action requirements for talent skills.

[0021] The graph uses all job action words as its root node set, with each distinct job action word corresponding to a root node in the target job skill tree graph. Each job skill object word is designated as a first-level child node, attached below the root node containing the corresponding job action word with its verb-object relationship, becoming a first-level child node of that root node. Word embedding expansion is performed on each job skill object word. A pre-trained word vector model maps the job skill object word to a dense vector. Based on this dense vector, multiple semantically closest extended words are retrieved in the semantic vector space. The retrieval method involves calculating the cosine similarity between the dense vector corresponding to the job skill object word and all word vectors in the word vector model. The top k words with a cosine similarity greater than a preset expansion threshold and consistent part of speech with the job skill object word are selected as second-level skill extension words, where k is the preset number of extensions. Each second-level skill extension word constitutes a second-level child node, attached below the first-level child node corresponding to the job skill object word.

[0022] When determining node connection weights, the initial connection weight between each first-level child node and the root node is determined based on the co-occurrence frequency of job action terms and job skill object terms in multiple job requirement core phrases. The number of times each root node's job action term and the corresponding job skill object term of its first-level child nodes co-occur in all job requirement core phrases is obtained; this number is the co-occurrence frequency of the job action term and the job skill object term.

[0023] Based on the semantic distance between each job skill object term and its corresponding secondary skill extension term, the extended connection weights between each secondary child node and its corresponding primary child node are determined. The dense vectors corresponding to the job skill object term and each secondary skill extension term are obtained, and the cosine distance between these two dense vectors is calculated as the semantic distance. The extended connection weights are normalized values ​​of the semantic distance. Specifically, the semantic distance values ​​corresponding to all secondary skill extension terms under a primary child node are normalized using the softmax function, ensuring that the sum of all extended connection weights under the same primary child node is 1. The root node, primary child nodes, secondary child nodes, initial connection weights, and extended connection weights are combined to form a hierarchical target job skill tree graph with weighted edges.

[0024] The data upload module for contract fulfillment performs key node reachability domain diffusion on the target job skill tree graph, matching multiple initial talent vectors from a pre-set talent vector database. For specific implementation details, please refer to [link / reference needed]. Figure 3All first-level child nodes are extracted from the target job skill tree graph, and each first-level child node is treated as a key skill node, with each key skill node corresponding to a job skill object word. Using each key skill node as the starting point for diffusion, a random walk diffusion is performed according to the connection weights in the target job skill tree graph. The connection weights include the initial connection weight between the first-level child node and the root node, and the extended connection weight between the second-level child node and its corresponding first-level child node.

[0025] During random walk diffusion, for each critical skill node, multiple independent random walks are performed. Each random walk starts from the critical skill node and jumps to other nodes along the edges of the target job skill tree graph. The maximum step size of the random walk is determined by the maximum depth of the target job skill tree graph, which is the number of edges from any root node to the farthest leaf node. When the number of steps reaches the maximum depth, the current walk diffusion is forcibly terminated. For each hop in each walk, the transition probability is determined by the connection weights between the current node and its neighbors. The transition probability is the ratio of the neighbor node's connection weight to the sum of all outgoing edge connection weights of the current node. Before each transition, the system generates a node within the interval... A uniformly distributed random number is generated between the nodes. If the random number is less than the stopping probability of the current step, the traversal stops at the current node; if the random number is greater than or equal to the stopping probability of the current step, the traversal continues to the next neighbor node. The stopping probability of each traversal is inversely proportional to the connection weight of the current node, which is the connection weight of the edge between the current node and the source node of the previous hop of the traversal. In specific implementation, the stopping probability is calculated using the following formula: in, Indicates the current node The probability of the walk stopping at a certain point; As a smoothing factor, The value is set to 0.5. The rationale for this value is to ensure that when the connection weight is greater than 0.5, the stopping probability is less than 0.5, and when the connection weight is less than 0.5, the stopping probability is greater than 0.5. This makes it more likely that the traversal will continue to spread on edges with larger weights and tend to stop on edges with smaller weights. Indicates starting from the previous node To the current node The connection weights, if the current node is the diffusion starting point, then Determine the initial connection weight between the critical skill node and its corresponding root node. Add all nodes visited during each random walk to the node set of that walk. After multiple independent random walks, take the union of all walked node sets as the set of reachable skill nodes for the corresponding critical skill node. Generate a set of reachable skill nodes for each critical skill node.

[0026] The reachable skill node set corresponding to each key skill node is unioned. This involves taking the union of all skill nodes in the reachable skill node sets corresponding to all key skill nodes to obtain the reachable domain of each key skill node. The reachable domain of a key node is the set of nodes encompassing all key skill nodes and their related reachable skill nodes. Each skill node in the reachable domain of a key node is converted into a query vector. This conversion uses a pre-trained word vector model to map the skill word corresponding to each skill node into a dense vector of fixed dimension. This dense vector serves as the query vector, and multiple skill nodes are converted into multiple query vectors. The dimension of the query vector is consistent with the dimension of the talent vectors in the preset talent vector database, and the dimension value is set to 300.

[0027] Multiple query vectors are used to construct a query matrix, where the number of rows in the query matrix corresponds to the number of query vectors, and the number of columns corresponds to the vector dimension. All talent vectors from a pre-defined talent vector database are used to construct a talent matrix, where the number of rows in the talent matrix corresponds to the number of talent vectors, and the number of columns corresponds to the vector dimension. Matrix multiplication is then performed on the query matrix and the talent matrix, multiplying the query matrix by the transpose of the talent matrix to obtain a similarity matrix. Each element in the similarity matrix represents the dot product similarity between a query vector and a human resources talent vector. The row index of the similarity matrix corresponds to the index of the query vector, and the column index corresponds to the index of the talent vector.

[0028] Iterate through each row of the similarity matrix, extracting the column index corresponding to the element with a dot product similarity greater than a preset initial screening threshold. The preset initial screening threshold is set to 0.7, based on a commonly used boundary point for determining high similarity in cosine similarity empirical thresholds, ensuring that only talent vectors highly matching the query vector's skills are included in the candidate range. The talent vector corresponding to each column index is used as a candidate talent vector. The candidate talent vectors from different rows are then deduplicated through a union. When the same human resource vector is extracted from multiple rows, only one instance is retained. After deduplication, multiple initial screening talent vectors are obtained, each corresponding to a different initial screening talent.

[0029] The salary calculation module, based on the target job skill tree graph, performs multi-dimensional skill decay mapping on multiple initially screened talent vectors to obtain multiple decayed talent vectors. It then calculates the depth-weighted distance between each decayed talent vector and the root node of the target job skill tree graph. In practice, the number of depth levels in the target job skill tree graph is determined. The number of depth levels is defined as the number of edges contained in the longest path from any root node to the farthest leaf node. Each skill node in the target job skill tree graph corresponds to a specific depth level value, which increments from 0 at the root node.

[0030] The preset skill decay function library stores multiple decay functions corresponding to depth levels. Each decay function is either an exponential decay function or a linear decay function with the depth level as the independent variable. The expression for the exponential decay function is as follows: ,in, The depth level is a non-negative integer. The attenuation coefficient factor, The value is set to 0.3, based on the fact that the decay coefficient is approximately 0.4 when the depth level is 3, thus moderately reducing the influence of medium-depth skills. The expression for the linear decay function is... ,in, Slope factor The value is 0.15, and when When the calculated result is less than 0.05, it is forcibly truncated to 0.05 to ensure that the decay rate of shallow skills is stable and that deep skills retain a weak correlation. Based on the number of depth levels in the target job skill tree graph, a decay function is selected for each depth level from a preset skill decay function library. The selection rule is as follows: a linear decay function is assigned to levels with a depth less than half of the maximum depth, and an exponential decay function is assigned to levels with a depth greater than or equal to half of the maximum depth. This ensures that the decay rate transitions from a gradual decrease in the first half to an accelerated decrease in the second half of the skill tree hierarchy.

[0031] For each initially screened talent vector, it is split into multiple skill sub-vectors according to skill dimensions. The initially screened talent vector is a dense vector with 300 dimensions, where each dimension corresponds to a specific skill feature. Based on a pre-stored mapping table between skill features and skill nodes in the target job skill tree graph, the dimensions in the initially screened talent vector are grouped. All dimensions belonging to the same skill node in the target job skill tree graph constitute a skill sub-vector, and each skill sub-vector corresponds to one skill node. After splitting, multiple skill sub-vectors are obtained, the number of which is consistent with the number of skill nodes covered by the initially screened talent vector.

[0032] Based on the depth level of the skill node corresponding to each skill sub-vector in the target job skill tree graph, a decay function corresponding to that depth level is selected to perform decay operations on the skill sub-vector. The specific process of decay operation is as follows: the decay coefficient is calculated using the decay function, which is the output value after the depth level is input to the decay function. Each element in the skill sub-vector is multiplied by the decay coefficient to obtain the decayed skill sub-vector. If the skill node corresponding to the skill sub-vector does not exist in the target job skill tree graph, its depth level is set to the maximum depth level plus 1, and the decay function corresponding to the maximum depth level is used. Multiple skill sub-vectors are decayed to obtain multiple decayed skill sub-vectors. The multiple decayed skill sub-vectors are concatenated according to their original dimensional order to form a decayed talent vector with the same dimensions as the initial talent vector.

[0033] For each decaying talent vector, a decay coefficient is extracted from each decaying skill sub-vector. This coefficient is calculated and stored during the decay operation and can be directly retrieved. The depth level corresponding to each decaying skill sub-vector is also recorded. The level contribution value for each decaying skill sub-vector is calculated by multiplying the decay coefficient by the inverse of the depth level, where the inverse of the depth level is taken as... ,in, Indicates the first For each decaying skill sub-vector, the depth level is incremented by 1 to prevent the denominator from being zero. This calculation method ensures that the closer a skill node is to the root node, the higher its level contribution value, and that the skill node with the larger the decay coefficient has the higher its influence weight. At the same time, it ensures that the root node with a depth level of 0 has the maximum level contribution.

[0034] The word vectors corresponding to all root nodes in the target job skill tree graph are averaged and then L2 normalized to obtain a reference vector of length 1. This reference vector represents the core comprehensive skill direction of the target job. The cosine of the angle between each decayed skill sub-vector in the same decayed talent vector and the reference vector is calculated. The cosine of the angle is obtained by dividing the inner product of the decayed skill sub-vector and the reference vector by the magnitude of both. Since the reference vector has been normalized and the decayed skill sub-vector is obtained by multiplying the elements by the decay coefficient, this cosine of the angle reflects the consistency of the vector space direction between the skill sub-vector and the core skill after the decay effect.

[0035] The depth-weighted distance value corresponding to each decaying skill sub-vector is obtained by multiplying its hierarchical contribution value by the corresponding cosine of the included angle, summing the results, and then dividing by the sum of the decay coefficients of all decaying skill sub-vectors in that decaying talent vector. The formula for calculating the depth-weighted distance value is as follows: in, This represents the depth-weighted distance value corresponding to a decaying talent vector; This represents the total number of decaying skill sub-vectors contained in the decaying talent vector; It refers to the first The decay coefficient of the th decay skill subvector, which is determined by the th... The decay function corresponding to the depth level of each decay skill sub-vector is calculated, and the value ranges from 0.05 to 1.0. It refers to the first The depth level of each decaying skill sub-vector in the target job skill tree graph, with values ​​being non-negative integers; It refers to the first The cosine value of the angle between the attenuation skill sub-vector and the reference vector ranges from -1 to 1. This is the inverse depth factor, used to adjust the contribution of a layer based on its depth level; the smaller the depth level, the larger this factor becomes. Through the above calculation process, each decaying talent vector yields a depth-weighted distance value, and all depth-weighted distance values ​​constitute multiple sets of depth-weighted distance values.

[0036] The risk control and audit module uses multiple depth-weighted distance values ​​as path constraints to collaboratively backtrack the paths of multiple initially screened talents, generating multiple collaborative recommendation paths. The talent corresponding to the endpoint of each collaborative recommendation path is then used as the target talent for recommendation, and a list of target talents is output. In specific implementation, the risk control and audit module receives multiple depth-weighted distance values ​​provided by the payroll calculation module, with each depth-weighted distance value corresponding one-to-one with an initially screened talent. A mapping relationship is established between multiple initially screened talents and multiple depth-weighted distance values ​​through the identity identifier of the initially screened talents.

[0037] From a pool of initial-screened talents, two distinct talent nodes are selected using a uniformly distributed random sampling method, designated as the first initial-screened talent and the second initial-screened talent, respectively. The depth-weighted distance value corresponding to the first initial-screened talent is recorded as the first depth-weighted distance value; the depth-weighted distance value corresponding to the second initial-screened talent is recorded as the second depth-weighted distance value. The absolute value of the difference between the first and second depth-weighted distance values ​​is used as the first path constraint threshold. The formula for calculating the first path constraint threshold is: in, Indicates the threshold value for the first path constraint; This represents the first depth-weighted distance value corresponding to the first batch of talents in the initial screening; This represents the second depth-weighted distance value corresponding to the second initial screening candidate. The depth-weighted distance value is a non-negative real number, and its magnitude reflects the comprehensive distance between the corresponding initial screening candidate and the root node of the target position in the vector space after skill decay.

[0038] Using the first-screened talent as the starting point, a breadth-first search is performed on the pre-defined talent relationship graph. The pre-defined talent relationship graph is a graph data structure pre-built and persistently stored during system initialization. Each talent is a node, and each edge represents a historical collaboration relationship or a fulfillment of skill similarity criteria between two talents. The edge weight is obtained by weighted summation of the historical collaboration strength and skill similarity between talents, with weight values ​​ranging from 0 to 1. The migration cost between nodes is calculated using a transformation function based on the edge weight, defined as taking the reciprocal of the edge weight and multiplying it by a fixed coefficient of 10. When the edge weight is at its minimum value of 0.1, the migration cost is 100; when the edge weight is at its maximum value of 1.0, the migration cost is 10. The fixed coefficient of 10 is set to map the reciprocal of the edge weight to a range with significant numerical differences, thereby effectively distinguishing between close and sparse relationships during the path search process.

[0039] In the breadth-first search process, starting from the path's originating node, neighboring nodes are visited layer by layer. A neighboring node is added to the search queue only if the migration cost from the current node to a neighboring node is less than the first path constraint threshold, and the directed edge from the current node to the neighboring node is recorded in the search tree. If the migration cost from the current node to a neighboring node is greater than or equal to the first path constraint threshold, the neighboring node is ignored and not included in the search scope. The breadth-first search continues until the search queue is empty or all reachable nodes satisfying the migration cost constraint have been visited. All nodes and their corresponding directed edges obtained through the breadth-first search constitute the first candidate path set. The first candidate path set records all feasible segments that can be reached from the first initially screened talent via edges with migration costs less than the first path constraint threshold.

[0040] Using the second initially screened talent as the path endpoint, a reverse path backtracking is performed in the first candidate path set to extract the complete path from the path start point to the path endpoint. The reverse path backtracking starts from the node corresponding to the second initially screened talent, searches for an edge with the current node as the endpoint in the directed edges stored in the first candidate path set, and obtains the starting node of that edge as the predecessor node for backtracking. This process is repeated upwards until the predecessor node of a certain backtracking step is the first initially screened talent node. If the backtracking successfully reaches the first initially screened talent node, all nodes traversed during the backtracking process are arranged in order from the first initially screened talent to the second initially screened talent, forming a collaborative recommendation path. If the reverse path backtracking cannot reach the first initially screened talent node, it indicates that there is no path connecting the path start point and the path endpoint in the first candidate path set, and no collaborative recommendation path is generated.

[0041] For each successfully generated collaborative recommendation path, calculate the total migration cost of that path. The total migration cost is defined as the sum of the migration costs between all adjacent talent nodes on the collaborative recommendation path, and the calculation formula is as follows: in, Represents a collaborative recommendation path The total migration cost; Represents collaborative recommendation path The total number of talent nodes included; and These represent collaborative recommendation paths. Upper The and the first Talent nodes; Indicates talent nodes With talent nodes The migration cost between the two nodes is determined by dividing the weight of the edge connecting the two nodes in the preset talent relationship graph by the edge weight and multiplying by a fixed coefficient of 10. After calculating the total migration cost of all collaborative recommendation paths, the total migration costs are sorted, and the collaborative recommendation path with the smallest total migration cost is selected. The talent corresponding to the endpoint node of this collaborative recommendation path is taken as the first recommended target talent. If multiple paths have the same smallest total migration cost, one is randomly selected.

[0042] After identifying the first batch of target talents, the first and second batches of initial-screened talents are removed from the initial-screen talent set, resulting in an updated initial-screen talent set. It is then determined whether at least two talent nodes remain in the updated initial-screen talent set. If fewer than two remain, the process is terminated. If two or more remain, two talent nodes are randomly selected from the updated initial-screen talent set to serve as the new first and second batches of initial-screened talents. The same steps are then used to calculate new path constraint thresholds, perform breadth-first search, and backtrack the path to obtain the second batch of target talents. This process of removal and random selection is repeated until fewer than two talents remain in the initial-screen talent set or no collaborative recommendation path can be generated from the current set. The first, second, and all subsequent target talents obtained through repeated executions are arranged strictly in chronological order of their generation to form a target talent list. This target talent list serves as the final output of the risk control and auditing module and is provided to the downstream recruitment decision-making process.

[0043] See Figure 4The historical data acquisition module collects historical job requirement texts and historical successful recruitment data from target companies. Historical job requirement texts are job description documents previously published by the target company, while historical successful recruitment data consists of talent files of employees ultimately hired and employed based on these historical job requirement texts. Each historical successful recruitment data set includes the talent's identity identifier, a set of skill tags, and a skill mastery score sequence. The historical data acquisition module dynamically constructs a job skill tree for each historical job requirement text. The construction process includes: word segmentation and part-of-speech tagging of the historical job requirement text; constructing a dependency syntax tree; extracting verb-object pairs as core phrases of the job requirement; constructing root nodes using job action words; constructing first-level child nodes using job skill object words; expanding the first-level child nodes through word embedding to obtain second-level skill extension words, which are then used as second-level child nodes; setting initial connection weights based on co-occurrence frequency; setting extended connection weights based on semantic distance; and finally combining them into a historical job skill tree graph. A historical job skill tree graph is generated for each historical job requirement text.

[0044] The historical distance calculation module converts historical successful recruitment data into historical successful recruitment vectors. The conversion process involves extracting the skill tag set from the historical successful recruitment data. For each skill tag, a pre-trained word vector model is used to map it into a fixed-dimensional 300-dimensional skill tag vector. The skill mastery score is used as a weighting coefficient for this skill tag vector. All skill tag vectors are weighted and summed, and the sum is L2 normalized to obtain the historical successful recruitment vector. The historical distance calculation module then obtains the historical job skill tree graph corresponding to the historical successful recruitment data. The average of the word vectors corresponding to all root nodes of the historical job skill tree graph is taken and L2 normalized to obtain the historical reference vector. Using the same multi-dimensional skill decay mapping and depth-weighted distance calculation method as in the salary calculation module, the historical successful recruitment vector is split into multiple historical skill sub-vectors according to the skill dimension. Based on the depth level of each historical skill sub-vector in the historical job skill tree graph, a corresponding decay function is selected for decay calculation, resulting in multiple historical decayed skill sub-vectors, which are then concatenated into the historical decayed recruitment vector. Extract the decay coefficient and depth level of each historical decay skill sub-vector from the historical decay talent vector, calculate the historical level contribution value, calculate the cosine of the angle between each historical decay skill sub-vector and the historical reference vector, sum the products of the historical level contribution value and the cosine of the angle, and divide by the sum of all decay coefficients to obtain the historical depth-weighted distance value. Each historical successfully recruited talent data point corresponds to a historical depth-weighted distance value.

[0045] The regression model training module constructs and trains a pre-defined depth-weighted distance regression model. This model employs a feedforward neural network, with a core architecture consisting of an input layer, three fully connected hidden layers, and an output layer. The input layer has a dimension of 300 and receives historical decay vectors. The first fully connected hidden layer contains 256 neurons using the ReLU activation function; the second contains 128 neurons using the ReLU activation function; and the third contains 64 neurons using the ReLU activation function. The output layer contains one neuron and outputs the predicted depth-weighted distance value directly without using an activation function. A dropout layer with a dropout ratio of 0.2 is added after each fully connected hidden layer to randomly discard some neuron outputs during training, suppressing overfitting.

[0046] The regression model training module uses each historical depth-weighted distance value as the training label and the historical decayed talent vector that generated that historical depth-weighted distance value as the training input sample to construct the training dataset. Each sample in the training dataset is a (historical decayed talent vector, historical depth-weighted distance value) pair. The training process uses mini-batch stochastic gradient descent, with each batch containing 32 samples. The loss function used is the mean squared error loss function, calculated as follows: in, This represents the mean squared error loss value; This indicates the number of training samples in each mini-batch. The value is 32; Indicates the first The true historical depth-weighted distance value of each training sample is calculated by the historical distance calculation module; This indicates that the preset depth-weighted distance regression model applies to the first... The predicted depth-weighted distance values ​​are output from each training sample. The optimizer used is the Adam optimizer, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. The training iteration cycle is set to 200 epochs, traversing all training samples once in each epoch. An early stopping mechanism is triggered when the training loss function value decreases by less than 0.0001 over 10 consecutive training epochs, terminating the iterative training. After training, the model parameters are saved, resulting in the trained depth-weighted distance regression model.

[0047] The depth distance output module is used during the inference phase. It receives multiple decayed talent vectors from the salary calculation module after performing multi-dimensional skill decay mapping on multiple initially screened talent vectors. These decayed talent vectors are then sequentially input into the trained depth-weighted distance regression model. The trained model performs forward propagation calculations on each input decayed talent vector: the data flows through the input layer, the first fully connected hidden layer with its ReLU activation and dropout, the second fully connected hidden layer with its ReLU activation and dropout, and the third fully connected hidden layer with its ReLU activation and dropout, finally outputting a scalar value directly through the output layer. This scalar value is the depth-weighted distance value corresponding to that decayed talent vector. This process is repeated for all decayed talent vectors, directly outputting multiple depth-weighted distance values, replacing the method of calculating depth-weighted distance values ​​using formulas. Through this method, the system can utilize historical recruitment success data to establish a regression mapping from decayed talent vectors to depth-weighted distance values, allowing the depth-weighted distance values ​​to reflect the overall distance distribution characteristics of successfully recruited talent in the target company's historical recruitment.

[0048] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI-driven intelligent talent matching and recommendation system oriented towards job requirements, characterized in that, The system includes: The contract template configuration module collects real-time job requirement texts from target companies, dynamically constructs a job skill tree from the real-time job requirement texts, and obtains a target job skill tree map. The contract performance data uplink module diffuses the reachability domain of key nodes in the target job skill tree graph and matches multiple preliminary talent vectors in the preset talent vector database. The multiple preliminary talent vectors correspond to multiple preliminary talents. The salary calculation module performs multi-dimensional skill decay mapping on the multiple initially screened talent vectors based on the target job skill tree graph to obtain multiple decayed talent vectors, and calculates the depth-weighted distance between each decayed talent vector and the root node of the target job skill tree graph to obtain multiple depth-weighted distance values. The risk control and audit module uses the multiple depth-weighted distance values ​​as path constraints to perform collaborative path backtracking on the multiple initially screened talents, generate multiple collaborative recommendation paths, and take the talents corresponding to the endpoint of each collaborative recommendation path as the recommended target talents, and output a list of recommended target talents. The process of collecting real-time job requirement texts from target enterprises and dynamically constructing a job skills tree from these texts to obtain a target job skills tree map includes: Dependency parsing was performed on the real-time job requirement text to extract multiple core job requirement phrases. Each core job requirement phrase contains job action words and job skill object words. Using the job action words as the root node of the graph, and each job skill object word as a first-level child node of the graph, word embedding expansion is performed on each job skill object word to obtain multiple second-level skill expansion words; Based on the co-occurrence frequency of job action words and job skill object words in the core phrases of the multiple job requirements, the initial connection weight between each first-level child node and the root node is determined. Based on the semantic distance between each job skill object word and the corresponding second-level skill extension word, the extended connection weight between each second-level child node and the corresponding first-level child node is determined. The root node, first-level child nodes, second-level child nodes, and the connection weights between each node are combined to form the target job skill tree graph.

2. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 1, characterized in that, The key node reachability domain of the target job skill tree graph is expanded, and multiple preliminary talent vectors are matched in a preset talent vector database, including: Extract all first-level child nodes from the target job skill tree graph as multiple key skill nodes, use each key skill node as the diffusion starting point, and perform random walk diffusion according to the connection weight in the target job skill tree graph to obtain the set of reachable skill nodes for each key skill node; Perform a union operation on the set of reachable skill nodes corresponding to each key skill node to obtain the reachable domain of the key node. Then, convert each skill node in the reachable domain of the key node into a query vector to obtain multiple query vectors. The multiple query vectors are respectively compared with each talent vector in the preset talent vector database by calculating the dot product similarity. The talent vectors corresponding to the query vectors with a dot product similarity greater than the preset initial screening threshold are used as the multiple initial screening talent vectors.

3. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 2, characterized in that, The maximum step size of the random walk diffusion is determined by the maximum depth of the target job skill tree graph, and the stopping probability of each walk is inversely proportional to the connection weight of the current node.

4. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 1, characterized in that, Based on the target job skill tree map, a multi-dimensional skill attenuation mapping is performed on the multiple initially screened talent vectors to obtain multiple attenuated talent vectors, including: Extract multiple decay functions from a preset skill decay function library that correspond to the number of depth levels in the target job skill tree graph, with each decay function corresponding to one depth level; For each initial talent vector, the initial talent vector is split into multiple skill sub-vectors according to the skill dimension, and each skill sub-vector corresponds to a skill dimension; Based on the depth level of each skill sub-vector in the target job skill tree graph, a corresponding decay function is selected to perform decay operation on the skill sub-vector to obtain multiple decayed skill sub-vectors. The multiple decayed skill sub-vectors are then concatenated to form the decayed talent vector.

5. The AI-driven intelligent matching and recommendation system for job requirements as described in claim 4, characterized in that, Each decay function in the preset skill decay function library is an exponential decay function or a linear decay function with depth level as the independent variable.

6. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 5, characterized in that, The process involves calculating the depth-weighted distance between each decaying talent vector and the root node of the target job skill tree graph, resulting in multiple depth-weighted distance values, including: For each decaying talent vector, a decay coefficient is extracted from each decaying skill sub-vector of the decaying talent vector. Based on the decay coefficient and the depth level corresponding to the decaying skill sub-vector, the level contribution value of the decaying skill sub-vector is calculated. Using the unit vector corresponding to the root node of the target job skill tree graph as a reference vector, calculate the cosine value of the angle between each decayed skill sub-vector of the decayed talent vector and the reference vector to obtain multiple angle cosine values. The depth-weighted distance value is obtained by multiplying the hierarchical contribution value of each decay skill sub-vector by the corresponding cosine of the included angle, summing the results, and dividing by the sum of the decay coefficients of all decay skill sub-vectors in the decay talent vector.

7. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 1, characterized in that, Using the multiple depth-weighted distance values ​​as path constraints, collaborative path backtracking is performed on the multiple initially screened talents to generate multiple collaborative recommendation paths, including: Randomly select a first preliminary screening candidate and a second preliminary screening candidate from the plurality of preliminary screening candidates, and use the difference between the depth-weighted distance value corresponding to the first preliminary screening candidate and the depth-weighted distance value corresponding to the second preliminary screening candidate as the first path constraint threshold. Using the first initially screened talent as the starting point of the path, a breadth-first search is performed in the preset talent relationship graph to search for all talent nodes that satisfy the migration cost between nodes being less than the first path constraint threshold, thereby obtaining the first candidate path set. Using the second initially screened talent as the path endpoint, a reverse path backtracking is performed in the first candidate path set, and the path that can reach the second initially screened talent is taken as the collaborative recommendation path.

8. The AI-driven intelligent talent matching and recommendation system based on job requirements as described in claim 7, characterized in that, The preset talent relationship graph is constructed with talents as nodes and historical collaboration relationships or skill similarity between talents as edge weights.

9. The AI-driven intelligent matching and recommendation system for job-specific needs as described in claim 8, characterized in that, The step of selecting the talent corresponding to the endpoint of each collaborative recommendation path as the target talent for recommendation and outputting a list of target talents includes: For each collaborative recommendation path, calculate the sum of migration costs between all adjacent talent nodes on the collaborative recommendation path to obtain the total migration cost of the collaborative recommendation path, and take the endpoint of the collaborative recommendation path with the minimum total migration cost as the first recommended target talent. After removing the first and second initial screening talents from the multiple initial screening talents, the step of using the multiple depth-weighted distance values ​​as path constraints is repeated to obtain the second recommended target talents. Arrange the first recommended target talent, the second recommended target talent, and all recommended target talents obtained through subsequent repeated executions in the order of acquisition to generate the recommended target talent list.

Images