AI intelligent recommendation engine construction method and system

Abstract

The invention belongs to the field of intelligent recommendation, and particularly relates to an AI intelligent recommendation engine construction method and system, and the method comprises the steps: carrying out the uplink compliance verification of user resume text information through a block chain and a preset compliance verification library, and generating an information verification matching chain; when a target recommendation demand is responded, compliance verification and text analysis are firstly performed on the demand, and the contribution degree of a field to a service is evaluated to obtain an effective matching field; on the basis of the effective fields, forward matching is carried out in combination with a user hierarchical portrait map and a skill growth model in the matching chain, and a matched user pool and accuracy are obtained; meanwhile, the user task completion degree is reversely evaluated through analogue simulation and a task-skill dependency graph, the matching strategy is dynamically adjusted in combination with time causal analysis till the recommendation requirement is met, bidirectional precise matching is achieved through the block chain in combination with the layered portrait and the skill growth model, and the recommendation efficiency is improved through reverse simulation and dynamic adjustment. And the man-post matching precision, efficiency and suitability are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recommendation, and in particular relates to a method and system for building an AI intelligent recommendation engine. Background Technology

[0002] Against the backdrop of rapid development in digital recruitment, enterprises are increasingly reliant on AI-powered intelligent recommendation engines, hoping to achieve efficient and accurate matching of people and jobs through technological means. However, traditional recommendation engines have revealed significant shortcomings in practical applications: On the one hand, resume information, as the core basis for matching, relies heavily on manual or simple rule verification for authenticity checks, leading to frequent issues such as identity fraud, false qualification reports, and distorted credit records, resulting in insufficient credibility of the basic matching data. On the other hand, user profiles are mostly simple aggregations of static tags, making it difficult to build a comprehensive understanding through multi-dimensional feature correlation analysis, nor can they quantify the growth rate of skills over time, making it difficult to predict the long-term suitability potential of candidates. More importantly, the matching process is mostly one-way screening, generating results only through a superficial comparison of resume fields and job requirements, lacking the ability to verify based on actual business scenarios. Furthermore, when the number of candidates meeting the requirements is insufficient, there is a lack of precise adjustment criteria that incorporate in-depth information such as skill gaps and growth cycles, ultimately resulting in low accuracy and poor suitability of recommendation results, making it difficult to balance enterprises' dual needs for recommendation quality and quantity. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an AI intelligent recommendation engine construction method and system. This method uses blockchain and a pre-set compliance verification library to perform on-chain compliance verification of user resume text information, generating an information verification matching chain. When responding to a target recommendation request, the request is first subjected to compliance verification and text parsing to evaluate the contribution of fields to the business and obtain effective matching fields. Based on the effective fields, forward matching is performed using a user hierarchical profile graph and a skill growth model in the matching chain to obtain a matching user pool and accuracy. Simultaneously, user task completion is evaluated inversely through simulation and a task-skill dependency graph, and the matching strategy is dynamically adjusted based on time causal analysis until the recommendation requirement is met. This application achieves bidirectional accurate matching through blockchain combined with hierarchical profiles and skill growth models. Through inverse simulation and dynamic adjustment, it significantly improves the accuracy, efficiency, and suitability of person-job matching.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for building an AI-powered intelligent recommendation engine includes:

[0006] Based on the text information of the user's resume to be matched, combined with blockchain and a pre-set compliance verification library, on-chain compliance verification is performed to obtain an information verification matching chain;

[0007] In response to the target recommendation requirements, the on-chain compliance verification and target requirement text parsing are performed on the target recommendation requirements. At the same time, based on the parsed target requirement text fields, the contribution of the business in the recommendation requirements is evaluated to obtain the effective matching fields.

[0008] Based on the effective matching fields and the information, the hierarchical profile map of the users to be matched in the matching chain is verified and the skill growth model is positively matched to obtain the matching user pool and the corresponding matching accuracy.

[0009] The hierarchical profile map is constructed by combining the corresponding user's skill proficiency, years of experience, and qualification verification results with a graph algorithm; the skill growth model is constructed by combining the corresponding user's change in technical proficiency, task completion rate, and the change in the time to complete the same amount of task with a time series analysis algorithm, and is used to measure the growth rate of the user's corresponding skills.

[0010] Specifically, the methods for building intelligent recommendation engines also include:

[0011] Based on the matching user pool and the corresponding matching accuracy, a reverse task completion simulation evaluation is performed using a simulation algorithm and a task-skill growth dependency graph to obtain the simulation completion score and evaluation accuracy for each user. The task-skill growth dependency graph is constructed by combining a skill growth model with the complexity of the business requirements and a reinforcement learning algorithm.

[0012] The system presets completion score threshold and evaluation accuracy threshold. When the number of users to be matched that simultaneously meet both the completion score threshold and the evaluation accuracy threshold meets the target recommendation requirement, the matched users are directly recommended to the corresponding target users.

[0013] When the number of users who meet both the completion score threshold and the evaluation accuracy threshold does not meet the target recommendation requirements, users who meet both the completion score threshold and the evaluation accuracy threshold will be recommended to users who meet the target requirements. At the same time, time causal analysis will be performed based on the simulation process information of users who do not meet the threshold requirements to obtain the inference matrix of information on unmatched users.

[0014] The inference matrix of mismatched user information and the number of missing users to be matched are fed back into the skill growth model to adjust the positive matching. At the same time, the reverse simulation process is repeated until the number of users that meet the threshold condition meets the number required for the target recommendation.

[0015] Specifically, the process of constructing the information verification matching chain includes:

[0016] Based on a pre-trained text preprocessing model combining the resume text of the user to be matched and the target recommendation requirement text, a set of fields for the user to be matched and a set of fields for the target recommendation are obtained. The set of fields for the user to be matched includes, but is not limited to, identity identification fields, skill tag fields, qualification certificate fields, credit record fields, and work history subfields. The target recommendation fields include subfields for enterprise qualifications, recruitment authority, historical recruitment compliance, business domain, task type, skill requirements, and compliance constraints.

[0017] Based on a single set of user fields to be matched, and with the historical success rate as the objective, we combine principal component analysis algorithm and principal component evaluation threshold to obtain a valid set of user fields to be matched.

[0018] Based on a single set of valid fields for a user to be matched, and combined with a clustering algorithm targeting field type, a cluster of valid fields for the user to be matched is obtained.

[0019] Based on the single user's effective field clustering cluster combined with graph algorithm, a single user cluster profile is obtained. At the same time, based on different user cluster profiles, clustering is performed with the historical authenticity score of the corresponding type field as the target to obtain the verification cluster set of key fields of different users. Combined with knowledge graph, a hierarchical profile graph of the user to be matched is constructed.

[0020] Specifically, the process of constructing the information verification matching chain also includes:

[0021] Based on the verification level clustering of different user key fields and combined with expert experience algorithms, the compliance verification level of different users under the verification clustering of each field type is obtained; the different levels of clustering subclusters include the clustering results of different users at the same compliance verification level in the same type of field, which are used to perform different levels of compliance verification according to the authenticity score of the user's corresponding type of field.

[0022] A first verification connection is established based on the compliance verification level of different users under different levels of cluster subclusters in each field type and the first compliance verification rule in the compliance verification library.

[0023] Based on the first verification connection, the user's resume text to be matched is combined with the first compliance verification rule to perform differential compliance verification. When all types of fields of the corresponding user pass the verification, the user is allowed to go on the chain, and the user allowed to go on the chain is marked with a chain authorization mark.

[0024] Specifically, the process of constructing the information verification matching chain also includes:

[0025] Based on the authorized user's on-chain authorization identifier, combined with the corresponding user's valid field set and compliance verification results, the SHA-256 algorithm generates a unique hash value for each verified user. The hash value, cluster identifier, and verification timestamp are then uploaded to the consortium blockchain through the PoS consensus mechanism, forming an on-chain record that includes field-level evidence storage nodes and cluster-level evidence storage nodes.

[0026] The first compliance verification rule includes verification rules for identity authenticity, validity of qualification certificate fields and work history sub-fields, and compliance of credit records;

[0027] Based on the target recommendation field set and field types, the contribution of each type of field to the required business information is obtained through factor analysis algorithm. Then, the target recommendation field set is preprocessed by combining the contribution of each type of field to the required business information with an attention sparse network, and the preprocessed target recommendation field set and field types are obtained.

[0028] Specifically, the process of constructing the information verification matching chain also includes:

[0029] Based on the preprocessed target recommendation field set and field type corresponding to different demand users with target recommendation needs, combined with the second compliance verification rules in the compliance verification library, a second verification connection is established through the establishment process of the first verification connection described above. The second compliance verification rules are used to verify the compliance of the demand user's enterprise qualification authenticity, recruitment authority scope, historical recruitment compliance sub-field scores, business field and task type, skill requirements and compliance constraint sub-fields.

[0030] Based on the second verification connection, combined with the target recommendation requirement text and the second compliance verification rules, the verified demand users are obtained. At the same time, based on the target recommendation field set of the demand users, the above process of on-chain storage of the matching users is repeated to generate a unique hash value and store it on the chain.

[0031] Based on the hash association relationship of on-chain evidence storage nodes, an information verification and matching chain is constructed through a Merkle tree structure. Specifically, the main matching node is the skill requirement cluster in the target requirements of different users in the target recommendation requirements, and the remaining clusters are auxiliary matching nodes. The nodes are associated with the nodes of each user cluster profile in the hierarchical profile map of the user to be matched. The hierarchical profile map of the user to be matched is used as the parent node, and the cluster nodes of each field type in the cluster profile of the user to be matched are associated. The cluster nodes are used as child nodes, and the field nodes of a single user to be matched within the cluster are associated. Each node contains the hash of the preceding node, the unique hash value of the identity information, the cluster identifier, and the verification result summary.

[0032] Specifically, the process of obtaining the matching user pool and the corresponding matching accuracy includes:

[0033] Based on the node information of the hierarchical profile graph of the user to be matched in the information verification and matching chain, the skill proficiency feature vector, experience years feature value, and qualification certificate verification result feature vector are extracted from the nodes of each user cluster profile to obtain the user's basic feature set. Among them, the skill proficiency feature vector corresponds to the cluster feature of the skill tag field in the hierarchical profile graph, the experience years feature value is the value of the work history subfield after time quantization, and the qualification certificate verification result feature vector is the binary identifier vector of the compliance verification result of the qualification certificate field.

[0034] Based on the skill proficiency feature vector in the user's basic feature set, the skill proficiency trend is predicted by time series analysis algorithm to obtain the skill growth prediction vector within the future preset period. At the same time, combined with the user's historical skill growth rate, a dynamic skill weight coefficient is generated.

[0035] Based on the preprocessed target recommendation field set and field type, the skill requirement subfield is converted into a demand skill feature vector through a word embedding model, and the business domain and task type subfield is converted into a business scenario feature vector to obtain the demand feature set.

[0036] Specifically, the process of obtaining the matching user pool and the corresponding matching accuracy includes:

[0037] Based on the association relationship of the Merkle tree structure in the information verification matching chain, the target skill weight distribution of demand users in the main matching node with skill demand cluster is extracted by graph traversal algorithm, and business scenario constraint parameters are extracted from the auxiliary matching node.

[0038] Based on the user's basic feature set and demand feature set, the first similarity between the skill proficiency feature vector and the demand skill feature vector and the user's business scenario adaptability are calculated by the cosine similarity algorithm. The first similarity is then weighted and corrected by the skill dynamic weight coefficient to obtain the skill matching score.

[0039] Based on the experience years feature value and qualification certificate verification result feature vector of the user to be matched, and the experience requirement threshold and qualification compliance threshold in the requirement feature set corresponding to the user, a logistic regression model is used to compare and output the experience matching score and label and the qualification matching score and label.

[0040] Based on the on-chain evidence verification results of the users to be matched in the information verification matching chain, the verified users are selected as a candidate set. The skill matching score, experience matching score and qualification matching score of the users in the candidate set are weighted and summed to obtain a comprehensive matching score and then identified accordingly.

[0041] A preset comprehensive matching score threshold is set. Users whose comprehensive matching scores are greater than the threshold are included in the matching user pool. The matching accuracy is obtained by dividing the number of users in the matching user pool who pass the subsequent reverse verification by the total number of users in the matching user pool.

[0042] An AI-powered intelligent recommendation engine building system includes: a matching chain module, a demand parsing module, and a positive matching module;

[0043] The matching chain module performs on-chain compliance verification based on the resume text information of the user to be matched, combined with the blockchain and a preset compliance verification library, to obtain an information verification matching chain; the compliance verification library contains verification rules for identity authenticity, qualification validity, and credit record compliance.

[0044] The requirement parsing module is used to respond to target recommendation requirements, perform on-chain compliance verification and target requirement text parsing on the target recommendation requirements, and evaluate the contribution of the business in the recommendation requirements based on the parsed target requirement text fields to obtain valid matching fields.

[0045] The positive matching module performs positive matching based on the effective matching fields and the information verification matching chain, combining the hierarchical profile map of the users to be matched with the skill growth model to obtain the matching user pool and the corresponding matching accuracy. The hierarchical profile map is constructed by combining the verification results of the corresponding user's skill proficiency, years of experience, and qualification certificates with a graph algorithm. The skill growth model is constructed by combining the change in the corresponding user's technical proficiency, the change in task completion rate, and the change in the time to complete the same amount of task with a time series analysis algorithm, and is used to measure the growth rate of the user's corresponding skills.

[0046] Specifically, the intelligent recommendation engine building system also includes a simulation evaluation module, an evaluation recommendation module, and a feedback adjustment module;

[0047] The simulation evaluation module performs reverse task completion simulation evaluation based on the matching user pool and the corresponding matching accuracy, combined with the simulation algorithm and the task-skill growth dependency graph, to obtain the simulation completion score and evaluation accuracy for each user; the task-skill growth dependency graph is constructed by combining the skill growth model with the complexity of the business requirements and the reinforcement learning algorithm.

[0048] The evaluation and recommendation module presets a completion score threshold and an evaluation accuracy threshold. When the number of users to be matched that simultaneously meet both the completion score threshold and the evaluation accuracy threshold is sufficient to meet the target recommendation requirements, the matched users are directly recommended to the users with the corresponding target requirements. When the number of users to be matched that simultaneously meet both the completion score threshold and the evaluation accuracy threshold is insufficient to meet the target recommendation requirements, the users who meet both the completion score threshold and the evaluation accuracy threshold are recommended to the users with the target requirements. At the same time, time causal analysis is performed based on the simulation process information of the users to be matched that do not meet the threshold requirements to obtain the inference matrix of information on the unmatched users.

[0049] The feedback adjustment module is used to feed back the inference matrix of mismatched user information and the number of missing users to be matched into the skill growth model to adjust the positive matching, and at the same time repeat the reverse simulation process until the number of users that meet the threshold condition meets the number required for the target recommendation.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] This invention addresses the shortcomings of existing technologies by combining blockchain with a compliance verification library to generate an information verification and matching chain. This ensures the authenticity and credibility of resume information, resolving issues of information distortion related to identity, qualifications, and credit from the source and improving the reliability of matching foundational data. Utilizing a hierarchical profile graph and a skills growth model, it comprehensively depicts a user's skill proficiency, years of experience, and qualification-related characteristics, quantifying skills growth rates and overcoming the limitations of static labels to achieve accurate assessment of a user's long-term suitability potential. Through a two-way mechanism of positive matching and reverse task completion simulation evaluation, combined with time causal analysis and closed-loop optimization, it ensures the accuracy of recommendation results while dynamically adjusting the matching strategy when the number of candidates is insufficient, balancing the needs for recommendation quality and quantity, and significantly improving the efficiency and suitability of person-job matching. Attached Figure Description

[0052] Figure 1 This is a flowchart of an AI intelligent recommendation engine construction method according to the present invention;

[0053] Figure 2 This is a structural diagram of an AI intelligent recommendation engine construction system according to the present invention. Detailed Implementation

[0054] Example 1

[0055] Please see Figure 1 The present invention provides an embodiment of an AI intelligent recommendation engine construction method, the steps of which include:

[0056] S1. Based on the text information of the resume of the user to be matched, combined with the blockchain and the preset compliance verification library, perform on-chain compliance verification to obtain the information verification matching chain;

[0057] S2. Respond to the target recommendation requirements, perform on-chain compliance verification and target requirement text parsing on the target recommendation requirements, and at the same time, evaluate the contribution of the business in the recommendation requirements based on the parsed target requirement text fields to obtain valid matching fields.

[0058] S3. Based on the effective matching fields and the information, verify the hierarchical profile map of the users to be matched in the matching chain and the skill growth model to perform positive matching, and obtain the matching user pool and the corresponding matching accuracy.

[0059] It should be further explained that the hierarchical profile map in this embodiment is constructed by combining the corresponding user's skill proficiency, years of experience, and qualification certificate verification results with a graph algorithm; the skill growth model in this embodiment is constructed by combining the change in the corresponding user's corresponding technical proficiency, the change in task completion rate, and the change in the time to complete the same amount of task with a time series analysis algorithm, and is used to measure the growth rate of the user's corresponding skills.

[0060] S4. Based on the matching user pool and the corresponding matching accuracy, combined with the simulation algorithm and the task-skill growth dependency graph, a reverse task completion simulation evaluation is performed to obtain the simulation completion score and evaluation accuracy for each user; the task-skill growth dependency graph is constructed by combining the skill growth model with the demand business complexity and the reinforcement learning algorithm.

[0061] S5. Preset completion score threshold and evaluation accuracy threshold. When the number of users to be matched that simultaneously meet the completion score threshold and evaluation accuracy threshold meets the number in the target recommendation requirement, the matched users will be directly recommended to the corresponding target users.

[0062] S6. When the number of users to be matched that simultaneously meet the completion score threshold and the evaluation accuracy threshold does not meet the number required by the target recommendation, users who meet the completion score threshold and the evaluation accuracy threshold will be recommended to the target users. At the same time, time causal analysis will be performed based on the simulation process information of users to be matched that do not meet the threshold requirements to obtain the inference matrix of information on unmatched users.

[0063] S7. Feed the inference matrix of mismatched user information and the number of missing users to be matched back into the skill growth model to adjust the positive matching, and repeat the reverse simulation process until the number of users that meet the threshold condition meets the number required for the target recommendation.

[0064] It should be further explained that the construction process of the information verification matching chain in this embodiment includes:

[0065] Based on a pre-trained text preprocessing model combining the resume text of the users to be matched and the target recommendation request text, a set of fields for the users to be matched and a set of fields for the target recommendation are obtained. The set of fields for the users to be matched includes, but is not limited to, identity identification fields, skill tag fields, qualification certificate fields, credit record fields, and work history subfields. The target recommendation fields include enterprise qualifications, recruitment authority, historical recruitment compliance subfields, business domain, task type, skill requirements, and compliance constraints subfields. It should be further noted that, in this embodiment, in the set of fields for the users to be matched, the identity identification field is used to confirm the uniqueness of the user's identity, the skill tag field reflects the user's professional ability, the qualification certificate field verifies the validity of the user's professional qualifications, the credit record field reflects the user's credit status, and the work history subfield presents the user's work experience. In the target recommendation fields, the enterprise qualifications, recruitment authority, and historical recruitment compliance subfields ensure the compliance of the requesting party, the business domain and task type clarify the requirement scenario, and the skill requirements and compliance constraints subfields define the matching standards. Both types of fields provide basic data support for subsequent compliance verification and accurate matching.

[0066] Based on a single set of user fields to be matched, and with the historical success rate as the objective, we combine principal component analysis algorithm and principal component evaluation threshold to obtain a valid set of user fields to be matched.

[0067] Based on a single set of valid fields for a user to be matched, and combined with a clustering algorithm targeting field type, a cluster of valid fields for the user to be matched is obtained.

[0068] Based on the single user's effective field clustering cluster combined with graph algorithm, a single user cluster profile is obtained. At the same time, based on different user cluster profiles, clustering is performed with the historical authenticity score of the corresponding type field as the target to obtain the verification cluster set of key fields of different users. Combined with knowledge graph, a hierarchical profile graph of the user to be matched is constructed.

[0069] Based on the verification level clustering of different user key fields and combined with expert experience algorithms, the compliance verification level of different users under the verification clustering of each field type is obtained; the different levels of clustering subclusters include the clustering results of different users at the same compliance verification level in the same type of field, which are used to perform different levels of compliance verification according to the authenticity score of the user's corresponding type of field.

[0070] A first verification connection is established based on the compliance verification level of different users under different levels of cluster subclusters in each field type and the first compliance verification rule in the compliance verification library.

[0071] Based on the first verification connection, the user's resume text to be matched is combined with the first compliance verification rule to perform differential compliance verification. When all types of fields of the corresponding user pass the verification, the user is allowed to go on the chain, and the user allowed to go on the chain is marked with a chain authorization mark.

[0072] Based on the authorized user's on-chain authorization identifier, combined with the corresponding user's valid field set and compliance verification results, the SHA-256 algorithm generates a unique hash value for each verified user. The hash value, cluster identifier, and verification timestamp are then uploaded to the consortium blockchain through the PoS consensus mechanism, forming an on-chain record that includes field-level evidence storage nodes and cluster-level evidence storage nodes.

[0073] The first compliance verification rule includes verification rules for identity authenticity, validity of qualification certificate fields and work history sub-fields, and compliance of credit records;

[0074] Based on the target recommendation field set and field types, the contribution of each type of field to the demand business information is obtained through factor analysis algorithm. The contribution of each type of field to the demand business information is used as a weight to combine with attention sparse network to perform sparsification preprocessing on the target recommendation field set, so as to obtain the preprocessed target recommendation field set and field types.

[0075] Based on the preprocessed target recommendation field set and field type corresponding to different demand users with target recommendation needs, combined with the second compliance verification rules in the compliance verification library, a second verification connection is established through the establishment process of the first verification connection described above. The second compliance verification rules are used to verify the compliance of the demand user's enterprise qualification authenticity, recruitment authority scope, historical recruitment compliance sub-field scores, business field and task type, skill requirements and compliance constraint sub-fields.

[0076] Based on the second verification connection, combined with the target recommendation requirement text and the second compliance verification rules, the verified demand users are obtained. At the same time, based on the target recommendation field set of the demand users, the above process of on-chain storage of the matching users is repeated to generate a unique hash value and store it on the chain.

[0077] Based on the hash association relationship of on-chain evidence storage nodes, an information verification and matching chain is constructed through a Merkle tree structure. Specifically, the main matching node is the skill requirement cluster in the target requirements of different users in the target recommendation requirements, and the remaining clusters are auxiliary matching nodes. The nodes are associated with the nodes of each user cluster profile in the hierarchical profile map of the user to be matched. The hierarchical profile map of the user to be matched is used as the parent node, and the cluster nodes of each field type in the cluster profile of the user to be matched are associated. The cluster nodes are used as child nodes, and the field nodes of a single user to be matched within the cluster are associated. Each node contains the hash of the preceding node, the unique hash value of the identity information, the cluster identifier, and the verification result summary.

[0078] It should be further explained that one way to construct the Merkle tree structure in this embodiment is as follows:

[0079] Based on the skill requirement clusters, remaining clusters, hierarchical profile nodes of users to be matched, field type clusters of users to be matched, and field nodes of individual users to be matched in the target recommendation requirements, the node hierarchy of the Merkle tree is determined as follows: the skill requirement clusters are the main matching nodes (first-level branch nodes), the remaining clusters are the auxiliary matching nodes (first-level branch nodes), the hierarchical profile nodes of users to be matched are the second-level branch nodes, the field type clusters are the third-level branch nodes, and the field nodes of individual users to be matched are the leaf nodes.

[0080] Based on the hash of the preceding node contained in a single user field node to be matched, namely the hash value of the corresponding field type cluster node, the unique hash identifier of identity information (the unique hash of the user field), the cluster identifier (the identifier of the field type cluster to which it belongs), and the verification result digest, the hash value of the leaf node is calculated by a hash algorithm to obtain the leaf node hash;

[0081] Based on the hash of the preceding node contained in the cluster node of the field type cluster, namely the hash value of the corresponding user hierarchical profile graph node to be matched, the unique hash identifier of identity information (unique hash of the cluster), the cluster identifier (identifier of the hierarchical profile graph to which it belongs), and the verification result summary, combined with the hash values ​​of all leaf nodes under the cluster, the hash value of the third-level branch node is calculated by the Merkle leaf node aggregation algorithm to obtain the hash of the third-level branch node.

[0082] Based on the hash of the preceding node (the hash value of the corresponding main matching node or auxiliary matching node), the unique hash identifier of the identity information (the unique hash of the hierarchical profile graph), the cluster identifier (the identifier of the demand user's needs), and the verification result summary contained in the hierarchical profile graph node of the user to be matched, combined with the hash values ​​of all third-level branch nodes under the graph, the hash value of the second-level branch node is calculated by the Merkle tree branch node aggregation algorithm to obtain the hash of the second-level branch node.

[0083] Based on the hash of the preceding node (the hash value of the root node of the demand user), the unique hash identifier of the identity information (the unique hash of the skill demand cluster), the cluster identifier (the unique identifier of the demand user), and the verification result summary contained in the main matching node (skill demand cluster), combined with the hash values ​​of all secondary branch nodes associated with the main matching node, the hash value of the main matching node is calculated by the Merkle tree first-level branch aggregation algorithm to obtain the hash of the main matching node.

[0084] Based on the hash of the preceding node (the hash value of the root node of the demand user), the unique hash identifier of the identity information (the unique hash of the remaining cluster), the cluster identifier (the unique identifier of the demand user), and the verification result digest contained in the auxiliary matching node (remaining cluster), combined with the hash values ​​of all secondary branch nodes associated with the auxiliary matching node, the hash value of the auxiliary matching node is calculated by the Merkle tree first-level branch aggregation algorithm to obtain the auxiliary matching node hash;

[0085] Based on the unique hash identifier of the user's identity information (the unique hash of the user), the cluster identifier (the unique identifier of the target recommendation requirement) and the verification result summary contained in the root node of the user, the root node hash is generated by combining the hash of the main matching node and the hash of the auxiliary matching node through the Merkle tree root node calculation algorithm, and the Merkle tree root hash of the information verification matching chain is obtained.

[0086] By hashing the root node, the hash values ​​of all level nodes are associated to form an information verification and matching chain based on the Merkle tree structure. This enables hierarchical hash association of the main matching node, auxiliary matching nodes, second-level branch nodes, third-level branch nodes, and leaf nodes, ensuring the traceability and immutability of information at each node.

[0087] It should be further explained that the specific description of the above nodes in this embodiment is as follows:

[0088] Main matching node: A first-level branch node that corresponds to the skill requirement cluster of the demand users in the target recommendation requirements. It is used to prioritize associating profile nodes of users to be matched that match the core skill requirements.

[0089] Auxiliary matching nodes: First-level branch nodes, corresponding to the remaining clusters of demand users (non-skill demand clusters) in the target recommendation requirements, used to associate profile nodes of users to be matched with non-core requirements.

[0090] Second-level branch nodes: These are the user hierarchical profile graph nodes to be matched, associated with the main matching node or auxiliary matching node, serving as the aggregation node for the overall user profile, and connected to field type cluster nodes.

[0091] Third-level branch nodes: These are field type cluster nodes, associated with second-level branch nodes, serving as cluster aggregation nodes for fields of the same type (such as skill tag fields and qualification certificate fields), and connected to individual user field nodes.

[0092] Leaf node: This is a single field node for the user to be matched. It is the lowest level node in the Merkle tree and corresponds to the specific field of the user to be matched (such as a skill tag or a qualification certificate).

[0093] Demand User Root Node: The top-level node of the Merkle tree, which connects the primary matching node and the secondary matching node, and serves as the root node of the Merkle tree for a specific target recommendation requirement.

[0094] It should be further explained that the attribute variables of the above-mentioned nodes in this embodiment are specifically as follows:

[0095] Preceding Node Hash: The hash value of the parent node of each node, used to reflect the hierarchical relationship. Specifically, the preceding node hash of a leaf node corresponds to the hash value of a third-level branch node; the preceding node hash of a third-level branch node corresponds to the hash value of a second-level branch node; the preceding node hash of a second-level branch node corresponds to the hash value of the main matching node or auxiliary matching node; and the preceding node hash of both the main matching node and auxiliary matching node corresponds to the hash value of the demand user's root node.

[0096] Unique Hash Identifier for Identity Information: Each node has a unique hash identifier used to distinguish different nodes. Specifically, the unique hash identifier for leaf nodes is a unique hash of a single user field; for third-level branch nodes, it is a unique hash of the field type cluster; for second-level branch nodes, it is a unique hash of the hierarchical profile of the user to be matched; for main matching nodes, it is a unique hash of the skill requirement cluster; for auxiliary matching nodes, it is a unique hash of the remaining clusters; and for the root node of the demand user, it is a unique hash of the demand user.

[0097] Cluster Identifier: The identifier of each node's cluster system, used to locate the node's position in the cluster structure. Specifically, the cluster identifier of a leaf node is the identifier of its field type cluster; the identifier of a third-level branch node is the identifier of its hierarchical profile map; the identifier of a second-level branch node is the identifier of the corresponding demand user's requirement; the identifiers of the main matching node and auxiliary matching node are unique identifiers of the demand user; and the unique identifier of the demand user root node is the unique identifier of the target recommendation requirement.

[0098] Verification Result Summary: A collection of compliance verification information for each node, including verification object type, verification type label, verification status identifier, non-compliance item field, verification level identifier, compliance verification level, verification timestamp, and verification node signature, used to record the verification process and results of the node.

[0099] It should be further explained that the specific process of hash calculation of variables in this embodiment is as follows:

[0100] Leaf node hash: The unique hash value of a leaf node is calculated by a hash algorithm based on the hash of its predecessor node, the unique hash identifier of its identity information, the cluster identifier, and the verification result summary.

[0101] Third-level branch node hash: The unique hash value of the third-level branch node is calculated by the Merkle leaf node aggregation algorithm based on the hash of the preceding node, the unique hash identifier of the identity information, the cluster identifier, the verification result digest, and the hash of all leaf nodes under it.

[0102] Second-level branch node hash: The unique hash value of the second-level branch node is calculated by the Merkle tree branch node aggregation algorithm based on the hash of the preceding node, the unique hash identifier of the identity information, the cluster identifier, the verification result digest, and the hash of all the third-level branch nodes below it.

[0103] Main matching node hash: The unique hash value of the main matching node is calculated by the Merkle tree first-level branch aggregation algorithm based on the hash of the preceding node, the unique hash identifier of the identity information, the cluster identifier, the verification result summary, and the hashes of all its associated second-level branch nodes.

[0104] Auxiliary matching node hash: The unique hash value of the auxiliary matching node is calculated by the Merkle tree first-level branch aggregation algorithm based on the hash of the preceding node, the unique hash identifier of the identity information, the cluster identifier, the verification result summary, and the hash of all its associated second-level branch nodes.

[0105] Root node hash: Based on the unique hash identifier, cluster identifier, and verification result digest of the demand user's root node identity information, combined with the hash of the main matching node and the hash of the auxiliary matching node, the unique hash value of the top level of the Merkle tree is generated by the Merkle tree root node calculation algorithm. It is used to associate all level nodes to form an information verification matching chain.

[0106] It should be further explained that the verification result summary in this embodiment includes variables such as verification object type, verification type label, verification status identifier, non-compliance item field, verification level identifier, compliance verification level, verification timestamp and verification node signature. It comprehensively records key information such as verification object, type, result, problem, level, grade, time and signature, providing a traceable and reliable verification basis for the information verification matching chain.

[0107] The process first extracts multi-dimensional fields of the users to be matched and the target recommendations through a pre-trained text preprocessing model, and then implements hierarchical compliance verification by combining a compliance verification library. The identity, skills, qualifications and other fields of the users to be matched and the enterprise qualifications, recruitment authority and other fields of the demand side must all be strictly verified. The verification results are stored on the blockchain to form an immutable on-chain record. The hierarchical association design of the Merkle tree structure uses skill requirement clusters as main matching nodes and remaining clusters as auxiliary matching nodes. These clusters are progressively associated with user hierarchical profiles, field clusters, and individual field nodes. Each node contains a preorder hash, a unique identifier, and a verification result summary. This ensures data traceability and verifies information integrity through hash association, fundamentally solving problems such as data falsification and tampering in traditional recommendations, and providing reliable basic data for matching. Secondly, principal component analysis is used to extract effective fields, and a hierarchical profile map is constructed by combining clustering and graph algorithms to achieve a three-dimensional presentation of user skill proficiency, years of experience, and qualification verification results. The skill growth model is based on time series analysis, dynamically capturing the trends of changes in user technical proficiency, task completion rate, and the time to complete the same task, quantifying the skill growth rate, breaking through the limitations of static labels, and accurately predicting the user's long-term adaptation potential. This dynamic characterization allows recommendations to go beyond current capability matching, uncovering the long-term fit between users and their needs and enhancing the foresight of the matching process. Furthermore, based on effective matching fields, a word embedding model transforms required skills and business scenarios into feature vectors. Combined with cosine similarity calculation and dynamic weight correction, accurate matching between users and their needs is achieved. Target skill weights are extracted from skill requirement clusters, and business constraints are extracted from auxiliary nodes. A weighted summation is used to generate a comprehensive matching score, ensuring that the matching results both align with skill requirements and adapt to business scenarios. Simultaneously, an accuracy evaluation mechanism for the matching user pool further optimizes matching quality and reduces invalid recommendations through reverse verification of results.

[0108] It should be further explained that the process of obtaining the matching user pool and the corresponding matching accuracy in this embodiment includes:

[0109] Based on the node information of the hierarchical profile graph of the users to be matched in the information verification and matching chain, the skill proficiency feature vector, experience years feature value, and qualification certificate verification result feature vector are extracted from the nodes of each user cluster profile to obtain the user's basic feature set. Among them, the skill proficiency feature vector corresponds to the cluster feature of the skill tag field in the hierarchical profile graph, the experience years feature value is the value of the work history sub-field after time quantization, and the qualification certificate verification result feature vector is the binary identifier vector of the compliance verification result of the qualification certificate field. It should be further noted that in this embodiment... For the clustering features of the skill tag field in the hierarchical profile map, it is necessary to clarify the normalization processing method of skill tags within the cluster. Specifically, based on the frequency of skill tags of all users in the same cluster, the skill tags are mapped to basic values ​​representing proficiency through min-max normalization; for skill tags that overlap with multiple clusters (i.e., the same skill belongs to multiple clusters), weighted fusion is performed through inter-cluster association weights (calculated based on inter-cluster co-occurrence frequency) to avoid redundant feature calculations; finally, redundant dimensions are eliminated through feature selection algorithms (such as the variance inflation factor method) to ensure the independence of feature vectors.

[0110] When extracting time spans based on the work history subfield in this embodiment, a time continuity verification mechanism needs to be introduced. Specifically, the start and end times in the resume are parsed using regular expressions. For abnormal resumes with overlapping or broken times, reasonable corrections are made based on the industry average length of service. For current work history without a clear end time, the cumulative duration is calculated based on the current time and marked with an "on-the-job" status. During the time quantification process, the original record with monthly precision needs to be retained and then converted into an annual feature value. At the same time, the time granularity identifier, such as "accurate to the month" or "estimated value", is stored.

[0111] Based on the skill proficiency feature vector in the user's basic feature set, a skill growth model constructed through time series analysis algorithm is used to predict the skill proficiency trend and obtain the skill growth prediction vector within a preset period in the future. At the same time, combined with the user's historical skill growth rate, a dynamic weight coefficient for the skill is generated. It should be further noted that the skill proficiency feature vector in this embodiment includes, but is not limited to, the change in technical proficiency and the change in task completion degree and the time to complete the same amount of task.

[0112] Based on the preprocessed target recommendation field set and field type, the skill requirement subfield is converted into a demand skill feature vector through a word embedding model, and the business domain and task type subfield is converted into a business scenario feature vector to obtain the demand feature set.

[0113] Based on the association relationship of the Merkle tree structure in the information verification matching chain, the target skill weight distribution of demand users in the main matching node with skill demand cluster is extracted by graph traversal algorithm, and business scenario constraint parameters are extracted from the auxiliary matching node.

[0114] Based on the user's basic feature set and demand feature set, the first similarity between the skill proficiency feature vector and the demand skill feature vector is calculated using the cosine similarity algorithm, along with the user's business scenario adaptability. The first similarity is then weighted and corrected using a dynamic skill weight coefficient to obtain a skill matching score. It should be further noted that in this embodiment, the model input features are the difference between the experience years feature value and the experience requirement threshold, and the matching ratio between the qualification verification result feature vector and the qualification compliance threshold. The training data consists of the correlation data between experience, qualifications, and successful matches in historical matching cases, and the model parameters are optimized using 5-fold cross-validation. The qualification matching score is calculated based on the proportion of verified qualification fields, and a veto mechanism is set for core qualification fields (specified by the demand side); if a core field fails, the qualification score is directly 0.

[0115] Based on the experience years feature value and qualification certificate verification result feature vector of the user to be matched, and the experience requirement threshold and qualification compliance threshold in the requirement feature set corresponding to the user, a logistic regression model is used to compare and output the experience matching score and label and the qualification matching score and label.

[0116] Based on the on-chain evidence verification results of the users to be matched in the information verification matching chain, the verified users are selected as a candidate set. The skill matching score, experience matching score and qualification matching score of the users in the candidate set are weighted and summed to obtain a comprehensive matching score and then identified accordingly.

[0117] A preset comprehensive matching score threshold is set. Users whose comprehensive matching scores are greater than the threshold are included in the matching user pool. The matching accuracy is obtained by dividing the number of users in the matching user pool who pass the subsequent reverse verification by the total number of users in the matching user pool.

[0118] It should be further explained that one implementation process of the reverse task completion simulation evaluation in this embodiment includes:

[0119] Based on the matching user pool, the hierarchical profile graph node information of users in the pool is extracted through field parsing tools. Specifically, this includes: for the skill proficiency feature vector, the skill tag field is converted into a 128-dimensional vector using one-hot encoding, with each dimension corresponding to the skill proficiency; for the experience years feature value, the time span in the work history sub-field is converted into a floating-point value, accurate to one month; for the qualification certificate verification result feature vector, the compliance verification result of the qualification certificate field is binary encoded, with a 1 for passing and a 0 for failing.

[0120] The user compliance identifier (0 / 1) is extracted from the cluster-level evidence storage nodes of the information verification and matching chain using a blockchain node query tool.

[0121] The time series analysis module is called from the skill growth model, and the output of the monthly average change in technical proficiency, task completion rate, and change in the time of completion of the same amount of task are summarized to form a basic dataset for user evaluation. In this embodiment, the time series analysis module is preferably a pre-trained ARIMA(2,1,1) model.

[0122] It should be further explained that in this embodiment, skill proficiency information at each time point is extracted based on skill assessment records stored in the user skill growth model over a period of time. An outlier identification mechanism is used to mark data that significantly deviates from the normal range in the assessment records, and outliers are replaced using trend interpolation of adjacent valid assessment records to ensure data continuity. The cleaned data is categorized by skill type, and for each skill type, the change in proficiency between different assessment time points is calculated to reflect the magnitude of skill improvement or decline in adjacent time periods. To eliminate the interference of short-term fluctuations on the trend, a sliding window technique is used to smooth the changes, weakening abrupt changes caused by accidental factors. Based on the smoothed changes, the average change level per unit time is comprehensively calculated to form the monthly average change value of skill proficiency, thereby quantifying the dynamic development trend of skills over time. Skill types include, but are not limited to, programming languages, project management, and professional tools.

[0123] From the work history sub-field of the information verification and matching chain, extract the task records that the user participated in over a period of time, including task type, planned period, actual completion status, and acceptance results. Based on the business domain of the target recommendation requirements, filter out core tasks related to the business through task type tag matching, and exclude irrelevant auxiliary tasks to ensure the relevance of the analysis object to the current matching scenario. Clarify the criteria for judging task completion: combine the matching degree between the task's planned period and the actual completion time, as well as the acceptance results stored on the chain, such as the acceptance pass mark. Only tasks that meet both the time requirement and the acceptance criteria are considered completed. For each task that meets the criteria, record its completion status, and then assign corresponding weights to different tasks based on the task's complexity, such as planned working hours and the range of skills involved. The more complex the task, the greater its impact on the overall completion rate. Based on the completion status and weight of a single task, calculate the average overall completion rate of historical tasks through weighted aggregation to reflect the stability of the user's historical performance in similar tasks.

[0124] From the task record database of the skill growth model, tasks completed by users within the past 6 months are screened. Task type label matching and complexity clustering algorithms, such as clustering based on task steps and skills involved, are used to identify tasks of the same type and with comparable complexity, defining them as samples of the same task volume to ensure consistency in the analysis. For the selected samples, the average completion time for this type of task is calculated monthly. If the sample size for a single month is insufficient, data from adjacent months are merged to ensure statistical validity, forming a continuous monthly average time series. Based on the monthly time series, the percentage change in completion time between adjacent months is calculated, i.e., the increase or decrease in time in the following month relative to the previous month, yielding the month-on-month change rate. An outlier detection mechanism identifies and removes extreme change rates caused by special interferences (including but not limited to sudden tasks and resource constraints), and trend interpolation is used to fill in the positions of outliers, maintaining the continuity of the change rate series. The processed month-on-month change rates are then comprehensively averaged to obtain the overall trend over the past 6 months. Positive values ​​indicate increased completion time (decreased efficiency), while negative values ​​indicate decreased completion time (increased efficiency), reflecting the dynamic changes in user efficiency within the same type of task.

[0125] Based on the task-skill growth dependency graph, a GraphSAGE graph neural network is used for three-level task decomposition. Specifically, the input business task text of the target recommendation requirement is encoded into a 768-dimensional vector through the BERT model. The first layer decomposes it into the main task, the second layer into a1 subtasks, and the third layer into the operation steps corresponding to the subtasks. The output is the task step decomposition result, including the subtask execution sequence and the core skill tags required for each subtask. The TextRank algorithm is used to extract the top 3 skills and task complexity coefficients from the subtask descriptions. In this embodiment, the task complexity coefficient is obtained by quantifying the development time of similar historical tasks.

[0126] For the user skill proficiency feature vector, the TF-IDF algorithm is used to weight the skill tags to generate a weighted skill vector; the core skill tags in the task step decomposition results are converted into 300-dimensional vectors through a pre-trained Word2Vec model; the cosine similarity algorithm is used to calculate the similarity between the weighted skill vector and the core skill vector to obtain the skill-task adaptation matrix; where the dimension is the number of user skills × the number of sub-tasks; combined with the skill growth rate output by the skill growth model, the numerical change of the adaptation matrix after 3 months is predicted by linear regression to generate an adaptation change curve, which is a discrete point sequence in weeks;

[0127] The Monte Carlo simulation algorithm is used to perform multiple rounds of task simulation. In each round of simulation, based on the complexity coefficient of the subtask, the fluctuation value of the task execution difficulty is generated by normal distribution and superimposed on the complexity coefficient.

[0128] Based on the user's years of experience, the task fault tolerance parameter is adjusted through a piecewise function, where different years of experience correspond to different fault tolerance rates; the specific settings here are to be made by those skilled in the art based on historical data and industry experience.

[0129] During simulation, if the skill-task fit reaches the preset threshold, the subtask completion time is calculated based on the base time combined with the difficulty fluctuation value and the fault tolerance rate; otherwise, the time is increased by the preset value. The completion quality score is calculated by combining the fit and the experience coefficient through the softmax function, and the subtask completion time and completion quality score for each round of simulation are output.

[0130] Based on the time consumption ratio of each subtask in the historical task, that is, the ratio of the time consumption of a single subtask to the total task time, the information entropy of each subtask is quantified by the entropy weight method (information entropy reflects the degree of fluctuation of the time consumption ratio). The weight of subtasks whose information entropy meets the preset low fluctuation condition (i.e., subtasks with more stable time consumption ratio) is increased to obtain the weight coefficient of each subtask in the total task.

[0131] For each round of simulation, the completion criteria are: the completion time does not exceed the expected time by a preset percentage and the quality score is not lower than a preset threshold. The number of rounds that meet the criteria is counted, and the ratio of the number of completed rounds to the total number of simulation rounds is calculated to obtain the subtask completion rate.

[0132] Extract the average quality score of each subtask in multiple rounds of simulation, and use the subtask weight as a coefficient to perform a weighted sum of the average quality scores of each subtask to obtain the weighted quality score.

[0133] The subtask completion rate and weighted quality score are weighted and aggregated according to a preset ratio to obtain the user's simulated completion score. The completion rate reflects the stability of task execution, and the weighted quality score reflects the quality level of task execution.

[0134] Based on the matching accuracy of the user pool and historical task simulation evaluation data (including the deviation rate between historical simulation scores and actual task completion results), the historical deviation rate of a single user's simulation completion score is calculated, and a Beta distribution is used as the prior distribution. Specifically, the distribution parameters are determined based on the number of samples whose historical deviation rate meets the preset conditions, and the credibility of the single user's simulation completion score is corrected by a Bayesian estimation model. The evaluation accuracy is calculated as: 1 - the upper limit of the confidence interval of the corrected deviation rate.

[0135] The above steps are executed in parallel for users in the matching user pool using a distributed computing framework. Each user is assigned an independent computing node, and the final output is a structured result table containing a unique hash identifier for the user, a simulation completion score, and an evaluation accuracy.

[0136] It should be further explained that the process of adjusting the forward matching and repeating the reverse simulation process in this embodiment includes:

[0137] Based on the inference matrix of mismatched user information, the core gap features are extracted by matrix factorization algorithm, including high-frequency skill gap labels, average growth cycle, and gap matching probability. The correlation between the number of missing users to be matched and the high-frequency skill gap is calculated by statistical function to generate a gap-quantity correlation table.

[0138] It should be further explained that the mismatched user information inference matrix includes the skill gap field, growth cycle field, and fit probability field for users who do not meet the threshold; the high-frequency skill gap label refers to the skill deficiency type that appears frequently, such as advanced Python development or distributed system experience; the average growth cycle refers to the estimated time to fill the gap, in months; the gap fit probability refers to the probability that a user will meet the threshold after growth; the correlation between the number of missing users to be matched and the high-frequency skill gap refers to the proportion of a certain type of skill gap in the number of missing users, for example, 60% of the missing users correspond to the advanced Python development gap.

[0139] Based on the gap-quantity correlation table, the parameters of the time series analysis module of the skill growth model are optimized. Specifically, for high-frequency skill gap labels, the coefficients of the autoregressive term and the moving average term of the model are adjusted by using the gradient descent algorithm. Specifically, training weights are added to the historical proficiency change data of the gap skills to reduce the prediction error of the growth rate of the skill. For the average growth cycle field, the time decay coefficient of the skill growth rate is corrected by using a linear regression model to improve the accuracy of long-term growth prediction.

[0140] It should be further explained that the time series analysis module uses the ARIMA model; the autoregressive term coefficient is the p-value in the ARIMA model, and the moving average term coefficient is the q-value in the ARIMA model; the time decay coefficient is adjusted, for example, for skills with a growth cycle of more than 6 months, the decay coefficient of the monthly average change value is adjusted from a2 to a3 to better reflect the actual trend of long-term growth.

[0141] Based on high-frequency skill gap labels, the dynamic weight coefficients of skills in positive matching are adjusted: the skill weight distribution is recalculated using the analytic hierarchy process to increase the weight value of gap skills; for subfields in the business scenario feature vector that are associated with gap skills, attention weights are increased using an attention sparse network to strengthen the influence of such features in positive matching.

[0142] It should be further explained that the Analytic Hierarchy Process (AHP) is a mathematical method for multi-criteria decision-making; the weighting of gap skills is increased, for example, the weighting of advanced Python development; sub-fields associated with gap skills include, for example, the concurrent processing capability in the backend development scenario of e-commerce platforms.

[0143] Based on the number of missing users to be matched and the quality distribution of users who currently meet the threshold, the threshold adjustment range is determined by fitting a normal distribution. For example, if the number of missing users does not exceed 20% of the total demand, the comprehensive matching score threshold will be reduced by 5 points; if the number of missing users exceeds 20%, the threshold will be reduced in stages, first by 3 points, and if it is still not satisfied, it will be reduced by another 2 points to expand the matching range.

[0144] The adjusted skill growth model parameters, feature weights, and score thresholds are applied to the positive matching process. The user's basic feature set is re-extracted, including the updated skill proficiency feature vector and growth prediction vector. The skill matching score, experience matching score, and qualification matching score are recalculated using the cosine similarity algorithm and logistic regression model to generate a new comprehensive matching score. Users are selected based on the new threshold to form an adjusted matching user pool, and its matching accuracy is calculated.

[0145] It should be further explained that the matching accuracy rate refers to the ratio of the number of users who pass the reverse verification in the new user pool to the total number of users in the user pool.

[0146] For the adjusted matching user pool, the reverse task completion simulation evaluation process is re-executed, specifically: the edge weights of the task-skill growth dependency graph are updated using the GraphSAGE graph neural network; when using Monte Carlo simulation, the task execution difficulty fluctuation value of the gap skills is reduced by a preset proportion; the simulation completion score and evaluation accuracy are recalculated, and the evaluation result table of the adjusted users is generated.

[0147] Compare the number of users who meet the threshold condition after adjustment with the number required for the target recommendation: if they meet the threshold, output the final recommendation result; if they do not meet the threshold, repeat the above steps until the number requirement is met or the maximum number of iterations is reached; after each iteration, use analysis of variance to check the adjustment effect and ensure that the fluctuation of user quality does not exceed the preset quality threshold.

[0148] It should be further explained that the construction and training process of the skill development model in this embodiment includes:

[0149] Based on the on-chain evidence data of users to be matched in the information verification matching chain, extract the change in user skill proficiency, i.e. the difference in skill proficiency at different time points, task completion rate, i.e. the weighted average of historical task completion status, and the change in completion time of the same amount of task, i.e. the month-on-month change rate of completion time in adjacent periods, to form a model training sample set.

[0150] Based on the sample set, a skill growth model is constructed using time series analysis algorithms. Specifically, the change in technical proficiency is used as the dependent variable, and the changes in task completion rate and the time to complete the same amount of task are used as covariates. The training set and validation set are divided by a sliding window, and the parameters of the autoregressive term, difference order, and moving average term of the model are optimized by the maximum likelihood estimation method.

[0151] The model's prediction accuracy is evaluated by using root mean square error and mean absolute error. For skill types with excessive prediction errors, L1 regularization constraints are introduced to correct the model parameters, ultimately resulting in a skill growth model that can output the skill growth rate, i.e., the increase in proficiency per unit time.

[0152] Based on user data that did not meet the threshold in the reverse task completion simulation evaluation, a mismatch user information inference matrix is ​​constructed: the matrix row dimension is the mismatch user, the column dimension is the skill gap field, i.e. the missing core skill tag, the growth cycle field, i.e. the estimated time to make up for the gap, and the adaptation probability field, i.e. the probability of meeting the threshold after growth.

[0153] The matrix is ​​reduced in dimensionality using a nonnegative matrix factorization algorithm to extract high-frequency skill gap labels, namely the top N skill gaps in frequency, average growth cycle, namely the time mean of similar gaps, and gap fit probability distribution, namely the fit probability interval of different cycles.

[0154] Based on the quantity requirements of the target recommendation needs and the number of users who currently meet the threshold, the number of missing users to be matched is calculated. The correlation between this number and the high-frequency skill gap label is calculated using the Pearson correlation coefficient, and a gap-quantity correlation table is generated to record the proportion of missing users corresponding to each type of gap.

[0155] Based on the gap-quantity correlation table, the parameters of the skill growth model are optimized. Specifically, for high-frequency skill gap labels, historical proficiency change data of the skill are extracted from the information verification matching chain, and its weight in model training is increased. For example, the autoregressive coefficient of the ARIMA model is adjusted by the gradient descent algorithm, so that the model's prediction deviation of the growth rate of the skill is reduced to a preset range.

[0156] Based on the average growth cycle, the time decay coefficient of the model is corrected through linear regression. Specifically, for skills with a long growth cycle (such as a cycle > preset value), the decay rate of the change in proficiency is reduced. For example, the decay coefficient is reduced by a preset ratio from the initial value to improve the prediction accuracy of long-term growth trends.

[0157] The adjusted model is applied to the user's basic feature set to re-output the skill growth prediction vector, which is the change in proficiency and the dynamic weight coefficient of skills in the future preset period, reflecting the weight value of skill growth potential.

[0158] Based on high-frequency skill gap labels, the dynamic weight coefficients of skills in positive matching are recalculated using the analytic hierarchy process. Specifically, the weight values ​​of gap skills are increased. For example, based on the correlation between the number of missing users to be matched and high-frequency skill gaps, i.e. the proportion of a certain type of skill gap in the number of missing users, the weights of sub-fields associated with gap skills in the business scenario feature vector are increased proportionally. For example, the business scenario constraints corresponding to gap skills are increased by increasing their attention weights through an attention sparse network.

[0159] Based on the distribution of the comprehensive matching scores of the number of missing users to be matched and the current users who meet the threshold, the adjustment range of the comprehensive matching score threshold is determined by kernel density estimation: if the number of missing users is less than or equal to the preset proportion of the total demand, the threshold is reduced, and the magnitude is positively correlated with the number of missing users; if the number of missing users is greater than the preset proportion, the threshold is reduced in stages, and the magnitude of each stage is a preset step size. At the same time, the on-chain evidence storage data of the information verification matching chain is used to ensure that the pass rate of the feature vector of the user's qualification certificate verification result is greater than or equal to the benchmark value after the adjustment, so as to maintain compliance.

[0160] This embodiment achieves a fundamental transformation in talent-job matching from static and passive to dynamic, proactive, and self-evolving through a sophisticated system that deeply integrates trusted data, dynamic modeling, forward-looking prediction, and closed-loop optimization. Its core benefits are first reflected in the construction of an unprecedentedly high-credibility decision-making foundation. The engine's starting point is not the vague self-description of a traditional resume, but rather the on-chain notarization and compliance verification of user identity, qualifications, and credit records using blockchain technology. This eliminates the risk of information falsification at its source, ensuring a solid and reliable data foundation for all subsequent analysis and matching, and greatly reducing recruitment decision errors and employment risks caused by information asymmetry. Building on this, the engine demonstrates a depth of insight and dynamic evaluation capabilities that surpass current industry standards. It does not simply treat user skills as static labels, but rather uses innovative hierarchical profile maps and skill growth models to deconstruct and reshape individual abilities in a three-dimensional and temporal manner, accurately capturing and predicting users' learning abilities and growth curves. This allows the system to not only assess what users can currently do, but also predict what they can grow into in the future, thus injecting a powerful potential assessment dimension into job recommendations. Furthermore, the engine achieves a paradigm shift from passive matching to proactive simulation and prediction. Its unique reverse task completion simulation and evaluation mechanism decomposes target tasks, simulates execution environments, and introduces multiple rounds of Monte Carlo simulation to preview candidates' performance in real projects, outputting quantitative completion scores and success probabilities. This elevates recruitment from subjective guesswork based on historical experience to objective prediction based on data simulation, greatly improving the accuracy of job matching predictions and providing unprecedented decision support for project risk control. The most revolutionary effect stems from its self-aware, self-optimizing closed-loop intelligence. When the initial matching results are unsatisfactory, the system accurately diagnoses the root causes of the gaps through in-depth time causal analysis, automatically adjusting the prediction parameters of the skill model, recalibrating the weight allocation in the matching algorithm, and dynamically optimizing the screening threshold, thus initiating a new round of matching simulation. This feedback-based continuous iteration mechanism enables the system to continuously learn and improve itself during operation, with matching accuracy and success rate increasing spirally with the number of uses. In summary, this engine ensures the reliability of decision-making by laying a foundation of trustworthy data, achieves penetrating insights into talent capabilities through deep dynamic modeling, significantly reduces hiring risks and project uncertainties through forward-looking task simulation, and ultimately endows the system with the vitality of continuous evolution by building a closed-loop optimization ecosystem. This not only greatly improves the efficiency and success rate of recruitment, but also fundamentally changes the mindset of talent assessment and matching, providing enterprises with a powerful, reliable, and future-oriented intelligent infrastructure for their talent strategy.

[0161] Example 2

[0162] Please see Figure 2Another embodiment of the present invention provides: an AI intelligent recommendation engine construction system, comprising: a matching chain module, a demand parsing module, a positive matching module, a simulation evaluation module, an evaluation recommendation module, and a feedback adjustment module;

[0163] The matching chain module performs on-chain compliance verification based on the resume text information of the user to be matched, combined with the blockchain and a preset compliance verification library, to obtain an information verification matching chain; the compliance verification library contains verification rules for identity authenticity, qualification validity, and credit record compliance.

[0164] The requirement parsing module is used to respond to target recommendation requirements, perform on-chain compliance verification and target requirement text parsing on the target recommendation requirements, and evaluate the contribution of the business in the recommendation requirements based on the parsed target requirement text fields to obtain valid matching fields.

[0165] The positive matching module performs positive matching based on the effective matching fields and the information verification matching chain, combining the hierarchical profile map of the users to be matched with the skill growth model to obtain the matching user pool and the corresponding matching accuracy. The hierarchical profile map is constructed by combining the corresponding user's skill proficiency, years of experience, and qualification certificate verification results with graph algorithms. The skill growth model is constructed by combining the change in the corresponding user's corresponding technical proficiency, the change in task completion rate and the change in the time to complete the same amount of task with time series analysis algorithms, and is used to measure the growth rate of the user's corresponding skills.

[0166] The simulation evaluation module performs reverse task completion simulation evaluation based on the matching user pool and the corresponding matching accuracy, combined with the simulation algorithm and the task-skill growth dependency graph, to obtain the simulation completion score and evaluation accuracy for each user; the task-skill growth dependency graph is constructed by combining the skill growth model with the complexity of the business requirements and the reinforcement learning algorithm.

[0167] The evaluation and recommendation module presets a completion score threshold and an evaluation accuracy threshold. When the number of users to be matched that simultaneously meet both the completion score threshold and the evaluation accuracy threshold is sufficient to meet the target recommendation requirements, the matched users are directly recommended to the users with the corresponding target requirements. When the number of users to be matched that simultaneously meet both the completion score threshold and the evaluation accuracy threshold is insufficient to meet the target recommendation requirements, the users who meet both the completion score threshold and the evaluation accuracy threshold are recommended to the users with the target requirements. At the same time, time causal analysis is performed based on the simulation process information of the users to be matched that do not meet the threshold requirements to obtain the inference matrix of information on the unmatched users.

[0168] The feedback adjustment module is used to feed back the inference matrix of mismatched user information and the number of missing users to be matched into the skill growth model to adjust the positive matching, and at the same time repeat the reverse simulation process until the number of users that meet the threshold condition meets the number required for the target recommendation.

[0169] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

[0170] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. An AI intelligent recommendation engine construction method, characterized in that, The method comprises the following steps: Based on the text information of the user resume to be matched, the blockchain and the preset compliance verification library are combined for on-chain compliance verification to obtain an information verification matching chain; In response to a target recommendation demand, the target recommendation demand is subjected to on-chain compliance verification and target demand text analysis, and the contribution of the business in the recommendation demand to the field is subjected to field effectiveness evaluation based on the analyzed target demand text field to obtain effective matching fields; Based on the effective matching fields, the user to be matched in the information verification matching chain is subjected to forward matching in combination with a hierarchical portrait graph and a skill growth model to obtain a matching user pool and a corresponding matching accuracy; The hierarchical portrait graph is constructed by combining the skill proficiency, experience years and qualification certificate verification results of the corresponding user with a graph algorithm; and the skill growth model is constructed by combining the change amount of the skill proficiency of the corresponding user, the task completion degree and the change state of the time required to complete the same task amount with a time series analysis algorithm, and is used to measure the growth rate of the corresponding skill of the user.

2. The AI intelligent recommendation engine construction method of claim 1, wherein, The method for constructing the intelligent recommendation engine further comprises the following steps: Based on the matching user pool and the corresponding matching accuracy, a simulation algorithm and a task-skill growth dependency graph are combined for reverse task completion degree simulation evaluation to obtain a simulation completion degree score and an evaluation accuracy corresponding to each user; the task-skill growth dependency graph is constructed by combining the skill growth model, the demand business complexity and a reinforcement learning algorithm; A preset completion degree score threshold and an evaluation accuracy threshold are set, and when the number of users to be matched that meet the completion degree score threshold and the evaluation accuracy threshold at the same time meets the number in the target recommendation demand, the matching users are directly recommended to the corresponding target demand users; When the number of users to be matched that meet the completion degree score threshold and the evaluation accuracy threshold at the same time does not meet the number in the target recommendation demand, the users to be matched that meet the completion degree score threshold and the evaluation accuracy threshold are recommended to the target demand users, and a time causal analysis is performed on the simulation process information of the users to be matched that do not meet the threshold demand to obtain a non-matching user information reasoning matrix; The non-matching user information reasoning matrix and the missing number of users to be matched are fed back to the skill growth model to adjust the forward matching, and the reverse simulation process is repeated until the number of users that meet the threshold condition meets the number in the target recommendation demand.

3. The AI intelligent recommendation engine construction method of claim 2, wherein, The construction process of the information verification matching chain comprises the following steps: Based on the text information of the user resume to be matched and the target recommendation demand text, a pre-trained text preprocessing model is combined to obtain a user field set to be matched and a target recommendation field set; the user field set to be matched includes but is not limited to an identity field, a skill label field, a qualification certificate field, a credit record field and a work experience subfield; the target recommendation field includes enterprise qualification, recruitment authority, historical recruitment compliance subfield, business field, task type, skill requirement and compliance constraint subfield; Based on a single user field set to be matched, a principal component analysis algorithm and a principal component evaluation threshold are combined with a historical success rate as a target to obtain a single user effective field set to be matched; Obtaining a single user clustering portrait based on a single user effective field clustering cluster to be matched and a graph algorithm, and simultaneously clustering based on different user clustering portraits, with a history authenticity score of a corresponding type field as a target, to obtain a verification clustering cluster set of different user key fields, and combining a knowledge graph to obtain a hierarchical portrait graph of the user to be matched. The construction process of the information verification matching chain also includes:

4. The AI intelligent recommendation engine construction method of claim 3, wherein, Obtaining a compliance verification level of a clustering sub-cluster of different levels under a corresponding verification clustering cluster of each field type of different users based on a verification level clustering cluster of different user key fields and an expert experience algorithm; the clustering sub-cluster of different levels includes a clustering result of the same compliance verification level of the same type field of different users, and is used for compliance verification of different levels according to the authenticity score of the corresponding type field of the user; Based on the compliance verification level of the clustering sub-cluster of different levels under the corresponding verification clustering cluster of each field type of different users and the first compliance verification rule in the compliance verification library, a first verification connection is established; Based on the first verification connection, difference compliance verification is performed on the first compliance verification rule and the resume text of the user to be matched, and when all type fields of the corresponding user are verified, the user is permitted to be on-chain, and the user on-chain is marked with an on-chain authorization identifier. The construction process of the information verification matching chain also includes:

5. The AI intelligent recommendation engine construction method of claim 4, wherein, Based on the on-chain authorization identifier of the user permitted to be on-chain, the corresponding user effective field set and the compliance verification result, a SHA-256 algorithm is used to generate a unique hash value of each user verified, and the hash value, clustering identifier and verification timestamp are uploaded to the alliance chain through a PoS consensus mechanism to form a chain record containing a field-level storage node and a cluster-level storage node; The first compliance verification rule includes verification rules of identity authenticity, validity of qualification certificate fields and work experience sub-fields, and credit record compliance; Based on the target recommendation field set and the field type, a factor analysis algorithm is used to obtain the contribution of each type field to the demand business information, and a sparse network is used to preprocess the target recommendation field set based on the contribution of each type field to the demand business information as a weight, to obtain the preprocessed target recommendation field set and the field type. The construction process of the information verification matching chain also includes:

6. The AI intelligent recommendation engine construction method of claim 5, wherein, Based on the preprocessed target recommendation field set and the field type of the corresponding demand user of the target recommendation demand, and the second compliance verification rule in the compliance verification library, a second verification connection is established through the establishment process of the first verification connection; the second compliance verification rule is used for compliance verification of enterprise qualification authenticity, recruitment authority range, historical recruitment compliance sub-field score, business field and task type, skill requirement and compliance constraint sub-field of the demand user. ​ Based on the second verification connection, the target recommendation demand text is combined with the second compliance verification rule to obtain a demand user that passes the verification, and based on the target recommendation field set of the demand user, the above-mentioned process of storing the evidence of the to-be-matched user on the chain is repeated to generate a unique hash value and store the evidence on the chain. Based on the hash association relationship of the on-chain storage node, an information verification matching chain is constructed through a Merkle tree structure, specifically: taking the skill requirement clustering cluster in the target demand of different demand users as the main matching node, and the remaining clustering clusters as the auxiliary matching nodes, and associating the nodes of each user clustering portrait in the to-be-matched user hierarchical portrait graph; taking the to-be-matched user hierarchical portrait graph as the parent node, and associating the clustering cluster nodes of each field type in the to-be-matched user clustering portrait; taking the clustering cluster node as the child node, and associating the field nodes of a single to-be-matched user in the cluster, each node containing a pre-node hash, an identity information unique hash value identifier, a clustering identifier, and a verification result digest.

7. The AI intelligent recommendation engine construction method of claim 6, wherein, The process of obtaining the matching user pool and the corresponding matching accuracy rate includes: Based on the node information of the to-be-matched user hierarchical portrait graph in the information verification matching chain, the skill proficiency feature vector, the experience limit feature value, and the qualification certificate verification result feature vector are extracted from the node of each user clustering portrait to obtain a user basic feature set; Based on the skill proficiency feature vector in the user basic feature set, a skill proficiency trend prediction is performed through a time series analysis algorithm to obtain a skill growth prediction vector in a future preset period, and a skill dynamic weight coefficient is generated in combination with the historical skill growth rate of the user; Based on the preprocessed target recommendation field set and field type, a word embedding model is used to convert the skill requirement sub-field into a demand skill feature vector and the business domain and task type sub-field into a business scenario feature vector to obtain a demand feature set.

8. The AI intelligent recommendation engine construction method of claim 7, wherein, The process of obtaining the matching user pool and the corresponding matching accuracy rate also includes: Based on the association relationship of the Merkle tree structure in the information verification matching chain, the target skill weight distribution of the demand user in the skill requirement clustering cluster as the main matching node is extracted through a graph traversal algorithm, and the business scenario constraint parameter in the auxiliary matching node is extracted; Based on the user basic feature set and the demand feature set, a first similarity between the skill proficiency feature vector and the demand skill feature vector and a user business scenario adaptation degree are calculated through a cosine similarity algorithm, the first similarity is weighted and corrected in combination with the skill dynamic weight coefficient to obtain a skill matching score; Based on the experience limit feature value and the qualification certificate verification result feature vector of the to-be-matched user and the experience requirement threshold and the qualification compliance threshold in the demand feature set corresponding to the demand user, a comparison is performed through a logistic regression model to output an experience matching score and an identifier and a qualification matching score and an identifier; Based on the on-chain storage verification result of the to-be-matched user in the information verification matching chain, a user that passes the verification is selected as a candidate set, the skill matching score, the experience matching score, and the qualification matching score of the user in the candidate set are weighted and summed to obtain a comprehensive matching score for corresponding identification. A preset comprehensive matching score threshold is set, users with a comprehensive matching score greater than the comprehensive matching score threshold are included in a matching user pool, and a matching accuracy is obtained according to the number of subsequent reverse verification passed users in the matching user pool / the total number of matching users in the matching user pool.

9. An AI intelligent recommendation engine construction system for implementing the AI intelligent recommendation engine construction method of any one of claims 1-8. Comprise: A matching chain module, a demand analysis module, and a forward matching module; The matching chain module performs on-chain compliance verification based on the resume text information of the user to be matched in combination with a preset compliance verification library, and obtains an information verification matching chain; The compliance verification library includes verification rules for identity authenticity, qualification validity, and credit record compliance; The demand analysis module is configured to respond to a target recommendation demand, perform on-chain compliance verification and target demand text analysis on the target recommendation demand, evaluate the contribution of the business in the recommendation demand based on the analyzed target demand text field, and obtain an effective matching field; The forward matching module performs forward matching based on the effective matching field in combination with the user-to-be-matched user hierarchical portrait graph and skill growth model in the information verification matching chain, and obtains a matching user pool and a corresponding matching accuracy; The hierarchical portrait graph is constructed by combining the skill proficiency, experience years, and qualification verification results of the corresponding user with a graph algorithm; and the skill growth model is constructed by combining the change amount of the skill proficiency of the corresponding user, the task completion degree, and the change state of the time required to complete the same amount of tasks with a time series analysis algorithm, and is used to measure the growth rate of the corresponding skill of the user.

10. The AI intelligent recommendation engine construction system of claim 9, wherein, The intelligent recommendation engine construction system further comprises a simulation evaluation module, an evaluation recommendation module, and a feedback adjustment module; The simulation evaluation module performs reverse task completion degree simulation evaluation based on the matching user pool and the corresponding matching accuracy in combination with a simulation algorithm and a task-skill growth dependency graph, and obtains a simulation completion degree score and an evaluation accuracy corresponding to each user; the task-skill growth dependency graph is constructed by combining the skill growth model, the demand business complexity, and a reinforcement learning algorithm; The evaluation recommendation module presets a completion degree score threshold and an evaluation accuracy threshold; when the number of users to be matched that meet the completion degree score threshold and the evaluation accuracy threshold meets the number in the target recommendation demand, the matching users are directly recommended to the corresponding target demand users; when the number of users to be matched that meet the completion degree score threshold and the evaluation accuracy threshold does not meet the number in the target recommendation demand, the users that meet the completion degree score threshold and the evaluation accuracy threshold are recommended to the target demand users, and a time causal analysis is performed on the simulation process information of the users to be matched that do not meet the threshold demand to obtain a non-matching user information reasoning matrix; The feedback adjustment module is configured to feed back the non-matching user information reasoning matrix and the missing number of users to be matched to the skill growth model to adjust the forward matching, and repeatedly perform the reverse simulation process until the number of users that meet the threshold condition meets the number in the target recommendation demand.

Images