A method and system for managing portrait information based on vector indexing

Abstract

A method and system for managing personnel profile information based on vector indexing is disclosed. The method first converts project manager profiles into high-dimensional vectors and constructs a vector index, with data consistency, fairness, and qualification compliance as multi-objective constraints. Next, real-time project requirements are converted into query vectors, and the vector index is used to match a set of candidate project managers. Finally, the MOPSO algorithm is used to search for feasible solutions, and non-dominated solutions are filtered using Pareto dominance relations, iteratively optimizing the optimal solution. The output is an interpretable result containing identifiers, matching scores, and constraint satisfaction status and reasons. This invention eliminates cross-system data conflicts through multi-feature comparison and source priority entity disambiguation mechanisms, ensuring the accuracy and timeliness of core information. Furthermore, relying on a multi-objective constraint set, optimization algorithms avoid irrelevant biases, ensuring that qualification review matching conforms to rules and regulatory fairness, thus comprehensively realizing the accuracy, fairness, and standardization of qualification information management.

Description

Technical Field

[0001] This invention relates to information management methods, belonging to the field of construction management technology, and particularly to a method and system for managing portrait information based on vector indexing. Background Technology

[0002] In the field of construction management, the personnel profiling information management method is gradually being implemented on the owner's side for internal evaluation and optimization of the project manager team. Under the premise of adhering to the current professional qualification system, this method integrates multi-dimensional data such as the project manager's work experience, certificate level, training records and performance evaluation to build a calculable and updatable digital profile, thereby assisting in core decisions such as project allocation, talent selection and resource allocation.

[0003] Currently, project manager profiling management still faces two core challenges: First, relevant information is scattered across multiple departments or systems, such as human resources, certification, training, and project management, resulting in inconsistent data definitions, heterogeneous formats, and delayed updates, making it difficult to achieve real-time synchronization and traceable management of cross-source data. Second, if the matching process relies on rule engines or simple machine learning models, it is easily affected by historical data distribution and feature selection, introducing implicit biases from irrelevant factors such as gender, age, and past project types, interfering with the objectivity of the evaluation results. Furthermore, existing solutions lack the ability to adapt to complex constraints such as qualification validity periods, geographical and professional categories, workload, risk levels, and construction windows, potentially leading to inaccurate matching results or delayed responses, thereby increasing management risks.

[0004] Some existing technologies introduce AI-based data fusion and fairness constraint mechanisms to address the above issues. However, in the process of engineering implementation, they still rely to some extent on preset rules or offline trained static models. They lack the ability to adapt to real-time business scenarios that are multi-source, heterogeneous, constantly changing, and highly concurrent. They are unable to stably achieve dynamic adjustments under complex business constraints and cannot balance consistency and fairness, resulting in inaccurate and unfair real-time matching of project manager qualifications. Summary of the Invention

[0005] The purpose of this invention is to overcome the aforementioned defects and problems in the prior art and to provide a method and system for managing personal profile information based on vector indexing, so as to accurately and fairly manage the information on project manager qualifications.

[0006] To achieve the above objectives, the technical solution of this invention is: a method for managing portrait information based on vector indexing, comprising:

[0007] Obtain the project manager's profile data, perform entity disambiguation and multi-feature comparison on the profile data, convert it into a high-dimensional vector, and build a vector index that supports incremental updates.

[0008] A multi-objective constraint set is constructed, with data consistency, fairness, and eligibility compliance as constraints.

[0009] Acquire real-time project requirement data and convert it into query vectors. Calculate the similarity between the query vectors and their high-dimensional vector representations based on the vector index to obtain a set of candidate project managers.

[0010] A multi-objective vector is constructed based on a set of candidate project managers as input. Feasible solutions are searched using the MOPSO algorithm. Non-dominated solutions are selected from the feasible solutions using Pareto dominance relations, and the external archive is updated. The optimal solution is iteratively optimized based on the external archive, and the output includes interpretable information such as project manager identification, matching score, satisfaction status of the multi-objective constraint set, and reasons for non-satisfaction.

[0011] Obtain the project manager's profile data, and perform entity disambiguation and multi-feature comparison on the profile data, specifically including:

[0012] Normalize the names, certificate numbers, employment relationships, and project participation records in the character profile data to build an existing entity database;

[0013] Based on the existing entity database, multi-feature comparison is performed sequentially through the candidate recall stage, hierarchical comparison stage, and decision stage.

[0014] During the candidate recall phase, when a new record is received from the project manager, the certificate number is used as the key to search the existing entity database first.

[0015] If a matching certificate number exists, the corresponding entity will be selected as the candidate entity.

[0016] If the certificate number is missing or not found, the search will be performed using the combination key of name and employment relationship, and records with overlapping employment relationships within a certain time window will be limited to candidate entities.

[0017] If no candidate entity is found, the project participation record will be used as a clue to recall entities with the same name and similar role type that have appeared in the same project as candidate entities.

[0018] During the hierarchical comparison phase, the new record is compared with each candidate entity item by item based on four features: name, certificate number, employment relationship, and project participation record. A consistency level is assigned to each feature. The consistency levels include strong consistency, general consistency, weak consistency, and inconsistency.

[0019] The consistency levels are combined into an overall matching level according to preset rules; the overall matching level includes three categories: high confidence that it is the same person, doubtful, and obviously different.

[0020] During the decision-making stage, if a candidate entity and a new record have an overall match level of high confidence that they are the same person, and there are no other candidate entities that reach the same level, then the new record will be merged under the name of that candidate entity.

[0021] If multiple candidate entities have an overall matching level of high confidence for the same person, they will not be merged for the time being, and the new record and related entities will be marked as awaiting manual review.

[0022] If the overall matching level of all candidate entities is significantly different, then create a new project manager entity for the new record and establish a corresponding primary key in the existing entity database;

[0023] For candidate entities whose overall matching level is questionable, a conservative manual approach is adopted, and manual review is requested.

[0024] The construction of a vector index that supports incremental updates specifically includes:

[0025] During the vector index initialization phase, all project managers' high-dimensional vectors are randomly or in descending order to serve as entry nodes for subsequent insertions and queries.

[0026] In the hierarchical graph construction phase, for each project manager's high-dimensional vector to be inserted, a highest level is determined for each high-dimensional vector to be inserted according to a preset probability distribution, and a hierarchical and navigable small world graph structure is constructed.

[0027] Starting from the highest-level entry node of the current vector index, perform a nearest neighbor search at each level:

[0028] Starting from the current node, traverse the neighboring nodes and compare the distance between each neighboring node and the high-dimensional vector to be inserted;

[0029] If a closer neighbor node exists, jump to that neighbor node and continue performing the nearest neighbor search until no closer node can be found in the current layer. Then, use that node as the entry node for the next layer and continue performing the nearest neighbor search downwards.

[0030] After traversing the nearest neighbor search of all layers, in each layer that the high-dimensional vector has access to, select several nearest neighbor nodes from the neighbor nodes visited during the nearest neighbor search as adjacency points, establish bidirectional connections between the high-dimensional vector to be inserted and the adjacency points, and delete redundant connections that are too far away from the connection list of the adjacency points.

[0031] During the incremental update phase, when new or changed profile data is added, a high-dimensional vector of the project manager is regenerated and inserted into the existing vector index according to the hierarchical mapping process. For updates to existing project managers, after locating the corresponding node in the vector index, the corresponding node is deleted and inserted as a new node, and the adjacency relationship is re-established. Alternatively, the node identifier is retained and the high-dimensional vector representation of the node is directly updated, and the adjacency relationship is re-established.

[0032] The step of calculating the similarity between the query vector and the high-dimensional vector representation based on the vector index to obtain a set of candidate project managers specifically includes:

[0033] For each candidate project manager and the query vector, an initial similarity is constructed based on the feature domain. Then, based on the initial similarity, the feature domains are quantized and truncated to [0,1] before aggregation to obtain a standardized similarity, the expression of which is as follows:

[0034] ;

[0035] in: For the feature domain For candidate project managers Standardized similarity; For the feature domain The original similarity; , respectively feature domain The lower and upper quantile values; For feature domain indexing; Index for candidate project managers; To truncate the input to the interval [0,1];

[0036] Based on source reliability, data coverage, and offline verification scores, non-negative raw weights are determined and normalized, specifically including:

[0037] ;

[0038] in: For the feature domain Normalized weights; For the feature domain The non-negative original weights; The number of feature domains participating in the aggregation; For summation index;

[0039] Based on standardized similarity and non-negative original weights, candidate scores are calculated, and their expression is as follows:

[0040] ;

[0041] in: Candidate Project Manager Candidate scores;

[0042] A dynamic threshold is determined based on the quantile threshold method, and shortlisted candidate scores are selected to obtain a set of candidate project managers; the expression for the dynamic threshold is as follows:

[0043] ;

[0044] in: For dynamic thresholds; This is the parameter for the proportion of shortlisted candidates; The set of all candidate scores; This is the set of candidate score indices for shortlisted candidates; This is a quantile function.

[0045] The non-negative original weights are determined based on source reliability, data coverage, and offline verification scores, and their expression is as follows:

[0046] Different data sources are classified into several reliability levels and mapped to... Within the interval, obtain the source reliability score;

[0047] Calculate the ratio between the number of project managers with valid values ​​and the total number of candidate project managers, or the ratio between the number of projects with valid values ​​and the total number of projects, and statistically analyze the coverage distribution and map it to... Within the interval, obtain the data coverage score;

[0048] A validation dataset was constructed based on historical project manager and project matching results. The hit rate or ranking correlation coefficient of the output results was evaluated after comparing matching calculations using feature domains alone or preferentially, and mapped to... Within the range, obtain the offline verification score;

[0049] Based on the source reliability score, data coverage score, and offline verification score, the non-negative original weight is calculated, and its expression is as follows:

[0050] ;

[0051] in: , , respectively feature domain Source reliability score, data coverage score, and offline verification score; , All are adjustment coefficients; , , These are the thresholds for source reliability score, data coverage score, and offline verification score, respectively. This is the upper limit of the truncation of the original weights.

[0052] The methods for obtaining the interpretable information specifically include:

[0053] Using the set of candidate project managers as the selection vector, and transforming the candidate scores and multi-objective constraint set into a multi-objective vector, the expression is as follows:

[0054] ;

[0055] in: It is a multi-objective vector; To select a vector; Candidate Project Manager The choice of variables; Candidate Project Manager Data consistency satisfaction; A group set for evaluating fairness; , In selecting vectors lower group with the group The inclusion ratio function; Candidate Project Manager The degree of compliance with eligibility requirements;

[0056] Using a multi-objective vector as input, feasible solutions are searched based on the MOPSO algorithm. The quality of feasible solutions is judged by Pareto dominance relations, and non-dominated solutions are identified and updated to the external archive. The expression is as follows:

[0057] ;

[0058] in: , All are feasible solutions; It is a Pareto dominance relationship; Index of the target vector; For the first One target vector;

[0059] The reference point method with weighted scalarization based on external archives is used to calculate the optimal solution for the current round, and its expression is as follows:

[0060] ;

[0061] in: This is the optimal solution for the current round; For external archives; For the first During the first iteration The weights of each target vector; The first interval normalized The values ​​of the target vectors;

[0062] The joint stopping rule is based on the supervolume increment of the external archive and the iteration upper limit. The expression for determining whether the optimal solution in the current round satisfies the stopping rule is as follows:

[0063] ;

[0064] in: For the first The super-volume increase of the wheel; For the first The overvolume of the external archives of the wheel at a fixed reference point; Incremental threshold; The number of stagnation rounds to tolerate; For iteration rounds; The upper limit of iterations;

[0065] like continuous wheel or If the iteration stops, the optimal solution of the current round is taken as the final solution; the final solution includes the project manager identifier, matching score, and explanatory information on the satisfaction and non-satisfaction of the multi-objective constraint set.

[0066] Conversely, if the concept drift signal is introduced to update the weights of the target vector, the optimal solution for the current round is repeatedly calculated; the weight update expression for the target vector is as follows:

[0067] ;

[0068] in: For the first During the first iteration The weights of each target vector; An operator that truncates each target vector into non-negative values ​​and normalizes them to a sum of 1; Update the step size for weights; For the first Business benchmark priority for each target vector; The drift response coefficient; Indicates the first The data distribution drift intensity index corresponding to each target vector.

[0069] The aforementioned multi-objective constraint set, constrained by data consistency, fairness, and eligibility compliance, specifically includes:

[0070] Data consistency constraints: Verify whether the validity period and update timestamp of the portrait data meet the time validity window required by the project, and select the valid record based on source priority and version number when data conflicts occur;

[0071] Fairness constraints: Without using protected attributes as the direct basis for decision-making, impose fairness index constraints at the group or individual level on the candidate project manager set, and re-rank the candidate project manager set if the constraints are not met.

[0072] Qualification and compliance constraints: Before generating matching results, the validity status of the certificates required for the project manager, the applicable profession, and the completion status of necessary training are verified in real time, and candidates that do not meet the mandatory compliance requirements are eliminated.

[0073] A vector-indexed portrait information management system, which applies the above-mentioned method, includes:

[0074] The vector index building module is used to obtain the project manager's profile data, convert the profile data into a high-dimensional vector after entity disambiguation and multi-feature comparison, and build a vector index that supports incremental updates.

[0075] The constraint building module is used to construct a multi-objective constraint set with data consistency, fairness, and eligibility compliance as constraints.

[0076] The candidate set acquisition module is used to acquire real-time project requirement data and convert it into query vectors, and calculate the similarity between the query vectors and the high-dimensional vector representations based on the vector index to obtain a set of candidate project managers;

[0077] The interpretable output module is used to construct a multi-objective vector based on the candidate project manager set as input, search for feasible solutions based on the MOPSO algorithm, filter non-dominated solutions from the feasible solutions using Pareto dominance relations and update the external archive, iteratively optimize the optimal solution based on the external archive, and output interpretable information including project manager identification, matching score, satisfaction status of multi-objective constraint set and reasons for non-satisfaction.

[0078] A vector-indexed personal portrait information management device includes:

[0079] The device includes a processor and a memory;

[0080] The memory is used to store computer program code and to transmit the computer program code to the processor;

[0081] The processor is used to execute the above-described vector-index-based portrait information management method according to the instructions in the computer program code.

[0082] A computer-readable storage medium storing computer-executable instructions, which, when executed on a computer, implement the aforementioned vector-index-based method for managing portrait information.

[0083] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0084] This invention discloses a method and system for managing project manager profiles based on vector indexes. The method first converts project manager profiles into high-dimensional vectors and constructs a vector index, using data consistency, fairness, and qualification compliance as multi-objective constraints. Next, real-time project requirements are converted into query vectors, and the vector index is used to match candidate project manager sets. Finally, the MOPSO algorithm is used to search for feasible solutions, and non-dominated solutions are filtered using Pareto dominance relations to iteratively optimize the optimal solution. An interpretable result containing identifiers, matching scores, and constraint satisfaction status and reasons is output. In application, this design effectively eliminates cross-system data conflicts through multi-feature comparison and source priority entity disambiguation mechanisms, ensuring that project managers... The accuracy and timeliness of core information such as work experience and certification levels lay a solid data foundation for qualification management. Furthermore, relying on a multi-objective constraint set encompassing data consistency, fairness, and qualification compliance, the algorithm optimizes algorithms to avoid irrelevant biases such as gender and age, ensuring that qualification review and matching strictly adhere to business rules and regulatory requirements, fully demonstrating management fairness. Simultaneously, the second-level similarity retrieval supported by vector indexes significantly improves the efficiency of qualification information matching, while the interpretable records of the output results facilitate audit supervision and provide support for model iteration, further strengthening the accuracy and compliance of qualification information management. This comprehensively achieves the accuracy, fairness, and standardization of project manager qualification information management. Attached Figure Description

[0085] Figure 1 This is a flowchart of the method of the present invention.

[0086] Figure 2 This is a system structure diagram of the present invention.

[0087] Figure 3 This is a structural diagram of the device of the present invention.

[0088] In the diagram: Index building module 1, constraint building module 2, candidate set acquisition module 3, interpretable output module 4, processor 5, memory 6, computer program code 61. Detailed Implementation

[0089] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0090] Example 1:

[0091] See Figure 1 A method for managing portrait information based on vector indexing, comprising:

[0092] Obtain the project manager's profile data, perform entity disambiguation and multi-feature comparison on the profile data, convert it into a high-dimensional vector, and build a vector index that supports incremental updates.

[0093] Furthermore, obtain the project manager's profile data, and perform entity disambiguation and multi-feature comparison on the profile data, specifically including:

[0094] Normalize the names, certificate numbers, employment relationships, and project participation records in the character profile data to build an existing entity database;

[0095] Based on the existing entity database, multi-feature comparison is performed sequentially through the candidate recall stage, hierarchical comparison stage, and decision stage.

[0096] During the candidate recall phase, when a new record is received from the project manager, the certificate number is used as the key to search the existing entity database first.

[0097] If a matching certificate number exists, the corresponding entity will be selected as the candidate entity.

[0098] If the certificate number is missing or not found, the search will be performed using the combination key of name and employment relationship, and records with overlapping employment relationships within a certain time window will be limited to candidate entities.

[0099] If no candidate entity is found, the project participation record will be used as a clue to recall entities with the same name and similar role type that have appeared in the same project as candidate entities.

[0100] During the hierarchical comparison phase, the new record is compared with each candidate entity item by item based on four features: name, certificate number, employment relationship, and project participation record. A consistency level is assigned to each feature. The consistency levels include strong consistency, general consistency, weak consistency, and inconsistency.

[0101] The consistency levels are combined into an overall matching level according to preset rules; the overall matching level includes three categories: high confidence that it is the same person, doubtful, and obviously different.

[0102] During the decision-making stage, if a candidate entity and a new record have an overall match level of high confidence that they are the same person, and there are no other candidate entities that reach the same level, then the new record will be merged under the name of that candidate entity.

[0103] If multiple candidate entities have an overall matching level of high confidence for the same person, they will not be merged for the time being, and the new record and related entities will be marked as awaiting manual review.

[0104] If the overall matching level of all candidate entities is significantly different, then create a new project manager entity for the new record and establish a corresponding primary key in the existing entity database;

[0105] For candidate entities whose overall matching level is questionable, a conservative manual approach is adopted, and manual review is requested.

[0106] In this embodiment, the entity disambiguation process specifically includes: identifying the same project manager entity across systems based on multi-feature comparison of name, certificate number, employment relationship and project participation record, and merging them according to preset source priority and timestamp when there is a conflict.

[0107] Specifically, source priority is set in the order of issuing authority or authoritative directory, internal data, historical snapshots, and temporary submissions; time priority is based on the updated timestamp and those that pass validity verification; if names are adjacent, strict identical matching after unified coding is used by default; if alias fields exist, normalized comparison is enabled; if certificate numbers are adjacent, strict identical matching is used; if employment relationship and project participation records appear in the same organizational domain and the same time window and are adjacent, the time window is set to 90 days by default (adjustable between 30 and 180 days), based on the verification results of mismatch rate and omission rate; if multiple sources conflict and cannot be decided by priority and time order, records with more covered fields and that pass verification are retained first, and records that are not adopted are kept as backups with the reasons for not being adopted to avoid repeated merging later.

[0108] Entity disambiguation multi-feature comparison is performed in the order of "candidate recall - hierarchical comparison - decision merging / splitting". First, fields such as name, certificate number, employment relationship and project participation record are preprocessed. Names are unified to the same encoding format, spaces and special characters are removed from certificate numbers, unit names in employment relationships are normalized according to standard organizational directories, and project names are unified according to project master data, so as to reduce the impact of writing differences between different systems on the comparison results.

[0109] During the candidate recall phase, when a new project manager record is received, the existing entity database is searched first using the certificate number as the keyword: if a completely identical certificate number exists, all corresponding entities are added to the candidate set; if the certificate number is missing or not found, the search is performed using the combination key "name + employer", and records with overlapping employment relationships within a certain time window are limited to candidate entities; if no candidate entity is found, the project participation records are used as clues to recall entities with the same name and similar role types that have appeared in the same project, in order to supplement any potentially missed matching objects.

[0110] During the hierarchical comparison phase, each new record is compared with every candidate entity based on four categories of features: name, certificate number, employment relationship, and project participation record. A consistency level is assigned to each feature category. For example, completely identical certificate numbers are categorized as "strongly consistent," minor format differences within the same issuing authority are categorized as "generally consistent," different numbers but belonging to the same certificate type are categorized as "weakly consistent," and clearly conflicting numbers are categorized as "inconsistent." Similarly, completely identical names after removing common aliases and differences between simplified and traditional characters are categorized as "strongly consistent," names that can only be matched through an alias lookup table are categorized as "generally consistent," names with similar spellings but not identifiable in the lookup table are categorized as "weakly consistent," and names that are clearly different are categorized as "inconsistent." Employment relationships and project participation records are assigned similar consistency levels based on factors such as whether they belong to the same organizational domain, whether time windows overlap, whether projects have the same name, and whether role types match. The consistency levels of these four feature categories are then combined according to preset rules to obtain the overall matching level of the candidate pair, for example, categorized into "high confidence in the same person," "questionable," and "clearly different."

[0111] During the decision-making phase, if a candidate entity has an overall match level of "high confidence same person" with a new record, and no other candidate entities reach the same level, the new record is automatically merged under that entity. During merging, data fields from multiple sources are merged according to preset source priority and update timestamp order, prioritizing fields from high-priority sources with the most recent update time and those passing validity and format checks. Conflicting values ​​are stored as backup fields, and the reason for rejection and the original source are recorded. If multiple candidate entities all reach "high confidence same person," it is assumed by default that there may be historical merging errors. The new record and related entities are marked as requiring manual review and are not automatically merged to avoid erroneous aggregation. If the overall match level of all candidate entities is "significantly different," a new project manager entity is created for the new record, and a corresponding primary key is established in the profile database. For candidate entities with a matching level of "questionable," a conservative strategy can be selected based on business configuration: one strategy is to temporarily treat the new record as an independent entity and record potential associations in the background for manual querying; another strategy is to not merge them during profile calculation, and only automatically merge or submit for manual confirmation when a higher confidence level is reached later. Through the above candidate recall, hierarchical comparison, and decision processing flow, the entity disambiguation step in this embodiment has a clear logical path of "how to compare first, how to merge if the comparison is successful, and how to handle uncertain or erroneous comparisons," which facilitates engineering implementation and audit verification.

[0112] Furthermore, a vector index supporting incremental updates is constructed, specifically including:

[0113] During the vector index initialization phase, all project managers' high-dimensional vectors are randomly or in descending order to serve as entry nodes for subsequent insertions and queries.

[0114] In the hierarchical graph construction phase, for each project manager's high-dimensional vector to be inserted, a highest level is determined for each high-dimensional vector to be inserted according to a preset probability distribution, and a hierarchical and navigable small world graph structure is constructed.

[0115] Starting from the highest-level entry node of the current vector index, perform a nearest neighbor search at each level:

[0116] Starting from the current node, traverse the neighboring nodes and compare the distance between each neighboring node and the high-dimensional vector to be inserted;

[0117] If a closer neighbor node exists, jump to that neighbor node and continue performing the nearest neighbor search until no closer node can be found in the current layer. Then, use that node as the entry node for the next layer and continue performing the nearest neighbor search downwards.

[0118] After traversing the nearest neighbor search of all layers, in each layer that the high-dimensional vector has access to, select several nearest neighbor nodes from the neighbor nodes visited during the nearest neighbor search as adjacency points, establish bidirectional connections between the high-dimensional vector to be inserted and the adjacency points, and delete redundant connections that are too far away from the connection list of the adjacency points.

[0119] During the incremental update phase, when new or changed profile data is added, a high-dimensional vector of the project manager is regenerated and inserted into the existing vector index according to the hierarchical mapping process. For updates to existing project managers, after locating the corresponding node in the vector index, the corresponding node is deleted and inserted as a new node, and the adjacency relationship is re-established. Alternatively, the node identifier is retained and the high-dimensional vector representation of the node is directly updated, and the adjacency relationship is re-established.

[0120] In this embodiment, the portrait data includes two types: text-based and data-based. The text-based portrait data is used to generate semantic vectors through a pre-trained language model. The data is standardized, and the semantic vectors and numerical vectors are aligned and fused to form a unified high-dimensional vector representation.

[0121] Specifically, text elements are vectorized using a pre-trained semantic model for general Chinese scenarios. Numerical elements are mapped to the same dimension after de-extinction, standardization, and missing value imputation. The vector dimension is 768 by default (it can be determined within the range of 300 to 1024 based on the trade-off between offline retrieval recall and latency). Text-side word segmentation uses a unified vocabulary and a fixed maximum sequence length of 256, with any exceeding 256 being truncated. Numerical missing values ​​are filled with the median of the past three months in the same domain, and outliers are truncated at 1% and 99%. When text elements are missing but numerical elements are complete, only numerical sub-vectors are output, and vice versa. In the fusion stage, the dimensions are aligned by concatenation and linear transformation to ensure that the output dimension is fixed.

[0122] In this embodiment, the vector index adopts a hierarchical navigable small world graph structure, and performs incremental insertion and index reconnection when adding or changing character portrait data, so as to maintain the stability of retrieval recall and query latency.

[0123] The construction of a vector index includes three stages: index initialization, hierarchical graph building, and incremental updates. In the index initialization stage, all high-dimensional vectors of project managers are arranged according to preset rules, such as random order or descending order of update time, and the first record is selected from them to generate the entry node of the index, which is used as the starting point for subsequent insertions and queries.

[0124] In the hierarchical graph construction phase, for each project manager vector to be inserted, a top-level is first determined according to a preset probability distribution, ensuring a small number of nodes are distributed at higher levels and the majority at lower levels, forming a hierarchical, navigable small-world graph structure. Then, starting from the highest-level entry node of the current index, a nearest neighbor search is performed in each level: traversing all neighboring nodes from the current node, comparing the distances between each neighboring node and the vector to be inserted, and jumping to that neighboring node when a closer neighbor exists, continuing the search until no closer node is found in the current level. This node is then used as the entry node for the next level, and the search continues downwards. After completing the search across all levels, in each level accessible to the vector, several nearest nodes are selected from those visited during the search as adjacency points. Bidirectional connections are established between the vector to be inserted and these adjacency points, and redundant connections that are too far apart are deleted from the adjacency point connection list to ensure that the number of connections for each node does not exceed a preset limit, thus balancing the sparsity and connectivity of the graph structure.

[0125] During the incremental update phase, when new or modified project manager profile data is added, a unified high-dimensional vector representation of the project manager is first regenerated, and then inserted into the existing index according to the hierarchical graph construction process described above. For updates to existing project manager profiles, after locating the corresponding node in the index, adjustments can be made in one of two ways: one is to treat it as deleting the old node and inserting a new node, re-establishing the relevant adjacency relationships; the other is to directly update the node's vector representation while retaining the node identifier, and re-perform nearest neighbor search and edge addition / deletion operations within the node's local neighborhood to correct unreasonable connection relationships caused by vector changes. Through the above index initialization, hierarchical graph construction, and incremental update process, this embodiment can maintain a stable level of recall and query latency for similarity retrieval without frequently rebuilding the entire index.

[0126] Specifically, the maximum number of connections per node in the vector index graph is 16 to 32 by default, the candidate neighborhood size during the construction phase is 100 to 200 by default, and the candidate neighborhood size during the query phase is 50 to 200 by default. The values ​​are based on the recall of the offline benchmark set and the 90th percentile query latency. Incremental updates are executed in batches every one to five minutes. When the change ratio exceeds one-tenth or a significant decrease in recall is detected, local reconnection is triggered. When it exceeds one-twentieth and lasts for three cycles, partition reconstruction is triggered. When memory resources are insufficient, the query candidate neighborhood size is reduced to no less than 20 and reconnection is temporarily suspended.

[0127] In this embodiment, the personnel profile data refers to traceable data records directly related to the project manager's qualifications, abilities, and resume; the project requirement data refers to the structured description of project-side requirements and constraints related to job matching; the high-dimensional vector representation refers to a fixed-length vector after alignment of textual semantics and numerical elements; the vector index refers to the approximate nearest neighbor structure established for the above vectors; the multi-objective constraint set refers to the set of rules that must be simultaneously satisfied or optimized as much as possible in the matching stage for data consistency, fairness, and qualification compliance; the interpretable information includes the project manager identifier, aggregate score, satisfaction markers of the three types of constraints, and main reasons for non-satisfaction. The score is retained to three decimal places, and the first 20 items are returned by default, which can be adjusted within the range of 5 to 50 items, based on the accuracy of offline verification and the workload of decision-making; the minimum executable caliber for missing tests is that at least one of the text domain or numerical domain exists and any key field in the certificate or training is present. If none of these are met, the candidate is not included and the reason is recorded to avoid misleading decisions.

[0128] A multi-objective constraint set is constructed, with data consistency, fairness, and eligibility compliance as constraints.

[0129] Furthermore, the multi-objective constraint set specifically includes:

[0130] Data consistency constraints: Verify whether the validity period and update timestamp of the portrait data meet the time validity window required by the project, and select the valid record based on source priority and version number when data conflicts occur;

[0131] Specifically, the time validity window is associated with key project milestones, and by default, data updated within 90 days before the planned start date is considered valid. It can be adjusted by binding it to the project type within 30 to 180 days. The validity of qualification certificates is verified based on the effective date and expiration date fields, combined with the suspension status mark. Training records are verified based on the completion date and qualification mark. If the timestamp is missing, the record with the newest version number and at least one time field is temporarily valid, and it will be automatically replaced after the complete timestamp is successfully obtained for the first time.

[0132] Fairness constraints: Without using protected attributes (such as gender or age) as the direct basis for decision-making, impose fairness index constraints at the group or individual level on the candidate project manager set, and re-rank the candidate project manager set if the constraints are not met.

[0133] Specifically, group fairness is primarily measured by the difference in the proportion of qualified applicants, with a default tolerance threshold of 5%, which can be adjusted within the range of 2% to 10% based on business and compliance requirements. Individual fairness can be referenced through nearest neighbor consistency, with nearest neighbors determined by locally similar samples in the vector space and a minimum sample size of no less than five. When the protected attribute records are incomplete or the sample size is insufficient to support stable estimation, only eligibility compliance and data consistency constraints are applied, and the fairness monitoring results are retained as a reference, without triggering rearrangement to avoid over-correction.

[0134] Qualification and compliance constraints: Before generating matching results, the validity status of the certificates required for the project manager, the applicable profession, and the completion status of necessary training are verified in real time, and candidates that do not meet the mandatory compliance requirements are eliminated.

[0135] Specifically, the qualification compliance check covers certificate status, applicable profession, issuing institution, and effective period; the training verification covers course list, completion status, and validity period. By default, the certificate is valid throughout the project start and end time and is not suspended or revoked, and the training is completed within the last 24 months and has not expired. The verification refresh cycle is once a day by default, and can be temporarily increased to once every four hours during the batch personnel and job scheduling period. When the external source is temporarily unavailable, the most recent valid verification result is used and marked as pending review, and will be automatically reviewed after it is restored.

[0136] Acquire real-time project requirement data and convert it into query vectors. Calculate the similarity between the query vectors and their high-dimensional vector representations based on the vector index to obtain a set of candidate project managers.

[0137] Furthermore, the step of calculating the similarity between the query vector and the high-dimensional vector representation based on the vector index to obtain a set of candidate project managers specifically includes:

[0138] For each candidate project manager, construct the original similarity based on the feature domain between the query vector and each candidate project manager;

[0139] To align the dimensions and distributions of different domains, the feature domains are aggregated after quantile calibration and truncation to [0,1] based on the original similarity, to obtain the standardized similarity, which is expressed as follows:

[0140] ;

[0141] in: For the feature domain For candidate project managers Standardized similarity; For the feature domain The original similarity; , respectively feature domain The lower and upper quantile values, when hour, This means that comparisons are possible between different domains; For feature domain indexing; Index for candidate project managers; To truncate the input to the interval [0,1];

[0142] Based on source reliability, data coverage, and offline verification scores, non-negative raw weights are determined and normalized, as expressed below:

[0143] ;

[0144] in: For the feature domain Normalized weights; For the feature domain The original weights; The number of feature domains participating in the aggregation; For summation index; when and This makes it easier for subsequent scoring to fall under a uniform standard.

[0145] Based on standardized similarity and non-negative original weights, candidate scores are calculated for ranking and filtering, and the expression is as follows:

[0146] ;

[0147] in: Candidate Project Manager Candidate scores; if each and ,but ;

[0148] To balance scale stability and noise suppression, a dynamic threshold is determined based on the quantile threshold method to screen shortlisted candidate scores, resulting in a set of candidate project managers. The threshold is recorded for auditing and adaptive testing. The expression for the dynamic threshold is as follows:

[0149] ;

[0150] in: For dynamic thresholds; This is the parameter for the proportion of shortlisted candidates; The set of all candidate scores; This is the set of candidate score indices for shortlisted candidates; This is a quantile function.

[0151] when When fixed, the number of shortlisted entries adapts to the score distribution; if the business requires an equal number of shortlisted entries, it can be changed to be based on... Sort by first One, at this time replace As a parameter for controlling the number of shortlisted entries.

[0152] Specifically, the standardization of quantile calibration takes into account the distribution differences and robustness of different feature domains, and limits the impact of outliers to the cutoff interval; the weighting process uses a non-negative normalized structure to carry source reliability, coverage and verification signals, which can support heuristic setting and facilitate offline learning iteration; the aggregation adopts a linear model to improve interpretability and audit friendliness, so that the final score is on a uniform scale and can work with subsequent consistency, fairness and compliance constraints in subsequent stages; the threshold selection obtains an adaptive inclusion scale that changes with the distribution through the quantile method, while retaining an equal number of Top-N as an alternative strategy.

[0153] In this embodiment, the default settings for the feature domains include a text domain, a numerical domain, a certificate and training domain, and a relational domain. The text domain uses cosine similarity, the numerical domain is mapped from distance to similarity, the certificate / training domain uses set similarity, and the relational domain uses path / co-occurrence index similarity mapping. The upper and lower boundaries of the quantile calibration are defaulted to 5% and 95%, respectively, and can be adjusted within the range of 2% to 99%. The adjustment is based on the on-the-job matching accuracy and bad case explanation complaint rate during the stable period. The weights are initially equalized and jointly calibrated using source reliability, coverage ratio, and offline cross-validation scores. The update cycle is daily to weekly, and the coverage ratio is calculated using a sliding window of approximately 30. If a domain is missing in a single candidate, its weight is proportionally allocated to the remaining available domains in the score aggregation of that candidate. If multiple domains are missing simultaneously and only one domain remains, its candidate score is determined solely by the available domain score, and the missing domain is marked in the interpretable information.

[0154] Furthermore, the determination of non-negative original weights based on source reliability, data coverage, and offline verification scores specifically includes:

[0155] Different data sources are classified into several reliability levels and mapped to... Within the interval, obtain the source reliability score;

[0156] Calculate the ratio between the number of project managers with valid values ​​and the total number of candidate project managers, or the ratio between the number of projects with valid values ​​and the total number of projects, and statistically analyze the coverage distribution and map it to... Within the interval, obtain the data coverage score;

[0157] A validation dataset was constructed based on historical project manager and project matching results. The hit rate or ranking correlation coefficient of the output results was evaluated after comparing matching calculations using feature domains alone or preferentially, and mapped to... Within the range, obtain the offline verification score;

[0158] Based on the source reliability score, data coverage score, and offline verification score, the non-negative original weight is calculated, and its expression is as follows:

[0159] ;

[0160] in: , , respectively feature domain Source reliability score, data coverage score, and offline verification score; , All are adjustment coefficients; , , These are the thresholds for source reliability score, data coverage score, and offline verification score, respectively. This is the upper limit of the truncation of the original weights.

[0161] In this embodiment, the acquisition of non-negative original weights includes three parts: source reliability score, data coverage score, and offline verification score. The source reliability score is used to reflect the credibility of the data source for the corresponding feature domain. Different data sources are pre-classified into several levels. For example, competent authorities or authoritative industry platforms are marked as the highest level, internal master data is marked as the intermediate level, and temporary reporting or manual import channels are marked as the general level. The source level is periodically adjusted in combination with indicators such as historical conflict rate and manual error correction records. Each piece of data automatically inherits its source level when it is aggregated and is converted into a source reliability score in the range of 0 to 1.

[0162] Data coverage score reflects the completeness of a specific feature domain within the current search scope. It can be calculated based on the ratio of project managers with valid values ​​to the total number of candidates, or the ratio of projects with valid values ​​to the total number of projects, to determine the coverage of that feature domain and compress the results to a range of 0 to 1. During the statistical process, factors such as field missing rate and outlier percentage can be considered, and lower coverage scores can be assigned to feature domains with severe missing values ​​or excessively high outlier percentages.

[0163] Offline validation scores are used to measure the contribution of the feature domain to historical matching results. Specifically, a validation dataset is constructed based on historical project manager-project matching results. Hit rate, ranking correlation coefficient, and other metrics are evaluated, using only or preferentially using the feature domain. The evaluation results are then standardized and mapped to the 0-1 range to obtain the corresponding offline validation score. This validation process can be performed monthly or quarterly to reflect the actual support capability of the feature domain for matching results at different times.

[0164] After obtaining the source reliability score, data coverage score, and offline verification score, non-negative original weights for domain d are formed using a pre-defined monotonic combination rule. For example, the source reliability score can be used as the base weight, with the data coverage score and offline verification score acting as amplification or reduction factors. This ensures that when a certain score is low, the original weight of that feature domain decreases accordingly; when any score is below the minimum threshold, the original weight of that feature domain can be directly set to zero, eliminating its influence from the aggregation process. To avoid an abnormally large weight for any feature domain, the combined original weights can be further truncated and smoothed. Finally, the original weights of all feature domains participating in the aggregation are normalized by summing them to obtain normalized weights for candidate score aggregation. This ensures that the weights are non-negative and sum to 1, while also making the normalized weights numerically reflect the combined influence of source reliability, data coverage, and offline verification scores.

[0165] A multi-objective vector is constructed based on the candidate project manager set as input, and feasible solutions are searched based on the MOPSO algorithm. Non-dominated solutions are selected from the feasible solutions using Pareto dominance relations and the external file is updated. Based on the external file, iterative optimization is performed, and the output includes interpretable information containing project manager identification, matching score, satisfaction status of multi-objective constraint set and reasons for non-satisfaction.

[0166] Using the set of candidate project managers as the selection vector, components are used to represent whether a project is shortlisted, and feasibility constraints such as size limit, number of positions, or cost can be added.

[0167] The candidate scores, along with data consistency constraints, fairness constraints, and eligibility compliance constraints, are transformed into a multi-objective vector for the MOPSO algorithm search. The expression for this vector is as follows:

[0168] ;

[0169] in: It is a multi-objective vector; To select a vector; Candidate Project Manager The choice of variables; Candidate Project Manager Data consistency satisfaction; A group set for evaluating fairness; , In selecting vectors lower group with the group The inclusion ratio function; Candidate Project Manager The degree of compliance with eligibility requirements;

[0170] Using a multi-objective vector as input, feasible solutions are searched based on the MOPSO algorithm. The quality of feasible solutions is judged by Pareto dominance relations, and non-dominated solutions are identified and updated to the external archive. The expression is as follows:

[0171] ;

[0172] in: , All are feasible solutions; It is a Pareto dominance relationship; Index of the target vector; For the first One target vector;

[0173] The reference point method with weighted scalarization based on external archives is used to calculate the optimal solution for the current round, and its expression is as follows:

[0174] ;

[0175] in: This is the optimal solution for the current round; For external archives; For the first During the first iteration The weights of each target vector; The first interval normalized The values ​​of the target vectors;

[0176] The joint stopping rule is based on the supervolume increment of the external archive and the iteration upper limit. The expression for determining whether the optimal solution in the current round satisfies the stopping rule is as follows:

[0177] ;

[0178] in: For the first The super-volume increase of the wheel; For the first The overvolume of the external archives of the wheel at a fixed reference point; Incremental threshold; The number of stagnation rounds to tolerate; For iteration rounds; The upper limit of iterations;

[0179] like continuous wheel or If the iteration stops, the optimal solution of the current round is taken as the final solution; the final solution includes the project manager identifier, matching score, and explanatory information on the satisfaction and non-satisfaction of the multi-objective constraint set.

[0180] Conversely, if the concept drift signal is introduced to update the weights of the target vector, the optimal solution for the current round is repeatedly calculated; the weight update expression for the target vector is as follows:

[0181] ;

[0182] in: For the first During the first iteration The weights of each target vector; An operator that truncates each target vector into non-negative values ​​and normalizes them to a sum of 1; Update the step size for weights; For the first Business benchmark priority for each target vector; The drift response coefficient; Indicates the first The data distribution drift intensity index corresponding to each target vector.

[0183] In this embodiment, the iteration time is... Weight of each objective The acquisition and updating of [objectives] includes three stages: weight initialization, concept drift monitoring, and smoothing update. In the weight initialization stage, the business side provides the basic priority of each objective, which serves as [the basis for the update]. The values ​​for these priorities can be varied. For example, the basic priority of the eligibility compliance and data consistency objectives can be set slightly higher than that of the similarity objective, while the basic priority of the fairness objective can be set no lower than the minimum level required by regulations. These basic priorities are then normalized and used as the initial weight vector. It is saved as a configuration and serves as the starting point for subsequent weight updates.

[0184] In the concept drift monitoring phase, the average performance and distribution of the actually adopted solutions across the four objectives are statistically analyzed daily or weekly within the most recent time window, and compared with historical baseline distributions. By comparing the mean, variance, or quantiles of the current period with those of the baseline period, the distribution drift intensity for each objective can be obtained. The drift intensity is scaled to a range of 0 to 1, and the drift direction is used to determine whether the importance of the objective needs to be increased or decreased. For example, when there is a significant increase in fairness-related issues in recent audits or user feedback, the drift intensity corresponding to the fairness objective will be adjusted. A larger positive value will be taken, and the weight of the fairness objective will be increased in subsequent weight updates accordingly. When an objective performs stably and meets business expectations over a long period of time, its drift intensity is close to zero, and the corresponding weight is mainly affected by the basic priority.

[0185] In the smooth update phase, before each round of optimization begins, the weights of the current target vector are used as the basis for the update. Business benchmark priority and drift strength Calculate new candidate weights. Intuitively, the new weights should fall between the previous round's weights and the base priority plus a drift correction term, and the weight update step size should be adjusted accordingly. Adjusting the degree of compromise between the two, A larger value indicates greater sensitivity to recent concept drift. After obtaining the candidate weight values, a normalization operator is applied to them. First, all negative components are truncated to zero. Then, the weight vector is scaled according to the sum of its components to make the updated weight vector... Each component is non-negative and the sum of the four target weights is 1. This smooth update mechanism avoids drastic weight oscillations between iterations and gradually reflects changes in data distribution and business focus.

[0186] In the final solution selection phase, the set of non-dominated feasible solutions maintained in the external archive and the weight vector obtained in the current iteration are used. Calculate the scalarized evaluation value. Specifically, for each candidate solution in the archive... First, the four target values ​​are normalized to fall on a uniform numerical scale. Then, the four targets are weighted and summed according to their current weights to obtain the comprehensive evaluation score of the candidate solution. Subsequently, the comprehensive evaluation scores of all candidate solutions are sorted, and the solution with the highest score is selected as the final output solution of this round of multi-objective optimization. The corresponding weight vector, target values, and satisfaction of key constraints are recorded for subsequent auditing and parameter adjustment. Through the above-mentioned weight initialization, concept drift monitoring, and smooth update mechanism, the selection of the optimal solution in this embodiment reflects the preset business priority and can adaptively respond to changes in data distribution and regulatory requirements over time.

[0187] Specifically, this scheme uses the selection vector as the decision variable, converges matching strength, data consistency, fairness, and eligibility compliance into four objectives, and employs a non-dominated frontier to maintain the diversity of feasible solutions. The optimal solution in the current round does not directly take a certain frontier point, but instead uses a reference point method with weighted scaling based on external archives to take into account multiple business focuses and ensure auditability. Furthermore, a concept drift signal is introduced when updating weights, and the step size and response coefficient are adjusted to a new equilibrium, so that the selection rules can adapt to changes in data distribution and regulatory orientation. In terms of convergence, a dual criterion of supervolume increment and iteration upper limit is used to avoid long-term search during the phase of slowing returns.

[0188] For example, the selection vector is represented in binary form to indicate whether a solution is included or not. The number of particles is 50 to 100 by default, the iteration limit is 300 to 500 by default, the individual guidance coefficient and the group guidance coefficient are each between 1 and 2 by default, the inertia factor is 0.6 to 0.9 by default, the external file size is 50 to 150 by default, and the reference point is set according to the uniform grid of the target space. The concept drift intensity is obtained by comparing the distribution difference between the current period and the reference period and linearly scaled to between 0 and 1. The weight update step size is 0.1 to 0.3 by default. When there are feasibility constraints such as the number of positions or cost, infeasible solutions are directly repaired to the nearest feasible boundary or discarded during the update process. If there is insufficient computing resources, the number of particles and the iteration limit are reduced first, and the external file size is kept at least 50 to maintain the diversity of the solution set.

[0189] The final solution includes project manager identification, matching score, explanations of whether the main constraints are met and the reasons for non-met requirements, and records key parameters of the matching process for auditing and subsequent model parameter updates;

[0190] For example, the interpretable information in the output includes the top three feature domains with the highest score contribution and their percentages, the types and reasons for unmet constraints, the differentiating points of similar candidates and the inclusion threshold (retained by default for 60 to 180 days for auditing); when the output is large, it is returned in pages and includes the timestamp and index version number of this search; if an anomaly occurs, the minimum necessary fields are returned and the reason is marked to ensure basic availability when resources are limited or external dependencies are interrupted.

[0191] Example 2:

[0192] See Figure 2 A vector-indexed portrait information management system, the system comprising:

[0193] Index building module 1 is used to obtain the project manager's profile data, convert the profile data into a high-dimensional vector after entity disambiguation and multi-feature comparison, and build a vector index that supports incremental updates.

[0194] Constraint Construction Module 2 is used to construct a multi-objective constraint set with data consistency, fairness, and eligibility compliance as constraints;

[0195] Candidate set acquisition module 3 is used to acquire real-time project requirement data and convert it into query vectors, and calculate the similarity between the query vectors and the high-dimensional vector representations based on the vector index to obtain a candidate project manager set;

[0196] The interpretable output module 4 is used to construct a multi-objective vector based on the candidate project manager set as input, search for feasible solutions based on the MOPSO algorithm, filter non-dominated solutions from the feasible solutions using Pareto dominance relations and update the external archive, iteratively optimize the optimal solution based on the external archive, and output interpretable information including project manager identification, matching score, satisfaction status of multi-objective constraint set and reasons for non-satisfaction.

[0197] Furthermore, for the specific implementation steps of the index construction module 1, constraint construction module 2, candidate set acquisition module 3, and interpretable output module 4, please refer to the corresponding description in Embodiment 1, which will not be repeated here.

[0198] Example 3:

[0199] See Figure 3 A vector-index-based human portrait information management device, the device including a processor 5 and a memory 6;

[0200] The memory 6 is used to store computer program code 61 and transmit the computer program code 61 to the processor 5;

[0201] The processor 5 is used to execute the vector index-based portrait information management method described in Embodiment 1 according to the instructions in the computer program code 61.

[0202] This embodiment also includes a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed on a computer, the vector index-based portrait information management method described in Embodiment 1 is implemented.

[0203] Generally, the computer instructions for implementing the method of the present invention can be carried on any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media can include any computer-readable medium except for the signal itself, which is temporarily propagating.

[0204] Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EKROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0205] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smarttalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. In particular, Python, suitable for neural network computation, and platform frameworks such as TensorFlow and PyTorch can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer or to an external computer (e.g., via the Internet using an Internet service provider) through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0206] The aforementioned equipment and non-transitory computer-readable storage media can be found in the detailed description of a vector-index-based method for managing portrait information and its beneficial effects, which will not be repeated here.

[0207] Although embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for managing portrait information based on vector indexing, characterized in that, include: Obtain the project manager's profile data, perform entity disambiguation and multi-feature comparison on the profile data, convert it into a high-dimensional vector, and build a vector index that supports incremental updates. A multi-objective constraint set is constructed, with data consistency, fairness, and eligibility compliance as constraints. Acquire real-time project requirement data and convert it into query vectors. Calculate the similarity between the query vectors and their high-dimensional vector representations based on the vector index to obtain a set of candidate project managers. A multi-objective vector is constructed based on the candidate project manager set as input, and feasible solutions are searched based on the MOPSO algorithm. Non-dominated solutions are selected from feasible solutions based on Pareto dominance relations and the external archive is updated; the optimal solution is iteratively optimized based on the external archive, and the output includes interpretable information including project manager identification, matching score, satisfaction status of multi-objective constraint set and reasons for non-satisfaction.

2. The method for managing portrait information based on vector indexing according to claim 1, characterized in that: Obtain the project manager's profile data, and perform entity disambiguation and multi-feature comparison on the profile data, specifically including: Normalize the names, certificate numbers, employment relationships, and project participation records in the character profile data to build an existing entity database; Based on the existing entity database, multi-feature comparison is performed sequentially through the candidate recall stage, hierarchical comparison stage, and decision stage. During the candidate recall phase, when a new record is received from the project manager, the certificate number is used as the key to search the existing entity database first. If a matching certificate number exists, the corresponding entity will be selected as the candidate entity. If the certificate number is missing or not found, the search will be performed using the combination key of name and employment relationship, and records with overlapping employment relationships within a certain time window will be limited to candidate entities. If no candidate entity is found, the project participation record will be used as a clue to recall entities with the same name and similar role type that have appeared in the same project as candidate entities. During the hierarchical comparison phase, the new record is compared with each candidate entity item by item based on four features: name, certificate number, employment relationship, and project participation record. A consistency level is assigned to each feature. The consistency levels include strong consistency, general consistency, weak consistency, and inconsistency. The consistency levels are combined into an overall matching level according to preset rules; the overall matching level includes three categories: high confidence that it is the same person, doubtful, and obviously different. During the decision-making stage, if a candidate entity and a new record have an overall match level of high confidence that they are the same person, and there are no other candidate entities that reach the same level, then the new record will be merged under the name of that candidate entity. If multiple candidate entities have an overall matching level of high confidence for the same person, they will not be merged for the time being, and the new record and related entities will be marked as awaiting manual review. If the overall matching level of all candidate entities is significantly different, then create a new project manager entity for the new record and establish a corresponding primary key in the existing entity database; For candidate entities whose overall matching level is questionable, a conservative manual approach is adopted, and manual review is requested.

3. The method for managing portrait information based on vector indexing according to claim 1, characterized in that: The construction of a vector index that supports incremental updates specifically includes: During the vector index initialization phase, all project managers' high-dimensional vectors are randomly or in descending order to serve as entry nodes for subsequent insertions and queries. In the hierarchical graph construction phase, for each project manager's high-dimensional vector to be inserted, a highest level is determined for each high-dimensional vector to be inserted according to a preset probability distribution, and a hierarchical and navigable small world graph structure is constructed. Starting from the highest-level entry node of the current vector index, perform a nearest neighbor search at each level: Starting from the current node, traverse the neighboring nodes and compare the distance between each neighboring node and the high-dimensional vector to be inserted; If a closer neighbor node exists, jump to that neighbor node and continue performing the nearest neighbor search until no closer node can be found in the current layer. Then, use that node as the entry node for the next layer and continue performing the nearest neighbor search downwards. After traversing the nearest neighbor search of all layers, in each layer that the high-dimensional vector has access to, select several nearest neighbor nodes from the neighbor nodes visited during the nearest neighbor search as adjacency points, establish bidirectional connections between the high-dimensional vector to be inserted and the adjacency points, and delete redundant connections that are too far away from the connection list of the adjacency points. During the incremental update phase, when new or changed profile data is added, a high-dimensional vector of the project manager is regenerated and inserted into the existing vector index according to the hierarchical mapping process. For updates to existing project managers, after locating the corresponding node in the vector index, the corresponding node is deleted and inserted as a new node, and the adjacency relationship is re-established. Alternatively, the node identifier is retained and the high-dimensional vector representation of the node is directly updated, and the adjacency relationship is re-established.

4. The method for managing portrait information based on vector indexing according to claim 1, characterized in that: The step of calculating the similarity between the query vector and the high-dimensional vector representation based on the vector index to obtain a set of candidate project managers specifically includes: For each candidate project manager and the query vector, an initial similarity is constructed based on the feature domain. Then, based on the initial similarity, the feature domains are quantized and truncated to [0,1] before aggregation to obtain a standardized similarity, the expression of which is as follows: ； in: For the feature domain For candidate project managers Standardized similarity; For the feature domain The original similarity; , respectively feature domain The lower and upper quantile values; For feature domain indexing; Index for candidate project managers; To truncate the input to the interval [0,1]; Based on source reliability, data coverage, and offline verification scores, non-negative raw weights are determined and normalized, as expressed below: ； in: For the feature domain Normalized weights; For the feature domain The non-negative original weights; The number of feature domains participating in the aggregation; For summation index; Based on standardized similarity and non-negative original weights, candidate scores are calculated, and their expression is as follows: ； in: Candidate Project Manager Candidate scores; A dynamic threshold is determined based on the quantile threshold method, and shortlisted candidate scores are selected to obtain a set of candidate project managers; the expression for the dynamic threshold is as follows: ； in: For dynamic thresholds; This is the parameter for the proportion of shortlisted candidates; The set of all candidate scores; This is the set of candidate score indices for shortlisted candidates; This is a quantile function.

5. The method for managing portrait information based on vector indexing according to claim 4, characterized in that: The determination of non-negative original weights based on source reliability, data coverage, and offline verification scores specifically includes: Different data sources are classified into several reliability levels and mapped to... Within the interval, obtain the source reliability score; Calculate the ratio between the number of project managers with valid values ​​and the total number of candidate project managers, or the ratio between the number of projects with valid values ​​and the total number of projects, and statistically analyze the coverage distribution and map it to... Within the interval, obtain the data coverage score; A validation dataset was constructed based on historical project manager and project matching results. The hit rate or ranking correlation coefficient of the output results was evaluated after comparing matching calculations using feature domains alone or preferentially, and mapped to... Within the range, obtain the offline verification score; Based on the source reliability score, data coverage score, and offline verification score, the non-negative original weight is calculated, and its expression is as follows: ； in: , , respectively feature domain Source reliability score, data coverage score, and offline verification score; , All are adjustment coefficients; , , These are the thresholds for source reliability score, data coverage score, and offline verification score, respectively. This is the upper limit of the truncation of the original weights.

6. The method for managing portrait information based on vector indexing according to claim 1, characterized in that: The methods for obtaining the interpretable information specifically include: Using the set of candidate project managers as the selection vector, and transforming the candidate scores and multi-objective constraint set into a multi-objective vector, the expression is as follows: ； in: It is a multi-objective vector; To select a vector; Candidate Project Manager The choice of variables; Candidate Project Manager Data consistency satisfaction; A group set for evaluating fairness; , In selecting vectors lower group with the group The inclusion ratio function; Candidate Project Manager The degree of compliance with eligibility requirements; Using a multi-objective vector as input, feasible solutions are searched based on the MOPSO algorithm. The quality of feasible solutions is judged by Pareto dominance relations, and non-dominated solutions are identified and updated to the external archive. The expression is as follows: ； in: , All are feasible solutions; It is a Pareto dominance relationship; Index of the target vector; For the first One target vector; The reference point method with weighted scalarization based on external archives is used to calculate the optimal solution for the current round, and its expression is as follows: ； in: This is the optimal solution for the current round; For external archives; For the first During the first iteration The weights of each target vector; The first interval normalized The values ​​of the target vectors; The joint stopping rule is based on the supervolume increment of the external archive and the iteration upper limit. The expression for determining whether the optimal solution in the current round satisfies the stopping rule is as follows: ； in: For the first The super-volume increase of the wheel; For the first The overvolume of the external archives of the wheel at a fixed reference point; Incremental threshold; The number of stagnation rounds to tolerate; For iteration rounds; The upper limit of iterations; like continuous wheel or If the iteration stops, the optimal solution of the current round is taken as the final solution; the final solution includes the project manager identifier, matching score, and explanatory information on the satisfaction and non-satisfaction of the multi-objective constraint set. Conversely, if the concept drift signal is introduced to update the weights of the target vector, the optimal solution for the current round is repeatedly calculated; the weight update expression for the target vector is as follows: ； in: For the first During the first iteration The weights of each target vector; An operator that truncates each target vector into non-negative values ​​and normalizes them to a sum of 1; Update the step size for weights; For the first Business benchmark priority for each target vector; This refers to the drift response coefficient; Indicates the first The data distribution drift intensity index corresponding to each target vector.

7. The method for managing portrait information based on vector indexing according to claim 1, characterized in that: The aforementioned multi-objective constraint set, constrained by data consistency, fairness, and eligibility compliance, specifically includes: Data consistency constraints: Verify whether the validity period and update timestamp of the portrait data meet the time validity window required by the project, and select the valid record based on source priority and version number when data conflicts occur; Fairness constraints: Without using protected attributes as the direct basis for decision-making, impose fairness index constraints at the group or individual level on the candidate project manager set, and re-rank the candidate project manager set if the constraints are not met. Qualification and compliance constraints: Before generating matching results, the validity status of the certificates required for the project manager, the applicable profession, and the completion status of necessary training are verified in real time, and candidates that do not meet the mandatory compliance requirements are eliminated.

8. A portrait information management system based on vector indexing, characterized in that, The system is applied to the method according to any one of claims 1-7, the system comprising: The vector index building module (1) is used to obtain the project manager's profile data, convert the profile data into a high-dimensional vector after entity disambiguation and multi-feature comparison, and build a vector index that supports incremental updates. The constraint construction module (2) is used to construct a multi-objective constraint set with data consistency, fairness and eligibility compliance as constraints; The candidate set acquisition module (3) is used to acquire real-time project requirement data and convert it into query vectors, and calculate the similarity between the query vectors and the high-dimensional vector representations based on the vector index to obtain a candidate project manager set; The interpretable output module (4) is used to construct a multi-objective vector based on the candidate project manager set as input, search for feasible solutions based on the MOPSO algorithm, filter non-dominated solutions from feasible solutions and update the external archive based on the Pareto dominance relation, iteratively optimize the optimal solution based on the external archive, and output interpretable information including project manager identification, matching score, satisfaction status of multi-objective constraint set and reasons for non-satisfaction.

9. A vector-indexed image information management device, characterized in that, include: The device includes a processor (5) and a memory (6); The memory (6) is used to store computer program code (61) and to transmit the computer program code (61) to the processor (5). The processor (5) is used to execute the vector index-based portrait information management method according to any one of the instructions in the computer program code (61).

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed on a computer, implement the vector-index-based portrait information management method according to any one of claims 1-7.

Images