GNN-based talent evaluation model and method and talent selection method

Abstract

The invention relates to the technical field of artificial intelligence and talent management, in particular to a talent evaluation model and method based on GNN and a talent selection method. The talent evaluation model comprises a dimension prediction model for training each attribute weight; the talent score is the weighted summation of each attribute and the attribute weight; the training process of the dimension prediction model comprises the following steps: acquiring full-attribute data as a training set; acquiring full-attribute data marked with expert scores as a verification set; the dimension prediction model is used for deleting a target attribute from the training set and speculating the target attribute and the weight thereof in combination with the residual attributes; traversing all attributes through the target attribute to obtain the weight of each attribute; on the verification set, the attribute scores of the talents and the corresponding weights are multiplied and summed to serve as predicted talent weights, and the attribute weights are verified through comparison of the talent weights and the expert weights; and training of the dimension prediction model and verification of the attribute weight are cycled until the attribute weight reaches a convergence condition. According to the method, the feature column needing to be learned is masked by using the cyclic mask method, and the masked column is learned by using other feature columns, so that the problems of insufficient data and non-uniform data can be efficiently solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and talent management technology, and in particular to a talent assessment model, method and talent selection method based on GNN. Background Art

[0002] With the rapid development of society and the intensification of technological competition, high-level scientific and technological talent has become a critical resource for the development of all industries. How to scientifically, impartially, and efficiently evaluate scientific and technological talent is not only directly related to an organization's talent selection and development strategies, but also has important implications for industrial innovation and the enhancement of technological competitiveness. To address this need, numerous research and practical efforts have sought to develop more scientific and effective talent evaluation methods.

[0003] Current technologies mainly focus on two aspects: evaluation models based on multiple indicators and evaluation models based on academic networks. However, both have certain limitations:

[0004] (1) The traditional multi-index weighted summation method relies on expert surveys to determine the weights of each indicator, which is highly subjective and leads to inconsistency and bias in the results. At the same time, this method has poor versatility and is difficult to adapt to the needs of different fields or specific talent evaluations.

[0005] (2) There is a lack of authoritative academic collaborator network data sets, and the results obtained from multiple data sources are inconsistent; and the data update cycle is long; this makes the basis for talent evaluation often lagging and difficult to fully reflect the actual situation of the scientific research environment.

[0006] (3) Existing evaluation models often use a fixed indicator system for scoring, making it difficult to flexibly adjust according to the specific skill requirements of talents. The talent requirements for different fields and positions vary greatly, and fixed indicators cannot cover all dimensions, limiting the accuracy and comprehensiveness of talent evaluation. Summary of the Invention

[0007] In order to overcome the defects of the talent evaluation model in the above-mentioned prior art, the present invention proposes a training method for a talent evaluation model based on GNN, which provides automatically optimized weight distribution and an efficient training mechanism, and can achieve more scientific and comprehensive talent evaluation.

[0008] The present invention proposes a training method for a talent assessment model based on a GNN. The talent assessment model includes a dimensional prediction model for training the weights of each attribute. The talent score is the weighted sum of each attribute and the attribute weight. The training of the dimensional prediction model includes the following steps:

[0009] The training set consists of all attribute data and the validation set of all attribute data labeled with expert ratings. The dimensional prediction model is used to remove a target attribute from the training set and infer the target attribute and its weight based on the remaining attributes. The target attribute is used to traverse all attributes to obtain the weight of each attribute. In the validation set, the attribute score of the talent is multiplied by the corresponding weight and the sum is used as the predicted talent weight. The attribute weight is then verified by comparing the talent weight with the expert weight.

[0010] The training of the dimensional prediction model and the verification of the attribute weights are repeated until the attribute weights reach the convergence condition.

[0011] Preferably, the dimensional prediction model includes a graph neural network and a dimensional prediction network; the graph neural network converts attribute data into graph structure data, and the dimensional prediction network contains nodes corresponding to each attribute one by one; the weight of each node in the dimensional prediction network is updated during the training of the dimensional prediction model; after the training is completed, the node weight of the dimensional prediction network is output as the attribute weight.

[0012] Preferably, the convergence condition of the attribute weights is that the cross entropy loss of the talent ratings and expert ratings calculated on the validation samples converges.

[0013] Preferably, the talent assessment model also includes an attribute completion model for completing missing attribute data; the attribute completion model includes a front-to-back connected graph neural network and a completion network, the graph neural network converts attribute data into graph structure data, and the completion network performs attribute completion on the graph structure data.

[0014] Preferably, the training method of the attribute completion model is: first obtain all attribute data and randomly delete some attributes to construct a training set; then extract training samples from the training set and input them into the attribute completion model and calculate the loss function, and update the completion network according to the loss function; and repeat the training until the attribute completion model converges.

[0015] Preferably, the process of training the attribute completion model includes the following steps:

[0016] First, training samples are selected from the training set, and the attribute completion model is made to learn the training samples; weak supervision loss is calculated on the training samples, and the completion network is updated according to the weak supervision loss;

[0017] Extract validation samples from the training set, let the attribute completion model complete the missing attribute data in the validation samples, and calculate the difference between the attribute data after the missing attribute data is completed and the corresponding full attribute data;

[0018] Determine whether the calculated difference has converged; if not, return to the training process of the attribute completion model; if so, fix the completion network and output the attribute completion model.

[0019] Preferably, the difference between the attribute data after the missing attribute data is completed and the corresponding full attribute data is represented by: the cross entropy loss or mean square error loss of the completed value and the original value of each attribute; or the distance between the completed attribute data and the corresponding full attribute data.

[0020] A talent assessment method proposed in the present invention first adopts the training method of the GNN-based talent assessment model to obtain a talent assessment model; the talent assessment model includes an attribute completion model and the weight of each attribute; then, the attribute data of the talent to be assessed is obtained, and it is determined whether the attribute data is complete; if not, the attribute data is completed using the attribute completion model, and the completed attribute data and the weight of each attribute are weighted and summed to obtain a talent score; if the attribute data of the talent is complete, the attribute data and the weight of each attribute are directly weighted and summed to obtain a talent score.

[0021] Preferably, talent attributes include: basic attributes and academic literacy; indicators of basic attributes include: highest academic degree, professional title / position, title, institution of employment, length of service, number of publications, reputation of publications, and talent cultivation; indicators of academic literacy include: number of papers, citations of papers, h&i10 indicators, paper value, academic cooperation, academic activity, academic breadth, and number of funds.

[0022] The present invention proposes a talent selection method, which first determines the target field and the talent's score on basic attributes; then determines the talent's academic literacy score in the target field; constructs the talent's attribute data based on the basic attribute score and the academic literacy score in the target field; then executes the talent evaluation method to obtain the talent scores of the candidate talents; and selects the required talent from the candidate talents according to the talent scores in descending order.

[0023] The advantages of the present invention are:

[0024] (1) The present invention proposes a training method for a talent assessment model based on a GNN. This method automatically learns and adjusts the weights of various evaluation indicators through a graph neural network (GNN) model, avoiding the weight allocation bias caused by expert subjectivity in traditional methods. By using a training method that traverses and deletes target attributes one by one and infers their weights based on the remaining attributes, the present invention can accurately capture the implicit correlations between attributes, avoiding the subjective bias of manual weighting while ensuring the objectivity and reliability of the weights, thereby improving the adaptability and accuracy of the model.

[0025] (2) The present invention masks the feature columns that need to be learned using a cyclic masking method, and uses other feature columns to learn the masked columns, which can effectively solve the problems of insufficient data and uneven data.

[0026] (3) The model of comprehensive talent evaluation based on graph neural network proposed in this invention uses node-to-edge and edge-to-node message passing mechanisms, combined with multi-layer perceptron (MLP), to iteratively represent and diffuse the features of nodes and edges, thereby achieving a comprehensive representation of graph structure data. When using this model for talent evaluation, low-dimensional indicators are first converted into feature vectors; then, the constructed feature vectors are input into the trained model and, after network transformation, capture the complex relationships between features. Finally, the trained model processes the input feature vectors and outputs the comprehensive score of the talent as the evaluation result, which comprehensively reflects the comprehensive ability of the talent under various indicators.

[0027] (4) This paper uses GNN to transform attribute data into a graph structure, effectively modeling nonlinear attribute relationships through information transfer between nodes. This approach is particularly suitable for learning to represent complex talent characteristics such as educational background and skill relationships. The corresponding design of the dimension prediction network and attribute nodes enables interpretability of weights, facilitating the analysis of key influencing factors.

[0028] (5) The attribute completion model uses a random deletion-reconstruction training strategy to give the model robust processing capabilities for incomplete data. In practical applications, it can automatically repair missing attributes, lower the threshold for data collection, reduce the amount of data and the requirements for data real-time performance, and solve the problems of insufficient academic collaboration network data and long update cycles in existing models. The use of an unsupervised training scheme based on the minimum conflict principle effectively solves the problems of insufficient data and uneven sample distribution, allowing the model to be fully trained even with less sample data, and ensuring high accuracy and generalization ability.

[0029] (6) The completion process uses dual optimization of weak supervision loss and verification difference to ensure that the completion results are consistent with the data distribution law and aligned with the expert annotation standards; thus ensuring the reliability of talent evaluation based on missing attribute data.

[0030] (7) The talent evaluation method proposed in this invention integrates attribute completion and weight calculation into a unified framework, realizing full process automation from data preprocessing to score output, and reducing error accumulation compared with traditional staged processing methods.

[0031] (8) The present invention is highly inclusive of attribute types, making it adaptable to various evaluation scenarios such as corporate recruitment and academic selection. Applications can be quickly migrated by simply adjusting the input attribute set.

[0032] (9) The present invention has designed a flexible system of comprehensive high- and low-dimensional indicators, which can dynamically select and score different evaluation indicators according to specific talent needs, thereby achieving personalized and accurate comprehensive quality assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Training flowchart for attribute completion model;

[0034] Figure 2 Flowchart for attribute weight training;

[0035] Figure 3 This is the graph neural network structure diagram;

[0036] Figure 4 A flow chart of talent evaluation methods;

[0037] Figure 5 Comparison between talent rating and expert rating for the embodiment. DETAILED DESCRIPTION

[0038] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] In this implementation, attributes for evaluating talent value are first set, and talents with complete attribute values are recorded as full-attribute talents; talents with missing attribute values are recorded as missing-attribute talents; the attribute data of full-attribute talents are recorded as full-attribute data, and the attribute data of missing-attribute talents are recorded as missing-attribute data.

[0040] This embodiment proposes a training method for a talent assessment model based on GNN. First, an attribute completion model and a dimension prediction model are constructed and trained. The attribute completion model is used to complete the missing attribute data to obtain full attribute data. The dimension prediction model is used to train the weights of each attribute, so that all attributes of each talent are multiplied by the corresponding weights and summed up to obtain a talent score.

[0041] The attribute completion model includes a graph neural network and a completion network connected front and back. After the attribute data of the talent with missing attributes is input into the attribute completion model, the graph neural network converts the attribute data into graph structure data. The completion model processes the graph structure data and outputs the attribute data of all attributes, thus completing the attributes of the talent with missing attributes.

[0042] The completion network can use multi-layer perceptron MLP, fully connected layer, convolutional neural network, etc.

[0043] Reference Figure 1 ,The training method of the attribute completion model includes the following steps:

[0044] S11. First, obtain attribute data of all-attribute talents and record them as full-attribute data. Then, randomly delete some attributes of each full-attribute data to form corresponding missing attribute data and full-attribute data as a first training set. The samples in the first training set are recorded as {missing attribute data, full-attribute data}.

[0045] S12. Calculate the weak supervision loss of the attribute completion model on the first training set, and update the completion network according to the weak supervision loss until the attribute completion model converges.

[0046] Specifically, step S12 is divided into the following sub-steps:

[0047] S121, randomly selecting training samples from the first training set and substituting them into the attribute completion model, allowing the attribute completion model to learn the training samples;

[0048] S122. Calculate the weak supervision loss on the training samples and update the completion network based on the weak supervision loss.

[0049] In this implementation, the specific task uses the unsupervised training technology of the evaluation model based on the minimum conflict principle to train the model. The weak supervision loss is:

[0050]

[0051] in, Refers to randomly deleting nodes with some attributes, i.e. talent attribute data samples, X i is the full attribute data of node i, is the attribute data of node i after completion.

[0052] S123. Extract validation samples from the training set, and use the attribute completion model to complete the missing attribute data in the validation samples. Compare the missing attribute data with the corresponding full attribute data, and calculate the completion deviation, that is, the difference between the completed attribute data and the corresponding full attribute data.

[0053] The attribute data after missing attribute data is completed is referred to as completed data.

[0054] In specific implementation, the completion bias can adopt the cross entropy loss or mean square error loss between the completed data and the corresponding full attribute data; it can also adopt the distance between the completed data and the corresponding full attribute data, specifically the Euclidean distance; it can also adopt the similarity between the completed data and the corresponding full attribute data, etc.

[0055] S124. Determine whether the completion deviation has converged; if not, return to step S121; if yes, fix the completion network and output the attribute completion model.

[0056] The dimensional prediction model includes a graph neural network and a dimensional prediction network. The dimensional prediction network is specifically a multi-layer perceptron MLP. After the attribute data of all-attribute talents are input into the dimensional prediction model, the graph neural network converts the attribute data into graph structure data. The dimensional prediction network contains nodes corresponding to each attribute.

[0057] Reference Figure 2 ,The training method of the dimension prediction model includes the following steps:

[0058] S21. Obtain all attribute data as a training set; obtain all attribute data annotated with expert ratings as a validation set;

[0059] S22. Randomly select all attribute data from the training set as training samples and input them into the dimensional prediction model. The graph neural network converts the all attribute data into graph structure data. After deleting one target attribute from the graph structure data, the graph structure data is input into the dimensional prediction model. The dimensional prediction model predicts the target attribute based on the remaining attributes in the graph structure data and trains the weight of the target attribute.

[0060] S23, determining whether the target attribute traverses all attributes;

[0061] If not, update the target attribute and then return to step S22;

[0062] If yes, then output the weight of each attribute and proceed to the next step;

[0063] S24. Extract validation samples from the validation set, and take the weighted sum of all attribute data of the validation samples and each attribute as the talent score; calculate the cross entropy loss based on the expert score and the talent score on the validation samples;

[0064] The talent score calculation formula is:

[0065] Pi=x i1 ×a1+x i2 ×a2+x i3 ×a3+…+x in ×a n +…+x iN ×a N

[0066] Among them, Pi represents the talent score of talent i, x in represents the score of talent i on the nth attribute, a n represents the weight of the nth attribute, N represents the total number of attributes, 1≤n≤N; that is, x i1 、x i2 、x i3 、x iN They represent the scores of talent i on the 1st, 2nd, 3rd, and Nth attributes, respectively. a1, a2, a3, a NRepresent the weights of the 1st, 2nd, 3rd, and Nth attributes respectively.

[0067] S25, determine whether the cross entropy loss converges; if not, return to step S22; if yes, output the attribute weights a1, a2, a3, ..., a n ,…,a N .

[0068] The convergence condition of the cross entropy loss can be set as follows: the range of the cross entropy loss calculated during the most recent M iterations is less than or equal to a set threshold. In specific implementations, M can be set to a value in the interval [3, 15] such as 3, 5, or 10, and the threshold can be set in the interval (0, 0.1), specifically 0.03, 0.5, etc.

[0069] It can be seen that the talent evaluation model proposed in the present invention includes an attribute completion model and attribute weights. The set of attribute weights is {a n |1≤n≤N}.

[0070] Graph neural networks are required in both the above attribute completion models and dimension prediction models. Graph neural networks adopt a meta-path-based heterogeneous network representation learning method, and use topological structures to obtain the topological representation of nodes, so as to capture high-order topological relationships between nodes as prior knowledge for attribute completion. The relationship between attribute-free nodes and directly connected existing attribute nodes is calculated based on the topological representation of the nodes, and the attributes of the attribute-free nodes are completed by weighted aggregation of the attributes of the existing attribute nodes. In addition to using multi-layer perceptrons to represent the features of nodes and edges, graph neural networks use node-to-edge (v→e) and edge-to-node (e→v) operations to represent the message diffusion of features between node and edge representations in graph neural networks, and after each feature movement, the model will superimpose multiple layers of MLP to represent the features. Specifically, given a graph G(V, E), where vertex v∈V, edge The single message passing operation from node to node in GNN is defined as follows:

[0071]

[0072] in, is node v i and node v j The embedding of the edge e(i, j) between is node v i In the embedding at layer l, is node v j In the embedding at layer l, is node v j In the embedding of layer l+1, x (i，j) is the initial feature of edge e(i, j), such as the type of edge; x j is node vj Initial features of N, such as node input. j Represents the index set of neighbor nodes connected by the input edge, that is, the node v j The input node set of the edge of the end point; [.,.] represents the connection of the vector. Function and The lth layer of neural networks for nodes and edges can be stacked using multiple linear layers. The message passing operation allows multiple rounds of stacking to form a model that maps from edges to node representations or from nodes to edges. The structure of the graph neural network is shown in Figure 3 .

[0073] Reference Figure 4 The talent scoring method using the talent assessment model proposed in this embodiment includes the following steps:

[0074] S1. Obtain talent information and extract attribute data;

[0075] S2. Determine whether the attribute data is missing attribute data or complete attribute data; if it is complete attribute data, input the attribute data into the attribute completion model to obtain the completed complete attribute data; if it is complete attribute data, proceed directly to the next step;

[0076] S3. Refer to the talent score calculation formula and perform weighted calculation on all attribute data and attribute weights to obtain the talent score.

[0077] In this embodiment, attribute data serving as model input is expressed as a feature vector, which is specifically composed of numerical values serving as indices of the attributes.

[0078] The present invention is verified below with reference to specific examples.

[0079] In this example, the talent attributes set include basic attributes and academic literacy. Basic attribute indicators include: highest academic degree, professional title / position, title, affiliation, length of service, number of publications, publication reputation, and talent cultivation. Academic literacy indicators include: number of papers, citations, H&I10 index, paper value, academic collaboration, academic activity, academic breadth, and funding. These 16 quantitative indicators reflect a talent's professional quality and moral character, the completeness of their knowledge system, and their scientific research capabilities.

[0080] Indicators of different dimensions may have different dimensions and ranges. Directly using these raw data may cause some indicators to have too much influence on the model during training, while other indicators have less influence. Therefore, by normalizing the indicators of each dimension to the same scale (such as [0, 1]), it can ensure that the contribution of each dimension to the model is balanced, thereby avoiding the indicator of a certain dimension dominating the output of the model. In the specific implementation of this solution, Min-Max normalization can be used to scale the indicator data of each dimension to the range of [0, 1], thereby constructing a normalized multidimensional indicator x i =[x i1 , x i2 ,...,x im ] as attribute data of talents.

[0081] In this embodiment, 16 indicators are set, so N=16; if m is less than N, it means that the attribute data [x i1 , x i2 ,...,x im ] is missing attribute data; if m=N, it means attribute data [x i1 , x i2 ,...,x im ] is the full attribute data.

[0082] In this embodiment, the quantification methods of different indicators are as follows:

[0083] (1) Highest academic qualification, based on a multi-classification gradient:

[0084] PhD (3 points); Master (2 points); Bachelor (1 point).

[0085] (2) Job title / position, scored based on a multi-category gradient:

[0086] Senior professor (4 points); Associate senior professor (3 points); Externally hired doctoral supervisor (2 points); Postdoctoral researcher (1 point); Others (0 points).

[0087] (3) Title, based on a multi-category gradient:

[0088] Academician (5 points); Outstanding Young Scientist / Changjiang Scholar (4 points); Four Young Scientists (3 points); Other important titles (leading talent, chief scientist, etc.) (2 points); General titles (1 point); Others (0 points).

[0089] (4) Employment unit, based on multi-category gradient scoring:

[0090] Top (4 points); High level (3 points); Intermediate (2 points); Average (1 point); Other (0 points).

[0091] (5) Length of service: points are assigned based on a multi-category gradient.

[0092] Length of service > 15 years (4 points); 10 years ≤ length of service ≤ 15 years (3 points); 5 years ≤ length of service ≤ 10 years (2 points); Length of service < 5 years (1 point).

[0093] The classification gradients of the above five indicators are specified by experts.

[0094] (6) Number of publications, based on multi-category gradient scoring.

[0095]

[0096] N1 is the number of publications of the talent, A k The author's position in the kth work is: first / chief editor: 4 points; second / deputy editor: 3 points; third: 2 points; fourth and subsequent editors: 1 point.

[0097] (7) Work reputation: Check the monograph ratings on websites such as Douban. The calculation formula is as follows.

[0098]

[0099] N1 is the number of publications by the talent, R k These are the evaluation scores from evaluation websites such as Douban, JD.com, and Dangdang.

[0100] (8) Educating people and cultivating talents: evaluating the performance of talents in guiding and cultivating students, including the quantity and quality of students guided.

[0101]

[0102] N2 is the number of students or scientific research talents in talent cultivation; S k T is the training duration or guidance depth of the k-th training object, reflecting the degree of investment of the talent in the training process; k is the achievement coefficient of the kth trainee, reflecting the subsequent performance of the trainee in scientific research or career, such as the number of papers published, honorary awards received, etc.; C k is the contribution coefficient of the kth trainee, which is reflected in the trainee's final destination, such as further study abroad, further study domestically, or domestic employment (central state-owned enterprises, high-tech industries, etc.).

[0103] (9) The number of papers is an important indicator to measure the academic contribution of talents. It is reflected in the number of journal papers published and the number of conference activities, which reflects their scientific research output capacity.

[0104] (10) Paper citations refer to the number of times a talent’s academic achievements are cited by their peers, reflecting the influence and recognition of their research work, which is expressed in the number of citations of journal papers and conference reports.

[0105] (11) h&i10 index, which evaluates the academic contribution and influence of scientific research talents by combining the H index and i10 index, providing a more comprehensive evaluation perspective.

[0106] (12) The value of the paper, combined with the paper impact factor IF and the paper citation cit p and author rank p , to evaluate the comprehensive value of talent journal (conference) papers r1, the calculation formula is as follows:

[0107]

[0108] Among them, T is the collection of signed papers of the talent to be evaluated, represents the number of citations of paper r1, Indicates the author rank of the talent to be evaluated in paper r1; e is the natural base number;

[0109] (13) Academic collaboration, taking into account the number of academic collaborators C and the number of co-authored works num co , collaborator academic influence Inf(co) and organization weight co The calculation formula is as follows:

[0110]

[0111] Among them, co represents any one of the co-authors of the talent to be evaluated, num co Inf(co) represents the number of co-authored works by the talent to be evaluated and the author, Inf(co) represents the academic influence of co, and weight co Indicates the organizational level to which co belongs; Inf(co) and weight co Annotated by experts.

[0112] (14) Academic activity reflects the degree of investment of talents in scientific research activities, which is reflected in the recent publication of academic achievements and the implementation of scientific research projects. It comprehensively considers the academic output of talents in the last Q years. q 、Influence IndexInf(T q ) and the corresponding weight of the new results in the calculation weight(q)=α (q) , the calculation formula is as follows:

[0113]

[0114] Among them, T q represents the number of academic outputs in the qth year before, and weight(q) represents the weight of academic outputs in the qth year before.

[0115] (15) Academic breadth: Diverse research fields may have different impacts on the scientific research performance of scholars in different fields. Based on the scientific research and academic data modeling, the research topic t and the number of achievements num for each talent are obtained. res , automatically classify academic achievements such as papers into each topic one by one, and calculate the topic distribution based on the assignment results:

[0116]

[0117] To further calculate the breadth of talent i, the formula is as follows:

[0118]

[0119] Among them, T is the research topic set, num res represents the number of achievements of talents in the subject set T, r i,t Represents the number of achievements of talents in topic t.

[0120] (16) The number of funds, whether the talent has received funding from the National Natural Science Foundation, reflects their scientific research capabilities and project competitiveness, as well as the national recognition of their research work, including general projects, key projects, major projects, major research plans, National Science Fund for Distinguished Young Scholars, Overseas, Hong Kong and Macao Young Scholars Collaborative Research Fund, Innovative Research Group Science Fund, National Basic Science Talent Training Fund, special projects, joint funding projects and international (regional) cooperation and exchange projects.

[0121] In this embodiment, the trained talent assessment model is used to score 50 talents respectively. The talent scores obtained by the method of the present invention are compared with the expert scores as shown in Table 1 below.

[0122] Table 1: Rating comparison

[0123]

[0124] Figure 5 The data trend analysis in Table 1 is shown above. Figure 5 It can be seen that the talent rating obtained by the method of the present invention is consistent with the expert rating trend.

[0125] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0126] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0127] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A training method for a talent assessment model based on GNN, characterized in that: The talent assessment model includes a dimensional prediction model for training attribute weights. The talent score is the weighted sum of each attribute and the attribute weight. Training the dimensional prediction model includes the following steps: The training set consists of all attribute data and the validation set of all attribute data labeled with expert ratings. The dimensional prediction model is used to remove a target attribute from the training set and infer the target attribute and its weight based on the remaining attributes. The target attribute is used to traverse all attributes to obtain the weight of each attribute. In the validation set, the attribute score of the talent is multiplied by the corresponding weight and the sum is used as the predicted talent weight. The attribute weight is then verified by comparing the talent weight with the expert weight. The training of the dimensional prediction model and the verification of the attribute weights are repeated until the attribute weights reach the convergence condition.

2. The training method of the GNN-based talent assessment model according to claim 1, characterized in that: The dimension prediction model includes a graph neural network and a dimension prediction network. The graph neural network converts attribute data into graph structure data, and the dimension prediction network contains nodes that correspond one-to-one to each attribute. During the dimension prediction model training process, the weights of each node in the dimension prediction network are updated. After training is completed, the node weights of the dimension prediction network are output as attribute weights.

3. The training method of the GNN-based talent assessment model according to claim 2, characterized in that: The convergence condition of attribute weights is that the cross entropy loss of talent ratings and expert ratings calculated on the validation samples converges.

4. The training method of the GNN-based talent assessment model according to claim 1, characterized in that: The talent assessment model also includes an attribute completion model for completing missing attribute data; the attribute completion model includes a front-to-back connected graph neural network and a completion network. The graph neural network converts attribute data into graph structure data, and the completion network completes attributes for the graph structure data.

5. The training method of the GNN-based talent assessment model according to claim 4, characterized in that: The training method of the attribute completion model is as follows: first, obtain all attribute data and randomly delete some attributes to construct a training set; then extract training samples from the training set and input them into the attribute completion model and calculate the loss function, and update the completion network according to the loss function; and repeat the training until the attribute completion model converges.

6. The training method of the GNN-based talent assessment model according to claim 5, characterized in that: The process of training an attribute completion model on training consists of the following steps: First, training samples are selected from the training set, and the attribute completion model is made to learn the training samples; weak supervision loss is calculated on the training samples, and the completion network is updated according to the weak supervision loss; Extract validation samples from the training set, let the attribute completion model complete the missing attribute data in the validation samples, and calculate the difference between the attribute data after the missing attribute data is completed and the corresponding full attribute data; Determine whether the calculated difference has converged; if not, return to the training process of the attribute completion model; if so, fix the completion network and output the attribute completion model.

7. The training method of the GNN-based talent assessment model according to claim 6, characterized in that: The difference between the attribute data after missing attribute data is completed and the corresponding full attribute data is represented by: the cross entropy loss or mean square error loss of the completed value and the original value of each attribute; or the distance between the completed attribute data and the corresponding full attribute data.

8. A talent assessment method using the training method of the GNN-based talent assessment model according to any one of claims 4 to 7, characterized in that: First, a talent evaluation model is obtained by using the training method of the GNN-based talent evaluation model; the talent evaluation model includes an attribute completion model and the weight of each attribute; then, the attribute data of the talent to be evaluated is obtained, and whether the attribute data is complete is determined; If not, the attribute data will be completed through the attribute completion model, and the completed attribute data will be weighted and summed with the weights of each attribute to obtain the talent score; if the attribute data of the talent is complete, the attribute data will be directly weighted and summed with the weights of each attribute to obtain the talent score.

9. The talent evaluation method according to claim 8, wherein: Talent attributes include: basic attributes and academic literacy; indicators of basic attributes include: highest academic degree, professional title / position, title, institution of employment, length of service, number of publications, reputation of publications, and talent cultivation; indicators of academic literacy include: number of papers, citations of papers, h&i10 indicators, paper value, academic cooperation, academic activity, academic breadth, and number of funds.

10. A talent selection method, characterized in that: First, determine the target field and the talent's scores on basic attributes; then determine the talent's academic literacy scores in the target field; Construct talent attribute data based on basic attribute scores and academic literacy scores in target fields; then execute the talent evaluation method as described in claim 8 or 9 to obtain talent scores of candidate talents; select the required talents from the candidate talents in descending order of talent scores.

Images