Intelligent talent information matching management system

Abstract

The invention relates to the technical field of information management, and discloses a talent information intelligent matching management system, which comprises a matching module, a sequence construction module and a sequence optimization module, and is characterized in that the matching module is used for constructing a demand vector matrix according to demand information, constructing an evaluation vector matrix according to talent information, and optimizing the evaluation vector matrix according to the talent information; the demand vector matrix and the evaluation vector matrix are input into a noise disturbance model generated through pre-training, the output matching degree is obtained, the sequence construction module is used for determining Q to-be-recommended targets according to the matching degree, constructing P to-be-recommended sequences based on the Q to-be-recommended targets and calculating the prejudice influence degree value corresponding to each to-be-recommended sequence, and finally, the prejudice influence degree value corresponding to each to-be-recommended sequence is obtained. The sequence optimization module is used for optimizing the P to-be-recommended sequences according to the prejudice influence degree value and the natural heuristic algorithm, the nonlinear relation between high-dimensional complex data is captured, the to-be-recommended sequences are optimized through the natural heuristic algorithm, and the problem of fairness deficiency in the recommendation process is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of information management technology, and more specifically, to an intelligent matching management system for talent information. Background Art

[0002] With the rapid development of the Internet and big data technologies, intelligent matching and evaluation systems play an increasingly important role in various platforms. These systems are widely used in online education platforms, recruitment websites, social networks, e-commerce platforms and other fields. They aim to intelligently recommend the most suitable talents, products or services based on user needs and characteristics. However, current intelligent matching and evaluation technologies still face many challenges in processing high-dimensional and complex information and ensuring the fairness of the recommendation process.

[0003] Existing technologies mainly rely on traditional machine learning models and basic deep learning models. However, these models have certain limitations in their ability to capture complex nonlinear relationships when processing high-dimensional data, resulting in matching accuracy and recommendation quality that are difficult to meet actual needs. In addition, these models pay little attention to fairness in the recommendation process, which makes the bias problem of recommendation results more serious.

[0004] In view of this, the present invention proposes an intelligent talent information matching management system to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent matching management system for talent information.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] Provides a talent information intelligent matching management system, including:

[0008] Matching module: used to obtain demand information and talent information, build a demand vector matrix based on the demand information, build an evaluation vector matrix based on the talent information, input the demand vector matrix and the evaluation vector matrix into the noise perturbation model generated by pre-training, and obtain the output matching degree;

[0009] Sequence construction module: used to determine Q targets to be recommended according to the matching degree, construct P sequences to be recommended based on the Q targets to be recommended, and calculate the bias influence value corresponding to each sequence to be recommended;

[0010] Sequence optimization module: used to optimize P recommended sequences according to the bias impact value and the natural inspiration algorithm to obtain the optimal recommended sequence.

[0011] In some embodiments, a method for constructing a demand vector matrix according to demand information includes:

[0012] Each element in the demand information is converted into a first vector, each first vector is clustered to obtain S first similarity clusters, each first similarity cluster is used as a row of the demand vector matrix, and the demand vector matrix is ​​constructed.

[0013] In some embodiments, the training method of the noise perturbation model includes:

[0014] Pair the historical demand vector matrix and the historical evaluation vector matrix with the corresponding historical matching degree labels to construct a training set. Use the historical demand vector matrix and the historical evaluation vector matrix as the input of the sub-model, and use the corresponding historical matching degree as the output of the sub-model.

[0015] The total number of diffusion steps is preset to T, and the noise increment of each step is defined as For each data point in the training set, the current diffusion step t and the corresponding noise data are generated by random sampling , construct a loss function according to the current diffusion step t and the noise data, wherein the distribution form of the noise data is Gaussian distribution;

[0016] A batch of data points are randomly selected from the training set, corresponding noise data is generated for each data point, the corresponding loss function value is calculated, the sub-model parameters are updated through the back-propagation algorithm and the minimization of the loss function value as the goal, the sub-model is iteratively trained, and the diffusion probability model is obtained.

[0017] In some embodiments, noise data is generated The methods include:

[0018] ;

[0019] in, is the forward process coefficient, is the cumulative coefficient, is the data point, is the standard normal distribution noise, It is represented as the identity matrix, For steps The corresponding cumulative coefficient is Characterized by normal distribution.

[0020] In some embodiments, according to the current diffusion step t and the noise data Methods for constructing loss functions include:

[0021] ;

[0022] in, Characterized as a single-step loss, Characterized as the total loss, Characterize the data points , standard normal distribution noise and the expected value of the current diffusion step t, It is represented by the square norm, represents the predicted normally distributed noise at the tth step.

[0023] In some embodiments, a method for constructing P sequences to be recommended based on Q targets to be recommended includes:

[0024] The Q targets to be recommended are randomly selected to obtain H combinations to be recommended, and the H combinations to be recommended are fully arranged to construct P sequences to be recommended.

[0025] In some embodiments, the method for calculating the bias influence value corresponding to each sequence to be recommended includes:

[0026] BIV ;

[0027] Among them, BIV is represented by the bias impact value, is the total number of elements in the sequence to be recommended, For the The gender code corresponding to the target to be recommended, is the average value of the gender encoding of all targets to be recommended, For the The matching degree of the target to be recommended is For the The age of the target to be recommended, is the average age of all targets to be recommended.

[0028] In some embodiments, the natural heuristic algorithm is a sparrow search algorithm, and the method of optimizing P sequences to be recommended according to the bias influence value and the natural heuristic algorithm to obtain the optimal recommended sequence includes:

[0029] S1: Randomly generate a group of sparrow individuals, each sparrow individual represents a sequence to be recommended, and randomly generate an initial position for each sparrow individual ;

[0030] S2: Construct the fitness function value of each sparrow individual :

[0031] = ;

[0032] In the formula, is the bias influence value corresponding to the sequence to be recommended;

[0033] S3: Perform exploration operations on some sparrow individuals and update their initial sound wave parameters ;

[0034] = ;

[0035] In the formula, For the A sparrow in the The position of the generation, For the A sparrow in the The position of the generation, is the maximum number of iterations, is a random angle, It is an exponential function with e as base, where e is a natural constant;

[0036] S4: Follow the remaining sparrows and update their initial sound wave parameters ;

[0037] S5: Calculate the fitness function value of each sparrow individual after the update , update the current optimal solution;

[0038] S6: Repeat S3-S5 until the preset maximum number of iterations is reached. Stop searching when , and output the recommended sequence corresponding to the current optimal solution, which is the optimal recommended sequence.

[0039] In some embodiments, the following operation is performed on the remaining sparrow individuals to update their initial sound wave parameters. The methods include:

[0040] = ;

[0041] In the formula, is a random number, is the location of the current optimal solution.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention first constructs a demand vector matrix according to demand information, constructs an evaluation vector matrix according to talent information, inputs the demand vector matrix and the evaluation vector matrix into a pre-trained noise perturbation model to obtain the output matching degree, and then determines Q targets to be recommended according to the matching degree, constructs P sequences to be recommended based on the Q targets to be recommended, calculates the bias influence value corresponding to each sequence to be recommended, and finally optimizes the P sequences to be recommended according to the bias influence value and a nature-inspired algorithm to obtain the optimal recommendation sequence. The present invention captures the nonlinear relationship between high-dimensional complex data and improves the matching accuracy by inputting the demand vector matrix and the evaluation vector matrix into a pre-trained noise perturbation model, combines matching degree screening and bias influence value calculation, and then uses a nature-inspired algorithm to optimize the sequence to be recommended, thereby effectively solving the problem of lack of fairness in the recommendation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a schematic diagram of the structure of an intelligent matching management system for talent information in the present invention;

[0045] Figure 2 A flow chart of a talent information intelligent matching management method in the present invention;

[0046] Figure 3 It is a flowchart of the method for constructing a demand vector matrix according to demand information in the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Unless otherwise defined, the technical or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0049] Example 1

[0050] See also Figure 1 As shown, this embodiment discloses a talent information intelligent matching management system, including:

[0051] Matching module: used to obtain demand information and talent information, build a demand vector matrix based on the demand information, build an evaluation vector matrix based on the talent information, input the demand vector matrix and the evaluation vector matrix into the noise perturbation model generated by pre-training, and obtain the output matching degree;

[0052] In this embodiment, the demand information represents the demand proposed by the user, specifically, the talent characteristics or requirements that need to be matched from the talent pool. The talent information represents the data used to describe the talent characteristics, usually including various attributes, skills, and experience related to the talent. For example, when applied to an online education system, the demand information includes learning goals, learning progress, learning preferences, and geographic location or time zone. The talent information includes the teacher's professional field and qualifications, teaching experience, teaching style or method, and language ability (such as whether bilingual teaching can be provided, etc.).

[0053] like Figure 3 As shown, the method of constructing a demand vector matrix according to demand information includes:

[0054] Each element in the demand information is converted into a first vector, each first vector is clustered to obtain S first similarity clusters, each first similarity cluster is used as a row of the demand vector matrix, and the demand vector matrix is ​​constructed.

[0055] It should be noted that the reason for converting each element in the demand information into a first vector is to digitize and standardize the unstructured or complex demand information to facilitate subsequent processing, clustering and matrix construction. The elements can be converted into vectors by encoding, and the first vectors can be clustered by K-means clustering or DBSCAN clustering. K-means clustering and DBSCAN clustering are both existing technologies, and this embodiment will not go into details.

[0056] Methods for constructing an evaluation vector matrix based on talent information include:

[0057] Each element in the talent information is converted into a second vector, each second vector is clustered to obtain W second similarity clusters, and each second similarity cluster is used as a row of the evaluation vector matrix, thereby constructing the evaluation vector matrix.

[0058] It can be understood that the method of constructing the evaluation vector matrix is ​​similar to the method of constructing the demand vector matrix, and the above-mentioned matching degree can be the similarity between the demand vector matrix and the evaluation vector matrix. For example, the matching degree is calculated in the following way:

[0059] ;

[0060] In the formula, Characterized by matching degree, Represented as a demand vector matrix, Represented as an evaluation vector matrix, is represented as the transposed matrix of the evaluation vector matrix, is the norm of the matrix.

[0061] In this embodiment, the matching degree is determined by calculating the cosine similarity between the demand vector matrix and the evaluation vector matrix. At the same time, the matching degree can also be obtained by calculating the Euclidean distance, but this embodiment will not go into details.

[0062] The training methods of the noise perturbation model include:

[0063] Pair the historical demand vector matrix and the historical evaluation vector matrix with the corresponding historical matching degree labels to construct a training set. Use the historical demand vector matrix and the historical evaluation vector matrix as the input of the sub-model and use the corresponding historical matching degree as the output of the sub-model. The sub-model can be U-Net or Transformer.

[0064] The total number of diffusion steps is preset to T, and the noise increment of each step is defined as For each data point in the training set, the current diffusion step t and the corresponding noise data are generated by random sampling , construct a loss function according to the current diffusion step t and the noise data, wherein the distribution form of the noise data is Gaussian distribution;

[0065] A batch of data points are randomly selected from the training set, corresponding noise data is generated for each data point, the corresponding loss function value is calculated, the sub-model parameters are updated through the back-propagation algorithm and the minimization of the loss function value as the goal, the sub-model is iteratively trained, and the diffusion probability model is obtained.

[0066] It should be noted that by obtaining the historical demand vector matrix, the historical evaluation vector matrix and the corresponding historical matching degree labels, a training set is formed to capture the matching rules between the real demand and the evaluation relationship; these data are used as the input and output of the sub-model (such as U-Net or Transformer), and the complex nonlinear mapping relationship between demand and evaluation is learned through the sub-model. At the same time, the total number of diffusion steps T is preset and the noise increment of each step is defined, and randomness is introduced into the training data to simulate the uncertainty that may exist in the actual scenario; further, data points are randomly selected from the training set and noise data is generated for each data point to expand the diversity of training samples. Finally, by constructing a single-step loss function and a cumulative total loss optimization model, the model can accurately capture the mapping rules between demand and evaluation. In the technical solution, this series of designs aims to enhance the generalization ability of the model through noise perturbation and multi-step diffusion, and ensure high-precision matching prediction under complex data conditions, thereby effectively addressing high-dimensional complex data and recommendation fairness issues.

[0067] Generate noisy data The methods include:

[0068] ;

[0069] in, is the forward process coefficient, is the cumulative coefficient, is the data point, is the standard normal distribution noise, It is represented as the identity matrix, For steps The corresponding cumulative coefficient is Characterized by normal distribution.

[0070] According to the current diffusion step t and the noise data Methods for constructing loss functions include:

[0071] ;

[0072] in, Characterized as a single-step loss, Characterized as the total loss, Characterize the data points , standard normal distribution noise and the expected value of the current diffusion step t, It is represented by the square norm, represents the predicted normally distributed noise at the tth step.

[0073] It should be added that the loss function construction method defines a single-step loss function to measure the difference between the predicted noise and the actual noise of the model in each diffusion step, and incorporates the noise disturbance characteristics into the loss evaluation, helping the model to gradually optimize in each diffusion step; at the same time, the cumulative algorithm comprehensively considers the single-step losses of all diffusion steps to optimize the global performance of the model. In the entire technical solution, this design enables the model to learn richer feature distributions during the diffusion process, thereby significantly improving the accuracy of matching predictions and the reliability of the overall technical solution, providing higher accuracy guarantees for the results of the recommendation system.

[0074] Sequence construction module: used to determine Q targets to be recommended according to the matching degree, construct P sequences to be recommended based on the Q targets to be recommended, and calculate the bias influence value corresponding to each sequence to be recommended;

[0075] In this embodiment, still taking the application in the online education system as an example, determining Q targets to be recommended according to the matching degree means that the potential targets in the online education system are sorted in descending order according to the corresponding matching degree, and the first Q targets are selected as the targets to be recommended. The targets to be recommended represent teachers.

[0076] The method for constructing P to-be-recommended sequences based on Q to-be-recommended targets includes:

[0077] The Q targets to be recommended are randomly selected to obtain H combinations to be recommended, and the H combinations to be recommended are fully arranged to construct P sequences to be recommended.

[0078] For example, three combinations to be recommended are randomly selected from five targets (TA1, TA2, TA3, TA4, TA5), for example:

[0079] Combination to be recommended 1: [TA1, TA2, TA3];

[0080] Combination 2 to be recommended: [TA2, TA4, TA5];

[0081] Combination to be recommended 3: [TA1, TA3, TA5];

[0082] The full permutation of the combination to be recommended 1 is [TA1, TA2, TA3], [TA1, TA3, TA2], ..., [TA3, TA2, TA1]. The same is true for the combination to be recommended 2 and 3, which will not be elaborated in this embodiment.

[0083] The method for calculating the bias impact value corresponding to each sequence to be recommended includes:

[0084] BIV ;

[0085] Among them, BIV is represented by the bias impact value, is the total number of elements in the sequence to be recommended, For the The gender code corresponding to the target to be recommended, is the average value of the gender encoding of all targets to be recommended, For the The matching degree of the target to be recommended is For the The age of the target to be recommended, is the average age of all targets to be recommended.

[0086] It can be understood that the gender coding mentioned above refers to the digital representation of the gender information of the recommendation target. For example, male is encoded as 2 and female is encoded as 1. By introducing gender coding, it is possible to quantitatively analyze whether there is an imbalance or deviation in the gender distribution in the recommendation sequence. From the above content, it can be seen that the bias impact value is the sum of the absolute values ​​of the deviations of the gender distribution and the age distribution. The larger the value, the greater the degree to which the gender and age distribution of the recommended sequence deviates from the overall distribution.

[0087] In this embodiment, the matching degree ranking of the target is first determined to provide accuracy guarantee for the sequence to be recommended. The gender and age distribution of each sequence to be recommended is quantitatively evaluated by calculating the bias influence value, which can achieve an efficient process from accurate matching of user needs to diversified recommendations. At the same time, the bias influence is effectively controlled in the process of generating the recommendation sequence, thereby improving the accuracy, fairness and user experience satisfaction of the recommendation results of the online education system.

[0088] Sequence optimization module: used to optimize P recommended sequences according to the bias impact value and the natural inspiration algorithm to obtain the optimal recommended sequence.

[0089] The natural heuristic algorithm can be a sparrow search algorithm. The method of optimizing P sequences to be recommended according to the bias influence value and the natural heuristic algorithm to obtain the optimal recommendation sequence includes:

[0090] S1: Randomly generate a group of sparrow individuals, each sparrow individual represents a sequence to be recommended, and randomly generate an initial position for each sparrow individual ;

[0091] Among them, by randomly generating the initial sparrow individuals and their initial positions, the algorithm starts from different solution space positions to avoid falling into the local optimum. Each sparrow individual corresponds to a sequence to be recommended.

[0092] S2: Construct the fitness function value of each sparrow individual :

[0093] = ;

[0094] In the formula, is the bias influence value corresponding to the sequence to be recommended.

[0095] In this embodiment, the bias influence value is used as the optimization target, so that the algorithm can adjust the recommendation sequence according to the bias influence value to generate a more balanced recommendation result, so that the optimization process of the recommendation sequence not only focuses on the recommendation quality, but also effectively controls the deviation of gender and age distribution, thereby improving the fairness of the recommendation system.

[0096] S3: Perform exploration operations on some sparrow individuals and update their initial sound wave parameters ;

[0097] = ;

[0098] In the formula, For the A sparrow in the The position of the generation, For the A sparrow in the The position of the generation, is the maximum number of iterations, is a random angle, It is an exponential function with e as base, where e is a natural constant.

[0099] In this embodiment, some sparrow individuals are made to search near the current solution space through exploration operation, and the diversity of solutions is increased by using exponential decay function and angle adjustment.

[0100] S4: Follow the remaining sparrows and update their initial sound wave parameters ;

[0101] In this embodiment, during the search process, the population is kept concentrated near the current optimal solution to accelerate the convergence speed, avoid falling into the local optimum, improve the optimization efficiency of the recommendation sequence, and enable the algorithm to quickly approach the global optimal solution while retaining a certain degree of search flexibility.

[0102] S5: Calculate the fitness function value of each sparrow individual after the update , update the current optimal solution;

[0103] S6: Repeat S3-S5 until the preset maximum number of iterations is reached. Stop searching when , and output the recommended sequence corresponding to the current optimal solution, which is the optimal recommended sequence.

[0104] Perform follow-up operations on the remaining sparrows and update their initial sound wave parameters The methods include:

[0105] = ;

[0106] In the formula, is a random number, is the location of the current optimal solution.

[0107] This embodiment first ensures the relevance of the recommendation target by screening the matching degree and generates a diverse sequence to be recommended. Then, these sequences to be recommended are optimized with the bias influence value as the target, which makes up for the imbalance in gender and age distribution that may be caused by the screening process. By combining the above steps in sequence, the recommendation system can realize a complete process from precise screening to fair optimization, which not only ensures the accuracy and diversity of the recommendation results, but also improves the fairness of the recommendation system and user satisfaction through the optimization of the bias influence value.

[0108] This embodiment first constructs a demand vector matrix based on demand information and an evaluation vector matrix based on talent information, inputs the demand vector matrix and the evaluation vector matrix into a pre-trained noise perturbation model to obtain an output matching degree, and then determines Q targets to be recommended based on the matching degree, constructs P sequences to be recommended based on the Q targets to be recommended, calculates the bias influence value corresponding to each sequence to be recommended, and finally optimizes the P sequences to be recommended based on the bias influence value and a nature-inspired algorithm to obtain the optimal recommendation sequence. This embodiment captures the nonlinear relationship between high-dimensional complex data and improves the matching accuracy by inputting the demand vector matrix and the evaluation vector matrix into a pre-trained noise perturbation model, combines matching degree screening and bias influence value calculation, and then uses a nature-inspired algorithm to optimize the sequence to be recommended, thereby effectively solving the problem of lack of fairness in the recommendation process.

[0109] Example 2

[0110] See also Figure 2 As shown, based on the same inventive concept, this embodiment discloses a method for intelligent matching management of talent information. For details not provided in this embodiment, please refer to the description of the relevant parts in Embodiment 1. The system includes:

[0111] S10: Obtain demand information and talent information, construct a demand vector matrix according to the demand information, construct an evaluation vector matrix according to the talent information, input the demand vector matrix and the evaluation vector matrix into the noise perturbation model generated by pre-training, and obtain the output matching degree;

[0112] In this embodiment, the demand information represents the demand proposed by the user, specifically, the talent characteristics or requirements that need to be matched from the talent pool. The talent information represents the data used to describe the talent characteristics, usually including various attributes, skills, and experience related to the talent. For example, when applied to an online education system, the demand information includes learning goals, learning progress, learning preferences, and geographic location or time zone. The talent information includes the teacher's professional field and qualifications, teaching experience, teaching style or method, and language ability (such as whether bilingual teaching can be provided, etc.).

[0113] Methods for constructing a demand vector matrix based on demand information include:

[0114] Each element in the demand information is converted into a first vector, each first vector is clustered to obtain S first similarity clusters, each first similarity cluster is used as a row of the demand vector matrix, and the demand vector matrix is ​​constructed.

[0115] Methods for constructing an evaluation vector matrix based on talent information include:

[0116] Each element in the talent information is converted into a second vector, each second vector is clustered to obtain W second similarity clusters, and each second similarity cluster is used as a row of the evaluation vector matrix, thereby constructing the evaluation vector matrix.

[0117] It can be understood that the method of constructing the evaluation vector matrix is ​​similar to the method of constructing the demand vector matrix, and the above-mentioned matching degree can be the similarity between the demand vector matrix and the evaluation vector matrix. For example, the matching degree is calculated in the following way:

[0118] ;

[0119] In the formula, Characterized by matching degree, Represented as a demand vector matrix, Represented as an evaluation vector matrix, is represented as the transposed matrix of the evaluation vector matrix, is the norm of the matrix.

[0120] In this embodiment, the matching degree is determined by calculating the cosine similarity between the demand vector matrix and the evaluation vector matrix. At the same time, the matching degree can also be obtained by calculating the Euclidean distance, but this embodiment will not go into details.

[0121] The training methods of the noise perturbation model include:

[0122] Pair the historical demand vector matrix and the historical evaluation vector matrix with the corresponding historical matching degree labels to construct a training set. Use the historical demand vector matrix and the historical evaluation vector matrix as the input of the sub-model and use the corresponding historical matching degree as the output of the sub-model. The sub-model can be U-Net or Transformer.

[0123] The total number of diffusion steps is preset to T, and the noise increment of each step is defined as For each data point in the training set, the current diffusion step t and the corresponding noise data are generated by random sampling , construct a loss function according to the current diffusion step t and the noise data, wherein the distribution form of the noise data is Gaussian distribution;

[0124] A batch of data points are randomly selected from the training set, corresponding noise data is generated for each data point, the corresponding loss function value is calculated, the sub-model parameters are updated through the back-propagation algorithm and the minimization of the loss function value as the goal, the sub-model is iteratively trained, and the diffusion probability model is obtained.

[0125] Generate noisy data The methods include:

[0126] ;

[0127] in, is the forward process coefficient, is the cumulative coefficient, is the data point, is the standard normal distribution noise, It is represented as the identity matrix, For steps The corresponding cumulative coefficient is Characterized by normal distribution.

[0128] According to the current diffusion step t and the noise data Methods for constructing loss functions include:

[0129] ;

[0130] in, Characterized as a single-step loss, Characterized as the total loss, Characterize the data points , standard normal distribution noise and the expected value of the current diffusion step t, It is represented by the square norm, represents the predicted normally distributed noise at the tth step.

[0131] S20: Determine Q targets to be recommended according to the matching degree, construct P sequences to be recommended based on the Q targets to be recommended, and calculate the bias influence value corresponding to each sequence to be recommended;

[0132] In this embodiment, still taking the application in the online education system as an example, determining Q targets to be recommended according to the matching degree means that the potential targets in the online education system are sorted in descending order according to the corresponding matching degree, and the first Q targets are selected as the targets to be recommended. The targets to be recommended represent teachers.

[0133] The method for constructing P to-be-recommended sequences based on Q to-be-recommended targets includes:

[0134] The Q targets to be recommended are randomly selected to obtain H combinations to be recommended, and the H combinations to be recommended are fully arranged to construct P sequences to be recommended.

[0135] The method for calculating the bias impact value corresponding to each sequence to be recommended includes:

[0136] BIV ;

[0137] Among them, BIV is represented by the bias impact value, is the total number of elements in the sequence to be recommended, For the The gender code corresponding to the target to be recommended, is the average value of the gender encoding of all targets to be recommended, For the The matching degree of the target to be recommended is For the The age of the target to be recommended, is the average age of all targets to be recommended.

[0138] It can be understood that the gender coding mentioned above refers to the digital representation of the gender information of the recommendation target. For example, male is encoded as 2 and female is encoded as 1. By introducing gender coding, it is possible to quantitatively analyze whether there is an imbalance or deviation in the gender distribution in the recommendation sequence. From the above content, it can be seen that the bias impact value is the sum of the absolute values ​​of the deviations of the gender distribution and the age distribution. The larger the value, the greater the degree to which the gender and age distribution of the recommended sequence deviates from the overall distribution.

[0139] S30: Optimize the P recommended sequences according to the bias influence value and the natural inspiration algorithm to obtain the optimal recommended sequence.

[0140] In the drawings of the embodiments of the present invention, only the structures involved in the embodiments of the present invention are involved, and other structures can refer to the general design. In the case of no conflict, the features in the same embodiment and different embodiments of the present invention can be combined with each other. The above are only specific implementation methods of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An intelligent matching management system for talent information, characterized in that: include: Matching module: used to obtain demand information and talent information, build a demand vector matrix based on the demand information, build an evaluation vector matrix based on the talent information, input the demand vector matrix and the evaluation vector matrix into the noise perturbation model generated by pre-training, and obtain the output matching degree; Sequence construction module: used to determine Q targets to be recommended according to the matching degree, construct P sequences to be recommended based on the Q targets to be recommended, and calculate the bias influence value corresponding to each sequence to be recommended; Sequence optimization module: used to optimize P recommended sequences according to the bias impact value and the natural inspiration algorithm to obtain the optimal recommended sequence.

2. The intelligent matching management system for talent information according to claim 1 is characterized in that: The method for constructing a demand vector matrix according to demand information includes: Each element in the demand information is converted into a first vector, each first vector is clustered to obtain S first similarity clusters, each first similarity cluster is used as a row of the demand vector matrix, and the demand vector matrix is ​​constructed.

3. The intelligent matching management system for talent information according to claim 2 is characterized in that: The training method of the noise disturbance model includes: Pair the historical demand vector matrix and the historical evaluation vector matrix with the corresponding historical matching degree labels to construct a training set. The historical demand vector matrix and the historical evaluation vector matrix are used as the input of the sub-model, and the corresponding historical matching degree is used as the output of the sub-model. The total number of diffusion steps is preset to T, and the noise increment of each step is defined as For each data point in the training set, the current diffusion step t and the corresponding noise data are generated by random sampling , construct a loss function according to the current diffusion step t and the noise data, wherein the distribution form of the noise data is Gaussian distribution; A batch of data points are randomly selected from the training set, corresponding noise data is generated for each data point, the corresponding loss function value is calculated, the sub-model parameters are updated through the back-propagation algorithm and the minimization of the loss function value as the goal, the sub-model is iteratively trained, and the diffusion probability model is obtained.

4. The intelligent matching management system for talent information according to claim 3 is characterized in that: The noise data is generated The methods include: ; in, is the forward process coefficient, is the cumulative coefficient, is the data point, is the standard normal distribution noise, It is represented as the identity matrix, For steps The corresponding cumulative coefficient is Characterized by normal distribution.

5. The intelligent matching management system for talent information according to claim 4 is characterized in that: According to the current diffusion step t and the noise data Methods for constructing loss functions include: ; in, Characterized as a single-step loss, Characterized as the total loss, Characterize the data points , standard normal distribution noise and the expected value of the current diffusion step t, It is represented by the square norm, represents the predicted normally distributed noise at the tth step.

6. The intelligent matching and management system for talent information according to claim 1 is characterized in that: The method for constructing P sequences to be recommended based on Q targets to be recommended includes: The Q targets to be recommended are randomly selected to obtain H combinations to be recommended, and the H combinations to be recommended are fully arranged to construct P sequences to be recommended.

7. The intelligent matching management system for talent information according to claim 1 is characterized in that: The method for calculating the bias influence value corresponding to each sequence to be recommended includes: BIV ; Among them, BIV is represented by the bias impact value, is the total number of elements in the sequence to be recommended, For the The gender code corresponding to the target to be recommended, is the average value of the gender encoding of all targets to be recommended, For the The matching degree of the target to be recommended is For the The age of the target to be recommended, is the average age of all targets to be recommended.

8. The intelligent matching management system for talent information according to claim 7 is characterized in that: The natural heuristic algorithm is a sparrow search algorithm. The method for optimizing P sequences to be recommended according to the bias influence value and the natural heuristic algorithm to obtain the optimal recommended sequence includes: S1: Randomly generate a group of sparrow individuals, each sparrow individual represents a sequence to be recommended, and randomly generate an initial position for each sparrow individual ; S2: Construct the fitness function value of each sparrow individual : = ; In the formula, is the bias influence value corresponding to the sequence to be recommended; S3: Perform exploration operations on some sparrow individuals and update their initial sound wave parameters ; = ; In the formula, For the A sparrow in the The position of the generation, For the A sparrow in the The position of the generation, is the maximum number of iterations, is a random angle, It is an exponential function with e as the base, where e is a natural constant; S4: Follow the remaining sparrows and update their initial sound wave parameters ; S5: Calculate the fitness function value of each sparrow individual after the update , update the current optimal solution; S6: Repeat S3-S5 until the preset maximum number of iterations is reached. Stop searching when , and output the recommended sequence corresponding to the current optimal solution, which is the optimal recommended sequence.

9. The intelligent matching management system for talent information according to claim 8 is characterized in that: The following operation is performed on the remaining sparrow individuals to update their initial sound wave parameters The methods include: = ; In the formula, is a random number, is the location of the current optimal solution.

Images