Occupational skill cognition level layering method based on firefly and deep belief network

Abstract

The invention discloses a vocational skill cognition level layering method based on fireflies and a deep belief network. The method comprises the following steps: S1, forming a preprocessed vocational skill data set; s2, constructing an improved deep belief network model, and forming a preliminary deep vocational skill feature expression structure; s3, performing supervised fine-tuning training on the improved deep belief network model based on the preliminary deep vocational skill feature expression structure and the training data containing the vocational skill evaluation label to form an improved deep belief network model subjected to fine-tuning optimization; s4, performing global search and adaptive optimization on key training parameters in the improved deep belief network model by introducing a firefly algorithm, and generating an improved deep belief network model subjected to global parameter optimization; and S5, performing reasoning and classification on the new vocational skill data set by using the improved deep belief network model subjected to global parameter optimization to obtain a vocational skill cognition level layering result of each tested subject. The method is high in occupational skill cognition evaluation ability.

Description

Technical Field

[0001] The present invention relates to the technical field of vocational skills, and particularly to a method for hierarchical classification of vocational skills cognitive level based on fireflies and deep belief networks. Background Art

[0002] With the development of informatization of human resource management, vocational skills assessment has gradually become a key link in enterprise talent selection and function matching. Especially in scenarios such as large-scale job allocation, vocational ability assessment, and employee development path planning, the accurate classification of an individual's vocational skills cognitive level is particularly important. Traditional vocational skills cognitive assessment methods mostly rely on manual interviews, paper-based assessments, or single-dimensional data judgments. For example, rough classification is carried out through surface indicators such as education level, years of work experience, and job experience. Not only is it highly subjective and has limited information dimensions, but it is also difficult to quantify an individual's true performance in complex cognitive tasks.

[0003] Currently, some systems attempt to introduce shallow machine learning methods, such as support vector machines, decision trees, or K-means clustering, to conduct a preliminary classification of vocational abilities on structured assessment data. Although this has improved the processing efficiency to a certain extent, there are the following obvious limitations: First, shallow models are difficult to mine high-order cognitive features in vocational skills data. Especially when facing cross-modal and multi-dimensional information, the feature extraction ability is weak, and the model generalization ability is insufficient. Second, current assessment models generally lack the ability to model the cognitive hierarchical mechanism and cannot reflect the differences and evolution laws between low-order and high-order cognitions. Third, parameter optimization generally adopts greedy or local gradient descent strategies, which are prone to falling into local optima, affecting the stability and accuracy of the final vocational cognitive judgment.

[0004] In addition, when facing new samples, most traditional models adopt a fixed model structure and static parameters, lacking the ability to quickly adapt to new data distributions. Especially in scenarios with diverse job backgrounds or drastic changes in task requirements, they perform poorly. Existing technologies have not fully utilized the latent semantic and cognitive feature correlations in large-scale data, resulting in the lack of hierarchy, interpretability, and strategic guiding significance in the final output of vocational skills cognitive results, and it is difficult to meet the requirements of intelligent talent management systems in refined and humanized assessments.

[0005] In summary, there is an urgent need for a more accurate, adjustable, adaptive, and cognitive hierarchical modeling-capable vocational skills cognitive assessment method to break through the existing limitations. Summary of the Invention

[0006] An object of the present invention is to propose a method for hierarchical classification of vocational skills cognitive level based on fireflies and deep belief networks, and the present invention has a strong vocational skills cognitive assessment ability.

[0007] A method for hierarchical classification of professional skill cognitive levels based on fireflies and deep belief networks according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect various professional skill data related to professional skill assessment to form a raw professional skill data set and perform data preprocessing to form a preprocessed professional skill data set;

[0009] S2. Construct an improved deep belief network model, and use the preprocessed professional skill data set to perform unsupervised pre-training on the improved deep belief network model. Adopt the restricted Boltzmann machine training method layer by layer to extract and construct multi-level representations reflecting the implicit characteristics of professional skills, and form a preliminary deep professional skill feature expression structure;

[0010] S3. Based on the preliminary deep professional skill feature expression structure and the training data containing professional skill assessment labels, perform supervised fine-tuning training on the improved deep belief network model, adjust the network parameters to optimize the discrimination ability of the improved deep belief network model for professional skill cognitive levels, and form an improved deep belief network model optimized by fine-tuning;

[0011] S4. Take the improved deep belief network model optimized by fine-tuning and the preprocessed professional skill data set as inputs, and introduce the firefly algorithm to perform global search and adaptive optimization on the key training parameters in the improved deep belief network model to generate an improved deep belief network model with globally optimized parameters;

[0012] S5. Use the improved deep belief network model with globally optimized parameters to perform inference and classification on the new professional skill data set to obtain the hierarchical classification results of the professional skill cognitive levels of each tested subject.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Collect various professional skill data related to professional skill assessment to construct a raw professional skill data set D raw :

[0015]

[0016] where d i represents the i-th original data entry of the evaluated object, N is the total number of raw professional skill data entries, represents the personal resume information of the i-th evaluated object, including educational background, work experience and skill certificate information, represents the performance evaluation result of the i-th evaluated object, including key performance indicator scores and task completion degrees, represents the score of the i-th evaluated object in the standardized assessment questionnaire, An embedded vector representing job ability requirement description information, used to characterize the ability matching requirements of the current job for this object;

[0017] S12. For the original occupational skill dataset D raw Perform unified format conversion processing. Convert personal resume information, performance evaluation results, test questionnaire scores, and job ability requirement description information into vector structures of fixed dimensions respectively, concatenate them into structured data entries in the unified field order, and maintain field consistency and vector dimension alignment to form a formatted occupational skill dataset with a unified format;

[0018] S13. Perform missing value filling and outlier removal operations on the formatted occupational skill dataset. Fill the missing feature dimensions in all fields with historical means or sample medians, and perform removal or replacement processing on outliers exceeding the set threshold. Delete logically conflicting and unrealistic entries to form a cleaned occupational skill dataset;

[0019] S14. Perform duplicate data merging processing on the cleaned occupational skill dataset. Identify redundant data entries with consistent content or similarity exceeding the set threshold through vector similarity or hash coding. Retain representative entries with complete information and quality higher than the threshold and delete the remaining entries to generate a deduplicated occupational skill dataset;

[0020] S15. Perform standardization processing on the deduplicated occupational skill dataset. Normalize each feature value in all fields to a unified interval [a, b], usually set to the interval [0, 1]. Use a linear mapping formula to transform each dimension of the feature so that the source fields are comparable in numerical scale to form a standardized occupational skill dataset D std 。

[0021] Optionally, the said S2 includes the following steps:

[0022] S21. Based on the standardized occupational skill dataset Construct an improved deep belief network model M′ introducing an occupational skill cognitive modulation factor DBN , the improved deep belief network model M′ DBN Consists of multiple layers of cognitive restricted Boltzmann machines:

[0023] M′ DBN ={CRBM l ∣l=1,2,...,L};

[0024] Among them, L is the total number of network layers, and CRBM l Is the l-th layer of cognitive restricted Boltzmann machine;

[0025] S22. Initialize the parameter set of each layer of cognitive restricted Boltzmann machine as Θ′ l={W l , b l , c l , γ (l)}, where W l is the weight matrix of the l-th layer, b l , c l are the bias vectors of the visible layer and the hidden layer, and γ (l) is the occupational skill cognitive modulation factor of the l-th layer. The occupational skill cognitive modulation factor is initially set according to the prior importance of each feature dimension in the occupational skill cognitive task;

[0026] S23. For each layer l, based on the preprocessed original data entries perform the input, and calculate the occupational skill cognitive feature vector of the i-th object

[0027]

[0028] where σ(·) is the activation function, and ⊙ represents the element-wise modulation operation. For l = 1,

[0029] S24. Reconstruct the current preprocessed original data entries and calculate the residuals. Define the reconstruction error of the l-th layer and use this reconstruction error to update the occupational skill cognitive modulation factor γ (l) :

[0030]

[0031] where η is the learning rate, λ is the cross-layer cognitive feedback coefficient, and Δ (l-1) represents the historical feedback of the residuals of the (l - 1)-th layer on the occupational skill cognition modulation of this layer;

[0032] S25. Repeat steps S23 to S24 until the training of the L-th layer is completed. Finally, fuse the highest-layer cognitive feature vector with the occupational skill cognitive modulation factors of each layer. Define the final cognitive feature vector of the i-th evaluated object as:

[0033]

[0034] where ω (l) is the modulation contribution weight of each layer, reflecting the influence of different cognitive levels on the final representation. Define the preliminary deep occupational skill feature expression structure of the final output as

[0035] Optionally, the S3 includes the following steps:

[0036] S31. Based on the preliminary deep occupational skill feature expression structure Construct a supervised training dataset D with training data containing occupational skill assessment labels sup ;

[0037] S32. Based on the supervised training dataset D sup Perform supervised fine-tuning training on the improved deep belief network model M′ DBN and use the weighted cross-entropy loss function to uniformly model the discrimination ability of various hierarchical labels:

[0038]

[0039] where y i,k represents the true label value of the i-th sample at the k-th cognitive level, represents the predicted output probability of the model, and α k is the importance weight of the k-th cognitive level, reflecting the influence degree of each cognitive level on the learning objective of the model. K represents the number of hierarchical levels of the occupational skill cognitive level;

[0040] S33. During the fine-tuning training process, perform supervised optimization on the model parameter set and use the gradient descent algorithm to update the model parameters based on the weighted cross-entropy loss function The update rule is:

[0041]

[0042] where η is the learning rate, represents the gradient of the total loss function of the model with respect to the parameter set;

[0043] S34. After each round of fine-tuning training, calculate the hierarchical classification accuracy index and the overall hierarchical accuracy index of the improved deep belief network model on the validation set. The overall hierarchical accuracy index is the proportion of the occupational skill cognitive level labels predicted by the improved deep belief network model that are exactly the same as the actual labels among all the evaluated objects, that is, the percentage of the number of samples with the predicted cognitive level matching the true cognitive level in the hierarchical results output by the improved deep belief network model to the total number of samples. The classification accuracy index is used as the basis for judging whether the improved deep belief network model has reached the convergence state. If the classification accuracy index no longer improves significantly in several consecutive training rounds, the fine-tuning process is terminated;

[0044] S35. Repeat steps S32 to S34 until the loss function converges or reaches the set number of training rounds, and finally form the improved deep belief network model optimized by supervised fine-tuning

[0045] Optionally, the S4 includes the following steps:

[0046] S41. Take the improved deep belief network model optimized by supervised fine-tuning and the preprocessed occupational skill dataset D std as the optimization input together;

[0047] S42. Based on the occupational skill cognitive stratification task, design the fitness evaluation function of the firefly algorithm, and define the fitting error between the model output and the true cognitive label as the light intensity function I i , which is used to measure the quality of the i-th individual parameter configuration in the current optimization round. The light intensity function I i is defined as:

[0048]

[0049] where, represents the weighted cross-entropy loss value generated by the model parameter configuration corresponding to the i-th individual on the validation dataset;

[0050] S43. Initialize the firefly population where, Θ i represents a set of key training parameters of the improved deep belief network corresponding to the i-th firefly individual, N f represents the number of the firefly population, and d is the dimension of the parameter space;

[0051] S44. Perform the mutual attraction and position update operations on any two individuals Θ i and Θ j . If the light intensity function I j > I i , then adjust the parameter position according to the following update rule:

[0052]

[0053] where, β0 is the initial attraction coefficient, γ is the light intensity attenuation factor, α is the perturbation control factor, and ∈ i is the perturbation term of the normal distribution;

[0054] S45. After updating all individuals, recalculate the light intensity function I i of each individual, retain the current optimal key training parameters Θ best of the improved deep belief network, and compare the occupational skill cognitive stratification result output by the corresponding improved deep belief network model with the true label, and update the global optimal stratification classification accuracy index and the overall stratification accuracy index;

[0055] S46. Repeat steps S44 to S45 until any of the following convergence conditions is met:

[0056] The global optimal fitness I bestThere is no improvement for several consecutive generations;

[0057] The parameter differences among all individuals converge to the set threshold;

[0058] The maximum number of iterations T is reached max ;

[0059] S47. Output the final improved deep belief network model after global search and adaptive optimization by the firefly algorithm The optimized parameter set corresponding to the improved deep belief network model

[0060] Optionally, the S5 includes the following steps:

[0061] S51. Apply the improved deep belief network model after global parameter optimization by the firefly algorithm to the new vocational skill dataset D new , for each tested subject in the new vocational skill dataset input it into the improved deep belief network model to obtain its vocational skill cognitive feature expression vector

[0062] S53. Based on the prediction probability vector output by the improved deep belief network model determine the vocational skill cognitive level label to which each tested subject belongs

[0063]

[0064] where K is the number of hierarchical levels of vocational skill cognition, is the prediction probability of the k-th cognitive level output by the model;

[0065] S54. Based on the weight parameter matrix corresponding to the output layer in the improved deep belief network model extract the set of key cognitive indicators corresponding to each cognitive level label k∈{1,2,...,K} The weight parameter matrix of the output layer of the model Each row in it corresponds to a specific cognitive level label. In the k-th row, the parameter values respectively correspond to the influence intensity of the input feature dimensions on the classification result of this cognitive level. Sort all the weight parameters in the k-th row according to the absolute value size, and select the top m with the greatest influence kThe feature dimensions, as the key cognitive indicators corresponding to the cognitive level k, each selected feature dimension is converted into an actually recognizable indicator item name according to the occupational skill feature semantic mapping rules defined in the training dataset, and finally forms a set of key cognitive indicators corresponding to the cognitive level label k

[0066] S55. Output the hierarchical results of the occupational skill cognitive levels of each tested subject;

[0067] S56. Summarize the hierarchical results of the occupational skill cognitive levels of all tested subjects and the set of key cognitive indicators, and construct a table of hierarchical results of occupational skill cognitive levels. The table of hierarchical results of occupational skill cognitive levels includes the identity identifier, predicted cognitive level of each tested subject Set of key cognitive indicators Predicted probability vector And visualization explanation information.

[0068] Optionally, the hierarchical results of the cognitive level include:

[0069] Cognitive level I (low-level cognitive level): The classification result where the maximum decision probability is located is Key cognitive indicators include: insufficient job matching degree, low activation degree of the basic knowledge memory dimension, and poor stability of skill operation;

[0070] Cognitive level II (mid-level cognitive level): The classification result where the maximum decision probability is located is Key cognitive indicators include: proficiency in performing some tasks, moderate weight of the problem-solving ability dimension, and fluctuations in knowledge transfer ability;

[0071] Cognitive level III (high-level cognitive level): The classification result where the maximum decision probability is located is Key cognitive indicators include: strong comprehensive task coordination ability, significant response value of the concept abstraction dimension, and job suitability score higher than the set threshold.

[0072] The beneficial effects of the present invention are:

[0073] (1) By introducing an improved deep belief network model with a cognitive modulation mechanism, the present invention significantly improves the expression ability of deep implicit features in occupational skill data. A cognitive modulation factor for occupational skills is introduced in each restricted Boltzmann machine layer, and the feature activation response is adaptively adjusted according to the prior importance and residual feedback of different dimensions, enabling the model to strengthen the ability factors highly related to the cognitive level.

[0074] (2) The present invention combines a supervised fine-tuning mechanism with a deep initial expression structure, achieving a stronger discrimination ability for occupational skill cognitive labels. Through supervised fine-tuning training based on the cross-entropy loss function, the differences between cognitive hierarchical labels are further compressed on the basis of deep features, significantly enhancing the model's discrimination ability for boundary samples. At the same time, a classification accuracy evaluation mechanism in multiple rounds of iteration is introduced to effectively prevent overfitting and improve the generalization of the model.

[0075] (3) The present invention introduces the firefly algorithm into the global search process of improving the key training parameters of the deep belief network, effectively overcoming the defect that traditional gradient optimization methods are prone to falling into local optima. Aiming at the problems of difficult training and sensitive initial parameters of the cognitive hierarchical model in the high-dimensional parameter space, a fitness function is constructed, with the error between the model output and the actual cognitive label as the light intensity function, and the key parameters are dynamically updated through the interactive attraction behavior between populations to achieve an adaptive optimal search process within the global scope. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0077] Figure 1 is a flowchart of a method for hierarchical classification of occupational skill cognitive levels based on fireflies and deep belief networks proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0078] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0079] Refer to Figure 1 , a method for hierarchical classification of occupational skill cognitive levels based on fireflies and deep belief networks, comprising the following steps:

[0080] S1. Collect various occupational skill data related to occupational skill assessment to form an original occupational skill data set and perform data preprocessing to form a preprocessed occupational skill data set;

[0081] S2. Construct an improved deep belief network model, and use the preprocessed occupational skill data set to perform unsupervised pre-training on the improved deep belief network model. Adopt the training method of the restricted Boltzmann machine layer by layer to extract and construct a multi-level representation reflecting the implicit features of occupational skills, and form a preliminary deep occupational skill feature expression structure;

[0082] S3. Based on the preliminary deep vocational skill feature expression structure and the training data containing vocational skill assessment tags, perform supervised fine-tuning training on the improved deep belief network model, adjust the network parameters to optimize the discrimination ability of the improved deep belief network model for the cognitive level of vocational skills, and form an improved deep belief network model optimized by fine-tuning.

[0083] S4. Take the improved deep belief network model optimized by fine-tuning and the preprocessed vocational skill data set as inputs, and through the introduction of the firefly algorithm, perform global search and adaptive optimization on the key training parameters in the improved deep belief network model to generate an improved deep belief network model with globally optimized parameters.

[0084] S5. Use the improved deep belief network model with globally optimized parameters to perform inference and classification on the new vocational skill data set, and obtain the hierarchical results of the cognitive level of vocational skills of each tested subject.

[0085] In this embodiment, S1 includes the following steps:

[0086] S11. Collect various vocational skill data related to vocational skill assessment, and construct the original vocational skill data set D raw :

[0087]

[0088] where d i represents the original data entry of the i-th evaluated object, N is the total number of original vocational skill data entries, represents the personal resume information of the i-th evaluated object, including educational background, work experience and skill certificate information, represents the performance evaluation result of the i-th evaluated object, including key performance indicator scores and task completion degrees, represents the score of the i-th evaluated object in the standardized assessment questionnaire, represents the embedded vector of the job ability requirement description information, which is used to describe the ability matching requirements of the current job for this object;

[0089] S12. Perform unified format conversion processing on the original vocational skill data set D raw Convert the personal resume information, performance evaluation result, assessment questionnaire score and job ability requirement description information into vector structures of fixed dimensions respectively, splice them into structured data entries in the unified field order, and keep the field consistency and vector dimension alignment to form a formatted vocational skill data set with a unified format.

[0090] S13. Perform missing value imputation and outlier removal operations on the formatted vocational skill dataset. For the feature dimensions with missing values in all fields, use historical means or sample medians for imputation. For outliers exceeding the set threshold, perform removal or replacement operations, and delete logically conflicting and unrealistic entries to form the cleaned vocational skill dataset;

[0091] S14. Perform duplicate data merging on the cleaned vocational skill dataset. Identify redundant data entries with consistent content or similarity exceeding the set threshold through vector similarity or hash encoding. Retain the representative entries with complete information and quality higher than the threshold and delete the remaining entries to generate the deduplicated vocational skill dataset;

[0092] S15. Perform standardization on the deduplicated vocational skill dataset. Normalize the feature values in all fields to a unified interval [a, b], usually set as the interval [0, 1]. Use the linear mapping formula to transform each dimension of the feature to make the source fields comparable in numerical scale, forming the standardized vocational skill dataset D std 。

[0093] In this embodiment, S2 includes the following steps:

[0094] S21. Based on the standardized vocational skill dataset Construct an improved deep belief network model M' introducing a vocational skill cognitive modulation factor DBN , the improved deep belief network model M' DBN Consists of multiple layers of cognitive restricted Boltzmann machines:

[0095] M' DBN ={CRBM l |l = 1, 2,..., L};

[0096] Among them, L is the total number of network layers, and CRBM l Is the l-th layer of cognitive restricted Boltzmann machine;

[0097] S22. Initialize the parameter set of each layer of cognitive restricted Boltzmann machine as Θ' l ={W l , b l , c l , γ (l)}, where W l Is the weight matrix of the l-th layer, b l , c l Are the bias vectors of the visible layer and the hidden layer, and γ (l) Is the vocational skill cognitive modulation factor of the l-th layer, and the vocational skill cognitive modulation factor is initially set according to the prior importance of each feature dimension in the vocational skill cognitive task;

[0098] S23. For each layer l, based on the preprocessed original data entries perform input, and calculate the vocational skill cognitive feature vector of the i-th object

[0099]

[0100] where, σ(·) is the activation function, and ⊙ represents the element-wise modulation operation. When l = 1,

[0101] S24. Reconstruct the current preprocessed original data entries and calculate the residuals, define the reconstruction error of the l-th layer and use this reconstruction error to update the vocational skill cognitive modulation factor γ (l) :

[0102]

[0103] where, η is the learning rate, λ is the cross-layer cognitive feedback coefficient, and Δ (l-1) represents the historical feedback of the residuals of the (l - 1)-th layer on the vocational skill cognition modulation of this layer;

[0104] S25. Repeat steps S23 to S24 until the training of layer l = L is completed. Finally, fuse the highest-layer cognitive feature vector with the vocational skill cognitive modulation factors of each layer, and define the final cognitive feature vector of the i-th evaluated object as:

[0105]

[0106] where, ω (l) is the modulation contribution weight of each layer, reflecting the influence of different cognitive levels on the final representation. Define the preliminary deep vocational skill feature expression structure of the final output as

[0107] In this embodiment, S3 includes the following steps:

[0108] S31. Based on the preliminary deep vocational skill feature expression structure and the training data containing vocational skill assessment labels, construct a supervised training dataset D sup ;

[0109] S32. Based on the supervised training dataset D sup perform supervised fine-tuning training on the improved deep belief network model M′ DBN and use the weighted cross-entropy loss function to uniformly model the discriminative ability of various hierarchical labels:

[0110]

[0111] where, yi,k represents the true label value of the \(i\)-th sample at the \(k\)-th cognitive level, represents the predicted output probability of the model, \(\alpha\) k is the importance weight of the \(k\)-th cognitive level, reflecting the influence degree of each cognitive level on the learning objective of the model, and \(K\) represents the number of hierarchical levels of the vocational skill cognitive level;

[0112] S33. During the fine-tuning training process, perform supervised optimization on the model parameter set and update the model parameters using the gradient descent algorithm based on the weighted cross-entropy loss function The update rule is:

[0113]

[0114] where \(\eta\) is the learning rate, represents the gradient of the total loss function of the model with respect to the parameter set;

[0115] S34. After each round of fine-tuning training, calculate the hierarchical classification accuracy index and the overall hierarchical accuracy index of the improved deep belief network model on the validation set. The overall hierarchical accuracy index is the proportion of the vocational skill cognitive level labels predicted by the improved deep belief network model that are exactly the same as the actual labels among all the evaluated objects, that is, the percentage of the number of samples with the predicted cognitive level matching the true cognitive level in the hierarchical results output by the improved deep belief network model to the total number of samples. The classification accuracy index is used as the basis for judging whether the improved deep belief network model reaches the convergence state. If the classification accuracy index no longer improves significantly in several consecutive training rounds, terminate the fine-tuning process;

[0116] S35. Repeat steps S32 to S34 until the loss function converges or reaches the set number of training rounds, and finally form the improved deep belief network model optimized by supervised fine-tuning

[0117] In this embodiment, S4 includes the following steps:

[0118] [[ID=3 / 5]]S41. Use the improved deep belief network model optimized by supervised fine-tuning and the preprocessed vocational skill dataset \(D\) std as the optimization input together;

[0119] S42. Design the fitness evaluation function of the firefly algorithm based on the vocational skill cognitive hierarchical task, and define the fitting error between the model output and the true cognitive label as the light intensity function \(I\) i , which is used to measure the quality of the \(i\)-th individual parameter configuration in the current optimization round. The light intensity function \(I\) i is defined as:

[0120]

[0121] Among them, represents the weighted cross-entropy loss value generated by the model parameter configuration corresponding to the i-th individual on the validation dataset;

[0122] S43. Initialize the firefly population Among them, Θ i represents a set of key training parameters for the improved deep belief network corresponding to the i-th firefly individual, and N f represents the number of the firefly population, and d is the dimension of the parameter space;

[0123] S44. For any two individuals Θ i and Θ j perform the mutual attraction and position update operations. If the light intensity function I j > I i , then adjust the parameter position according to the following update rules:

[0124]

[0125] Among them, β0 is the initial attraction coefficient, γ is the light intensity attenuation factor, α is the perturbation control factor, and ∈ i is the perturbation term of the normal distribution;

[0126] S45. After updating all individuals, recalculate the light intensity function I i for each individual, retain the current optimal key training parameters Θ best of the improved deep belief network, and compare the occupational skill cognitive stratification results output by the corresponding improved deep belief network model with the true labels to update the global optimal stratification classification accuracy index and the overall stratification accuracy index;

[0127] S46. Repeat steps S44 to S45 until any of the following convergence conditions are met:

[0128] The global optimal fitness I best no longer improves for several consecutive generations;

[0129] The parameter differences among all individuals converge to the set threshold;

[0130] The maximum number of iterations T max is reached;

[0131] S47. Output the final improved deep belief network model after the global search and adaptive optimization of the firefly algorithm The optimized parameter set corresponding to the improved deep belief network model

[0132] In this embodiment, S5 includes the following steps:

[0133] S51. Apply the improved deep belief network model optimized by the global parameters of the firefly algorithm to the new vocational skill dataset D new , and for each tested subject in the new vocational skill dataset input it into the improved deep belief network model to obtain its vocational skill cognitive feature expression vector

[0134] S53. Based on the prediction probability vector output by the improved deep belief network model determine the vocational skill cognitive level label to which each tested subject belongs where K is the number of hierarchical levels of vocational skill cognitive levels,

[0135]

[0136] is the prediction probability of the k-th cognitive level output by the model;

[0137] S54. Based on the weight parameter matrix corresponding to the output layer in the improved deep belief network model extract the set of key cognitive indicators corresponding to each cognitive level label k ∈ {1, 2,..., K} Each row in the weight parameter matrix of the output layer of the model corresponds to a specific cognitive level label. In the k-th row, the parameter values respectively correspond to the influence intensity of the input feature dimensions on the classification result of this cognitive level. Sort all the weight parameters in the k-th row according to their absolute value sizes, and select the top m feature dimensions with the greatest influence as the key cognitive indicators corresponding to the cognitive level k. Each selected feature dimension is converted into the name of an actually recognizable index item according to the vocational skill feature semantic mapping rules defined in the training dataset, and finally the set of key cognitive indicators corresponding to the cognitive level label k is formed k

[0138] S55. Output the vocational skill cognitive level stratification results of each tested subject;

[0139] S56. Summarize the vocational skill cognitive level stratification results and the set of key cognitive indicators of all tested subjects, and construct a vocational skill cognitive level stratification result table. The vocational skill cognitive level stratification result table includes the identity identifier, predicted cognitive level set of key cognitive indicators prediction probability vector​​​ and visual interpretation information.

[0140] In this embodiment, the cognitive level stratification results include:

[0141] Cognitive level I (low-level cognitive level): The classification result where the maximum decision probability lies is The key cognitive indicators include: insufficient job matching degree, low activation degree of the basic knowledge memory dimension, and poor stability of skill operation;

[0142] Cognitive level II (mid-level cognitive level): The classification result where the maximum decision probability lies is The key cognitive indicators include: proficiency in performing some tasks, medium weight of the problem-solving ability dimension, and fluctuations in knowledge transfer ability;

[0143] Cognitive level III (high-level cognitive level): The classification result where the maximum decision probability lies is The key cognitive indicators include: strong comprehensive task coordination ability, significant response value of the concept abstraction dimension, and job suitability score higher than the set threshold.

[0144] Example 1:

[0145] In the intelligent human resource management platform of Company A Limited, a hierarchical analysis process task request was automatically triggered. The request number is [RQS-20241114-00312], the request source is the company's human resources department minister, and the task objective is "to intelligently identify and stratify the professional skill cognitive levels of 286 front-line core employees in the technology department, quality inspection department, and production line process department, and provide a basis for the job rotation and performance incentive policy grading in December".

[0146] The following key information is listed in the analysis request form filled in the platform:

[0147] Data source: HRMS system + quarterly performance form + online assessment platform;

[0148] Analysis time window: Q2 to Q3 in 2024;

[0149] Stratification level requirement: Divided into three layers based on cognitive characteristics;

[0150] Purpose of stratification: Job competency ranking + skill matching warning;

[0151] Output required: level label, probability value, key cognitive indicators, and interpretability report.

[0152] The system then started the "MDBN-FA intelligent cognitive modeling process" corresponding to the method of the present invention at 08:36, with the process number [MDBNFA-P-20241114-A01].

[0153] The data extraction process is completed, and a total of 286 data entries are retrieved, including:

[0154] There are 286 pieces of resume data (education background, qualification certificates, project resume), with 29 field dimensions;

[0155] There are 286 pieces of performance scores (quarterly task achievement rate, process quality consistency), with 15 dimensions;

[0156] The scores of the standard assessment papers (cognitive reasoning test, operation step simulation test) were all implemented in October 2024;

[0157] The job skill requirements embedding codes are retrieved from the HR system and vectors are automatically generated through the BERT model.

[0158] After preprocessing, the system completes the standardized feature fusion, with the data dimension of 78 dimensions, forming a dataset.

[0159] The system automatically starts the training process of the improved deep belief network MDBN. The model is composed of 3 layers of CRBM, and each layer introduces a cognitive modulation factor, where:

[0160] The first layer focuses on capturing the stability of task execution and proficiency in the standardized process;

[0161] The second layer emphasizes problem-solving ability and knowledge transfer performance;

[0162] The third layer integrates high-order abstract logical features and job adaptation semantic scores.

[0163] The training data is divided in a ratio of 8:2. The training set has 229 people, and the test set has 57 people.

[0164] The model completes pre-training and starts the supervised fine-tuning with the weighted cross-entropy function. After the fine-tuning is completed, the model reaches the following performance indicators: hierarchical accuracy: 91.5%; high-order cognitive level recognition accuracy: 89.4; Rand Index hierarchical stability index: 0.943.

[0165] Subsequently, the system calls the "Firefly-Opt module" to execute the parameter optimization process based on the firefly algorithm, with the population size Nf = 20 and the maximum number of iterations Tmax = 50.

[0166] In the 38th iteration, the fitness function reaches the optimum, and the optimized parameter Θ_best is output and stored in

/ hrmodel / mdbnfa_opt / param_v38.json

[0167] The system began performing inference classification tasks on 286 employees, and the model output for each input entry:

[0168] Cognitive level label (I / II / III);

[0169] Level confidence probability vector ([0.02, 0.31, 0.67]);

[0170] Key cognitive indicators that lead to grade classification (task coordination ability, job suitability score, and skill transferability);

[0171] Explainability report PDF.

[0172] After all the reasoning tasks are completed, the system generates hierarchical result tables, classification reports, grade distribution charts and employee personal PDF reports in the directory.

[0173] Take employee number EMP-114 as an example:

[0174] Basic information: College degree, process team leader, 7 years of employment;

[0175] Resume embedding score: job resume coverage rate 0.74;

[0176] Assessment score: Logical reasoning score 78 / 100, operational consistency average score 92;

[0177] Model output: Cognitive level = III, probability of higher-order cognition = 0.91;

[0178] Key identification indicators: Job suitability score = 0.93, Abstract dimension response value = 1.37 × standard deviation, Cross-task transfer performance is ranked in the top 10% within the group.

[0179] The system interpretation results pointed out that it "has the ability to control complex and multiple process parameters in a coordinated manner, can quickly build emergency step logic in the event of sudden process changes, has a fast response speed, and a low operating error rate." The conclusion was reviewed and approved by the supervisor and was used to recommend it to a newly added product collaborative development position.

[0180] To verify the accuracy and effectiveness of this method, the HR department conducted a comparative test using both a traditional manual method and a shallow model method (Multi-layer Perceptron MLP). The test set included 57 employees. The comparison results are as follows:

[0181] Table 1 Employee data of the present invention, traditional manual method and shallow model method

[0182]

[0183]

[0184] Table 2 Comparison of the overall prediction accuracy between the present invention, the traditional manual method, and the shallow model method

[0185] Method Accuracy High-order recall rate Interpretability score (HR score) Traditional manual scoring method 73.6％ 65.3％ Medium MLP shallow model 82.1％ 71.4％ Low The present invention (MDBN+FA) 92.8％ 88.6％ High

[0186] Through the implementation of the first batch of job adjustments based on the output results of this system, a total of 37 people were transferred. In the subsequent performance tracking, 86.4% of the transferred personnel were rated as "good" or above for two consecutive months in the new positions. The internal research of HR showed that the supervisor satisfaction increased to 92%.

[0187] By introducing an improved deep belief network model with a cognitive modulation mechanism, the present invention significantly improves the expression ability of deep implicit features in vocational skill data. A vocational skill cognitive modulation factor is introduced into each restricted Boltzmann machine layer, and the feature activation response is adaptively adjusted according to the prior importance and residual feedback of different dimensions, enabling the model to strengthen the ability factors highly related to the cognitive level.

[0188] The present invention combines a supervised fine-tuning mechanism with a deep initial expression structure to achieve a stronger discrimination ability for vocational skill cognitive labels. Through supervised fine-tuning training based on the cross-entropy loss function, the differences between cognitive hierarchical labels are further compressed on the basis of deep features, significantly enhancing the model's discrimination ability for boundary samples. At the same time, a classification accuracy evaluation mechanism in multiple rounds of iteration is introduced to effectively prevent overfitting and improve the model's generalization ability.

[0189] The present invention introduces the firefly algorithm into the global search process of the key training parameters of the improved deep belief network, effectively overcoming the defect that the traditional gradient optimization method is prone to falling into local optima. A fitness function is constructed for the problems of high training difficulty and initial parameter sensitivity of the cognitive hierarchical model in the high-dimensional parameter space. Taking the error between the model output and the actual cognitive label as the light intensity function, the key parameters are dynamically updated through the interactive attraction behavior between populations to achieve the adaptive optimal search process within the global scope.

[0190] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A hierarchical method for the cognitive level of professional skills based on fireflies and deep belief networks, characterized in that, It includes the following steps: S1. Collect various vocational skill data related to vocational skill assessment to form an original vocational skill data set and perform data preprocessing to form a preprocessed vocational skill data set; S2. Construct an improved deep belief network model, and use the preprocessed vocational skill data set to perform unsupervised pre-training on the improved deep belief network model. Adopt the restricted Boltzmann machine training method layer by layer to extract and construct multi-level representations reflecting the implicit characteristics of vocational skills, and form a preliminary deep vocational skill feature expression structure; S3. Based on the preliminary deep vocational skill feature expression structure and the training data containing vocational skill assessment labels, perform supervised fine-tuning training on the improved deep belief network model, adjust the network parameters to optimize the discrimination ability of the improved deep belief network model for the cognitive level of vocational skills, and form an improved deep belief network model optimized by fine-tuning; S4. Use the improved deep belief network model optimized by fine-tuning and the preprocessed vocational skill data set as inputs, and introduce the firefly algorithm to globally search and adaptively optimize the key training parameters in the improved deep belief network model to generate an improved deep belief network model with globally optimized parameters; S5. Use the improved deep belief network model with globally optimized parameters to perform inference and classification on the new vocational skill data set to obtain the hierarchical results of the vocational skill cognitive levels of each tested subject.

2. The hierarchical method for professional skill cognitive level based on fireflies and deep belief network according to claim 1, characterized in that The S1 includes the following steps: S11. Collect various types of vocational skill data related to vocational skill assessment and construct an original vocational skill dataset D raw ; S12. Perform unified format conversion, missing value filling, outlier removal, duplicate data merging, and data standardization on the original vocational skill dataset D raw to form a standardized vocational skill dataset D std .

3. A hierarchical method for cognitive level of professional skills based on fireflies and deep belief networks according to claim 1, characterized in that, The S2 includes the following steps: S21. Based on the standardized vocational skill dataset Construct an improved deep belief network model M' by introducing a vocational skill cognitive modulation factor DBN , the improved deep belief network model M' DBN is composed of multiple layers of cognitive restricted Boltzmann machines: M′ DBN ={CRBM l | l = 1, 2, ..., L}; where L is the total number of network layers, and CRBM l is the l-th layer of the Cognitive Restricted Boltzmann Machine; S22. Initialize the parameter set of each layer of the cognitive restricted Boltzmann machine as Θ l ′ ={W l , b l , c l , γ (l)}, where W l is the weight matrix of the l-th layer, b l , c l are the bias vectors of the visible layer and the hidden layer, and γ (l) is the occupational skill cognitive modulation factor of the l-th layer. The occupational skill cognitive modulation factor is initially set according to the prior importance of each feature dimension in the occupational skill cognitive task; S23. For each layer l, based on the preprocessed original data entry d i (1) perform an input and calculate the vocational skill cognitive feature vector of the i-th object where σ(·) is the activation function, N is the total number of original entries of occupational skills data, ⊙ represents the element-wise modulation operation. For l = 1, S24. Reconstruct the original data entry after current preprocessing and calculate the residual, and define the reconstruction error of the l-th layer And use this reconstruction error to update the vocational skill cognitive modulation factor γ (l) : where η is the learning rate, λ is the cross-layer cognitive feedback coefficient, and Δ (l-1) represents the historical feedback of the residual of the (l-1)-th layer on the cognitive modulation of the professional skills of this layer; S25. Repeat steps S23 to S24 until the training of layer l = L is completed. Finally, fuse the highest-level cognitive feature vector with the vocational skill cognitive modulation factors of each layer, and define the final cognitive feature vector of the i-th evaluated object as: Among them, ω (l) is the contribution weight of each layer of modulation, reflecting the influence of different cognitive levels on the final representation. Define the preliminary deep professional skill feature expression structure of the final output as 4. A hierarchical method for cognitive level of professional skills based on fireflies and deep belief network according to claim 3, characterized in that The S3 includes the following steps: S31. Based on the preliminary deep career skill feature expression structure and the training data containing career skill assessment labels, construct a supervised training dataset D sup , where represents the standardized data entry of the i-th evaluated object; S32. Based on the supervised training dataset D sup Perform supervised fine-tuning training on the improved deep belief network model M′ DBN using the weighted cross-entropy loss function to uniformly model the discrimination ability of various hierarchical labels; S33. During the fine-tuning training process, for the set of model parameters perform supervised optimization, and use the gradient descent algorithm to update the model parameters based on the weighted cross-entropy loss function The update rule is: where η is the learning rate, represents the gradient of the total loss function of the model with respect to the parameter set; S34. After each round of fine-tuning training, calculate the hierarchical classification accuracy index and the overall hierarchical accuracy index of the improved deep belief network model on the validation set. The overall hierarchical accuracy index is the proportion of the vocational skill cognitive level labels predicted by the improved deep belief network model that are exactly the same as the actual labels among all evaluated objects, that is, the percentage of the number of samples with the predicted cognitive level matching the true cognitive level in the hierarchical results output by the improved deep belief network model in the total number of samples. The classification accuracy index is used as the basis for judging whether the improved deep belief network model has reached the convergence state. If the classification accuracy index no longer improves significantly in several consecutive training rounds, terminate the fine-tuning process; S35. Repeat steps S32 to S34 until the loss function converges or the set number of training rounds is reached, and finally an improved deep belief network model optimized by supervised fine-tuning is formed 5. A hierarchical method for professional skill cognitive level based on fireflies and deep belief network according to claim 4, characterized in that, The S4 includes the following steps: S41. Use the improved deep belief network model optimized by supervised fine-tuning and the preprocessed vocational skill dataset D std as the optimization input together; S42. Design the fitness evaluation function of the firefly algorithm based on the hierarchical tasks of vocational skills cognition, and define the fitting error between the model output and the true cognition label as the light intensity function I i , which is used to measure the quality of the parameter configuration of the i-th individual in the current optimization round; S43. Initialize the firefly population where Θ i represents a set of key training parameters of the improved deep belief network corresponding to the i-th firefly individual, N f represents the number of the firefly population, and d is the dimension of the parameter space; S44. For any two individuals Θ i and Θ j perform mutual attraction and position update operations. If the light intensity function I j >I i , then adjust the parameter positions according to the following update rules: where β0 is the initial attraction coefficient, γ is the light intensity attenuation factor, α is the perturbation control factor, and ∈ i is a perturbation term of normal distribution; After updating all individuals, recalculate the light intensity function I of each individual i , retain the key training parameters Θ of the current optimal improved deep belief network best , and compare the occupational skill cognitive stratification results output by its corresponding improved deep belief network model with the true labels, and update the global optimal stratification classification accuracy index and the overall stratification accuracy index; S46. Repeat steps S44 to S45 until any of the following convergence conditions are met: Global optimal fitness I best It has not been improved for several consecutive generations; The parameter differences among all individuals converge to the set threshold; Reach the maximum number of iterations T max ; S47. Output the final improved deep belief network model after global search and adaptive optimization by the firefly algorithm The optimized parameter set corresponding to the improved deep belief network model 6. A hierarchical method for cognitive level of professional skills based on fireflies and deep belief networks according to claim 5, characterized in that The S5 includes the following steps: S51. Apply the improved deep belief network model optimized by the global parameters of the firefly algorithm to the new vocational skill dataset D new , and for each tested subject in the new vocational skill dataset input it into the improved deep belief network model to obtain its vocational skill cognitive feature expression vector S53. Based on the improved deep belief network model The predicted probability vector output Determine the occupational skill cognitive level label to which each subject under test belongs S54. Based on the improved deep belief network model The weight parameter matrix corresponding to the output layer Extract the set of key cognitive indicators corresponding to each cognitive level label k∈{1,2,...,K} S55. Output the hierarchical results of the vocational skill cognitive levels of each tested subject; Summarize the hierarchical results of the occupational skill cognitive levels and the set of key cognitive indicators of all the tested subjects, and construct a table of hierarchical results of occupational skill cognitive levels. The table of hierarchical results of occupational skill cognitive levels includes the identity identifier of each tested subject and the predicted cognitive level Set of key cognitive indicators Predicted probability vector And visualization explanation information.

7. A method for hierarchical classification of professional skill cognitive levels based on fireflies and deep belief networks according to claim 6, characterized in that, The hierarchical results of the cognitive level include: Cognitive level I (low-level cognitive level): The classification result where the maximum decision probability lies is The key cognitive indicators include: insufficient job matching, low activation degree in the dimension of basic knowledge memory, and poor stability in skill operation; Cognitive level II (intermediate cognitive level): The classification result where the maximum decision probability lies is The key cognitive indicators include: proficiency in performing some tasks, a medium weight in the problem-solving ability dimension, and fluctuations in knowledge transfer ability; Cognitive Level III (High-Order Cognitive Level): The classification result with the maximum decision probability is The key cognitive indicators include: strong comprehensive task coordination ability, significant response value in the concept abstraction dimension, and job suitability score higher than the set threshold.

Images