Occupational skill interest cultivation method based on leapfrog optimization and sparse automatic encoder

Abstract

The invention discloses an occupational skill interest cultivation method based on leapfrog optimization and a sparse automatic encoder. The method comprises the following steps: S1, obtaining a standardized multi-source user behavior data set; s2, establishing an initial sparse automatic encoder model to obtain a preliminary user deep interest feature vector set; s3, constructing a leapfrog optimization module to obtain an optimized sparse automatic encoder model; s4, establishing a user-skill similarity matrix, and generating a personalized vocational skill interest cultivation path candidate set according to the user-skill similarity matrix; and S5, pushing the updated personalized vocational skill interest cultivation path set to a user side and continuously monitoring an execution effect to complete a vocational skill interest cultivation closed loop. According to the invention, the personalized response capability of the professional skill cultivation system and the learning satisfaction degree of the user are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of vocational skills technology, and in particular to a method for cultivating vocational skills interest based on frog-jump optimization and sparse autoencoders. Background Technology

[0002] With the development of artificial intelligence and optimization algorithms, the field of vocational skills training is gradually introducing intelligent methods for personalized learning path recommendation and user interest analysis. As an important part of skills improvement and career development, vocational skills interest cultivation is of great significance for accurately identifying users' potential interests and dynamically matching training content. However, in traditional vocational training systems, the identification of individual users' interests and the formulation of skills cultivation paths mainly rely on questionnaires, expert evaluations, or simple statistical modeling based on shallow features. These methods have significant limitations.

[0003] Currently, most existing vocational skills training platforms employ static rule matching or recommendation methods based on surface data clustering, which lack effective mining of users' deep-seated interests and potential skill states. Because these methods often cannot handle high-dimensional, multi-source, and heterogeneous user behavior data, and are easily interfered with by data noise and redundant information during feature extraction, the expression of users' individual characteristics is insufficient, and the recommendation results tend to be homogeneous, making it difficult to achieve truly precise and differentiated training. In addition, in traditional models, the system response is slow when facing dynamic changes in user behavior, and the model structure and parameter updates lag behind, making it unable to adapt to changes in user interests or skill development stages in a timely manner, thus making the recommendation path outdated and the level of personalization declining.

[0004] In recent years, some studies have introduced deep learning methods, such as autoencoders, for feature extraction. However, existing technologies mostly remain at the stage of fixed training and fixed parameter application, lacking optimization mechanisms tailored to the specific needs of cultivating vocational skills and interests. On the one hand, traditional sparse autoencoders are prone to overfitting or sparsity out of control in high-noise data environments, resulting in insufficient model interpretability and generalization ability. On the other hand, the model optimization process generally lacks a dynamic optimization strategy that combines global search and local adaptation, making it easy to get stuck in local optima during training, affecting the accuracy of the final interest feature extraction and skill matching.

[0005] Therefore, there is an urgent need for a new method that can deeply extract user interest features from massive behavioral data and generate dynamic adaptive training paths through intelligent optimization, in order to solve the problems of insufficient accuracy, lagging updates, and limited recommendations in existing vocational skills interest training systems. Summary of the Invention

[0006] One objective of this invention is to propose a method for cultivating vocational skills interest based on frog-jump optimization and sparse autoencoders. This invention significantly improves the personalized responsiveness of the vocational skills training system and user learning satisfaction.

[0007] A method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source user behavior and basic information, construct a multi-source user behavior dataset, and preprocess it to obtain a standardized multi-source user behavior dataset;

[0009] S2. Establish an initial sparse autoencoder model and train it on a standardized multi-source user behavior dataset to obtain a preliminary set of user deep interest feature vectors;

[0010] S3. Construct a frog-jump optimization module, and iteratively optimize the initial sparse autoencoder model through the frog-jump optimization module to obtain an optimized sparse autoencoder model;

[0011] S4. Utilize an optimized sparse autoencoder model to re-encode the standardized multi-source user behavior dataset, obtain an optimized user deep interest feature vector set, combine the optimized user deep interest feature vector set with the skill module feature vector library, establish a user-skill similarity matrix, and generate a personalized vocational skill interest cultivation path candidate set based on the user-skill similarity matrix.

[0012] S5. During the process of users executing the candidate set of personalized career skills and interest cultivation paths, real-time interaction log data is collected to form a dynamic feedback data stream. The dynamic feedback data stream is input into the optimized sparse autoencoder model for incremental learning, and the updated optimized sparse autoencoder model and the updated set of personalized career skills and interest cultivation paths are output. The updated set of personalized career skills and interest cultivation paths is pushed to the user terminal and the execution effect is continuously monitored to complete the closed loop of career skills and interest cultivation.

[0013] Optionally, S1 includes the following steps:

[0014] S11. Collect multi-source user behavior and basic information, including user course click behavior, course browsing duration, test answer records, historical training feedback scores, and interest assessment results, and construct a multi-source user behavior dataset D. raw :

[0015]

[0016] Where, d i Let t represent the data vector of the i-th user. i b represents the timestamp of user data records.i This represents a user's course browsing behavior vector, including click frequency and browsing duration features, f i s represents the feature vector of a user's historical feedback score. i q represents the feature vector of a user's skill test score. i This represents the feature vector of the user's interest questionnaire results, where N is the total number of users.

[0017] S12. Perform format unification and structure normalization processing on the multi-source user behavior dataset, embed heterogeneous feature vectors from different sources in a unified manner, normalize features of different dimensions, and perform noise removal and missing data completion processing. For data dimensions with abnormal fluctuations or missing data, use local weighted interpolation or sliding window averaging to obtain the noise-removed multi-source user behavior dataset.

[0018] S13. Perform redundant feature elimination processing on the noise-removed multi-source user behavior dataset, and construct the feature correlation matrix R based on correlation analysis:

[0019]

[0020] Among them, R i,j Let x represent the Pearson correlation coefficient between the i-th and j-th dimensions. k,i This represents the feature value of the k-th user in the i-th dimension. Let represent the average value of the i-th dimension across all users. If the Pearson correlation coefficient is greater than a preset value, then the i-th and j-th dimensions are considered to have high redundancy. Based on preset rules, the dominant dimension is retained, and redundant dimensions are removed, resulting in a standardized multi-source user behavior dataset D. norm .

[0021] Optionally, S2 includes the following steps:

[0022] S21. Let the standardized multi-source user behavior dataset be... Where z i This represents the standardized feature vector of the i-th user;

[0023] S22. Construct a mapping matrix of occupational skill categories K represents the number of occupational skill categories, M ′ Using the feature dimension, we obtain the skill category feature vector c for each user. i =Cz i ;

[0024] S23. Construct a domain constraint mask matrix M based on the occupational skill category mapping matrix C. The domain constraint mask matrix is ​​used to define the connection relationship between each node of the hidden layer and the input feature dimension, so that each hidden node is only connected to the input dimension feature of its corresponding skill category, thus forming a structured skill perception connection pattern.

[0025] M = [m jl ] H×M′ ;

[0026]

[0027] Where H is the number of hidden layer nodes, m jl This indicates whether the j-th node in the hidden layer is connected to the l-th feature in the input feature dimension. Its value is either 0 or 1, where 1 indicates a connection and 0 indicates no connection. C kl This indicates whether there is a semantic relationship between the k-th skill category and the l-th input feature in the occupational skill category mapping matrix. j≡k(mod H / K) means that the hidden layer nodes are divided equally according to the number of occupational skill categories K, with each category corresponding to H / K hidden nodes, and structural constraints are established so that the j-th hidden node is only connected to the input feature dimension related to its corresponding category k.

[0028] S24. Input the sparse autoencoder into the weight matrix W. (1) Element-wise multiplication with the domain constraint mask matrix M yields the weighted input weight matrix.

[0029] S25. Construct the user interest and attention matrix A with the standardized feature vector z of the i-th user. i Attention-weighted calculations are performed to generate the interest attention vector γ. i Each component of the interest attention vector is used to measure the degree of attention that the corresponding node in the hidden layer pays to the user.

[0030] γ i =softmax(Az) i );

[0031] S26. Calculate the sparse coding representation of the hidden layer based on the domain-constrained mask matrix and the interest attention matrix:

[0032]

[0033] Where σ(·) is the Sigmoid activation function, b (1) This is the hidden layer bias vector;

[0034] S27. Based on the sparse coding representation of the hidden layer h i,jFor each node in the hidden layer, calculate its average activation value across all user samples. Where j represents the hidden layer node number;

[0035] S28. Constructing the comprehensive loss function

[0036] S29. The gradient descent algorithm is used to iteratively optimize the comprehensive loss function. During training, the weighted input weight matrix, output weight matrix, hidden layer bias vector and output bias vector are dynamically adjusted, and the learning rate is decayed according to the current iteration number u.

[0037] S210. When the convergence result of the comprehensive loss function is lower than the set error threshold, output the preliminary user deep interest feature vector set F. init Each vector h in the initial set of deep user interest feature vectors i This represents the deeper expression of the i-th user in the dimension of career skills and interests.

[0038] Optionally, the comprehensive loss function includes: a reconstruction error term, used to measure the difference between the reconstructed vector and the original input vector; a KL divergence-based term, used to measure the deviation of the average activation value of the hidden layer nodes from the target sparsity ρ, adjusting the sparsity of the hidden layer representation; and an interest separation regularization term, used to encourage good distinguishability of interest representations among different users in the semantic space, improving the model's ability to express differences.

[0039]

[0040] in, To reconstruct the vector, W (2) To output the weight matrix, b (2) is the output bias vector, β is the sparse regularization coefficient, ρ is the target sparsity, λ is the contrast regularization coefficient, and Δ is the interest separation threshold.

[0041] Optionally, S3 includes the following steps:

[0042] S31. Construct a frog-jumping optimization module to optimize the initial sparse autoencoder model. Set the total size of the frog swarm to P, and define each frog as a candidate sparse autoencoder model containing the input weight matrix W. (1) Output weight matrix W (2) And user interest attention matrix A;

[0043] S32. Divide the frog swarm into m subgroups, each subgroup containing several individuals. For each candidate sparse autoencoder model, construct a multi-objective fitness function F based on the generated preliminary user deep interest feature vector set.

[0044] S33. For each subgroup, sort the individuals in the subgroup from best to worst according to the multi-objective fitness function F, and select the individual with the best multi-objective fitness function value as the best individual of that subgroup. The individual with the worst multi-objective fitness function value is selected as the worst individual in the subgroup. A jump vector is constructed based on the difference between the two, the jump magnitude is scaled by the jump factor r, and the parameters of the worst individual are updated.

[0045] S34. When the updated individual using the jump vector is better than the original worst individual in the multi-objective fitness function value, the updated individual replaces the worst individual. If the update does not bring about optimization, a new set of individuals with Gaussian perturbation is generated according to the structural perturbation principle and used to replace the worst individual in the subgroup.

[0046] S35. After all subgroups have executed the local jump strategy, select the individual with the best multi-objective fitness function value from the entire frog swarm as the global elite individual. The core model parameters of the global elite individual will be injected into the individual with the worst fitness in all subgroups to build an elite-driven global accelerated convergence mechanism.

[0047] S36. Combine the user deep interest feature vectors output by the encoders in all current subgroups to calculate the average interest representation mutual information between different subgroups, which is defined as the global interest diversity index Ψ. When the global interest diversity index is lower than the preset diversity threshold, it indicates that the global search has shown a trend of convergence and concentration, and the system automatically triggers the jump perturbation enhancement mechanism.

[0048] S37. The process of local hopping in the subgroup, global elite propagation, and interest diversity regulation will be continuously iterated until one of the following termination conditions is met: one of which is that the number of iterations reaches the maximum number of iterations T. max Secondly, the multi-objective fitness function value of the globally optimal individual converges to below the error threshold ∈.

[0049] S38. The final output is the set of individual parameters with the best global fitness, which includes the optimal input weight matrix W. (1)* Optimal output weight matrix W (2)* With the optimal user interest attention matrix A * Individuals are used to optimize the sparse autoencoder model, resulting in an optimized sparse autoencoder model.

[0050] Optionally, the multi-objective fitness function includes a reconstruction error term, which evaluates the ability of the candidate sparse autoencoder model to retain the original user input feature information, and is measured by the mean square error between each user's input vector and the vector reconstructed by the sparse autoencoder; an interest feature separation term, which measures the discriminative power of the deep interest vectors generated between different users, and calculates the difference in the inner product of interest vectors among all users; and a vocational skill mapping error term, which evaluates the effectiveness of the sparse autoencoder in maintaining the vocational skill category feature representation ability, and is based on the Euclidean distance between the user's original skill mapping result and the skill representation derived from the reconstruction result. The sub-objectives are weighted and integrated by weight parameters.

[0051] Optionally, S4 includes the following steps:

[0052] S41. Utilize the optimal input weight matrix W in the optimized sparse autoencoder model. (1)* Optimal output weight matrix W (2)* Optimal User Interest Attention Matrix A * For each user feature vector z i Sparse encoding is performed to obtain the optimized user deep interest feature vector.

[0053]

[0054] Where σ(·) is the Sigmoid activation function, b (1)* is the corresponding optimal hidden layer bias vector, and ⊙ represents the element-wise multiplication operation;

[0055] S42. Construct an optimized user deep interest feature vector set F using the optimized user deep interest feature vector. * Construct a feature vector library for vocational skills modules Where s j Let K represent the skill foundation assessment result of the j-th vocational skill module in the feature space, and K represent the total number of skill modules. The vocational skill module feature vector library is constructed by professionals or obtained by semantic mapping from skill experts' scores to the interest feature space.

[0056] S43. Based on optimizing the user's deep interest feature vector set F * Calculate the user-skill similarity matrix R using the skill module feature vector library S, where R is the element of the user-skill similarity matrix. i,j The cosine similarity represents the interest similarity between the i-th user and the j-th skill module.

[0057]

[0058] S44. For each user, based on their corresponding cosine similarity vector r i ={R i,1 ,R i,2 ,...,R i,K The similarity values ​​are sorted, and the professional skill modules with similarity values ​​higher than a set threshold τ are selected to form a set of user interest-related skills S. i ={s j |R i,j ≥τ,1≤j≤K};

[0059] S45. Based on the user interest-related skill set S for each user i Combined with the interest similarity R of each skill module i,j User's current basic skill assessment results i Based on the corresponding dimension scores and the coverage of course resources available from the system, a candidate set P for personalized career skills and interest cultivation paths is generated. i Each path p in the candidate set of personalized career skills and interest development paths ik This is a series of skills—a recommended course sequence.

[0060] The beneficial effects of this invention are:

[0061] (1) This invention introduces a sparse autoencoder based on occupational skill category domain constraints and interest attention mechanism. In the feature extraction stage, the hidden layer connection structure is subject to targeted constraints. Combined with user interest preferences to guide sparse coding, the accuracy of extracting potential interest features in complex multi-source user behavior data is significantly improved. By constructing an occupational skill category mapping matrix and a domain constraint mask matrix, it is ensured that the connection between the hidden layer nodes and the input features has a clear skill semantic association. At the same time, the hidden layer output is weighted and enhanced by combining the interest attention matrix, so that the model can focus on the user's personalized interest features and effectively suppress the interference of irrelevant and redundant features. Experimental results show that in the standard user behavior dataset, the average reconstruction error in the feature extraction stage of this invention is significantly reduced and the interest vector discrimination index is improved.

[0062] (2) This invention adopts a multi-objective frog-jump optimization algorithm to jointly and dynamically optimize the input weight matrix, output weight matrix and interest attention matrix of the sparse autoencoder. By introducing three fitness indices of reconstruction error, interest feature separation and occupational skill mapping error, an evolutionary strategy combining global optimization and local refinement is established. Compared with the traditional training method that relies on single reconstruction error optimization, it can simultaneously improve the reconstruction accuracy, interest feature expression resolution and skill matching consistency of the model during the training process of the sparse autoencoder, and avoid the problem of getting trapped in local optima. Through local subgroup fine search and global elite injection mechanism, the model optimization convergence process is accelerated.

[0063] (3) This invention constructs a sparse autoencoder incremental learning and periodic frog-jump optimization retraining mechanism driven by dynamic feedback. It can capture interest shifts and skill development trends in real time during the continuous change of user behavior data, realize adaptive updates of interest features and training paths, collect user learning behavior data in real time during the execution of training paths, judge model mismatch based on cumulative reconstruction error monitoring mechanism, and automatically trigger frog-jump optimization retraining when the model performance decline exceeds the set threshold, thereby maintaining the system's sensitivity and adaptability to changes in user interests. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of a method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1 A method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders includes the following steps:

[0068] S1. Collect multi-source user behavior and basic information, construct a multi-source user behavior dataset, and preprocess it to obtain a standardized multi-source user behavior dataset;

[0069] S2. Establish an initial sparse autoencoder model and train it on a standardized multi-source user behavior dataset to obtain a preliminary set of user deep interest feature vectors;

[0070] S3. Construct a frog-jump optimization module, and iteratively optimize the initial sparse autoencoder model through the frog-jump optimization module to obtain an optimized sparse autoencoder model;

[0071] S4. Utilize an optimized sparse autoencoder model to re-encode the standardized multi-source user behavior dataset, obtain an optimized user deep interest feature vector set, combine the optimized user deep interest feature vector set with the skill module feature vector library, establish a user-skill similarity matrix, and generate a personalized vocational skill interest cultivation path candidate set based on the user-skill similarity matrix.

[0072] S5. During the process of users executing the candidate set of personalized career skills and interest cultivation paths, real-time interaction log data is collected to form a dynamic feedback data stream. The dynamic feedback data stream is input into the optimized sparse autoencoder model for incremental learning, and the updated optimized sparse autoencoder model and the updated set of personalized career skills and interest cultivation paths are output. The updated set of personalized career skills and interest cultivation paths is pushed to the user terminal and the execution effect is continuously monitored to complete the closed loop of career skills and interest cultivation.

[0073] In this embodiment, S1 includes the following steps:

[0074] S11. Collect multi-source user behavior and basic information, including user course click behavior, course browsing duration, test answer records, historical training feedback scores, and interest assessment results, and construct a multi-source user behavior dataset D. raw :

[0075]

[0076] Where, d i Let t represent the data vector of the i-th user. i b represents the timestamp of user data records. i This represents a user's course browsing behavior vector, including click frequency and browsing duration features, f i s represents the feature vector of a user's historical feedback score. i q represents the feature vector of a user's skill test score. i This represents the feature vector of the user's interest questionnaire results, where N is the total number of users.

[0077] S12. Perform format unification and structure normalization processing on the multi-source user behavior dataset, embed heterogeneous feature vectors from different sources in a unified manner, normalize features of different dimensions, and perform noise removal and missing data completion processing. For data dimensions with abnormal fluctuations or missing data, use local weighted interpolation or sliding window averaging to obtain the noise-removed multi-source user behavior dataset.

[0078] S13. Perform redundant feature elimination processing on the noise-removed multi-source user behavior dataset, and construct the feature correlation matrix R based on correlation analysis:

[0079]

[0080] Among them, R i,j Let x represent the Pearson correlation coefficient between the i-th and j-th dimensions. k,i This represents the feature value of the k-th user in the i-th dimension. Let represent the average value of the i-th dimension across all users. If the Pearson correlation coefficient is greater than a preset value, then the i-th and j-th dimensions are considered to have high redundancy. Based on preset rules, the dominant dimension is retained, and redundant dimensions are removed, resulting in a standardized multi-source user behavior dataset D. norm .

[0081] In this embodiment, S2 includes the following steps:

[0082] S21. Let the standardized multi-source user behavior dataset be... Where z i This represents the standardized feature vector of the i-th user;

[0083] S22. Construct a mapping matrix of occupational skill categories K represents the number of occupational skill categories, M ′ Using the feature dimension, we obtain the skill category feature vector c for each user. i =Cz i ;

[0084] S23. Construct a domain constraint mask matrix M based on the occupational skill category mapping matrix C. The domain constraint mask matrix is ​​used to define the connection relationship between each node of the hidden layer and the input feature dimension, so that each hidden node is only connected to the input dimension feature of its corresponding skill category, thus forming a structured skill perception connection pattern.

[0085] M = [m jl ] H×M′ ;

[0086]

[0087] Where H is the number of hidden layer nodes, m jl This indicates whether the j-th node in the hidden layer is connected to the l-th feature in the input feature dimension. Its value is either 0 or 1, where 1 indicates a connection and 0 indicates no connection. C kl This indicates whether there is a semantic relationship between the k-th skill category and the l-th input feature in the occupational skill category mapping matrix. j≡k(mod H / K) means that the hidden layer nodes are divided equally according to the number of occupational skill categories K, with each category corresponding to H / K hidden nodes, and structural constraints are established so that the j-th hidden node is only connected to the input feature dimension related to its corresponding category k.

[0088] S24. Input the sparse autoencoder into the weight matrix W. (1) Element-wise multiplication with the domain constraint mask matrix M yields the weighted input weight matrix.

[0089] S25. Construct the user interest and attention matrix A with the standardized feature vector z of the i-th user.i Attention-weighted calculations are performed to generate the interest attention vector γ. i Each component of the interest attention vector is used to measure the degree of attention that the corresponding node in the hidden layer pays to the user.

[0090] γ i =softmax(Az) i );

[0091] S26. Calculate the sparse coding representation of the hidden layer based on the domain-constrained mask matrix and the interest attention matrix:

[0092]

[0093] Where σ(·) is the Sigmoid activation function, b (1) This is the hidden layer bias vector;

[0094] S27. Based on the sparse coding representation of the hidden layer h i,j For each node in the hidden layer, calculate its average activation value across all user samples. Where j represents the hidden layer node number;

[0095] S28. Constructing the comprehensive loss function

[0096] S29. The gradient descent algorithm is used to iteratively optimize the comprehensive loss function. During training, the weighted input weight matrix, output weight matrix, hidden layer bias vector and output bias vector are dynamically adjusted, and the learning rate is decayed according to the current iteration number u.

[0097] S210. When the convergence result of the comprehensive loss function is lower than the set error threshold, output the preliminary user deep interest feature vector set F. init Each vector h in the initial set of deep user interest feature vectors i This represents the deeper expression of the i-th user in the dimension of career skills and interests.

[0098] In this implementation, the comprehensive loss function includes: a reconstruction error term, used to measure the difference between the reconstructed vector and the original input vector; a KL divergence-based term, used to measure the deviation of the average activation value of the hidden layer nodes from the target sparsity ρ, adjusting the sparsity of the hidden layer representation; and an interest separation regularization term, used to encourage good distinguishability of interest representations among different users in the semantic space, improving the model's ability to express differences.

[0099]

[0100] in, To reconstruct the vector, W (2) To output the weight matrix, b(2) is the output bias vector, β is the sparse regularization coefficient, ρ is the target sparsity, λ is the contrast regularization coefficient, and Δ is the interest separation threshold.

[0101] In this embodiment, S3 includes the following steps:

[0102] S31. Construct a frog-jumping optimization module to optimize the initial sparse autoencoder model. Set the total size of the frog swarm to P, and define each frog as a candidate sparse autoencoder model containing the input weight matrix W. (1) Output weight matrix W (2) And user interest attention matrix A;

[0103] S32. Divide the frog swarm into m subgroups, each subgroup containing several individuals. For each candidate sparse autoencoder model, construct a multi-objective fitness function F based on the generated preliminary user deep interest feature vector set.

[0104] S33. For each subgroup, sort the individuals in the subgroup from best to worst according to the multi-objective fitness function F, and select the individual with the best multi-objective fitness function value as the best individual of that subgroup. The individual with the worst multi-objective fitness function value is selected as the worst individual in the subgroup. A jump vector is constructed based on the difference between the two, the jump magnitude is scaled by the jump factor r, and the parameters of the worst individual are updated.

[0105] S34. When the updated individual using the jump vector is better than the original worst individual in the multi-objective fitness function value, the updated individual replaces the worst individual. If the update does not bring about optimization, a new set of individuals with Gaussian perturbation is generated according to the structural perturbation principle and used to replace the worst individual in the subgroup.

[0106] S35. After all subgroups have executed the local jump strategy, select the individual with the best multi-objective fitness function value from the entire frog swarm as the global elite individual. The core model parameters of the global elite individual will be injected into the individual with the worst fitness in all subgroups to build an elite-driven global accelerated convergence mechanism.

[0107] S36. Combine the user deep interest feature vectors output by the encoders in all current subgroups to calculate the average interest representation mutual information between different subgroups, which is defined as the global interest diversity index Ψ. When the global interest diversity index is lower than the preset diversity threshold, it indicates that the global search has shown a trend of convergence and concentration, and the system automatically triggers the jump perturbation enhancement mechanism.

[0108] S37. The process of local hopping in the subgroup, global elite propagation, and interest diversity regulation will be continuously iterated until one of the following termination conditions is met: one of which is that the number of iterations reaches the maximum number of iterations T. max Secondly, the multi-objective fitness function value of the globally optimal individual converges to below the error threshold ∈.

[0109] S38. The final output is the set of individual parameters with the best global fitness, which includes the optimal input weight matrix W. (1)* Optimal output weight matrix W (2)* With the optimal user interest attention matrix A * Individuals are used to optimize the sparse autoencoder model, resulting in an optimized sparse autoencoder model.

[0110] In this embodiment, the multi-objective fitness function includes a reconstruction error term, which evaluates the ability of the candidate sparse autoencoder model to retain the original user input feature information. The metric is the mean square error between each user's input vector and the vector reconstructed by the sparse autoencoder. An interest feature separation term is used to measure the discriminative power of the deep interest vectors generated between different users. The difference in the inner product of the interest vectors among all users is compared and calculated. A vocational skill mapping error term is used to evaluate the effectiveness of the sparse autoencoder in maintaining the vocational skill category feature representation ability. The metric is the Euclidean distance between the user's original skill mapping result and the skill representation derived from the reconstruction result. The sub-objectives are weighted and integrated using weight parameters.

[0111] In this embodiment, S4 includes the following steps:

[0112] S41. Utilize the optimal input weight matrix W in the optimized sparse autoencoder model. (1)* Optimal output weight matrix W (2)* Optimal User Interest Attention Matrix A * For each user feature vector z i Sparse encoding is performed to obtain the optimized user deep interest feature vector.

[0113]

[0114] Where σ(·) is the Sigmoid activation function, b (1)* is the corresponding optimal hidden layer bias vector, and ⊙ represents the element-wise multiplication operation;

[0115] S42. Construct an optimized user deep interest feature vector set F using the optimized user deep interest feature vector. * Construct a feature vector library for vocational skills modules Where s jLet K represent the skill foundation assessment result of the j-th vocational skill module in the feature space, and K represent the total number of skill modules. The vocational skill module feature vector library is constructed by professionals or obtained by semantic mapping from skill experts' scores to the interest feature space.

[0116] S43. Based on optimizing the user's deep interest feature vector set F * Calculate the user-skill similarity matrix R using the skill module feature vector library S, where R is the element of the user-skill similarity matrix. i,j The cosine similarity represents the interest similarity between the i-th user and the j-th skill module.

[0117]

[0118] S44. For each user, based on their corresponding cosine similarity vector r i ={R i,1 ,R i,2 ,...,R i,K The similarity values ​​are sorted, and the professional skill modules with similarity values ​​higher than a set threshold τ are selected to form a set of user interest-related skills S. i ={s j |R i,j ≥τ,1≤j≤K};

[0119] S45. Based on the user interest-related skill set S for each user i Combined with the interest similarity R of each skill module i,j User's current basic skill assessment results i Based on the corresponding dimension scores and the coverage of course resources available from the system, a candidate set P for personalized career skills and interest cultivation paths is generated. i Each path p in the candidate set of personalized career skills and interest development paths ik This is a series of skills—a recommended course sequence.

[0120] Example 1:

[0121] SkillPlus, a large-scale comprehensive online vocational skills training platform, has launched a new generation of vocational skills interest cultivation system. The system is the first to fully deploy the personalized skill recommendation technology based on frog-jump optimization and sparse autoencoder proposed in this invention. The pilot coverage area is Xuhui District, Shanghai, covering 12 popular vocational skills modules including artificial intelligence engineer, data analyst, and network security engineer.

[0122] Within the first month of the system's official operation, the SkillPlus platform received data input from a total of 3,274 registered users from different colleges and enterprises undergoing transformation. By filling in basic information and interest questionnaires, users were able to access the system's backend records of their behavior over the past three months, including watching courses, completing tests, and participating in training projects on the platform. This resulted in a preliminary multi-source user behavior dataset with an initial data volume of 18GB.

[0123] During the data standardization process, platform engineers discovered that some user records contained abnormally long browsing durations (over 10 hours per session) or abnormally fluctuating test scores (score differences exceeding 60 points within the same day). The system automatically initiated an anomaly removal and completion mechanism, reducing the proportion of abnormal records from the initial 2.8% to 0.6%, resulting in a final standardized effective feature dimension of 108 dimensions.

[0124] Subsequently, based on the occupational skill classification standard, the system maps the input features to 12 occupational skill categories, constructing a skill category mapping matrix C. A domain constraint mask matrix M is generated based on the mapping results, and a user interest attention matrix A is dynamically generated by combining user click preferences and answer preferences. For example, for user ID U_0589, the feature attention weight in the skill category "Data Analysis" reaches 0.82, significantly higher than the average of 0.45.

[0125] Based on the above structure, SkillPlus deployed an initial sparse autoencoder model with 96 hidden layers. During the initial training phase, the system evaluated the model performance every 100 iterations. In the first evaluation, the reconstruction error was 0.091, and the interest feature separation index was only 0.58, prompting the system to automatically activate the frog-jump optimization module.

[0126] The frog swarm size was set to 100 frogs, with 8 subgroups. A local jump strategy and a global elite propagation mechanism were executed in parallel. In the 7th subgroup iteration, the system found that in subgroup 3, individual model P_3_7, after a single jump, reduced its reconstruction error from 0.084 to 0.072, and its interest feature separation index increased from 0.62 to 0.69, demonstrating significant optimization and becoming a local elite. By the 40th round, the global interest diversity index Ψ dropped to 0.35, below the preset threshold of 0.4, triggering the jump perturbation enhancement mechanism and successfully avoiding early convergence. Finally, the model globally converged in the 53rd round, and the indicators of the best-fit individual stabilized.

[0127] Based on an optimized sparse autoencoder model, the platform completed interest feature encoding for 3274 users. The system uses a skill module library (containing feature vectors for 112 skill modules) to calculate the skill-interest similarity for each user. For example, user U_1062's similarity to the "Machine Learning Practice Project" module reached 0.78, exceeding the recommendation threshold of 0.65; therefore, the system included it in that user's training path.

[0128] In a specific case, user U_1451 is a third-year student majoring in electrical engineering. The initial recommended learning path was: "Python Basics → Data Cleaning and Visualization → Machine Learning Algorithm Practice → Big Data Processing". When generating the recommendation, the platform referenced dynamic feedback data showing that the user's completion rate for "data processing" courses was as high as 93% and the test accuracy rate remained above 87% within the past month.

[0129] After the user's learning path is pushed, the system continuously tracks the learning process. Taking user U_1451 as an example, from July 12th to August 30th, 2025, this user consistently studied for more than 2 hours per day, completing an average of 2 small modules per week. The system collects dynamic feedback data every two weeks and found that by the fourth week, the user's accuracy rate in answering questions in the "Machine Learning Algorithm Practice" course dropped to 62%, and the learning time shortened to 50 minutes per day. The system automatically recorded the cumulative reconstruction error for this stage as 0.092, which is higher than the set threshold of 0.08.

[0130] Upon detecting the anomaly, the system automatically triggered a leapfrog optimization module to retrain and update the sparse autoencoder. After rapid optimization and adjustment, the reconstruction error decreased to 0.069, the interest vector was re-encoded, and the system determined that the user's interest bias had changed, shifting from "algorithm theory" to "data application engineering." Therefore, the platform automatically adjusted the recommended path, replacing subsequent courses with content on "data engineering project practice" and "advanced SQL applications," and pushed updated learning plan notifications via the app.

[0131] In a comprehensive performance evaluation after three months of system operation, SkillPlus compiled comparative data between the method of this invention and traditional static recommendation methods:

[0132] Table 1 Comparison data between the method of this invention and traditional static recommendation methods

[0133] index Traditional methods Method of the present invention cultivation pathway completion rate 59.3％ 78.5％ Interest change capture rate 21.4％ 45.7％ Average recommendation accuracy (skills assessment pass rate) 72.6％ 84.1％ Average path update response time 6 days 12 hours 23 hours

[0134] The improved interest change capture rate stems from the introduction of dynamic reconstruction error monitoring and periodic model optimization, which significantly enhances the timeliness of user interest drift identification. For example, when user U_2456 was executing the second recommended path, their interest in the "cybersecurity" module rapidly increased due to a short-term change in job requirements. The system adjusted the recommended content within 48 hours to ensure that the learning plan kept pace with the user's actual needs in real time.

[0135] Examples of training sample instances are as follows:

[0136] User U_0567, initial feature encoding vector part:

[0137] [0.15, 0.03, 0.72, 0.05, 0.55, ..., 0.12];

[0138] Encoding vector optimized by frog-jumping algorithm:

[0139] [0.22,0.01,0.79,0.03,0.63,…,0.09];

[0140] The reconstruction error decreased from 0.085 to 0.067.

[0141] Example of dynamic feedback behavior for user U_0892:

[0142] July 20, 2024: Course completion rate 85%, quiz accuracy rate 88%;

[0143] August 3, 2024: Course completion rate 72%, correct answer rate 59%, reconstruction error increased to 0.095;

[0144] Early morning of August 4, 2024: Automatic model re-optimization triggered.

[0145] The data results show that the method of this invention realizes a closed-loop management of dynamic monitoring, interest update and path re-recommendation, which greatly improves the personalized response capability of the vocational skills training system and user learning satisfaction.

[0146] In summary, this embodiment 1 fully verifies in a real-world scenario that the method proposed in this invention not only has significant advantages in personalized interest capture and cultivation path matching, but also has excellent adaptive update capabilities in dynamic environments, significantly solving the prominent problems of existing technologies in terms of accuracy, real-time performance, and system stability.

[0147] This invention introduces a sparse autoencoder based on occupational skill category domain constraints and an interest attention mechanism. During the feature extraction stage, it applies targeted constraints to the hidden layer connection structure and incorporates user interest preferences to guide sparse coding. This significantly improves the accuracy of extracting latent interest features from complex, multi-source user behavior data. By constructing an occupational skill category mapping matrix and a domain constraint mask matrix, it ensures that the connections between hidden layer nodes and input features have clear skill semantic associations. Simultaneously, the interest attention matrix is ​​used to weight and enhance the hidden layer output, enabling the model to focus on personalized user interest features and effectively suppress interference from irrelevant and redundant features. Experimental results show that, on a standard user behavior dataset, the average reconstruction error during the feature extraction stage of this invention is significantly reduced, and the interest vector discriminancy index is improved.

[0148] This invention employs a multi-objective frog-leap optimization algorithm to jointly and dynamically optimize the input weight matrix, output weight matrix, and interest / attention matrix of a sparse autoencoder. By introducing a triple fitness index—reconstruction error, interest feature separability, and vocational skill mapping error—an evolutionary strategy combining global optimization and local refinement is established. Compared to traditional training methods that rely on single reconstruction error optimization, this approach can simultaneously improve the model's reconstruction accuracy, interest feature expression resolution, and skill matching consistency during the training process of the sparse autoencoder, avoiding the problem of getting trapped in local optima. Through local subgroup fine-grained search and global elite injection mechanisms, the model optimization convergence process is accelerated.

[0149] This invention constructs a dynamic feedback-driven incremental learning and periodic frog-leap optimization retraining mechanism for sparse autoencoders. It can capture interest shifts and skill development trends in real time as user behavior data changes continuously, and achieve adaptive updates of interest features and training paths. By collecting user learning behavior data in real time during the execution of the training path, it judges model mismatch based on a cumulative reconstruction error monitoring mechanism. When the model performance decline exceeds a set threshold, it automatically triggers frog-leap optimization retraining, thereby maintaining the system's sensitivity and adaptability to changes in user interests.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders, characterized in that, Includes the following steps: S1. Collect multi-source user behavior and basic information, construct a multi-source user behavior dataset, and preprocess it to obtain a standardized multi-source user behavior dataset; S2. Establish an initial sparse autoencoder model and train it on a standardized multi-source user behavior dataset to obtain a preliminary set of user deep interest feature vectors; S3. Construct a frog-jump optimization module, and iteratively optimize the initial sparse autoencoder model through the frog-jump optimization module to obtain an optimized sparse autoencoder model; S4. Utilize an optimized sparse autoencoder model to re-encode the standardized multi-source user behavior dataset, obtain an optimized user deep interest feature vector set, combine the optimized user deep interest feature vector set with the skill module feature vector library, establish a user-skill similarity matrix, and generate a personalized vocational skill interest cultivation path candidate set based on the user-skill similarity matrix. S5. During the process of users executing the candidate set of personalized career skills and interest cultivation paths, real-time interaction log data is collected to form a dynamic feedback data stream. The dynamic feedback data stream is input into the optimized sparse autoencoder model for incremental learning, and the updated optimized sparse autoencoder model and the updated set of personalized career skills and interest cultivation paths are output. The updated set of personalized career skills and interest cultivation paths is pushed to the user terminal and the execution effect is continuously monitored to complete the closed loop of career skills and interest cultivation.

2. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 1, characterized in that, S1 includes the following steps: S11. Collect multi-source user behavior and basic information, including user course click behavior, course browsing duration, test answer records, historical training feedback scores, and interest assessment results, and construct a multi-source user behavior dataset D. raw ; S12. Perform format unification and structure normalization processing on the multi-source user behavior dataset, embed heterogeneous feature vectors from different sources in a unified manner, normalize features of different dimensions, and perform noise removal and missing data completion processing. For data dimensions with abnormal fluctuations or missing data, use local weighted interpolation or sliding window averaging to obtain the noise-removed multi-source user behavior dataset. S13. Redundant feature elimination processing is performed on the noise-removed multi-source user behavior dataset. A feature correlation matrix R is constructed based on correlation analysis, and the Pearson correlation coefficient is obtained. If the Pearson correlation coefficient is greater than a preset value, it is determined that the i-th and j-th dimensions have high redundancy. The dominant dimension is retained and the redundant dimensions are removed according to preset rules to obtain the standardized multi-source user behavior dataset. Among them, z i Let N represent the standardized feature vector of the i-th user, and N be the total number of users.

3. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 2, characterized in that, S2 includes the following steps: S21. Construct a mapping matrix of occupational skill categories K represents the number of occupational skill categories, M ′ Using the feature dimension, we obtain the skill category feature vector c for each user. i =Cz i ; S22. Construct a domain constraint mask matrix M based on the occupational skill category mapping matrix C. The domain constraint mask matrix is ​​used to define the connection relationship between each node of the hidden layer and the input feature dimension, so that each hidden node is only connected to the input dimension feature of its corresponding skill category, thus forming a structured skill perception connection pattern. Where H is the number of hidden layer nodes. Indicates whether the j-th node in the hidden layer is related to the j-th node in the input feature dimension. Establish connections based on these features. This represents the relationship between the k-th skill category and the 6-th skill category in the occupational skill category mapping matrix. Whether there is a semantic relationship between the input features, j≡k(mod H / K) means that the hidden layer nodes are divided equally according to the number of occupational skill categories K, each category corresponds to H / K hidden nodes, and structural constraints are established so that the j-th hidden node is only connected to the input feature dimension related to its corresponding category k. S23. Input the sparse autoencoder into the weight matrix W. (1) Element-wise multiplication with the domain constraint mask matrix M yields the weighted input weight matrix. S24. Construct the user interest and attention matrix A with the standardized feature vector z of the i-th user. i Attention-weighted calculations are performed to generate the interest attention vector γ. i Each component of the interest attention vector is used to measure the degree of attention that the corresponding node in the hidden layer pays to the user. S25. Calculate the sparse coding representation of the hidden layer based on the domain-constrained mask matrix and the interest attention matrix: Where σ(·) is the Sigmoid activation function, b (1) This is the hidden layer bias vector; S26. Based on the sparse coding representation of the hidden layer h i,j For each node in the hidden layer, calculate its average activation value across all user samples. Where j represents the hidden layer node number; S27. Constructing the comprehensive loss function S28. The gradient descent algorithm is used to iteratively optimize the comprehensive loss function. During the training process, the weighted input weight matrix, output weight matrix, hidden layer bias vector and output bias vector are dynamically adjusted, and the learning rate is decayed according to the current iteration number u. S29. When the convergence result of the comprehensive loss function is lower than the set error threshold, output the preliminary user deep interest feature vector set F. init Each vector h in the initial set of deep user interest feature vectors i This represents the deeper expression of the i-th user in the dimension of career skills and interests.

4. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 3, characterized in that, The comprehensive loss function includes: a reconstruction error term, used to measure the difference between the reconstructed vector and the original input vector; and a KL divergence-based term, used to measure the deviation of the average activation value of the hidden layer nodes from the target sparsity ρ, thereby adjusting the sparsity of the hidden layer representation. Interest separation regularization term: Used to encourage the interest representations of different users to have good distinguishability in semantic space, thereby improving the model's ability to express differences. in, To reconstruct the vector, W (2) To output the weight matrix, b (2) is the output bias vector, β is the sparse regularization coefficient, ρ is the target sparsity, λ is the contrast regularization coefficient, and Δ is the interest separation threshold.

5. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 4, characterized in that, S3 includes the following steps: S31. Construct a frog-jumping optimization module to optimize the initial sparse autoencoder model. Set the total size of the frog swarm to P, and define each frog as a candidate sparse autoencoder model containing the input weight matrix W. (1) Output weight matrix W (2) And user interest attention matrix A; S32. Divide the frog swarm into m subgroups, each subgroup containing several individuals. For each candidate sparse autoencoder model, construct a multi-objective fitness function F based on the generated preliminary user deep interest feature vector set. S33. For each subgroup, sort the individuals in the subgroup from best to worst according to the multi-objective fitness function F, and select the individual with the best multi-objective fitness function value as the best individual of that subgroup. The individual with the worst multi-objective fitness function value is selected as the worst individual in the subgroup. A jump vector is constructed based on the difference between the two, the jump magnitude is scaled by the jump factor r, and the parameters of the worst individual are updated. S34. When the updated individual using the jump vector is better than the original worst individual in the multi-objective fitness function value, the updated individual replaces the worst individual. If the update does not bring about optimization, a new set of individuals with Gaussian perturbation is generated according to the structural perturbation principle and used to replace the worst individual in the subgroup. S35. After all subgroups have executed the local jump strategy, select the individual with the best multi-objective fitness function value from the entire frog swarm as the global elite individual. The core model parameters of the global elite individual will be injected into the individual with the worst fitness in all subgroups to build an elite-driven global accelerated convergence mechanism. S36. Combine the user deep interest feature vectors output by the encoders in all current subgroups to calculate the average interest representation mutual information between different subgroups, which is defined as the global interest diversity index Ψ. When the global interest diversity index is lower than the preset diversity threshold, it indicates that the global search has shown a trend of convergence and concentration, and the system automatically triggers the jump perturbation enhancement mechanism. S37. The process of local hopping in the subgroup, global elite propagation, and interest diversity regulation will be continuously iterated until one of the following termination conditions is met: one of which is that the number of iterations reaches the maximum number of iterations T. max Secondly, the multi-objective fitness function value of the globally optimal individual converges to below the error threshold ∈. S38. The final output is the set of individual parameters with the best global fitness, which includes the optimal input weight matrix W. (1)* Optimal output weight matrix W (2)* With the optimal user interest attention matrix A * Individuals are used to optimize the sparse autoencoder model, resulting in an optimized sparse autoencoder model.

6. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 5, characterized in that, The multi-objective fitness function includes a reconstruction error term, which evaluates the ability of the candidate sparse autoencoder model to retain the original user input feature information. The metric is the mean square error between each user's input vector and the vector reconstructed by the sparse autoencoder. An interest feature separation term is used to measure the discriminative power of the deep interest vectors generated between different users. The difference in the inner product of the interest vectors among all users is calculated by comparing the differences. A vocational skill mapping error term is used to evaluate the effectiveness of the sparse autoencoder in maintaining the vocational skill category feature representation. The metric is the Euclidean distance between the user's original skill mapping result and the skill representation derived from the reconstruction result. The sub-objectives are weighted and integrated using weight parameters.

7. The method for cultivating vocational skills and interests based on frog-jump optimization and sparse autoencoders according to claim 5, characterized in that, S4 includes the following steps: S41. Utilize the optimal input weight matrix W in the optimized sparse autoencoder model. (1)* Optimal output weight matrix W (2)* Optimal User Interest Attention Matrix A / For each user feature vector z i Sparse encoding is performed to obtain the optimized user deep interest feature vector. S42. Construct an optimized user deep interest feature vector set F using the optimized user deep interest feature vector. * Construct a feature vector library for vocational skills modules Where s j Let K represent the skill foundation assessment result of the j-th vocational skill module in the feature space, and K represent the total number of skill modules. The vocational skill module feature vector library is constructed by professionals or obtained by semantic mapping from skill experts' scores to the interest feature space. S43. Based on optimizing the user's deep interest feature vector set F * Calculate the user-skill similarity matrix R using the skill module feature vector library S, where R is the element of the user-skill similarity matrix. i,j This represents the interest similarity between the i-th user and the j-th skill module; S44. For each user, based on their corresponding cosine similarity vector r i ={R i,1 ,R i,2 ,...,R i,K The similarity values ​​are sorted, and the professional skill modules with similarity values ​​higher than a set threshold τ are selected to form a set of user interest-related skills S. i ={s j |R i,j ≥τ,1≤j≤K}; S45. Based on the user interest-related skill set S for each user i Combined with the interest similarity R of each skill module i,j User's current basic skill assessment results i Based on the corresponding dimension scores and the coverage of course resources available from the system, a candidate set P for personalized career skills and interest cultivation paths is generated. i Each path p in the candidate set of personalized career skills and interest development paths ik This is a series of skills—a recommended course sequence.

Images