Personalized learning path recommendation method based on deep learning

Abstract

The invention discloses a personalized learning path recommendation method based on deep learning, and the method comprises the following steps: S1, collecting and standardizing a learning log, resource data and a knowledge point first repair relation, and generating a training data set and a knowledge graph; s2, constructing a deep Boltzmann machine model based on the training data set and initializing parameters; s3, performing model improvement based on meta learning; s4, training a strategy prior network and a value network; s5, screening resources according to the knowledge graph and the learner state to generate a candidate set; s6, adopting a Monte Carlo tree with prior guidance to search and optimize path selection, and generating candidate learning paths; and S7, selecting an optimal path according to the candidate path scores, and outputting final model parameters after judging convergence. According to the method, accurate recommendation of personalized learning paths is realized, and the learning efficiency and the intelligent adaptability of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and in particular to a personalized learning path recommendation method based on deep learning. Background Technology

[0002] With the development of online education, personalized learning path recommendation systems have become an important means to improve learning outcomes. Traditional learning path recommendation methods mostly rely on rules and collaborative filtering algorithms. These methods primarily predict learner needs based on historical learning data; however, they cannot effectively address the dynamic changes in learner needs. Traditional methods have significant limitations in the accuracy and personalization of recommendations, especially when faced with complex learning tasks, often failing to accurately reflect the learner's knowledge mastery.

[0003] Deep learning technology has been introduced into personalized recommendation systems, using neural networks to train learner behavior data and improving the intelligence of these systems. However, existing deep learning methods still have problems. Most methods rely on static training data and cannot adjust to learners' personalized needs in real time. Furthermore, deep learning models have long training times and high computational costs, making it difficult to meet the real-time recommendation requirements of practical applications.

[0004] While some methods have incorporated optimization algorithms such as reinforcement learning or Monte Carlo tree search to improve path recommendation performance, these methods still have limitations. For example, reinforcement learning algorithms typically require a large number of samples and time to converge, and struggle to balance exploration and exploitation in complex learning tasks. Although Monte Carlo tree search can optimize path selection, traditional methods cannot dynamically adjust to learner needs and still rely on coarse-based pre-defined models, failing to accurately address personalized recommendation requirements.

[0005] Therefore, how to provide a personalized learning path recommendation method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a personalized learning path recommendation method based on deep learning. This invention integrates a deep Boltzmann machine with an improved Monte Carlo tree search algorithm to construct an individualized state modeling and dynamic path optimization mechanism, which has the advantages of high recommendation accuracy, strong adaptability and high training efficiency.

[0007] A personalized learning path recommendation method based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect learning interaction logs, learning resource data and knowledge point prerequisite relationships, perform data preprocessing, and generate training datasets and knowledge graphs; S2. Construct a deep Boltzmann machine model based on the training dataset, and complete the parameter initialization and structure setting of the deep Boltzmann machine model; S3. The process of improving the deep Boltzmann machine model is to optimize the parameter sharing for different learning tasks in the first training layer, and to perform fast parameter updates based on learner samples in the second training layer, generating an individualized parameter set and outputting the learner state vector. S4. Use the state representation and reward results output by the deep Boltzmann machine model to train the policy prior network and value network; S5. Based on the knowledge graph and learner state vector, select learning resources, generate a candidate resource set, and use the Monte Carlo tree search algorithm to perform node selection, expansion, simulation and backtracking. S6. Perform the AlphaZero-style prior-guided improvement process of Monte Carlo tree search, call the policy prior network and value network parameters to select nodes, call the individualized parameter set to calculate the completion probability and learning benefit, search and update the path benefit according to the PUCT rule, and generate candidate learning paths. S7. Select the target learning path based on the candidate learning path scores, determine whether the deep Boltzmann machine model has converged, if it has converged, output the final deep Boltzmann machine parameter set, policy prior network parameters and value network parameters and end; if it has not converged, record the execution result and return to S1.

[0008] Optionally, the training dataset and knowledge graph in S1 specifically include: S11. Perform missing value imputation, outlier removal and format standardization on the collected data to generate a training dataset containing learner features, knowledge point features and resource features. S12. Construct a knowledge graph based on the prerequisite relationships of knowledge points, forming a directed graph structure with nodes representing knowledge points and edges representing prerequisite relationships.

[0009] Optionally, the first training layer in S3 specifically includes: S311. The first training layer receives input including the training dataset and the set of initial parameters for the deep Boltzmann machine. S312. Extract task sample sets from the training dataset according to task division, divide the task sample sets into support samples and validation samples, load the initial parameter set of the deep Boltzmann machine, perform forward propagation calculation on learner features, knowledge point features and resource features in the support samples, pass through the input layer, hidden layer and output layer in sequence, and calculate the output prediction value. S313. Calculate the absolute error of the support samples as the prediction error of the support samples. Take the square of the prediction error of the support samples and average it according to the number of samples to obtain the mean square error result. Add the sum of the squares of the parameter weights multiplied by the regularization coefficient to the loss calculation to form the regularization term. Add the regularization term to the mean square error result to obtain the total loss value. S314. Calculate the parameter gradients based on the total loss value, use the gradient descent algorithm in the optimized configuration, multiply the gradient of each parameter by the learning rate and then perform a weighted update, and perform gradient clipping operation on the updated parameters. S315. Use the validation samples to perform forward computation and obtain the prediction error of the validation samples; when the decrease in validation error is lower than the set threshold for several consecutive times, the training is considered to have converged. S316. After the convergence condition is met, the first training layer outputs the shared parameter set and the input standardized statistics. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

[0010] Optionally, the second training layer in S3 specifically includes: S321. The second training layer receives input including a shared parameter set, standardized input statistics, and a sample of the target learners. S322. Based on the input standardized statistical information, perform standardization processing on the learner features, knowledge point features and resource features in the target learner sample, and keep the distribution consistent with the input feature distribution of the first training layer to generate a standardized learner sample set. S323. Load the shared parameter set as the initial parameters of the second training layer, perform forward propagation calculation on the target learner sample, and obtain the second prediction value; S324. Calculate the absolute error of the target learner sample as the second prediction error, average the square of the second prediction error according to the number of samples, add the second regularization term, and obtain the second total loss value. S325. Perform backpropagation calculation based on the second total loss value to obtain the second gradient set of parameters of each layer. Use the gradient descent algorithm to multiply the gradient of each parameter by the second learning rate for weighted update, and perform the second gradient pruning operation. S326. When the second verification error decreases below a set threshold for several consecutive times, the training of the second training layer is determined to be converged. S327. After the convergence condition is met, the second training layer outputs an individualized parameter set and a learner state vector. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

[0011] Optionally, S4 specifically includes: S41. Map learner state vectors to benefit results to generate a set of correspondences between learner features and benefit results; benefit results refer to quantitative indicators of learning effectiveness calculated based on learners' learning completion status, test performance, and resource usage feedback in historical interaction logs. S42. Based on the set of correspondences, train the policy prior network and the value network, and establish the mapping relationship from learning state to policy distribution and from learning state to benefit evaluation, respectively. S43. During training, simultaneously optimize the parameters of the policy prior network and the value network until the outputs of the two networks reach the convergence condition, and output the trained policy prior network and value network.

[0012] Optionally, S5 specifically includes: S51. Calculate the set of knowledge points that the learner has not yet mastered based on the knowledge graph and the learner's state vector; S52. Based on the prerequisite relationships of knowledge points in the knowledge graph, select learning resources from the learning resource library that are related to the knowledge points that have not been mastered and whose prerequisite conditions are met. S53. Calculate the fit score for each resource by combining the learner's historical learning behavior records and resource usage preferences, and select learning resources with fit scores higher than a preset threshold to generate a candidate resource set.

[0013] Optionally, S6 specifically includes: S61, Node Selection Phase: The learner's state vector and candidate resource set are input into the policy prior network, and the prior probability of each candidate action is output. The learner's state vector and candidate resource set are input into the individualized parameter set, and the mastery improvement value is calculated. The historical reward results of the candidate resource set are extracted from the correspondence set, and the mastery improvement value and historical reward value are weighted and summed according to preset weights to obtain the immediate reward. The exploration gain is calculated by combining the node visit count, and the prior probability, immediate reward, and exploration gain are weighted to form a comprehensive score. The candidate action with the highest comprehensive score is selected to enter the next node. S62, Node Expansion Phase: Generate new nodes based on candidate actions, construct corresponding state vectors, and initialize node visit counts and average path revenue; S63. Simulation Phase: Starting from the state vector of the new node, the policy prior network is called step by step to generate action probabilities and actions are sampled according to the probabilities; the individualized parameter set is called to calculate the immediate benefits of each step, and the immediate benefits are weighted and accumulated according to the path depth decay coefficient to obtain the cumulative immediate benefits; the value network is called to input the simulation end state vector to obtain the end value assessment, and the cumulative immediate benefits and the end value assessment are weighted according to the weights to form the simulation return value. S64. Backtracking Phase: The simulation reward value is backtracked sequentially from the leaf node to the root node along the current path. The number of visits to each node and the average path revenue are updated. The average path revenue is updated by weighting the current average path revenue and the simulation reward value. The number of visits to the nodes is weighted by the prior probability of the action and the comprehensive score is refreshed. S65. Termination and Output Phase: Repeat S61 to S64 until the maximum number of simulations is reached or the average path reward change between two consecutive searches is lower than the convergence threshold, at which point the search is terminated; calculate the average path reward of each action under the root node as the path score, generate a set of candidate learning paths, sort them by path score, select the target learning path, and output it.

[0014] Optionally, S7 specifically includes: S71. Sort the candidate learning paths according to their path scores, and select the learning path with the highest path score as the target learning path. S72. Determine whether the deep Boltzmann machine model has converged. The convergence criterion is: if the path score changes by less than 0.01 in 5 consecutive training sessions, the deep Boltzmann machine model is considered to have converged. S73. If the deep Boltzmann machine model converges, output the final set of deep Boltzmann machine parameters, policy prior network parameters, and value network parameters, and end the learning path recommendation process. S74. If the deep Boltzmann machine model fails to converge, record the current execution result and return to S1 for the next round of training until the convergence condition is met or the preset maximum training round of 1000 is reached.

[0015] The beneficial effects of this invention are: (1) Through a two-level training mechanism, the first training layer realizes global modeling of shared knowledge between different tasks, and the second training layer quickly adapts to the individual differences of the target learners, effectively improving the model's ability to characterize individual learning features and enhancing the accuracy of personalized recommendations.

[0016] (2) Introducing a strategy prior and value guidance mechanism into Monte Carlo tree search, integrating learner status, resource characteristics, mastery improvement and historical gains, and comprehensively scoring to guide the path search direction, significantly improving the rationality and effectiveness of the generated learning path.

[0017] (3) By using optimization methods such as standardized input, gradient clipping, and early stopping mechanism, the training stability is improved and overfitting and oscillation are avoided. At the same time, the search loop is controlled by path scoring and convergence judgment to shorten the recommendation time and improve the overall system efficiency. Attached Figure Description

[0018] 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:

[0019] Figure 1 This is a flowchart of a personalized learning path recommendation method based on deep learning proposed in this invention; Figure 2 This is a structural diagram of an improved deep Boltzmann machine model for a personalized learning path recommendation method based on deep learning proposed in this invention. Figure 3 This is a diagram of the improved Monte Carlo tree search structure for a personalized learning path recommendation method based on deep learning proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figure 1-3 A personalized learning path recommendation method based on deep learning includes the following steps: S1. Collect learning interaction logs, learning resource data and knowledge point prerequisite relationships, perform data preprocessing, and generate training datasets and knowledge graphs; S2. Construct a deep Boltzmann machine model based on the training dataset, and complete the parameter initialization and structure setting of the deep Boltzmann machine model; S3. The process of improving the deep Boltzmann machine model is to optimize the parameter sharing for different learning tasks in the first training layer, and to perform fast parameter updates based on learner samples in the second training layer, generating an individualized parameter set and outputting the learner state vector. S4. Use the state representation and reward results output by the deep Boltzmann machine model to train the policy prior network and value network; S5. Based on the knowledge graph and learner state vector, select learning resources, generate a candidate resource set, and use the Monte Carlo tree search algorithm to perform node selection, expansion, simulation and backtracking. S6. Perform the AlphaZero-style prior-guided improvement process of Monte Carlo tree search, call the policy prior network and value network parameters to select nodes, call the individualized parameter set to calculate the completion probability and learning benefit, search and update the path benefit according to the PUCT rule, and generate candidate learning paths. S7. Select the target learning path based on the candidate learning path scores, determine whether the deep Boltzmann machine model has converged, if it has converged, output the final deep Boltzmann machine parameter set, policy prior network parameters and value network parameters and end; if it has not converged, record the execution result and return to S1.

[0022] In this embodiment, the training dataset and knowledge graph in S1 specifically include: S11. Perform missing value imputation, outlier removal and format standardization on the collected data to generate a training dataset containing learner features, knowledge point features and resource features. S12. Construct a knowledge graph based on the prerequisite relationships of knowledge points, forming a directed graph structure with nodes representing knowledge points and edges representing prerequisite relationships.

[0023] In this embodiment, the first training layer in S3 specifically includes: S311. The first training layer receives input including the training dataset and the set of initial parameters for the deep Boltzmann machine. S312. Extract task sample sets from the training dataset according to task division, divide the task sample sets into support samples and validation samples, load the initial parameter set of the deep Boltzmann machine, perform forward propagation calculation on learner features, knowledge point features and resource features in the support samples, pass through the input layer, hidden layer and output layer in sequence, and calculate the output prediction value. S313. Calculate the absolute error of the support samples as the prediction error of the support samples. Take the square of the prediction error of the support samples and average it according to the number of samples to obtain the mean square error result. Add the sum of the squares of the parameter weights multiplied by the regularization coefficient to the loss calculation to form the regularization term. Add the regularization term to the mean square error result to obtain the total loss value. S314. Calculate the parameter gradients based on the total loss value, use the gradient descent algorithm in the optimized configuration, multiply the gradient of each parameter by the learning rate and then perform a weighted update, and perform gradient clipping operation on the updated parameters. S315. Use the validation samples to perform forward computation and obtain the prediction error of the validation samples; when the decrease in validation error is lower than the set threshold for several consecutive times, the training is considered to have converged. S316. After the convergence condition is met, the first training layer outputs the shared parameter set and the input standardized statistics. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

[0024] In this embodiment, the second training layer in S3 specifically includes: S321. The second training layer receives input including a shared parameter set, standardized input statistics, and a sample of the target learners. S322. Based on the input standardized statistical information, perform standardization processing on the learner features, knowledge point features and resource features in the target learner sample, and keep the distribution consistent with the input feature distribution of the first training layer to generate a standardized learner sample set. S323. Load the shared parameter set as the initial parameters of the second training layer, perform forward propagation calculation on the target learner sample, and obtain the second prediction value; S324. Calculate the absolute error of the target learner sample as the second prediction error, average the square of the second prediction error according to the number of samples, add the second regularization term, and obtain the second total loss value. S325. Perform backpropagation calculation based on the second total loss value to obtain the second gradient set of parameters of each layer. Use the gradient descent algorithm to multiply the gradient of each parameter by the second learning rate for weighted update, and perform the second gradient pruning operation. S326. When the second verification error decreases below a set threshold for several consecutive times, the training of the second training layer is determined to be converged. S327. After the convergence condition is met, the second training layer outputs an individualized parameter set and a learner state vector. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

[0025] In this embodiment, S4 specifically includes: S41. Map learner state vectors to benefit results to generate a set of correspondences between learner features and benefit results; benefit results refer to quantitative indicators of learning effectiveness calculated based on learners' learning completion status, test performance, and resource usage feedback in historical interaction logs. S42. Based on the set of correspondences, train the policy prior network and the value network, and establish the mapping relationship from learning state to policy distribution and from learning state to benefit evaluation, respectively. S43. During training, simultaneously optimize the parameters of the policy prior network and the value network until the outputs of the two networks reach the convergence condition, and output the trained policy prior network and value network.

[0026] In this embodiment, S5 specifically includes: S51. Calculate the set of knowledge points that the learner has not yet mastered based on the knowledge graph and the learner's state vector; S52. Based on the prerequisite relationships of knowledge points in the knowledge graph, select learning resources from the learning resource library that are related to the knowledge points that have not been mastered and whose prerequisite conditions are met. S53. Calculate the fit score for each resource by combining the learner's historical learning behavior records and resource usage preferences, and select learning resources with fit scores higher than a preset threshold to generate a candidate resource set.

[0027] In this embodiment, S53 specifically includes: The learning behavior score is obtained by weighted summation of the completion rate, accuracy rate, and learning time percentage of a resource, with weights of 0.5, 0.3, and 0.2, respectively; the resource preference is the learner's frequency of use of different types, divided into video, text, and quizzes; the knowledge relevance is obtained from the knowledge graph, with 1 for a perfect match, 0.5 for a partial match, and 0 for a no match; the fit score is obtained by weighted summation of the learner's behavior score, resource preference, and knowledge point relevance, with weights of 0.4, 0.3, and 0.3, respectively.

[0028] In this embodiment, S6 specifically includes: S61, Node Selection Phase: The policy prior network is invoked, inputting the learner's state vector and the candidate resource set, and outputting the prior probability of each candidate action; the individualized parameter set is invoked, inputting the learner's state vector and the candidate resource set, and calculating the mastery improvement value; historical returns of the candidate resource set are extracted from the correspondence set, and the mastery improvement value and historical returns are weighted and summed with weights of 0.6 and 0.4 to obtain the immediate return; exploration gain is calculated based on the number of node visits, and the prior probability, immediate return, and exploration gain are weighted with weights of 0.45, 0.4, and 0.15 to form a comprehensive score, selecting the candidate action with the highest comprehensive score to proceed to the next node; S62, Node Expansion Phase: Generate new nodes based on candidate actions, construct corresponding state vectors, and initialize node visit counts and average path revenue; S63. Simulation Phase: Starting from the state vector of the new node, the policy prior network is called step by step to generate action probabilities and actions are sampled according to the probabilities; the individualized parameter set is called to calculate the immediate reward at each step, and the immediate reward is weighted and accumulated according to the path depth decay coefficient to obtain the cumulative immediate reward; the value network is called to input the simulation end state vector to obtain the end value assessment, and the cumulative immediate reward and the end value assessment are weighted with weights of 0.7 and 0.3 to form the simulation reward value; S64. Backtracking Phase: The simulation reward value is backtracked sequentially from the leaf node to the root node along the current path. The number of visits to each node and the average path revenue are updated. The average path revenue is updated by weighting the current average path revenue and the simulation reward value. The number of visits to the nodes is weighted by the prior probability of the action and the comprehensive score is refreshed. S65. Termination and Output Phase: Repeat S61 to S64 until the maximum number of simulations is reached or the average path reward change between two consecutive searches is lower than the convergence threshold, at which point the search is terminated; calculate the average path reward of each action under the root node as the path score, generate a set of candidate learning paths, sort them by path score, select the target learning path, and output it.

[0029] In this embodiment, the specific formula for the exploration reward in S61 is as follows: the ratio of the logarithm of the number of visits to the parent node to the number of visits to the current node is taken as the square root of the ratio and multiplied by the exploration factor to obtain the exploration reward.

[0030] In this embodiment, the path depth attenuation coefficient in S63 is specifically composed of an attenuation factor set to 0.9 and an attenuation coefficient that is the attenuation factor raised to the power of the path depth.

[0031] In this embodiment, the simulation end state vector in S63 represents the final state of the system when it reaches the leaf node from the root node during the path search process.

[0032] In this embodiment, S64 specifically includes: Each node's access count is incremented by 1; the average path revenue is obtained by weighting the current node's historical average revenue and the simulated return value by (1-a) and a, where a is the reciprocal of the current node's access count plus one.

[0033] In this embodiment, S7 specifically includes: S71. Sort the candidate learning paths according to their path scores, and select the learning path with the highest path score as the target learning path. S72. Determine whether the deep Boltzmann machine model has converged. The convergence criterion is: if the path score changes by less than 0.01 in 5 consecutive training sessions, the deep Boltzmann machine model is considered to have converged. S73. If the deep Boltzmann machine model converges, output the final set of deep Boltzmann machine parameters, policy prior network parameters, and value network parameters, and end the learning path recommendation process. S74. If the deep Boltzmann machine model fails to converge, record the current execution result and return to S1 for the next round of training until the convergence condition is met or the preset maximum training round of 1000 is reached.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a personalized learning path planning scenario in an educational intelligent recommendation system. In this scenario, learners study knowledge points through an online learning platform, and the system dynamically recommends the optimal learning path based on their historical learning behavior, test scores, and resource usage records. Existing systems generally suffer from homogenized recommended content, lagging path updates, and insufficient response to learners' differentiated characteristics. This invention introduces individualized modeling based on deep Boltzmann machines and an AlphaZero-style Monte Carlo tree search algorithm to achieve global and local collaborative optimization in learner feature modeling, strategy guidance, and path search, thereby significantly improving the accuracy and dynamic response capability of learning path recommendations.

[0035] In the application, learning logs and test data from 1200 learners were first collected, containing approximately 86,000 learning records and 12,000 knowledge point resources. The system utilizes a deep Boltzmann machine to construct learner state representations. The shared parameters are optimized through a first training layer, and then rapidly updated for individual samples by a second training layer, generating individualized parameter sets. Subsequently, a mapping from learning states to rewards is established by combining a policy prior network and a value network, and path search and policy updates are completed using an AlphaZero-style Monte Carlo tree search algorithm. After sorting the learning paths according to reward scores, the optimal recommendation is output.

[0036] In the experiment, the recommendation performance of traditional collaborative filtering algorithms, deep reinforcement learning algorithms, and the method of this invention was compared. Average learning gain, path accuracy and resource matching degree, and learning completion rate were used as the main evaluation indicators. Experimental results show that the method of this invention achieves significant improvements at different learning stages. Especially under complex knowledge graph structures, path accuracy is improved by approximately 15.7%, and learning gain is increased by 22.3%, indicating that this method can better capture learner state changes and dynamically adjust path recommendation results.

[0037] In further experiments, the test group and the control group each consisted of 300 learners, with the difficulty and number of learning resources kept consistent. After comparing learning records for eight consecutive weeks, the test group achieved an average learning completion rate of 92.4%, an increase of 18.9% compared to the control group; the average improvement in knowledge mastery was 0.73, higher than the control group's 0.56. Learning time utilization improved by approximately 20%, and learners' path adjustment response time was shortened by 32%.

[0038] These results demonstrate that the method of this invention has significant advantages in dynamic recommendation and personalized response, effectively improving learning efficiency and recommendation accuracy. From a system perspective, this invention maintains computational efficiency even in scenarios with large-scale concurrent users. The average computation time for each path recommendation is reduced from 3.6 seconds in traditional algorithms to 2.1 seconds, and system resource utilization is reduced by approximately 25%. Even with more than 5000 learners, the algorithm maintains stability above 98%, and recommendation latency fluctuations are less than 0.4 seconds, exhibiting good scalability and real-time performance. Overall, the results show that this invention not only outperforms existing methods in personalized learning recommendation accuracy but also improves computational performance and model robustness.

[0039] The table below summarizes the experimental results of the method of this invention and the control algorithm.

[0040] Table 1. Performance Comparison Results of Learning Path Recommendation

[0041] As can be seen from Table 1, the method of the present invention outperforms traditional collaborative filtering and deep reinforcement learning methods in all key indicators, especially in terms of path accuracy and learning completion rate.

[0042] The path accuracy improved from 74.6% to 89.2%, indicating that the method can accurately identify the learner's optimal learning path under complex knowledge graph conditions; the average learning gain increased by 22.3%, reflecting the advantages of the method in knowledge mastery and learning transfer.

[0043] Improved learning time utilization and path response time demonstrate that the system possesses higher real-time performance and computational efficiency. The algorithm maintains a stability of over 98% under large-scale concurrency environments, fully proving the feasibility, stability, and superior performance of this invention in personalized learning recommendations.

[0044] 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 personalized learning path recommendation method based on deep learning, characterized in that, Includes the following steps: S1. Collect learning interaction logs, learning resource data and knowledge point prerequisite relationships, perform data preprocessing, and generate training datasets and knowledge graphs; S2. Construct a deep Boltzmann machine model based on the training dataset, and complete the parameter initialization and structure setting of the deep Boltzmann machine model; S3. The process of improving the deep Boltzmann machine model is to optimize the parameter sharing for different learning tasks in the first training layer, and to perform fast parameter updates based on learner samples in the second training layer, generating an individualized parameter set and outputting the learner state vector. S4. Use the state representation and reward results output by the deep Boltzmann machine model to train the policy prior network and value network; S5. Based on the knowledge graph and learner state vector, select learning resources, generate a candidate resource set, and use the Monte Carlo tree search algorithm to perform node selection, expansion, simulation and backtracking. S6. Perform the AlphaZero-style prior-guided improvement process of Monte Carlo tree search, call the policy prior network and value network parameters to select nodes, call the individualized parameter set to calculate the completion probability and learning benefit, search and update the path benefit according to the PUCT rule, and generate candidate learning paths. S7. Select the target learning path based on the candidate learning path scores, determine whether the deep Boltzmann machine model has converged, and if it has converged, output the final deep Boltzmann machine parameter set, policy prior network parameters and value network parameters and end. If convergence is not achieved, record the execution result and return to S1.

2. The personalized learning path recommendation method based on deep learning according to claim 1, characterized in that, The training dataset and knowledge graph in S1 specifically include: S11. Perform missing value imputation, outlier removal and format standardization on the collected data to generate a training dataset containing learner features, knowledge point features and resource features. S12. Construct a knowledge graph based on the prerequisite relationships of knowledge points, forming a directed graph structure with nodes representing knowledge points and edges representing prerequisite relationships.

3. The personalized learning path recommendation method based on deep learning according to claim 2, characterized in that, The first training layer in S3 specifically includes: S311. The first training layer receives input including the training dataset and the set of initial parameters for the deep Boltzmann machine. S312. Extract task sample sets from the training dataset according to task division, divide the task sample sets into support samples and validation samples, load the initial parameter set of the deep Boltzmann machine, perform forward propagation calculation on learner features, knowledge point features and resource features in the support samples, pass through the input layer, hidden layer and output layer in sequence, and calculate the output prediction value. S313. Calculate the absolute error of the support samples as the prediction error of the support samples. Take the square of the prediction error of the support samples and average it according to the number of samples to obtain the mean square error result. Add the sum of the squares of the parameter weights multiplied by the regularization coefficient to the loss calculation to form the regularization term. Add the regularization term to the mean square error result to obtain the total loss value. S314. Calculate the parameter gradients based on the total loss value, use the gradient descent algorithm in the optimized configuration, multiply the gradient of each parameter by the learning rate and then perform a weighted update, and perform gradient clipping operation on the updated parameters. S315. Use the validation samples to perform forward computation and obtain the prediction error of the validation samples; when the decrease in validation error is lower than the set threshold for several consecutive times, the training is considered to have converged. S316. After the convergence condition is met, the first training layer outputs the shared parameter set and the input standardized statistics. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

4. The personalized learning path recommendation method based on deep learning according to claim 3, characterized in that, The second training layer in S3 specifically includes: S321. The second training layer receives input including a shared parameter set, standardized input statistics, and a sample of the target learners. S322. Based on the input standardized statistical information, perform standardization processing on the learner features, knowledge point features and resource features in the target learner sample, and keep the distribution consistent with the input feature distribution of the first training layer to generate a standardized learner sample set. S323. Load the shared parameter set as the initial parameters of the second training layer, perform forward propagation calculation on the target learner sample, and obtain the second prediction value; S324. Calculate the absolute error of the target learner sample as the second prediction error, average the square of the second prediction error according to the number of samples, add the second regularization term, and obtain the second total loss value. S325. Perform backpropagation calculation based on the second total loss value to obtain the second gradient set of parameters of each layer. Use the gradient descent algorithm to multiply the gradient of each parameter by the second learning rate for weighted update, and perform the second gradient pruning operation. S326. When the second verification error decreases below a set threshold for several consecutive times, the training of the second training layer is determined to be converged. S327. After the convergence condition is met, the second training layer outputs an individualized parameter set and a learner state vector. If the convergence condition is not met, the next round of training continues until the preset training round or convergence judgment condition is reached.

5. The personalized learning path recommendation method based on deep learning according to claim 4, characterized in that, S4 specifically includes: S41. Map learner state vectors to benefit results to generate a set of correspondences between learner features and benefit results; benefit results refer to quantitative indicators of learning effectiveness calculated based on learners' learning completion status, test performance, and resource usage feedback in historical interaction logs. S42. Based on the set of correspondences, train the policy prior network and the value network, and establish the mapping relationship from learning state to policy distribution and from learning state to benefit evaluation, respectively. S43. During training, simultaneously optimize the parameters of the policy prior network and the value network until the outputs of the two networks reach the convergence condition, and output the trained policy prior network and value network.

6. The personalized learning path recommendation method based on deep learning according to claim 5, characterized in that, S5 specifically includes: S51. Calculate the set of knowledge points that the learner has not yet mastered based on the knowledge graph and the learner's state vector; S52. Based on the prerequisite relationships of knowledge points in the knowledge graph, select learning resources from the learning resource library that are related to the knowledge points that have not been mastered and whose prerequisite conditions are met. S53. Calculate the fit score for each resource by combining the learner's historical learning behavior records and resource usage preferences, and select learning resources with fit scores higher than a preset threshold to generate a candidate resource set.

7. The personalized learning path recommendation method based on deep learning according to claim 6, characterized in that, S6 specifically includes: S61, Node Selection Phase: The learner's state vector and candidate resource set are input into the policy prior network, and the prior probability of each candidate action is output. The learner's state vector and candidate resource set are input into the individualized parameter set, and the mastery improvement value is calculated. The historical reward results of the candidate resource set are extracted from the correspondence set, and the mastery improvement value and historical reward value are weighted and summed according to preset weights to obtain the immediate reward. The exploration gain is calculated by combining the node visit count, and the prior probability, immediate reward, and exploration gain are weighted to form a comprehensive score. The candidate action with the highest comprehensive score is selected to enter the next node. S62, Node Expansion Phase: Generate new nodes based on candidate actions, construct corresponding state vectors, and initialize node visit counts and average path revenue; S63. Simulation Phase: Starting from the state vector of the new node, the policy prior network is called step by step to generate action probabilities and actions are sampled according to the probabilities; the individualized parameter set is called to calculate the immediate benefits of each step, and the immediate benefits are weighted and accumulated according to the path depth decay coefficient to obtain the cumulative immediate benefits; the value network is called to input the simulation end state vector to obtain the end value assessment, and the cumulative immediate benefits and the end value assessment are weighted according to the weights to form the simulation return value. S64. Backtracking Phase: The simulation reward value is backtracked sequentially from the leaf node to the root node along the current path. The number of visits to each node and the average path revenue are updated. The average path revenue is updated by weighting the current average path revenue and the simulation reward value. The number of visits to the nodes is weighted by the prior probability of the action and the comprehensive score is refreshed. S65. Termination and Output Phase: Repeat S61 to S64 until the maximum number of simulations is reached or the average path reward change between two consecutive searches is lower than the convergence threshold, at which point the search is terminated; calculate the average path reward of each action under the root node as the path score, generate a set of candidate learning paths, sort them by path score, select the target learning path, and output it.

8. The personalized learning path recommendation method based on deep learning according to claim 7, characterized in that, Specifically, S7 includes: S71. Sort the candidate learning paths according to their path scores, and select the learning path with the highest path score as the target learning path. S72. Determine whether the deep Boltzmann machine model has converged. The convergence criterion is: if the path score changes by less than 0.01 in 5 consecutive training sessions, the deep Boltzmann machine model is considered to have converged. S73. If the deep Boltzmann machine model converges, output the final set of deep Boltzmann machine parameters, policy prior network parameters, and value network parameters, and end the learning path recommendation process. S74. If the deep Boltzmann machine model fails to converge, record the current execution result and return to S1 for the next round of training until the convergence condition is met or the preset maximum training round of 1000 is reached.

Images