A machine learning-based personalized recommendation method for educational resources
By combining labeled SOM neural networks and gradient boosting machines, the problems of high computational cost and poor recommendation performance in personalized recommendation of educational resources are solved, achieving efficient and refined personalized recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SANYUN (HUBEI) DIGITAL TECHNOLOGY CO LTD
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-09
Smart Images
Figure CN122173706A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of personalized recommendation technology, and in particular to a method for personalized recommendation of educational resources based on machine learning. Background Technology
[0002] The number of educational resources on digital platforms continues to grow, requiring students to choose from a vast array of options during their learning process. Traditional manual methods struggle to provide highly tailored learning content based on individual student differences. Existing personalized recommendation technologies primarily rely on collaborative filtering or static feature-based matching methods, generating recommendation lists through student behavior records or student profiles. Collaborative filtering methods depend on historical similarities between students or resources; their effectiveness declines when student data is sparse or resources are frequently updated. Static feature-based matching methods struggle to capture dynamic changes in resource content structure and student learning status, resulting in insufficient recommendation accuracy.
[0003] Regarding resource organization, existing technologies mostly rely on unsupervised clustering methods to construct resource classification structures, but fail to utilize resource label information. This results in clustering results that fail to reflect the semantic differences between resources, impacting the learning performance of subsequent recommendation models. During recommendation model training, existing gradient boosting models typically perform a complete weak learner iteration on all training samples, failing to distinguish between fully learned samples and those still requiring training. This leads to redundant computation, reducing model training efficiency and inference speed. Especially with a large number of educational resources and diverse student behavior data, this uniform iterative approach rapidly increases the model's computational load, making it difficult to adapt to the real-time recommendation needs of large-scale educational resource scenarios.
[0004] Therefore, how to provide a machine learning-based personalized recommendation method for educational resources is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a personalized recommendation method for educational resources based on machine learning. This invention utilizes a labeled SOM neural network to construct a semantic structure for resources and combines it with a gradient boosting machine with an early termination mechanism to achieve efficient matching of students and resources. This method can improve recommendation accuracy and reduce computational costs in large-scale resource scenarios and has the advantages of clear structure and fast response speed.
[0006] A personalized recommendation method for educational resources based on machine learning according to an embodiment of the present invention includes the following steps: S1. Acquire student data and educational resource data, perform preprocessing, and form student feature sets and resource feature sets; S2. Construct a SOM neural network based on the resource feature set, initialize the topology, node weight vector and node label distribution vector, and write some resource labels into the corresponding node label distribution vector. S3. Input the resource features into the SOM neural network, update the node weight vector, and update the neighborhood node label distribution vector according to the resource label to obtain the node representation. S4. Cluster resources using node representation to generate a set of resource clusters, and record the identifier of the resource cluster to which each resource belongs; S5. Construct the first gradient booster based on the student feature set, resource feature set and resource cluster identifier, and initialize the weak learner sequence, loss function and sample-level termination decision variable; S6. During the training of the first gradient booster, the prediction error and the sample-level termination decision variable determine whether to stop the subsequent weak learners from updating the current sample. After the training of the first gradient booster is completed, the matching score between the student and each resource cluster is output. S7. Select the target resource cluster based on the matching score, construct the second gradient booster using the student feature set and the resource feature set within the target resource cluster, train and generate the matching score between the student and the resources within the target resource cluster, and output the personalized recommendation results of educational resources according to the matching score.
[0007] Optionally, S2 specifically includes: S21. Set the number of nodes, node arrangement, and neighborhood function parameters of the SOM neural network, and initialize the weight vector of each node as an initial vector consistent with the resource feature dimension. S22. Establish a node label distribution vector for each node in the SOM neural network, set the dimension of the node label distribution vector to the resource label dimension, and initialize all elements of the node label distribution vector to zero. S23. Based on the resource feature set and the corresponding resource tags, write some resource tags into the node tag distribution vector according to the distribution of resources in the feature space.
[0008] Optionally, S3 specifically includes: S31. Read the resource feature vector and resource tag information sequentially from the resource feature set; S32. For a single resource feature vector, calculate the distance between the resource feature vector and the weight vectors of all nodes in the SOM neural network, and determine the node with the smallest distance as the best matching node for the current resource feature vector. S33. Determine the set of nodes that have a neighborhood relationship with the best matching node based on the neighborhood function parameters, and record it as the set of neighborhood nodes corresponding to the current resource feature vector; S34. For the best matching node and each node in the set of neighboring nodes, update the node weight vector according to the preset learning rate parameter and the value of the neighborhood function. S36. When the resource feature vector contains resource label information, for the best matching node and each node in the neighborhood node set, calculate the label update coefficient based on the neighborhood function value, and perform weighted cumulative update on the corresponding position of the resource label in the node label distribution vector according to the label update coefficient. S37. When the resource feature vector does not contain resource label information, obtain the label distribution vector of each node in the neighborhood node set, calculate the weighted average result of the label distribution based on the neighborhood function value, and write the weighted average result of the label distribution into the label distribution vector of the best matching node to form the label distribution update result corresponding to the unlabeled resource. S38. Repeat S31 to S37 until all resource feature vectors in the resource feature set have been processed to obtain the updated node weight vector and node label distribution vector, and generate the node representation.
[0009] Optionally, S4 specifically includes: A K-means clustering method based on node weight vectors is adopted, and the number of clusters is set. After clustering is completed, the cluster number of each node is used as the node cluster identifier; according to the corresponding best matching node, the resource feature vector is assigned to the resource cluster with the same cluster number as the best matching node, and the cluster number is recorded as the resource cluster identifier.
[0010] Optionally, S5 specifically includes: S51. Construct training samples for the first gradient booster based on student feature set, resource feature set and resource cluster identifier, set the number of weak learners, weak learner structure and loss function type of the gradient booster, and initialize the weak learner sequence. S52. Generate the initial value of the sample-level termination determination variable and initialize all elements of the sample-level termination determination variable to the non-termination state. S53. Initialize the learning rate parameters, iteration round parameters, and residual calculation method of the first gradient booster to form the initial model of the first gradient booster.
[0011] Optionally, the calculation process of the termination determination variable in S6 specifically includes: At the start of the t-th iteration of the first gradient booster, read the sample-level termination decision variables corresponding to all training samples, and perform subsequent calculations on the training samples that are in the non-termination state. For training samples that are not terminated, according to the feature splitting structure of the weak learner in the t-th round, record the splitting node sequence of the training sample in the weak learner in the t-th round, form the path vector in the t-th round, and store the path vector in the path recording sequence. For training samples that are not terminated, read the path vector of round t and the path vector of round t-1 in the path record sequence, calculate the path change between the two path vectors, compare the path change with the path stability threshold, and generate a path change judgment result. For training samples that are not terminated, based on the leaf node identifiers that appear in the path vector of round t, query the recorded weak learner leaf node coverage table. When the leaf node in the path vector of round t has appeared in the previous weak learner, and the path change judgment results of several consecutive rounds all indicate that the path change amount is lower than the path stability threshold, update the sample-level termination judgment variable to the termination state.
[0012] Optionally, S6 specifically includes: S61. In each iteration of the first gradient booster, for training samples that are not terminated, the predicted value of the training sample is calculated based on the current weak learner, and the residual of the training sample is calculated based on the predicted value and the target value. S62. For training samples that are not terminated, generate corresponding residual labels based on the residuals and write the residual labels into the residual information record of the weak learner. S63. For training samples in the terminated state, write the predicted value from the previous iteration into the prediction record as the predicted value for the current iteration, and do not perform the residual label calculation process. S64. After processing all training samples, update the parameters of the current round weak learner based on the residual labels of the training samples that are not in the terminated state, complete the training of the current round weak learner and write it into the iteration record.
[0013] Optionally, the convergence conditions for the first gradient booster specifically include: Calculate the change in predicted value of training samples that are not yet terminated in the current round, and calculate the average change in predicted value of all training samples that are not yet terminated. When the average change is lower than the preset change convergence threshold in several consecutive iterations, the first gradient booster is determined to have reached the convergence condition. When the number of training samples in the unterminated state remains constant for several consecutive iterations, the first gradient booster is considered to have reached the convergence condition. When the first gradient booster reaches any convergence condition, the training of subsequent weak learners is stopped, and the training result of the first gradient booster is output.
[0014] Optionally, S7 specifically includes: After the first gradient booster reaches the convergence condition, the target resource cluster is selected based on the student-resource cluster matching score output by the first gradient booster. Training samples for the second gradient booster are constructed based on the student feature set and the resource feature set within the target resource cluster. The number of weak learners, the structure of the weak learners, the type of loss function, and the initialization method of the sample-level termination decision variable of the second gradient booster are set to the same model structure as the first gradient booster. The second gradient booster is trained using the same training process as the first gradient booster, generating matching scores between students and resources within the target resource cluster, and outputting personalized educational resource recommendations based on the matching scores.
[0015] The beneficial effects of this invention are: (1) By introducing a labeled SOM neural network to construct a resource semantic structure, the resource clustering results can simultaneously reflect the resource content features and label features, improve the accuracy and interpretability of resource organization, provide a more stable resource structure foundation for subsequent recommendation models, and overcome the shortcomings of existing unsupervised clustering in reflecting the semantic differences of resources.
[0016] (2) By designing a sample-level early termination mechanism based on path vector changes and leaf node coverage tables, fine-grained control of the gradient booster iteration process is achieved, so that the fully learned samples automatically stop iterating during the training process, significantly reducing the amount of training computation and model redundancy updates, improving model training efficiency, and solving the problem of high computational cost caused by unified iteration in the existing technology.
[0017] (3) By constructing a two-level matching structure of the first gradient booster and the second gradient booster, matching scores are performed at the resource cluster level and the resource level respectively, which improves the refinement and personalization of the recommendation results, and ensures the accuracy of the recommendation while maintaining the computational efficiency, thus meeting the real-time recommendation needs in large-scale educational resource scenarios. 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: Figure 1 This is a flowchart of a machine learning-based personalized recommendation method for educational resources proposed in this invention. Figure 2 This is a schematic diagram of the updated mechanism of the labeled SOM in the personalized recommendation method for educational resources based on machine learning proposed in this invention. Figure 3 This is a diagram of the first gradient boosting machine training structure for a machine learning-based personalized recommendation method for educational resources proposed in this invention. Detailed Implementation
[0019] 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.
[0020] refer to Figure 1-3 A personalized recommendation method for educational resources based on machine learning includes the following steps: S1. Acquire student data and educational resource data, perform preprocessing, and form student feature sets and resource feature sets; S2. Construct a SOM neural network based on the resource feature set, initialize the topology, node weight vector and node label distribution vector, and write some resource labels into the corresponding node label distribution vector. S3. Input the resource features into the SOM neural network, update the node weight vector, and update the neighborhood node label distribution vector according to the resource label to obtain the node representation. S4. Cluster resources using node representation to generate a set of resource clusters, and record the identifier of the resource cluster to which each resource belongs; S5. Construct the first gradient booster based on the student feature set, resource feature set and resource cluster identifier, and initialize the weak learner sequence, loss function and sample-level termination decision variable; S6. During the training of the first gradient booster, the prediction error and the sample-level termination decision variable determine whether to stop the subsequent weak learners from updating the current sample. After the training of the first gradient booster is completed, the matching score between the student and each resource cluster is output. S7. Select the target resource cluster based on the matching score, construct the second gradient booster using the student feature set and the resource feature set within the target resource cluster, train and generate the matching score between the student and the resources within the target resource cluster, and output the personalized recommendation results of educational resources according to the matching score.
[0021] In this embodiment, S1 specifically includes: Acquire student data and educational resource data. Student data consists of student profile data and student behavior data. Student profile data includes grade, major, learning stage, learning preference vector, and historical academic performance. Student behavior data includes learning duration records, access frequency records, resource click records, learning sequence records, question answer records, answer accuracy records, and learning time distribution records. Educational resource data consists of resource content features and resource attribute features. Resource content features include text content features, video content features, question content features, and knowledge point encoding features. Resource attribute features include resource type, resource difficulty, knowledge point category, applicable grade, historical completion rate, historical accuracy rate, and resource tags. After acquiring student data and educational resource data, preprocessing is performed on the student data and educational resource data. The preprocessing process includes missing value handling, outlier handling, normalization, feature encoding, and vectorization.
[0022] In this embodiment, S2 specifically includes: S21. Set the number of nodes, node arrangement, and neighborhood function parameters of the SOM neural network. Initialize the weight vector of each node to an initial vector consistent with the resource feature dimension. Set the number of nodes in the network according to the number of samples in the resource feature set. Set the number of nodes to a matrix structure formed by combining the number of rows and columns obtained by rounding up the square root of the number of samples. Set the initial neighborhood radius to half of the larger value between the number of rows and columns of the node, and use a Gaussian neighborhood function based on topological distance as the neighborhood calculation method. During training, the neighborhood radius is updated linearly from the initial value to 1 according to the training rounds. Set the initial learning rate to 10%, and gradually reduce the learning rate according to the preset decay strategy during training. During the initialization of the node weight vector, set the dimension of the weight vector of each node to be consistent with the resource feature dimension. Calculate the minimum and maximum values for each dimension of the resource feature set, and generate the node weight vector in a uniform random manner between the minimum and maximum values.
[0023] S22. Establish a node label distribution vector for each node in the SOM neural network, set the dimension of the node label distribution vector to the resource label dimension, and initialize all elements of the node label distribution vector to zero. S23. Based on the resource feature set and the corresponding resource tags, write some resource tags into the node tag distribution vector according to the distribution of resources in the feature space.
[0024] In this embodiment, S3 specifically includes: S31. Read the resource feature vector and resource tag information sequentially from the resource feature set; S32. For a single resource feature vector, calculate the distance between the resource feature vector and the weight vectors of all nodes in the SOM neural network, and determine the node with the smallest distance as the best matching node for the current resource feature vector; in this embodiment, Euclidean distance is used. S33. Determine the set of nodes that have a neighborhood relationship with the best matching node based on the neighborhood function parameters, and record it as the neighborhood node set corresponding to the current resource feature vector; when the distance between a node and the best matching node in the topology grid is less than the neighborhood radius, the node is included in the neighborhood node set; the initial neighborhood radius is set to half of the larger value between the number of rows and columns of the node grid; S34. For the best matching node and each node in the neighborhood node set, update the node weight vector according to the preset learning rate parameter and the neighborhood function value; perform the update sequentially for each node according to the single iteration order; the learning rate is set to 10% and decreases as the training process progresses.
[0025] S36. When the resource feature vector contains resource tag information, for the best matching node and each node in the neighborhood node set, calculate the tag update coefficient based on the neighborhood function value, and perform weighted cumulative update on the corresponding position of the resource tag in the node tag distribution vector according to the tag update coefficient; a Gaussian neighborhood function is used. S37. When the resource feature vector does not contain resource label information, obtain the label distribution vector of each node in the neighborhood node set, calculate the weighted average result of the label distribution based on the neighborhood function value, and write the weighted average result of the label distribution into the label distribution vector of the best matching node to form the label distribution update result corresponding to the unlabeled resource. S38. Repeat S31 to S37 until all resource feature vectors in the resource feature set have been processed to obtain the updated node weight vector and node label distribution vector, and generate the node representation; the node representation consists of the node weight vector and the node label distribution vector.
[0026] In this embodiment, S4 specifically includes: A K-means clustering method based on node weight vectors is adopted, and the number of clusters is set; in this embodiment, the number of clusters is set to the square root of the number of nodes. After clustering is completed, the cluster number of each node is used as the node cluster identifier; according to the corresponding best matching node, the resource feature vector is assigned to the resource cluster with the same cluster number as the best matching node, and the cluster number is recorded as the resource cluster identifier.
[0027] In this embodiment, S5 specifically includes: S51. Construct training samples for the first gradient booster based on student feature set, resource feature set and resource cluster identifier. Set the number of weak learners in the gradient booster to 50, the structure of the weak learners to be a decision tree with a depth of 4 and the loss function type to be squared error, and initialize the weak learner sequence. S52. Generate the initial value of the sample-level termination determination variable and initialize all elements of the sample-level termination determination variable to the non-termination state. S53. Initialize the learning rate parameters, iteration round parameters, and residual calculation method of the first gradient booster to form the initial model of the first gradient booster; the learning rate is 10%, decays to 0.8, and the residual calculation method is set to "training sample predicted value minus the true value", and the calculation result is used as the target value of the weak learner in this round; when the training sample is in the termination state, the residual is not calculated, and the prediction value of the previous round is directly used.
[0028] In this embodiment, the calculation process of the termination determination variable in S6 specifically includes: At the start of the t-th iteration of the first gradient booster, read the sample-level termination decision variables corresponding to all training samples, and perform subsequent calculations on the training samples that are in the non-termination state. For training samples that are not terminated, according to the feature splitting structure of the weak learner in the t-th round, record the splitting node sequence of the training sample in the weak learner in the t-th round, form the path vector in the t-th round, and store the path vector in the path recording sequence. For training samples that are not terminated, read the path vector of round t and the path vector of round t-1 in the path record sequence, calculate the path change between the two path vectors, compare the path change with the path stability threshold, and generate a path change judgment result. For training samples in an unterminated state, based on the leaf node identifiers appearing in the path vector of round t, the recorded weak learner leaf node coverage table is queried. If a leaf node in the path vector of round t has already appeared in a previous weak learner, and the path change judgment results for several consecutive rounds indicate that the path change amount is lower than the path stability threshold, the sample-level termination judgment variable is updated to the termination state; the consecutive rounds are 3 rounds; the leaf node recurrence condition is that the leaf node number in the path vector of round t has already appeared in the leaf node coverage table; the path change amount uses Hamming distance; path stability is defined as the path change amount not exceeding one node in the current round.
[0029] In this embodiment, S6 specifically includes: S61. In each iteration of the first gradient booster, for training samples that are not terminated, the predicted value of the training sample is calculated based on the current weak learner, and the residual of the training sample is calculated based on the predicted value and the target value. S62. For training samples that are not terminated, generate corresponding residual labels based on the residuals and write the residual labels into the residual information record of the weak learner. S63. For training samples in the terminated state, write the predicted value from the previous iteration into the prediction record as the predicted value for the current iteration, and do not perform the residual label calculation process. S64. After processing all training samples, update the parameters of the current round weak learner based on the residual labels of the training samples that are not in the terminated state, complete the training of the current round weak learner and write it into the iteration record.
[0030] In this embodiment, the convergence conditions of the first gradient booster specifically include: Calculate the change in predicted value of training samples that are not yet terminated in the current round, and calculate the average change in predicted value of all training samples that are not yet terminated. When the average change is lower than the preset change convergence threshold for several consecutive iterations, the first gradient booster is determined to have reached the convergence condition; the preset change convergence threshold is set to 1 / 1000 for 5 consecutive iterations. The first gradient booster is considered to have reached convergence when the number of training samples in the unterminated state remains constant for several consecutive iterations; here, this is set to 10 consecutive iterations. When the first gradient booster reaches any convergence condition, the training of subsequent weak learners is stopped, and the training result of the first gradient booster is output.
[0031] In this embodiment, S7 specifically includes: After the first gradient booster reaches the convergence condition, the target resource cluster is selected based on the student-resource cluster matching score output by the first gradient booster. Training samples for the second gradient booster are constructed based on the student feature set and the resource feature set within the target resource cluster. The initialization method of the number of weak learners, the structure of the weak learners, the type of loss function, and the sample-level termination decision variable of the second gradient booster are set to the same model structure as the first gradient booster. The training samples of the second gradient booster are composed of the student feature vector and the resource feature vector within the target resource cluster. The label of the second gradient booster is the historical interaction with the resource.
[0032] The second gradient booster is trained using the same training process as the first gradient booster, generating matching scores between students and resources within the target resource cluster, and outputting personalized educational resource recommendations based on the matching scores.
[0033] The input to the first gradient booster consists of the student feature vector, the resource feature vector, and the resource cluster identifier. After training, the first gradient booster outputs a matching score between the student and the resource cluster, and uses the matching score to determine the target resource cluster corresponding to the student. The input to the second gradient booster consists of the student feature vector and the resource feature vectors of each resource in the target resource cluster. After training, the second gradient booster outputs a matching score between the student and each resource in the target resource cluster, and uses the matching score to obtain the recommendation ranking for specific resources.
[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-world educational application scenario: an intelligent system for recommending learning resources. This system deals with a large number of students and a vast amount of educational resources. Student characteristics include multi-dimensional data such as knowledge point mastery, learning speed, learning preference tags, learning stage, question answering records, and reading behavior records. Educational resource characteristics include resource type, knowledge point coverage, difficulty level, resource text vectors, resource behavior tags, and known classification tags for some resources. With over 3,000 students and over 50,000 resources, traditional single-stage recommendation models based on content similarity are insufficient to meet personalized needs and large-scale processing requirements at this scale.
[0035] In this embodiment, student features and educational resource features are standardized. Educational resource features are input into a SOM neural network to complete topology training and form resource clusters, ensuring that resources with similar content are assigned to the same category. Based on this, student features, resource features, and resource cluster identifiers are input into a first gradient booster to train students to match resource clusters with scores. This score is used to select the target resource clusters from which students are most likely to be interested. Student features and resource features from the target resource clusters are jointly input into a second gradient booster to learn the relationship between students and resources at a fine-grained level. In this two-level model structure, the sample-level termination decision variable, path vector recording method, and iterative stability rules are all executed according to the aforementioned claims, ensuring consistency throughout the training process.
[0036] To verify the application effect of this invention, it was compared with existing technologies. The existing technology uses a single-stage recommendation method based on content similarity as a control method, which relies solely on the similarity calculation between resource content and student profile vectors to complete the recommendation. Below, 10 students were randomly selected from the experimental results as examples to compare the performance of the method of this invention and the control method on the following indicators: Resource cluster matching score: The student-resource cluster matching score output by the first gradient booster of this invention; Resource matching score: Student's resource matching score output by the second gradient booster; Recommendation accuracy: The percentage of clicks made by students among the top 10 recommended items; Average search volume: The average number of resources a student needs to browse before obtaining the target resource.
[0037] Table 1: Performance Comparison of the Invention and the Baseline Model
[0038] As shown in Table 1, the method of this invention maintains a high level in both resource cluster matching score and resource matching score. The resource cluster matching score among the 10 students ranges from 0.76 to 0.85, while the resource matching score ranges from 0.86 to 0.94, indicating that the matching performance at the resource level is higher after resource cluster filtering. Regarding recommendation accuracy, the method of this invention uses the click-through rate of the top 10 recommended items as the observation indicator. The click-through rate among the 10 students ranges from 0.87 to 0.95, while the click-through rate of the baseline method ranges from 0.68 to 0.78, showing a numerical difference between the two technologies in this indicator.
[0039] Regarding average search volume, the method of this invention ranged from 105 to 145 among 10 students, while the baseline method ranged from 235 to 310. Average search volume measures the number of resources a student needs to browse before obtaining useful information; the comparison results show that the two methods output different data ranges. By comparing resource matching scores, click-through rates of the top 10 recommendations, and average search volume, it can be observed that the two-level recommendation structure and the baseline technology exhibit different data distribution characteristics under multiple student input scenarios.
[0040] Overall, this invention outputs quantifiable resource cluster scores and resource scores in a test sample of 10 students, and exhibits numerical performance different from the baseline technology in terms of click-through rate and search volume, reflecting the actual operating characteristics and data output patterns of the two-level model structure under different combinations of student characteristics.
[0041] 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 recommendation method for educational resources based on machine learning, characterized in that, Includes the following steps: S1. Acquire student data and educational resource data, perform preprocessing, and form student feature sets and resource feature sets; S2. Construct a SOM neural network based on the resource feature set, initialize the topology, node weight vector and node label distribution vector, and write some resource labels into the corresponding node label distribution vector. S3. Input the resource features into the SOM neural network, update the node weight vector, and update the neighborhood node label distribution vector according to the resource label to obtain the node representation. S4. Cluster resources using node representation to generate a set of resource clusters, and record the resource cluster identifier for each resource. S5. Construct the first gradient booster based on the student feature set, resource feature set, and resource cluster identifier, and initialize the weak learner sequence, loss function, and sample-level termination decision variable. S6. During the training of the first gradient booster, the prediction error and the sample-level termination decision variable determine whether to stop the subsequent weak learners from updating the current sample. After the training of the first gradient booster is completed, the matching score between the student and each resource cluster is output. S7. Select the target resource cluster based on the matching score, construct the second gradient booster using the student feature set and the resource feature set within the target resource cluster, train and generate the matching score between the student and the resources within the target resource cluster, and output the personalized recommendation results of educational resources according to the matching score.
2. The personalized recommendation method for educational resources based on machine learning according to claim 1, characterized in that, S2 specifically includes: S21. Set the number of nodes, node arrangement, and neighborhood function parameters of the SOM neural network, and initialize the weight vector of each node as an initial vector consistent with the resource feature dimension. S22. Establish a node label distribution vector for each node in the SOM neural network, set the dimension of the node label distribution vector to the resource label dimension, and initialize all elements of the node label distribution vector to zero. S23. Based on the resource feature set and the corresponding resource tags, write some resource tags into the node tag distribution vector according to the distribution of resources in the feature space.
3. The personalized recommendation method for educational resources based on machine learning according to claim 2, characterized in that, S3 specifically includes: S31. Read the resource feature vector and resource tag information sequentially from the resource feature set; S32. For a single resource feature vector, calculate the distance between the resource feature vector and the weight vectors of all nodes in the SOM neural network, and determine the node with the smallest distance as the best matching node for the current resource feature vector. S33. Determine the set of nodes that have a neighborhood relationship with the best matching node based on the neighborhood function parameters, and record it as the set of neighborhood nodes corresponding to the current resource feature vector. S34. For the best matching node and each node in the set of neighboring nodes, update the node weight vector according to the preset learning rate parameter and the value of the neighborhood function. S36. When the resource feature vector contains resource label information, for the best matching node and each node in the neighborhood node set, calculate the label update coefficient based on the neighborhood function value, and perform weighted cumulative update on the corresponding position of the resource label in the node label distribution vector according to the label update coefficient. S37. When the resource feature vector does not contain resource label information, obtain the label distribution vector of each node in the neighborhood node set, calculate the weighted average result of the label distribution based on the neighborhood function value, and write the weighted average result of the label distribution into the label distribution vector of the best matching node to form the label distribution update result corresponding to the unlabeled resource. S38. Repeat S31 to S37 until all resource feature vectors in the resource feature set have been processed to obtain the updated node weight vector and node label distribution vector, and generate the node representation.
4. The personalized educational resource recommendation method based on machine learning according to claim 3, characterized in that, S4 specifically includes: A K-means clustering method based on node weight vectors is adopted, and the number of clusters is set. After clustering is completed, the cluster number of each node is used as the node cluster identifier; according to the corresponding best matching node, the resource feature vector is assigned to the resource cluster with the same cluster number as the best matching node, and the cluster number is recorded as the resource cluster identifier.
5. The personalized recommendation method for educational resources based on machine learning according to claim 4, characterized in that, S5 specifically includes: S51. Construct training samples for the first gradient booster based on student feature set, resource feature set and resource cluster identifier, set the number of weak learners, weak learner structure and loss function type of the gradient booster, and initialize the weak learner sequence. S52. Generate the initial value of the sample-level termination determination variable and initialize all elements of the sample-level termination determination variable to the non-termination state. S53. Initialize the learning rate parameters, iteration round parameters, and residual calculation method of the first gradient booster to form the initial model of the first gradient booster.
6. The personalized recommendation method for educational resources based on machine learning according to claim 5, characterized in that, The calculation process for the termination determination variable in S6 specifically includes: At the start of the t-th iteration of the first gradient booster, read the sample-level termination decision variables corresponding to all training samples, and perform subsequent calculations on the training samples that are in the non-termination state. For training samples that are not terminated, according to the feature splitting structure of the weak learner in the t-th round, record the splitting node sequence of the training sample in the weak learner in the t-th round, form the path vector in the t-th round, and store the path vector in the path recording sequence. For training samples that are not terminated, read the path vector of round t and the path vector of round t-1 in the path record sequence, calculate the path change between the two path vectors, compare the path change with the path stability threshold, and generate a path change judgment result. For training samples that are not terminated, based on the leaf node identifiers that appear in the path vector of round t, query the recorded weak learner leaf node coverage table. When the leaf node in the path vector of round t has appeared in the previous weak learner, and the path change judgment results of several consecutive rounds all indicate that the path change amount is lower than the path stability threshold, update the sample-level termination judgment variable to the termination state.
7. The personalized recommendation method for educational resources based on machine learning according to claim 6, characterized in that, S6 specifically includes: S61. In each iteration of the first gradient booster, for training samples that are not terminated, the predicted value of the training sample is calculated based on the current weak learner, and the residual of the training sample is calculated based on the predicted value and the target value. S62. For training samples that are not terminated, generate corresponding residual labels based on the residuals and write the residual labels into the residual information record of the weak learner. S63. For training samples in the terminated state, write the predicted value from the previous iteration into the prediction record as the predicted value for the current iteration, and do not perform the residual label calculation process. S64. After processing all training samples, update the parameters of the current round weak learner based on the residual labels of the training samples that are not in the terminated state, complete the training of the current round weak learner and write it into the iteration record.
8. The personalized recommendation method for educational resources based on machine learning according to claim 7, characterized in that, The convergence conditions for the first gradient booster specifically include: Calculate the change in predicted value of training samples that are not yet terminated in the current round, and calculate the average change in predicted value of all training samples that are not yet terminated. When the average change is lower than the preset change convergence threshold in several consecutive iterations, the first gradient booster is determined to have reached the convergence condition. When the number of training samples in the unterminated state remains constant for several consecutive iterations, the first gradient booster is considered to have reached the convergence condition. When the first gradient booster reaches any convergence condition, the training of subsequent weak learners is stopped, and the training result of the first gradient booster is output.
9. The personalized recommendation method for educational resources based on machine learning according to claim 8, characterized in that, Specifically, S7 includes: After the first gradient booster reaches the convergence condition, the target resource cluster is selected based on the student-resource cluster matching score output by the first gradient booster. Training samples for the second gradient booster are constructed based on the student feature set and the resource feature set within the target resource cluster. The number of weak learners, the structure of the weak learners, the type of loss function, and the initialization method of the sample-level termination decision variable of the second gradient booster are set to the same model structure as the first gradient booster. The training process for the second gradient booster is the same as that for the first gradient booster. The matching scores between students and resources within the target resource cluster are generated, and personalized recommendations for educational resources are output based on the matching scores.