An online course intelligent recommendation method and system based on big data analysis

Abstract

This invention discloses an intelligent online course recommendation method and system based on big data analysis, comprising the following steps: S1, collecting multi-source data and preprocessing it; S2, extracting learning state features, constructing course nodes and course connection edges, and building a course knowledge graph based on the association strength; S3, performing vector representation learning in the course knowledge graph, constructing a course transition probability matrix, and determining the initial recommendation set corresponding to the current learning state; S4, performing node propagation calculation in the course knowledge graph starting from course nodes, and screening the candidate course set; S5, performing candidate path search and sorting the candidate paths to form a recommendation priority sequence; S6, compiling the course recommendation list and collecting feedback data to incrementally update the course knowledge graph. This invention enables collaborative analysis in the online course recommendation process, improving the matching accuracy of course recommendations.

Description

Technical Field

[0001] This invention relates to the field of course recommendation technology, and in particular to an intelligent online course recommendation method and system based on big data analysis. Background Technology

[0002] With the development of online education platforms, the number of course resources is constantly increasing, and learners generate a large amount of learning behavior data during the process of browsing, watching, evaluating, and providing feedback on courses. Existing online course recommendation methods usually rely on historical click records, course tags, popularity, or simple similarity to recommend courses. By statistically analyzing learners' existing behaviors, they recommend courses that are similar to their historical browsing content or have high access volumes.

[0003] Existing technologies also include recommendation methods that utilize big data analytics to model learning behavior. For example, they determine learning preferences based on data such as learning duration, course completion rate, and assessment results, and then combine this with collaborative filtering, content matching, or ranking models to generate recommendation results. While these methods can achieve a certain degree of personalized recommendation, they often rely on single behavioral features or static course tags, making it difficult to express the semantic relationships between courses and the matching relationship between learning status and course content.

[0004] Meanwhile, existing methods for handling course resource relationships mostly rely on course classification, keyword tags, or manually configured relationships, lacking a unified model for course nodes, course connection edge weights, and the strength of associations with learning states. Because courses do not form a propagable and computable graph structure, the recommendation process easily remains at the level of similar course selection, making it difficult to propagate nodes and search paths based on the current learning state. This results in insufficient continuity of recommendation results and unclear association paths.

[0005] Furthermore, learners' feedback after accepting recommendations is typically only used for simple statistics or re-ranking, lacking an incremental update mechanism for course connection edge weights and learning state matching parameters. When learning states change, the recommendation model struggles to promptly correct course relationships and recommendation paths, easily leading to problems such as fixed recommendation results, duplicated and concentrated candidate courses, and insufficient course suitability.

[0006] Therefore, how to provide an intelligent recommendation method and system for online courses based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent recommendation method and system for online courses based on big data analysis. This invention fully utilizes course knowledge graphs, graph embedding algorithms, random walk algorithms with restarts, K-shortest path search algorithms, and recursive least squares update methods. It describes in detail the intelligent recommendation process of constructing a course knowledge graph based on a learning behavior sample set, learning course node vector representations, performing node propagation calculations, searching for candidate recommendation paths, and incrementally updating based on learning feedback. It has the advantages of strong recommendation relevance, high course matching degree, clear recommendation paths, and strong dynamic update capabilities.

[0008] An intelligent online course recommendation method based on big data analysis according to an embodiment of the present invention includes the following steps: S1. Collect multi-source data from online course platforms and preprocess it to generate a learning behavior sample set; S2. Extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. S3. In the course knowledge graph, a graph embedding algorithm is used to learn the vector representation of course nodes and learning state features, and a course transition probability matrix is ​​constructed based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. S4. Based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, the node propagation calculation is performed in the course knowledge graph using the random walk algorithm with restart, and the candidate course set is screened according to the access probability of the course nodes. S5. Based on the candidate course set, the K-shortest path search algorithm is used to search for candidate paths in the course transition probability matrix, and the candidate paths are sorted according to the path entropy value and the course node access probability to form a recommendation priority sequence. S6. Arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and use the least squares method to incrementally update the course knowledge graph.

[0009] Optionally, the preprocessing specifically includes: associating and merging multi-source data in the online course platform according to the same learning object and the same course object; using a time window to segment the merged data into behavioral segments; filling in missing items in the behavioral segments with neighboring segments; using the interquartile range method to remove abnormal behavioral segments that deviate from the normal distribution range; and rearranging the retained behavioral segments according to a uniform time granularity to form a learning behavior sample set.

[0010] Optionally, S2 specifically includes: S21. The learning behavior sample set is continuously divided according to time windows. The learning duration, completion change and assessment change within each time window are statistically analyzed. The statistical results of adjacent time windows are decayed and fused using the exponential weighted moving average method to obtain the learning status characteristics. The learning duration is obtained by weighting the effective learning time within the time window with the behavior interval. The completion change is obtained by window merging of the course progress increment. The assessment change is obtained by moving difference of the difference between adjacent assessment results. S22. The course content text is segmented into terms, weighted and encoded. The word vectors are weighted and fused according to the term weights. The contribution ratio of the word vectors is adjusted according to the semantic importance of the terms to obtain the course enhanced semantic vector. S23. Based on the semantic association value, the connection relationship between courses is filtered by threshold, and the local sensitive hashing method is used to remove low similarity connection relationships to obtain the course connection edge and the corresponding course connection edge weight. S24. The learning state features are transformed into the course semantic vector space using a linear projection method. The matching value between the projected learning state features and the course semantic vector is calculated. Based on the matching value, the learning state features, course nodes, course connection edges, and course connection edge weights are combined into a course knowledge graph. The matching value is jointly determined by the cosine similarity between the projected learning state features and the course semantic vector, the behavioral intensity corresponding to the learning duration, and the state offset corresponding to the evaluation change.

[0011] Optionally, S22 specifically includes: S221. Segment the course content text into terms, count the frequency of terms in the course content text, calculate the term weights using the TF-IDF algorithm, and select terms whose weights reach the preset weight threshold to form a course keyword sequence. S222. Encode the terms in the course keyword sequence with word vectors, and aggregate the word vectors according to the term weights to form the initial semantic vector of the course. S223. Using the TextRank algorithm, the semantic importance of the terms in the course keyword sequence is iterated, and the contribution ratio of the word vectors in the initial semantic vector of the course is corrected according to the semantic importance to obtain the course enhanced semantic vector. S224. Calculate the cosine similarity between the enhanced semantic vectors of different courses, and use the cosine similarity as the semantic association value between courses to determine the semantic association relationship between courses.

[0012] Optionally, S3 specifically includes: S31. In the course knowledge graph, construct a weighted adjacency matrix based on the correlation strength between learning state features and course nodes, and project the learning state features into state guidance vectors according to the correlation strength. S32. A graph convolutional network is used to perform restricted message passing on the weighted adjacency matrix, and based on the correlation strength between the learning state features and the course nodes, the state guidance vector is introduced into the node update process to obtain the course node vector and the learning state vector. S33. Calculate the node vector similarity between course node vectors, and weight and fuse the node vector similarity with the course connection edge weight to obtain the transition strength between course nodes. Then, use the Softmax function to probabilize the transition strength corresponding to the same course node and construct the course transition probability matrix. S34. Calculate the matching value between the learning state vector and the course node vector, perform threshold screening on the course nodes according to the matching value, and encode the screened course nodes into the initial recommendation set corresponding to the current learning state.

[0013] Optionally, S4 specifically includes: S41. Construct an initial probability distribution based on the matching values ​​of course nodes in the initial recommendation set, and load the initial probability distribution into the row and column positions of the course nodes corresponding to the course transition probability matrix. S42. The initial probability distribution is iteratively propagated using a random walk algorithm with restart. In each round of propagation, node probability diffusion is performed based on the course transition probability matrix, and a portion of the propagated probability is reverted to the initial probability distribution through the restart coefficient. S43. Calculate the difference between the probability vectors of two adjacent rounds of propagation. Stop iterating when the difference is lower than the convergence threshold to obtain the stable propagation probability of the course node. S44. Based on the stable propagation probability, the course nodes are screened, and the nearest duplicate nodes are removed by combining the shortest association distance between the course node and the initial recommendation set to obtain the candidate course set.

[0014] Optionally, S42 specifically includes: S421. Determine the restart baseline vector based on the initial probability distribution, and assign initial probability values ​​to the course nodes corresponding to the initial recommendation set in the restart baseline vector, and assign zero probability values ​​to the remaining course nodes. S422. In each round of propagation, the propagation probability of the course nodes in the previous round and the course transition probability matrix are iterated to obtain the current round of diffusion probability according to the transition relationship between course nodes. S423. Based on the restart coefficient, the current round diffusion probability is controlled by a retention ratio, and the remaining probability ratio is reset to the restart reference vector to obtain the current round course node propagation probability; S424. Adjust the total probability of the current round of course node propagation probability and use the adjusted course node propagation probability as the input for the next round of propagation.

[0015] Optionally, S5 specifically includes: S51. Set the course nodes in the initial recommendation set as the starting point of the path search, set the course nodes in the candidate course set as the ending point of the path search, and determine the reachable connection relationship between the starting point and the ending point based on the course transition probability matrix. S52. Based on reachable connectivity, the K-shortest path search algorithm is used to search for candidate paths between the initial recommendation set and the candidate course set. The path cost is determined based on transition values ​​and node vector similarity, resulting in a candidate path set, specifically including: Based on reachable connectivity, the path search range between course nodes in the initial recommendation set and course nodes in the candidate course set is determined, and the connection relationships of course nodes with transition values ​​in the course transition probability matrix are included in the path search range. Cost calibration is performed on the connection relationships of course nodes within the path search range. The transition values ​​in the course transition probability matrix are converted into basic path costs. The basic path costs are then corrected by combining the node vector similarity between corresponding course nodes to obtain the one-sided path cost of the connection relationship between course nodes. Using the course nodes in the initial recommendation set as the starting point of the path and the course nodes in the candidate course set as the ending point of the path, the K-shortest path search algorithm is used to expand the path layer by layer according to the one-sided path cost. During the path expansion process, loop paths that repeatedly pass through the same course node are eliminated. The search paths leading to the same candidate course node are accumulated in terms of cost. The top-ranked search paths are retained according to the accumulated path cost to obtain the candidate path set corresponding to each candidate course node. S53. Calculate the entropy value of the node transfer distribution of each candidate path in the candidate path set to obtain the path entropy value, and calibrate the propagation dispersion of the candidate path based on the path entropy value. S54. The candidate paths are sorted by combining path cost, path entropy value and learning state matching value, and a recommendation priority sequence is formed according to the course nodes corresponding to the sorted candidate paths.

[0016] Optionally, S6 specifically includes: S61. Extract the top-ranked course nodes according to the recommendation priority sequence, and arrange the course recommendation list according to the course content corresponding to the course nodes; S62. Collect learning feedback data corresponding to the course recommendation list, and associate the learning feedback data with the course nodes in the course recommendation list to form a feedback association sample. The learning feedback data is obtained by merging the learning behavior records generated by the recommended courses after recommendation with the assessment change records according to the feedback occurrence time. S63. Determine the feedback deviation amount corresponding to the course node based on the feedback correlation sample, and use the recursive least squares method to recursively correct the course connection edge weight and learning state matching parameter. S64. Based on the corrected course connection edge weights and learning state matching parameters, update the connection relationships between course nodes in the course knowledge graph and the correlation strength between learning state features and course nodes.

[0017] According to an embodiment of the present invention, an intelligent online course recommendation system based on big data analysis includes: The sample generation module is used to collect multi-source data from the online course platform and preprocess it to generate a sample set of learning behaviors. The graph construction module is used to extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. The vector embedding module is used to learn vector representations of course nodes and learning state features in the course knowledge graph using graph embedding algorithms, and to construct a course transition probability matrix based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. The node propagation module is used to perform node propagation calculations in the course knowledge graph based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, and using a random walk algorithm with restart. It also selects the candidate course set based on the access probability of the course nodes. The path sorting module is used to search for candidate paths in the course transition probability matrix based on the candidate course set using the K-shortest path search algorithm, and sort the candidate paths according to the path entropy value and the access probability of course nodes to form a recommendation priority sequence. The feedback update module is used to arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and to incrementally update the course knowledge graph using the least squares method.

[0018] The beneficial effects of this invention are: First, this invention extracts learning state features from a learning behavior sample set and constructs a course knowledge graph by combining course semantic relationships. This unifies learning behaviors with course nodes, course connection edges, and association strength into the same graph structure, avoiding reliance solely on historical clicks, popular courses, or static tags for recommendations.

[0019] Secondly, this invention uses a graph embedding algorithm to learn vector representations of course nodes and learning states, and performs random walk computation with restart based on the course transition probability matrix. This allows for the screening of a set of candidate courses closely related to the current learning state in the course knowledge graph. Then, a recommendation priority sequence is formed by K-shortest path search and path entropy sorting, resulting in better course coherence in the recommendation results.

[0020] Finally, this invention collects learning feedback data corresponding to the course recommendation list and uses recursive least squares method to incrementally update the course connection edge weights and learning state matching parameters, so that the course knowledge graph can be continuously corrected with learning feedback, reducing the problems of fixed recommendation results and duplicate concentration of candidate courses. Attached Figure Description

[0021] 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 an overall flowchart of an intelligent online course recommendation method based on big data analysis proposed in this invention; Figure 2 This is a flowchart of the random walk node propagation process with restart for an online course intelligent recommendation method based on big data analysis proposed in this invention. Figure 3 This is a module structure diagram of an intelligent online course recommendation system based on big data analysis proposed in this invention. Detailed Implementation

[0022] 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.

[0023] refer to Figures 1-2 A method for intelligent recommendation of online courses based on big data analysis includes the following steps: S1. Collect multi-source data from online course platforms and preprocess it to generate a learning behavior sample set; S2. Extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. S3. In the course knowledge graph, a graph embedding algorithm is used to learn the vector representation of course nodes and learning state features, and a course transition probability matrix is ​​constructed based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. S4. Based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, the node propagation calculation is performed in the course knowledge graph using the random walk algorithm with restart, and the candidate course set is screened according to the access probability of the course nodes. S5. Based on the candidate course set, the K-shortest path search algorithm is used to search for candidate paths in the course transition probability matrix, and the candidate paths are sorted according to the path entropy value and the course node access probability to form a recommendation priority sequence. S6. Arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and use the least squares method to incrementally update the course knowledge graph.

[0024] In this embodiment, the preprocessing specifically includes: associating and merging multi-source data in the online course platform according to the same learning object and the same course object; using a time window to segment the merged data into behavioral segments; filling in missing items in the behavioral segments with neighboring segments; using the interquartile range method to remove abnormal behavioral segments that deviate from the normal distribution range; and rearranging the retained behavioral segments according to a uniform time granularity to form a learning behavior sample set.

[0025] In this embodiment, S2 specifically includes: S21. The learning behavior sample set is continuously divided according to time windows. The learning duration, completion change and assessment change within each time window are statistically analyzed. The statistical results of adjacent time windows are decayed and fused using the exponential weighted moving average method to obtain the learning status characteristics. The learning duration is obtained by weighting the effective learning time within the time window with the behavior interval. The completion change is obtained by merging the course progress increment through the window. The assessment change is obtained by the difference between adjacent assessment results through the moving difference. S22. The course content text is segmented into terms, weighted and encoded. The word vectors are weighted and fused according to the term weights. The contribution ratio of the word vectors is adjusted according to the semantic importance of the terms to obtain the course enhanced semantic vector. S23. Based on semantic association values, threshold filtering is applied to the connection relationships between courses, and the locality-sensitive hashing method is used to remove low-similarity connections, obtaining course connection edges and their corresponding weights, specifically including: Hash signatures are generated based on the course-enhanced semantic vectors corresponding to each course, and courses with similar hash signatures are grouped into the same candidate bucket using the locality-sensitive hashing method. Calculate the semantic association value between course augmentation semantic vectors within the same candidate bucket, and remove course relationships with semantic association values ​​lower than the connection threshold from the candidate bucket; The preserved course relationships are confirmed, course connection edges are established between corresponding courses, and semantic association values ​​are assigned to the course connection edges to form course connection edge weights; Repeated course connection edges are merged, and course connection edges are screened according to their weights to obtain a set of course connection edges used to construct the course knowledge graph. S24. The learning state features are transformed into the course semantic vector space using a linear projection method. The matching value between the projected learning state features and the course semantic vector is calculated. Based on the matching value, the learning state features, course nodes, course connection edges and course connection edge weights are combined into a course knowledge graph. The matching value is determined by the cosine similarity between the projected learning state features and the course semantic vector, the behavioral intensity corresponding to the learning duration, and the state offset corresponding to the assessment change.

[0026] In this embodiment, S22 specifically includes: S221. Segment the course content text into terms, count the frequency of terms in the course content text, calculate the term weights using the TF-IDF algorithm, and select terms whose weights reach the preset weight threshold to form a course keyword sequence. S222. Encode the terms in the course keyword sequence with word vectors, and aggregate the word vectors according to the term weights to form the initial semantic vector of the course. S223. Using the TextRank algorithm, the semantic importance of the terms in the course keyword sequence is iterated, and the contribution ratio of the word vectors in the initial semantic vector of the course is corrected according to the semantic importance to obtain the course enhanced semantic vector. S224. Calculate the cosine similarity between the enhanced semantic vectors of different courses, and use the cosine similarity as the semantic association value between courses to determine the semantic association relationship between courses.

[0027] In this embodiment, S3 specifically includes: S31. In the course knowledge graph, construct a weighted adjacency matrix based on the correlation strength between learning state features and course nodes, and project the learning state features into state guidance vectors according to the correlation strength. S32. A graph convolutional network is used to perform restricted message passing on the weighted adjacency matrix. Based on the correlation strength between the learning state features and the course nodes, a state guidance vector is introduced into the node update process to obtain the course node vector and the learning state vector, specifically including: The adjacency propagation range between course nodes is determined based on the weighted adjacency matrix, and a message passing mask is formed within the adjacency propagation range according to the weight of the course connection edges. A graph convolutional network is used to perform neighborhood convolution operations on the semantic representation of course nodes under message passing mask constraints. The convolution result is then nonlinearly transformed by an activation function to obtain the neighborhood response representation of the course nodes. Based on the correlation strength between learning state features and course nodes, state-associated course nodes are selected, the state guidance vector is embedded into the neighborhood response representation of the state-associated course node, and a gating update mechanism is used to adjust the participation ratio of the state guidance vector in node update. Perform residual connections on the gated updated node representations to obtain course node vectors, and generate learning state vectors based on the participation of the state guidance vector in the gated update. S33. Calculate the node vector similarity between course node vectors, and weight and fuse the node vector similarity with the course connection edge weight to obtain the transition strength between course nodes. Then, use the Softmax function to probabilize the transition strength corresponding to the same course node and construct the course transition probability matrix. S34. Calculate the matching value between the learning state vector and the course node vector, perform threshold screening on the course nodes according to the matching value, and encode the screened course nodes into the initial recommendation set corresponding to the current learning state.

[0028] In this embodiment, S4 specifically includes: S41. Construct an initial probability distribution based on the matching values ​​of course nodes in the initial recommendation set, and load the initial probability distribution into the row and column positions of the course nodes corresponding to the course transition probability matrix. S42. The initial probability distribution is iteratively propagated using a random walk algorithm with restart. In each round of propagation, node probability diffusion is performed based on the course transition probability matrix, and a portion of the propagated probability is reverted to the initial probability distribution through the restart coefficient. S43. Calculate the difference between the probability vectors of two adjacent rounds of propagation. Stop iterating when the difference is lower than the convergence threshold to obtain the stable propagation probability of the course node. S44. Based on the stable propagation probability, the course nodes are screened, and the nearest duplicate nodes are removed by combining the shortest association distance between the course node and the initial recommendation set to obtain the candidate course set.

[0029] In this embodiment, S42 specifically includes: S421. Determine the restart baseline vector based on the initial probability distribution, and assign initial probability values ​​to the course nodes corresponding to the initial recommendation set in the restart baseline vector, and assign zero probability values ​​to the remaining course nodes. S422. In each round of propagation, the propagation probability of the course nodes in the previous round and the course transition probability matrix are iterated to obtain the current round of diffusion probability according to the transition relationship between course nodes. S423. Based on the restart coefficient, the current round diffusion probability is controlled by a retention ratio, and the remaining ratio of probability is reset to the restart reference vector to obtain the current round course node propagation probability. The restart coefficient is set according to the dispersion of the matching values ​​of course nodes in the initial recommendation set. When the matching values ​​are concentrated, the restart coefficient is reduced, and when the matching values ​​are dispersed, the restart coefficient is increased. S424. Adjust the total probability of the current round of course node propagation probability and use the adjusted course node propagation probability as the input for the next round of propagation.

[0030] In this embodiment, S5 specifically includes: S51. Set the course nodes in the initial recommendation set as the starting point of the path search, set the course nodes in the candidate course set as the ending point of the path search, and determine the reachable connection relationship between the starting point and the ending point based on the course transition probability matrix. S52. Based on reachable connectivity, the K-shortest path search algorithm is used to search for candidate paths between the initial recommendation set and the candidate course set. The path cost is determined based on transition values ​​and node vector similarity, resulting in a candidate path set, specifically including: Based on reachable connectivity, the path search range between course nodes in the initial recommendation set and course nodes in the candidate course set is determined, and the connection relationships of course nodes with transition values ​​in the course transition probability matrix are included in the path search range. Cost calibration is performed on the connection relationships of course nodes within the path search range. The transition values ​​in the course transition probability matrix are converted into basic path costs. The basic path costs are then corrected by combining the node vector similarity between corresponding course nodes to obtain the one-sided path cost of the connection relationship between course nodes. Using the course nodes in the initial recommendation set as the starting point of the path and the course nodes in the candidate course set as the ending point of the path, the K-shortest path search algorithm is used to expand the path layer by layer according to the one-sided path cost. During the path expansion process, loop paths that repeatedly pass through the same course node are eliminated. The search paths leading to the same candidate course node are accumulated in terms of cost. The top-ranked search paths are retained according to the accumulated path cost to obtain the candidate path set corresponding to each candidate course node. S53. Calculate the entropy value of the node transfer distribution of each candidate path in the candidate path set to obtain the path entropy value, and calibrate the propagation dispersion of the candidate path based on the path entropy value. S54. The candidate paths are sorted by combining path cost, path entropy value and learning state matching value, and a recommendation priority sequence is formed according to the course nodes corresponding to the sorted candidate paths.

[0031] In this embodiment, S6 specifically includes: S61. Extract the top-ranked course nodes according to the recommendation priority sequence, and arrange the course recommendation list according to the course content corresponding to the course nodes. The course recommendation list consists of the top-ranked course nodes, the course content corresponding to the course nodes, and the course associations in the candidate paths, and is arranged according to the recommendation priority sequence. S62. Collect learning feedback data corresponding to the course recommendation list, and associate the learning feedback data with the course nodes in the course recommendation list to form a feedback association sample. The learning feedback data is obtained by merging the learning behavior records generated by the recommended courses after recommendation with the assessment change records according to the feedback occurrence time. S63. Determine the feedback deviation amount corresponding to the course node based on the feedback correlation sample, and use the recursive least squares method to recursively correct the course connection edge weight and learning state matching parameter. S64. Based on the corrected course connection edge weights and learning state matching parameters, update the connection relationships between course nodes in the course knowledge graph and the correlation strength between learning state features and course nodes.

[0032] refer to Figure 3 An intelligent online course recommendation system based on big data analysis includes: The sample generation module is used to collect multi-source data from the online course platform and preprocess it to generate a sample set of learning behaviors. The graph construction module is used to extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. The vector embedding module is used to learn vector representations of course nodes and learning state features in the course knowledge graph using graph embedding algorithms, and to construct a course transition probability matrix based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. The node propagation module is used to perform node propagation calculations in the course knowledge graph based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, and using a random walk algorithm with restart. It also selects the candidate course set based on the access probability of the course nodes. The path sorting module is used to search for candidate paths in the course transition probability matrix based on the candidate course set using the K-shortest path search algorithm, and sort the candidate paths according to the path entropy value and the access probability of course nodes to form a recommendation priority sequence. The feedback update module is used to arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and to incrementally update the course knowledge graph using the least squares method.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a course recommendation scenario on an online learning platform. This platform has a large number of courses, and learners generate a significant amount of data during course browsing, viewing, assessment, and feedback. Traditional recommendation methods primarily rely on historical clicks and popular courses, which can easily lead to problems such as duplicate recommendations of similar courses, mismatches between recommended courses and current learning status, and low learner completion rates.

[0034] In this scenario, the platform collects and preprocesses multi-source learning data from learners to form a learning behavior sample set. Then, it statistically analyzes learning duration, completion variability, and assessment variability according to time windows, and uses an exponentially weighted moving average method to fuse these data to obtain learning status features. For course-side data, the system performs term segmentation, TF-IDF weight calculation, word vector encoding, and TextRank semantic importance correction on the course content text to obtain enhanced semantic vectors for the courses. These vectors are then combined with semantic association values ​​and locality-sensitive hashing to identify course connection edges and construct a course knowledge graph.

[0035] During the recommendation process, the system employs a graph convolutional network to learn vector representations of the course knowledge graph and constructs a course transition probability matrix based on the weights of course connection edges and the similarity of node vectors. For the current learner, the system determines an initial recommendation set based on their learning state, and then uses a random walk algorithm with restart to propagate nodes and screen the candidate course set. Subsequently, the system uses a K-shortest path search algorithm to obtain candidate paths in the course transition probability matrix, and sorts them according to path cost, path entropy, and learning state matching value to form a recommendation priority sequence. After the course recommendation list is arranged, the system collects learning feedback data and uses recursive least squares method to incrementally update the course connection edge weights and learning state matching parameters.

[0036] During implementation, a sample of students who continuously generated valid learning records within the platform was selected for statistical analysis. Using traditional recommendation methods, the click-through rate for recommended courses was 31.8%, the average completion rate was 54.3%, and the proportion of courses with repetitive semantics reached 22.4%. Some learners skipped or briefly exited the platform after receiving similar courses consecutively.

[0037] After adopting this invention, the click-through rate of recommended courses increased to 42.6%, the average completion rate increased to 68.7%, the average learning time of recommended courses increased from 18.6 minutes to 26.4 minutes, the average improvement in evaluation scores after recommendation increased from 5.2 points to 8.9 points, and the proportion of courses with repetitive semantics decreased to 9.6%. Simultaneously, the average entropy of candidate course paths increased from 0.41 to 0.63, indicating that recommended paths are no longer concentrated on a few similar courses; the course node matching stability rate increased from 76.2% to 89.4%, and the parameter update success rate after feedback increased from 73.5% to 91.2%, demonstrating that the course knowledge graph can be continuously corrected based on subsequent learning feedback.

[0038] Table 1. Statistical Table of the Implementation Effect of Intelligent Course Recommendation

[0039] As can be seen from Table 1, this invention, through learning state feature extraction, course knowledge graph construction, node propagation calculation, candidate path search, and feedback incremental update, makes the matching between recommended courses and the current learning state more stable, significantly reduces repeated recommendations and short-term exits, and improves course completion rate, assessment improvement, and feedback update effect.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation of online courses based on big data analysis, characterized in that, Includes the following steps: S1. Collect multi-source data from online course platforms and preprocess it to generate a learning behavior sample set; S2. Extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. S3. In the course knowledge graph, a graph embedding algorithm is used to learn the vector representation of course nodes and learning state features, and a course transition probability matrix is ​​constructed based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. S4. Based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, the node propagation calculation is performed in the course knowledge graph using the random walk algorithm with restart, and the candidate course set is screened according to the access probability of the course nodes. S5. Based on the candidate course set, the K-shortest path search algorithm is used to search for candidate paths in the course transition probability matrix, and the candidate paths are sorted according to the path entropy value and the course node access probability to form a recommendation priority sequence. S6. Arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and use the least squares method to incrementally update the course knowledge graph.

2. The online course intelligent recommendation method based on big data analysis according to claim 1, characterized in that, The preprocessing specifically includes: associating and merging multi-source data in the online course platform according to the same learning object and the same course object; using time windows to segment the merged data into behavioral segments; filling in missing items in the behavioral segments with neighboring segments; using the interquartile range method to remove abnormal behavioral segments that deviate from the normal distribution range; and rearranging the retained behavioral segments according to a uniform time granularity to form a learning behavior sample set.

3. The intelligent online course recommendation method based on big data analysis according to claim 1, characterized in that, S2 specifically includes: S21. The learning behavior sample set is continuously divided according to time windows. The learning duration, completion change and assessment change within each time window are statistically analyzed. The statistical results of adjacent time windows are decayed and fused using the exponential weighted moving average method to obtain the learning status characteristics. The learning duration is obtained by weighting the effective learning time within the time window with the behavior interval. The completion change is obtained by window merging of the course progress increment. The assessment change is obtained by moving difference of the difference between adjacent assessment results. S22. The course content text is segmented into terms, weighted and encoded. The word vectors are weighted and fused according to the term weights. The contribution ratio of the word vectors is adjusted according to the semantic importance of the terms to obtain the course enhanced semantic vector. S23. Based on the semantic association value, the connection relationship between courses is filtered by threshold, and the local sensitive hashing method is used to remove low similarity connection relationships to obtain the course connection edge and the corresponding course connection edge weight. S24. The learning state features are transformed into the course semantic vector space using a linear projection method. The matching value between the projected learning state features and the course semantic vector is calculated. Based on the matching value, the learning state features, course nodes, course connection edges, and course connection edge weights are combined into a course knowledge graph. The matching value is jointly determined by the cosine similarity between the projected learning state features and the course semantic vector, the behavioral intensity corresponding to the learning duration, and the state offset corresponding to the evaluation change.

4. The intelligent online course recommendation method based on big data analysis according to claim 3, characterized in that, S22 specifically includes: S221. Segment the course content text into terms, count the frequency of terms in the course content text, calculate the term weights using the TF-IDF algorithm, and select terms whose weights reach the preset weight threshold to form a course keyword sequence. S222. Encode the terms in the course keyword sequence with word vectors, and aggregate the word vectors according to the term weights to form the initial semantic vector of the course. S223. Using the TextRank algorithm, the semantic importance of the terms in the course keyword sequence is iterated, and the contribution ratio of the word vectors in the initial semantic vector of the course is corrected according to the semantic importance to obtain the course enhanced semantic vector. S224. Calculate the cosine similarity between the enhanced semantic vectors of different courses, and use the cosine similarity as the semantic association value between courses to determine the semantic association relationship between courses.

5. The online course intelligent recommendation method based on big data analysis according to claim 1, characterized in that, S3 specifically includes: S31. In the course knowledge graph, construct a weighted adjacency matrix based on the correlation strength between learning state features and course nodes, and project the learning state features into state guidance vectors according to the correlation strength. S32. A graph convolutional network is used to perform restricted message passing on the weighted adjacency matrix, and based on the correlation strength between the learning state features and the course nodes, the state guidance vector is introduced into the node update process to obtain the course node vector and the learning state vector. S33. Calculate the node vector similarity between course node vectors, and weight and fuse the node vector similarity with the course connection edge weight to obtain the transition strength between course nodes. Then, use the Softmax function to probabilize the transition strength corresponding to the same course node and construct the course transition probability matrix. S34. Calculate the matching value between the learning state vector and the course node vector, perform threshold screening on the course nodes according to the matching value, and encode the screened course nodes into the initial recommendation set corresponding to the current learning state.

6. The intelligent online course recommendation method based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Construct an initial probability distribution based on the matching values ​​of course nodes in the initial recommendation set, and load the initial probability distribution into the row and column positions of the course nodes corresponding to the course transition probability matrix. S42. The initial probability distribution is iteratively propagated using a random walk algorithm with restart. In each round of propagation, node probability diffusion is performed based on the course transition probability matrix, and a portion of the propagated probability is reverted to the initial probability distribution through the restart coefficient. S43. Calculate the difference between the probability vectors of two adjacent rounds of propagation. Stop iterating when the difference is lower than the convergence threshold to obtain the stable propagation probability of the course node. S44. Based on the stable propagation probability, the course nodes are screened, and the nearest duplicate nodes are removed by combining the shortest association distance between the course node and the initial recommendation set to obtain the candidate course set.

7. The online course intelligent recommendation method based on big data analysis according to claim 6, characterized in that, S42 specifically includes: S421. Determine the restart baseline vector based on the initial probability distribution, and assign initial probability values ​​to the course nodes corresponding to the initial recommendation set in the restart baseline vector, and assign zero probability values ​​to the remaining course nodes. S422. In each round of propagation, the propagation probability of the course nodes in the previous round and the course transition probability matrix are iterated to obtain the current round of diffusion probability according to the transition relationship between course nodes. S423. Based on the restart coefficient, the current round diffusion probability is controlled by a retention ratio, and the remaining probability ratio is reset to the restart reference vector to obtain the current round course node propagation probability; S424. Adjust the total probability of the current round of course node propagation probability and use the adjusted course node propagation probability as the input for the next round of propagation.

8. The intelligent online course recommendation method based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Set the course nodes in the initial recommendation set as the starting point of the path search, set the course nodes in the candidate course set as the ending point of the path search, and determine the reachable connection relationship between the starting point and the ending point based on the course transition probability matrix. S52. Based on reachable connectivity, the K-shortest path search algorithm is used to search for candidate paths between the initial recommendation set and the candidate course set. The path cost is determined based on transition values ​​and node vector similarity, resulting in a candidate path set, specifically including: Based on reachable connectivity, the path search range between course nodes in the initial recommendation set and course nodes in the candidate course set is determined, and the connection relationships of course nodes with transition values ​​in the course transition probability matrix are included in the path search range. Cost calibration is performed on the connection relationships of course nodes within the path search range. The transition values ​​in the course transition probability matrix are converted into basic path costs. The basic path costs are then corrected by combining the node vector similarity between corresponding course nodes to obtain the one-sided path cost of the connection relationship between course nodes. Using the course nodes in the initial recommendation set as the starting point of the path and the course nodes in the candidate course set as the ending point of the path, the K-shortest path search algorithm is used to expand the path layer by layer according to the one-sided path cost. During the path expansion process, loop paths that repeatedly pass through the same course node are eliminated. The search paths leading to the same candidate course node are accumulated in terms of cost. The top-ranked search paths are retained according to the accumulated path cost to obtain the candidate path set corresponding to each candidate course node. S53. Calculate the entropy value of the node transfer distribution of each candidate path in the candidate path set to obtain the path entropy value, and calibrate the propagation dispersion of the candidate path based on the path entropy value. S54. The candidate paths are sorted by combining path cost, path entropy value and learning state matching value, and a recommendation priority sequence is formed according to the course nodes corresponding to the sorted candidate paths.

9. The intelligent online course recommendation method based on big data analysis according to claim 1, characterized in that, S6 specifically includes: S61. Extract the top-ranked course nodes according to the recommendation priority sequence, and arrange the course recommendation list according to the course content corresponding to the course nodes; S62. Collect learning feedback data corresponding to the course recommendation list, and associate the learning feedback data with the course nodes in the course recommendation list to form a feedback association sample. The learning feedback data is obtained by merging the learning behavior records generated by the recommended courses after recommendation with the assessment change records according to the feedback occurrence time. S63. Determine the feedback deviation amount corresponding to the course node based on the feedback correlation sample, and use the recursive least squares method to recursively correct the course connection edge weight and learning state matching parameter. S64. Based on the corrected course connection edge weights and learning state matching parameters, update the connection relationships between course nodes in the course knowledge graph and the correlation strength between learning state features and course nodes.

10. An intelligent online course recommendation system based on big data analysis, comprising executing the intelligent online course recommendation method based on big data analysis as described in any one of claims 1 to 9, characterized in that, include: The sample generation module is used to collect multi-source data from the online course platform and preprocess it to generate a sample set of learning behaviors. The graph construction module is used to extract learning state features from the learning behavior sample set, construct course nodes and course connection edges by combining course semantic relationships, and construct a course knowledge graph based on the correlation strength between learning state features and course nodes. The vector embedding module is used to learn vector representations of course nodes and learning state features in the course knowledge graph using graph embedding algorithms, and to construct a course transition probability matrix based on the course connection edge weights and node vector similarity to determine the initial recommendation set corresponding to the current learning state. The node propagation module is used to perform node propagation calculations in the course knowledge graph based on the course transition probability matrix, starting from the course nodes in the initial recommendation set, and using a random walk algorithm with restart. It also selects the candidate course set based on the access probability of the course nodes. The path sorting module is used to search for candidate paths in the course transition probability matrix based on the candidate course set using the K-shortest path search algorithm, and sort the candidate paths according to the path entropy value and the access probability of course nodes to form a recommendation priority sequence. The feedback update module is used to arrange the course recommendation list based on the recommendation priority sequence and collect learning feedback data, and to incrementally update the course knowledge graph using the least squares method.

Images