Education course recommendation method and system based on data mining

Abstract

The invention discloses an education course recommendation method and system based on data mining, and the specific technical scheme is as follows: collecting popularity index data such as course page view and completion rate, and processing the data through a deep learning model to obtain a course popularity value; and when the popularity value exceeds a threshold value, increasing and optimizing a course dynamic weight coefficient. And fusing the optimized weight coefficient and the user characteristics, judging the matching degree by using an adaptive learning algorithm, and generating a preliminarily recommended course list. And performing vector conversion by using a word embedding technology, calculating a semantic matching score, and customizing course content if the semantic matching score exceeds a threshold value. And analyzing a knowledge dependency relationship and a difficulty progressive relationship of the customized curriculum, optimizing a learning sequence by adopting a path planning algorithm, extracting sequence effect data, and adjusting a curriculum position, and finally forming an accurate education curriculum recommendation scheme. According to the method, the accuracy of course recommendation and the scientificity of the learning path are remarkably improved, and the learning experience and effect of the user are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of educational course recommendation technology, and in particular relates to an educational course recommendation method and system based on data mining. Background Technology

[0002] In recent years, online education platforms have flourished and become a significant force in the education sector. Against this backdrop, personalized course recommendations have become crucial for improving learning outcomes. Users' course needs vary based on their interests, learning goals, and prior knowledge; therefore, recommendation systems must accurately match course content and optimize learning paths to meet these diverse needs.

[0003] However, existing recommender systems have significant shortcomings in dynamically integrating multidimensional data to achieve accurate recommendations and optimize learning paths. Specifically, existing recommendation systems struggle with data perception and integration. They fail to capture and analyze dynamic changes in course popularity in real time and lack the ability to integrate multi-source data such as user behavior logs and course metadata. This can lead to a disconnect between recommended courses and actual user needs, resulting in an inability to respond to real-time changes and inefficient resource utilization. In terms of semantic understanding and deep analysis, there is a general lack of deep semantic analysis of course content and user learning objectives. Traditional methods such as collaborative filtering and content recommendation often ignore the complex semantic relationships and knowledge structures between courses. This not only exacerbates cold start and data sparsity problems but also affects matching accuracy, making it difficult for the system to effectively handle recommendations for new users or courses. Regarding learning path optimization, many systems fail to fully consider the knowledge dependencies and difficulty progression between courses in their recommendation paths, resulting in a lack of systematicity and coherence in the generated learning paths, thereby reducing learning effectiveness and user experience. In terms of algorithm models and real-time performance, existing methods often rely on fixed rules or offline models. They have poor adaptability to sudden traffic spikes or peak periods such as the release of popular courses or the start of a new semester, and cannot dynamically adjust recommendation strategies based on real-time user behavior and feedback, potentially leading to increased congestion or even overflow problems. These issues collectively prevent recommendation systems from fully meeting users' personalized learning needs, thus impacting the overall learning experience and effectiveness of online education platforms.

[0004] To address the problems in existing technologies, there is an urgent need to propose a data mining-based method and system for recommending educational courses. This system should be able to dynamically perceive learning needs and deeply optimize learning paths to improve recommendation accuracy and learning outcomes, thereby promoting the intelligent development of online education platforms. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a data mining-based method and system for recommending educational courses, thereby resolving the issues present in the existing technologies.

[0006] To achieve the above objectives, this invention provides a data mining-based method for recommending educational courses, comprising the following steps:

[0007] Collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate;

[0008] If the course popularity value exceeds the preset threshold, the dynamic weight coefficient of the corresponding course will be increased, and the weight coefficient will be optimized based on the correlation between the course popularity value and the weight coefficient.

[0009] An adaptive learning algorithm is used to fuse optimized weight coefficients and user features to determine the matching degree between user features and course learning objectives, thereby obtaining a preliminary recommended course list.

[0010] The initial recommended course list is vectorized based on word embedding technology. The semantic matching score is obtained by calculating the vector similarity to match the learning target corresponding to the user features.

[0011] If the semantic matching score is higher than the preset threshold, the matching course is selected for content customization, and the customized course version is determined.

[0012] The knowledge dependencies and difficulty progression between customized course versions are analyzed, and a path planning algorithm is used to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence.

[0013] Extract sequence effect data from the optimized learning sequence, compare the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence, and obtain the final educational course recommendation scheme.

[0014] Optionally, the process of collecting popularity index data and processing the popularity index data using a deep learning model to obtain the course popularity value includes:

[0015] Process course visit and completion rate data to obtain cleaned popularity index data;

[0016] Based on the heat index data after cleaning, a convolutional neural network is used to extract the spatial correlation features between the indicators to obtain the correlation feature vector;

[0017] If the associated feature vectors meet the preset threshold, the associated feature vectors are processed by a long short-term memory network to obtain the time series prediction results.

[0018] Based on the time series prediction results, principal component analysis was used for dimensionality reduction to obtain the dimensionality-reduced trend vector.

[0019] The trend vectors after dimensionality reduction are grouped using a clustering algorithm to obtain popularity trend classification labels;

[0020] Calculate the course popularity score based on the popularity trend classification tags.

[0021] Optionally, if the course popularity value exceeds a preset threshold, the dynamic weight coefficient of the corresponding course is increased, and the process of optimizing the weight coefficient based on the correlation between the course popularity value and the weight coefficient includes:

[0022] If the course popularity value is greater than the preset threshold, the course identifiers that meet the conditions are obtained by judging the threshold, and a set of filtered courses is obtained.

[0023] Based on the selected course set, the product of the course popularity value and the dynamic weight is calculated to obtain the preliminary weight coefficient;

[0024] Based on the initial weighting coefficients, the coefficient distribution in the course set is processed by mean normalization to obtain standardized weighting coefficients;

[0025] If the standardized weight coefficients exceed the preset range, the coefficient range is adjusted through linear mapping to obtain the optimized weight coefficients.

[0026] Optionally, the process of using an adaptive learning algorithm to fuse optimized weight coefficients and user features, determining the matching degree between user features and course learning objectives, and obtaining a preliminary recommended course list includes:

[0027] An adaptive learning algorithm is used to fuse optimized weight coefficients with user feature vectors, calculate a weighted matching score, and obtain a sorted course list.

[0028] The sorted course list is deduplicated to obtain the processed recommended course list.

[0029] Based on user interests and preferences, the order of course recommendations is adjusted to generate an adjusted list of recommended courses;

[0030] If the number of courses in the adjusted recommended course list is less than the preset threshold, then courses with a relevance higher than the preset threshold are extracted from the original course content and added to the recommendation results to obtain a complete preliminary recommended course list.

[0031] Optionally, the process of vectorizing the initial recommended course list based on word embedding technology and obtaining a semantic matching score by calculating vector similarity to match the learning objective corresponding to the user features includes:

[0032] The course content descriptions are obtained from the initial recommended course list, and the text is segmented using a word segmentation tool to obtain a standardized word sequence.

[0033] A standardized word sequence is converted into a vector representation by a pre-trained word embedding model, generating a vector set of course content.

[0034] Obtain the user's learning objective description and convert it into a target vector representation using a pre-trained word embedding model;

[0035] The initial matching score for each course is obtained by calculating the cosine similarity between the course content vector set and the target vector representation.

[0036] If the initial matching score is greater than the preset threshold, the corresponding course vector is retained and a candidate course matching set is generated.

[0037] Based on the candidate course matching set, a weighted average method is used to merge the course data to obtain the final matching score.

[0038] Optionally, if the semantic matching score is higher than a preset threshold, then the matching course is selected for content customization. The process of determining the customized course version includes:

[0039] Retrieve courses whose matching scores are higher than a preset threshold to obtain a set of matching courses;

[0040] A word segmentation tool is used to segment the description of the matching course set, generating a sequence of segmented words;

[0041] Extract knowledge point identifiers from the segmented word sequence to obtain a knowledge point list;

[0042] Calculate the cosine similarity between the knowledge point list and the user feature vector to obtain a set of similarity values;

[0043] If the values ​​in the similarity value set exceed the preset similarity threshold, the corresponding knowledge points are retained, and a subset of retained knowledge points is generated.

[0044] Based on the retained subset of knowledge points, a hierarchical clustering algorithm is used to group the knowledge points to obtain the knowledge point grouping results;

[0045] From the knowledge point grouping results, evaluate the complexity index of each group and generate a complexity evaluation set;

[0046] By adjusting the presentation order of knowledge point groups using the complexity evaluation set, we obtain the adjusted order list.

[0047] For the adjusted order list, the preference elements in the user feature vector are merged to generate a merged preference list;

[0048] Based on the list of integration preferences, construct a sequence of course modules to obtain a set of module sequences;

[0049] By combining the module sequence set, a complete course framework is formed, and a customized course version is determined.

[0050] Optionally, the process of analyzing the knowledge dependencies and difficulty progression among customized course versions, and using path planning algorithms to optimize the learning sequence to cover the learning objectives, to obtain an optimized learning sequence includes:

[0051] Construct a knowledge dependency graph based on the knowledge point structure in the course version;

[0052] Calculate the progressive difficulty sequence using a knowledge dependency graph;

[0053] The A* algorithm is used to optimize the learning order. The input is a knowledge dependency graph and a difficulty progression sequence, and the output is a preliminary path.

[0054] Based on the coverage of the learning objectives by the initial path, adjust the path to generate an optimized path;

[0055] The overall effect of the optimization path is judged by simulating the learning process. If the simulated score is higher than the threshold, the optimization path is determined.

[0056] Based on the optimization path, generate an optimization learning sequence.

[0057] Optionally, the process of extracting sequence effect data from the optimized learning sequence, adjusting the position of courses in the sequence by comparing the sequence effect with a preset maximum effect standard, and obtaining the final educational course recommendation scheme includes:

[0058] The effectiveness data is obtained from the optimized learning sequence. By analyzing the completion rate and target matching rate of each course in the sequence, an effectiveness data set is obtained.

[0059] For the effect dataset, the cosine similarity algorithm is used to compare the effect data with the preset maximum effect standard, and it is determined whether the similarity reaches the preset threshold to obtain the comparison result.

[0060] If the comparison result is lower than the preset threshold, the course position adjustment range is calculated based on the deviation value in the effect data, and the adjusted course position set is generated.

[0061] By adjusting the set of course locations and combining the dependencies between knowledge points, the learning sequence is rearranged using a topological sorting algorithm to generate a preliminary recommended path.

[0062] The path coverage rate is extracted from the initial recommended path, and the path validation results are obtained by comparing the degree of matching between the path coverage rate and the learning objective.

[0063] If the path verification results meet the target requirements, then the final recommended learning path is generated based on the preliminary recommended path.

[0064] Based on the final recommended learning path, a path description including the course sequence and the association of knowledge points is generated, resulting in an educational course recommendation scheme.

[0065] This invention also proposes an educational course recommendation system based on data mining, for implementing the method, the system comprising:

[0066] The course popularity prediction module is used to collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate;

[0067] The weight adjustment module is used to increase the dynamic weight coefficient of the corresponding course when the course popularity value exceeds the preset threshold, and to optimize the weight coefficient based on the correlation between the course popularity value and the weight coefficient.

[0068] The preliminary recommendation module is used to fuse optimized weight coefficients and user features using an adaptive learning algorithm, determine the matching degree between user features and course learning objectives, and obtain a preliminary recommended course list.

[0069] The semantic matching module is used to perform vector transformation on the initial recommended course list based on word embedding technology. It calculates the vector similarity to match the learning target corresponding to the user features and obtains the semantic matching score.

[0070] The course customization module is used to select matching courses for content customization when the semantic matching score is higher than a preset threshold, and to determine the customized course version.

[0071] The learning sequence optimization module is used to analyze the knowledge dependencies and difficulty progression relationships between customized course versions, and uses a path planning algorithm to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence.

[0072] The final path generation module is used to extract sequence effect data from the optimized learning sequence, and by comparing the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence to obtain the final educational course recommendation scheme.

[0073] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0074] Compared with the prior art, the present invention has the following advantages and technical effects:

[0075] This invention addresses the challenges of course popularity prediction, personalized user matching, and learning path optimization in online education, proposing an integrated solution. First, it analyzes popularity metrics such as course views and completion rates using a deep learning model to predict course popularity and dynamically adjust weighting coefficients. Then, combining user interests and preferences, it employs an adaptive learning algorithm to assess the match between user characteristics and course goals, generating a preliminary recommended course list. Next, it uses word embedding technology to calculate the semantic matching score between course content and user learning objectives, further filtering courses. Finally, it utilizes a path planning model to analyze course knowledge dependencies and difficulty progression, optimizing the learning sequence and adjusting course positions to maximize learning effectiveness. Through multi-level data fusion and dynamic optimization, this invention significantly improves the accuracy of course recommendations and the scientific nature of learning paths, enhancing the user learning experience and outcomes. Attached Figure Description

[0076] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0077] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0078] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0079] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0080] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0081] Example 1

[0082] like Figure 1 As shown, this embodiment provides a data mining-based educational course recommendation method, including the following steps:

[0083] Collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate;

[0084] If the course popularity value exceeds the preset threshold, the dynamic weight coefficient of the corresponding course will be increased, and the weight coefficient will be optimized based on the correlation between the course popularity value and the weight coefficient.

[0085] An adaptive learning algorithm is used to fuse optimized weight coefficients and user features to determine the matching degree between user features and course learning objectives, thereby obtaining a preliminary recommended course list.

[0086] The initial recommended course list is vectorized based on word embedding technology. The semantic matching score is obtained by calculating the vector similarity to match the learning target corresponding to the user features.

[0087] If the semantic matching score is higher than the preset threshold, the matching course is selected for content customization, and the customized course version is determined.

[0088] The knowledge dependencies and difficulty progression between customized course versions are analyzed, and a path planning algorithm is used to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence.

[0089] Extract sequence effect data from the optimized learning sequence, compare the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence, and obtain the final educational course recommendation scheme.

[0090] As a specific implementation method, the following steps are included:

[0091] Step S101: By collecting popularity metrics such as course views and completion rates, a deep learning model is used to process these metrics to predict popularity trends and obtain the course popularity value.

[0092] The process involves processing course view counts and completion rates to obtain cleaned popularity metrics. Based on this cleaned data, a convolutional neural network is used to extract spatial correlation features between metrics, resulting in correlation feature vectors. If these feature vectors meet a preset threshold, a long short-term memory network is used to process them, yielding time series prediction results. Principal component analysis (PCA) is then used for dimensionality reduction, resulting in dimensionality-reduced trend vectors. These trend vectors are then grouped using a clustering algorithm to obtain popularity trend classification labels. Finally, based on these labels, the course popularity value is calculated, yielding the final popularity score.

[0093] For example, in online education platforms, processing course visit volume and completion rate data to assess course popularity is a key business scenario. Suppose a platform has 100 courses, and each course records its daily visit volume and completion rate. Visit volume reflects the number of times users click on the course, e.g., a course has 500 daily visits; completion rate reflects the percentage of users who complete the course, e.g., a course has a completion rate of 60%. When cleaning the data, outliers need to be removed; for example, a sudden increase in visit volume to 5000 might be a system error, and the data should be filled with the mean or median. The popularity metric can be defined as a weighted average of visit volume and completion rate, for example, popularity = 0.7 × visit volume + 0.3 × completion rate, thus obtaining the daily popularity value for each course.

[0094] One possible implementation involves using a convolutional neural network to extract spatial correlation features between popularity metrics. Course popularity may be influenced by factors such as course category and release time, exhibiting spatial correlations.

[0095] For example, the popularity of programming courses may be related to other technical courses. Convolutional neural networks scan the popularity data matrix through convolutional kernels to extract the correlation patterns between courses. For instance, if a programming course and a data science course show similar trends in popularity, a correlation feature vector can be generated. This vector captures the implicit connections between courses, helping to discover clusters of popular courses and improve the accuracy of recommendation systems.

[0096] Specifically, Long Short-Term Memory (LSTM) networks are used to process associated feature vectors to predict time series data. If the Euclidean distance between the associated feature vectors is greater than a preset threshold of 0.8, it indicates a strong correlation between the courses and is suitable for time series prediction.

[0097] For example, if the daily popularity value sequence of a course group is [50, 55, 60, ...], a Long Short-Term Memory (LSTM) network can capture long-term trends and short-term fluctuations through a memory gate mechanism to predict the popularity value for the next 7 days, such as [62, 65, ...]. This prediction helps the platform adjust resource allocation in advance, such as increasing server bandwidth to cope with high traffic.

[0098] Principal component analysis (PCA) is used to reduce the dimensionality of prediction results. Assuming the prediction results contain multidimensional features, such as popularity values ​​and growth rates, PCA compresses these features into a 2-3 dimensional trend vector, retaining more than 90% of the information variance.

[0099] For example, the trend vector of a course might reflect its continuously rising popularity or significant fluctuations. Dimensionality reduction makes the data easier to visualize and analyze, reducing subsequent computational complexity.

[0100] In one possible implementation, the K-means clustering algorithm is used to group the trend vectors to obtain popularity trend classification labels. Assume the clusters are divided into three categories: high popularity (stable), medium popularity (fluctuating), and low popularity (decreasing).

[0101] For example, a course's trend vector indicates that its popularity remains consistently high, classifying it as "high and stable popularity." This type of categorization helps the platform identify popular courses and optimize course ranking and promotion strategies.

[0102] Specifically, the final popularity score is calculated based on the category tags.

[0103] For example, courses with stable high popularity are assigned a value of 0.9, those with fluctuating medium popularity are assigned 0.6, and those with declining low popularity are assigned 0.3. A course in the stable high popularity category will have a final popularity value of 0.9, reflecting its market appeal. These popularity values ​​can be used to dynamically adjust course recommendation weights, increasing user engagement and platform revenue.

[0104] It should be noted that this multi-step technical solution forms a complete analysis chain from raw data to final popularity through data cleaning, feature extraction, time prediction, dimensionality reduction, and clustering. Each step supports the others, ensuring the accuracy and practicality of the results.

[0105] The features extracted by convolutional neural networks provide high-quality input for subsequent predictions, while dimensionality reduction and clustering simplify the interpretation of complex data. This approach not only improves the accuracy of course popularity assessment but also provides data-driven decision support for platform operations, enhancing user experience and resource utilization efficiency.

[0106] Step S102: Based on the obtained course popularity value, if the popularity value exceeds a preset threshold, the dynamic weight coefficient of the corresponding course is increased. The correlation between the popularity value and the weight coefficient is processed by calculating the weight adjustment formula to determine the optimized weight coefficient.

[0107] If the course popularity value exceeds a preset threshold, the threshold is used to identify eligible courses, resulting in a filtered course set. Based on this set, the product of the course popularity value and the dynamic weight is calculated using an adjustment formula to obtain preliminary weight coefficients. These preliminary weight coefficients are then normalized using mean normalization to obtain standardized weight coefficients. If the standardized weight coefficients exceed a preset range, the range is adjusted using linear mapping to obtain optimized weight coefficients.

[0108] For example, in online education platforms, selecting courses and optimizing weighting coefficients is a key business scenario for improving the accuracy of recommendation systems. Suppose a platform has 100 courses, and the popularity score of each course is calculated by weighting its views and completion rate. The platform needs to select popular courses based on their popularity scores and adjust their recommendation weights accordingly.

[0109] One possible implementation involves using a threshold to identify eligible courses. Assuming a popularity threshold of 0.8, a programming course with a popularity score of 0.85 exceeds the threshold and is included in the filtered course set; another language course with a popularity score of 0.75, below the threshold, is discarded. The filtered set might contain 30 courses, categorized as programming, data science, etc. The key to threshold determination is setting reasonable criteria to ensure that truly popular courses are selected, providing a high-quality data foundation for subsequent weight adjustments.

[0110] For example, based on the filtered course set, an adjustment formula is used to calculate the product of the popularity value and the dynamic weight, resulting in an initial weight coefficient. The dynamic weight can change dynamically according to the course category or release time. For instance, programming courses, due to high market demand, have a weight of 0.9; newly released courses, to attract users, have a weight of 0.95. Assuming a programming course has a popularity value of 0.85 and a dynamic weight of 0.9, the initial weight coefficient is 0.85 × 0.9 = 0.765. This method, by introducing dynamic weights, flexibly reflects the market appeal of courses, facilitating accurate recommendations.

[0111] In one possible implementation, the initial weight coefficients are normalized to obtain standardized weight coefficients. Assume the initial weight coefficients for 30 courses range from 0.5 to 0.9, with a mean of 0.7. The initial weight coefficient of a certain course, 0.765, after normalization, might be adjusted to (0.765-0.5) / (0.9-0.5) = 0.6625. Mean normalization makes the coefficient distribution more uniform, avoids extreme values ​​affecting the fairness of the recommendation, and improves the stability of the recommendation system.

[0112] For example, if the standardized weight coefficients exceed the preset range, the coefficient range needs to be adjusted through linear mapping to obtain optimized weight coefficients. Suppose the reasonable range for standardized weight coefficients is 0.2 to 0.8, and the coefficient for a certain course is 0.9, exceeding the range. Through linear mapping, it can be compressed to 0.8, as shown in the formula: New coefficient = 0.2 + (0.8 - 0.2) × (0.9 - minimum value) / (maximum value - minimum value). Linear mapping ensures that the coefficients are within a reasonable range, preventing the recommendation system from overly favoring certain courses and improving the diversity of recommendation results.

[0113] In one possible implementation, the optimized weighting coefficients can be directly used for course recommendation ranking.

[0114] For example, a programming course might have an optimized weighting coefficient of 0.78, higher than the data science course's 0.65, and would be prioritized for recommendation to users. This method employs a multi-step optimization process, from filtering to weight adjustment, forming a complete logical chain to ensure the recommendation system is both accurate and fair, effectively improving user experience and platform course click-through rates.

[0115] Step S103: Obtain the optimized weight coefficients. For user feature data such as interests and preferences, use an adaptive learning algorithm to fuse the weight coefficients and user features, determine the matching degree between user features and course learning objectives, and obtain a preliminary recommended course list.

[0116] User data is acquired, and user features and interests are extracted to generate user feature vectors. The cosine similarity algorithm is used to calculate the matching degree between user features and course learning objectives using these feature vectors and course content, resulting in a preliminary matching score. If the preliminary matching score is greater than a preset threshold, courses meeting the criteria are selected based on course identifiers, generating a candidate course set. An adaptive learning algorithm is used to fuse optimized weight coefficients with user feature vectors to calculate a weighted matching score, resulting in a sorted course list. The sorted course list is then deduplicated, removing duplicate course identifiers to obtain a processed recommended course list. Based on the processed recommended course list and user interests, the course recommendation order is adjusted, generating an adjusted recommended course list. If the number of courses in the adjusted recommended course list is less than a preset threshold, highly relevant supplementary courses are extracted from the course content and added to the recommendation results, resulting in a complete preliminary recommended course list.

[0117] For example, in online education platforms, acquiring user data and extracting user characteristics and interests are the foundation for building personalized recommendations.

[0118] In one possible implementation, user data includes browsing history, study time, and rating feedback.

[0119] Specifically, for a programming enthusiast user, the data shows frequent visits to programming course pages, over 10 hours of study time, and a preference for technical content. By analyzing this data, features such as the interest category being programming and the preference level being advanced are extracted to generate a user feature vector, such as a vector containing a programming weight of 0.8 and a data science weight of 0.3. This vectorization facilitates subsequent matching and improves the targeting of recommendations.

[0120] It should be noted that by using the cosine similarity algorithm to calculate the degree of matching between user feature vectors and course content, the fit between users and course objectives can be quantified.

[0121] For example, a programming course's learning objective vector emphasizes algorithm optimization. If a user's vector shows a similarity of 0.75, a preliminary matching score of 0.75 is obtained. This calculation considers vector direction rather than absolute magnitude, ensuring that the matching prioritizes shared interests.

[0122] In one embodiment, if the initial matching score is greater than a preset threshold of 0.6, then courses that meet the criteria are selected based on the course identifier to generate a candidate course set.

[0123] For example, if a programming user's score of 0.75 exceeds the threshold, the program will be filtered out as PRO-101 and the data science course as DS-202, forming a set of 5 courses. This filtering ensures that only highly relevant courses are retained, reducing invalid recommendations.

[0124] Preferably, an adaptive learning algorithm is used to fuse the optimized weight coefficients with the user feature vector to calculate the weighted matching score.

[0125] For example, the optimized weight coefficient for course PRO-101 is 0.78. After fusing it with the user vector, the weighted score is 0.75 × 0.78 = 0.585, resulting in a sorted list of courses, with PRO-101 ranking first. This fusion dynamically adjusts the score, enhancing the accuracy of the recommendations.

[0126] Specifically, duplicate course identifiers are removed by deduplicating the sorted course list.

[0127] For example, if PRO-101 appears twice in the list, redundancy is removed by identifier comparison, resulting in a processed list of recommended courses. This process avoids duplicate exposure and improves the simplicity of the list.

[0128] For example, the order of recommended courses can be adjusted based on the processed list and user interests. Suppose that PRO-101 has the highest score in the list, but the user prefers newer courses, then that course is placed at the top, generating an adjusted list of recommended courses. This adjustment reflects the user's real-time interests and improves click-through rates.

[0129] In one embodiment, if the number of recommended courses in the adjusted list is less than a preset threshold of 3, then supplementary courses with high relevance are extracted from the course content.

[0130] For example, if the original list only has two courses, a related algorithm course can be added to complete the initial recommended course list. This supplementation mechanism ensures the diversity of recommendations and prevents the list from being too short, which could negatively impact user choices.

[0131] Step S104: Extract course content descriptions from the obtained preliminary recommended course list, convert the descriptions into vector representations using word embedding technology, and obtain semantic matching scores by calculating vector similarity to match the learning objectives corresponding to user features.

[0132] Course content descriptions are obtained from the initial recommended course list. A word segmentation tool is used to segment the descriptions into standardized word sequences. These standardized word sequences are then converted into vector representations using a pre-trained word embedding model, generating a course content vector set. User learning target descriptions are obtained and converted into target vector representations using the same word embedding model. An initial matching score for each course is obtained by calculating the cosine similarity between the course content vector set and the target vector representation. If the initial matching score is greater than a preset threshold, the corresponding course vector is retained, generating a candidate course matching set. Based on the candidate course matching sets, a weighted average method is used to fuse course metadata, resulting in a final, ranked matching score.

[0133] In one possible implementation, course content descriptions are obtained from the initial recommended course list, and the descriptions are segmented using a word segmentation tool to obtain standardized word sequences. This process helps transform the detailed course descriptions into computable units, facilitating subsequent vectorization operations.

[0134] For example, in the initial recommendation list for programming enthusiasts, a course description like "Advanced Algorithm Optimization and Data Structure Applications" would be segmented by a word segmentation tool into standardized word sequences such as "advanced," "algorithm," "optimization," "data structure," and "applications." This serialization process ensures the structure of the text, facilitating model processing and improving matching accuracy.

[0135] Specifically, a pre-trained word embedding model is used to convert standardized word sequences into vector representations, generating a set of course content vectors. Word embedding models such as Word2Vec can capture the semantic relationships between words, mapping each word to a vector in a multi-dimensional space, and then obtaining the overall course vector by averaging or summing.

[0136] For example, the word sequence of this programming course is converted into a 256-dimensional vector, where the algorithm-related dimension scores highly, reflecting the core theme of the course. This vector representation quantifies abstract text, facilitating similarity calculation.

[0137] It should be noted that the user's learning objective description is obtained and converted into a target vector representation using the same word embedding model. , This consistent model ensures the uniformity of the vector space.

[0138] For example, if a user's goal is described as "mastering algorithm optimization skills," then after word segmentation and embedding, a target vector is generated, highlighting the algorithm dimension and echoing the user's interest vector. This transformation bridges the user's intent with the course content.

[0139] In one embodiment, an initial matching score for each course is obtained by calculating the cosine similarity between the course content vector set and the target vector representation. Cosine similarity measures the proximity of vector directions and has a value between 0 and 1.

[0140] For example, the similarity score between the PRO-101 course vector and the target vector is 0.82, indicating a high match, while the similarity score for another course, DS-202, is 0.65. This quantitative score provides an objective basis for selection and avoids subjective bias.

[0141] Preferably, if the initial matching score is greater than a preset threshold, the corresponding course vector is retained, and a candidate course matching set is generated. Setting a threshold such as 0.7 can filter out low-relevance items.

[0142] For example, PRO-101 with a value of 0.82 is retained because it exceeds the threshold, while DS-202 with a value of 0.65 is discarded, forming a refined candidate set. This retention mechanism reduces noise and improves recommendation quality.

[0143] For example, based on the candidate course matching set, a weighted average method is used to integrate course metadata to obtain the final matching score after ranking. Metadata includes course difficulty, duration, etc., and the initial score is weighted by weights such as difficulty weight of 0.4 and duration weight of 0.3.

[0144] For example, PRO-101 initially scores 0. Analyzing a user request—which involves retrieving course content descriptions from the initial recommended course list—results in a score of 0.82. After merging, this score becomes 0.82 × 0.7 + metadata adjustment = 0.85, resulting in a higher final score. This fusion integrates multiple factors, optimizes the ranking, and ensures that recommendations better match the user's actual needs, thereby improving learning efficiency.

[0145] Step S105: Based on the obtained semantic matching score, if the score is higher than a preset threshold, select a matching course for content customization. Adjust the course difficulty and knowledge point structure by analyzing user characteristics to determine the customized course version.

[0146] Courses with matching scores higher than a preset threshold are selected to form a course set. For this set, course descriptions are segmented using a word segmentation tool to generate a sequence of segmented words. Knowledge point identifiers are extracted from the segmented word sequences to obtain a knowledge point list. User feature descriptions are converted into feature vector representations using a pre-trained language model to generate user feature vectors. The cosine similarity between the knowledge point list and the user feature vectors is calculated to obtain a similarity value set. If the value in the similarity value set exceeds a preset similarity threshold, the corresponding knowledge point is retained, generating a retained knowledge point subset. Based on the retained knowledge point subset, a hierarchical clustering algorithm is used to group the knowledge points, resulting in knowledge point grouping results. The complexity index of each group is evaluated from the knowledge point grouping results, generating a complexity evaluation set. The presentation order of the knowledge point groups is adjusted using the complexity evaluation set, resulting in an adjusted order list. For the adjusted order list, preference elements from the user feature vectors are integrated to generate a fused preference list. Based on the fused preference list, a course module sequence is constructed, resulting in a module sequence set. The module sequence set is then combined to form a complete course framework, determining the customized course version.

[0147] In one possible implementation, a course set is selected from courses with matching scores higher than a preset threshold. For example, if the threshold is 0.75, courses PRO-101 and DS-301 might be selected. This selection ensures that highly relevant courses from the initial screening proceed to subsequent processing, improving the overall matching accuracy.

[0148] Specifically, for the selected course set, a word segmentation tool is used to segment the course descriptions and generate a sequence of segmented words.

[0149] For example, the description of PRO-101, "Advanced Algorithm Optimization and Data Structure Applications," is segmented into sequences such as "Advanced," "Algorithm," "Optimization," "Data Structure," and "Applications." This segmentation breaks the text down into basic units, facilitating the extraction of core elements. Knowledge point identifiers are extracted from the segmented word sequences to obtain a list of knowledge points.

[0150] For example, "algorithm optimization" and "data structure application" can be identified from a sequence as knowledge points and listed. This extraction focuses on the core concepts of the course, facilitating subsequent correlation analysis.

[0151] It should be noted that, through a pre-trained language model, user feature descriptions are converted into feature vector representations, generating user feature vectors.

[0152] For example, the user characteristic "programming enthusiast, preference for algorithm skills" is transformed into a 512-dimensional vector by the BERT model, with a higher value for the algorithm-related dimension. This transformation captures the semantic essence of the user's intent.

[0153] In one embodiment, the cosine similarity between the knowledge point list and the user feature vector is calculated to obtain a set of similarity values.

[0154] For example, the similarity between the knowledge point "algorithm optimization" and the user vector is 0.85, while that between "data structure application" and "user vector" is 0.78. This calculation quantifies the semantic fit between the knowledge point and the user's features.

[0155] Preferably, if the values ​​in the similarity value set exceed a preset similarity threshold, such as 0.7, the corresponding knowledge points are retained, generating a subset of retained knowledge points. For example, analyzing a user request—which involves retrieving a description from a course list and calculating a matching score—values ​​of 0.85 and 0.78 both exceed the threshold and are retained, forming a subset. This retention filters out low-relevance knowledge points, refining the content focus. Based on the subset of retained knowledge points, a hierarchical clustering algorithm is used to group the knowledge points, obtaining the knowledge point grouping results.

[0156] "Algorithm optimization" and related points are grouped into "Optimization Groups," and "Data Structure Applications" are grouped into "Structure Groups." This grouping reveals the inherent connections between knowledge points and facilitates modular organization. The complexity metrics of each group are evaluated from the knowledge point grouping results to generate a complexity evaluation set.

[0157] The "optimization group" has a medium complexity metric, assessed at level 3 based on the number of knowledge points and dependencies. This assessment considers the gradual progression of the learning path. The presentation order of the knowledge point groups is adjusted using the complexity assessment set to obtain the adjusted order list.

[0158] The system first presents low-complexity "structure groups," followed by "optimization groups." This arrangement of groupings from easy to difficult optimizes learning fluency. For the adjusted list, preference elements from the user's feature vectors are then integrated to generate a fused preference list.

[0159] By incorporating a user's preference weight of 0.6 for the algorithm, the priority of the "optimization group" is increased. This personalized ranking enhances user satisfaction. Based on the integrated preference list, a sequence of course modules is constructed, resulting in a module sequence set.

[0160] The sequence is "basic structure module - optimization module," forming a coherent path. The complete course framework is assembled from these module sequences, determining the customized course version.

[0161] The framework integrates the PRO-101 module with the DS-301 extension to generate customized versions. This combination provides a comprehensive learning solution.

[0162] Step S106: Obtain the finalized course version, analyze the knowledge dependencies and difficulty progression between courses, use a path planning model to optimize the learning sequence to cover the learning objectives, determine whether the optimized sequence improves the overall effect, and obtain the optimized learning sequence:

[0163] Obtain the course version and determine the knowledge point structure within it. Based on this structure, construct a knowledge dependency graph, where nodes represent knowledge points and edges represent dependencies. Calculate a difficulty progression sequence from the knowledge dependency graph, sorted by the difficulty value of each knowledge point. Optimize the learning order using the A* algorithm, taking the knowledge dependency graph and difficulty progression sequence as input and outputting an initial path. Adjust the path based on its coverage of the learning objectives to generate an optimized path. Simulate the learning process to assess the overall effectiveness of the optimized path; if the simulated score exceeds a threshold, the optimized path is confirmed. Generate an optimized learning sequence based on the optimized path.

[0164] In one embodiment, when obtaining the course version and determining the knowledge point structure, core knowledge points can be extracted by analyzing the course version's teaching objectives and syllabus. Assuming the course version is DS-301, and the teaching objective is to master data structures and algorithm applications, its knowledge point structure includes linear structures, tree structures, and graph structures. Core knowledge points can be further refined into array operations, binary tree traversal, and graph search algorithms. This extraction method ensures a clear knowledge point structure, facilitating subsequent dependency analysis. In constructing the knowledge dependency graph, nodes represent knowledge points, and edges represent the pre- and post-learning dependencies.

[0165] For example, array operations, as fundamental nodes, have dependency edges with binary tree traversal nodes because understanding array operations helps in grasping the recursive implementation of tree structures. Graph search algorithm nodes depend on binary tree traversal nodes because search algorithms require recursive thinking. In this way, the knowledge dependency graph clearly presents the logical order of learning content, which helps optimize the learning path. When calculating the progressive difficulty sequence, the knowledge points are sorted based on their difficulty values.

[0166] For example, array operations have a difficulty value of 1, binary tree traversal has a difficulty value of 2, and graph search algorithm has a difficulty value of 3. After sorting, the resulting sequence is: array operations - binary tree traversal - graph search algorithm. This sequence ensures learners progress gradually from simple to complex, improving learning efficiency. When using the A* algorithm to optimize the learning order, the input is a knowledge dependency graph and a progressively increasing difficulty sequence, and the output is an initial path.

[0167] For example, the A* algorithm evaluates the dependency weights and difficulty values ​​between nodes, generating a path: array operations → binary tree traversal → graph search algorithm. This path considers dependencies and increasing difficulty, ensuring smooth learning. When adjusting the path to generate an optimized path, coverage analysis is performed in conjunction with the learning objective.

[0168] For example, if the learning objective is to master the application of graph algorithms, the initial path already covers array operations and binary tree traversal, but needs adjustment to better highlight graph search algorithms. The graph search algorithm module can be prioritized, generating an optimized path: array operations → graph search algorithm → binary tree traversal. This adjustment ensures that the key objective is covered first. When evaluating the effectiveness of the optimized path through simulated learning, a simulated score threshold of 0.8 is set.

[0169] For example, the simulated learner's mastery rate under the optimized path was 0.85, which is higher than the threshold, indicating that the path is effective. The simulation process considers learning time and knowledge point mastery rate, and the optimized path can reduce the learner's cognitive burden. When generating the optimized learning sequence, modules are integrated according to the optimized path to form the final sequence.

[0170] For example, the sequence is: array operations module - graph search algorithm module - binary tree traversal module. This sequence provides a clear learning path, and the modular design facilitates a gradual learning process.

[0171] It should be noted that the modules in the sequence can be further broken down into smaller tasks. For example, the array operation module includes array sorting and searching tasks, which improves the flexibility of learning.

[0172] In one embodiment, if learners prefer hands-on learning, practical tasks can be added to the sequence.

[0173] For example, a task implementing depth-first search can be added after the graph search algorithm module. This approach enhances learners' hands-on skills and improves the practical application of knowledge points.

[0174] Step S107: Extract sequence effect data from the obtained optimized learning sequence, adjust the position of courses in the sequence by comparing the sequence effect with the preset maximum effect standard, and obtain the final recommended learning path:

[0175] The learning sequence is optimized to obtain performance data. By analyzing the completion rate and target matching degree of each course in the sequence, a performance data set is obtained. For this performance data set, a cosine similarity algorithm is used to compare the performance data with a preset maximum performance standard to determine if the similarity reaches a preset threshold. If the comparison result is lower than the preset threshold, the course position adjustment range is calculated based on the deviation value in the performance data, generating an adjusted course position set. Using the adjusted course position set and knowledge point dependencies, a topological sorting algorithm is used to rearrange the learning sequence, generating a preliminary recommended path. Path coverage is extracted from the preliminary recommended path. By comparing the path coverage with the learning objective, it is determined whether the path meets the target requirements, obtaining a path verification result. If the path verification result meets the target requirements, a final recommended learning path is generated based on the preliminary recommended path. Based on the final recommended learning path, a path description containing the course order and knowledge point associations is generated, resulting in a recommended learning plan.

[0176] For example, when analyzing the completion rate and goal matching degree of each course in an optimization learning sequence, an effectiveness dataset can be obtained by analyzing learners' actual learning performance. Completion rate can be based on the learner's score percentage in the course tasks, while goal matching degree measures the degree of alignment between the course content and the learning objectives. Assume the learning sequence includes module A (linked list operations), module B (binary search tree), and module C (dynamic programming). By recording learners' test scores and task completion status in each module, an effectiveness dataset can be obtained.

[0177] For example, the completion score for Module A is 0.9, and the goal matching score is 0.95; the completion score for Module B is 0.85, and the goal matching score is 0.9; and the completion score for Module C is 0.75, and the goal matching score is 0.8. These data reflect the differences in learners' performance across different modules, facilitating subsequent analysis.

[0178] In one possible implementation, when comparing the performance data with a preset maximum performance standard using the cosine similarity algorithm, the performance data set can be treated as a vector and compared with the ideal standard vector. The preset maximum performance standard is a completion rate of 1.0 and a target matching rate of 1.0, with a threshold set to 0.9. Assume that the cosine similarity of module A is 0.92, module B is 0.88, and module C is 0.82. Module C's similarity is below the threshold, indicating that its performance has not met expectations. The deviation value can be calculated by comparing the actual completion rate with the target value; the deviation value for module C is 0.25. This analysis method intuitively reflects the gap in course effectiveness.

[0179] Specifically, if the comparison result is lower than the threshold, the course position adjustment range is calculated based on the deviation value.

[0180] For example, a deviation value of 0.25 for Module C indicates that its difficulty or content arrangement may not be suitable for its current position in the sequence. Adjustments can be made based on the percentage of deviation; it is suggested that Module C be moved forward to a more fundamental position in the sequence to reduce learning difficulty. The adjusted course position set is Module A, Module C, and Module B. This adjustment takes into account the learners' actual performance and the course difficulty.

[0181] In one embodiment, when rearranging the learning sequence using a topological sorting algorithm based on knowledge point dependencies, a dependency graph can be constructed. Linked list operations are the foundation of binary search trees, which are prerequisites for dynamic programming. Topological sorting ensures that module A precedes module C, and module C precedes module B, generating a preliminary recommended path: module A → module C → module B. This path respects the logical dependencies between knowledge points.

[0182] For example, when extracting path coverage from the initial recommended path, the degree to which the knowledge points in the path cover the learning objective can be statistically analyzed. Assuming the learning objective is to master the application of dynamic programming, path coverage is evaluated by checking whether module C adequately covers the dynamic programming knowledge points. If the coverage rate is 0.9, it indicates that the path basically meets the objective requirements. After confirming the path validation results, the final recommended learning path is Module A → Module C → Module B.

[0183] In one possible implementation, when generating recommended learning plans, the path description needs to clearly explain the course order and the association between knowledge points.

[0184] For example, a recommended learning plan might be described as follows: First, learn linked list operations in module A to master basic data structures; then learn dynamic programming in module C to understand complex algorithm design; finally, learn binary search trees in module B to solidify knowledge of tree structures. This clear and modular approach helps learners progress step by step.

[0185] Example 2

[0186] like Figure 2 As shown, based on the same overall inventive concept, this invention also provides an educational course recommendation system based on data mining. The educational course recommendation system based on data mining provided by this invention is described below. The system described below can be referred to in correspondence with the method described above. The system includes:

[0187] The course popularity prediction module is used to collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate;

[0188] The weight adjustment module is used to increase the dynamic weight coefficient of the corresponding course when the course popularity value exceeds the preset threshold, and to optimize the weight coefficient based on the correlation between the course popularity value and the weight coefficient.

[0189] The preliminary recommendation module is used to fuse optimized weight coefficients and user features using an adaptive learning algorithm, determine the matching degree between user features and course learning objectives, and obtain a preliminary recommended course list.

[0190] The semantic matching module is used to perform vector transformation on the initial recommended course list based on word embedding technology. It calculates the vector similarity to match the learning target corresponding to the user features and obtains the semantic matching score.

[0191] The course customization module is used to select matching courses for content customization when the semantic matching score is higher than a preset threshold, and to determine the customized course version.

[0192] The learning sequence optimization module is used to analyze the knowledge dependencies and difficulty progression relationships between customized course versions, and uses a path planning algorithm to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence.

[0193] The final path generation module is used to extract sequence effect data from the optimized learning sequence, and by comparing the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence to obtain the final educational course recommendation scheme.

[0194] Example 3

[0195] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0196] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for recommending educational courses based on data mining, characterized in that, Includes the following steps: Collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate; If the course popularity value exceeds the preset threshold, the dynamic weight coefficient of the corresponding course will be increased, and the weight coefficient will be optimized based on the correlation between the course popularity value and the weight coefficient. An adaptive learning algorithm is used to fuse optimized weight coefficients and user features to determine the matching degree between user features and course learning objectives, thereby obtaining a preliminary recommended course list. The initial recommended course list is vectorized based on word embedding technology. The semantic matching score is obtained by calculating the vector similarity to match the learning target corresponding to the user features. If the semantic matching score is higher than the preset threshold, the matching course is selected for content customization, and the customized course version is determined. The knowledge dependencies and difficulty progression between customized course versions are analyzed, and a path planning algorithm is used to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence. Extract sequence effect data from the optimized learning sequence, compare the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence, and obtain the final educational course recommendation scheme.

2. The method according to claim 1, characterized in that, The process of collecting popularity index data, processing the popularity index data using a deep learning model, and obtaining the course popularity value includes: Process course visit and completion rate data to obtain cleaned popularity index data; Based on the heat index data after cleaning, a convolutional neural network is used to extract the spatial correlation features between the indicators to obtain the correlation feature vector; If the associated feature vectors meet the preset threshold, the associated feature vectors are processed by a long short-term memory network to obtain the time series prediction results. Based on the time series prediction results, principal component analysis was used for dimensionality reduction to obtain the dimensionality-reduced trend vector. The trend vectors after dimensionality reduction are grouped using a clustering algorithm to obtain popularity trend classification labels; Calculate the course popularity score based on the popularity trend classification tags.

3. The method according to claim 1, characterized in that, If a course's popularity score exceeds a preset threshold, the corresponding course's dynamic weight coefficient is increased. Based on the correlation between course popularity and the weight coefficient, the process of optimizing the weight coefficient includes: If the course popularity value is greater than the preset threshold, the course identifiers that meet the conditions are filtered out based on the threshold, and a set of filtered courses is obtained. Based on the selected course set, the product of the course popularity value and the dynamic weight is calculated to obtain the preliminary weight coefficient; Based on the initial weighting coefficients, the coefficient distribution in the course set is processed by mean normalization to obtain standardized weighting coefficients; If the standardized weight coefficients exceed the preset range, the coefficient range is adjusted through linear mapping to obtain the optimized weight coefficients.

4. The method according to claim 1, characterized in that, The process of using an adaptive learning algorithm to fuse optimized weight coefficients and user features, determining the matching degree between user features and course learning objectives, and obtaining a preliminary recommended course list includes: An adaptive learning algorithm is used to fuse optimized weight coefficients with user feature vectors, calculate a weighted matching score, and obtain a sorted course list. The sorted course list is deduplicated to obtain the processed recommended course list. Based on user interests and preferences, the order of course recommendations is adjusted to generate an adjusted list of recommended courses; If the number of courses in the adjusted recommended course list is less than the preset threshold, then courses with a relevance higher than the preset threshold are extracted from the original course content and added to the recommendation results to obtain a complete preliminary recommended course list.

5. The method according to claim 1, characterized in that, The process of performing vector transformation on the initial recommended course list based on word embedding technology, and obtaining a semantic matching score by calculating vector similarity to match the learning objectives corresponding to user features, includes: The course content descriptions are obtained from the initial recommended course list, and the text is segmented using a word segmentation tool to obtain a standardized word sequence. A standardized word sequence is converted into a vector representation by a pre-trained word embedding model, generating a vector set of course content. Obtain the user's learning objective description and convert it into a target vector representation using a pre-trained word embedding model; The initial matching score for each course is obtained by calculating the cosine similarity between the course content vector set and the target vector representation. If the initial matching score is greater than the preset threshold, the corresponding course vector is retained and a candidate course matching set is generated. Based on the candidate course matching set, a weighted average method is used to merge the course data to obtain the final matching score.

6. The method according to claim 1, characterized in that, If the semantic matching score is higher than a preset threshold, the matching course is selected for content customization. The process of determining the customized course version includes: Retrieve courses whose matching scores are higher than a preset threshold to obtain a set of matching courses; A word segmentation tool is used to segment the description of the matching course set, generating a sequence of segmented words; Extract knowledge point identifiers from the segmented word sequence to obtain a list of knowledge points; Calculate the cosine similarity between the knowledge point list and the user feature vector to obtain a set of similarity values; If the values ​​in the similarity value set exceed the preset similarity threshold, the corresponding knowledge points are retained, and a subset of retained knowledge points is generated. Based on the retained subset of knowledge points, a hierarchical clustering algorithm is used to group the knowledge points to obtain the knowledge point grouping results; From the knowledge point grouping results, evaluate the complexity index of each group and generate a complexity evaluation set; By adjusting the presentation order of knowledge point groups using the complexity evaluation set, we obtain the adjusted order list. For the adjusted order list, the preference elements in the user feature vector are merged to generate a merged preference list; Based on the list of integration preferences, construct a sequence of course modules to obtain a set of module sequences; By combining the module sequence set, a complete course framework is formed, and a customized course version is determined.

7. The method according to claim 1, characterized in that, The process of analyzing the knowledge dependencies and difficulty progression among customized course versions, and using path planning algorithms to optimize the learning sequence to cover learning objectives, includes: Construct a knowledge dependency graph based on the knowledge point structure in the course version; Calculate the progressive difficulty sequence using a knowledge dependency graph; The A* algorithm is used to optimize the learning order. The input is a knowledge dependency graph and a difficulty progression sequence, and the output is a preliminary path. Based on the coverage of the learning objectives by the initial path, adjust the path to generate an optimized path; The overall effect of the optimization path is judged by simulating the learning process. If the simulated score is higher than the threshold, the optimization path is determined. Based on the determined optimization path, an optimization learning sequence is generated.

8. The method according to claim 1, characterized in that, The process of extracting sequence performance data from optimized learning sequences, adjusting the positions of courses within the sequence by comparing the sequence performance with a preset maximum performance standard, and obtaining the final educational course recommendation scheme includes: The effectiveness data is obtained from the optimized learning sequence. By analyzing the completion rate and target matching rate of each course in the sequence, an effectiveness data set is obtained. For the effect dataset, the cosine similarity algorithm is used to compare the effect data with the preset maximum effect standard, and it is determined whether the similarity reaches the preset threshold to obtain the comparison result; If the comparison result is lower than the preset threshold, the course position adjustment range is calculated based on the deviation value in the effect data, and the adjusted course position set is generated. By adjusting the set of course locations and combining the dependencies between knowledge points, the learning sequence is rearranged using a topological sorting algorithm to generate a preliminary recommended path. The path coverage rate is extracted from the initial recommended path, and the path validation results are obtained by comparing the degree of matching between the path coverage rate and the learning objective. If the path verification results meet the target requirements, then the final recommended learning path is generated based on the preliminary recommended path. Based on the final recommended learning path, a path description including the course sequence and the association of knowledge points is generated, resulting in an educational course recommendation scheme.

9. An educational course recommendation system based on data mining, characterized in that, The system for implementing the method according to any one of claims 1-8, the system comprising: The course popularity prediction module is used to collect popularity index data, process the popularity index data using a deep learning model, and obtain the course popularity value; wherein, the popularity index data includes course views and completion rate; The weight adjustment module is used to increase the dynamic weight coefficient of the corresponding course when the course popularity value exceeds the preset threshold, and to optimize the weight coefficient based on the correlation between the course popularity value and the weight coefficient. The preliminary recommendation module is used to fuse optimized weight coefficients and user features using an adaptive learning algorithm, determine the matching degree between user features and course learning objectives, and obtain a preliminary recommended course list. The semantic matching module is used to perform vector transformation on the initial recommended course list based on word embedding technology. It calculates the vector similarity to match the learning target corresponding to the user features and obtains the semantic matching score. The course customization module is used to select matching courses for content customization when the semantic matching score is higher than a preset threshold, and to determine the customized course version. The learning sequence optimization module is used to analyze the knowledge dependencies and difficulty progression relationships between customized course versions, and uses a path planning algorithm to optimize the learning order to cover the learning objectives, thereby obtaining an optimized learning sequence. The final path generation module is used to extract sequence effect data from the optimized learning sequence, and by comparing the sequence effect with the preset maximum effect standard, adjust the position of courses in the sequence to obtain the final educational course recommendation scheme.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-8.

Images