Course pushing method for intelligent analysis of learning habits of continuous education students
By constructing a learning habit analysis model and a hybrid intelligent recommendation algorithm, the problem of inaccurate course recommendations in traditional course push methods is solved, personalized course recommendations are realized, and learning effect and resource utilization are improved.
Patent Information
- Application Number
- CN202510220293.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-27
- Publication Date
- 2025-08-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional course push methods cannot accurately meet the personalized learning needs of continuing education students, and the course recommendations lack accuracy. The difficulty of VR/AR courses does not match the students' learning style and needs, resulting in poor learning results.
By collecting students' learning behavior data in real time, building a learning habit analysis model, combining course classification and VR/AR course adaptability assessment, and using a hybrid intelligent recommendation algorithm to provide students with personalized course recommendations.
It has achieved accurate personalized course recommendations, improved learning effect and satisfaction, and given priority to recommend VR/AR simulation experience courses, enriched the learning resources form, and improved the quality of continuing education.
Smart Images
Figure CN120407910A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of educational technology, and specifically provides a course push method for intelligent analysis of the learning habits of continuing education students. Background Art
[0002] With the continuous development of educational technology, the field of continuing education is facing more and more challenges and opportunities. Due to diverse backgrounds and different needs, continuing education students have different requirements for course selection and learning methods. Traditional course push methods often rely on fixed course lists or simple classifications, making it difficult to accurately meet the personalized learning needs of students.
[0003] Traditional technologies have drawbacks. Firstly, course recommendations lack accuracy and often cannot be recommended according to the actual learning progress and interests of students. Secondly, the course adaptability is low. The difficulty and application scenario characteristics of VR / AR courses may not match the learning styles and needs of students, resulting in poor learning effects.
[0004] In summary, traditional course push methods have many deficiencies when facing the personalized learning needs of continuing education students. Therefore, it is particularly important to develop a course push method for intelligent analysis of the learning habits of continuing education students. Summary of the Invention
[0005] The purpose of the present invention is to make up for the deficiencies of the prior art and provide a course push method for intelligent analysis of the learning habits of continuing education students. It can collect and analyze the learning behavior data of students in real time, build an accurate learning habit analysis model, and comprehensively consider the classification and detailed annotation information of courses as well as the adaptability evaluation of VR / AR courses to provide personalized course recommendations for students.
[0006] To solve the above technical problems, the present invention provides the following technical solutions: A course push method for intelligent analysis of the learning habits of continuing education students. The specific steps of this method are as follows:
[0007] S1. Collection of Learning Habit Data
[0008] Through multiple channels such as the learning management system, online learning platform, and mobile learning client, collect the learning behavior data of continuing education students in real time, including course browsing records, specifically recording the timestamp, browsing duration, and course chapter jump situation of each course browsing by students, learning duration data, accurately counting the cumulative learning time of students in each course to the second, homework completion situation, covering the homework submission time, answer correct rate, wrong question distribution, and time spent on each question, and exam score information, including the exam score, scores of each question type, and exam time used;
[0009] Meanwhile, by using online questionnaires, offline interviews, and learning community interactions, subjective data on learners' learning preferences and interest areas are collected. In the questionnaire survey, a series of structured and semi-structured questions are set. In the offline interviews, in-depth exchanges are conducted with some typical learners to explore their potential learning needs. The learning community interactions analyze the speech content and participation frequency of learners in the forum and group discussions to obtain their focus of attention and learning tendencies.
[0010] S2. Data preprocessing
[0011] The collected raw data is cleaned to identify and process the noise and outliers in the data. A statistics-based method is used to perform deduplication on the data. Each record is uniquely identified through a hash algorithm, and duplicate records are removed. Normalization processing is carried out. For numerical data with different dimensions, the Min-Max normalization method is used to map the data to the [0, 1] interval. The formula is: where x is the raw data, x min and x max are the minimum and maximum values of this feature respectively. The data from different sources is integrated to construct a learner learning habit dataset indexed by the learner's unique identifier.
[0012] S3. Construction of the learning habit analysis model
[0013] Deep learning model - Adaptive Learning Habit Analysis Network. This model consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the preprocessed learner learning habit data and converts it into a vector form. The hidden layer adopts a multi-layer perceptron structure, with each layer containing several neurons. The neurons are connected by weights. The weight W ij represents the connection strength between the j-th neuron in the i-th layer and the neurons in the i + 1-th layer. Its initial value is determined by the Xavier initialization method. An attention mechanism is introduced between the hidden layers to automatically learn the importance of different features in learning habit analysis. Specifically, for the input feature X, the attention weight α is calculated: where e = W a ·tanh(W b ·X + b), W a and W are trainable parameters, b is the bias term, n is the number of features. The features are weighted and summed through the attention weight to obtain a more representative feature representation. The output layer predicts the learner's learning style, learning progress, and knowledge mastery level labels based on the output of the hidden layer. The cross-entropy loss function is used to train the model, where N is the number of samples, C is the number of label categories, y ij is the true label. It is a prediction label, and the model parameters are updated through the stochastic gradient descent algorithm to minimize the loss function;
[0014] S4. Course Classification and Annotation
[0015] Comprehensively classify the existing continuing education courses. According to the dimensions of subject area, course nature, and teaching objectives, the courses are divided into major categories such as medicine, engineering, humanities and social sciences, management, and art, and each major category is further subdivided into several minor categories;
[0016] Organize professional education experts, course teachers, and field scholars to conduct detailed annotation on each course. The annotation content includes the keywords of the course. By analyzing the course syllabus and teaching content, extract the vocabulary that can accurately reflect the core content of the course, the difficulty level. According to the knowledge depth, skill requirements, and learning time factors of the course, use a five-level scoring system, simple, relatively simple, medium, relatively difficult, difficult for evaluation, the applicable scenarios, clarify whether the course is applicable to beginners, advanced learners, or senior professionals at different levels of learners. For courses containing VR / AR experiences, specifically mark their VR / AR application characteristics, applicable learning fields, and the required hardware devices;
[0017] S5. VR / AR Course Adaptation
[0018] For course fields with strong practicality, give priority to recommending courses with VR / AR simulation experiences. Based on the results of the learning habit analysis model, construct a course adaptation degree evaluation function. Let the learning style vector of the learner be S, the learning progress be P, the knowledge mastery level be K, the difficulty level of the course be D, the applicable scenario vector be A, and the VR / AR application characteristic vector be V. Then the course adaptation degree F is: F = w1·cosine(S,A) + w2·(P - D) + w3·(K - D) + w4·cosine(V,S), where w1, w2, w3, and w4 are weights. Sort the courses according to the course adaptation degree F and recommend the VR / AR courses with the highest adaptation degree for the learners;
[0019] S6. Course Push
[0020] Integrate the learning habit analysis results of the learners, the course classification and annotation information, and the VR / AR course adaptation situation, and use a hybrid intelligent recommendation algorithm that combines collaborative filtering and content-based recommendation algorithms. For the collaborative filtering part, by calculating the similarity between learners, find neighbor learners with similar interests to the target learner, and use the cosine similarity algorithm to calculate the similarity sim(i,j) between learner i and learner j: where x i k and x jk is the learning behavior feature vector of student i and student j for course k, and m is the number of courses. Based on the course preferences of neighbor students, recommend to the target students the courses they like but the target students have not yet studied. For the content-based recommendation part, calculate the similarity between courses according to the annotation information of the courses, process the course keywords using the TF-IDF algorithm to obtain the text feature vector of the courses, and then calculate the similarity between courses through cosine similarity. Let the text feature vectors of course a and course b be T a and T b , then the course similarity Recommend courses similar to those already studied by the target students. Finally, fuse the collaborative filtering and content-based recommendation results, and adjust the proportion of the two recommendation results through weights. Let the weight of the collaborative filtering recommendation result be The weight of the content-based recommendation result is w n , and w c + w n = 1. Determine the values of w c and w n through experiments and cross-validation to obtain the final course recommendation list. In the pushed course list, highlight the courses with VR / AR experiences and provide detailed course introductions, highlights of VR / AR experiences, learning objectives, and time required information;
[0021] S7. Feedback and Optimization
[0022] Establish a perfect student feedback mechanism. Through multiple channels such as online questionnaires, course evaluation systems, and customer service feedback, collect students' evaluations of the pushed courses and feedback on learning effects. The evaluation content includes the practicality of course content, the effectiveness of teaching methods, the quality of VR / AR experiences, and the matching degree with their own learning needs. The learning effect feedback includes the improvement of learning achievements, the mastery of knowledge and skills, and the change of learning interest. Analyze the feedback data, extract key information, and optimize the learning habit analysis model, course adaptability evaluation function, and hybrid intelligent recommendation algorithm according to the feedback information. For the learning habit analysis model, retrain the model according to the new feedback data and adjust the model parameters to improve the prediction accuracy of the model for students' learning habits. For the course adaptability evaluation function, adjust the weights w1, w2, w3, w4 according to students' feedback to make the course adaptability more accurately reflect students' needs. For the hybrid intelligent recommendation algorithm, adjust the fusion weights w c and w n of the collaborative filtering and content-based recommendation results to optimize the accuracy and diversity of the recommendation results.
[0023] Furthermore, during the process of collecting learning habit data, for the collection of learning community interaction data, it also includes analyzing the resources posted by students in the learning community, extracting the theme, knowledge point coverage, and quality evaluation information of the resources, and incorporating this information into the student learning habit dataset to more comprehensively reflect the learning behavior and knowledge contribution of students. Through natural language processing technology, lexical analysis, syntactic analysis, and semantic understanding are performed on the text content posted by students to extract key information and enrich the student learning habit portrait.
[0024] Even further, in the cleaning step of data preprocessing, the Local Outlier Factor (LOF) algorithm is also introduced to detect and process outliers in the data. For each sample point in the dataset, its LOF value is calculated. If the LOF value is greater than the set threshold, then the sample point is determined to be an outlier. The calculation formula for the LOF value is: where p is the sample point, and N k (p) is the set of k-nearest neighbors of p, and reach-dist k (o,p) is the reachable distance from point o to point p, and k-dist(p) is the k-distance of point p. By this method, outliers in the data can be more accurately identified and processed, improving the data quality.
[0025] Even further, in the construction of the learning habit analysis model, the activation function of the hidden layer uses the Rectified Linear Unit (RelU) function - Adaptive Rectified Linear Unit (AReLU). The traditional RelU function outputs 0 when the input is less than 0, which may lead to the problem of neuron death. The AReLU function is defined as: where α is the adaptive parameter, which is dynamically adjusted according to the activation situation of neurons during the training process. When a neuron remains inactive in multiple training batches, the value of α is gradually decreased to increase the activation probability of the neuron. When a neuron is frequently activated, the value of α is increased to avoid over-activation. This can effectively alleviate the problem of neuron death and improve the training effect and generalization ability of the model.
[0026] Even further, during the process of course classification and annotation, for the extraction of course keywords, an automatic extraction method based on word vectors is also adopted. First, a pre-trained word vector model is used to convert each word in the course text into a vector representation. Then, a clustering algorithm is used to cluster similar word vectors into one category, and a representative word is selected from each category as a keyword. At the same time, in combination with the teaching syllabus and knowledge point structure of the course, the automatically extracted keywords are manually reviewed and adjusted to ensure that the keywords can accurately reflect the core content and key knowledge of the course.
[0027] Further, in the VR / AR course adaptation, the learning time preference and learning environment factors of the trainees are also introduced. Let the learning time preference vector of the trainees be T, the learning environment vector be E, the teaching time arrangement vector of the course be T c , and the learning environment requirement vector required by the course be E c , then the course adaptation degree F is updated as: F = w1·cosine(S, A) + w2·(P - D) + w3·(K - D) + w4·cosine(V, S) + w5·cosine(T, T c ) + w6·cosine(E, E e ), where w5 and w6 are the newly introduced weights, which are determined by analyzing the learning time and environment preference data of the trainees and combining expert experience, and w1 + w2 + w3 + w4 + w5 + w6 = 1. By considering these factors, the adaptability between the course and the trainees can be evaluated more comprehensively, and the accuracy of the recommendation can be improved.
[0028] Further, in the hybrid intelligent recommendation algorithm for course push, in order to improve the accuracy and diversity of the recommendation, a recommendation strategy adjustment mechanism based on reinforcement learning is introduced. The recommendation process is regarded as a Markov decision-making process. The state space is the learning habit characteristics of the trainees and the information of the courses they have learned. The action space is the list of recommended courses. The reward function is defined according to the feedback of the trainees on the recommended courses. Through the reinforcement learning algorithm, continuously learn the optimal recommendation strategy and dynamically adjust the fusion weights w c and w n , as well as the sorting method of the recommended courses to adapt to the needs and learning scenarios of different trainees.
[0029] Further, in the feedback and optimization process, the course classification and annotation system is also updated regularly. With the development of the education field and the application of new technologies, the course content and teaching objectives will change continuously. Therefore, it is necessary to regularly organize experts to re-evaluate and annotate the courses. At the same time, according to the feedback of the trainees and market demands, new course classifications are added or adjusted to ensure that the course classification and annotation system can accurately reflect the actual situation of continuing education courses and provide more accurate basic data for course push.
[0030] Compared with the prior art, the course push method for intelligent analysis of the learning habits of continuing education trainees has the following beneficial effects:
[0031] 1. This method collects the learning behavior data of continuing education students in real time through multiple channels, conducts in-depth preprocessing and intelligent analysis, and constructs an accurate learning habit analysis model. This model can comprehensively consider the learning styles, learning progress, and knowledge mastery of students, combine the classification and detailed annotation information of courses, as well as the adaptability evaluation of VR / AR courses, and provide personalized course recommendations for students. This highly personalized recommendation method not only meets the diverse learning needs of students, but also significantly improves the accuracy of course recommendations, enabling students to find learning resources suitable for themselves more quickly, thereby enhancing learning effects and satisfaction.
[0032] 2. This method uses a hybrid intelligent recommendation algorithm, which combines collaborative filtering and content-based recommendation algorithms, comprehensively considers the historical learning behaviors of students and the annotation information of courses, and generates the final course recommendation list. This method not only improves the diversity and coverage of recommendations, but also can continuously optimize the learning habit analysis model, course adaptability evaluation function, and recommendation algorithm according to the feedback and learning effects of students. Through this continuous learning and optimization, this method can continuously optimize the allocation of educational resources, enabling more high-quality courses to be accurately pushed to students in need. At the same time, for course fields with strong practicality, this method also preferentially recommends courses with VR / AR simulation experiences, further enriching the forms and contents of learning resources and enhancing the overall quality and effects of continuing education.
[0033] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.
[0035] Figure 1 It is a flowchart operation diagram of a course push method for intelligent analysis of the learning habits of continuing education students. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0036] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention objective, the following will, in conjunction with the accompanying drawings and preferred embodiments, describe in detail the specific embodiments, structures, features, and their effects of the present invention as follows.
[0037] Example 1
[0038] This example describes a large medical training institution that provides continuing education courses for many medical staff, including multiple fields such as clinical skills improvement, medical research progress, and medical management. It hopes to improve the learning effect and satisfaction of students through accurate course delivery.
[0039] Through a dedicated online learning platform, the learning behaviors of students in various medical courses are recorded. For example, when a doctor is learning the course "New Progress in the Diagnosis and Treatment of Cardiovascular Diseases", the platform records the timestamp of each login and learning duration. If the doctor learns in 3 sessions, with each session lasting 30 minutes, 45 minutes, and 60 minutes respectively, and the total learning duration is 135 minutes, it also records the jump situation of course chapters, such as directly jumping from the basic pathology chapter to the latest treatment method chapter. At the same time, through online questionnaires, the learning preferences of doctors are understood, such as the preference for surgical video teaching and case analysis discussion forms. In the learning community interaction, the discussion content and participation frequency of doctors on cardiovascular disease-related topics are analyzed. For example, a certain doctor actively participates in the discussion of the treatment plan for arrhythmia, and the speech content shows his attention to the application of new technologies.
[0040] The collected data is cleaned to remove duplicate records, such as the same browsing records submitted by the same doctor multiple times. The numerical data of learning duration is normalized. Assuming the minimum learning duration is 0 minutes and the maximum is 180 minutes, the 135-minute learning duration of a certain doctor is normalized to Integrate data from different sources to construct a student learning habit dataset.
[0041] An adaptive learning habit analysis network is adopted. The input layer receives the preprocessed learning habit data and converts it into a vector. The hidden layer adopts a multi-layer perceptron structure, and the weights between neurons are determined by the Xavier initialization method. An attention mechanism is introduced to calculate the attention weights. For example, for features such as learning duration, chapter jump times, and community discussion participation, calculate their importance weights. The output layer predicts the learning style, progress, and knowledge mastery of students. The cross-entropy loss function is used to train the model. For example, for the learning data samples of 100 students, calculate the difference between the predicted label and the true label, and update the model parameters through the stochastic gradient descent algorithm.
[0042] The medical courses are divided into small categories such as internal medicine, surgery, obstetrics and gynecology, and pediatrics. Medical experts and course teachers are organized to annotate keywords for the course "New Advances in the Diagnosis and Treatment of Cardiovascular Diseases", such as "cardiovascular diseases", "new diagnostic and treatment technologies", and "clinical applications". The difficulty level is rated as medium, and the applicable scenarios are marked as suitable for doctors with certain clinical experience who hope to improve their level of diagnosing and treating cardiovascular diseases. If the course includes VR / AR experiences, mark its application characteristics and required hardware devices in terms of heart structure display and surgical simulation operations.
[0043] For highly practical courses, such as surgical skills training courses, VR / AR simulation experience courses are preferentially recommended. Let the learning style vector of the trainee be S, the learning progress be P, the knowledge mastery level be K, the course difficulty level be D, the applicable scenario vector be A, and the VR / AR application characteristic vector be V. The course adaptation degree F = w1·cosine(S,A) + w2·(P - D) + w3·(K - D) + w4·cosine(V,S). The courses are sorted and recommended according to the adaptation degree.
[0044] Using a hybrid intelligent recommendation algorithm, the collaborative filtering part calculates the similarity between trainees, such as the learning behavior feature vectors of two doctors for multiple medical courses. The similarity is calculated through the cosine similarity algorithm, and courses are recommended based on the course preferences of neighbor doctors. For the content-based recommendation part, the TF-IDF algorithm is used to process the course keywords to obtain text feature vectors, and the course similarity is calculated, such as the similarity between the course "New Advances in the Diagnosis and Treatment of Cardiovascular Diseases" and the course "Cardiac Intervention Therapy Technology". Similar courses are recommended based on the characteristics of the courses already studied by the trainees. The two recommendation results are integrated, and the weights w ct and w n are adjusted to highlight the VR / AR experience courses and provide detailed information.
[0045] Embodiment 2
[0046] This embodiment describes the continuing education department of an engineering and technology college, which provides professional courses covering multiple disciplines such as mechanical engineering, electrical engineering, and civil engineering for in-service engineers, aiming to help engineers update their knowledge and improve their skills to meet the needs of industry development, and improve the learning efficiency and professional competitiveness of trainees by accurately pushing appropriate courses.
[0047] Collect the learning behavior data of trainees using the learning management system and mobile learning client. Taking the example of mechanical engineers learning the "CNC Machining Technology" course, record the timestamps and browsing durations of their course browsing. For example, the cumulative learning duration of an engineer within a week is 240 minutes, and the jumping situations between different chapters are also recorded, such as frequently jumping from the programming basics chapter to the tool path planning chapter. At the same time, understand the trainees' learning interest areas through offline interviews. For example, they are interested in the optimization of automated processing equipment. In the learning community interaction, analyze the mechanical design case resources shared by trainees, including the theme, knowledge point coverage, and quality evaluation. For example, the quality of the high-precision mold design case resources shared by a certain trainee is high and the knowledge points are comprehensive.
[0048] When cleaning the data, use the Local Outlier Factor (LOF) algorithm to detect outliers. For example, the record of a trainee completing a high-difficulty course in an extremely short time may be determined as an outlier. Normalize the learning duration data. Assume the minimum learning duration is 0 minutes and the maximum is 300 minutes. The 240-minute learning duration of this engineer after normalization is Integrate the data to build a dataset.
[0049] Build an adaptive learning habit analysis network. The input layer processes the data. The hidden layer adopts a multi-layer perceptron structure, and the activation function uses the Adaptive ReLU (ARelU) function. Dynamically adjust the α value according to the neuron activation situation. For example, if a certain neuron is less activated in the initial training batch, gradually decrease the α value to increase the activation probability. The output layer predicts the trainees' learning status, and train the model with the cross-entropy loss function and optimize the parameters through the stochastic gradient descent algorithm.
[0050] Subdivide engineering courses by discipline. For example, mechanical engineering courses are divided into subcategories such as manufacturing technology, design principles, and automation control. For the "CNC Machining Technology" course, the keywords marked by experts are "CNC programming", "processing technology", and "machine tool operation". The difficulty level is rated as relatively difficult, and the applicable scenario is marked as suitable for engineers with a certain mechanical processing foundation who want to deeply learn CNC technology. If there is a VR / AR experience course, mark its characteristics and hardware requirements in aspects such as machine tool simulation operation and processing process visualization.
[0051] For manufacturing technology courses with strong practicality, give priority to recommending VR / AR courses. Considering the trainees' learning style S, progress P, knowledge mastery level K, course difficulty D, applicable scenario A, VR / AR application characteristics V, introduce the trainees' learning time preference vector T and learning environment vector E. The course adaptability F = w1·cosine(S,A)+w2·(P - D)+w3·(K - D)+w4·cosine(V,S)+w5·cosine(T,T c )+w6·cosine(E,E e ) and recommend according to the adaptability ranking.
[0052] Adopt a hybrid intelligent recommendation algorithm. For the collaborative filtering part, calculate the similarity of students. For example, calculate the cosine similarity based on the learning behavior feature vectors of mechanical engineers for multiple courses, and recommend based on the course preferences of neighboring students. For the content-based recommendation part, process the course keywords through the TF-IDF algorithm to obtain text feature vectors and calculate the course similarity. For example, combine the course similarities of "Numerical Control Machining Technology" and "Precision Machinery Manufacturing Technology" with the characteristics of the courses that the students have already studied for recommendation. Integrate the two recommendation results, introduce reinforcement learning to adjust the weights w ct 、w n and the sorting of recommended courses.
[0053] Collect feedback through various channels. For example, students feedback that the VR / AR machine tool simulation course helps to understand the processing principle, but the operation interface is not friendly enough. Retrain the model, adjust the course adaptation weight, and optimize the recommendation algorithm weight according to the feedback. Regularly update the course classification and annotation. For example, add the course classification of "Additive Manufacturing Technology" according to the new technologies in the industry.
[0054] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the equivalent embodiments with equivalent changes within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.
Claims
1. A course push method for intelligent analysis of the learning habits of continuing education students, characterized in that, The specific steps of this method are as follows: S1. Collection of learning habit data Through multiple channels such as learning management systems, online learning platforms, and mobile learning clients, real-time collection of learning behavior data of continuing education students, specifically including course browsing records, detailed recording of the timestamp, browsing duration, and course chapter jump situation of each course browsing by students, learning duration data, accurately counting the cumulative learning time of students in each course down to the second, homework completion situation, covering homework submission time, answer accuracy rate, wrong question distribution, and time spent on each question, and exam score information, including exam scores, scores of each question type, and exam time used; At the same time, use online questionnaires, offline interviews, and learning community interactions to collect subjective data on students' learning preferences and interest areas. In the questionnaire survey, a series of structured and semi-structured questions are set. Offline interviews are conducted for in-depth communication with some typical students to explore their potential learning needs. Learning community interactions analyze the speech content and participation frequency of students in forums and group discussions to obtain their focus of attention and learning tendencies; S2. Data preprocessing Clean the collected raw data, identify and process the noise and outliers in the data. Adopt a statistics-based method to deduplicate the data. Use the hash algorithm to uniquely identify each record, remove duplicate records, and perform normalization. For numerical data with different dimensions, use the Min-Max normalization method to map the data to the interval [0, 1]. The formula is: where x is the raw data, x min and x max are the minimum and maximum values of this feature respectively. Integrate data from different sources to construct a dataset of students' learning habits indexed by the unique identifier of the students; S3. Construction of learning habit analysis model Deep learning model - Adaptive Learning Habit Analysis Network. This model consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the preprocessed learning habit data of students and converts it into vector form. The hidden layer adopts a multi-layer perceptron structure, with each layer containing a number of neurons. The neurons are connected by weights, and the weight W ij represents the connection strength between the j-th neuron in the i-th layer and the neurons in the i+1-th layer. Its initial value is determined by the Xavier initialization method. An attention mechanism is introduced between the hidden layers to automatically learn the importance of different features in learning habit analysis. Specifically, for the input feature X, the attention weight α is calculated as follows: where e = W a ·tanh(W b ·X + b), W a and W are trainable parameters, b is the bias term, n is the number of features. The features are weighted and summed through the attention weights to obtain a more representative feature representation. The output layer predicts the learning style, learning progress, and knowledge mastery level labels of students based on the output of the hidden layer, and uses the cross-entropy loss function to train the model, where N is the number of samples, C is the number of label categories, y ij is the true label, is the predicted label, and the model parameters are updated through the stochastic gradient descent algorithm to minimize the loss function; S4. Course classification and annotation Conduct a comprehensive classification of existing continuing education courses. According to the dimensions of subject area, course nature, and teaching objectives, the courses are divided into major categories such as medicine, engineering, humanities and social sciences, management, and art, and each major category is further subdivided into several minor categories; Organize professional education experts, course teachers, and field scholars to conduct detailed annotation of each course. The annotation content includes the keywords of the course, extracting the vocabulary that can accurately reflect the core content of the course through the analysis of the course syllabus and teaching content, difficulty level, using a five-level scoring system, simple, relatively simple, medium, relatively difficult, difficult for evaluation according to the knowledge depth, skill requirements, and learning time factors of the course, applicable scenarios, clarifying whether the course is applicable to beginners, advanced learners, or senior professionals at different levels of students. For courses with VR / AR experiences, specifically mark their VR / AR application characteristics, applicable learning fields, and required hardware devices; S5. VR / AR course adaptation For course fields with strong practicality, give priority to recommending courses with VR / AR simulation experiences. Based on the results of the learning habit analysis model, construct a course adaptation degree evaluation function. Let the learning style vector of the student be S, the learning progress be P, the knowledge mastery degree be K, the difficulty level of the course be D, the applicable scenario vector be A, and the VR / AR application characteristic vector be V. Then the course adaptation degree F is: F = w1·cosine(S,A) + w2·(P - D) + w3·(K - D) + w4·cosine(V,S), where w1, w2, w3, and w4 are weights. Sort the courses according to the course adaptation degree F and recommend the VR / AR courses with the highest adaptation degree to students; S6. Course push Based on the analysis results of students' learning habits, course classification and annotation information, as well as the VR / AR course live matching situation, the hybrid intelligent recommendation algorithm that combines collaborative filtering and content-based recommendation algorithm is used. In the collaborative filtering part, by calculating the similarity between students, neighbor students with similar interests to the target student are found. The cosine similarity algorithm is used to calculate the similarity sim(i,j) between student i and student j: where x i k and x j k are the learning behavior feature vectors of student i and student j for course k respectively, and m is the number of courses. According to the course preferences of neighbor students, courses that they like but the target student has not learned are recommended to the target student. In the content-based recommendation part, according to the annotation information of the courses, the similarity between courses is calculated. The TF-IDF algorithm is used to process the course keywords to obtain the text feature vectors of the courses, and then the similarity between courses is calculated through cosine similarity. Let the text feature vectors of course a and course b be T a and T b , then the course similarity According to the characteristics of the courses that the target student has learned, courses similar to them are recommended. Finally, the collaborative filtering and content-based recommendation results are fused, and the proportion of the two recommendation results is adjusted through weights. Let the weight of the collaborative filtering recommendation result be The weight of the content-based recommendation result is w n , and w c +w n = 1. The values of w c and w n are determined through experiments and cross-validation to obtain the final course recommendation list. In the pushed course list, the courses with VR / AR experience are highlighted, and detailed course introductions, VR / AR experience highlights, learning objectives, and required time information are provided; S7. Feedback and optimization Establish a perfect learner feedback mechanism. Through various channels such as online questionnaires, course evaluation systems, and customer service feedback, collect learners' evaluations of the pushed courses and feedback on learning effects. The evaluation content includes the practicality of the course content, the effectiveness of teaching methods, the quality of VR / AR experiences, and the matching degree with their own learning needs. The learning effect feedback includes the improvement of learning achievements, the mastery of knowledge and skills, and the change of learning interest. Analyze the feedback data, extract key information, and optimize the learning habit analysis model, course adaptability evaluation function, and hybrid intelligent recommendation algorithm according to the feedback information. For the learning habit analysis model, retrain the model based on the new feedback data, adjust the model parameters, and improve the prediction accuracy of the model for learners' learning habits. For the course adaptability evaluation function, adjust the weights w1, w2, w3, w4 according to learners' feedback to make the course adaptability more accurately reflect learners' needs. For the hybrid intelligent recommendation algorithm, adjust the fusion weights w c and w n and w to optimize the accuracy and diversity of the recommendation results.
2. The course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, wherein During the process of collecting learning habit data, for the collection of learning community interaction data, it also includes analyzing the resources published by students in the learning community, extracting the theme, knowledge point coverage, and quality evaluation information of the resources, and incorporating this information into the student learning habit dataset to more comprehensively reflect the learning behavior and knowledge contribution of students. Through natural language processing techniques, lexical analysis, syntactic analysis, and semantic understanding are performed on the text content published by students to extract key information and enrich the student learning habit portrait.
3. The course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, characterized in that In the cleaning step of data preprocessing, the Local Outlier Factor (LOF) algorithm is also introduced to detect and process outliers in the data. For each sample point in the dataset, its LOF value is calculated. If the LOF value is greater than the set threshold, the sample point is determined to be an outlier. The calculation formula for the LOF value is as follows: where p is the sample point, and N k (p) is the set of k-nearest neighbors of p, and reach-dist k (o, p) is the reachable distance from point o to point p, and k-dist(p) is the k-distance of point p.
4. The course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, wherein In the construction of the learning habit analysis model, the activation function of the hidden layer adopts the RelU function --- Adaptive RelU (AReLU). The traditional RelU function outputs 0 when the input is less than 0, which may lead to the problem of neuron death. The AReLU function is defined as: where α is an adaptive parameter, which is dynamically adjusted according to the activation of neurons during the training process. When the neurons are continuously in the non-activated state in multiple training batches, the value of α is gradually decreased to increase the activation probability of the neurons. When the neurons are frequently activated, the value of α is increased to avoid over-activation.
5. A course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, characterized in that, During the process of course classification and annotation, for the extraction of course keywords, an automatic extraction method based on word vectors is also adopted. First, the pre-trained word vector model is used to convert each word in the course text into a vector representation, and then the similar word vectors are clustered into one category through a clustering algorithm. Representative words are selected from each category as keywords. At the same time, in combination with the teaching syllabus and knowledge point structure of the course, manual review and adjustment are performed on the automatically extracted keywords.
6. The course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, wherein, In the VR / AR course adaptation, the learning time preference and learning environment factors of the trainees are also introduced. Let the learning time preference vector of the trainees be T, the learning environment vector be E, the teaching time arrangement vector of the course be T c , and the learning environment requirement vector required by the course be E c . Then the course adaptation degree F is updated as: F = w1·cosine(S, A) + w2·(P - D) + w3·(K - D) + w4·cosine(V, S) + w5·cosine(T, T c ) + w6·cosine(E, E e ), where w5 and w6 are the newly introduced weights, which are determined by analyzing the learning time and environment preference data of the trainees and combining expert experience, and w1 + w2 + w3 + w4 + w5 + w6 = 1.
7. A course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, characterized in that, In the hybrid intelligent recommendation algorithm for course push, a recommendation strategy adjustment mechanism based on reinforcement learning is introduced. The recommendation process is regarded as a Markov decision process, where the state space is the learning habit characteristics of students and the information of the courses they have learned, the action space is the list of recommended courses, and the reward function is defined according to the feedback of students on the recommended courses. Through the reinforcement learning algorithm, continuously learn the optimal recommendation strategy and dynamically adjust the fusion weights w c and w n , as well as the sorting method of recommended courses, to adapt to the needs and learning scenarios of different students.
8. A course push method for intelligent analysis of the learning habits of continuing education students according to claim 1, characterized in that, During the process of feedback and optimization, the course classification and annotation system is also updated regularly. With the development of the education field and the application of new technologies, the course content and teaching objectives will change continuously. Therefore, it is necessary to regularly organize experts to re-evaluate and annotate the courses. At the same time, new or adjusted course classifications are added according to the feedback from students and market demands.