Method, system and device for recommending personalized learning path in education
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG GONGSHANG UNIVERSITY
- Filing Date
- 2026-04-17
- Publication Date
- 2026-08-07
AI Technical Summary
[0004]本发明针对现有个性化学习路径推荐在单一模型框架下无法兼顾多任务、不同类型数据融合困难及难以满足实时要求的缺点，提供了一种教育个性化学习路径推荐方法、系统及装置
通过本发明的方法解决了现有技术无法对多维数据进行精准建模的问题，并且支持知识点诊断及资源推荐两类差异化任务，避免采用多个模型分别进行训练带来的高开发与维护成本，通过对不同维度下的向量切片进行计算训练，降低大规模资源库场景下的学习路径推荐的计算延迟，实现实时的个性化学习路径推荐。
Smart Images
Figure CN122529264A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart education technology, specifically to a method, system, and device for recommending personalized learning paths in education. Background Technology
[0002] Existing personalized recommendation methods for smart education mainly include: 1. Recommendation methods based on fixed-dimensional vectors. These methods use vectors of the same dimension to represent students and learning resources, and perform similarity matching recommendations through matrix factorization or deep learning based on vectors; 2. Student modeling methods based on a single data source. These methods mainly use students' answer records to model their knowledge mastery status, and evaluate students' knowledge mastery by obtaining indicators such as answer accuracy, ignoring other dimensions of students' online learning data, such as video viewing behavior and discussion participation; 3. Recommendation methods based on full-scale high-dimensional vector matching. These methods perform full-scale similarity calculations on all knowledge points and student status.
[0003] Based on this, the existing methods have the following drawbacks: 1. Fixed-dimensional recommendation methods cannot adapt to the differentiated needs of multiple tasks. For example, knowledge point diagnosis tasks and resource recommendation tasks have different requirements for vector accuracy and computational efficiency, and need to be oriented towards candidate sets of different sizes. Currently, fixed-dimensional recommendation methods cannot take both types of tasks into account under a single model framework, which increases development and maintenance costs; 2. Due to the significant differences in information density between different types of data, it is difficult for existing single-dimensional recommendation methods to achieve the fusion of multi-dimensional heterogeneous learning behavior data. Unified-dimensional encoding methods, through simple concatenation, lead to over-parameterization of simple data and insufficient representation of complex data, and cannot extract the value of multi-dimensional learning behavior data; 3. The inference latency of recommendation methods that match full high-dimensional vectors increases linearly with the scale of resources, making it difficult to meet the response requirements of real-time recommendations in large-scale scenarios. Summary of the Invention
[0004] This invention addresses the shortcomings of existing personalized learning path recommendation methods, such as the inability to handle multiple tasks, difficulties in integrating different types of data, and the inability to meet real-time requirements within a single model framework. It provides a method, system, and apparatus for recommending personalized learning paths in education.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0006] A method for recommending personalized learning paths in education includes the following steps: Learning behavior data is acquired and features are extracted to obtain learning behavior features. These learning behavior features are then fused and preprocessed to obtain student nested vectors. Acquire knowledge point data and perform feature extraction and feature concatenation to obtain knowledge point nested vectors. Based on student nested vectors and knowledge point nested vectors, obtain the set of students' weak knowledge points. The course resource vector is obtained through the course content. Based on the learning behavior data, positive and negative course samples of students are constructed. Combining the course resource vector, student nested vector, and knowledge point nested vector, knowledge mastery alignment function and resource ranking loss function of different dimensions are constructed. Recommended resource sequence is obtained through training. The recommended resource sequence is filtered based on student nesting vectors and course resource vectors to obtain course recommendation results; Based on the course recommendations and the students' weak knowledge points, a learning path recommendation is generated for each student.
[0007] As one possible implementation, the process of acquiring learning behavior data and extracting features to obtain learning behavior features, and then fusing and preprocessing these features to obtain a student nested vector, includes the following steps: Learning behavior features are fused based on attention weights to obtain fused behavior features, which include answer record features, video viewing features, and discussion participation features. The student nested vector is obtained by preprocessing the fusion behavior features, where the preprocessing includes data normalization.
[0008] As one possible implementation, the process of acquiring knowledge point data, performing feature extraction and feature concatenation to obtain a knowledge point nesting vector, and obtaining a set of students' weak knowledge points based on the student nesting vector and the knowledge point nesting vector includes the following steps: Acquire knowledge point data and perform feature encoding to obtain an initial knowledge point vector, wherein the knowledge point data includes the knowledge point name and a brief description of the knowledge point; Based on the knowledge point data, feature extraction is performed to obtain knowledge point features, which include chapter affiliation features, difficulty level features, and knowledge point sequence number features. The initial knowledge point vector and knowledge point features are concatenated to form a nested knowledge point vector. Obtain the similarity between the student nested vector and the knowledge point nested vector, and use the similarity to obtain the set of students' weak knowledge points.
[0009] As one possible implementation, the steps of acquiring learning behavior data and extracting features to obtain learning behavior features, fusing and preprocessing the learning behavior features to obtain student nested vectors, and acquiring knowledge point data and extracting and concatenating features to obtain knowledge point nested vectors include: Obtain students' answer accuracy rate, and construct positive and negative answer samples based on preset accuracy thresholds, answer accuracy rates, and corresponding knowledge points; A multi-scale loss function with different dimensions is constructed. Based on slices of positive answer samples, negative answer samples, student nested vectors of different dimensions, and knowledge point nested vectors, training is performed using the multi-scale loss function. The multi-scale loss function with different dimensions is as follows:
[0010] in, express Multi-scale loss function of dimension This represents the number of positive responses. This represents the number of negative samples in the responses. This indicates a positive sample of the answer. This represents a negative sample of the answers. This indicates the first positive or negative sample of answers. Student nested vectors of each student Dimensional slicing This indicates the first positive or negative sample of answers. Nested vectors of knowledge points Dimensional slicing.
[0011] As one possible implementation, the method involves obtaining a course resource vector from the course content, constructing positive and negative course samples for students based on learning behavior data, and combining the course resource vector with student nested vectors and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions of different dimensions. The method then trains these functions to obtain a recommended resource sequence, including the following steps: The system obtains students' correct answer rates for knowledge points, and calculates target similarity and training weights. It also obtains course resource vectors based on course content, which include course video data, course exercise data, and course handout data. Based on students' course resource vectors, we construct knowledge mastery alignment functions in different dimensions by slicing student nested vectors and knowledge point nested vectors under different dimensions, using target similarity, training weights, and student nested vectors and knowledge point nested vectors under different dimensions. Based on students' learning behavior data, construct positive and negative samples for the course, and then construct positive and negative sample vectors based on the positive and negative samples respectively. Based on slices of positive sample vectors, negative sample vectors, course resource vectors, and student nested vectors in different dimensions, resource ranking loss functions of different dimensions are constructed, and recommended resource sequences are obtained through training.
[0012] As one possible implementation, the target similarity and training weights are:
[0013]
[0014] The alignment function for knowledge mastery across different dimensions is as follows:
[0015] The resource ranking loss functions for different dimensions are:
[0016] in, Indicates the first The student and the first Target similarity between knowledge points Indicates the first The student on the first The accuracy rate of answering questions on each knowledge point Indicates the first The student and the first Training weights between knowledge points Indicates the exponential parameter. express The knowledge mastery alignment function of the dimension Represents a set of students. Represents a set of course resource vectors. Indicates the first Each course resource vector in Dimensional slices, Indicates the first Student nested vectors of each student Dimensional slicing This represents the resource ranking loss function. Indicates the first A positive sample of students' courses. Indicates the first A negative sample of students' courses. This represents the first positive sample of the course. A number of positive sample vectors Indicates the first negative sample of the course One negative sample vector, This represents the similarity calculation function.
[0017] As one possible implementation, the process of filtering the recommended resource sequence based on student nested vectors and course resource vectors to obtain course recommendation results includes the following steps: Obtain course resource vectors of different dimensions in the recommended resource sequence, and obtain resource scores of the corresponding dimensions based on the similarity between course resource vectors of different dimensions and student nested vectors of the corresponding dimensions. The course recommendation results are obtained by sorting the course resource vectors of different dimensions in the recommended resource list according to their corresponding resource ratings.
[0018] A personalized learning path recommendation system for education includes: The student vector construction module acquires learning behavior data and extracts features to obtain learning behavior features. It then fuses and preprocesses these learning behavior features to obtain nested student vectors. The weak knowledge point acquisition module acquires knowledge point data and performs feature extraction and feature concatenation to obtain a knowledge point nested vector. Based on the student nested vector and the knowledge point nested vector, the student's set of weak knowledge points is obtained. The resource sequence acquisition module obtains course resource vectors through course content, constructs positive and negative course samples for students based on learning behavior data, and combines course resource vectors, student nested vectors, and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions in different dimensions, and obtains recommended resource sequences through training. The course recommendation module filters the recommended resource sequence based on student nested vectors and course resource vectors to obtain course recommendation results; The learning path recommendation module generates learning path recommendations for students based on course recommendations and the students' weak knowledge points.
[0019] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any one of the following ways: Learning behavior data is acquired and features are extracted to obtain learning behavior features. These learning behavior features are then fused and preprocessed to obtain student nested vectors. Acquire knowledge point data and perform feature extraction and feature concatenation to obtain knowledge point nested vectors. Based on student nested vectors and knowledge point nested vectors, obtain the set of students' weak knowledge points. The course resource vector is obtained through the course content. Based on the learning behavior data, positive and negative course samples of students are constructed. Combining the course resource vector, student nested vector, and knowledge point nested vector, knowledge mastery alignment function and resource ranking loss function of different dimensions are constructed. Recommended resource sequence is obtained through training. The recommended resource sequence is filtered based on student nesting vectors and course resource vectors to obtain course recommendation results; Based on the course recommendations and the students' weak knowledge points, a learning path recommendation is generated for each student.
[0020] A personalized learning path recommendation device includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the method described in any one of the following ways: Learning behavior data is acquired and features are extracted to obtain learning behavior features. These learning behavior features are then fused and preprocessed to obtain student nested vectors. Acquire knowledge point data and perform feature extraction and feature concatenation to obtain knowledge point nested vectors. Based on student nested vectors and knowledge point nested vectors, obtain the set of students' weak knowledge points. The course resource vector is obtained through the course content. Based on the learning behavior data, positive and negative course samples of students are constructed. Combining the course resource vector, student nested vector, and knowledge point nested vector, knowledge mastery alignment function and resource ranking loss function of different dimensions are constructed. Recommended resource sequence is obtained through training. The recommended resource sequence is filtered based on student nesting vectors and course resource vectors to obtain course recommendation results; Based on the course recommendations and the students' weak knowledge points, a learning path recommendation is generated for each student.
[0021] This invention, by adopting the above technical solutions, has significant technical effects: The method of this invention solves the problem that existing technologies cannot accurately model multidimensional data, and supports two differentiated tasks: knowledge point diagnosis and resource recommendation. It avoids the high development and maintenance costs caused by training multiple models separately. By calculating and training vector slices under different dimensions, it reduces the computational latency of learning path recommendation in large-scale resource library scenarios and realizes real-time personalized learning path recommendation. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a schematic diagram of the overall system of the present invention. Detailed Implementation
[0024] The present invention will be further described in detail below with reference to the embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.
[0025] Example 1: A method for recommending personalized learning paths in education, such as Figure 1 As shown, it includes the following steps: S100. Obtain learning behavior data and extract features to obtain learning behavior features. Then, fuse and preprocess the learning behavior features to obtain student nested vectors. S200: Obtain knowledge point data and perform feature extraction and feature concatenation to obtain knowledge point nesting vectors. Based on student nesting vectors and knowledge point nesting vectors, obtain the set of students' weak knowledge points. S300. Obtain course resource vectors through course content, construct positive and negative course samples for students based on learning behavior data, combine course resource vectors with student nested vectors and knowledge point nested vectors, construct knowledge mastery alignment functions and resource ranking loss functions of different dimensions, and obtain recommended resource sequences through training. S400. Based on the student nested vector and the course resource vector, the recommended resource sequence is filtered to obtain the course recommendation results; S500: Based on the course recommendation results and the students' weak knowledge points, a learning path recommendation result is formed for the students.
[0026] In this application, multidimensional heterogeneous learning behavior data is acquired, and different types of data are fully mined and fused to solve the problem that the information density of different types of data is significantly different and it is difficult to perform feature fusion. Since the knowledge point diagnosis task is oriented towards a small-scale candidate set and the resource recommendation task is oriented towards a large-scale candidate set, this application takes into account both knowledge point diagnosis and resource recommendation tasks in the process of recommending learning paths, avoiding the training of multiple independent models, and thus obtaining personalized learning path recommendation results.
[0027] S100. Obtain learning behavior data and extract features to obtain learning behavior features. Fuse and preprocess the learning behavior features to obtain student nested vectors. This includes the following steps: Step S1001: In this embodiment, students' learning behavior data is collected based on the MOOC platform. The learning behavior data includes answer records, video viewing data, discussion participation data, and lecture note download data. Step S1002: Extract features from the learning behavior data to obtain learning behavior features, including answer record features, video viewing features, and discussion participation features. In this embodiment, the answer record features include answer accuracy rate, average answer time, number of answers, temporal variation features, variation trend, total number of answers, total accuracy rate, total answer time, and average answer speed. These features are then projected onto a unified 2048-dimensional fusion space through independent linear transformations and activation functions. In this embodiment, the answer record features are taken as a 365-dimensional feature vector. The answer record features are as follows:
[0028] Step S1003: Extract features from the video viewing data to obtain video control features. In this embodiment, the video viewing features include total video viewing time, average completion rate, and number of views. Taking a 245-dimensional feature vector as an example, the video viewing features are projected onto a unified 2048-dimensional fusion space through independent linear transformations and activation functions. The video viewing features are as follows:
[0029] Step 1004: In this embodiment, the GPT-5 large language model is used to automatically analyze each student's discussion participation data, annotate the relevant knowledge points, and score them from five dimensions: content quality, depth of understanding, clarity of expression, initiative, and collaboration. A 475-dimensional feature vector is extracted, including the number of questions asked, the number of answers, the content quality score, the depth of understanding score, global statistical features, and the global average quality features. These features are then projected onto a unified 2048-dimensional fusion space through independent linear transformations and activation functions. The discussion participation features are as follows:
[0030] Step S1005: After projecting the three types of features into the same space, this embodiment uses learnable attention weights for adaptive fusion to obtain a fused feature vector, which is then normalized to obtain a student nested vector. The fused feature vector is:
[0031] In this embodiment, the normalization method is L2 normalization, and the student nested vector is:
[0032] in, , , These represent the mapped features of the answer record, video viewing, and discussion participation, respectively. , , These represent the corresponding weight matrices. , , This indicates the corresponding bias term. , , These represent the corresponding attention weights. Represents the fused feature vector. This represents a nested vector of students.
[0033] S200. Obtain knowledge point data and perform feature extraction and feature concatenation to obtain a knowledge point nested vector. Based on the student nested vector and the knowledge point nested vector, obtain the set of students' weak knowledge points, including the following steps: Step S2001: In this embodiment, the text-embedding-3-large model is used to encode the knowledge point name and description. Combined with the structured features of the knowledge point's chapter affiliation, difficulty level, and sequence number, the data is concatenated to obtain a concatenated knowledge point vector, which is then mapped to a 2048-dimensional nested vector space. The mapping process is as follows:
[0034]
[0035] Step S2002: Perform L2 normalization on the mapped vector to obtain the nested knowledge point vector, as shown below:
[0036] Step S2003: In this embodiment, the set of students' weak knowledge points is obtained by using the inner product similarity between the student nested vector and the knowledge point nested vector. The inner product is equivalent to the cosine similarity. The higher the inner product score, the better the student's mastery of the current knowledge point. The knowledge points with lower scores are selected as the set of students' weak knowledge points. The inner product similarity is:
[0037] in, Represents hidden layer features. , Represents the weight matrix. , Indicates the bias term. This represents the mapped knowledge point vector. Represents a nested vector of knowledge points. Indicates the first The student and the first Inner product similarity between knowledge points Indicates the first Student nesting vectors of each student. Indicates the first A knowledge point nested vector.
[0038] This embodiment also includes a process of training learnable parameters in the calculation of student nested vectors and knowledge point nested vectors. This process first obtains the student's answer accuracy rate and presets an accuracy rate threshold. If the student's answer accuracy rate is greater than the threshold, positive answer samples are formed based on the corresponding student nested vectors and knowledge point nested vectors; otherwise, negative answer samples are formed. To train parameters in different dimensions, this embodiment selects a dimension level. Multi-scale loss functions are constructed based on three different dimensions, where the multi-scale loss functions for each dimension are as follows:
[0039] For multi-scale loss functions with different dimensions, this embodiment obtains the total loss function through weighted calculation. Based on the constructed positive and negative answer samples, student nested vectors of corresponding dimensions, and slices of knowledge point vectors, training is performed using the multi-scale loss function. This ensures that the similarity between the student nested vectors and the positive answer samples is high, and the similarity between the student nested vectors and the negative answer samples is low. This yields learnable parameters in the nested vector calculation process, thereby achieving accurate diagnostic capabilities and flexible switching of nested dimensions. The learnable parameters include weight matrices and attention weights in the feature calculation process. The update process uses an end-to-end joint update via the Adam optimizer, with a learning rate of 0.001, a batch size of 256, a maximum training epoch count of 100, and an early stopping strategy. The total loss function is:
[0040] in, express Multi-scale loss function of dimension This represents the number of positive responses. This represents the number of negative samples in the responses. This indicates a positive sample of the answer. This represents a negative sample of the answers. This indicates the first positive or negative sample of answers. Student nested vectors of each student Dimensional slicing This indicates the first positive or negative sample of answers. Nested vectors of knowledge points Dimensional slicing Represents the total loss function. , , Represents the weight parameters. , , These represent multi-scale loss functions with dimensions of 128, 512, and 2048, respectively.
[0041] S300. Obtain course resource vectors from course content, construct positive and negative course samples for students based on learning behavior data, and combine course resource vectors, student nested vectors, and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions of different dimensions. Then, obtain recommended resource sequences through training, including the following steps: Step S3001: Based on the students' correct answer rate for the knowledge points, calculate the target similarity and training weight. This process ensures that weak knowledge points receive high target similarity and high training weight, while mastered knowledge points receive lower target similarity and lower training weight, thus prioritizing the recommendation of weak knowledge points. The target similarity and training weight are as follows:
[0042]
[0043] Step S3002: Construct a course resource vector based on the course content, wherein the course resource vector includes course video data, course exercise data, and course handout data; Step S3003: Based on the student's course resource vector, construct knowledge mastery alignment functions of different dimensions through target similarity, training weights, and slicing of student nested vectors and knowledge point nested vectors under different dimensions. The knowledge mastery alignment functions of different dimensions are as follows:
[0044] Step S3004: In this embodiment, the student's learning behavior data also includes lecture note download data. Based on the student's learning behavior data, positive samples and negative samples of the course are constructed. For example, if a student's video viewing time exceeds 60% of the total learning video time, it is marked as a positive sample; if a student answers a practice question incorrectly, it is marked as a positive sample; if a lecture note is successfully downloaded, it is marked as a positive sample; otherwise, it is marked as a negative sample. Positive samples of the course are formed based on the positive sample data, and negative samples of the course are formed based on the negative sample data. Positive sample vectors and negative sample vectors are formed through the positive samples and negative samples of the course, respectively. Step S3005: Based on slices of positive sample vectors, negative sample vectors, course resource vectors, and student nested vectors at different dimensions, construct resource ranking loss functions of different dimensions. The resource ranking loss functions of different dimensions are as follows:
[0045] Step S3006: Based on the knowledge mastery alignment function, resource ranking loss function and corresponding weight parameters of different dimensions, obtain the total knowledge mastery alignment function and the total resource ranking loss function. Train based on the loss function to obtain the recommended resource sequence formed by knowledge points. in, Indicates the first The student and the first Target similarity between knowledge points Indicates the first The student on the first The accuracy rate of answering questions on each knowledge point Indicates the first The student and the first Training weights between knowledge points Indicates the exponential parameter. express The knowledge mastery alignment function of the dimension Represents a set of students. Represents a set of course resource vectors. Indicates the first Each course resource vector in Dimensional slices, Indicates the first Student nested vectors of each student Dimensional slicing This represents the resource ranking loss function. Indicates the first A positive sample of students' courses. Indicates the first A negative sample of students' courses. This represents the first positive sample of the course. A number of positive sample vectors Indicates the first negative sample of the course One negative sample vector, This represents the similarity calculation function.
[0046] S400. Based on the student nested vector and course resource vector, the recommended resource sequence is filtered to obtain course recommendation results. In this embodiment, in order to obtain accurate course recommendation results, a three-stage progressive architecture is adopted for filtering, including three stages: initial ranking, intermediate ranking, and fine ranking. The filtering of high-dimensional vector slice results is gradually narrowed down to a smaller candidate set. This process includes the following steps: Step S4001: Initial ranking stage. In this embodiment, student nested vectors and course resource vectors are sliced in 128 dimensions to calculate initial similarity. Based on the initial similarity, the recommended resource sequence is filtered to obtain an initial candidate set. The initial similarity is:
[0047] Step S4002: In the intermediate ranking stage, based on the slices of student nested vectors and course resource vectors in 512 dimensions, calculate the intermediate similarity. Based on the intermediate similarity, filter from the initial candidate set to obtain the intermediate candidate set. The intermediate similarity is:
[0048] Step S4003: In the fine-tuning stage, based on the slices of student nested vectors and course resource vectors in the 2048-dimensional dimension, high-level similarity is calculated. Based on the intermediate similarity, the initial candidate set is filtered to obtain the course recommendation results. The high-level similarity is:
[0049] In this embodiment, a three-stage sorting architecture is used to gradually narrow down the candidate set. The total computational cost of the three-stage sorting is:
[0050] In the formula for calculating the total amount of calculation, when At this time, the main computational load is the computational load of the low-dimensional initial arrangement, compared to the computational load of traditional high-dimensional matching. The total computational cost of the three-stage sorting in this application is approximately equal to the computational cost of high-dimensional matching. In other words, the method in this embodiment achieves a theoretical speedup of 16 times, and the high-dimensional ranking only applies to the selected candidate set, giving the system sublinear computational growth characteristics.
[0051] This embodiment differentiates student behavior data through encoding and designs adapted networks for three types of data with different information densities: answer record features, video viewing features, and discussion participation features. Combined with a learnable attention fusion mechanism, this allows multi-dimensional data to fully realize their respective values. Compared to baseline methods that only apply answer data, this embodiment improves the F1 score of the weak knowledge point set by 9.5%, a relative improvement of 15.2%. Furthermore, this method supports flexible switching between multiple dimensions, significantly reducing system development and maintenance costs. A two-stage training strategy is constructed: the first stage prioritizes matching resource vectors to students' weak knowledge points, and the second stage aligns with students' actual learning behavior preferences. A three-stage progressive filtering mechanism is used to adjust the resource library size as needed. The computational cost of the architecture is reduced by more than 90% while maintaining a high level of recommendation quality. This resolves the contradiction between the difficulty of achieving both recommendation quality and computational efficiency, enabling the recommendation results to effectively fill knowledge gaps and conform to students' personalized learning preferences, thereby helping to improve students' learning motivation and learning outcomes.
[0052] Example 2: A personalized learning path recommendation system for education, such as Figure 2 As shown, it includes: The student vector construction module 100 acquires learning behavior data and extracts features to obtain learning behavior features. It then fuses and preprocesses the learning behavior features to obtain nested student vectors. The weak knowledge point acquisition module 200 acquires knowledge point data and performs feature extraction and feature concatenation to obtain a knowledge point nested vector. Based on the student nested vector and the knowledge point nested vector, the student's set of weak knowledge points is obtained. The resource sequence acquisition module 300 obtains course resource vectors through course content, constructs positive and negative course samples for students based on learning behavior data, and combines course resource vectors, student nested vectors, and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions in different dimensions, and obtains recommended resource sequences through training. The course recommendation module 400 filters the recommended resource sequence based on student nested vectors and course resource vectors to obtain course recommendation results; The learning path recommendation module 500 generates learning path recommendations for students based on course recommendation results and the students' weak knowledge points.
[0053] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.
[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0055] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0058] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0059] It should be noted that: The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.
[0060] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.
Claims
1. A method for recommending personalized learning paths in education, characterized in that, Includes the following steps: Learning behavior data is acquired and features are extracted to obtain learning behavior features. These learning behavior features are then fused and preprocessed to obtain student nested vectors. Acquire knowledge point data and perform feature extraction and feature concatenation to obtain knowledge point nested vectors. Based on student nested vectors and knowledge point nested vectors, obtain the set of students' weak knowledge points. The course resource vector is obtained through the course content. Based on the learning behavior data, positive and negative course samples of students are constructed. Combining the course resource vector, student nested vector, and knowledge point nested vector, knowledge mastery alignment function and resource ranking loss function of different dimensions are constructed. Recommended resource sequence is obtained through training. The recommended resource sequence is filtered based on student nesting vectors and course resource vectors to obtain course recommendation results; Based on the course recommendations and the students' weak knowledge points, a learning path recommendation is generated for each student.
2. The personalized learning path recommendation method according to claim 1, characterized in that, The process of acquiring learning behavior data and extracting features to obtain learning behavior features, and then fusing and preprocessing these features to obtain student nested vectors, includes the following steps: Learning behavior features are fused based on attention weights to obtain fused behavior features, which include answer record features, video viewing features, and discussion participation features. The student nested vector is obtained by preprocessing the fusion behavior features, where the preprocessing includes data normalization.
3. The personalized learning path recommendation method according to claim 1, characterized in that, The process of acquiring knowledge point data, extracting and concatenating features to obtain a nested knowledge point vector, and then obtaining a set of students' weak knowledge points based on the student nested vector and the knowledge point nested vector includes the following steps: Acquire knowledge point data and perform feature encoding to obtain an initial knowledge point vector, wherein the knowledge point data includes the knowledge point name and a brief description of the knowledge point; Based on the knowledge point data, feature extraction is performed to obtain knowledge point features, which include chapter affiliation features, difficulty level features, and knowledge point sequence number features. The initial knowledge point vector and knowledge point features are concatenated to form a nested knowledge point vector. Obtain the similarity between the student nested vector and the knowledge point nested vector, and use the similarity to obtain the set of students' weak knowledge points.
4. The personalized learning path recommendation method according to claim 1, characterized in that, The steps of acquiring learning behavior data and extracting features to obtain learning behavior features, fusing and preprocessing the learning behavior features to obtain student nested vectors, and acquiring knowledge point data and extracting and concatenating features to obtain knowledge point nested vectors include: Obtain students' answer accuracy rate, and construct positive and negative answer samples based on preset accuracy thresholds, answer accuracy rates, and corresponding knowledge points; A multi-scale loss function with different dimensions is constructed. Based on slices of positive answer samples, negative answer samples, student nested vectors of different dimensions, and knowledge point nested vectors, training is performed using the multi-scale loss function. The multi-scale loss function with different dimensions is as follows: in, express Multi-scale loss function of dimension This represents the number of positive responses. This represents the number of negative samples in the responses. This indicates a positive sample of the answer. This represents a negative sample of the answers. This indicates the first positive or negative sample of answers. Student nested vectors of each student Dimensional slicing This indicates the first positive or negative sample of answers. Nested vectors of knowledge points Dimensional slicing.
5. The personalized learning path recommendation method according to claim 1, characterized in that, The process involves obtaining course resource vectors from course content, constructing positive and negative course samples for students based on learning behavior data, and combining course resource vectors with student nested vectors and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions of different dimensions. Recommended resource sequences are then obtained through training, including the following steps: The system obtains students' correct answer rates for knowledge points, and calculates target similarity and training weights. It also obtains course resource vectors based on course content, which include course video data, course exercise data, and course handout data. Based on students' course resource vectors, we construct knowledge mastery alignment functions in different dimensions by slicing student nested vectors and knowledge point nested vectors under different dimensions, using target similarity, training weights, and student nested vectors and knowledge point nested vectors under different dimensions. Based on students' learning behavior data, construct positive and negative samples for the course, and then construct positive and negative sample vectors based on the positive and negative samples respectively. Based on slices of positive sample vectors, negative sample vectors, course resource vectors, and student nested vectors in different dimensions, resource ranking loss functions of different dimensions are constructed, and recommended resource sequences are obtained through training.
6. The personalized learning path recommendation method according to claim 5, characterized in that, The target similarity and training weights are as follows: The alignment function for knowledge mastery across different dimensions is as follows: The resource ranking loss functions for different dimensions are: in, Indicates the first The student and the first Target similarity between knowledge points Indicates the first The student on the first The accuracy rate of answering questions on each knowledge point Indicates the first The student and the first Training weights between knowledge points Indicates the exponential parameter. express The knowledge mastery alignment function of the dimension Represents a set of students. Represents a set of course resource vectors. Indicates the first Each course resource vector in Dimensional slices, Indicates the first Student nested vectors of each student Dimensional slicing This represents the resource ranking loss function. Indicates the first A positive sample of students' courses. Indicates the first A negative sample of students' courses. This represents the first positive sample of the course. A number of positive sample vectors Indicates the first negative sample of the course One negative sample vector, This represents the similarity calculation function.
7. The personalized learning path recommendation method according to claim 1, characterized in that, The process of filtering the recommended resource sequence based on student nested vectors and course resource vectors to obtain course recommendation results includes the following steps: Obtain course resource vectors of different dimensions in the recommended resource sequence, and obtain resource scores of the corresponding dimensions based on the similarity between course resource vectors of different dimensions and student nested vectors of the corresponding dimensions. The course recommendation results are obtained by sorting the course resource vectors of different dimensions in the recommended resource list according to their corresponding resource ratings.
8. A personalized learning path recommendation system for education, characterized in that, include: The student vector construction module acquires learning behavior data and extracts features to obtain learning behavior features. It then fuses and preprocesses these learning behavior features to obtain nested student vectors. The weak knowledge point acquisition module acquires knowledge point data and performs feature extraction and feature concatenation to obtain a knowledge point nested vector. Based on the student nested vector and the knowledge point nested vector, the student's set of weak knowledge points is obtained. The resource sequence acquisition module obtains course resource vectors through course content, constructs positive and negative course samples for students based on learning behavior data, and combines course resource vectors, student nested vectors, and knowledge point nested vectors to construct knowledge mastery alignment functions and resource ranking loss functions in different dimensions, and obtains recommended resource sequences through training. The course recommendation module filters the recommended resource sequence based on student nested vectors and course resource vectors to obtain course recommendation results; The learning path recommendation module generates learning path recommendations for students based on course recommendations and the students' weak knowledge points.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.
10. A personalized learning path recommendation device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.