Personalized learning path recommendation method and system based on lightweight collaborative filtering

Abstract

The present application relates to the technical field of wisdom education, in particular to a personalized learning path recommendation method and system based on lightweight collaborative filtering, comprising a model definition module, a data processing module, an embedding initialization module, a model training module and a recommendation generation module; the model definition module defines a lightweight deep learning collaborative filtering model, fuses student features and knowledge graph information, and recommends a learning path; the data processing module loads and preprocesses student data and knowledge graph data; the embedding initialization module initializes knowledge point embedding; the model training module is used for training the model; the recommendation generation module generates a personalized learning path for the student and provides learning effect evaluation; the recommended path generated by the system not only meets the personal preferences and abilities of the student, but also ensures the logical coherence and learnability between knowledge points, improves the recommendation quality, and can make reasonable recommendations based on the semantic information of the knowledge points in the case of sparse student interaction data.

Description

Technical Field

[0001] This invention relates to the field of smart education technology, and in particular to a method and system for recommending personalized learning paths based on lightweight collaborative filtering. Background Technology

[0002] The innovative application of technologies such as big data and artificial intelligence in education is driving the development of educational informatization towards personalization, intelligence, and refinement. Integrating knowledge graph technology into the university teaching process is widely considered an important direction for the informatization and intelligentization of education. How to recommend the optimal learning path (i.e., the order of learning knowledge points) based on students' individual abilities, interests, and learning history is key to improving learning efficiency and effectiveness. Existing learning path recommendation technologies mainly suffer from the following shortcomings: Traditional collaborative filtering methods mainly rely on student-knowledge point interaction matrices (such as answer records), lacking modeling of complex semantic relationships between knowledge points (such as prerequisite relationships and difficulty progression), resulting in inconsistent recommendation paths in terms of knowledge logic.

[0003] Rule-based or simple graph-based methods: Although knowledge graphs are introduced to structure knowledge, the recommendation logic often relies on predefined hard rules (such as learning A before learning B), lacking in-depth mining of students' personalized characteristics and adaptive matching capabilities, resulting in poor flexibility.

[0004] Existing deep learning methods: Some studies have attempted to use deep learning models, but these models typically have a large number of parameters, high training costs, and fail to effectively integrate structured information from knowledge graphs with personalized information from student profiles, resulting in limited recommendation accuracy and interpretability.

[0005] While some existing technology literature utilizes knowledge graphs or large language models to achieve personalized recommendations, their adaptability in the education field remains significantly insufficient. For example, the invention patent CN202511714266.7 discloses an API intelligent recommendation method and system based on semantic knowledge graphs, realizing microservice / API recommendation scenarios. However, it does not consider core requirements such as the hierarchical association of knowledge points in the education field and the adaptation to students' learning progress, making it unsuitable for direct transfer to learning path recommendation. The invention patent CN202511563249.8 discloses a personalized recommendation method and system based on the collaboration of large language models and domain models, achieving accurate adaptation of real-time user needs and personalized recommendations across scenarios. However, the model architecture is complex, and it lacks knowledge point embedding optimization and learning effect evaluation mechanisms designed for education scenarios, making it difficult to balance recommendation accuracy and inference efficiency, and also unable to dynamically match changes in students' learning status.

[0006] In summary, existing technologies for personalized learning path recommendation based on knowledge graphs in the education field generally suffer from poor scenario adaptability, an imbalance between model efficiency and accuracy, and a lack of closed-loop evaluation of learning outcomes, making it difficult to meet the differentiated learning needs in educational scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a personalized learning path recommendation method and system based on lightweight collaborative filtering, which solves the problems of poor scenario adaptability, imbalance between model efficiency and accuracy, and lack of closed-loop evaluation of learning effect in traditional learning path recommendation technology.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a personalized learning path recommendation system based on lightweight collaborative filtering, including a model definition module, a data processing module, an embedding initialization module, a model training module, and a recommendation generation module; The model definition module is used to define a lightweight deep learning collaborative filtering model that integrates student features and knowledge graph information to achieve learning path recommendation. The data processing module is used to load and preprocess student data and knowledge graph data, providing structured input for model training; The embedding initialization module initializes the knowledge point embeddings based on the pre-trained language model, thereby enhancing the model's knowledge expression capabilities. The model training module is used for training the model; The recommendation generation module generates personalized learning paths for students based on the trained model and provides learning performance evaluation.

[0009] The lightweight deep learning collaborative filtering model includes a student feature encoding layer, a knowledge graph feature fusion layer, and a recommendation prediction layer. The student feature encoding layer uses a fully connected neural network to encode students' learning history, academic performance, and interest preferences, outputting a student feature vector. The knowledge graph feature fusion layer uses a lightweight graph attention mechanism to extract the features and relationships of knowledge point nodes, fusing the knowledge graph information with the student feature vector to output a fused feature vector. The recommendation prediction layer uses the fused feature vector and a softmax function to predict the student's learning fit for each knowledge point.

[0010] The student data includes student interests, learning behavior data, and academic performance data. The student data preprocessing process includes data cleaning, feature normalization, and feature encoding. The knowledge graph data includes knowledge point nodes and knowledge point relationships. The knowledge graph data preprocessing process includes constructing a knowledge graph adjacency matrix and standardizing knowledge point attributes.

[0011] The model training process includes data partitioning, loss calculation, parameter optimization, and early stopping mechanism.

[0012] Secondly, the present invention also provides a personalized learning path recommendation method based on lightweight collaborative filtering, applied to the personalized learning path recommendation system based on lightweight collaborative filtering as described in the first aspect above, comprising the following steps: The structured data output from the S1 data processing module is input into the embedding initialization module and the model definition module, respectively. The embedding initialization module in S2 inputs the generated knowledge point embedding matrix into the model definition module. After the model definition module is trained by the model training module, it outputs the optimal model to the recommendation generation module. The S3 electric recommendation generation module generates learning paths and evaluates their effectiveness, then feeds the evaluation results back to the model training module to achieve dynamic optimization of model parameters.

[0013] This invention discloses a personalized learning path recommendation system based on lightweight collaborative filtering. The system comprises a model definition module that defines a lightweight deep learning collaborative filtering model, integrating student features and knowledge graph information to recommend learning paths. A data processing module loads and preprocesses student and knowledge graph data, providing structured input for model training. An embedding initialization module initializes knowledge point embeddings based on a pre-trained language model, enhancing the model's knowledge representation capabilities. A model training module trains the model. A recommendation generation module generates personalized learning paths for students based on the trained model and provides learning performance evaluation. This system proposes a lightweight collaborative filtering model that integrates student features and knowledge graph information, simplifying the model structure while ensuring recommendation accuracy and improving inference efficiency. It uses a pre-trained language model to initialize knowledge point embeddings, enhancing the model's knowledge representation and association capture capabilities. It integrates learning performance evaluation and dynamic adjustment. The system integrates various mechanisms to achieve real-time optimization of personalized learning paths. A fully automated data processing module provides high-quality, structured input for model training, improving training efficiency and stability. Through a knowledge-aware fusion mechanism, the generated recommended paths not only align with students' individual preferences and abilities but also ensure logical coherence and learnability between knowledge points, enhancing recommendation quality. Pre-trained language models are used to initialize knowledge point embeddings, enabling the model to make reasonable recommendations based on semantic information of knowledge points even when student interaction data is sparse. Since the recommendation results explicitly link to relationships in the knowledge graph, it can explain to students that "we recommend learning knowledge point B because you have already mastered its prerequisite knowledge point A, and B is related to your area of ​​interest C," improving system credibility and user experience. This addresses the problems of poor scenario adaptability, imbalance between model efficiency and accuracy, and lack of closed-loop evaluation of learning outcomes inherent in traditional learning path recommendation technologies. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0016] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the strain data, acceleration data, displacement data, pressure data, and video data involved in this application were all obtained with full authorization.

[0017] Figure 1 This is the overall architecture diagram of the personalized path recommendation learning model of this invention. It shows the connection relationship of the model's five core modules (data processing module, embedding initialization module, model definition module, model training module, and recommendation generation module). Each module works in sequence and collaboratively to realize the entire process from data input to path recommendation and effect evaluation. The structured data output from the data processing module is input to the embedding initialization module and the model definition module respectively. The knowledge point embedding matrix generated by the embedding initialization module is input to the model definition module. After being trained by the model training module, the model definition module outputs the optimal model to the recommendation generation module. The recommendation generation module generates the learning path and performs effect evaluation. The evaluation results can be fed back to the model training module to realize dynamic optimization of model parameters.

[0018] Figure 2This is a flowchart illustrating the personalized learning path generation and effectiveness evaluation process. First, student data and knowledge graph data are input. The model calculates the student's mastery probability score for all knowledge points, sorts them by score, and selects high-scoring knowledge points as the starting nodes for the learning paths. Then, utilizing the knowledge graph structure information, a greedy algorithm expands the path along the edges connecting knowledge points until a predetermined path length is reached, generating multiple personalized learning paths. After the student learns according to the recommended path, relevant learning data is collected for multi-dimensional effectiveness evaluation, including grade improvement rate (calculated as a percentage improvement compared to historical grades, reflecting learning progress), learning ability rating (divided into five levels based on the most recent grade: excellent, good, average, pass, and need improvement, visually displaying the current learning level), and learning satisfaction rating (considering the student's acceptance of the learning process and learning experience, evaluating path suitability). After evaluation, if the result meets a preset threshold, the next stage of learning path is recommended; otherwise, the path is readjusted and fed back to the model, achieving dynamic path optimization and model iteration.

[0019] Figure 3 This is a schematic diagram of the internal structure of the Model Definition Module (LightweightCF). The LightweightCF model includes a student embedding layer, a knowledge point embedding layer, a student feature processing layer, a graph attention layer, and a prediction layer. The student embedding layer maps student IDs to a low-dimensional embedding space, capturing students' potential learning preference patterns. The knowledge point embedding layer embeds knowledge points, facilitating the model's understanding and processing of relationships between them. The student feature processing layer extracts and transforms students' multimodal features using a multilayer perceptron, ensuring consistency with the output dimension of the embedding layer. The graph attention layer uses a graph attention mechanism to model the relationships between knowledge points in the knowledge graph, enhancing the semantic information of the knowledge point embeddings. The prediction layer concatenates the student embedding, student feature embedding, and knowledge point embedding, and then uses a multilayer neural network to predict the student's mastery of the knowledge points.

[0020] Figure 4This is a data processing flowchart, which includes: student data loading and cleaning, knowledge graph data loading and filtering, interaction record construction, student feature encoding, knowledge point encoding, and knowledge graph construction. First, student information, including student ID, learning style, interests, and historical grades, is read from an Excel file, and missing and outlier values ​​are processed. Next, the node and edge information of the knowledge graph is read, and knowledge points related to the target subject (software design patterns) are filtered out. Then, positive sample interaction records (student-knowledge point mastery relationship) are generated based on the knowledge points the student has mastered, and negative sample interaction records are generated by sampling knowledge points the student has not mastered. One-hot encoding is performed on the student's non-numerical features (such as learning style), and these are merged with numerical features (such as historical grades) to construct a student feature vector. Finally, knowledge points are encoded, and edge indices of the knowledge graph are constructed for input to the graph attention network.

[0021] Figure 5 This is a schematic diagram of model training.

[0022] Figure 6 This is a schematic diagram of the personalized learning path recommendation system based on lightweight collaborative filtering provided by the present invention.

[0023] Figure 7 This is a flowchart of the personalized learning path recommendation method based on lightweight collaborative filtering provided by the present invention.

[0024] In the diagram: 1-Model definition module, 2-Data processing module, 3-Embedding initialization module, 4-Model training module, 5-Recommendation generation module. Detailed Implementation

[0025] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0026] Please see Figures 1-7 Firstly, the personalized learning path recommendation system based on lightweight collaborative filtering provided by this invention includes a model definition module 1, a data processing module 2, an embedding initialization module 3, a model training module 4, and a recommendation generation module 5. The model definition module 1 is used to define a lightweight deep learning collaborative filtering model that integrates student features and knowledge graph information to achieve learning path recommendation. The data processing module 2 is used to load and preprocess student data and knowledge graph data, providing structured input for model training; The embedding initialization module 3 initializes the knowledge point embedding based on the pre-trained language model, thereby enhancing the model's knowledge expression capability. The model training module 4 is used for training the model; The recommendation generation module 5 generates personalized learning paths for students based on the trained model and provides learning performance evaluation.

[0027] In this embodiment of the invention, the model definition module 1 defines a lightweight deep learning collaborative filtering model that integrates student features and knowledge graph information to achieve learning path recommendation. The data processing module 2 loads and preprocesses student data and knowledge graph data, providing structured input for model training. The embedding initialization module 3 initializes knowledge point embeddings based on a pre-trained language model, enhancing the model's knowledge expression capabilities. The model training module 4 is used for model training. The recommendation generation module 5 generates personalized learning paths for students based on the trained model and provides learning effect evaluation. This system proposes a lightweight collaborative filtering model that integrates student features and knowledge graph information, simplifying the model structure while ensuring recommendation accuracy and improving reasoning efficiency. It uses a pre-trained language model to initialize knowledge point embeddings, enhancing the model's knowledge expression and association capture capabilities. It integrates learning effect evaluation and dynamic adjustment mechanisms to achieve… Real-time optimization of personalized learning paths; the construction of a fully automated data processing module 2 to provide high-quality, structured input for model training, improving model training efficiency and stability; through a knowledge-aware fusion mechanism, the generated recommended paths not only conform to students' personal preferences and abilities but also ensure the logical coherence and learnability between knowledge points, improving recommendation quality; the use of pre-trained language models to initialize knowledge point embeddings enables the model to make reasonable recommendations based on the semantic information of knowledge points even when student interaction data is sparse; since the recommendation results are explicitly linked to the relationships in the knowledge graph, it can explain to students that "the recommended learning knowledge point B is because you have already mastered its prerequisite knowledge point A, and B is related to your interest area C", improving the credibility of the system and user experience, thereby solving the problems of poor scenario adaptability, imbalance between model efficiency and accuracy, and lack of closed-loop evaluation of learning effects in traditional learning path recommendation technologies.

[0028] Furthermore, the lightweight deep learning collaborative filtering model includes a student feature encoding layer, a knowledge graph feature fusion layer, and a recommendation prediction layer. The student feature encoding layer uses a fully connected neural network to encode students' learning history, academic performance, and interest preferences, outputting student feature vectors with dimensions of 64-128. The knowledge graph feature fusion layer extracts the features and relationships of knowledge point nodes through a lightweight graph attention mechanism, fusing knowledge graph information with student feature vectors to output a fused feature vector. The recommendation prediction layer, based on the fused feature vector, uses a softmax function to predict the student's learning fit for each knowledge point. The number of model parameters is controlled within 1 million to ensure lightweightness and efficiency.

[0029] Furthermore, the student data includes student interests and hobbies, learning behavior data (such as study time, answer accuracy, knowledge point access records), and academic performance data. The preprocessing process includes data cleaning (removing missing values ​​and outliers), feature normalization (mapping continuous features such as grades and study time to the [0,1] interval), and feature encoding (one-hot encoding of categorical features such as gender and major). The knowledge graph data includes knowledge point nodes (including knowledge point ID, name, difficulty coefficient, and chapter), and knowledge point relationships (such as prerequisite knowledge points, similar knowledge points, and derived knowledge points). The preprocessing process includes constructing a knowledge graph adjacency matrix and standardizing knowledge point attributes. Finally, a structured student feature dataset and knowledge graph dataset are output, providing standardized input for model training.

[0030] Furthermore, the embedding initialization module 3 uses a BERT pre-trained language model to encode the textual descriptions of knowledge points, generating initial knowledge point embedding vectors. It extracts the vectors corresponding to the [CLS] tags output by BERT as the overall semantic representation of the knowledge points, and truncates and normalizes them to adapt to the model's embedding dimension requirements. The initialized embedding vectors are assigned to the knowledge point embedding layer, and these embedding vectors are fine-tuned during model training to better adapt to the current learning path recommendation task.

[0031] Furthermore, the model training module 4 includes data partitioning, loss calculation, parameter optimization, and an early stopping mechanism. A binary cross-entropy loss function (BCELoss) is used to measure the difference between the model's predicted probability of students mastering knowledge points and the true labels. The AdamW optimizer is selected, and the model parameters are updated by combining the learning rate (lr) and weight decay (weight_decay) parameters to minimize the loss function value. During training, when the validation set loss no longer decreases for 15 consecutive rounds, training is stopped, and the best-performing model parameters on the validation set are loaded to prevent the model from overfitting the training data.

[0032] Furthermore, the recommendation generation process of the recommendation generation module 5 is as follows: Inputting student characteristics and knowledge graph data, the model outputs a learning fit score for each knowledge point. Combining the relationships between knowledge points with the student's current learning progress, a greedy algorithm expands the learning path along the edges connecting knowledge points until a predetermined path length is reached, generating multiple personalized learning paths for each student. The learning effectiveness evaluation mechanism collects real-time data on student learning path completion, knowledge point test scores, and learning time to construct multi-dimensional evaluation indicators (mastery, efficiency, fit), and dynamically adjusts the learning path based on the evaluation results to adapt to changes in the student's learning status.

[0033] To better understand this technical solution, the following embodiments are provided for further explanation: Example: Taking the teaching scenario of the course "Software Testing Technology" as an example. I. System Deployment and Data Preparation: This system is deployed on a local server, developed using Python, and based on the PyTorch deep learning framework and the NetworkX graph processing library. The core tools and parameters required for implementation are as follows: Hardware environment: Equipped with NVIDIA GeForce RTX 3060 (12GB GDDR6 VRAM) / RTX 4060 (8GB / 16GB GDDR6X VRAM), 16GB / 32GB DDR4 / DDR5 memory, and 512GB NVMe SSD.

[0034] Software environment: Python 3.8, PyTorch 1.12, Transformers 4.18.

[0035] Initial data: A CSV file containing the historical learning records of 80 software engineering students; a JSON-formatted knowledge graph file describing software testing knowledge points and their relationships (containing core knowledge point entities such as "unit testing", "integration testing", "white-box testing", "black-box testing", and "test case design", as well as relational edges such as "prerequisites", "inclusion", and "association").

[0036] II. Data Processing Flow (in conjunction with...) Figure 4 ) Student data loading and cleaning: Read the IDs, historical test scores, learning interests, and learning preference tags (such as "theory-oriented" and "practice-oriented") of 80 students from the student_records.csv file. Fill in missing score data with the mean and perform one-hot encoding on non-numerical features such as learning preferences.

[0037] Knowledge graph data loading and filtering: Load all knowledge point nodes and relationship edges in the testing_knowledge_graph.json file, and filter out the core knowledge point set directly related to the "Software Testing Technology" course, such as filtering out non-core testing concepts such as "Agile Development".

[0038] Interaction record construction: Based on the chapter quizzes and experimental projects that students have passed, positive sample interaction records are generated, for example (Student_01, Mastery, Unit Test). At the same time, negative sample interaction records are generated by randomly sampling knowledge points that have never appeared in the student's history.

[0039] Student feature coding and knowledge point coding: The cleaned student numerical features (such as average scores) are concatenated with the one-hot encoded categorical features to form an 80x64-dimensional student feature matrix. A unique ID is assigned to each knowledge point, and its text description is prepared (e.g., black-box testing: a testing method based on specifications that does not focus on internal logic).

[0040] Knowledge Graph Construction: Using the NetworkX library, a directed graph structure for the software testing knowledge graph is constructed based on the loaded node and edge information. Nodes in the graph represent knowledge points, and edges represent relationships such as "prerequisite". This graph structure will be used in the subsequent graph attention layer.

[0041] III. Model Building and Training (Combined) Figure 3 , Figure 5 ) Pre-trained embedding initialization: A pre-trained language model (BERT) is used as the semantic encoder. The text description of each knowledge point is input into the model, and the hidden state marked with [CLS] is taken as the 768-dimensional initial embedding vector of that knowledge point to form a knowledge point embedding matrix.

[0042] Model training strategy: Student embedding layer: Maps 80 student IDs to dense vectors of 80x32 dimensions. Knowledge point embedding layer receives a 768-dimensional knowledge point vector generated by embedding initialization module 3 and projects it to 32 dimensions through a linear layer to align with the student embedding dimensions.

[0043] Student Feature Processing Layer: The original 64-dimensional student feature vector is transformed and output as a refined 32-dimensional student feature embedding through a multilayer perceptron (MLP) containing one hidden layer (128 neurons).

[0044] Graph Attention Layer: This is the core layer that distinguishes it from existing technologies. This layer receives knowledge point embeddings and edge indices from the knowledge graph. It employs a single-head graph attention network (GAT) to calculate the attention coefficients between each knowledge point and its first-order neighbors in the knowledge graph, and aggregates neighbor information to output the knowledge point context embedding after knowledge structure enhancement. This process ensures that the embedding of the "integration test" node includes the semantic information of its predecessor node, the "unit test".

[0045] Prediction layer: The student embedding and the feature embedding output from the student feature processing layer are concatenated with the enhanced embedding of the target knowledge point to obtain a 96-dimensional fusion vector. This vector is then passed through a fully connected layer containing the `ReLU` activation function to output a scalar score, which is the predicted mastery probability of the student for this knowledge point.

[0046] Model Training Module 4: The Binary Cross-Entropy Loss (BCELoss) function is used to measure the difference between the model's predicted probability of students mastering knowledge points and the true labels. The Adam optimizer is used, with an initial learning rate of 0.001 and a batch size of 32. During training, the loss is calculated on the validation set in each epoch. Training stops when the validation set loss no longer decreases after 15 consecutive epochs, and the parameters of the best-performing model on the validation set are loaded to prevent the model from overfitting the training data.

[0047] IV. Personalized Path Generation and Evaluation (combined with...) Figure 1 , Figure 2 ) Recommended to generate module 5 and execute as follows Figure 2 The process shown is as follows: For the target student (student_01), the trained LightweightCF model calculates the probability score of their mastery of all candidate software test knowledge points.

[0048] Select the knowledge point with the highest score (such as "boundary value analysis") as the starting node of the learning path.

[0049] Combining the structural constraints of the knowledge graph, a greedy algorithm is used for path expansion. The algorithm prioritizes knowledge points that have a "prerequisite" or "strong association" relationship with the end node of the current path and have high model prediction scores as the next node. For example, it expands from "boundary value analysis" to "equivalence class partitioning" because they are both key methods under the "test case design" node.

[0050] Repeat step three until an ordered learning path containing 5 knowledge points is generated, for example: [Boundary Value Analysis → Equivalence Class Partitioning → Scenario Graph Method → ​​Black-box Testing → System Testing].

[0051] Multi-dimensional effect evaluation: After the student (student_01) learns according to the recommended path above, the system collects data such as the subsequent improvement rate of grades, learning ability rating, and learning satisfaction score to conduct a multi-dimensional evaluation: Performance Improvement Rate: Comparing the scores in the "Test Case Design" related tests before and after the learning recommended path, the improvement rate was calculated to be 15%.

[0052] Learning ability rating: Based on the most recent comprehensive test score, the system automatically rates it as "good".

[0053] Learning satisfaction rating: Students rated the logicality of the learning path and the suitability of the difficulty level through a post-class questionnaire. The average score was 4.5 / 5.0.

[0054] Feedback and Optimization: The above evaluation data is summarized into new interaction records and fed back to model training module 4. The system periodically (e.g., every two weeks) uses the accumulated new data to incrementally train the LightweightCF model, realizing dynamic optimization of model parameters and continuous iteration of the learning path recommendation strategy.

[0055] Implementation Results: The system successfully generated differentiated software testing learning paths for 50 students. Practice shows that the group of students using the system's recommended paths demonstrated significantly higher test case design completeness and defect discovery rates in their final practical projects compared to the group following a fixed syllabus. This verifies the effectiveness of this invention in addressing the issues of fragmented and unsuitable software testing teaching resources. Its closed-loop framework of "data processing - knowledge enhancement modeling - path generation - evaluation and feedback" can provide a reference for personalized teaching reforms in computer science or software engineering courses, and has significant and far-reaching implications for promoting innovation in higher education teaching models and improving the quality of talent cultivation.

[0056] The above-disclosed embodiments are merely preferred embodiments of the personalized learning path recommendation method and system based on lightweight collaborative filtering in this application, and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A personalized learning path recommendation system based on lightweight collaborative filtering, characterized in that, It includes a model definition module, a data processing module, an embedding initialization module, a model training module, and a recommendation generation module; The model definition module is used to define a lightweight deep learning collaborative filtering model that integrates student features and knowledge graph information to achieve learning path recommendation. The data processing module is used to load and preprocess student data and knowledge graph data, providing structured input for model training; The embedding initialization module initializes the knowledge point embeddings based on the pre-trained language model, thereby enhancing the model's knowledge expression capabilities. The model training module is used for training the model; The recommendation generation module generates personalized learning paths for students based on the trained model and provides learning performance evaluation.

2. The personalized learning path recommendation system based on lightweight collaborative filtering as described in claim 1, characterized in that, The lightweight deep learning collaborative filtering model includes a student feature encoding layer, a knowledge graph feature fusion layer, and a recommendation prediction layer. The student feature encoding layer uses a fully connected neural network to encode students' learning history, academic performance, and interest preferences, outputting a student feature vector. The knowledge graph feature fusion layer uses a lightweight graph attention mechanism to extract the features and relationships of knowledge point nodes, fusing the knowledge graph information with the student feature vector to output a fused feature vector. The recommendation prediction layer, based on the fused feature vector, uses a softmax function to predict the student's learning fit for each knowledge point.

3. The personalized learning path recommendation system based on lightweight collaborative filtering as described in claim 1, characterized in that, The student data includes student interests, learning behavior data, and academic performance data. The student data preprocessing process includes data cleaning, feature normalization, and feature encoding. The knowledge graph data includes knowledge point nodes and knowledge point relationships. The knowledge graph data preprocessing process includes constructing a knowledge graph adjacency matrix and standardizing knowledge point attributes.

4. The personalized learning path recommendation system based on lightweight collaborative filtering as described in claim 1, characterized in that, The model training process includes data partitioning, loss calculation, parameter optimization, and early stopping mechanism.

5. A personalized learning path recommendation method based on lightweight collaborative filtering, applied to the personalized learning path recommendation system based on lightweight collaborative filtering as described in any one of claims 1-4, characterized in that, Includes the following steps: The structured data output from the data processing module is input into the embedding initialization module and the model definition module, respectively. The embedding initialization module inputs the generated knowledge point embedding matrix into the model definition module. After the model definition module is trained by the model training module, it outputs the optimal model to the recommendation generation module. The electric recommendation generation module generates learning paths and evaluates their effectiveness, then feeds the evaluation results back to the model training module to achieve dynamic optimization of model parameters.

Images