Learning resource recommendation method based on knowledge graph and learner portrait

Abstract

The invention discloses a learning resource recommendation method based on a knowledge graph and a learner portrait, effectively alleviates the information overload problem caused by massive resources, constructs a learning content portrait and a learner portrait, introduces an education knowledge graph as auxiliary information, and utilizes a preference propagation thought to mine the potential preference of a learner, thereby improving the learning resource recommendation efficiency. The problem of data sparsity is relieved; the learning content portrait, the learner portrait and the knowledge graph preference are fused, the deep belief network is adopted to perform feature learning, and the preference of the learner to the non-interactive learning resources is predicted, so that more accurate, dynamic and comprehensive personalized learning resource recommendation is provided for the learner.

Description

Technical Field

[0001] This invention belongs to the field of recommendation technology, and specifically relates to a learning resource recommendation method based on knowledge graphs and learner profiles. Background Technology

[0002] Knowledge graph is an important branch of artificial intelligence technology. It was proposed by Google in 2012. It is a structured semantic knowledge base used to describe concepts and their relationships in the physical world in symbolic form. Its basic building blocks are the "entity-relationship-entity" triple and the entity and its related attribute-value pairs. Entities are interconnected through relations to form a network of knowledge structures.

[0003] Learner profiling refers to the process of analyzing and describing a specific group of learners from multiple dimensions to better understand their needs, habits, backgrounds, goals, and other characteristics, thereby providing a basis for the design and implementation of education and training. With the development of educational technology, learner profiling is increasingly widely used in various training and educational environments, especially in areas such as corporate training, online education, and personalized learning, becoming an important means of improving learning effectiveness and efficiency. In today's highly competitive business environment, the importance of corporate training is becoming increasingly prominent.

[0004] In today's information age, the quantity and diversity of learning resources are abundant. Learners face countless choices. However, this also raises a question: how to find the most suitable resources from such a vast pool? This necessitates a personalized recommendation system to help users find the most appropriate resources.

[0005] Traditional recommendation algorithms, such as collaborative filtering, often perform poorly when faced with the sparsity of learner-learning resource interaction data prevalent in large-scale online education platforms. These methods often struggle to accurately depict learners' true mastery of a particular knowledge domain and their potential learning interests, and they also fail to effectively utilize the complex semantic network relationships between learning resources.

[0006] While progress has been made in building learner profiles, challenges remain in handling large-scale, multi-dimensional, and heterogeneous learning process data. Traditional profile building methods may suffer from insufficient utilization of raw data features, limited dimensionality, and an inability to comprehensively characterize learner abilities, particularly in capturing learners' deep implicit preferences.

[0007] Furthermore, even with the introduction of educational knowledge graphs, most solutions have failed to fully leverage their advantages in knowledge association, preference propagation, and semantic understanding, making it difficult to achieve accurate and dynamic planning and recommendation of learning paths. Summary of the Invention

[0008] To address the problems existing in current technologies, this invention constructs learning content profiles and learner profiles, and introduces educational knowledge graphs as auxiliary information. It utilizes the concept of preference propagation to mine learners' potential preferences, thus alleviating the data sparsity problem. Furthermore, it integrates the learning content profiles, learner profiles, and knowledge graph preferences, employing deep belief networks for feature learning to predict learners' preferences for non-interactive learning resources. This provides learners with more accurate, dynamic, and comprehensive personalized learning resource recommendations, specifically including the following steps:

[0009] S1. Multi-dimensional profile construction: Based on the large-scale practice-oriented online learning dataset MOOPer, we construct a learning content profile with three dimensions: level quality, level difficulty, and knowledge point coverage, as well as a learner profile with three dimensions: learning interest, learning behavior, and learning status.

[0010] In step S1, the process of constructing the learning content profile is as follows:

[0011] S11. Conduct sentiment analysis on learners' comments on levels and integrate them with learners' ratings of levels to construct level quality indicators.

[0012] S12. Select the three-dimensional features of the average score of the level, the average practice time of the learner, and the average number of times the learner views the answer, and use the K-Means clustering method and the TOPS IS entropy weight method to quantify the level difficulty index.

[0013] S13. Use the TF-IDF method to extract keywords from course names, classify courses by type, and calculate the percentage overlap between the knowledge points contained in a level and the total knowledge points of its course category to obtain the knowledge point coverage index. The TF-IDF formula is as follows:

[0014]

[0015] Where ft,d represents the frequency of term t in document d, measuring the importance of term t in a single document; N is the total number of levels; n t Let be the number of levels containing term t. The process of constructing the learner profile is as follows:

[0016] S14. Extract learning characteristics of learners’ learning interests, learning behaviors and learning status from learners’ online learning data;

[0017] S15. The learned features are standardized using the z-score standardization method, and then reconstructed using an autoencoder. The z-score standardization formula is as follows:

[0018]

[0019] Where Xi represents the original data points before standardization, Mean represents the mean of the original data set, and Std represents the standard deviation of the original data set.

[0020] S16. The reconstructed 3D learner features are clustered using the quadratic K-Means clustering method to obtain a 3D profile that reflects the learner's characteristics.

[0021] S2. Knowledge Graph Construction: Construct an educational knowledge graph for personalized recommendations, including knowledge unit nodes such as courses, exercises, levels, and knowledge points, as well as four semantic relationships—inclusion, dependency, relevance, and mapping—existing between level nodes. This includes the following steps:

[0022] S21. Extract course, level, exercise and knowledge point information;

[0023] S22. Using courses, exercises, levels, and knowledge points as knowledge unit nodes, construct a knowledge graph that starts with courses and radiates outward along the path of "course-exercise-level-knowledge point".

[0024] S23. Construct four semantic relationships—inclusion, dependency, relevance, and mapping—between the level nodes. These relationships are built based on the level names and question descriptions to describe the specific connections between different levels. The formula for determining relevance is as follows:

[0025] Sim(A, B) = ω1Sim name (A, B) + ω2Sim content (A, B) ≥ T corr

[0026] Among them, Sim name (A, B) represent name similarity; Sim content (A, B) represent content similarity; ω1 and ω2 are the weights of name similarity and content similarity; Vec name and Vec content The result of vectorizing the level names and question descriptions in S1 using TF-IDF; T corr This is a preset relevance similarity threshold.

[0027] The formula for the overlap ratio when constructing containment relationships is as follows:

[0028]

[0029] Among them, K A K B These are the keyword sets for level A and level B extracted in S1, respectively.

[0030] S3. Latent Preference Learning: Based on the knowledge graph, the learner's preference information for already interacted levels is propagated to uninteracted level nodes using the preference propagation concept, in order to mine the learner's latent preferences and thus alleviate the data sparsity problem. Step S3 is as follows:

[0031] S31. Based on the learner-level interaction records, obtain the learner's explicit preference information for the levels already interacted with, as the initial information for preference propagation;

[0032] S32. Combining the knowledge graph, the initial preference information is propagated along the semantic relationships (inclusion, dependence, relevance, and mapping) between level nodes to non-interactive level nodes through the preference propagation idea, thereby inferring the learner's potential preferences for these non-interactive levels.

[0033] S33. The potential preference information obtained after the preference propagation is fused with the explicit preference information in the learner profile to form the final learner preference model.

[0034] S4. Personalized learning resource recommendation: Integrating the learning content profile, learner profile, and knowledge graph preferences, a deep belief network (DBN) is used for feature learning to predict learners' preferences for non-interactive learning resources, and a personalized learning resource recommendation list is generated based on the predicted scores. The process of step S4 is as follows:

[0035] S41. The learning content profile, learner profile, and learner-level interaction records after preference propagation are fused to form the input feature matrix of a deep belief network (DBN).

[0036] S42. A deep belief network is used to perform unsupervised pre-training and supervised fine-tuning on the input feature matrix. In the unsupervised pre-training stage, a layer-by-layer greedy training method is used to train the Restricted Boltzmann Machine (RBM) to learn the feature representation of the data. In the supervised fine-tuning stage, the learned feature representation is input into a Softmax classifier for training.

[0037] S43. The output of the deep belief network is used as the learner's preference score for non-interactive learning resources, and a personalized list of learning resources is dynamically generated and recommended to the learner based on the preference score.

[0038] Beneficial effects:

[0039] This invention introduces an educational knowledge graph and employs the concept of preference propagation to effectively propagate learners' preference information from a small number of interacted levels to more uninteracted levels through the rich semantic relationships between levels in the knowledge graph. This significantly expands the coverage and depth of learners' preference information, compensating for the information shortages in traditional recommendation systems when facing the sparsity and cold start scenarios prevalent in large-scale online learning data, thereby improving the robustness and coverage of the recommendation system.

[0040] This invention constructs learner profiles from three dimensions: learning interest, learning behavior, and learning status. At the same time, it constructs learning content profiles from three dimensions: level quality, level difficulty, and knowledge point coverage. This breaks through the single-dimensional limitations of traditional profiles. This dynamic updating and refined characterization based on dual profiles lays a solid foundation for providing highly accurate personalized recommendations.

[0041] At the recommendation method level, this invention proposes a personalized recommendation framework based on deep belief networks (DBNs), which deeply integrates learning content profiles, learner profiles, and knowledge graph preference propagation results. DBNs utilize multi-layer restricted Boltzmann machines for layer-by-layer feature extraction and nonlinear transformation, effectively handling high-dimensional sparse data and discovering deep distribution characteristics of the data.

[0042] Through the aforementioned technical means, this invention significantly improves the personalization and adaptability of learning, enabling learners to obtain a more precise and satisfying learning experience. In practical applications, this invention can effectively improve academic performance, completion rates, and learning efficiency, representing the cutting-edge innovative possibilities of intelligent education technology. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the overall recommendation process of a learning resource recommendation method based on knowledge graphs and learner profiles according to the present invention.

[0044] Figure 2 This is a diagram illustrating the content profile structure of a learning resource recommendation method based on knowledge graphs and learner profiles.

[0045] Figure 3 This is a learner profile structure diagram for a learning resource recommendation method based on knowledge graphs and learner profiles.

[0046] Figure 4 A flowchart illustrating the knowledge graph construction process for a learning resource recommendation method based on knowledge graphs and learner profiles. Detailed Implementation

[0047] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0048] like Figure 1As shown, a learning resource recommendation method based on knowledge graphs and multi-dimensional profiles is proposed. This method aims to address challenges faced by learners in existing online education platforms, such as information overload, data sparsity and cold start problems in traditional recommendation algorithms, insufficient comprehensiveness and dynamism of learner profiles, and inadequate utilization of knowledge graphs. The goal is to provide learners with more accurate, dynamic, and comprehensive personalized learning resource recommendations. The overall recommendation flowchart includes the following steps: constructing learner profiles and learning content profiles, using Deep Belief Networks (DBNs) for feature learning, calculating learner level matching degrees, and generating a TOP-N recommendation set.

[0049] like Figure 2 As shown, the learning content profile is constructed based on user-practice interaction data and user review data from the MOOPer large-scale online open practice dataset. The learning content profile includes three dimensions: card quality, level difficulty, and knowledge point coverage.

[0050] Level Quality Metrics: A pre-trained Beet model is used to analyze user reviews and extract the average sentiment bias from them. Generally, a positive sentiment bias in user reviews indicates high level quality; conversely, a predominantly negative sentiment bias suggests relatively low quality. Simultaneously, the average user rating for each level is calculated as another important objective dimension for measuring level quality, forming a comprehensive quality metric together with the average sentiment bias in user reviews.

[0051] Level Difficulty Metrics: Level difficulty is quantified across three dimensions: average score, average practice time for learners, and whether users viewed answers, to construct a comprehensive level difficulty assessment system. After determining these three dimensions of level difficulty, K-Means clustering is performed on these three features, initially dividing the practice difficulty into four difficulty levels. Based on this, to obtain a more refined and comprehensive level difficulty metric, a comprehensive difficulty evaluation method based on the TOPSIS entropy weight method is further applied within each difficulty category to obtain the final level difficulty metric.

[0052] Knowledge point coverage metric: First, data cleaning and keyword extraction were performed. Considering the relatively small amount of text in the course names, only simple cleaning was needed, mainly removing punctuation marks and special characters to ensure text purity. Next, the TF-IDF algorithm was used to extract keywords from the course names. By analyzing the keywords of each course and combining them with the professional divisions within the field of computer science, ten course categories were manually set: Python, Java Programming, C & C++, Database, Big Data Processing, Algorithms, Web Development, Computer Fundamentals, Practical Exercises, and Other Application-Based Courses.

[0053] After categorizing the courses, we can further calculate the knowledge point coverage for each level. For a given level L, its course category is C. L Let T be the set of knowledge points contained in level L. L The level belongs to course category C. L The set of all knowledge points included is The knowledge point coverage (Rcoverage) of this level can then be defined as:

[0054]

[0055] in, This indicates that the knowledge points contained in level L belong to its course category C. L The number of knowledge points. Indicates course category C L The total number of all knowledge points included.

[0056] like Figure 3 As shown, the constructed learner profile includes three dimensions: learning interest, learning behavior, and learning status. These three dimensions are scored using an autoencoder-based and quadratic K-Means clustering method. The learning interest score is composed of the number of comments, the number of levels passed, and the average score; the learning behavior score is composed of the number of times answers were viewed and the number of levels retried; and the learning status score is composed of the average time spent, level pass rate, average score, and knowledge point mastery rate. The total score is calculated based on the overall learning characteristics.

[0057] To eliminate the influence of different dimensions and magnitudes in the learning feature data on the clustering results, and also to improve the convergence speed and reduce the training time during autoencoder training, this paper uses the z-score normalization method to normalize the learning features of each dimension before feeding the data into the autoencoder for feature reconstruction. The normalization formula is as follows.

[0058]

[0059] The standardized learned features are then fed into an autoencoder for reconstruction. The core idea is to encode the input data into a low-dimensional representation using a neural network, and then reconstruct the original data from this representation to obtain the key features. The reconstruction loss function is defined as follows:

[0060]

[0061] in, Both gθ(z) and gθ(z) are sigmoid functions. θ and θ are the parameters of the encoder and decoder, respectively, where n is the number of samples and x is the number of samples. i This is the input sample.

[0062] After the learners' 3D data is processed by the autoencoder, this section will use the quadratic K-Means clustering method to perform multi-dimensional clustering of learners' learning features. The quadratic K-Means clustering method has proven to be excellent in learner profiling. First, K-Means is used to perform the first clustering of learning features. The average score within each category is calculated and then sorted. Scores are assigned sequentially to obtain the first clustering score for each learner. The above steps are repeated for newly obtained categories to obtain the second clustering score for each learner. The first clustering score is multiplied by 10 and added to the second clustering score to obtain the learner's final score, resulting in a 3D profile that accurately and comprehensively reflects the learner's learning characteristics.

[0063] like Figure 4 As shown, this educational knowledge graph includes the import of knowledge entities and the import of semantic relationships. The knowledge graph radiates outwards along a hierarchical path of "course-exercise-level-knowledge point," and constructs four semantic relationships—inclusion, dependency, relevance, and mapping—between the level nodes. Specifically, this embodiment implements this step as follows:

[0064] First, nodes are imported and established. The identified and organized courses, exercises, levels, and knowledge points are imported as basic entities for the educational knowledge graph. Each entity contains a unique ID, name, and attributes such as keywords extracted from step S1. A natural hierarchical relationship exists between entities: each course contains multiple exercises, each exercise contains multiple levels, and each level focuses on testing one or more knowledge points. Furthermore, different levels sharing a common knowledge point or different courses sharing a single exercise enriches the relationships between entities. These entities will serve as the basic building blocks of the knowledge graph, providing the foundation for subsequent relationship construction. The next step is to construct semantic relationships between levels. Based on level names, question descriptions, and keywords, four semantic relationships are established between levels: inclusion, dependency, relevance, and mapping. These relationships aim to capture more complex connections between levels beyond their respective courses, thereby enriching the structure of the knowledge graph and enhancing its semantic representation capabilities.

[0065] After node import and the construction of four semantic relationships, all entities and relationships will be stored in the Neo4j graph database. Graph databases can efficiently store and query complex network structures, providing strong data support for subsequent preference propagation and recommendation algorithms.

[0066] Traditional recommender systems typically only capture learners' explicit preferences, i.e., the direct relationships between learners and already interacted learning resources. However, compared to explicit preferences, learners' latent needs and interests are often hidden within related resources of already interacted learning resources, and are of greater value to recommender systems. This paper analyzes learners' latent preferences using the idea of ​​preference propagation on a constructed educational knowledge graph.

[0067] Given a set of learners U = {u1, u2, ...} and a set of learning resources V = {v1, v2, ...}, construct a co-occurrence matrix Y = {y uv When learner u interacts with learning resource v, y |u∈U,v∈V} uv =1, otherwise y uv =0. For a learner u, extract all y values ​​of that learner in the co-occurrence matrix. uv Learning resources with a value of 1 represent the learner's historical interaction records, which are recorded as a set of positive feedback items. All entities within the knowledge graph are taken as starting nodes. Each node propagates to its nearest neighbors, with each propagation denoted as a hop. The set of entities corresponding to the k-th hop can then be represented as:

[0068]

[0069] in, Let (h, r, t) represent the knowledge graph constructed above, where (h, r, t) is a triple, h and t are the head and tail entities, and r represents the four semantic relations constructed in 4.2.2. From this, a new set can be derived:

[0070]

[0071] in, This represents the set of entities corresponding to all learning materials interacted with by learner u after the k-th hop. This represents all triples corresponding to the (k-1)th jump and the kth jump.

[0072] When k=0, entity This represents the learner's historical interaction records; when k > 0, This indicates the learner's latent preferences. Preference propagation is caused by... This is reflected in the fact that as k continuously increases, The scope spreads outwards.

[0073] Based on the constructed learning content profile and learner profile, a learner-level matrix can be obtained. This matrix reflects the relationship between learners' learning abilities and level features. Using these learners and level features as input to a learning resource recommendation algorithm can improve the quality of the recommendation results. A deep belief model is used to generate predicted level scores, enabling the recommendation system to make recommendations to learners based on these predicted scores.

[0074] The training process of DBN follows a greedy, layer-by-layer unsupervised learning mechanism. First, the feature matrix X is input into the first layer RBM for pre-training, learning the low-level feature representation h(1) of the data. The RBM is trained through backpropagation to obtain its weights w1 and bias b1. Then, matrix X is multiplied by w1 and the bias b1 is added. Finally, the result is passed through the sigmoid activation function to obtain the activation value a(1) of the first layer, as shown in the following formula.

[0075] a(1)=sigmoid(X*w1+b1)

[0076] The training of subsequent RBM layers is similar to that of the first layer. The feature representation a(1) obtained from the first layer is used as input and fed into the second RBM layer for the same pre-training process. The weights w2 and biases b2 are obtained through the backpropagation algorithm, and the activation value a(2) of the second layer is calculated. This layer-by-layer training process continues until the last layer L of the DBN network, thus obtaining the final high-level feature representation h(L), calculated as shown in the following formula.

[0077] a(L)=sigmoid(a(L-1)*w L +b L )

[0078] The final feature representation h(L) is then input into a Softmax classifier for supervised fine-tuning. This classifier multiplies h(L) by a weight matrix W and adds a bias vector b before feeding it into the Softmax function. This yields the probability that each sample belongs to each category, thus predicting learner preferences. Here, K represents the number of categories, as shown below.

[0079]

[0080] Finally, the recommendation set generated by DBN is sorted according to the similarity level, and personalized learning resources are recommended to learners.

[0081] The proposed method was tested on the MOOPe dataset. A collaborative filtering algorithm based on learned content profiling was used as the baseline model, and the proposed recommendation model based on deep belief networks was compared with it. The results are shown in Table 1. The proposed recommendation model based on deep belief networks combines learned content profiling, learner profiling, and knowledge graph preference propagation results as feature inputs to the DBN model, resulting in significant improvements in precision, recall, and F1 score.

[0082] With its powerful feature learning capabilities, DBN can efficiently process and integrate multi-dimensional feature combinations in this paper, and effectively handle the matrix sparsity problem caused by large-scale interactive data. Furthermore, experimental results further demonstrate the close correlation between the constructed profile features and knowledge graph information; they can effectively characterize the deep features of each entity in the recommendation system, providing high-quality input for the DBN model.

[0083] As shown in Table 2, in order to verify the contribution of knowledge graph preference propagation to improving the performance of DBN recommendation system, this study designed a set of ablation experiments. The ablation model removed the knowledge graph preference module and used only the original profile features of learners and learning content as input.

[0084] In terms of precision, recall, and F1 score, the complete model outperformed the ablation model, strongly demonstrating the importance of knowledge graph preference propagation in improving model performance. Specifically, knowledge graph preference propagation effectively expands the model's exploratory capabilities, helping it discover more resources that learners may be potentially interested in but have not directly interacted with. Simultaneously, it can more accurately identify learners' true interests, effectively filtering out irrelevant information and reducing unnecessary recommendations.

[0085] Table 1 Comparison of Model Results

[0086]

[0087] Table 2 Ablation Experiment Results:

[0088]

Claims

1. A learning resource recommendation method based on knowledge graphs and learner profiles, characterized in that, The method specifically includes the following steps: S1. Multi-dimensional profile construction: Based on the large-scale practice-oriented online learning dataset mOOPer, we construct learning content profiles with three dimensions: level quality, level difficulty, and knowledge point coverage, as well as learner profiles with three dimensions: learning interest, learning behavior, and learning status. S2. Knowledge Graph Construction: Construct an educational knowledge graph for personalized recommendations, which includes knowledge unit nodes such as courses, exercises, levels, and knowledge points, as well as four semantic relationships between level nodes: inclusion, dependency, correlation, and mapping. S3. Latent Preference Learning: Based on the knowledge graph, the learner's preference information for the already interacted levels is propagated to the uninteracted level nodes through the idea of ​​preference propagation, so as to explore the learner's latent preferences and thus alleviate the data sparsity problem. S4. Personalized learning resource recommendation: Integrating the learning content profile, learner profile, and knowledge graph preferences, a deep belief network is used for feature learning to predict learners' preferences for non-interactive learning resources, and a personalized learning resource recommendation list is generated based on the predicted scores.

2. The learning resource recommendation method based on knowledge graphs and learner profiles according to claim 1, characterized in that, In step S1, the process of constructing the learning content profile is as follows: S11. Conduct sentiment analysis on learners' comments on levels and integrate them with learners' ratings of levels to construct level quality indicators. S12. Select the three-dimensional features of the level's average score, the learner's average practice time, and the learner's average number of times to view the answer, and use the K-Means clustering method and the TOPSIS entropy weight method to quantify the level difficulty index; S13. Use the TF-IDF method to extract keywords from course names, classify courses by type, and calculate the overlap ratio between the knowledge points contained in a level and the total knowledge points of its course category to obtain the knowledge point coverage index. The TF-IDF formula is as follows: Among them, f t,d The frequency of term t in document d measures the importance of term t in a single document; N is the total number of levels; n t This represents the number of levels containing the term t.

3. The learning resource recommendation method based on knowledge graphs and learner profiles according to claim 1, characterized in that, In step S1, the process of constructing the learner profile is as follows: S14. Extract learning characteristics of learners’ learning interests, learning behaviors and learning status from learners’ online learning data; S15. The learned features are standardized using the z-score standardization method, and then reconstructed using an autoencoder. The z-score standardization formula is as follows: Where Xi represents the original data points before standardization, Mean represents the mean of the original data set, and Std represents the standard deviation of the original data set. S16. The reconstructed 3D learner features are clustered using the quadratic K-Means clustering method to obtain a 3D profile that reflects the learner's characteristics.

4. The learning resource recommendation method based on knowledge graphs and learner profiles according to claim 1, characterized in that, The process of step S2 is as follows: S21. Extract course, level, exercise and knowledge point information; S22. Using courses, exercises, levels, and knowledge points as knowledge unit nodes, construct a knowledge graph that starts with courses and radiates outward along the path of "course-exercise-level-knowledge point". S23. Construct four semantic relationships—inclusion, dependency, correlation, and mapping—between the level nodes. These relationships are constructed based on the level name and question description to describe the specific connections between different levels. The formula for determining correlation is as follows: Sim(A,B)=ω1Sim name (A,B)+ω2Sim content (A,B)≥T corr (3) Among them, Sim name (A, B) represent name similarity; Sim content (A, B) represent content similarity; ω1 and ω2 are the weights of name similarity and content similarity; Vec name and Vec content The result of vectorizing the level names and question descriptions in S1 using TF-IDF; T corr To preset the relevance similarity threshold; The formula for the overlap ratio when constructing containment relationships is as follows: Among them, K A K B These are the keyword sets for level A and level B extracted in S1, respectively.

5. The learning resource recommendation method based on knowledge graphs and learner profiles according to claim 1, characterized in that, The process of step S3 is as follows: S31. Based on the learner-level interaction records, obtain the learner's explicit preference information for the levels already interacted with, as the initial information for preference propagation; S32. Combining the knowledge graph, the initial preference information is propagated along the semantic relationships between level nodes to non-interactive level nodes through the preference propagation idea, thereby inferring the learner's potential preferences for these non-interactive levels. S33. The potential preference information obtained after the preference propagation is fused with the explicit preference information in the learner profile to form the final learner preference model.

6. The learning resource recommendation method based on knowledge graphs and learner profiles according to claim 1, characterized in that, The process of step S4 is as follows: S41. The learning content profile, learner profile, and learner-level interaction records after preference propagation are fused to form the input feature matrix of the deep belief network. S42. A deep belief network is used to perform unsupervised pre-training and supervised fine-tuning on the input feature matrix. In the unsupervised pre-training stage, a layer-by-layer greedy training method is used to train the restricted Boltzmann machine to learn the feature representation of the data. In the supervised fine-tuning stage, the learned feature representation is input into the Softmax classifier for training. S43. The output of the deep belief network is used as the learner's preference score for non-interactive learning resources, and a personalized list of learning resources is dynamically generated and recommended to the learner based on the preference score.

Images