Life-life learning course recommendation method based on knowledge graph and life-life learning platform

Abstract

The embodiment of the invention relates to the technical field of informatization teaching, in particular to a lifelong learning course recommendation method based on a knowledge graph and a lifelong learning platform. According to the embodiment of the invention, the method comprises the steps: constructing a user portrait and a knowledge graph, extracting a course-knowledge point sub-graph related to a current user from the knowledge graph, processing the course-knowledge point sub-graph through a GNN model to generate a course vector, and carrying out the multi-modal matching according to the similarity between the course vector and the user portrait; and iteratively selecting courses through an MMR algorithm to obtain a candidate course list. According to the method, the user portrait and the knowledge graph are constructed according to the user data and the course data, and the candidate course list is made according to the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more matched with the current user, and the course recommendation accuracy and the integrity of a user learning path are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of information-based teaching technology, and specifically to a lifelong learning course recommendation method and a lifelong learning platform based on a knowledge graph. Background Art

[0002] Lifelong learning refers to the continuous learning process that occurs throughout a person's life, designed to adapt to social development and achieve individual growth. This is often referred to as "living and learning" or "learning never ends." The concept of lifelong learning emerged within a unique social, educational, and life context. It is characterized by its lifelong, universal, and widespread nature. Since the concepts of lifelong education and lifelong learning were introduced, they have been widely recognized and actively practiced. Lifelong learning inspires us to cultivate the concept of lifelong education, enabling students to learn how to learn, and more importantly, to cultivate positive habits of initiative, continuous exploration, self-renewal, application of learning, and the optimization of knowledge.

[0003] There is no existing course recommendation system for lifelong learning. Other course recommendation systems rely primarily on collaborative filtering (CF), content-based filtering, or simple rule engines, which present the following key issues:

[0004] Inaccurate recommended content: Traditional recommendation systems (such as collaborative filtering) rely on historical user behavior data but lack a deep understanding of the logical connections between knowledge points. This can lead to recommended courses that may not meet users' actual learning needs. Content-based recommendations only consider course tag matching, ignoring the prerequisite relationships between knowledge points and the learner's knowledge level, and are prone to recommending repetitive or overly simple / difficult content.

[0005] Incomplete learning paths: Most recommendation systems only provide discrete lists of courses, rather than a systematic learning path. Users must plan their own learning sequence, increasing the cognitive burden. The lack of knowledge graph support makes it difficult to build a comprehensive, multi-level, and interdisciplinary knowledge system, hindering learning outcomes.

[0006] Lack of dynamic adaptability: Existing systems usually adopt static recommendation strategies and are unable to dynamically adjust course recommendation methods according to user learning progress, resulting in lower accuracy of recommended courses after repeated use. Summary of the Invention

[0007] In view of the above problems, the embodiments of the present invention provide a lifelong learning course recommendation method and a lifelong learning platform based on knowledge graph, which are used to solve the problems of inaccurate recommendation content, lack of dynamic adaptability and incomplete learning paths in the learning course recommendation system in the prior art.

[0008] According to one aspect of an embodiment of the present invention, a method for recommending lifelong learning courses based on a knowledge graph is provided, wherein the method includes:

[0009] Obtain user data, course data, and device data, where user data includes user behavior data, user history data, and user learning goals;

[0010] Build a user profile of the current user based on the user history data and the user's learning goals, and build a knowledge graph based on the user behavior data and course data;

[0011] Extracting a course-knowledge point subgraph related to the current user from the knowledge graph, processing the course-knowledge point subgraph using a GNN model to generate a course vector, and performing multimodal matching based on the similarity between the course vector and the user profile; iteratively selecting courses using an MMR algorithm until a list of candidate courses is output;

[0012] Obtain user feedback from the current user on the candidate course list, and optimize and update the knowledge graph based on the user feedback.

[0013] In an optional manner, the constructing of a knowledge graph based on the user behavior data and course data specifically includes:

[0014] Extracting knowledge point entities from the course data using a BERT-NER model;

[0015] Build triples of knowledge graphs through the REBEL end-to-end model;

[0016] By using the TransE embedding model to predict missing relations, the triples of the knowledge graph are completed to generate a complete knowledge graph.

[0017] In an optional manner, extracting knowledge point entities from the course data using a BERT-NER model specifically includes:

[0018] The BERT encoder of the BERT-NER model generates a context-dependent representation of the input sequence X = (x1, ..., xn): H = BERT(X), H∈R n×d , where d is the BERT hidden layer dimension and n is the input sequence length; the fully connected layer of the BERT-NER model calculates the label probability of each token, and its formula is: P = Softmax (HW T +b), P∈R n·∣T∣ , where |T| is the number of label types, W is the label projection matrix, and b is the bias term; the output layer of the BERT-NER model outputs a list of knowledge point entities and their types.

[0019] The triples of the knowledge graph are established by the REBEL end-to-end model, specifically including: generating a triple sequence T according to the knowledge point entity S, and the formula is specifically: Among them, θ is the total parameters of the BERT-NER model, t <t is the generated partial sequence;

[0020] The missing relations are predicted by TransE, and the triples of the knowledge graph are completed to generate a complete knowledge graph, specifically including: f(h,r,t)=||h+rt|| p , where h,r,t∈R d is the d-dimensional embedding vector of the head entity, relation, and tail entity, ||h+rt|| p For L p norm; if ||h+rt||<preset threshold, then it is inferred that there is a relationship r; and a supplementary knowledge graph is generated based on the existence of the relationship r.

[0021] In an optional manner, the user historical data includes historical test data and historical behavior data, and constructing a user profile based on the user historical data and the user's learning goals specifically includes:

[0022] Knowledge mastery is calculated based on historical test data and historical behavior data; user vectors are generated by aggregating courses that users have interacted with based on the GNN model; recommendation weights are adjusted based on learning goals to output user profiles, where the user profiles include courses learned and knowledge mastery vectors.

[0023] In an optional manner, extracting a course-knowledge point subgraph related to the user from the knowledge graph specifically includes:

[0024] Use Cypher query to locate the nodes of the user's learned courses, expand the subgraph based on the relationship type; use the PageRank algorithm to calculate the importance of the nodes in the subgraph, using the formula: Keep the Top-K nodes and output the course-knowledge point subgraph; where d is the damping coefficient, Bu is the set of nodes pointing to u, and L(v) is the number of outbound links of v.

[0025] In an optional method, multimodal matching is performed by calculating the similarity between the course vector and the user vector, specifically including:

[0026] The similarity between the course vector and the user vector is calculated using the enhanced similarity formula, where the enhanced similarity formula is: Where u∈R d is the user vector; e cis a course vector, is a similarity weight, is a distance attenuation coefficient, and d(ku, kc) is the shortest path edge number from the known knowledge point ku to the required kc of the course.

[0027] The multi-modal matching is performed according to an environment adaptation formula: FinalScore(c) = Score(u, c) x (1 + gamma II(device data)), where gamma is a mobile terminal reward coefficient, and II(·) is an indication function.

[0028] In an optional manner, the courses are iteratively selected through the MMR algorithm until a candidate course list is selected, and the method specifically comprises the following steps:

[0029] initializing an empty result set R,

[0030] judging whether the number of courses in the result set R reaches N;

[0031] if yes, outputting R; otherwise, calculating all candidate MMR scores, selecting a course with the highest score to join the result set R, and re-judging whether the number of courses in the updated result set R reaches N, until all candidate courses are outputted.

[0032] In an optional manner, user feedback and course learning data of a current user are acquired in real time, entity relationships of a knowledge graph are corrected according to the user feedback, and the knowledge graph is updated and optimized according to the course learning data, and the method specifically comprises the following steps:

[0033] user feedback is collected from user queries, log data or external knowledge sources, entity relationships that need to be corrected are identified, and the entity relationships of the knowledge graph are corrected through any one or more of deleting incorrect relationships, adding new relationships and updating relationship weights;

[0034] candidate course lists and learning progress of the current user are acquired in real time, and the knowledge graph is continuously trained according to the candidate course lists and the learning progress of the current user.

[0035] According to still another aspect of the embodiment of the present application, a lifelong learning platform is provided, and the lifelong learning platform comprises:

[0036] a data acquisition module configured to acquire user data, course data and device data, wherein the user data comprises user behavior data, user historical data and user learning goals;

[0037] a user portrait construction module configured to construct a user portrait of a current user according to the user historical data and the user learning goals;

[0038] a knowledge graph construction module configured to construct a knowledge graph according to the user behavior data and the course data;

[0039] The course recommendation module is used to extract the course-knowledge point subgraph related to the current user from the knowledge graph, process the course-knowledge point subgraph using the GNN model to generate a course vector, and perform multimodal matching based on the similarity between the course vector and the user profile; iteratively select courses using the MMR algorithm until a list of candidate courses is output;

[0040] And a feedback update module is used to obtain user feedback from the current user on the candidate course list, and optimize and update the knowledge graph based on the user feedback.

[0041] In an optional manner, the system further includes:

[0042] The teaching management module is used to perform one or more operations of adding, deleting, checking and modifying course data;

[0043] The educational management module is used to obtain all user data of the current course in real time and to visualize the learning progress and mastery of all users of the current course;

[0044] The user management module is used to perform one or more operations including account management, user authentication, and course management on any one or more of the system user module, learning user module, and teaching user module;

[0045] Data storage module, used to store user data, course data and device data.

[0046] The embodiment of the present invention constructs a user profile and a knowledge graph, extracts a course-knowledge point subgraph related to the current user from the knowledge graph, processes the course-knowledge point subgraph through a GNN model to generate a course vector, and performs multimodal matching based on the similarity between the course vector and the user profile; it iteratively selects courses through the MMR algorithm to obtain a list of candidate courses.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] 1. The present invention combines the construction of user portraits and knowledge graphs based on user data and course data, and makes a list of candidate courses based on the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more consistent with the current user, thereby improving the accuracy of course recommendations and the integrity of the user's learning path.

[0049] 2. The present invention obtains the user feedback and course learning data of the current user in real time, corrects the entity relationship of the knowledge graph according to the user feedback, and optimizes and updates the knowledge graph according to the course learning data; further improves the dynamic adaptability of the course recommendation method and further improves the accuracy of course recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are only used to illustrate the embodiments and are not to be considered as limiting the present invention. In addition, the same reference symbols are used to represent the same components throughout the drawings. In the drawings:

[0051] Figure 1 A flowchart of a first embodiment of a method for recommending lifelong learning courses based on a knowledge graph provided by the present invention is shown;

[0052] Figure 2 A schematic diagram of the process of constructing a knowledge graph based on user behavior data and course data provided by the present invention is shown;

[0053] Figure 3 A schematic diagram of a process for iteratively selecting courses by using the MMR algorithm provided by the present invention until a list of candidate courses is selected is shown;

[0054] Figure 4 It shows a structural schematic diagram of the lifelong learning platform provided by the present invention. DETAILED DESCRIPTION

[0055] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0056] Example 1:

[0057] Figure 1 A first embodiment of a method for recommending lifelong learning courses based on a knowledge graph is shown. The method includes:

[0058] 110. Obtain user data, course data, and device data. User data includes user behavior data, user history data, and user learning goals. In step 110, user data can be generated based on the user's personal information and learning information entered into the lifelong learning platform, as well as information such as usage logs generated by the current user through their use of the lifelong learning platform. Course data is all data related to courses uploaded to the lifelong learning platform. Device data refers to data on various teaching devices, including teaching terminals, mobile phones, computers, and tablets. Course data can also include Coursera open course description text. User behavior data can include user clicks, favorites, and learning time. The format of user behavior data can be: user ID, course ID, behavior type, and timestamp.

[0059] 120. Build a user profile for the current user based on the user's historical data and learning goals, and build a knowledge graph based on the user's behavior data and course data. In step 120, the user's historical data includes historical test data and historical behavior data. Building a user profile based on the user's historical data and learning goals specifically includes: calculating knowledge mastery based on the historical test data and historical behavior data; generating a user vector by aggregating the courses the user has interacted with based on the GNN model; and adjusting the recommendation weight based on the learning goals to output a user profile, where the user profile includes the courses already learned and the knowledge mastery vector. Building a knowledge graph based on the user's behavior data and course data specifically includes: extracting knowledge point entities from the course data using the BERT-NER model; establishing triples of the knowledge graph using the REBEL end-to-end model; and completing the triples of the knowledge graph by predicting missing relationships using the TransE embedding model to generate a complete knowledge graph.

[0060] 130, extract the course-knowledge point subgraph related to the current user from the knowledge graph, process the course-knowledge point subgraph through the GNN model to generate a course vector, and perform multimodal matching based on the similarity between the course vector and the user portrait; iteratively select courses through the MMR algorithm until a list of candidate courses is output; in step 130, the present invention combines the construction of user portraits and knowledge graphs based on user data and course data, and makes a list of candidate courses based on the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more consistent with the current user, thereby improving the accuracy of course recommendations and the integrity of the user learning path.

[0061] 140. Obtain user feedback on the candidate course list from the current user and optimize and update the knowledge graph based on the user feedback. In step 140, user feedback is collected from user queries, log data, or external knowledge sources, entity relationships that need to be corrected are identified, and entity relationships in the knowledge graph are corrected by deleting incorrect relationships, adding new relationships, and updating relationship weights. The current user's candidate course list and learning progress are obtained in real time, and the knowledge graph is continuously trained based on the current user's candidate course list and learning progress.

[0062] In this embodiment, the embodiment of the present invention constructs a user portrait and a knowledge graph, extracts a course-knowledge point subgraph related to the current user from the knowledge graph, processes the course-knowledge point subgraph through the GNN model to generate a course vector, and performs multimodal matching based on the similarity between the course vector and the user portrait; iteratively selects courses through the MMR algorithm to obtain a list of candidate courses. The present invention combines the construction of user portraits and knowledge graphs based on user data and course data, and makes a list of candidate courses based on the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more consistent with the current user, improving the accuracy of course recommendations and the integrity of the user learning path. The present invention obtains user feedback and course learning data of the current user in real time, corrects the entity relationship of the knowledge graph based on the user feedback, and optimizes and updates the knowledge graph based on the course learning data; further improves the dynamic adaptability of the course recommendation method, and further improves the accuracy of course recommendations.

[0063] Example 2:

[0064] This embodiment is based on Example 1. This embodiment discloses a first embodiment of a method for recommending lifelong learning courses based on a knowledge graph. This embodiment further refines steps 110-140 in Example 1.

[0065] Specifically, in step 120, a knowledge graph is constructed based on user behavior data and course data. Figure 2 , specifically including:

[0066] 210, extracting knowledge point entities from the course data using the BERT-NER model; In step 210, extracting knowledge point entities from the course data using the BERT-NER model specifically includes: the BERT encoder of the BERT-NER model generates a context-related representation for the input sequence X = (x1, ..., xn): H = BERT(X), H∈R n×d , where d is the BERT hidden layer dimension and n is the input sequence length; the fully connected layer of the BERT-NER model calculates the label probability of each token, and its formula is: P = Softmax (HW T +b), P∈R n·∣T∣ , where |T| is the number of label types, W is the label projection matrix, and b is the bias term. The output layer of the BERT-NER model outputs a list of knowledge point entities and their types. The sequence labeling model based on the BERT-NER model outputs BIOES labels. At the same time, a loss function can be calculated using cross-entropy to fine-tune the entity list and types. In a specific example, the fine-tuned BERT-NER model is used to extract knowledge points, extracting "linear algebra" from "This course covers the basics of linear algebra."

[0067] 220, establishing triples of the knowledge graph through the REBEL end-to-end model; In step 220, establishing triples of the knowledge graph through the REBEL end-to-end model specifically includes: generating a triple sequence T based on the knowledge point entity S, and the formula is specifically: Among them, θ is the total parameters of the BERT-NER model, t <t This is a partial sequence that has already been generated. In this example, the REBEL end-to-end model supports over 200 relationship types, avoiding the error accumulation of traditional pipelines. Furthermore, relationships can be supplemented through the rule engine, such as extracting specific structures based on dependency analysis. Specifically, entity association can be calculated using co-occurrence frequency and PMI.

[0068] 230, by using the TransE embedding model to predict missing relations, the triples of the knowledge graph are completed to generate a complete knowledge graph. In step 230, by using TransE to predict missing relations, the triples of the knowledge graph are completed to generate a complete knowledge graph, specifically including: f(h,r,t)=||h+rt|| p , where h,r,t∈R d is the d-dimensional embedding vector of the head entity, relation, and tail entity, ||h+rt|| p For L p norm; if ||h+rt||<preset threshold, then infer the existence of relationship r; generate a supplementary knowledge graph based on the existence of relationship r. In step 230, the existing triples + user behavior logs are input into the TransE embedding model to obtain the existence of relationship r, and the existence of relationship r is added to the triples. The supplemented triples are the knowledge graph. In a specific example, the TransE model predicts the missing relationship "linear algebra → machine learning", which can be obtained by ||h+rt|| p =0.8<threshold 1.0 for completion.

[0069] In addition, the TransE embedding model uses rule-based reasoning engines such as Drools to check for logical contradictions in triples. For example, if "Course A → Course B" and "Course B → Course A" exist, the low-confidence relationship is deleted. If a circular dependency (e.g., k1 → k2 → k3 → k1) is detected, a manual review process is triggered.

[0070] In steps 210-230, the BERT-NER model is adaptively trained using educational texts such as textbooks and course outlines based on the general BERT pre-training model, with the annotated entity types including knowledge points, courses, and subjects. The training data can be: public datasets such as CoNLL format annotated data in the education field, or core knowledge points extracted from course descriptions, such as manual annotation of "Newton's law" as a knowledge point. In terms of hyperparameter settings, the learning rate can be set to 2e -5 , the batch size is set to 32, the training rounds are set to 10, and the loss function is cross entropy loss.

[0071] The threshold in the TransE embedding model can be determined by optimizing the validation set and setting the threshold to ||h+rt|| p <1.0, the training process can adopt a negative sampling strategy such as a negative sampling ratio of 1:10, an Adam optimizer, a learning rate of 0.001, and an embedding dimension d = 200.

[0072] In step 120, the user's historical data includes historical test data and historical behavior data, and a user profile is constructed based on the user's historical data and the user's learning goals, specifically including: calculating knowledge mastery based on historical test data and historical behavior data; generating user vectors by aggregating the courses that the user has interacted with based on the GNN model; adjusting the recommendation weights in combination with the learning goals, thereby outputting a user profile, wherein the user profile includes the courses that have been learned and the knowledge mastery vector. In this embodiment, the user portrait construction can specifically include the following steps: standardizing the test data and quantifying the behavior data, extracting the standardization of the test data and behavior data; calculating the knowledge mastery based on the fusion model; constructing a line graph structure, and then generating user vectors through GNN aggregation, adjusting the learning goal weights, and generating a user profile. Through the above steps, the system can dynamically generate accurate user portraits, wherein the knowledge mastery quantifies the user's current level; the user vector encodes the long-term learning model; and the target weight ensures that the recommendation is aligned with the user's intention.

[0073] In the above scheme, according to the network structure of the GNN model: the GraphSAGE model is adopted, the aggregation function is mean pooling, the number of layers is 2, the activation function is ReLU, and the training process is: optimizing the similarity between the user vector and the course vector through contrastive learning. The loss function is: L = \sum_{(u,c^+)}\max(0,\epsilon-\text{sim}(u,c^+)+\text{sim}(u,c^-)), where c^+ is a positive sample that can be expressed as the course the user has learned, and c^- is a negative sample that can be expressed as a randomly sampled course that has not been learned, and ε = 0.2.

[0074] In step 130, the course-knowledge point subgraph related to the user is extracted from the knowledge graph, specifically including: locating the user's learned course nodes through Cypher query, expanding the subgraph based on the relationship type; calculating the importance of the nodes in the subgraph through the PageRank algorithm, using the formula: Keep the Top-K nodes and output the course-knowledge point subgraph; where d is the damping coefficient, Bu is the set of nodes pointing to u, and L(v) is the number of outlinks of v.

[0075] In this step, the PageRank algorithm uses a damping coefficient of d = 0.85, following the classic setting. The number of retained Top-K nodes, K = 50, was adjusted experimentally to balance computational efficiency and coverage. The enhanced similarity formula, α_sim = 0.7, was determined through A / B testing, with a weighting bias toward user-course similarity. The distance decay coefficient, λ = 0.5, was set based on the average path length between knowledge points in the knowledge graph. The environmental adaptation formula is γ = 0.2, where γ is the mobile reward coefficient, derived from the statistical impact of user device type on completion rate.

[0076] In addition, the PageRank algorithm can be improved in the course-knowledge point subgraph extraction process as follows: When calculating node importance, assign a higher weight to "prerequisite relationship" edges, such as a weight of 2.0, to ensure that prerequisite knowledge points are not filtered out. Dynamically adjust Top-K: Adaptively adjust the K value based on the complexity of the user's learning goal, for example, K = 30 for basic learning goals and K = 70 for advanced goals.

[0077] In step 130, multimodal matching is performed by calculating the similarity between the course vector and the user vector, specifically including: calculating the similarity between the course vector and the user vector by using an enhanced similarity formula, wherein the enhanced similarity formula is: Where u∈R d is the user vector; e c is the course vector, β is the similarity weight, which can be set to 0.7 and can be obtained through training. λ is the distance decay coefficient, which can be 0.5. d(ku,kc) is the number of edges in the shortest path from the known knowledge point ku to the required course kc. Multimodal matching is performed according to the environment adaptation formula: FinalScore(c) = Score(u,c) × (1 + γ·II(device data)), where γ is the mobile terminal reward coefficient, which can be 0.2, and II(·) is the indicator function.

[0078] In step 130, the course selection is iterated by the MMR algorithm until a list of candidate courses is selected. Figure 3 , specifically including:

[0079] 310, initialize the empty result set R,

[0080] 320, determine whether the number of courses in the result set R reaches N;

[0081] 330, if the output result set R;

[0082] 340, otherwise calculate all candidate MMR scores and select the course with the highest score to add to the result set R, and continue to execute step 320 until a list of all candidate courses is output.

[0083] In the embodiment shown in steps 310-340, the formula for calculating the candidate course list is:

[0084] c*=argmax c∈C / R [λ·FinalScore(c)-(1-λ)max r∈R sim(e c ,e r )],in,

[0085] λ = 0.7, where λ is the correlation weight; sim(·,·) is the cosine similarity; R is the set of selected courses, and C is the total number of candidate courses. Using the above formula, we can continuously update the result set R and calculate all candidate courses.

[0086] In a specific example, the knowledge graph and user profile can be reflected through knowledge point mapping rules. Specifically, questions in the user's historical test data are associated with knowledge points, such as by binding question IDs to knowledge point IDs. Knowledge mastery is then calculated using the "question-knowledge point" edge in the knowledge graph. The knowledge mastery formula is as follows: M_k = \frac{\sum_{q\in Q_k}\text{correctness}(q)}{|Q_k|}, where Q_k is the set of questions associated with knowledge point k.

[0087] The iterative course selection process using the MMR algorithm until a list of candidate courses is obtained also includes the following: Diversity metric: The difference in knowledge points between courses is introduced into the MMR algorithm: \text{MMR}(c)=\arg\max_{c_i\in C}\left[\alpha\cdot\text{sim}(u,c_i)-(1-\alpha)\cdot\max_{c_j\in R}\text{dist}(c_i,c_j)\right]

[0088] dist(c_i,c_j) is the Jaccard distance between the knowledge points of courses i and j: \text{dist}(c_i,c_j)=1-\frac{|K_{c_i}\cap K_{c_j}|}{|K_{c_i}\cup K_{c_j}|}

[0089] The parameter α is set to 0.6, where α can balance relevance and diversity through user research.

[0090] In an optional method, user feedback and course learning data of the current user are obtained in real time, entity relationships of the knowledge graph are modified based on user feedback, and the knowledge graph is optimized and updated based on course learning data, specifically including:

[0091] Collect user feedback from user queries, log data or external knowledge sources, identify entity relationships that need to be corrected, and correct the entity relationships in the knowledge graph by deleting any one or more of the following: deleting incorrect relationships, adding new relationships and updating relationship weights. In this embodiment, correcting entity relationships in the knowledge graph based on user feedback specifically includes: (1) Data collection and feedback aggregation: Collect feedback information from user queries, log data or external knowledge sources to identify entity relationships that need to be corrected. Aggregate feedback information, such as counting the number of times users have corrected a certain entity relationship or the confidence score. (2) Entity relationship correction: Delete incorrect relationships: If the feedback indicates that a certain relationship is not valid, delete the relationship from the knowledge graph. Add new relationships: If the feedback indicates that a certain relationship is valid but does not exist in the knowledge graph, add the relationship. Update relationship weights: Adjust the relationship weights, such as the confidence score, based on the confidence score of the feedback. (3) Consistency check: Check whether the corrected knowledge graph meets consistency constraints such as transitivity and symmetry. Use inference rules such as rule engines to automatically repair potential conflicts.

[0092] Obtain the current user's candidate course list and learning progress in real time, and continue to train the knowledge graph based on the current user's candidate course list and learning progress. Optimizing and updating the knowledge graph based on course learning data specifically includes: (1) Incremental data preparation, extracting updated entities and relationships from the knowledge graph, and constructing an incremental dataset ΔG = (ΔV, ΔE), where ΔV is the newly added / modified entity and ΔE is the newly added / modified relationship. (2) Initializing or loading the model, loading a pre-trained GNN model such as GCN, GAT, R-GCN, etc., with parameters θ. If the amount of incremental data is small, you can choose to fix some model parameters such as the bottom embedding layer and only fine-tune the high-level parameters. (3) Incremental training: Use the incremental data ΔG to fine-tune the model, and the optimization goal is to minimize the loss function L(θ). The loss function usually includes: entity embedding loss such as triple loss and contrastive learning loss. Relationship prediction loss such as cross entropy loss. (4) Embedding update, updating the entity and relationship embedding representations hv and hr to reflect the latest knowledge graph structure.

[0093] In a specific example, user feedback processing involves mapping user ratings (1-5) to updated weights of course nodes in the knowledge graph: w_{\text{new}} = w_{\text{old}} + \beta\cdot(\text{rating} - 3), where β = 0.1 is the feedback coefficient. A rating greater than 3 increases the relationship weight, while a rating less than 3 decreases it. Furthermore, newly added relationships are validated using a rule engine such as Apache Jena to avoid circular dependencies. For example, when "Course A → Course B" conflicts with "Course B → Course A," relationships with high confidence are retained.

[0094] In addition, the user rated the recommended course "Introduction to Machine Learning" with 4 points, triggering an update of the weight of the course node in the knowledge graph, such as w_{\text{Introduction to Machine Learning}}=0.8+0.1\cdot(4-3)=0.9.

[0095] Example 3:

[0096] Figure 4 An embodiment of a lifelong learning platform of the present invention is shown. This embodiment is based on embodiment 1 or 2. The lifelong learning platform of this embodiment includes a data acquisition module 410, a user portrait construction module 420, a knowledge graph construction module 430, a course recommendation module 440 and a feedback update module 450.

[0097] Specifically, data acquisition module 410 is configured to execute step 110 in Example 1, including: user data can be generated based on the user's personal information and learning information entered into the knowledge graph-based lifelong learning course recommendation system, as well as information such as usage logs generated by the current user through the knowledge graph-based lifelong learning course recommendation system; course data is all data uploaded to the knowledge graph-based lifelong learning course recommendation system. Device data refers to data on different teaching devices, including teaching terminals, mobile phones, computers, tablets, etc.

[0098] The user portrait construction module 420 is used to execute step 120 in Example 1, specifically including: user historical data including historical test data and historical behavior data, constructing a user portrait based on the user historical data and the user learning goals, specifically including: calculating knowledge mastery based on the historical test data and historical behavior data; generating a user vector by aggregating the courses that the user has interacted with based on the GNN model; adjusting the recommendation weight in combination with the learning goals, thereby outputting a user portrait, wherein the user portrait includes the courses that have been learned and the knowledge mastery vector.

[0099] The knowledge graph construction module 430 is used to execute step 120 in Example 1, specifically including: constructing a knowledge graph based on user behavior data and course data, specifically including: extracting knowledge point entities from course data through the BERT-NER model; establishing triples of the knowledge graph through the REBEL end-to-end model; and completing the triples of the knowledge graph by predicting missing relationships using the TransE embedding model to generate a complete knowledge graph.

[0100] The course recommendation module 440 is used to execute step 130 in Example 1, specifically including: the present invention combines the construction of user portraits and knowledge graphs based on user data and course data, and makes a list of candidate courses based on the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more consistent with the current user, thereby improving the accuracy of course recommendations and the completeness of the user's learning path.

[0101] The feedback update module 450 is used to execute step 140 in Example 1, specifically including: collecting user feedback from user queries, log data or external knowledge sources, identifying entity relationships that need to be corrected, and correcting the entity relationships of the knowledge graph by deleting any one or more of erroneous relationships, adding new relationships and updating relationship weights; obtaining the current user's candidate course list and learning progress in real time, and continuing to train the knowledge graph based on the current user's candidate course list and learning progress.

[0102] The embodiment of the present invention constructs a user portrait and a knowledge graph, extracts a course-knowledge point subgraph related to the current user from the knowledge graph, processes the course-knowledge point subgraph through the GNN model to generate a course vector, and performs multimodal matching based on the similarity between the course vector and the user portrait; it iteratively selects courses through the MMR algorithm to obtain a list of candidate courses. The present invention combines the construction of user portraits and knowledge graphs based on user data and course data, and makes a list of candidate courses based on the relationship between the knowledge graph and the user portrait, so that the recommended candidate course list is more consistent with the current user, improving the accuracy of course recommendations and the integrity of the user learning path. The present invention obtains user feedback and course learning data of the current user in real time, corrects the entity relationship of the knowledge graph based on the user feedback, and optimizes and updates the knowledge graph based on the course learning data; further improves the dynamic adaptability of the course recommendation method, and further improves the accuracy of course recommendations.

[0103] In an alternative mode, the system further comprises: a teaching management module, configured to perform any one or more of adding, deleting, inquiring, and modifying the course data; a teaching administration management module, configured to acquire all user data of the current course in real time, and to visually display the learning progress and mastery of all users of the current course; a user management module, configured to perform any one or more of account management, user authentication, and course management on any one or more of the system user module, the learning user module, and the teaching user module; and a data storage module, configured to store the user data, the course data, and the device data. The system, through the teaching management module, the teaching administration management module, the user management module, and the data storage module, constitutes a complete lifelong learning system. The lifelong learning platform can be an open and integrated education project. The system, through deep implementation of educational technology, simplifies the information acquisition process, and promotes the continuous development and comprehensive application of knowledge. The platform aims to meet the learning needs of people of different ages and backgrounds, and to promote the overall improvement of the education level of the society.

[0104] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Furthermore, embodiments of the application are not described with reference to any particular programming language.

[0105] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application can be practiced without these specific details. In other instances, well-known methods, structures and techniques have not been described in detail in order to not obscure the understanding of this description. Like numbers refer to like elements throughout. Similarly, while operations are presented and described sequentially, embodiments can be operated in an order other than that which is presented.

[0106] Those skilled in the art will appreciate that modules in the apparatuses in the embodiments can be adapted and placed in one or more apparatuses other than the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and further can be divided into multiple sub-modules or sub-units or sub-components. At least some of such features and / or processes or units are mutually exclusive.

[0107] It should be noted that the above embodiments illustrate rather than limit the invention, and that alternative embodiments may be devised by a person skilled in the art without departing from the scope of the appended claims. In the claims, any reference signs placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising several different elements and by means of appropriately programmed computers. In a unit claim enumerating several systems, several of these systems may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names. The steps in the above embodiments should not be understood as limiting the order of execution unless otherwise specified.

Claims

1. A lifelong learning course recommendation method based on knowledge graph, characterized in that: The method comprises: Obtain user data, course data, and device data, where user data includes user behavior data, user history data, and user learning goals; Build a user profile of the current user based on the user history data and the user's learning goals, and build a knowledge graph based on the user behavior data and course data; Extracting a course-knowledge point subgraph related to the current user from the knowledge graph, processing the course-knowledge point subgraph using a GNN model to generate a course vector, and performing multimodal matching based on the similarity between the course vector and the user profile; iteratively selecting courses using an MMR algorithm until a list of candidate courses is output; Obtain user feedback and course learning data of the current user in real time, modify the entity relationship of the knowledge graph based on the user feedback, and optimize and update the knowledge graph based on the course learning data.

2. The method for recommending lifelong learning courses based on knowledge graph according to claim 1, characterized in that: The constructing of a knowledge graph based on the user behavior data and course data specifically includes: Extracting knowledge point entities from the course data using a BERT-NER model; Build triples of knowledge graphs through the REBEL end-to-end model; By using the TransE embedding model to predict missing relations, the triples of the knowledge graph are completed to generate a complete knowledge graph.

3. The method for recommending lifelong learning courses based on knowledge graph according to claim 2, characterized in that: The extracting of knowledge point entities from the course data using the BERT-NER model specifically includes: The BERT encoder of the BERT-NER model generates a context-dependent representation of the input sequence X = (x1, ..., xn): H = BERT(X), H∈R n×d , where d is the BERT hidden layer dimension and n is the input sequence length; the fully connected layer of the BERT-NER model calculates the label probability of each token, and its formula is: P = Softmax (HW T +b), P∈R n·∣T∣ , where |T| is the number of label types, W is the label projection matrix, and b is the bias term; the output layer of the BERT-NER model outputs a list of knowledge point entities and their types. The triples of the knowledge graph are established by the REBEL end-to-end model, specifically including: generating a triple sequence T according to the knowledge point entity S, and the formula is specifically: Among them, θ is the total parameters of the BERT-NER model, t ＜t is the generated partial sequence; The missing relations are predicted by TransE, and the triples of the knowledge graph are completed to generate a complete knowledge graph, specifically including: f(h,r,t)=||h+rt|| p , where h,r,t∈R d is the d-dimensional embedding vector of the head entity, relation, and tail entity, ||h+rt|| p For L p norm; if ||h+rt||<preset threshold, then it is inferred that there is a relationship r; and a supplementary knowledge graph is generated based on the existence of the relationship r.

4. The method for recommending lifelong learning courses based on knowledge graph according to claim 1, characterized in that: The user historical data includes historical test data and historical behavior data. The user profile is constructed based on the user historical data and the user's learning goals, specifically including: Knowledge mastery is calculated based on historical test data and historical behavior data; user vectors are generated by aggregating courses that users have interacted with based on the GNN model; recommendation weights are adjusted based on learning goals to output user profiles, where the user profiles include courses learned and knowledge mastery vectors.

5. The method for recommending lifelong learning courses based on knowledge graph according to claim 1, characterized in that: The process of extracting the course-knowledge point subgraph related to the user from the knowledge graph specifically includes: Use Cypher query to locate the nodes of the user's learned courses, expand the subgraph based on the relationship type; use the PageRank algorithm to calculate the importance of the nodes in the subgraph, using the formula: Keep the Top-K nodes and output the course-knowledge point subgraph; where d is the damping coefficient, Bu is the set of nodes pointing to u, and L(v) is the number of outbound links of v.

6. The method for recommending lifelong learning courses based on knowledge graph according to claim 5, characterized in that: Multimodal matching is performed by calculating the similarity between the course vector and the user vector, specifically including: The similarity between the course vector and the user vector is calculated using the enhanced similarity formula, where the enhanced similarity formula is: Where u∈R d is the user vector; e c is the course vector, β is the similarity weight, λ is the distance decay coefficient, and d(ku,kc) is the number of edges in the shortest path from the known knowledge point ku to the required course kc; Multimodal matching is performed according to the environment adaptation formula, which is: FinalScore(c)=Score(u,c)×(1+γ·II(device data)), where γ is the mobile terminal reward coefficient and II(·) is the indicator function.

7. The method for recommending lifelong learning courses based on knowledge graph according to claim 6, characterized in that: The iterative selection of courses using the MMR algorithm until a list of candidate courses is selected specifically includes: Initialize an empty result set R, Determine whether the number of courses in the result set R reaches N; If yes, output R; otherwise, calculate all candidate MMR scores and select the highest-scoring course to add to the result set R, and re-judge whether the number of courses in the updated result set R reaches N, until a list of all candidate courses is output.

8. The method for recommending lifelong learning courses based on knowledge graph according to claim 1, characterized in that: Obtain user feedback and course learning data of the current user in real time, modify entity relationships in the knowledge graph based on the user feedback, and optimize and update the knowledge graph based on the course learning data, specifically including: Collect user feedback from user queries, log data, or external knowledge sources, identify entity relationships that need to be corrected, and correct the entity relationships in the knowledge graph by deleting incorrect relationships, adding new relationships, and updating relationship weights. Obtain the current user's candidate course list and learning progress in real time, and continue to train the knowledge graph based on the current user's candidate course list and learning progress.

9. A lifelong learning platform, characterized by: The lifelong learning platform includes: A data acquisition module is used to acquire user data, course data, and device data, wherein user data includes user behavior data, user history data, and user learning goals; A user portrait building module is used to build a user portrait of the current user based on the user's historical data and user learning goals; A knowledge graph construction module, used to construct a knowledge graph based on the user behavior data and course data; The course recommendation module is used to extract the course-knowledge point subgraph related to the current user from the knowledge graph, process the course-knowledge point subgraph using the GNN model to generate a course vector, and perform multimodal matching based on the similarity between the course vector and the user profile; iteratively select courses using the MMR algorithm until a list of candidate courses is output; And a feedback update module is used to obtain user feedback from the current user on the candidate course list, and optimize and update the knowledge graph based on the user feedback.

10. The knowledge graph-based lifelong learning course recommendation system according to claim 9, characterized in that: The system further comprises: The teaching management module is used to perform one or more operations including adding, deleting, checking and modifying course data; The educational management module is used to obtain all user data of the current course in real time and to visualize the learning progress and mastery of all users of the current course; The user management module is used to perform one or more operations including account management, user authentication, and course management on any one or more of the system user module, learning user module, and teaching user module; Data storage module, used to store user data, course data and device data.

Images