Method for accurately positioning and breaking through subject weak points based on knowledge graph

Abstract

The invention discloses a knowledge graph-based method for accurately positioning and breaking through discipline weak points, which comprises the following steps of: predicting knowledge points possibly accessed by students in the next step by using a deep learning model based on an updated knowledge graph, capturing a multi-dimensional dependency relationship among the knowledge points through a graph neural network, and combining historical learning track data of the students to obtain a knowledge point database; generating a personalized learning resource recommendation list; dynamically adjusting the priority of the recommendation list according to the comprehensive association degree of the subject weak point and the key path, and improving the priority of the knowledge point of the key path highly associated with the subject weak point; and according to the priority of the recommendation list, adopting a multi-level caching strategy to pre-cache learning resources, maintaining a hotspot knowledge point index in a memory, storing secondary hotspot resources in a solid state disk, storing cold data in a mechanical hard disk, and dynamically adjusting cache content of each level through a cache replacement algorithm.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for accurately locating and breaking through weak points of a subject based on a knowledge graph. Background Art

[0002] In the process of breaking through the weak points of the subject, we face a series of complex technical challenges, including real-time processing of a large amount of learning behavior data, accurate identification of students' weak knowledge points and providing personalized learning recommendations. This involves the coordination of multiple links, including data processing, knowledge graph construction, learning trajectory analysis and resource recommendation. Efficient processing of real-time data streams in a distributed environment is the key, while ensuring the dynamic update of the knowledge graph. When identifying weak points, we need to balance computational efficiency and accuracy, quickly locate relevant knowledge points, and consider multi-dimensional correlations. When generating personalized recommendations, we need to comprehensively consider students' learning history, the complex dependencies between knowledge points, and weak point breakthrough strategies, which increases the complexity of the recommendation algorithm. In order to improve the system response speed, it is also necessary to design a multi-level cache strategy, but the dynamic adjustment of cache content will affect the real-time and accuracy of the recommendation. These interlocking issues jointly determine the effectiveness of locating and breaking through the weak points of the subject. Summary of the invention

[0003] The purpose of the present invention is to solve the problems in the prior art and to propose a method based on knowledge graph to accurately locate and break through the weak points of the subject.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for accurately locating and breaking through the weak points of a subject based on a knowledge graph, comprising the following steps:

[0006] S1, obtains students' learning behavior data stream from the education platform and learning system, uses a distributed stream processing framework to process students' learning behavior data stream, calculates the access frequency and residence time of knowledge points, dynamically adjusts the association strength of knowledge points according to the preset weight function, and obtains a real-time updated knowledge graph;

[0007] S2, based on the updated knowledge graph, uses a deep learning model to predict the knowledge points that students may visit next, captures the multi-dimensional dependencies between knowledge points through a graph neural network, and combines students' historical learning trajectory data to generate a personalized learning resource recommendation list;

[0008] S3, continuously monitors students’ learning behavior according to the recommendation list, records the time students spend on knowledge points, and triggers the graph traversal algorithm if the time students spend on a certain knowledge point exceeds the preset threshold. It uses a bidirectional breadth-first search to traverse the knowledge graph forward and backward at the same time, quickly locates related preceding and subsequent knowledge points, and constructs a local knowledge network.

[0009] S4, for the constructed local knowledge network, use the graph compression algorithm to reduce the number of nodes, retain important nodes and key paths, determine the weak points of the subject according to the students' stay time and access frequency at the knowledge points, and associate them with the important nodes in the local knowledge network;

[0010] S5, obtain the knowledge point data of the subject's weak points and key paths, calculate the comprehensive correlation through attribute fusion, and if the comprehensive correlation exceeds the preset threshold, it is judged that the subject's weak points and key paths have a strong connection;

[0011] S6, dynamically adjust the priority of the recommendation list according to the comprehensive correlation between the subject weaknesses and the key path, and increase the priority of the knowledge points of the key path that are highly correlated with the subject weaknesses;

[0012] S7, based on the priority of the recommendation list, adopts a multi-level cache strategy to pre-cache learning resources, maintains hot knowledge point indexes in memory, stores secondary hot resources in solid-state drives, stores cold data in mechanical hard drives, and dynamically adjusts cache content at all levels through cache replacement;

[0013] S8, according to the set weight evaluation system, calculate the importance of knowledge points in the cached content, access frequency, knowledge link integrity and correlation factors of subject weaknesses to obtain the weight score of each cached content. When the cache space is insufficient, eliminate the cached content with the lowest weight.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] The present invention discloses a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph, which solves the problem that it is difficult to capture students' learning status in real time, accurately identify weak points, and recommend learning resources in a personalized manner in the traditional education system. By processing the student learning behavior data stream in real time and dynamically updating the knowledge graph, the method can timely reflect the student's learning progress. By using deep learning and graph neural network technology, the student's learning trajectory can be accurately predicted, and personalized learning resource recommendations can be provided. The innovative use of bidirectional graph traversal and graph compression algorithms can quickly locate the weak points of the subject and construct a key knowledge network, effectively improving the accuracy and efficiency of weak point identification. By comprehensively calculating the correlation and dynamically adjusting the recommendation priority, a precise breakthrough in the weak points of the subject is achieved. Finally, through the application of multi-level cache strategy and intelligent cache replacement mechanism, the system response speed is greatly improved, providing students with a smooth learning experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method proposed by the present invention for accurately locating and breaking through the weak points of a subject based on a knowledge graph;

[0017] Figure 2 A schematic diagram of a method proposed by the present invention for accurately locating and breaking through the weak points of a subject based on a knowledge graph;

[0018] Figure 3 Another schematic diagram of a method proposed by the present invention for accurately locating and overcoming weaknesses in a subject based on a knowledge graph. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0020] like Figure 1-Figure 3 This embodiment discloses a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph, which may specifically include:

[0021] S1. Obtain students’ learning behavior data streams from education platforms and learning systems, including access logs, dwell time, and interaction records. Use a distributed stream processing framework to process students’ learning behavior data streams, calculate the access frequency and dwell time of knowledge points through a sliding window algorithm, and dynamically adjust the association strength of knowledge points according to a preset weight function to obtain a real-time updated knowledge graph.

[0022] Specifically, learning behavior data carrying a student's unique identification number is received. The learning behavior data is collected in real time by the education platform when the student is learning online. According to the learning behavior data, the data is transmitted to the Flink distributed stream processing framework in real time through the Kafka (distributed stream processing platform) message queue. The sliding time window function is used to aggregate the data transmitted to Flink to obtain the heat index of each knowledge point in the preset time window. If the heat index of a knowledge point exceeds the preset threshold, the Apriori algorithm is used to mine the association rules between the knowledge point and other knowledge points, calculate the association strength of the knowledge point, and store the mined association rules in the Neo4j graph database. The association relationship between knowledge points is stored in the graph database Neo4j to construct an initial knowledge graph. The DeepWalk algorithm is used to perform random walks on the knowledge graph stored in Neo4j to generate a knowledge point sequence. The knowledge point sequence is used to learn the vector representation of knowledge points using the Skip-gram model, capture the long-range dependencies between knowledge points, update the association strength between knowledge points, and obtain a real-time updated knowledge graph. According to the vector representation of the learned knowledge points, the PathSim algorithm is used to calculate the similarity between students and knowledge points, and generate a personalized learning path recommendation sequence for students.

[0023] For example, student access logs and interaction records are obtained from the education platform in real time, including the timestamp of each click, the length of stay, the content of questions and answers, and the scores of homework. The Kafka message queue partitions the data by the knowledge point ID to ensure that the data of the same knowledge point is sent to the same Flink task. Flink uses a 5-minute sliding time window, which slides every 30 seconds, to calculate the number of visits and the average stay time of the knowledge point in the window. For the number of visits, it is directly accumulated; for the stay time, the weighted average method is used, with the weight of the most recent data being 0.6 and the weight of the historical data being 0.4. When normalizing, the number of visits is divided by the maximum number of visits in the window, and the stay time is divided by the preset longest effective learning time, such as 30 minutes. The Apriori algorithm sets the minimum support to 0.1 and the minimum confidence to 0.5, mines frequent item sets from the student's learning sequence, and generates association rules such as IF knowledge point A THEN knowledge point B (support = 0.15, confidence = 0.7). In Neo4j, knowledge points are nodes, association rules are directed edges, and the weight of the edges is the confidence value. The DeepWalk algorithm performs random walks on the knowledge graph, sets the walk length to 10, and repeats each knowledge point 100 times to generate a sequence of knowledge points. Using the Skip-gram model, the window size is set to 5, and the knowledge points are mapped to a 50-dimensional vector space. The association strength of the knowledge points is updated by calculating the vector cosine similarity. When the cosine similarity of two knowledge point vectors is greater than 0.8, the association strength is increased. The PathSim algorithm calculates the similarity between students and knowledge points based on the meta-path student-learning-knowledge point-association-knowledge point. Set the difficulty range of the knowledge point to 1-5, and use the difficulty as the weight factor for PathSim calculation. According to the knowledge points that the students have mastered, select the knowledge points with the highest similarity and moderate difficulty (not exceeding the average difficulty of the knowledge points mastered + 1) to form a personalized learning path. After each knowledge point is learned, the recommendation results are recalculated and updated.

[0024] S2. Based on the updated knowledge graph, use the deep learning model to predict the knowledge points that students may visit next, capture the multi-dimensional dependencies between knowledge points through the graph neural network, and combine the students' historical learning trajectory data to generate a personalized learning resource recommendation list.

[0025] Specifically, the node features and edge relationship information of knowledge points in the knowledge graph database are obtained, and a feature vector containing adjacent nodes, edge attributes, and node types is constructed for each knowledge point; according to the feature vector, a graph attention network based on a meta-path is used to represent and learn the knowledge points, capture the multi-dimensional dependencies between knowledge points, and obtain a low-dimensional dense vector representation of the knowledge points; the student's historical learning trajectory data is obtained, including the sequence of visited knowledge points, residence time, and number of interactions, and combined with student portrait information such as age, major, learning ability, etc., a long short-term memory network is used to model student interests, and the padding and truncation methods are used to process learning sequences of different lengths, and the student's current interest state vector is output; the knowledge point vector representation and the student's interest are combined The state vectors are concatenated and input into a multilayer perceptron, which uses historical student access data for supervised learning, optimizes model parameters using the cross-entropy loss function, and calculates the student's access probability to each knowledge point; if the access probability is greater than the preset threshold, the corresponding knowledge point is added to the candidate recommendation set; based on the difficulty of the knowledge points in the candidate recommendation set, the student's current learning progress, and the preset learning path generated by the topological sorting result of the knowledge graph constructed based on the knowledge of domain experts, the candidate knowledge points are sorted using the weighted sum method, and the difficulty of the knowledge points, the correlation with the current learning progress, and the predicted access probability are comprehensively considered to generate the final personalized learning resource recommendation list, which includes the recommended knowledge point ID, the corresponding learning resource link, and the reason for the recommendation.

[0026] Exemplarily, knowledge point features are extracted from the knowledge graph database, and a 100-dimensional feature vector is constructed for each knowledge point, including 20 adjacent node IDs, 50 edge attribute values, and 30 node type one-hot codes. A 3-layer graph attention network is used, each layer contains 5 attention heads, and the meta-path is set to knowledge point-predecessor relationship-knowledge point and knowledge point-related relationship-knowledge point. The learning rate is 0.001, and 100 rounds of training are performed to obtain a 64-dimensional knowledge point vector representation. The learning trajectory of students in the past three months is obtained, including up to 500 visited knowledge point IDs, corresponding residence time in seconds, and number of interactions, combined with portrait information such as student age, major one-hot coding, and learning ability score (1-10 points). A 2-layer long short-term memory network is used, with a hidden layer size of 128 and a dropout rate of 0.5. The learning sequence is padded or truncated to a length of 500, and a 128-dimensional student interest state vector is output. The 64-dimensional knowledge point vector and the 128-dimensional interest vector are concatenated and input into a 3-layer multilayer perceptron with layer sizes of 256, 128, and 1, respectively. One million student access records from the past 6 months are used for training, with a batch size of 64, a learning rate of 0.0005, and 200 rounds of training. The access probability of 10,000 knowledge points is calculated, and the top 100 are selected as the candidate recommendation set. The difficulty range of the knowledge points is set to 1-5, and the current learning progress is the percentage of the number of completed knowledge points to the total knowledge points. Based on the 500,000 relationships between 5,000 knowledge points defined by experts, the Kahn algorithm is used to generate a topological sort as the preset learning path. The weighted sum method is used to sort the candidate knowledge points, with the weights being: difficulty adaptation (the inverse of the difference between the current difficulty and the recommended difficulty) 0.3, learning progress relevance (the inverse of the distance in the preset path) 0.4, and predicted access probability 0.3. Finally, 10 recommended knowledge points are generated, each containing the knowledge point ID, the corresponding learning resource URL, and the recommendation reason based on the weight contribution.

[0027] S3. Continuously monitor students’ learning behavior according to the recommendation list and record the time students spend on knowledge points. If the time students spend on a certain knowledge point exceeds the preset threshold, the graph traversal algorithm is triggered, and a bidirectional breadth-first search is used to traverse the knowledge graph forward and backward at the same time to quickly locate related preceding and subsequent knowledge points and build a local knowledge network.

[0028] Specifically, receive the student learning behavior data carrying the student ID, the current learning knowledge point ID and the start learning timestamp in the Kafka data stream; use the time window function to calculate the student's cumulative stay time at each knowledge point based on the student learning behavior data; store the cumulative stay time in the Redis distributed cache; when it is detected that the student ID or knowledge point ID has changed, reset the timer and write the cumulative time of the previous knowledge point into Redis. Set a knowledge point residence time threshold to determine whether the cumulative residence time exceeds the preset threshold. If the cumulative residence time exceeds the preset threshold, the Cypher query statement of the Neo4j graph database is triggered to implement the graph traversal function, taking the current knowledge point as the starting node, and traversing two layers of relationships forward and backward respectively; obtain the query results returned by the Neo4j graph database, and extract the traversed knowledge point ID list from the query results; calculate the knowledge point relationship strength based on the knowledge point ID list, determine the degree of association between knowledge points based on co-occurrence frequency or predefined weights, and select knowledge points with relationship strength greater than the preset threshold to build a local knowledge network; store the local knowledge network in the graph database; update the student's recommendation list based on the local knowledge network, and for the knowledge points in the local knowledge network that already exist in the original recommendation list, update their positions instead of adding them repeatedly; sort the knowledge points in the local knowledge network from high to low according to the relationship strength, and insert them to the front of the original recommendation list to keep the total length of the recommendation list unchanged.

[0029] For example, Kafka is configured with three partitions, each of which receives a type of data: student ID, knowledge point ID, and timestamp. A 5-minute sliding time window is used to calculate the cumulative residence time every 30 seconds. Redis uses a Hash structure for storage, with the key being student: knowledge point ID and the value being the cumulative time in seconds. When a student switches a knowledge point, the cumulative time of the old knowledge point is written to Redis, and the timer of the new knowledge point is reset to 0. The knowledge point residence time threshold is set to 15 minutes. Neo4j query uses the apoc.path.subgraphNodes function of the APOC library, with the maxLevel parameter set to 2 and the relationshipFilter set to PRE_REQUISITE|NEXT_TOPIC. The co-occurrence frequency method is used to calculate the relationship strength of knowledge points. For two adjacent knowledge points A and B, the relationship strength = (the number of learning sequences in which A and B co-occur) / (the total number of learning sequences in which A appears), and the threshold is set to 0.3. In the local knowledge network, edges with a relationship strength greater than 0.3 are retained. When updating the recommendation list, the double pointer method is used for deduplication, and the time complexity is O(n). New knowledge points are inserted at the front of the original list in descending order of relationship strength, keeping the total length of the list at 20. During the whole process, Kafka processes 1,000 messages per second, the average response time of Redis is less than 5 milliseconds, the average execution time of Neo4j query is 200 milliseconds, and the update operation of the recommendation list takes an average of 50 milliseconds.

[0030] S4. For the constructed local knowledge network, use graph compression algorithm to reduce the number of nodes and retain important nodes and key paths. According to the students' stay time and access frequency at the knowledge points, determine the weak points of the subject and associate them with the important nodes in the local knowledge network.

[0031] Specifically, the PageRank algorithm is used to calculate the importance score of each node in the local knowledge network to obtain the PageRank score of the node; the quartile method is used to determine the importance threshold, and the quartiles of the PageRank scores of all nodes are calculated based on the PageRank scores, and the first quartile is used as the importance threshold; it is determined whether the PageRank score of the node is lower than the importance threshold, and if it is higher than the importance threshold, the node is marked as an important node to obtain a compressed knowledge network; the shortest path between all node pairs in the compressed knowledge network is calculated based on the Johnson algorithm to obtain the path between the node pairs; the path of all nodes on the path is calculated. The geometric mean of the PageRank score is used to obtain the path importance value; determine whether the path importance value is higher than the average value. If it is higher than the average value, the path is marked as a critical path, and the nodes on the critical path and the edges directly connected to it are retained to form a streamlined local knowledge network; obtain the student's residence time and access frequency data on each knowledge point from the learning behavior database; calculate the knowledge point mastery score based on the residence time, access frequency, and expected learning time. The score calculation formula is: Mastery = Access frequency × 0.4 + (residence time / expected learning time) × 0.6, where the expected learning time is based on the difficulty level and content volume of the knowledge point, and is predicted using a linear regression model. Determine whether the mastery score is lower than the preset threshold. If it is lower than the preset threshold, the knowledge point is marked as a subject weakness. Match the weak points of the discipline with the nodes in the streamlined local knowledge network; determine whether the weak points of the discipline exist in the streamlined local knowledge network. If so, increase the weight of the node; if not, add the weak points of the discipline as new nodes to the streamlined local knowledge network; calculate the cosine similarity between the new node and the existing nodes in the streamlined local knowledge network; determine whether the cosine similarity is higher than the similarity threshold. If so, establish an association relationship between the new node and the existing node to obtain an updated network.

[0032] Exemplarily, the PageRank algorithm uses a damping factor of 0.85 and iterates 20 times to calculate the importance of the node. In the quartile method, assuming that the score distribution is: Q1 = 0.2, Q2 = 0.5, Q3 = 0.8, 0.2 is used as the importance threshold. The Johnson algorithm runs on a graph represented by an adjacency matrix with a time complexity of O(N^2logN+NE), where N is the number of nodes and E is the number of edges. In the calculation of path importance, if the path contains 5 nodes and the PageRank scores are 0.3, 0.5, 0.4, 0.6, and 0.2 respectively, the geometric mean is (0.30.50.40.60.2)^(1 / 5)≈0.37. In the calculation of knowledge point mastery, the expected learning time is predicted by a linear regression model. The independent variables include difficulty levels such as 1-5 and content volume (unit: thousands of characters). The regression equation is: expected time = 10+5 difficulty + 2 content volume (unit: minutes). The mastery threshold is set to 0.6, and values ​​below this value are marked as weak points. When adding new nodes, use TF-IDF to vectorize the knowledge point content, calculate the cosine similarity, and set the threshold to 0.7. If the similarity between the new node quadratic function and the existing node quadratic equation is 0.75, an association is established; if the similarity with cell division is 0.1, no association is established. The updated network is stored in the Neo4j graph database and operated using the Cypher query language.

[0033] S5. Obtain the detailed attributes of knowledge point coverage, correlation strength, mastery, number of exercises, error frequency and difficulty coefficient of the subject's weak points and key paths, and calculate the comprehensive correlation through the attribute fusion algorithm. If the comprehensive correlation exceeds the preset threshold, it is judged that the subject's weak points and key paths have a strong connection.

[0034] The knowledge point data of the weak points and key paths of the subject in the knowledge graph database are obtained. The knowledge point data include coverage, correlation strength, mastery level, number of exercises, error frequency and difficulty coefficient; the weight coefficient of each attribute is determined by the hierarchical analysis method according to the knowledge point data, and each attribute value is mapped to the range of 0-1 through weighted normalization processing; the t-distribution random neighborhood embedding method is used to perform nonlinear dimensionality reduction processing on the weighted normalized attribute data, and the main eigenvectors are extracted and the eigenvalues ​​are calculated; if the cumulative contribution rate reaches the preset threshold, the first N features are selected as the basis for calculating the comprehensive correlation; according to the selected features, a multi-layer perceptron network is constructed as the attribute fusion algorithm. The number of nodes in the input layer of the multi-layer perceptron network is N, and the hidden layer structure is 64 nodes-32 nodes-16 nodes. The ReLU activation function is used, and the output layer is a single node to represent the comprehensive correlation. The network is trained using sample data automatically generated by expert annotation and clustering algorithm, and the model performance is evaluated through cross-validation. The grid search method is used to optimize the hyperparameters to obtain a comprehensive correlation calculation model. Input the attribute data of the knowledge points of the subject's weak points and key paths into the trained multi-layer perceptron network to obtain the comprehensive correlation value; determine whether the comprehensive correlation value is greater than the preset correlation threshold; if the comprehensive correlation value is greater than the preset correlation threshold, it is determined that the subject's weak points and key paths have a strong correlation; store the judgment results in the knowledge graph database. For newly emerging attributes or knowledge points, the incremental learning method is used to dynamically update the model to ensure model adaptability.

[0035] Exemplarily, data was extracted from the Neo4j knowledge graph database, including 100 subject weaknesses and 200 key path knowledge points. The attribute weights were determined using the hierarchical analysis method: coverage 0.2, association strength 0.25, mastery 0.15, number of exercises 0.1, error frequency 0.2, and difficulty coefficient 0.1. After weighted normalization, the t-SNE algorithm was applied for dimensionality reduction, with the perplexity set to 30, the learning rate set to 200, and 2000 iterations to obtain a 2-dimensional feature space. The top 10 features with a cumulative contribution rate of 85% were selected as input. A multilayer perceptron network was constructed with 10 nodes in the input layer, 64-32-16 nodes in the hidden layer, and 1 node in the output layer. 5000 expert-annotated data and 10,000 samples automatically generated by K-means clustering (K=5) were used for training. A 5-fold cross validation was used, with 100 rounds of training per fold, a batch size of 64, and a learning rate of 0.001. Grid search was used to optimize hyperparameters, with the search range: number of hidden layer nodes [32, 64, 128], and learning rate [0.0001, 0.001, 0.01]. The final model achieved an F1 score of 0.92 on the test set. For new attributes, such as video viewing time, elastic network regression was used to estimate the initial weight of 0.08, and the model was updated every 100 new samples through online learning methods. The comprehensive correlation threshold was set to 0.7, and the judgment results showed that 80% of the subject weaknesses had strong connections with the key paths, which were updated to the Neo4j database, with an average query response time of 50ms.

[0036] S5 also includes extracting the coverage and correlation strength of each knowledge point from the learning behavior data stream of the education platform, obtaining the mastery level and number of exercises of the knowledge points based on the students' historical learning trajectory data, analyzing the students' performance on related questions, obtaining the error frequency of the knowledge points, and determining the difficulty coefficient of each knowledge point based on the complexity index in the knowledge graph. These attributes are input into the weighted sum model to assign weights to each attribute in the weighted sum model to calculate the comprehensive correlation.

[0037] Specifically, the learning behavior data stream is obtained, and the coverage of knowledge points is extracted using text analysis tools. The association rule mining algorithm is used to calculate the association strength between knowledge points based on the extracted knowledge point coverage. Relevant records are queried from the student's historical learning trajectory database. The Bayesian knowledge tracking model is used to analyze the student's mastery of each knowledge point. The number of exercises for each knowledge point is counted to obtain the student's answer records on related questions. The error frequency of each knowledge point is calculated to construct a domain knowledge graph, and the hierarchical relationship and dependency relationship between knowledge points are defined. Based on this, the complexity index of each knowledge point is determined as the difficulty coefficient. The knowledge point coverage, association strength, mastery, number of exercises, error frequency and difficulty coefficient are input into the weighted sum model. Weights are assigned to each attribute in the weighted sum model. The comprehensive association score of the knowledge point is calculated. The knowledge points are sorted according to the comprehensive association score, and personalized knowledge point learning sequence suggestions are generated. The learning sequence suggestions are compared with the student's current learning progress. Develop a targeted learning plan. Continuously collect students' learning behavior data during the learning process, regularly update knowledge point attributes and comprehensive associations, and dynamically adjust the learning plan.

[0038] Exemplarily, in the process of acquiring the learning behavior data stream, a data collection module can be embedded in the learning platform to record students' clicks, browsing, answering and other operations at intervals of 5 seconds to form a structured behavior log. Use natural language processing tools to perform text analysis on learning materials, extract keywords through the TF-IDF algorithm, and then combine with the domain knowledge graph to determine the knowledge points covered by each learning material. Use the Apriori association rule mining algorithm, set the minimum support to 05 and the minimum confidence to 8, mine the association rules between knowledge points, and calculate the Lift value to represent the association strength. Extract the practice records of each student on the relevant knowledge points from the historical learning database, use the Bayesian knowledge tracking model BKT, set the mastery threshold to 8, and infer the students' mastery of each knowledge point. Count the number of exercises and error rate of each knowledge point, and mark the knowledge points with an error rate of more than 3 as weak knowledge points. Combined with the domain knowledge graph, consider the prerequisite relationship and semantic distance of the knowledge points, and calculate the importance of each knowledge point through the PageRank algorithm as the difficulty coefficient. In the weighted sum model, set the knowledge point coverage weight 2, the association strength weight 1, the mastery weight 3, the number of exercises weight 1, the error rate weight 2, and the difficulty coefficient weight 1 to calculate the comprehensive association score. Sort the comprehensive association scores, generate a personalized learning sequence, compare it with the student's current knowledge mastery progress, formulate a key reinforcement plan for weak knowledge points, and formulate a consolidation and improvement plan for the mastered knowledge points. In the subsequent learning process, continue to obtain learning behavior data, update the knowledge point attributes and comprehensive association once a week, and use the reinforcement learning algorithm DQN to dynamically adjust the learning plan and optimize the learning path.

[0039] S6. Dynamically adjust the priority of the recommendation list based on the comprehensive correlation between the subject weaknesses and the key paths. Increase the priority of the key path knowledge points that are highly correlated with the subject weaknesses.

[0040] Obtain the comprehensive correlation data of the weak points of the subject and the key path in the knowledge graph database, and filter out the knowledge points that are higher than the threshold according to the preset correlation threshold; use the heap sort algorithm to sort the filtered knowledge points, and obtain a list of knowledge points arranged from high to low according to the comprehensive correlation; design a comprehensive score formula: score = comprehensive correlation * 0.5 + (1-difficulty coefficient) * 0.3 + learning progress * 0.2, and calculate the score of each knowledge point in the knowledge point list according to the pre-established comprehensive score formula. The comprehensive score formula includes comprehensive correlation, difficulty coefficient and learning progress; determine whether the knowledge point score is higher than the preset threshold. If it is higher than the preset threshold, adjust the priority of the knowledge point according to the score. Use a user-based collaborative filtering algorithm to generate a personalized recommendation list. The collaborative filtering algorithm determines the similarity between students through cosine similarity calculation; select the learning records of the 10 students with the highest similarity as the basis for recommendation. The TF-IDF algorithm is used to vectorize the knowledge point content, calculate the similarity between knowledge points, and solve the cold start problem of the personalized recommendation list; the personalized recommendation list is stored in the distributed cache system, and the data expiration time of the distributed cache system is set; the signal that the student has completed the learning of the knowledge point is received, the local update is triggered, and the recommendation items related to the completed knowledge point are recalculated.

[0041] Exemplarily, the comprehensive correlation data of 1000 knowledge points are extracted from the Neo4j knowledge graph database, and a threshold of 0.7 is set to screen out 734 highly correlated knowledge points. Using the heap sort algorithm, the time complexity is O(nlogn), and the 734 knowledge points are sorted, which takes about 0.05 seconds. The comprehensive score formula is then applied, where the difficulty coefficient ranges from 0-1 and the learning progress ranges from 0-100%. If the comprehensive correlation of knowledge point A is 0.85, the difficulty coefficient is 0.6, and the learning progress is 40%, then its score is 0.85*0.5+(1-0.6)0.3+0.11=0.655. In the user-based collaborative filtering algorithm, the cosine similarity is used to calculate the similarity between 10,000 students, and the 10 students with the highest similarity are selected, and the average calculation time is 0.2 seconds. The TF-IDF algorithm processes the content of 1,000 knowledge points, generates a 1,000-dimensional vector, and calculates the cosine similarity between knowledge points, which takes about 1 second. The recommendation list is stored in the Redis distributed cache, and the expiration time is set to 3,600 seconds. After the students complete the knowledge point learning, a local update is triggered, which takes an average of 0.1 seconds. The click-through rate and completion rate are calculated daily. If the click-through rate is 25% and the completion rate is 80%, the collaborative filtering weight is adjusted to 0.6 and the content-based recommendation weight is adjusted to 0.4. Throughout the process, the average response time of Neo4j queries is 50ms, and the average delay of Redis read and write operations is less than 5ms, ensuring the real-time and high efficiency of the recommendation system.

[0042] S7. According to the priority of the recommendation list, a multi-level cache strategy is used to pre-cache learning resources, maintain hot knowledge point indexes in memory, store secondary hot resources in solid-state drives, store cold data in mechanical hard drives, and dynamically adjust cache contents at all levels through cache replacement algorithms.

[0043] Specifically, knowledge points are divided into three categories: hotspots, secondary hotspots, and cold data according to their priorities. The priority is determined by the priority of the knowledge points in the recommendation list. The priority threshold of hotspot data is set to 0.8, and the priority threshold of secondary hotspot data is set to 0.5. Data below 0.5 is considered cold data. A hash table is built in memory to store the hotspot knowledge point index. The key of the hash table is the knowledge point ID, and the value is the resource storage location. The hotspot resources are searched through the hash table, and the content of the hotspot resources is loaded into the memory cache. At the same time, a cache preheating mechanism is implemented to preload the top 100 high-priority hotspot resources at startup. The LRU-K algorithm is used to manage secondary hotspot resources. The K value of the LRU-K algorithm is dynamically adjusted according to the average access interval of the resource, ranging from 2 to 5. The timestamp of the most recent K accesses of each resource is recorded, and the priority of the resource is calculated based on the timestamp. When the cache space is insufficient, the lowest priority resource is eliminated first. A distributed file system is used to store cold data. The consistent hashing algorithm is used to evenly distribute resources to multiple storage nodes. The cold data storage location is located in combination with the prefix tree index. The sliding time window count is used to record the number of accesses to each resource in the past 24 hours. Set a dynamic promotion threshold with an initial value of 100 times / day. It is automatically adjusted every hour according to the overall system load. When the access frequency exceeds the threshold, the resource is promoted to the secondary hotspot or hotspot cache. Perform cache optimization tasks regularly and check the cache hit rate every hour. If the cache hit rate is less than 85%, the cache adjustment is triggered. According to the access frequency and priority of the resource, data is migrated between different cache levels to ensure cache efficiency.

[0044] For example, the recommendation list contains 10,000 knowledge points, and the priority queue is implemented using a max-heap with a time complexity of O(logn). The initial division results are: 2,000 hot spots (priority>0.8), 3,000 secondary hot spots (0.5<priority≤0.8), and 5,000 cold data (priority≤0.5). The hash table in memory uses open addressing to resolve conflicts, the load factor is set to 0.75, and the average search time is O(1). The cache preheats and loads the first 100 hot resources, which takes about 2 seconds. In the LRU-K algorithm, the K value is initially set to 3 and is adjusted once an hour according to the average access interval. If the average access interval of a resource is 10 minutes, the K value is adjusted to 2; if it is 2 hours, it is adjusted to 4. The cache capacity of the solid-state drive is 100GB, and elimination is triggered when the usage rate exceeds 90%. The distributed file system uses 5 storage nodes, and the resources are stored in a distributed manner through consistent hashing. The number of virtual nodes is set to 200 to reduce uneven data distribution. The prefix tree index depth is limited to 5 layers, and the average search time is <10ms. The sliding time window size is 24 hours, and the count is updated every 5 minutes. The initial value of the dynamic increase threshold is 100 times / day, and it is adjusted according to the CPU utilization. When the CPU utilization exceeds 80%, the threshold is increased by 20%. The cache optimization task runs once an hour. When the hit rate is detected to be lower than 85%, the second hot spot resources with the top 10% of the access frequency are promoted to hot spots, and the hot spot resources with the bottom 10% are demoted to secondary hot spots to achieve dynamic balance. The average response time of the entire multi-level cache system is <50ms, and the cache hit rate is maintained above 90%.

[0045] S8. According to the set weight evaluation system, the cache content including the importance of knowledge points, access frequency, knowledge link integrity, and subject weakness correlation factors are calculated to obtain the weight score of each cache content. When the cache space is insufficient, the cache content with the lowest weight is eliminated.

[0046] Obtain the attribute data of the cached content in the knowledge graph database, which includes the importance of knowledge points, access frequency, knowledge link integrity and the correlation of subject weaknesses; normalize the attribute data to obtain normalized attribute values ​​mapped to the 0-1 interval; set them to 0.3, 0.3, 0.2, and 0.2 respectively according to the normalized attribute values ​​and the preset weight coefficients. Based on the association relationship in the knowledge graph, calculate the proportion of knowledge points covered by the current cached content in the entire knowledge link as a quantitative indicator of knowledge link integrity. Calculate the score value of the cached content, and use a linear weighted calculation method to score each content in the cache. The scoring formula is: score = Σ (attribute value × weight coefficient). Store the scoring results in the priority queue, and the queue is sorted from high to low by score. Set the score update cycle to 1 hour, or trigger an update when the access frequency changes by more than 20% to ensure the real-time nature of the score, monitor the cache space usage, set the initial usage rate threshold to 85%, and dynamically adjust it once an hour according to the load, with an adjustment range of 80%-90%. Determine whether the cache space usage exceeds the preset threshold. If it exceeds the preset threshold, the cache content with the lowest score is taken out from the priority queue for elimination. Each time, the 5% of the content with the lowest score is taken out from the priority queue for elimination until the cache space meets the requirements. The metadata of the eliminated content is saved in persistent storage for possible subsequent recovery.

[0047] For example, 10,000 knowledge points are stored in the Neo4j knowledge graph database, and the cache content contains 2,000 hot knowledge points. The min-max method is used for attribute normalization. For example, the original score range of the importance of knowledge points is 0-100, which becomes 0-1 after normalization. Knowledge link integrity calculation method: Assuming that a complete knowledge link contains 20 knowledge points, and the current cache covers 15 of them, the integrity is 75%. Linear weighted calculation is implemented using vector dot multiplication, and the calculation speed is about 1,000 / second. The priority queue is implemented using a binary heap, and the time complexity of insertion and deletion operations is O(logn). Score update trigger condition: The access frequency changes by more than 20%, which is achieved by maintaining an access counter and reset once an hour. The proportional-integral-differential (PID) control algorithm is used for dynamic adjustment of cache space utilization, and adjustments are made according to CPU utilization. The adjustment cycle is 1 hour, and the initial P=0.5, I=0.1, and D=0.05. Gradual elimination deletes 5% of the lowest-scoring content each time, about 100 knowledge points, to avoid sudden large-scale elimination. The eliminated metadata (including knowledge point ID, last access time, and score) is stored in Redis, and the records of the last 24 hours are retained. The entire scoring and elimination process takes an average of 200ms, and the cache hit rate is increased from the original 85% to 92%.

[0048] S8 also includes determining the importance of each knowledge point based on the knowledge graph, obtaining the access frequency of each knowledge point based on the student learning behavior data stream, and evaluating the knowledge link integrity of the knowledge point based on the local knowledge network. Using the comprehensive correlation, determine the degree of correlation between the knowledge point and the weak points of the subject, input the factors into the pre-set weight evaluation system, and calculate the weight score of each cached content.

[0049] According to the pre-constructed knowledge graph, the importance score of each knowledge point is calculated through the connection relationship and node degree of the knowledge points in the graph. The access behavior log data of students to each knowledge point on the learning platform is obtained, and the access frequency of each knowledge point is counted. For each knowledge point, the predecessor knowledge points and successor knowledge points related to it are retrieved in the local knowledge network to determine whether the knowledge link of the knowledge point is complete. If the predecessor or successor knowledge point is missing, the knowledge link integrity score is reduced. The association rule mining algorithm is used to mine the association rules between the students' subject weaknesses and the comprehensive behavior data of the knowledge points, and the correlation between each knowledge point and the subject weaknesses is calculated. Multiple factors such as the importance of knowledge points, access frequency, link integrity and correlation of weak points are input into the pre-set weighted scoring model, and different weights are assigned to each factor. The comprehensive weight score of each knowledge point is obtained by weighted summation.

[0050] Exemplarily, in the knowledge graph, the importance of each knowledge point is measured by calculating the PageRank value. The PageRank algorithm takes into account the in-degree and out-degree of the knowledge point. After the iterative calculation converges, the importance score is obtained, and the score range is 0-1. The students' access behavior data is extracted from the learning platform log, and the data is cleaned and counted using the MapReduce framework to obtain the number of visits to each knowledge point. The access frequency = the number of visits / the total number of visits. Using the Cypher query language of the Neo4j graph database, with the current knowledge point as the center, its predecessor and successor knowledge points are retrieved. If the predecessor or successor knowledge point is missing, the integrity score is reduced proportionally, and the proportional coefficient is set to 8. Using the Apriori association rule mining algorithm, the minimum support is set to 05 and the minimum confidence is set to 5. The association rules of knowledge points and weak points are mined, and the association degree = support × confidence. Finally, the scores of each factor are input into the weighted scoring model, and the weights of importance, access frequency, integrity, and association are set to 4, 3, 2, and 1 respectively. The comprehensive weight score is calculated using vector dot multiplication. The entire process is implemented using the Spark memory computing framework, with 1,000 parallel tasks set up, each task processing 100 knowledge points on average, a single task execution time of about 20 seconds, and a total time consumption of about 330 seconds. Finally, the comprehensive weight score is stored in the Redis cache, with an expiration time of 1 day, for the intelligent recommendation algorithm to call.

[0051] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for accurately locating and breaking through the weak points of a subject based on a knowledge graph, characterized in that: The following steps are involved: S1, obtains students' learning behavior data stream from the education platform and learning system, uses a distributed stream processing framework to process students' learning behavior data stream, calculates the access frequency and residence time of knowledge points, dynamically adjusts the association strength of knowledge points according to the preset weight function, and obtains a real-time updated knowledge graph; S2, based on the updated knowledge graph, uses a deep learning model to predict the knowledge points that students may visit next, captures the multi-dimensional dependencies between knowledge points through a graph neural network, and combines students' historical learning trajectory data to generate a personalized learning resource recommendation list; S3, continuously monitors students’ learning behavior according to the recommendation list, records the time students spend on knowledge points, and triggers the graph traversal algorithm if the time students spend on a certain knowledge point exceeds the preset threshold. It uses a bidirectional breadth-first search to traverse the knowledge graph forward and backward at the same time, quickly locates related preceding and subsequent knowledge points, and constructs a local knowledge network. S4, for the constructed local knowledge network, use the graph compression algorithm to reduce the number of nodes, retain important nodes and key paths, determine the weak points of the subject according to the students' stay time and access frequency at the knowledge points, and associate them with the important nodes in the local knowledge network; S5, obtain the knowledge point data of the subject's weak points and key paths, calculate the comprehensive correlation through attribute fusion, and if the comprehensive correlation exceeds the preset threshold, it is judged that the subject's weak points and key paths have a strong connection; S6, dynamically adjust the priority of the recommendation list according to the comprehensive correlation between the subject weaknesses and the key path, and increase the priority of the knowledge points of the key path that are highly correlated with the subject weaknesses; S7, based on the priority of the recommendation list, adopts a multi-level cache strategy to pre-cache learning resources, maintains hot knowledge point indexes in memory, stores secondary hot resources in solid-state drives, stores cold data in mechanical hard drives, and dynamically adjusts cache content at all levels through cache replacement; S8, according to the set weight evaluation system, calculate the importance of knowledge points in the cached content, access frequency, knowledge link integrity and correlation factors of subject weaknesses to obtain the weight score of each cached content. When the cache space is insufficient, eliminate the cached content with the lowest weight.

2. According to claim 1, a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph is characterized in that: Said S1 comprises: Receive learning behavior data with a student's unique identification number, including access logs, stay time, and interaction records. The learning behavior data is collected in real time by the education platform when students are learning online; Based on the learning behavior data, the data is transmitted to the Flink distributed stream processing framework in real time through the Kafka message queue; The sliding time window function is used to aggregate the data transmitted to Flink to obtain the popularity index of each knowledge point within the preset time window; If the heat index of a knowledge point exceeds the preset threshold, the Apriori algorithm is used to mine the association rules between the knowledge point and other knowledge points, and the mined association rules are stored in the Neo4j graph database. The association relationship between knowledge points is stored in the graph database Neo4j to construct the initial knowledge graph. The DeepWalk algorithm is used to perform random walks on the knowledge graph stored in Neo4j to generate a sequence of knowledge points. The knowledge point sequence is used for the Skip-gram model to learn the vector representation of knowledge points, capture the long-range dependencies between knowledge points, update the association strength between knowledge points, and obtain a real-time updated knowledge graph.

3. According to claim 1, a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph is characterized in that: The S2 comprises: Obtain the node features and edge relationship information of knowledge points in the knowledge graph database, and construct a feature vector containing adjacent nodes, edge attributes and node types for each knowledge point; According to the feature vector, a meta-path-based graph attention network is used to learn the representation of the knowledge point, and a low-dimensional dense vector representation of the knowledge point is obtained; Obtain students’ historical learning trajectory data, including the sequence of knowledge points visited, residence time and number of interactions. Combined with student portrait information, use long short-term memory network to model students’ interests and output students’ current interest state vector. The knowledge point vector representation and the student interest state vector are concatenated and input into the multi-layer perceptron. The multi-layer perceptron uses the cross entropy loss function to optimize the model parameters and calculate the probability of students visiting each knowledge point. If the access probability is greater than the preset threshold, the corresponding knowledge point is added to the candidate recommendation set; According to the difficulty of the knowledge points in the candidate recommendation set, the students' current learning progress and the preset learning path, the candidate knowledge points are sorted using the weighted summation method to generate the final personalized learning resource recommendation list.

4. According to claim 1, a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph is characterized in that: The S3 includes: Receive student learning behavior data in the Kafka data stream that carries the student ID, current learning knowledge point ID, and the start learning timestamp; According to the students' learning behavior data, the time window function is used to calculate the students' cumulative stay time at each knowledge point; Store the accumulated stay time in the Redis distributed cache; Determine whether the cumulative stay time exceeds the preset threshold. If the cumulative stay time exceeds the preset threshold, trigger the Cypher query statement of the Neo4j graph database; Get the query results returned by the Neo4j graph database, and extract the traversed knowledge point ID list from the query results; The knowledge point relationship strength is calculated based on the knowledge point ID list, and knowledge points with relationship strength greater than a preset threshold are selected to construct a local knowledge network.

5. According to claim 1, a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph is characterized in that: The S4 comprises: The PageRank algorithm is used to calculate the importance score of each node in the local knowledge network and obtain the PageRank score of the node; Calculate the quartiles of all nodes’ PageRank scores based on the PageRank scores, and use the first quartile as the importance threshold; Determine whether the PageRank score of the node is lower than the importance threshold. If it is higher than the importance threshold, mark the node as an important node to obtain a compressed knowledge network. The shortest paths between all pairs of nodes in the compressed knowledge network are calculated based on the Johnson algorithm to obtain the paths between the pairs of nodes. Calculate the geometric mean of the PageRank scores of all nodes on the path to get the path importance value; Determine whether the path importance value is higher than the average value. If it is higher than the average value, mark the path as a critical path, retain the nodes on the critical path and the edges directly connected to it, and form a simplified local knowledge network; Obtain the students' stay time and access frequency data at each knowledge point from the learning behavior database; Calculate the knowledge point mastery score based on the residence time, visit frequency and expected learning time; Determine whether the mastery score is below a preset threshold. If so, mark the knowledge point as a subject weakness; Match the weak points of the disciplines with the nodes in the streamlined local knowledge network; Determine whether the weak point of the subject exists in the simplified local knowledge network. If so, increase the weight of the node. If not, add the weak point of the subject as a new node to the simplified local knowledge network. Calculate the cosine similarity between the new node and the existing nodes in the streamlined local knowledge network; It is determined whether the cosine similarity is higher than the similarity threshold. If so, an association relationship is established between the new node and the existing node to obtain an updated network.

6. According to claim 1, a method for accurately locating and breaking through the weak points of a subject based on a knowledge graph is characterized in that: The S5 comprises: Obtain the knowledge point data of the weak points and key paths of the subject in the knowledge graph database. The knowledge point data includes coverage, association strength, mastery level, number of exercises, error frequency and difficulty coefficient; According to the knowledge point data, the weight coefficient of each attribute is determined by using the hierarchical analysis method, and each attribute value is processed by weighted normalization; The weighted normalized attribute data is processed by nonlinear dimensionality reduction using the t-distribution random neighborhood embedding method to extract the main eigenvectors and calculate the eigenvalues. If the cumulative contribution rate reaches the preset threshold, the first N features are selected as the basis for calculating the comprehensive correlation; According to the selected features, a multi-layer perceptron network is constructed as an attribute fusion algorithm, and the number of input layer nodes of the multi-layer perceptron network is N; Input the attribute data of knowledge points of subject weaknesses and key paths into the trained multi-layer perceptron network to obtain the comprehensive correlation value; Determine whether the comprehensive correlation value is greater than a preset correlation threshold; If the comprehensive correlation value is greater than the preset correlation threshold, it is determined that the subject weakness has a strong correlation with the critical path; The judgment results are stored in the knowledge graph database.

7. The method for accurately locating and breaking through the weak points of a subject based on a knowledge graph according to claim 6 is characterized in that: The S5 further includes: Extract the coverage and correlation strength of each knowledge point from the learning behavior data stream of the education platform, obtain the mastery degree and number of practice of the knowledge points based on the students' historical learning trajectory data, analyze the students' performance on related questions, obtain the error frequency of the knowledge points, and determine the difficulty coefficient of each knowledge point based on the complexity index in the knowledge graph. Input these attributes into the weighted sum model, assign weights to each attribute in the weighted sum model, and calculate the comprehensive correlation; Specifically include: Obtain learning behavior data stream, use text analysis tools to extract knowledge point coverage, and use association rule mining algorithms to calculate the association strength between knowledge points based on the extracted knowledge point coverage; Query relevant records from the student history learning trajectory database, use the Bayesian knowledge tracking model to analyze the student's mastery of each knowledge point, count the number of exercises for each knowledge point, and obtain the student's answer records on related questions; Calculate the error frequency of each knowledge point to build a domain knowledge graph, define the hierarchical relationship and dependency relationship between knowledge points, and determine the complexity index of each knowledge point as the difficulty coefficient. Input the knowledge point coverage, association strength, mastery level, number of exercises, error frequency and difficulty coefficient into the weighted sum model to assign weights to each attribute in the weighted sum model. Calculate the comprehensive relevance score of the knowledge points.

8. The method for accurately locating and breaking through the weak points of a subject based on a knowledge graph according to claim 1, characterized in that: The S6 comprises: Obtain the comprehensive correlation data between the weak points of the subject and the key paths in the knowledge graph database, and filter out the knowledge points above the threshold according to the preset correlation threshold; Sort the selected knowledge points to obtain a list of knowledge points arranged from high to low in terms of comprehensive relevance; Calculate the score of each knowledge point in the knowledge point list according to a pre-established comprehensive score formula, which includes comprehensive relevance, difficulty coefficient and learning progress; Determine whether the knowledge point score is higher than a preset threshold, and if so, adjust the priority of the knowledge point.

9. The method for accurately locating and breaking through the weak points of a subject based on a knowledge graph according to claim 1, characterized in that: The S7 comprises: According to the priority of knowledge points, they are divided into three categories: hot spots, secondary hot spots, and cold data. The priority is determined by the priority of the knowledge points in the recommendation list; Construct a hash table to store hot knowledge point indexes. The key of the hash table is the knowledge point ID, and the value is the resource storage location. Search for hot resources through the hash table and load the hot resource content into the memory cache; The LRU-K algorithm is used to manage secondary hotspot resources. The K value of the LRU-K algorithm is dynamically adjusted according to the average access interval of the resources. Record the timestamps of the most recent K accesses to each resource and calculate the priority of the resource based on the timestamps; Use a distributed file system to store cold data and evenly distribute resources to multiple storage nodes through a consistent hashing algorithm; Combine prefix tree index to locate cold data storage location, and use sliding time window count to record the number of times each resource has been accessed in the past few hours; Perform cache optimization tasks and check cache hit ratio; If the cache hit rate is lower than the preset value, cache adjustment is triggered to migrate data between different cache levels based on the access frequency and priority of the resource.

10. The method for accurately locating and breaking through the weak points of a subject based on a knowledge graph according to claim 9, characterized in that: The S8 comprises: Obtain attribute data of cached content in the knowledge graph database, including the importance of knowledge points, access frequency, knowledge link integrity, and correlation of subject weaknesses; Normalize the attribute data to obtain normalized attribute values; Calculate the score of the cached content based on the normalized attribute value and the preset weight coefficient; The score values ​​are stored in a priority queue, and the priority queue is sorted from high to low according to the score values; Determine whether the cache space usage exceeds a preset threshold. If so, remove the cache content with the lowest score from the priority queue for elimination. Among them, the preset threshold is dynamically adjusted according to the load; The S8 further includes: According to the knowledge graph, the importance of each knowledge point is determined. Based on the student learning behavior data stream, the access frequency of each knowledge point is obtained. According to the local knowledge network, the knowledge link integrity of the knowledge point is evaluated. The degree of association between the knowledge point and the weak points of the subject is determined by using the comprehensive correlation. The factors are input into the pre-set weight evaluation system to calculate the weight score of each cached content. Specifically include: According to the pre-built knowledge graph, the importance score of each knowledge point is calculated through the connection relationship and node degree index of the knowledge points in the graph; Obtain students' access behavior log data for each knowledge point on the learning platform and count the access frequency of each knowledge point; For each knowledge point, the predecessor knowledge point and the successor knowledge point related to it are retrieved in the local knowledge network to determine whether the knowledge link of the knowledge point is complete. If the predecessor or successor knowledge point is missing, the knowledge link integrity score is reduced; Adopting association rule mining algorithm, we can mine the association rules between students’ subject weaknesses and knowledge points from their comprehensive behavior data, and calculate the correlation between each knowledge point and the subject weaknesses. Multiple factors such as the importance of knowledge points, access frequency, link integrity and weak point correlation are input into a pre-set weighted scoring model, and different weights are assigned to each factor. The comprehensive weight score of each knowledge point is obtained through weighted summation.

Images