Children cognition-based content recommendation method and system

Abstract

The invention is suitable for the technical field of education, and provides a content recommendation method and system based on children cognition, and the method comprises the steps: obtaining historical learning data, carrying out the construction of a knowledge graph based on the historical learning data, and generating a knowledge structure graph; performing vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes by using a graph embedding algorithm to generate a knowledge vector set; according to the knowledge vector set, calculating a semantic association degree among the plurality of knowledge nodes, and if the semantic association degree is greater than a preset threshold value, extracting a key knowledge node from the plurality of knowledge nodes by using a hierarchical clustering algorithm; determining a learning sequence corresponding to the key knowledge nodes according to a preset knowledge learning strategy, and generating a knowledge progressive sequence; generating a target knowledge node sequence based on the knowledge progressive sequence and the historical learning data by using a preset reinforcement learning algorithm; and the target recommendation content is generated according to the target knowledge node sequence and recommended to the user, so that more appropriate learning content can be recommended.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education, in particular to a content recommendation method and system based on children's cognition. BACKGROUND

[0002] The continuous progress of artificial intelligence and machine learning technology provides strong technical support for content recommendation in the field of children's education. These technologies can process and analyze a large amount of data, such as children's learning behavior data, content data, etc., to achieve more accurate recommendations.

[0003] Traditional recommendation methods usually rely on simple interest tags to recommend content. For example, if a child is interested in "animals", all learning content related to "animals" is recommended; or behavior data-based recommendations, i.e. analyzing children's behavior data such as video viewing duration, number of completed exercises, etc. to recommend similar content. For example, if a child often watches math animations, more math animations are recommended.

[0004] However, the traditional method only recommends based on interest tags or behavior data, which is difficult to accurately match children's cognitive ability and knowledge structure, resulting in recommended content that may not meet the actual learning situation of children. Therefore, how to recommend more suitable learning content to users has become a technical problem to be solved. SUMMARY

[0005] The embodiments of the present application provide a content recommendation method and system based on children's cognition, which can solve the technical problem of how to recommend more suitable learning content to users.

[0006] In a first aspect, the embodiments of the present application provide a content recommendation method based on children's cognition, comprising: obtaining historical learning data and constructing a knowledge graph based on the historical learning data to generate a knowledge structure graph, the knowledge structure graph including a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes; performing vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes using a graph embedding algorithm to generate a knowledge vector set; calculating semantic correlation between the plurality of knowledge nodes according to the knowledge vector set, and if the semantic correlation is greater than a preset threshold, extracting key knowledge nodes from the plurality of knowledge nodes using a hierarchical clustering algorithm; determining a learning order corresponding to the key knowledge nodes according to a preset knowledge learning strategy to generate a knowledge progression sequence; generating a target knowledge node sequence based on the knowledge progression sequence and the historical learning data using a preset reinforcement learning algorithm; generating target recommendation content according to the target knowledge node sequence and recommending the target recommendation content to the user.

[0007] In some embodiments, calculating the semantic relevance between multiple knowledge nodes based on the knowledge vector set includes: Calculate the cosine similarity between each vector in the knowledge vector set; Divide the vectors whose cosine similarity exceeds the preset similarity threshold into a group to obtain a semantic association set; The weighted average method is used to calculate the similarity weighted value of each group of vectors in the semantic association set to obtain the semantic association. The weight value corresponding to the similarity of each group of vectors is positively correlated with the frequency of common occurrence of each group of vectors.

[0008] In some embodiments, if the semantic relevance is greater than a preset threshold, a hierarchical clustering algorithm is used to extract key knowledge nodes from multiple knowledge nodes, including: Mark the knowledge nodes corresponding to semantic relevance greater than a preset threshold as relevant knowledge nodes, and obtain a set of relevant knowledge nodes; Using a hierarchical clustering algorithm to group the knowledge nodes in the related knowledge node set, a related knowledge node group is obtained; When the relevant knowledge node group includes at least two knowledge nodes, key knowledge nodes are extracted from the at least two knowledge nodes according to the appearance frequencies of the knowledge nodes.

[0009] In some embodiments, the learning order corresponding to the key knowledge nodes is determined according to a preset knowledge learning strategy to generate a knowledge progression sequence, including: Determine the directional relationship between key knowledge nodes based on the preset learning strategy; Generate a directed graph based on the directional relationship between key knowledge nodes. The edges in the directed graph include edge weights, and the edge weights are positively correlated with the frequency of co-occurrence of the corresponding key knowledge nodes. The sequence generation algorithm is used to sort the key knowledge nodes based on the numerical value of the edge weight, determine the learning order corresponding to the key knowledge nodes, and generate a knowledge progression sequence.

[0010] In some embodiments, a target knowledge node sequence is generated based on a knowledge progression sequence and historical learning data using a preset reinforcement learning algorithm, including: Determine the user's current learning progress based on historical learning data and generate an initial knowledge node sequence; Obtain user learning feedback data, including correct answer rate and learning time; Combining learning feedback data and knowledge progression sequence, the reinforcement learning algorithm is used to adjust the initial knowledge node sequence to generate the target knowledge node sequence.

[0011] In some embodiments, when the target knowledge node sequence includes knowledge nodes other than the initial knowledge node sequence, the method further includes: marking the knowledge nodes other than the initial knowledge node sequence as extended knowledge nodes; converting the extended knowledge nodes into extended knowledge vectors by using a vector generation tool, the extended knowledge vectors comprising a learning difficulty feature and a learning correlation degree feature; calculating a vector value of the extended knowledge vector, and if the vector value exceeds a preset vector threshold, adjusting the learning difficulty feature and / or the learning correlation degree feature to obtain an optimized extended knowledge vector; adjusting the initial knowledge node sequence according to the optimized extended knowledge vector to generate a target knowledge node sequence.

[0012] In some embodiments, the method further comprises: calculating an initial attention degree of the knowledge nodes according to historical learning data; calculating an appropriateness score of the knowledge nodes in different age groups according to a preset knowledge node appropriateness score rule, and if the appropriateness score is less than a preset score threshold, reducing the weight of the initial attention degree of the knowledge nodes to obtain an optimized attention degree; based on the optimized attention degree, calculating an appropriateness weight of each knowledge node in different age groups in combination with the age of the user to obtain a knowledge node appropriateness weight matrix in different age groups; performing clustering processing on the knowledge node appropriateness weight matrix in different age groups by using a K-means clustering algorithm to generate an optimized extended direction vector; adjusting the initial knowledge node sequence according to the optimized extended knowledge vector to generate a target knowledge node sequence.

[0013] In some embodiments, the target recommendation content is generated according to the target knowledge node sequence, comprising: calculating a semantic matching degree between the target knowledge node sequence and a candidate recommendation content in a preset content library; calculating an association weight between the candidate recommendation content and the target knowledge node according to the semantic matching degree; calculating an appropriateness weight of the candidate recommendation content in different age groups according to a preset knowledge node appropriateness score rule if the association weight exceeds a preset weight threshold; performing weighted calculation on the candidate recommendation content according to the semantic matching degree, the association weight and the appropriateness weight to obtain a recommendation score corresponding to the candidate recommendation content; sorting the candidate recommendation content according to the numerical value of the recommendation score to obtain sorted recommendation content; determining the target recommendation content according to the sorted recommendation content.

[0014] In some embodiments, the target recommendation content is determined according to the sorted recommendation content, comprising: determining the learning preference of the user according to historical learning data; a content recommendation score corresponding to the ranking recommended content is calculated, the content recommendation score being related to content complexity; an attention mechanism algorithm is used to adjust the content recommendation score in combination with learning preferences to obtain an optimized content recommendation score; the ranking recommended content is re-ranked according to the numerical value of the optimized content recommendation score to obtain optimized recommended content; if the number of contents in the optimized recommended content exceeds a preset number threshold, contents with scores lower than a preset score range are deleted to make the number of contents in the optimized recommended content meet the preset number threshold, and target recommended content is obtained.

[0015] In a second aspect, the embodiments of the present application provide a content recommendation system based on children's cognition, which has the function of implementing the method in the first aspect or any possible implementation manner thereof. Specifically, the system includes units implementing the method in the first aspect or any possible implementation manner thereof.

[0016] In one of the embodiments, the system includes: an acquisition unit configured to acquire historical learning data and construct a knowledge graph based on the historical learning data to generate a knowledge structure graph, the knowledge structure graph including a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes; a mapping unit configured to perform vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes by using a graph embedding algorithm to generate a knowledge vector set; a processing unit configured to calculate semantic correlation degrees between the plurality of knowledge nodes according to the knowledge vector set, and extract key knowledge nodes from the plurality of knowledge nodes by using a hierarchical clustering algorithm if the semantic correlation degrees are greater than a preset threshold; the processing unit is further configured to determine a learning order corresponding to the key knowledge nodes according to a preset knowledge learning strategy to generate a knowledge progression sequence; the processing unit is further configured to generate a target knowledge node sequence based on the knowledge progression sequence and the historical learning data by using a preset reinforcement learning algorithm; the processing unit is further configured to generate target recommended content according to the target knowledge node sequence and recommend the target recommended content to a user.

[0017] In a third aspect, the embodiments of the present application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, when the processor executes the computer program, the computer device implements the method of any one of the implementation manners of the first aspect.

[0018] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, which stores a computer program. The computer program is executed by a computer device, so that the computer device implements the method in any one of the implementation manners of the first aspect.

[0019] In a fifth aspect, the embodiments of the present application provide a computer program product. When the computer program product is run on a computer device, the computer device executes the method in any one of the implementation manners of the first aspect.

[0020] Compared with the prior art, the embodiments of the present application have the beneficial effects that: the knowledge graph is constructed by analyzing the historical learning data of the user, the knowledge points are abstracted as nodes, and the logical relationship between the knowledge points is abstracted as edges, which is conducive to comprehensively understanding the learning system of the user and avoiding isolated content recommendation; the knowledge nodes (knowledge points) and edges (relationships) are mapped to a low-dimensional vector space, which is conducive to capturing the implicit semantics of the knowledge points; the knowledge nodes with a semantic correlation degree higher than a threshold are screened, the strongly correlated knowledge nodes are grouped through hierarchical clustering, and the key knowledge nodes are formed, which can filter low-correlation knowledge points and focus on core content; a preset learning strategy generates a knowledge progression sequence, and a learning order is formulated for the key knowledge nodes, which can ensure that the recommended content is logically progressive and avoids knowledge gaps; according to the real-time feedback data of the user, a reinforcement learning algorithm is used for dynamic adjustment, real-time adaptation to the change of the user's ability, and avoidance of path rigidity; the target knowledge node sequence is converted into specific learning content (such as learning videos, exercises, etc.), target recommended content is generated, and the target recommended content is recommended to the user, so as to meet the learning needs of the user. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 FIG. 1 is a flowchart of a content recommendation method based on child cognition provided by the embodiments of the present application; Figure 2 FIG. 2 is a structural diagram of a content recommendation system based on child cognition provided by the embodiments of the present application; Figure 3 FIG. 3 is a structural diagram of a computer device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0022] In order to further illustrate the technical solutions of the present application, specific embodiments are used for illustration.

[0023] Figure 1 FIG. 1 is a flowchart of a content recommendation method based on child cognition provided by the embodiments of the present application.

[0024] As shown in FIG. 1, the above method includes the following steps S101-S106. Figure 1

[0025] ​S101, acquire historical learning data, and construct a knowledge graph based on the historical learning data to generate a knowledge structure graph.

[0026] The knowledge structure graph includes a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes.

[0027] It can be understood that the knowledge structure graph is a graphical representation method that presents the knowledge points mastered by the child and their mutual relationships in a structured form by constructing a knowledge graph. One knowledge node can correspond to one knowledge point, and the edges between a plurality of knowledge nodes can be understood as the interactive relationships performed by the child on the plurality of knowledge points.

[0028] As an example but not limitation, the historical learning data of the child is acquired, including interaction records on the learning platform, completed homework, test scores, watched learning videos, etc. The basic information of the child, such as age, grade, learning preference, etc. is also collected. In short, the historical learning data here can be understood as data related to the child's learning, which will not be listed one by one here.

[0029] A graph database is used to store the knowledge graph. Each knowledge point is a node, and the relationship between knowledge points is an edge. For example, nodes representing knowledge points such as "addition" and "subtraction" are created in the database, and edges representing relationships such as "relevance" and "learning order" between them are created. According to the data in the graph database, a knowledge structure graph is generated using graph visualization technology.

[0030] In one implementation, the historical learning data is acquired from the learning data and interaction records of the child user, invalid information is removed using data cleaning techniques, and structured knowledge point sets are obtained by analyzing knowledge points and interactive behaviors using text segmentation and entity recognition techniques.

[0031] For example, when the child completes a math problem on the learning platform, the record may contain irrelevant operations such as page refresh or repeated submission of answers. Data cleaning filters out timestamp anomalies or null records to ensure that only valid test data such as user ID, question ID, and test time is retained. The cleaned data is more accurate and provides a reliable basis for subsequent analysis.

[0032] For the structured knowledge point set, a relationship extraction technique is used to analyze the association between knowledge points. If the knowledge points frequently appear together in the interaction records, edges between the knowledge points are generated, the weights of the edges are determined by calculating the interaction frequency, and a set of edges of the knowledge graph is obtained.

[0033] Specifically, the relationship extraction technique analyzes the association between knowledge points to generate edges and weights through interaction frequency.

[0034] For example, in children's mathematics learning, the knowledge points "addition" and "subtraction" often appear in the same exercise, and the interaction record shows that the frequency of their co-occurrence is 80%. Then a "addition-subtraction" edge is generated, and the weight is set to 0.8. The edge with high weight reflects the strong association between knowledge points, which helps to accurately recommend related learning content.

[0035] According to the knowledge point set and the edge set, the knowledge graph is stored in the graph database, the knowledge points are mapped into nodes by the node generation technology, and the connections between the knowledge points are generated in combination with the edge set to obtain a preliminary knowledge structure; for the preliminary knowledge structure, a personalized knowledge structure graph is generated by using a structure visualization technology.

[0036] Preferably, the graph database stores the knowledge graph, the node generation technology maps the knowledge points into nodes, and the edge set generates the connections. For example, the graph database such as Neo4j stores the nodes of "score", "decimal", and "percentage", and the edges such as "score-decimal", and the edge weight is 0.7. After the preliminary knowledge structure is formed, the relationship between the knowledge points can be quickly queried to support personalized recommendation. The efficient query of the graph database improves the system response speed.

[0037] As an example but not limitation, the nodes and edges are labeled according to the degree of mastery of the knowledge points by children. If the degree of mastery of the child user is higher than a preset threshold, the corresponding nodes and edges are highlighted in the personalized knowledge structure graph to obtain a personalized knowledge structure graph reflecting the degree of mastery. For example, if a child has a good mastery of a certain knowledge point, it is labeled in green; if it has not been mastered, it is labeled in red.

[0038] For example, the structure visualization technology generates a personalized knowledge structure graph, and the nodes and edges are highlighted when the degree of mastery of a child is higher than a threshold. Assuming that the mastery threshold is 70%, and the degree of mastery of a child for "score" is 85%, the "score" node and the "score-decimal" edge are highlighted in green in the graph.

[0039] S102, vector mapping processing is performed on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes by using a graph embedding algorithm to generate a knowledge vector set.

[0040] The graph embedding algorithm is a technology for mapping graph structure data (such as nodes and edges in a knowledge graph) into a low-dimensional vector space while preserving the structure and semantic information in the graph as much as possible.

[0041] In one implementation, the knowledge nodes and topological relationships are obtained from the structure graph data, and the knowledge nodes and topological relationships are cleaned and formatted by using a data preprocessing method to obtain a standardized knowledge node set and a topological relationship matrix. For the knowledge node set and the topological relationship matrix, the node features and connection relationships are mapped by using a graph embedding algorithm to generate a low-dimensional vector representation to obtain a vector representation set.

[0042] For example, when obtaining knowledge nodes and topological relations from the structural diagram data of a children learning platform, the knowledge points in the historical learning data can be regarded as nodes, and the associations in the interaction behaviors can be regarded as topological relations. In the data preprocessing stage, invalid data such as repeated records or irrelevant interactions are cleaned up, the nodes are formatted into uniform fields such as “score operation” and “geometric figure”, and the topological relations are sorted into a matrix to record the interaction frequency between knowledge points.

[0043] For example, the score operation and the geometric figure appear together 10 times in a learning task, and the matrix corresponding element value is 10. Such cleaning and formatting ensure data consistency and facilitate subsequent analysis.

[0044] Specifically, for the knowledge node set and the topological relation matrix, the graph embedding algorithm can map high-dimensional nodes and relations into low-dimensional vectors. In one possible implementation, the Node2Vec algorithm is used to generate node sequences through random walk to capture the topological structure. For example, the score operation node can generate a vector [0.3, -0.5, 0.8], and the geometric figure node can generate a vector [0.4, -0.6, 0.7]. These vectors retain the structural information between nodes and reduce the computational complexity. Such mapping helps to discover hidden semantic associations.

[0045] S103, calculating the semantic association degree between multiple knowledge nodes according to the knowledge vector set, and if the semantic association degree is greater than a preset threshold, extracting key knowledge nodes from the multiple knowledge nodes by using a hierarchical clustering algorithm.

[0046] In some embodiments, calculating the semantic association degree between multiple knowledge nodes according to the knowledge vector set includes: calculating the cosine similarity between vectors in the knowledge vector set; dividing vectors with a cosine similarity greater than a preset similarity threshold into a group to obtain a semantic association set; and calculating the similarity weighted value of each group of vectors in the semantic association set by using a weighted average method to obtain the semantic association degree, wherein the weight value corresponding to the similarity of each group of vectors is positively correlated with the frequency of co-occurrence of each group of vectors.

[0047] In one implementation, according to the vector representation set, the cosine similarity method is used to calculate the distance between vectors (L2 normalization can be performed on each vector before calculation), and if the distance is less than a preset threshold, it is determined that there is a semantic association between the corresponding knowledge nodes, and a semantic association set is obtained. The similarity value of each pair of knowledge nodes is obtained from the semantic association set, the similarity value is quantified by using a weighted average method, and the semantic association degree between knowledge points is determined.

[0048] For example, cosine similarity method can quantify the semantic closeness between nodes when calculating the distance between vectors. Assuming that the score operation and the vector cosine similarity of geometric figures are 0.95, and the preset threshold is 0.9, it is determined that there is a semantic association between the two. This association may reflect that children often involve area calculation of geometric figures when learning score operation. The semantic association set records all node pairs that meet the threshold, such as “score operation-geometric figure” and “algebraic equation-logical reasoning”, and provides a basis for subsequent analysis.

[0049] Specifically, after obtaining the similarity value from the semantic association set, the weighted average method can determine the association strength. Assuming that the similarity between score operation and geometric figure is 0.95, and the co-occurrence frequency is 10 times, the weighted strength value may be 9.5; the similarity between algebraic equation and logical reasoning is 0.85, and the frequency is 5 times, and the strength value is 4.25. This quantitative method considers both semantics and interaction frequency, and accurately reflects the closeness between knowledge points. The strength value can be used to optimize the learning path and preferentially recommend knowledge points with high strength association.

[0050] For example, cleaning and formatting takes into account the fragmented nature of children's interaction records, graph embedding algorithm adapts to the complex association between knowledge points, and cosine similarity and weighted average quantify both semantic and behavioral dimensions. These methods together build a precise knowledge association analysis framework, which helps to generate personalized learning recommendations.

[0051] In some embodiments, if the semantic association degree is greater than a preset threshold, a hierarchical clustering algorithm is used to extract key knowledge nodes from a plurality of knowledge nodes, including: marking the knowledge nodes corresponding to the semantic association degree greater than the preset threshold as related knowledge nodes to obtain a related knowledge node set; grouping the knowledge nodes in the related knowledge node set using a hierarchical clustering algorithm to obtain a related knowledge node group; when the related knowledge node group includes at least two knowledge nodes, extracting a key knowledge node from the at least two knowledge nodes according to the occurrence frequency of the knowledge nodes.

[0052] Assuming that “matrix” is mapped to vector [0.2, 0.5, -0.1], and “vector” is mapped to [0.3, 0.4, 0.2], the similarity value of the two vectors is calculated by cosine similarity to be 0.85. If the preset threshold is 0.8, “matrix” and “vector” are marked as related knowledge points and included in the related knowledge point set. This method captures the internal relationship between knowledge points through the geometric distance of semantic vectors and is suitable for large-scale knowledge point screening.

[0053] In one implementation, when performing hierarchical clustering on the related knowledge point set (before hierarchical clustering, the vectors can be standardized by Z-score), a distance matrix can be constructed based on the Euclidean distance between the vectors.

[0054] For example, the "matrix" "vector" "determinant" form a set, calculate the distance of each pair of knowledge point vectors, and generate a matrix, then group it through a hierarchical clustering algorithm. Suppose the result divides the "matrix" "vector" into a group and the "determinant" into a separate group. Within each group, determine the key knowledge nodes by analyzing the frequency or semantic center of the knowledge points, such as the key knowledge node of the "matrix" group. The advantage of hierarchical clustering is that it can dynamically adjust the grouping granularity to adapt to the complexity of different course knowledge.

[0055] In S104, a learning order corresponding to the key knowledge nodes is determined according to a preset knowledge learning strategy, and a knowledge progressive sequence is generated.

[0056] The preset knowledge learning strategy is a rule that determines the order of key knowledge nodes in the learning process. This strategy usually takes into account factors such as logical dependency of knowledge, difficulty level, etc. For example, the learning order can be determined using the relationships in the knowledge graph. Suppose, in the knowledge graph, if knowledge point A is the prerequisite of knowledge point B, then A should be learned before B.

[0057] As an example but not limitation, the preset knowledge learning strategy can be formulated according to the learning content of a textbook, for example, the learning order corresponding to the key knowledge nodes can be determined according to the official textbook of a certain region and grade. For another example, the preset knowledge learning strategy can also be determined according to the learning plan defined by the teacher or parents. Thus, the learning order corresponding to the key knowledge nodes is determined. The specific settings can be made according to the actual situation, which will not be listed one by one here.

[0058] In some embodiments, determining the learning order corresponding to the key knowledge nodes according to the preset knowledge learning strategy to generate the knowledge progressive sequence includes: determining the directed relationship between the key knowledge nodes according to the preset learning strategy; generating a directed graph according to the directed relationship between the key knowledge nodes, the edges in the directed graph including edge weights, and the edge weights having a positive correlation with the frequency of the corresponding key knowledge nodes appearing together; and sorting the key knowledge nodes based on the numerical value of the edge weights using a sequence generation algorithm to determine the learning order corresponding to the key knowledge nodes and generate the knowledge progressive sequence.

[0059] For example, when constructing a concept relationship graph, a directed graph can be used to represent the logical relationship between key knowledge nodes. In the "linear algebra" scenario, "matrix" points to "vector", indicating that the former is the basis of the latter. The weight of the edge can be determined based on the co-occurrence frequency or semantic similarity. Suppose the edge weight from "matrix" to "vector" is 0.9, indicating a strong progressive relationship. This directed graph clearly shows the dependency order of the knowledge points, facilitating subsequent sequence generation.

[0060] In one implementation, the sequence generation algorithm can determine the knowledge progressive sequence through topological sorting.

[0061] For example, based on the directed graph, the algorithm analyzes the weights and directions of the edges, prioritizing the progressive relationships with high weights. The result can generate a sequence: matrix → vector → determinant. This sequence reflects the logical order of the learning path, ensuring that the knowledge points are presented in a progressive manner.

[0062] For example, if the business involves knowledge recommendation on an educational platform, the above method can generate personalized learning paths. Suppose a student queries "matrix", the system recommends "vector" based on the relevant knowledge point set, and prompts the subsequent learning of "determinant" through the progressive sequence. This way, through semantic association and logical ordering, the coherence of knowledge points is enhanced, making it easier for users to understand complex course content.

[0063] S105, generating a target knowledge node sequence based on the knowledge progression sequence and historical learning data using a preset reinforcement learning algorithm.

[0064] In combination with the relevant principles, it can be understood that the reinforcement learning algorithm learns how to take actions to maximize cumulative rewards through the interaction of the agent (Agent) and the environment (Environment). The agent selects actions based on the current state, and the environment rewards or punishes based on the actions, and the agent updates the strategy based on the reward signal.

[0065] In this embodiment, the knowledge progression sequence is used as a path guide for the agent to learn, and the historical learning data (such as the learner's past learning behavior, mastered knowledge points, etc.) is combined to provide personalized learning states for the agent.

[0066] As an example but not limited to, the state of the agent is defined, including the knowledge points the current learner has mastered, the learning progress, interest preferences, etc. The action of the agent is defined, i.e. selecting the next target knowledge node. The reward function is designed to reward or punish based on the learner's response to the action (selected knowledge point) (such as learning effect, interest change, etc.). The historical learning data is used to train the reinforcement learning model to optimize the agent's strategy so that it can select the optimal next knowledge point based on the current state. In practical applications, the agent uses the trained reinforcement learning model to select the optimal next knowledge point based on the current learner's state, and gradually generates a target knowledge node sequence.

[0067] The reinforcement learning algorithm can dynamically adjust the learning path according to the learner's learning progress and feedback, ensuring that the difficulty of the learning content matches the learner's actual level.

[0068] S106, generating target recommendation content based on the target knowledge node sequence and recommending the target recommendation content to the user.

[0069] It can be understood that the target knowledge node sequence represents a series of knowledge points that the user needs to learn next, which are arranged in a certain logical order to ensure the coherence and progression of learning.

[0070] Each knowledge point in the target knowledge node sequence is matched with the learning content in the preset content library. Each content item in the preset content library is marked with a corresponding knowledge point label, which identifies the knowledge point covered by the content.

[0071] According to the matching result, the content related to the target knowledge node is filtered out. The content can be sorted using a recommendation algorithm (such as collaborative filtering, content-based recommendation, etc.) to ensure that the recommended content meets the user's learning progress and maintains the user's interest.

[0072] As an example but not limited, according to the user's historical learning data and preferences, the recommended content is personalized adjusted. For example, if the user prefers video-form learning content, related videos can be preferentially recommended. The filtered and sorted content is generated into a recommendation list and displayed to the user in a certain format (such as list, card, etc.). This ensures that the learning content received by the user matches their learning goals and current knowledge level, helping users learn more effectively.

[0073] In some embodiments, the target knowledge node sequence is generated based on the knowledge progression sequence and historical learning data using a preset reinforcement learning algorithm, including: determining the user's current learning progress based on the historical learning data to generate an initial knowledge node sequence; obtaining the user's learning feedback data, which includes the correct answer rate and learning duration; combining the learning feedback data and the knowledge progression sequence, and using the reinforcement learning algorithm to adjust the initial knowledge node sequence to generate the target knowledge node sequence.

[0074] As known from the foregoing, the historical learning data includes the user's past learning records, including completed learning tasks, test scores, learning duration, etc.

[0075] The current learning progress can be understood as the user's position on the "learning path", reflecting the knowledge points the user has mastered and the content the user has not yet learned.

[0076] The initial knowledge node sequence is a list of knowledge points generated based on the user's current learning progress, representing the initial content the user needs to learn next.

[0077] The learning feedback data refers to the user's performance data during the learning process, which is used to evaluate the learning effect. For example, feedback data is collected in real time during the user's learning process and recorded in the user's learning profile.

[0078] As an example but not limited, the user's learning progress and mastered knowledge points are taken as the state of reinforcement learning. The next knowledge point is selected as the action. The reward function is designed according to the learning feedback data, such as giving positive rewards for high accuracy. The strategy of selecting knowledge points is adjusted according to user feedback. The initial knowledge node sequence is adjusted according to the updated strategy to generate the target knowledge node sequence.

[0079] In an implementation, the age stage and learning progress data of the child are obtained, the cognitive level and concept sequence corresponding to the age stage are extracted from a pre-established database, and the initial knowledge node sequence is obtained by selecting the basic concept sequence if the age stage is lower than a preset threshold value. The initial knowledge node sequence and the learning progress are input into a reinforcement learning algorithm, and the concept sequence is dynamically adjusted in combination with feedback mechanism data.

[0080] In combination with application scenarios, for example, in the field of children's education, personalized learning content recommendation for different age stages and learning progress is a core requirement. When obtaining the age stage and learning progress data of the child, a questionnaire filled out by the parents or interactive records of the learning platform can be used. Assuming that a 5-year-old child has a learning progress showing that the basic digital concept has been mastered but the addition concept is not familiar. The pre-established database stores the cognitive level of each age stage and the corresponding concept sequence, for example, the cognitive level of a 5-year-old child usually includes basic numbers, shapes, and simple logical reasoning. Based on this, it is determined by logical judgment that the age of the child is lower than the preset threshold value of 6 years old, and therefore the basic concept sequence is selected as the initial knowledge node sequence, including number recognition, counting, and simple addition.

[0081] In a possible implementation, the reinforcement learning algorithm adjusts the priority of the concept sequence in combination with the feedback mechanism data of the child in the learning process, such as the correct rate of answering questions or the learning time (normalization processing is needed to unify the dimensions of the correct rate of answering questions and the learning time, for example, a conventional normalization algorithm can be used, which is not limited here).

[0082] Suppose the child's correct rate in counting practice reaches 80%, but only 50% in addition practice, the algorithm will preferentially recommend more basic addition practice content.

[0083] The learning path is dynamically adjusted according to real-time user feedback to ensure that the content matches the current level of the user, improving learning effectiveness and efficiency.

[0084] In some embodiments, when the target knowledge node sequence includes knowledge nodes other than the initial knowledge node sequence, the method further includes: marking the knowledge nodes other than the initial knowledge node sequence as expansion knowledge nodes; converting the expansion knowledge nodes into expansion knowledge vectors using a vector generation tool, the expansion knowledge vectors including learning difficulty features and learning correlation degree features; calculating the vector values of the expansion knowledge vectors, and if the vector values exceed a preset vector threshold, adjusting the learning difficulty features and / or the learning correlation degree features to obtain optimized expansion knowledge vectors; and adjusting the initial knowledge node sequence according to the optimized expansion knowledge vectors to generate the target knowledge node sequence.

[0085] The expansion knowledge nodes refer to knowledge points other than the initial knowledge node sequence, which can be used as further learning and expansion content for the user.

[0086] The vector generation tool refers to a tool or algorithm for converting knowledge points into vector representations, and typically uses Word2Vec, BERT or graph embedding algorithms.

[0087] The learning difficulty features reflect the learning difficulty of the knowledge points, which can be the complexity of the knowledge points, the difficulty of the user's mastery, etc., and can be quantified by text complexity or historical learning success rate.

[0088] The learning correlation degree features reflect the correlation degree of the knowledge points with the current learning content of the user, which can be quantified based on co-occurrence frequency or knowledge graph edge weight.

[0089] As an example but not limitation, the learning difficulty and learning correlation degree features of each expansion knowledge node are extracted, and these features are converted into vector representations using a vector generation tool (which needs to be normalized and have a uniform dimension), so that the abstract features of the knowledge points are quantified into computable vectors for subsequent processing.

[0090] In one implementation, a reinforcement learning algorithm can be used to generate a predicted expansion direction to determine the expansion knowledge nodes. A vector generation tool is used to convert the candidate set into an expansion direction vector, and if the vector value exceeds a preset vector threshold, the learning difficulty features and / or the learning correlation degree features are adjusted to obtain an optimized expansion direction vector. The initial knowledge node sequence is adjusted according to the optimized expansion knowledge vector to generate the target knowledge node sequence.

[0091] Suppose that based on the initial knowledge node sequence of "addition-subtraction", the predicted expansion direction is "comparison of size" or "simple multiplication". A candidate set is formed according to these expansion directions, and a vector generation tool is used to convert the candidate set into an expansion direction vector.

[0092] For example, "compare size" can be quantified as a vector with a difficulty coefficient of 0.6 and a cognitive correlation of 0.8. If the vector value exceeds the preset vector threshold of 0.9, the difficulty coefficient will be adjusted, for example, to 0.5, thereby generating an optimized expansion direction vector.

[0093] By way of example but not limitation, the difficulty coefficient can be adjusted by replacing simpler learning content, for example, replacing "compare score size" with "compare integer size", thereby reducing the difficulty coefficient.

[0094] By way of example but not limitation, the learning correlation can also be adjusted for optimization, for example, there are two expansion knowledge nodes, which are quantified as a difficulty coefficient of 0.6 and a cognitive correlation of 0.8 (first vector) and a difficulty coefficient of 0.2 and a cognitive correlation of 0.7 (second vector). Assuming that the first vector exceeds the preset vector threshold and the second vector does not exceed, the second vector can be used as the corresponding expansion knowledge node. That is, according to the user's learning situation, the knowledge node with slightly lower learning correlation but more suitable difficulty can be preferentially recommended.

[0095] By way of example but not limitation, based on the optimized expansion direction vector, the initial knowledge node sequence is adjusted to generate a target knowledge node sequence. For example, "addition - subtraction - compare integer size".

[0096] In one implementation, based on the optimized expansion direction vector, learning content matching the cognitive level of children is extracted from a content matching database, which can include animated teaching videos, interactive games, and graphic exercises. The content matching tool will sort these resources and preferentially select interactive games related to "compare size" because their form is more suitable for the attention characteristics of 5-year-old children. The final cognitive expansion content can be a combination of a 5-minute animated video and 10 "compare size" exercises. This way, through dynamic adjustment and accurate matching, it ensures that the learning content not only meets the current ability of children, but also gradually guides their cognitive development.

[0097] In some embodiments, the method further comprises: calculating an initial attention degree of the knowledge node according to historical learning data; calculating the suitability score of the knowledge node in different age groups according to a preset knowledge node suitability score rule, and if the suitability score is less than a preset score threshold, reducing the weight of the initial attention degree of the knowledge node to obtain an optimized attention degree; based on the optimized attention degree, combining the age of the user, calculating the suitability weight of each knowledge node in different age groups to obtain a knowledge node suitability weight matrix in different age groups; using a K-means clustering algorithm to cluster the knowledge node suitability weight matrix in different age groups to generate an optimized expansion direction vector; and adjusting the initial knowledge node sequence according to the optimized expansion knowledge vector to generate a target knowledge node sequence.

[0098] In an implementation, the interaction record and learning duration data are obtained from the child history interaction behavior database, the interaction frequency and average learning duration of each knowledge point are calculated by using the behavior data analysis method, and the initial attention degree of the knowledge node is obtained. According to the initial attention degree of the knowledge node, in combination with the pre-established age adaptability standard table, if the age adaptability score of the knowledge point is lower than the preset threshold, the attention degree weight of the knowledge point is reduced, and the adjusted knowledge point attention degree (optimized attention degree) is obtained. By using the adjusted knowledge point attention degree, the age suitability weight of each knowledge point in different age groups is calculated by using the content recommendation logic, and the age suitability weight matrix is obtained. For the age suitability weight matrix, the K-means clustering algorithm is used to generate the optimized expansion direction vector, and the knowledge point recommendation priority of each age group is determined.

[0099] Illustratively, when obtaining the interaction record and learning duration data from the child history interaction behavior database, the specific operation of the child on the learning platform can be focused on.

[0100] For example, a 6-year-old child has interacted with the "basic mathematics" module 50 times in the past 3 months through clicking, watching videos, completing exercises, etc., and has interacted with the "language literacy" module 30 times, with a total learning duration of 400 minutes. The behavior data analysis method can calculate the initial attention degree of the knowledge node by counting the interaction frequency and average learning duration.

[0101] Specifically, the interaction frequency of basic mathematics is 50 times, and the average learning time is 12 minutes each time. The initial value of the attention degree can be set as the weighted sum of the frequency and the duration, for example, 60. The initial value of the attention degree of language literacy is 45. This method intuitively reflects the interest tendency of the child to the knowledge point, providing data support for subsequent recommendation.

[0102] In a possible implementation, the suitability of the knowledge point is evaluated in combination with the pre-established age adaptability standard table.

[0103] For example, the age adaptability standard table stipulates that the suitability score of the "basic mathematics" knowledge point for a 6-year-old child is 90, and the suitability score of the "complex geometry" is 40. If the threshold is set to 70, the suitability score of the complex geometry is lower than the threshold, and the attention degree weight thereof needs to be reduced.

[0104] Preferably, assuming that the initial attention degree is 50, the weight is reduced by 20% after adjustment, and the adjusted attention degree is 40. This adjustment ensures that the recommended content is more suitable for the cognitive ability of the child, and avoids the decline of learning interest due to too high difficulty.

[0105] It should be noted that based on the adjusted knowledge point attention, the content recommendation logic is used to calculate the age suitability weight matrix, which can introduce age difference analysis.

[0106] For example, for the two age groups of 4-6 years old and 7-9 years old, the suitability weight of basic mathematics for 4-6 years old is 0.9, and for 7-9 years old is 0.7; and the language literacy is 0.8 and 0.85 respectively. The weight matrix is generated by the product of attention and suitability score, reflecting the recommendation priority of knowledge points for different age groups. This matrix method makes the recommendation logic more specific, which helps to accurately match learning content.

[0107] In an embodiment, for the age suitability weight matrix, a K-means clustering algorithm is used to generate an optimized expansion direction vector.

[0108] For example, the knowledge points are clustered into high, medium and low priority groups according to the weight matrix. Assuming that basic mathematics and language literacy are clustered into the high priority group, the expansion direction vector can be represented as [0.9, 0.8], and the low priority knowledge point such as complex geometry is [0.4]. The clustering result determines the recommended priority of 4-6 year-old children as basic mathematics and language literacy. This method optimizes the recommendation direction by grouping to ensure that the learning content is highly consistent with the age and interest of children.

[0109] This multi-dimensional collaborative solution can effectively improve the attractiveness and adaptability of learning content, and provide children with a better learning experience.

[0110] In some embodiments, generating target recommendation content according to a target knowledge node sequence includes: calculating the semantic matching degree between the target knowledge node sequence and the candidate recommendation content in the preset content library; calculating the association weight between the candidate recommendation content and the target knowledge node according to the semantic matching degree; calculating the suitability weight of the candidate recommendation content in different age groups according to the preset knowledge node suitability scoring rule if the association weight exceeds the preset weight threshold; calculating the recommendation score of the candidate recommendation content according to the semantic matching degree, the association weight and the suitability weight; sorting the candidate recommendation content according to the numerical value of the recommendation score to obtain the sorted recommendation content; and determining the target recommendation content according to the sorted recommendation content.

[0111] It can be understood that the BERT algorithm is used to generate semantic vectors of the target knowledge node sequence and the candidate recommendation content, and the cosine similarity between the two is calculated to determine the semantic matching degree between the two.

[0112] According to the semantic matching degree, the association weight between the candidate recommendation content and the target knowledge point is calculated, and if the association weight is higher than the preset threshold, the candidate recommendation content is retained to obtain the candidate recommendation content set.

[0113] The age suitability label corresponding to the candidate recommended content in the candidate recommended content set is obtained, the user age range is combined, a preset suitability score rule is used to calculate the suitability weight of the candidate recommended content, and if the suitability weight is higher than a preset threshold, the content is retained to determine the suitability content set.

[0114] Through the suitability content set, a sorting algorithm is used to perform weighted calculation on the semantic matching degree, knowledge association weight and suitability weight of the content to obtain a comprehensive score, the content with the highest score is screened, a final candidate recommended content set is determined, and a sorted recommended content is obtained. The target recommended content is determined in turn from high to low according to the score.

[0115] Exemplarily, the process of extracting text features from the content library through the pre-established semantic analysis tool can be based on natural language processing technology. The semantic analysis tool usually performs word segmentation, stop word removal and other preprocessing on the text in the content library to generate features that can reflect the semantics of the text.

[0116] For example, there is a popular science article about “animal habitats” in the content library, and the tool will extract keywords such as “forest”, “ocean” and “adaptability”, and generate text features in combination with the context. When generating a semantic vector using the BERT algorithm, these features are input into the pre-trained BERT model to output a high-dimensional vector.

[0117] For example, the “animal habitat” article may generate a 512-dimensional semantic vector reflecting its core semantics. The input text of the expansion direction, such as the user input “animal knowledge suitable for children”, will also be processed in the same way to generate a corresponding semantic vector. When calculating the cosine similarity between the two, the smaller the angle between the vectors, the higher the matching degree. Assuming that the similarity between the “animal habitat” article and the input is 0.85, which is higher than the threshold 0.7, it indicates that the semantics are highly related.

[0118] In one possible implementation, the knowledge point label is obtained from the content library according to the semantic matching degree, and the association weight is calculated in combination with the knowledge point weight table.

[0119] For example, the article in the content library has “biodiversity” and “habitat” labels, and the weight table has a “biodiversity” weight of 0.9 and a “habitat” weight of 0.8. If the association weight of the article with the target knowledge point “animal adaptability” is calculated to be 0.82, which is higher than the threshold 0.75, the article is retained. This method ensures the closeness of the content to the target knowledge point. The knowledge association content set thus contains a series of articles highly related to animal knowledge, such as “survival strategies of marine organisms” and “protection mechanisms of forest animals”.

[0120] It should be noted that, in order to obtain the age suitability label for the knowledge-related content set, and to calculate the suitability weight according to the user age range, the pre-set suitability scoring rule needs to be relied on.

[0121] For example, assuming that the user is an 8-year-old child, the rule stipulates that the content suitable for children aged 8-10 years old needs to contain simple vocabulary and vivid examples. The article "Survival Strategies of Marine Life" in the content library is suitable for children because it uses simple language and cartoon illustrations, and the suitability score is 0.9, which is higher than the threshold value 0.8, so it is retained. On the contrary, an academic article "Biological Evolution" is excluded because the terminology is complex and the score is only 0.6. The suitability content set is therefore more suitable for the cognitive level of children.

[0122] Specifically, when calculating the comprehensive score through the suitability content set, the sorting algorithm will comprehensively consider the semantic matching degree, the knowledge correlation weight, and the suitability weight.

[0123] For example, assuming that the semantic matching degree of the article "Survival Strategies of Marine Life" is 0.85, the knowledge correlation weight is 0.82, and the suitability weight is 0.9, the weighted calculation comprehensive score is 0.86. Another article "Protection Mechanism of Forest Animals" has a score of 0.83. After sorting, the article with the highest score is preferentially entered into the candidate recommended content set. This weighted sorting ensures that the recommended content is both relevant and suitable for children.

[0124] It can be understood that the formation of the candidate recommended content set emphasizes the rigor of multi-dimensional screening. For example, "Survival Strategies of Marine Life" becomes the preferred recommended content because of its vivid description and high compatibility with children's interests. In contrast, other content may be excluded because of insufficient suitability or lower semantic matching degree. This method ensures the accuracy and applicability of the recommended content through layer-by-layer screening, and provides high-quality learning resources for children.

[0125] In some embodiments, determining the target recommended content according to the sorted recommended content includes: determining the learning preference of the user according to the historical learning data; calculating the content recommendation score corresponding to the sorted recommended content, the content recommendation score being related to the content complexity; adjusting the content recommendation score by using an attention mechanism algorithm combined with the learning preference to obtain an optimized content recommendation score; re-sorting the sorted recommended content according to the numerical size of the optimized content recommendation score to obtain an optimized recommended content; and if the number of contents in the optimized recommended content exceeds a pre-set number threshold, deleting the contents with a score lower than a pre-set score range to make the number of contents in the optimized recommended content meet the pre-set number threshold, to obtain the target recommended content.

[0126] In one implementation, the historical behavior data of the child is obtained from a behavior database, and a clustering algorithm is used to analyze the historical behavior data to obtain the preference characteristics of the child.

[0127] The content complexity of the content features corresponding to the ranked recommended content is determined by a vector calculation method to obtain a content recommendation score (which can also be understood as a content complexity score).

[0128] According to the content complexity score and the child preference feature, an attention mechanism is used to calculate the recommended content to obtain an optimized content recommendation score; and the ranked recommended content is re-ranked according to the numerical value of the optimized content recommendation score to obtain optimized recommended content.

[0129] If the number of the optimized recommended content exceeds the preset upper limit, low-weight content is removed by a truncation method to obtain the final recommended content (target recommended content).

[0130] Exemplarily, when the child historical behavior data is obtained from the behavior database, the behavior dataset can be constructed by analyzing the interaction records of the child in the learning application, such as the type of video watched, the time spent, the frequency of completing tasks, and the like. Assuming that a child A watched math-related animations 10 times in the past 30 days, with an average time spent of 5 minutes, and completed math mini-games 20 times, while watching language-related content only 2 times, with a time spent of less than 2 minutes. Through these data, it can be preliminarily judged that the child A is more interested in math content. The collection of behavior data needs to ensure privacy compliance, and only the features related to recommendation, such as content type and interaction time, are extracted to avoid involving sensitive information. The advantage of this is that the learning preferences of children can be accurately captured to provide a data basis for subsequent recommendations.

[0131] In a possible implementation, when the historical behavior data is analyzed by using a clustering algorithm, a K-means clustering algorithm can be used to divide the behavior features of the children into several categories.

[0132] For example, the children are divided into three categories: “math preference”, “language preference”, and “comprehensive interest”. Assuming that the behavior data of child A shows that the proportion of math-related interactions is 80%, the proportion of language-related interactions is 10%, and the proportion of other interactions is 10%, through clustering analysis, child A is classified into the “math preference” category. The core of the clustering algorithm is to group children with similar behavior patterns to extract preference features.

[0133] It should be noted that the data needs to be standardized before clustering to ensure that the data in different dimensions are comparable. The advantage of this method is that it can quickly summarize the interest tendencies of children to provide a basis for personalized recommendation.

[0134] Specifically, when the content features of the recommended content are extracted from the content library, the text, image, or interactive elements of the content can be analyzed.

[0135] For example, a certain math animation content contains the features of "addition and subtraction", "animation explanation", and "interactive question and answer". These features are converted into vector representations by a pre-trained semantic analysis model. The content complexity score can be calculated by the distance between vectors. If the distance between a certain content vector and the "basic math" template vector is closer, the score is lower, indicating that the content is simple. If the distance between the content vector and the "advanced math" template is farther, the score is higher, indicating that the content is complex.

[0136] For example, content B is "addition within 10 animation" with a complexity score of 2, and content C is "fraction operation explanation" with a complexity score of 8. The advantage of this scoring method is that it can quantify the difficulty of the content, which is convenient for matching with children's preferences.

[0137] In one embodiment, when the attention mechanism is used for weighted calculation of recommended content, the content weight can be dynamically adjusted according to the children's preference features. Assuming that the preference features of child A show high interest in math, the attention mechanism will assign a higher weight to math-related content.

[0138] For example, the initial matching degree of content B is 0.7, and the score is increased to 0.9 after weight adjustment because it is highly related to the math preference. Content D is a Chinese content with an initial matching degree of 0.6, and the weight is adjusted to 0.4. The advantage of the attention mechanism is that it can highlight the content that interests the child and improve the accuracy of the recommendation.

[0139] Preferably, if the number of optimized recommended content exceeds the preset upper limit, such as 10 recommended content and 15 optimized recommended content, the low-weight content can be removed by the truncation method, for example, removing the 5 content with the lowest score and keeping the top 10.

[0140] As an example but not limited to, the truncation method can also consider the diversity of content, such as animation, game, and exercise, to avoid all single-type content. The advantage of this method is that it can control the number of recommendations, improve user experience, and maintain content diversity.

[0141] For example, the final target recommended content received by child A may include "addition within 10 animation", "simple math game", and "basic math exercise", which are highly matched with the preference and have moderate difficulty, providing better user experience.

[0142] In one embodiment, the content hierarchical labeling technology is used to classify the target recommended content by difficulty and theme, obtaining classified recommended content.

[0143] As an example but not limited to, a preset theme classification rule is used to classify the target recommended content by keyword matching method to generate a first content set with theme labels, obtaining a theme classification result.

[0144] For the first content set, the text length and vocabulary complexity of each content are obtained, and a weighted calculation method is used, wherein the text length accounts for w1, the vocabulary complexity accounts for w2, and w1+w2=1, the content difficulty score is calculated, the second content set with difficulty classification is generated, and the difficulty classification result is determined.

[0145] According to the preset user preference data, a collaborative filtering algorithm is used to sequence sort the second content set, and a sequence sorting result is obtained.

[0146] Through the theme association rule, the content not matching the user preference is removed from the sequence sorting result, and a final recommendation sequence is generated.

[0147] A preset theme classification rule is used to classify the target recommendation content through a keyword matching method. For example, the mathematics theme includes "addition", "score", and "geometry", and the Chinese theme includes "pinyin", "reading", and "composition". Each content is matched with a keyword library through text analysis to generate a first content set with a theme label.

[0148] For example, an animation title "addition explanation within 10" is marked as a "mathematics" theme because it contains the "addition" keyword, and a reading material title "fairy tale" is marked as a "Chinese" theme because it contains the "reading" keyword. This method can quickly classify the content by theme and provide structured data for subsequent processing.

[0149] In one possible implementation, when calculating the content difficulty score for the first content set, the text length and vocabulary complexity of each content can be analyzed.

[0150] For example, the text length of a certain mathematics animation is 200 words, the vocabulary complexity is low, and basic words such as "addition" and "number" are mainly used, which is suitable for low-grade children; the text length of a certain Chinese reading material is 500 words, and it contains complex words such as "meaning" and "philosophy", which is suitable for high-grade children. A weighted calculation method is used, assuming that the text length accounts for w1 of 0.4 and the vocabulary complexity accounts for w2 of 0.6, and the difficulty score is calculated. The scoring formula is the standardized text length multiplied by 0.4 plus the vocabulary complexity multiplied by 0.6, generating a score of 0 to 10.

[0151] For example, the above-mentioned mathematics animation score is about 3, and the Chinese reading material score is about 7, and after generating the second content set with difficulty classification, the content is classified into "simple", "medium", and "difficult". This scoring method can accurately reflect the suitable age range of the content.

[0152] For example, a collaborative filtering algorithm can be used to sequence rank the second content set based on the user preference data. Suppose the preference data of child B shows that it prefers math content, especially simple math animations. The collaborative filtering algorithm recommends similar content by analyzing the interaction records of other children with similar preferences as child B.

[0153] For example, child B has watched simple math animations 10 times, medium math exercises 3 times, and Chinese reading only once in the past. Accordingly, the algorithm prioritizes simple math content, such as "addition within 10 animations", and ranks it high in the sequence, while ranking Chinese content, such as "pinyin exercises", low in the sequence. The generated preliminary recommendation sequence contains 20 pieces of content, ranked from high to low in matching degree. This method ensures that the recommended content is close to the user's interests.

[0154] By way of example and not limitation, the theme association rule can be to arrange content of the same theme in increasing difficulty.

[0155] For example, filter out math content from the sequence ranking results to ensure that a progressive path from "simple addition animations" to "medium addition exercises" is included. Using the content filtering method, remove content that does not match the preferences, such as Chinese content that child B is not interested in. The final recommendation sequence may contain 8 pieces of content, such as "simple addition animations", "simple addition games", and "medium addition exercises", which not only ensure the theme, but also meet the preferences. This filtering method can improve the coherence and learning effect of the recommended content.

[0156] The above mainly introduces the method of the embodiments of the present application in combination with the drawings. It should be understood that although each step in the flowchart involved in each of the above embodiments is shown in sequence, these steps are not necessarily executed in the order shown in the figure. Unless explicitly stated herein, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each of the above embodiments can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0157] Figure 2 FIG. 1 is a structural schematic diagram of a content recommendation system based on child cognition provided by an embodiment of the present application.

[0158] As Figure 2As shown, the apparatus 1000 includes the following units.

[0159] The acquisition unit 1001 is configured to acquire historical learning data, and perform knowledge graph construction based on the historical learning data to generate a knowledge structure graph, the knowledge structure graph including a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes. The mapping unit 1002 is configured to perform vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes by using a graph embedding algorithm to generate a knowledge vector set. The processing unit 1003 is configured to calculate semantic correlation degrees between the plurality of knowledge nodes according to the knowledge vector set, and extract key knowledge nodes from the plurality of knowledge nodes by using a hierarchical clustering algorithm if the semantic correlation degrees are greater than a preset threshold. The processing unit 1003 is further configured to determine a learning order corresponding to the key knowledge nodes according to a preset knowledge learning strategy to generate a knowledge progression sequence. The processing unit 1003 is further configured to generate a target knowledge node sequence based on the knowledge progression sequence and the historical learning data by using a preset reinforcement learning algorithm. The processing unit 1003 is further configured to generate target recommendation content according to the target knowledge node sequence, and recommend the target recommendation content to a user.

[0160] It should be noted that the information interaction between the above units, the execution process and the like, since based on the same concept as the method embodiments of the present application, the specific functions and the technical effects brought by it can be referred to the method embodiment part, and will not be repeated here.

[0161] Figure 3 is a structural schematic diagram of a computer device provided by the present application. As shown in the figure, Figure 3 The computer device 3000 of the present embodiment includes at least one processor 3100 (only one processor is shown), a memory 3200, and a computer program 3210 stored in the memory 3200 and executable on the at least one processor 3100, and the processor 3100 executes the computer program 3210, so that the computer device implements the steps in the above embodiments. Figure 3

[0162] ​The processor 3100 can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0163] The memory 3200 can be an internal storage unit of the computer device 3000 in some embodiments, for example, a hard disk or a memory of the computer device 3000. The memory 3200 can also be an external storage device of the computer device 3000 in other embodiments, for example, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the computer device 3000. Further, the memory 3200 can include both the internal storage unit and the external storage device of the computer device 3000. The memory 3200 is used to store an operating system, application programs, Boot Loader data, and other programs, for example, program codes of computer programs, etc. The memory 3200 can also be used to temporarily store data that has been output or will be output.

[0164] It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, only the division of the above functional units is taken as an example, and in actual application, the above functions can be completed by different functional units or modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit in the embodiment can be integrated in one processing unit, or each unit can exist physically independently, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit are only for the convenience of mutual distinction, and are not used to limit the protection scope of the present application. The specific working process of the unit in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0165] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a computer device, so that the computer device implements steps in each method embodiment.

[0166] The embodiment of the present application provides a computer program product. When the computer program product is run on a computer device, the computer device can implement the above-mentioned methods.

[0167] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the present application implements all or part of the processes in the above-mentioned embodiment methods, which can be completed by instructing related hardware through a computer program. The computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the computer device can implement the steps of each method embodiment. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium at least includes any entity or device capable of carrying the computer program code to the photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal and a software distribution medium. For example, a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc. In some jurisdictions, according to legislation and patent practice, the computer readable medium cannot be an electrical carrier signal and a telecommunications signal.

[0168] It should be understood that the size of the serial number of each step in the above-mentioned embodiment does not mean the order of execution. The execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the present application. In the description, specific details such as specific system structures, technologies are proposed for illustration but not for limitation, so as to thoroughly understand the embodiments of the present application. However, it should be clear for those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed description of well-known systems, devices, circuits and methods is omitted to avoid unnecessary details that hinder the description of the present application.

[0169] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or sets thereof.

[0170] It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, and that the term "at least one of" followed by a list of two or more items means any single one of the items in the list, and that the term "one or more of" followed by a list of two or more items means any single one or plurality of the items in the list.

[0171] In addition, the terms "comprise", "comprising", "include", "including", "have", "has", "contain", "containing", "provide", "providing", "involving", "carry", "carrying", "utilize", "utilizing", "characterized by" and the like, when used in the description and the following claims, are each intended to mean that the respective item is an element of the described or claimed item, but does not exclude the presence or addition of one or more other items.

[0172] In the above embodiments, the description of each embodiment is focused on a certain aspect, and the parts not described in detail or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0173] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0174] In the embodiments provided in the present application, it should be understood that the disclosed devices, computer devices and methods can be implemented in other ways. For example, the above-described device and computer device embodiments are only schematic. The division of units is only a logical function division. In actual implementation, another division mode can be used. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0175] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A content recommendation method based on children's cognition, characterized in that: include: Acquire historical learning data, and construct a knowledge graph based on the historical learning data to generate a knowledge structure graph, wherein the knowledge structure graph includes a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes; Performing vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes using a graph embedding algorithm to generate a knowledge vector set; Calculating the semantic relevance between the plurality of knowledge nodes based on the knowledge vector set, and extracting key knowledge nodes from the plurality of knowledge nodes using a hierarchical clustering algorithm if the semantic relevance is greater than a preset threshold; Determine the learning order corresponding to the key knowledge nodes according to the preset knowledge learning strategy and generate a knowledge progression sequence; Generate a target knowledge node sequence based on the knowledge progression sequence and the historical learning data using a preset reinforcement learning algorithm; Generate target recommended content according to the target knowledge node sequence, and recommend the target recommended content to the user.

2. The method according to claim 1, characterized in that The calculating the semantic association between the plurality of knowledge nodes according to the knowledge vector set includes: Calculating the cosine similarity between each vector in the knowledge vector set; Divide the vectors whose cosine similarity exceeds the preset similarity threshold into a group to obtain a semantic association set; The weighted average method is used to calculate the similarity weight value of each group of vectors in the semantic association set to obtain the semantic association, and the weight value corresponding to the similarity of each group of vectors is positively correlated with the frequency of common occurrence of each group of vectors.

3. The method according to claim 1, characterized in that If the semantic relevance is greater than a preset threshold, extracting key knowledge nodes from the multiple knowledge nodes using a hierarchical clustering algorithm, including: Marking the knowledge nodes corresponding to the semantic relevance greater than a preset threshold as relevant knowledge nodes to obtain a set of relevant knowledge nodes; Using a hierarchical clustering algorithm to group the knowledge nodes in the relevant knowledge node set to obtain relevant knowledge node groups; When the relevant knowledge node group includes at least two knowledge nodes, key knowledge nodes are extracted from the at least two knowledge nodes according to the occurrence frequencies of the knowledge nodes.

4. The method according to claim 3, characterized in that The step of determining the learning order corresponding to the key knowledge nodes according to the preset knowledge learning strategy and generating a knowledge progression sequence includes: Determine the directional relationship between the key knowledge nodes according to a preset learning strategy; generating a directed graph according to the directional relationship between the key knowledge nodes, wherein the edges in the directed graph include edge weights, and the edge weights are positively correlated with the frequency of co-occurrence of the corresponding key knowledge nodes; A sequence generation algorithm is used to sort the key knowledge nodes based on the numerical values ​​of the edge weights, determine the learning order corresponding to the key knowledge nodes, and generate a knowledge progression sequence.

5. The method according to claim 1, wherein The generating of a target knowledge node sequence based on the knowledge progression sequence and the historical learning data by using a preset reinforcement learning algorithm includes: Determine the user's current learning progress based on the historical learning data and generate an initial knowledge node sequence; Obtaining learning feedback data from users, including correct answer rate and learning time; Combining the learning feedback data and the knowledge progression sequence, the initial knowledge node sequence is adjusted using a reinforcement learning algorithm to generate a target knowledge node sequence.

6. The method according to claim 5, characterized in that When the target knowledge node sequence includes knowledge nodes other than the initial knowledge node sequence, the method further includes: Marking the knowledge nodes outside the initial knowledge node sequence as extended knowledge nodes; Using a vector generation tool to convert the expanded knowledge node into an expanded knowledge vector, the expanded knowledge vector including a learning difficulty feature and a learning relevance feature; Calculating a vector value of the expanded knowledge vector, and if the vector value exceeds a preset vector threshold, adjusting the learning difficulty feature and / or the learning association feature to obtain an optimized expanded knowledge vector; The initial knowledge node sequence is adjusted according to the optimized and expanded knowledge vector to generate a target knowledge node sequence.

7. The method according to claim 6, characterized in that The method further comprises: Calculating the initial attention of the knowledge node according to the historical learning data; Calculating the suitability scores of the knowledge nodes in different age groups according to preset knowledge node suitability scoring rules; if the suitability scores are less than a preset scoring threshold, reducing the weight of the initial attention of the knowledge node to obtain optimized attention; Based on the optimized attention and the user's age, the suitability weight of each knowledge node in different age groups is calculated to obtain the suitability weight matrix of knowledge nodes in different age groups; Using the K-means clustering algorithm to perform clustering on the suitability weight matrices of the knowledge nodes of different age groups, and generate optimized expansion direction vectors; The initial knowledge node sequence is adjusted according to the optimized and expanded knowledge vector to generate a target knowledge node sequence.

8. The method according to claim 7, characterized in that Generating target recommended content according to the target knowledge node sequence includes: Calculating the semantic matching degree between the target knowledge node sequence and the candidate recommended content in the preset content library; Calculating the association weight between the candidate recommended content and the target knowledge node according to the semantic matching degree; Calculate the suitability weights of the candidate recommended content for different age groups whose association weights exceed the preset weight threshold according to the preset knowledge node suitability scoring rules; Perform weighted calculation on the candidate recommendation content according to the semantic matching degree, association weight, and suitability weight to obtain the recommendation score corresponding to the candidate recommendation content; Sorting the candidate recommended contents according to the numerical values ​​of the recommendation scores corresponding to the candidate recommended contents to obtain ranked recommended contents; The target recommended content is determined according to the ranked recommended content.

9. The method according to claim 8, characterized in that The determining the target recommended content according to the ranked recommended content includes: determining the user's learning preferences based on the historical learning data; Calculating a content recommendation score corresponding to the ranked recommended content, where the content recommendation score is related to content complexity; Using an attention mechanism algorithm and combining the learning preference to adjust the content recommendation score to obtain an optimized content recommendation score; Re-sorting the ranked recommended contents according to the numerical values ​​of the optimized content recommendation scores to obtain optimized recommended contents; If the number of contents in the optimized recommended contents exceeds a preset number threshold, contents with scores lower than a preset score range are deleted so that the number of contents in the optimized recommended contents meets the preset number threshold, thereby obtaining target recommended contents.

10. A content recommendation system based on children's cognition, characterized in that: include: an acquisition unit, configured to acquire historical learning data, and construct a knowledge graph based on the historical learning data to generate a knowledge structure graph, wherein the knowledge structure graph includes a plurality of knowledge nodes and edges connecting the plurality of knowledge nodes; a mapping unit, configured to perform vector mapping processing on the plurality of knowledge nodes and the edges connecting the plurality of knowledge nodes using a graph embedding algorithm to generate a knowledge vector set; a processing unit, configured to calculate the semantic relevance between the plurality of knowledge nodes based on the knowledge vector set, and extract key knowledge nodes from the plurality of knowledge nodes using a hierarchical clustering algorithm if the semantic relevance is greater than a preset threshold; The processing unit is further configured to determine the learning order corresponding to the key knowledge nodes according to a preset knowledge learning strategy, and generate a knowledge progression sequence; The processing unit is further configured to generate a target knowledge node sequence based on the knowledge progression sequence and the historical learning data using a preset reinforcement learning algorithm; The processing unit is further configured to generate target recommended content according to the target knowledge node sequence, and recommend the target recommended content to the user.

Images