AI-based personalized learning resource recommendation system

Abstract

This invention provides an AI-based personalized learning resource recommendation system that improves the accuracy, adaptability, and continuous optimization capabilities of learning resource recommendations. First, a multi-dimensional user profile is constructed by collecting users' static information, behavioral sequences, and external states. Next, learning resources are tagged and modeled based on a three-tiered semantic structure of concepts, skills, and context, constructing a dynamically adjustable knowledge graph. Finally, by combining reinforcement learning and periodic feedback mechanisms, the recommendation strategy is continuously optimized, achieving closed-loop control of recommendation, evaluation, and adjustment. This invention, by constructing a state vector and policy network, dynamically perceives user changes, significantly improving resource matching and individual adaptability, providing technical support for achieving intelligent and precise education.

Description

Technical Field

[0001] This invention relates to the field of learning resource recommendation, specifically a personalized learning resource recommendation system based on artificial intelligence. Background Technology

[0002] With the continuous advancement of educational informatization, personalized learning has become an important development direction for improving teaching efficiency and meeting differentiated learning needs. Traditional methods are mainly based on rule matching (such as expert knowledge bases) or collaborative filtering algorithms (such as UserCF and ItemCF). These methods can provide a certain degree of personalized service in the early stages. However, they have significant limitations. On the one hand, they rely too much on historical behavioral data, which makes it impossible for the system to capture the dynamic changes in user interests in a timely manner. On the other hand, they lack the ability to model learner behavior sequences and cannot accurately depict the state evolution in the user's learning path. At the same time, the recommendation goals are too singular, usually only using click-through rate or access frequency as optimization goals, ignoring key educational indicators such as learning effectiveness, knowledge mastery, and learning load.

[0003] To address the aforementioned issues, existing technologies have incorporated machine learning and reinforcement learning, using a user feedback-based policy update mechanism to dynamically adjust recommendation strategies. This has improved the adaptive capabilities and interactive intelligence of recommendation systems to some extent, enabling them to update recommendation strategies based on user interaction data and exhibiting a degree of intelligence and flexibility. However, existing user modeling technologies fail to integrate multi-dimensional information such as static attributes, behavioral timelines, and external states, resulting in weak profile representation capabilities. Furthermore, the optimization objectives of recommendation strategies are singular, relying solely on immediate feedback for local updates and lacking a comprehensive evaluation of long-term learning benefits. Therefore, establishing a structurally sound, semantically driven, and dynamically adjustable recommendation system that achieves accurate modeling of learner states and closed-loop feedback optimization has become a key direction for current technological development and the core problem that this invention aims to solve. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a personalized learning resource recommendation system based on artificial intelligence.

[0005] Therefore, the technical solution adopted by the present invention is as follows:

[0006] An AI-based personalized learning resource recommendation system, comprising:

[0007] M1: User information processing module, including a collection unit and a modeling unit. The collection unit acquires the user's basic information, and the modeling unit performs feature modeling based on the user's basic information to generate a personalized user profile.

[0008] M2: The learning resource construction module performs semantic annotation and resource tagging on learning resources based on a three-level semantic structure of concepts, skills, and contexts, and establishes a knowledge graph; according to the user profile, it adjusts the weights and association paths of resource tags in the knowledge graph.

[0009] M3: Decision module, based on the user profile and the adjusted knowledge graph, generates an initial recommendation strategy, further obtains the user's initial feedback data, and combines reinforcement learning to dynamically update and adjust the initial recommendation strategy to generate personalized learning resource recommendation results;

[0010] M4: Intervention module, which comprehensively evaluates the learning resource recommendation results based on the user's learning indicators within a preset period, and automatically triggers the adjustment of the user profile based on the evaluation results to generate new learning resource recommendation results;

[0011] The basic information includes static information, dynamic behavioral information, and external state information; these three information are concatenated to form a comprehensive feature, which is then input into a multilayer perceptron to generate a personalized user profile. .

[0012] Furthermore, the specific process of establishing the knowledge graph is as follows:

[0013] 1) Extract concept tags for each learning resource Skill tags and context labels These are combined to form a comprehensive tag for the corresponding learning resource, represented as:

[0014]

[0015] in, Indicates learning resources Comprehensive tags;

[0016] 2) After extracting the tags from all learning resources, map them to a set of concept tag nodes. Skill tag node set and context tag node set And based on the semantic relevance of the learning resources, an edge set is established. To form a knowledge graph , is represented as:

[0017]

[0018] in, This represents the set of all nodes.

[0019] Furthermore, the resource tag weights are the node weights of the knowledge graph, and the correlation between user profiles and nodes is calculated using cosine similarity, with the following formula:

[0020]

[0021] in, Represents each node in the knowledge graph The corresponding embedding representation, The similarity between the user profile and the node representation is used to represent the similarity; the node weights are updated based on the cosine similarity calculation results. , is represented as:

[0022]

[0023] in, Indicates the weight of the node after the update; This indicates that the step size coefficient is being updated;

[0024] After the node weights of the knowledge graph are updated, for a connection node... and Edges, edge weights The data will be updated synchronously, as shown below:

[0025]

[0026] in, This represents the updated edge weights; This represents the edge weight adjustment coefficient; The function representing the influence of the weights of the nodes at both ends of an edge on the edge is expressed as:

[0027]

[0028] By updating the edge weights, changes in node importance are propagated to the association paths of the knowledge graph. These paths are then adjusted based on user profiles, ultimately forming an adjusted knowledge graph. , is represented as:

[0029]

[0030] in, This represents the new set of node weights; This represents the updated set of edge weights.

[0031] Furthermore, the specific technical steps of the reinforcement learning are as follows:

[0032] 1) Define the state ,action and rewards ,

[0033] The reward is generated based on the user's initial feedback data and is represented as follows:

[0034]

[0035] in, This indicates a score for learning efficiency; This indicates an improvement in knowledge acquisition; Indicators representing learning fatigue levels; and These represent the weighting coefficients for knowledge mastery and learning fatigue, respectively.

[0036] 2) Define the objective function , is represented as:

[0037]

[0038] in, This means that the cumulative rewards for all learning paths experienced by the user are weighted and averaged according to the initial recommendation strategy. This represents the total number of time steps in the initial learning phase; Indicates the first Discount factor for each time step; Indicates time step Instant rewards;

[0039] 3) Update parameters using the policy gradient method The objective function is optimized, and the policy gradient formula is:

[0040]

[0041] in, Describe the objective function For strategy parameters The gradient; Representation strategy In state Select action The gradient of the logarithmic probability;

[0042] The formula for updating the parameters is:

[0043]

[0044] in, This indicates the updated policy parameters; This represents the learning rate; through multiple iterations, parameters are obtained to recommend the most suitable learning resources to users, generating personalized learning resource recommendation results.

[0045] Furthermore, the learning metrics are respectively the degree of improvement in learning efficiency. Increase in knowledge mastery and learning fatigue The three learning metrics were standardized, and a comprehensive score was calculated. The formula is:

[0046]

[0047] in, , and This represents the weighting coefficients for the three learning metrics;

[0048] When the overall score falls below a set score threshold, the user profile is adjusted using the following formula:

[0049]

[0050] in, This indicates the adjusted user profile; This represents the step size coefficient for updating the user profile. This indicates the amount of user profile correction obtained based on learning metrics within a set period.

[0051] Compared with the prior art, the advantages of the present invention are as follows:

[0052] 1. This invention introduces the policy gradient method from reinforcement learning to dynamically adjust the recommendation strategy in order to maximize long-term cumulative rewards. Compared with static rules or traditional collaborative filtering algorithms, it can more effectively respond to changes in user interests, continuously learn user preferences, and improve the relevance and personalization of recommended content.

[0053] 2. This invention introduces a three-level semantic structure tagging system of concepts, skills, and contexts to construct a computable and reasonable knowledge graph, and dynamically adjusts the weights of nodes and edges in conjunction with user profiles to achieve semantically driven personalized resource filtering.

[0054] 3. This invention fully integrates users' static information, dynamic behavioral sequences, and external state features, and generates multi-dimensional user profiles through temporal modeling and attention mechanisms, which significantly improves the accuracy of characterizing learners' interests and states, laying a data foundation for personalized recommendations. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1This is a flowchart of the personalized learning resource recommendation system of the present invention;

[0057] Figure 2 This is a flowchart of the user information processing module of the present invention;

[0058] Figure 3 This is a flowchart of the learning resource construction module of the present invention. Detailed Implementation

[0059] To achieve the above objectives, the present invention provides a personalized learning resource recommendation system based on artificial intelligence. Please refer to the following technical solution. Figures 1 to 3 The system includes:

[0060] M1: User information processing module, including a data acquisition unit and a modeling unit. The data acquisition unit obtains basic user information, and the modeling unit performs feature modeling based on the basic user information to generate personalized user profiles.

[0061] Basic user information includes static information, dynamic behavior information, and external status information.

[0062] Static information refers to a user's static attributes, such as age, professional background, learning goals, historical grades, and areas of interest. Static information is represented by vectors. express;

[0063] Dynamic behavioral information refers to a user's learning behavior within a learning cycle, such as task completion rate, question accuracy, learning duration, and interaction frequency at each learning stage. Dynamic behavioral information is time-series data; in a time series, the... Information from each time step is represented by a vector. It is stated that there are a total of learning cycles. At each time step, the dynamic behavior information is represented as:

[0064]

[0065] in, Indicates from the 1st to the 1st Dynamic behavior at each time step;

[0066] External state information includes the user's attention index, cognitive load level, and learning fatigue level, obtained through interaction logs or wearable devices. External state information is represented by vectors. express.

[0067] The modeling unit utilizes time-series encoding and attention mechanisms to integrate basic user information into personalized user profiles. The specific steps are as follows:

[0068] 1) Dynamic behavioral information is processed through a Bidirectional Long Short-Term Memory (BiLSTM) network. Encode the data and extract time-dependent features, which are represented as follows:

[0069]

[0070] in, Indicates time step The hidden state vector; BiLSTM can capture the dependencies between consecutive time steps, thus better reflecting the changing trend of dynamic behavior over time;

[0071] 2) To differentiate the importance of different time steps, an attention mechanism is introduced to calculate the weight of each time step, using the following formula:

[0072]

[0073] in, Indicates time step The weights; Represents a vector of learnable attention parameters; This indicates that for all time steps The attention scores are summed for normalization to ensure that the sum of the weights at all time steps equals 1. Based on the weights, the hidden state vectors at all time steps are weighted and summed to obtain the aggregated representation of the dynamic behavior, as shown in the formula:

[0074]

[0075] in, A weighted representation of the entire learning cycle, used to summarize the dynamic behavioral information characteristics of users;

[0076] 3) Combine the three types of basic user information to form a comprehensive feature. The combined features are then input into a multilayer perceptron (MLP) to generate a personalized user profile, represented as follows:

[0077]

[0078] in, Indicates parameters Neural networks; A user profile is a comprehensive representation of a user's static information, dynamic behavioral information, and external status information.

[0079] M2: The learning resource construction module, based on a three-tiered semantic structure of concepts, skills, and context, performs semantic annotation and resource tagging on learning resources, and establishes a knowledge graph; according to user profiles, it adjusts the weights and association paths of resource tags in the knowledge graph.

[0080] This module is used to semantically annotate learning resources, construct knowledge graphs, and dynamically adjust node weights and resource association paths based on user profiles to support personalized learning resource recommendations.

[0081] The learning resources undergo semantic structuring, mapping them to a three-tiered tagging system of Concept, Skill, and Environment. Based on this, a knowledge graph is constructed, providing structured input for subsequent knowledge graph optimization and recommendation decisions.

[0082] Each learning resource is broken down into three categories of tags: concept tags, skill tags, and context tags. Concept tags represent the knowledge points or topics covered by the learning resource; skill tags represent the type of skill training the resource targets; and context tags represent the application scenario, difficulty level, or learning environment of the learning resource. Combining these three tags yields the comprehensive tag for the learning resource, represented as follows:

[0083]

[0084] in, Indicates learning resources Comprehensive tags; Indicates learning resources Conceptual tags; Indicates learning resources Skill tags; Indicates learning resources Contextual tags.

[0085] After all the labels for the learning resources have been extracted, the label vectors are processed into nodes and graph structures.

[0086] Map the concept tags, skill tags, and context tags of all learning resources to a set of concept tag nodes. Skill tag node set and context tag node set Then, define the relationships between nodes and establish an edge set based on the semantic relevance of the learning resources. To form a knowledge graph , is represented as:

[0087]

[0088] in, This represents the set of all nodes; each node is associated with a corresponding learning resource.

[0089] In this embodiment, there are learning resources A, B, and C. The concept label extracted from learning resource A is "function," the skill label is "differentiation," and the context label is "college entrance examination." The concept label extracted from learning resource B is "limit," the skill label is "computation," and the context label is "university." The concept label extracted from learning resource C is "probability," the skill label is "modeling," and the context label is "competition." After extracting labels from all learning resources, the concept labels are mapped to node sets. Skill tags are mapped to a set of nodes. Context labels are mapped to a set of nodes. An edge set is established based on the semantic relevance between learning resource tags. and An edge is established between them to indicate that they belong to the same learning resource. Multiple edges are also established in combination with other learning resources and tags to form the final knowledge graph.

[0090] In this module, based on the generated personalized user profiles, the weights and association paths of nodes (i.e., resource tags) in the knowledge graph are dynamically adjusted to ensure that the recommended learning resources better match the user's personalized needs.

[0091] Each node in the knowledge graph Each has a corresponding embedded representation The correlation between user profiles and nodes is calculated using cosine similarity, with the following formula:

[0092]

[0093] in, This represents the similarity between the user profile and the node representation, with a value range of... The larger the value, the stronger the correlation; update the node weights based on the calculation results. , is represented as:

[0094]

[0095] in, Indicates the weight of the node after the update; This indicates that the step size coefficient is updated to control the magnitude of weight adjustment. Through the node weight update operation, the nodes in the knowledge graph will automatically be promoted or demoted according to their matching degree with the user profile, making personalized recommendations more accurate.

[0096] After the node (resource tag) weight is updated, the weights of the edges in the knowledge graph will also be updated. For a connection node... and The edge, its weight The update formula is expressed as:

[0097]

[0098] in, This represents the updated edge weights; This represents the edge weight adjustment coefficient, which controls the update magnitude. The function representing the influence of the weights of the nodes at both ends of an edge on the edge, expressed as an average weight, reflects the average importance of the connected nodes.

[0099]

[0100] By updating edge weights, changes in node importance are propagated to the path structure, enabling the knowledge graph's associated paths to be dynamically adjusted based on user profiles.

[0101] Finally, the updated set of node weights With edge weight set Together, they form an adjusted knowledge graph, which is a user-specific personalized knowledge graph version, represented as:

[0102]

[0103] This knowledge graph serves as input to the recommendation decision module, guiding the recommendation strategy to prioritize traversing high-weight nodes and paths, thereby improving the accuracy and individual suitability of recommended resources.

[0104] M3: The decision-making module generates an initial recommendation strategy based on user profiles and the adjusted knowledge graph. It further obtains initial user feedback data and, combined with reinforcement learning, dynamically updates and adjusts the initial recommendation strategy to generate personalized learning resource recommendations.

[0105] This module combines user profiles and an adjusted knowledge graph to generate an initial set of learning resource recommendation strategies. Based on user interaction feedback, it dynamically optimizes these strategies using reinforcement learning methods, ultimately outputting personalized learning resource recommendations.

[0106] The user profile is integrated with the adjusted knowledge graph into a unified state representation vector. , used to generate recommendation strategies, is represented as:

[0107]

[0108] in, The state encoding function is represented here. In this embodiment, a graph neural network (GNN) is used as the state encoding function. Each node in the adjusted knowledge graph is assigned its latest weight and embedding vector representation. The user profile serves as the global feature input of the graph neural network. It undergoes multiple rounds of message propagation and aggregation with the feature information of each node in the graph. Through the attention mechanism, it automatically identifies tag nodes and paths that are highly related to the current user profile, so that the aggregated feature vector better reflects the user's focus in the current knowledge graph. Finally, after multiple layers of aggregation and encoding by the GNN, a unified state representation vector is output. This vector comprehensively reflects the personalized matching between the user profile and the knowledge graph.

[0109] Based on the state vector, a policy network is used to generate an initial recommendation probability distribution for candidate learning resources, thus forming an initial recommendation policy.

[0110] After generating the initial recommendation strategy, the decision-making module further updates the recommendation strategy dynamically using user feedback data and reinforcement learning methods, generating personalized learning resource recommendations.

[0111] As users engage in the initial learning phase based on the initial recommendation strategy, real-time feedback data is collected, including learning efficiency, knowledge mastery, and learning fatigue. This feedback data is quantified into immediate reward signals. , is represented as:

[0112]

[0113] in, This indicates a score for learning efficiency; This indicates an improvement in knowledge acquisition; Indicators representing learning fatigue levels; and These represent weighted coefficients for knowledge mastery and learning fatigue, used to balance the effects of knowledge improvement and fatigue.

[0114] The initial recommendation strategy is dynamically updated and adjusted using a reinforcement learning algorithm. The specific steps are as follows:

[0115] 1) Define the state ,action and rewards ,

[0116] state Generated based on the fusion of user profiles and adjusted knowledge graphs; Actions This indicates the learning resources selected based on the initial recommendation strategy; rewards. Constructed based on initial user feedback data;

[0117] 2) The optimization objective of the recommendation strategy is to maximize the cumulative reward for users during the initial learning phase. Based on this optimization objective, the objective function is defined. , is represented as:

[0118]

[0119] in, This represents the expected value of a random variable, i.e., a weighted average. This means that the cumulative rewards for all learning paths experienced by the user are weighted and averaged according to the initial recommendation strategy. This represents the total number of time steps in the initial learning phase; Indicates the first A discount factor for each time step is used to consider both short-term effects and long-term returns when optimizing a strategy, but without giving unlimited attention to the distant future. Indicates time step Instant rewards;

[0120] 3) Update parameters using the policy gradient method To optimize the objective function, the policy gradient formula is:

[0121]

[0122] in, Describe the objective function For strategy parameters The gradient; Representation strategy In state Select action The gradient of the logarithmic probability is used to measure the direction of improvement of the current policy;

[0123] The formula for updating the parameters is:

[0124]

[0125] in, This indicates the updated policy parameters; The learning rate controls the magnitude of each parameter update. Through multiple iterations, the strategy gradually learns the parameters that can recommend the most suitable learning resources to users under different learning states, thereby generating personalized learning resource recommendation results.

[0126] M4: The intervention module comprehensively evaluates the learning resource recommendation results based on the user's learning metrics within a preset period. Based on the evaluation results, it automatically triggers adjustments to the user profile and generates new learning resource recommendations.

[0127] The main function of this module is to comprehensively evaluate the user's learning performance at the end of a learning cycle, and automatically adjust the user profile, knowledge graph and recommendation strategy parameters based on the evaluation results, thereby generating new learning resource recommendation results and realizing closed-loop optimization of the recommendation system.

[0128] Set a fixed learning cycle At the end of the period, learning metrics are extracted from the user's learning records, namely the improvement in learning efficiency. Increase in knowledge mastery and learning fatigue At the same time, the three learning metrics were standardized for comprehensive evaluation.

[0129] During the learning cycle After completion, a comprehensive score is calculated using learning indicators. The formula is:

[0130]

[0131] in, , and This represents the weighting coefficients of the three learning indicators, which respectively control the importance of learning efficiency, knowledge mastery, and fatigue in the overall score; the overall score measures the user's overall learning effect, with the score increasing when learning efficiency and knowledge mastery improve, and decreasing when learning fatigue increases;

[0132] Set a rating threshold When the overall score falls below the score threshold, the user profile is automatically adjusted using the following formula:

[0133]

[0134] in, This indicates the adjusted user profile; This indicates the step size coefficient for updating user profiles, controlling the adjustment range; This represents the user profile correction amount obtained based on learning metrics within a set period. The user profile correction amount is a vector offset value that measures the difference between the user's actual performance and the user's state as understood by the system. It indicates the direction in which the user profile should be adjusted. First, the changing trends are extracted from the user's three learning metrics to form a standardized evaluation vector. , is represented as:

[0135]

[0136] in, This indicates that each item undergoes normalization; this step unifies the three indicators into a vector structure with strong numerical comparability. Then, a low-dimensional to high-dimensional mapping function is introduced to... Mapped to an adjustment vector consistent with the user profile dimension, it is represented as:

[0137]

[0138] in, This represents a simple fully connected neural network, trained using historical user information, with the aim of mapping learning biases to adjust the profile direction.

[0139] After the user profile is updated, the resource tag weights and association paths in the knowledge graph are adjusted according to the new user profile.

[0140] Finally, the parameters of the recommendation strategy network are fine-tuned using the new user profiles and the knowledge graph adjusted based on these profiles. To improve recommendation effectiveness, after updating user profiles, adjusting knowledge graphs, and fine-tuning strategy parameters, new learning resource recommendation results are generated. These newly generated results will serve as input for the next cycle, forming a closed-loop process of evaluation, adjustment, and recommendation.

[0141] This invention proposes an AI-based personalized learning resource recommendation system that achieves precise delivery of learning resources by constructing multi-dimensional user profiles, semantically structured knowledge graphs, and reinforcement learning recommendation strategies. The invention comprises a user information processing module, a learning resource construction module, a decision-making module, and an intervention module. Relying on time series modeling, attention mechanisms, dynamic graph adjustment, and strategy gradient optimization methods, it constructs a closed-loop adaptive recommendation framework, improving the accuracy of personalized recommendations and the scientific rigor of educational interventions.

[0142] In summary, this invention not only responds to changes in user interests in real time, but also has advantages such as sustainable optimization of recommendation strategies, interpretable modeling of resource structures, and dynamic intervention of user states. It solves the core technical problems of existing recommendation systems in terms of accuracy, dynamism, and personalized services, and has significant application and promotion value and research prospects.

[0143] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A personalized learning resource recommendation system based on artificial intelligence, characterized in that, The system includes: M1: User information processing module, including a collection unit and a modeling unit. The collection unit acquires the user's basic information, and the modeling unit performs feature modeling based on the user's basic information to generate a personalized user profile. M2: The learning resource construction module performs semantic annotation and resource tagging on learning resources based on a three-level semantic structure of concepts, skills, and contexts, and establishes a knowledge graph; according to the user profile, it adjusts the weights and association paths of resource tags in the knowledge graph. M3: Decision module, based on the user profile and the adjusted knowledge graph, generates an initial recommendation strategy, further obtains the user's initial feedback data, and combines reinforcement learning to dynamically update and adjust the initial recommendation strategy to generate personalized learning resource recommendation results; M4: Intervention module, which comprehensively evaluates the learning resource recommendation results based on the user's learning indicators within a preset period, and automatically triggers the adjustment of the user profile based on the evaluation results to generate new learning resource recommendation results; The basic information includes static information, dynamic behavioral information, and external state information; these three information are concatenated to form a comprehensive feature, which is then input into a multilayer perceptron to generate a personalized user profile. .

2. The personalized learning resource recommendation system based on artificial intelligence according to claim 1, characterized in that, The specific process of establishing the knowledge graph is as follows: 1) Extract concept tags for each learning resource Skill tags and context labels These are combined to form a comprehensive tag for the corresponding learning resource, represented as: in, Indicates learning resources Comprehensive tags; 2) After extracting the tags from all learning resources, map them to a set of concept tag nodes. Skill tag node set and context tag node set And based on the semantic relevance of the learning resources, an edge set is established. To form a knowledge graph , represented as: in, This represents the set of all nodes.

3. The personalized learning resource recommendation system based on artificial intelligence according to claim 2, characterized in that, The resource tag weights are the node weights of the knowledge graph. The correlation between user profiles and nodes is calculated using cosine similarity, with the following formula: in, Represents each node in the knowledge graph The corresponding embedding representation, The similarity between the user profile and the node representation is used to represent the similarity; the node weights are updated based on the cosine similarity calculation results. , represented as: in, Indicates the weight of the node after the update; This indicates that the step size coefficient is being updated; After the node weights of the knowledge graph are updated, for a connection node... and Edges, edge weights The data will be updated synchronously, as shown below: in, This represents the updated edge weights; This represents the edge weight adjustment coefficient; The function representing the influence of the weights of the nodes at both ends of an edge on the edge is expressed as: By updating the edge weights, changes in node importance are propagated to the association paths of the knowledge graph. These paths are then adjusted based on user profiles, ultimately forming an adjusted knowledge graph. , represented as: in, This represents the new set of node weights; This represents the updated set of edge weights.

4. The personalized learning resource recommendation system based on artificial intelligence according to claim 3, characterized in that, The specific technical steps of the reinforcement learning are as follows: 1) Define the state ,action and rewards , The reward is generated based on the user's initial feedback data and is represented as follows: in, This indicates a score for learning efficiency; This indicates an improvement in knowledge acquisition; Indicators representing learning fatigue levels; and These represent the weighting coefficients for knowledge mastery and learning fatigue, respectively. 2) Define the objective function , represented as: in, This means that the cumulative rewards for all learning paths experienced by the user are weighted and averaged according to the initial recommendation strategy. This represents the total number of time steps in the initial learning phase; Indicates the first Discount factor for each time step; Indicates time step Instant rewards; 3) Update parameters using the policy gradient method The objective function is optimized, and the policy gradient formula is: in, Describe the objective function For strategy parameters The gradient; Representation Strategy In state Select action The gradient of the logarithmic probability; The formula for updating the parameters is: in, This indicates the updated policy parameters; This represents the learning rate; through multiple iterations, parameters are obtained to recommend the most suitable learning resources to users, generating personalized learning resource recommendation results.

5. The personalized learning resource recommendation system based on artificial intelligence according to claim 4, characterized in that, The learning metrics are respectively the improvement in learning efficiency. Increase in knowledge mastery and learning fatigue The three learning metrics were standardized, and a comprehensive score was calculated. The formula is: in, , and This represents the weighting coefficients for the three learning metrics; When the overall score falls below a set score threshold, the user profile is adjusted using the following formula: in, This indicates the adjusted user profile; This represents the step size coefficient for updating the user profile. This indicates the amount of user profile correction obtained based on learning metrics within a set period.

Images