Music teaching intelligent recommendation method and system based on knowledge graph

Abstract

The application discloses a music teaching intelligent recommendation method and system based on a knowledge graph, belongs to the technical field of knowledge graphs, and comprises the following steps: constructing a knowledge graph with a three-layer node structure according to evidence center design and a music skill ontology based on multi-modal interaction data collected from target users; binding evolution attribute parameters to the nodes and establishing a time axis index, and performing correlation strength prediction and edge updating through a graph attention network; loading an improved forgetting curve model distinguishing between declarative memory and procedural memory, and calculating a cognitive load comprehensive coefficient by fusing physiological signal and subjective scale data; generating personalized recommendation results through three parallel modules of forgetting compensation, skill progression and cognitive load adjustment; and generating a push scheme according to the recommendation results and performing closed-loop updating. Through deep integration of music education theory, cognitive science principles and graph neural network technology, the application significantly improves the scientific nature and effectiveness of music teaching.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph technology, specifically referring to a music teaching intelligent recommendation method and system based on knowledge graphs. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent education systems have been widely used in the field of music teaching. In existing technologies, teaching recommendation methods based on knowledge graphs usually adopt static knowledge representation, constructing music knowledge points such as scales, chords, and musical forms into a fixed relationship network, and recommending resources through preset rules or simple similarity calculations.

[0003] However, existing intelligent recommendation methods for music teaching still have certain shortcomings. Existing technologies suffer from simplistic knowledge representation, often focusing only on explicit music theory knowledge points. They struggle to effectively model and track the developmental patterns of procedural knowledge and sensorimotor coordination abilities, such as performance skills, rhythm perception, and pitch control, leading to a disconnect between recommended content and actual abilities. Furthermore, their knowledge graphs lack temporal evolution capabilities and pedagogical modeling of skill development sequences and transfer relationships. Simultaneously, existing methods neglect cognitive science principles, failing to distinguish between the different forgetting patterns of declarative and procedural memory. Cognitive load measurement relies solely on behavioral indicators, and recommendation strategies employ single-dimensional similarity matching, lacking multi-objective collaborative optimization for review and consolidation, skill advancement, and load adjustment. Moreover, existing patent texts lack specific parameter settings, dataset descriptions, and controlled experimental designs, and technical terminology definitions are vague. Therefore, this paper proposes an intelligent recommendation method and system for music teaching based on knowledge graphs. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph-based intelligent recommendation method and system for music teaching, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a knowledge graph-based intelligent recommendation method for music teaching, comprising the following steps:

[0006] S1. Based on the multimodal interaction data collected from the target users, perform semantic alignment and feature extraction to form a set of multimodal teaching feature vectors;

[0007] S2. Based on the set of multimodal teaching feature vectors, a framework based on the evidence center is designed. Supervised feature mapping is performed through skill behavior indicators predefined by domain experts. A knowledge graph with a three-layer node structure is constructed with reference to the music skill ontology to generate a multimodal fusion knowledge graph for music teaching.

[0008] S3. Based on the multimodal fusion knowledge graph of music teaching, bind evolution attribute parameters to each node and establish a time axis index to construct a dynamic evolution knowledge graph. Use an incremental graph update strategy to construct the dynamic evolution knowledge graph.

[0009] S4. Based on the dynamically evolving knowledge graph and the multimodal teaching feature vector set, perform association strength prediction and edge update through graph attention network (GAT) to complete implicit skill association reasoning.

[0010] S5. Based on the updated implicit skill node structure and dynamically evolving knowledge graph, load an improved forgetting curve model that distinguishes between declarative memory and procedural memory, and integrate physiological signals and subjective scale data to calculate the cognitive load comprehensive coefficient, and perform cognitive state perception and cognitive load measurement.

[0011] S6. Based on the generated cognitive state vector and the updated implicit skill node structure, personalized recommendation results are generated through the weighted fusion of three parallel modules: forgetting compensation, skill advancement, and cognitive load adjustment.

[0012] S7. Based on the generated personalized recommendation results, generate a push scheduling plan and push resources to the target users. At the same time, use the real-time collected feedback data as a graph update trigger signal to input the dynamic evolution knowledge graph and execute dynamic evolution push and closed-loop update.

[0013] Preferably, in step S1, the target user's performance audio data, performance posture video data, touch interaction log data, and physiological signal data are simultaneously collected through a microphone array, camera equipment, touch screen interactive terminal, and wearable physiological sensor, respectively. The audio data is subjected to short-time Fourier transform to extract Mel frequency cepstral coefficients, pitch deviation sequences, and rhythmic stability features. The video data is subjected to posture keypoint detection to extract hand shape and sitting posture standardization features. The touch interaction data is subjected to temporal pattern recognition to extract practice duration distribution, fingering error frequency, and accuracy sequences. The physiological signals are extracted to obtain time-domain and frequency-domain indicators of heart rate variability (HRV), as well as phasic and tonic components of skin conductance response (GSR). The above three types of features are timestamped and semantically mapped according to preset teaching label fields. The multimodal features within the same time window are spliced ​​and fused into a unified multimodal teaching feature vector set through a cross-modal attention mechanism.

[0014] Preferably, in step S2, based on the generated multimodal teaching feature vector set, explicit music theory knowledge point labels, teaching repertoire labels, and performance technique labels are extracted from the feature vectors. These labels are used as the first layer of explicit knowledge nodes. Explicit semantic association edges are established based on the music education ontology constructed with reference to the NAFME skills standard and my country's compulsory education art curriculum standards, clarifying relationships such as pre-dependencies, developmental sequences, positive transfer, and negative transfer between skills. An evidence-centered design framework is adopted: first, music education experts define observable behavioral indicators corresponding to implicit skills; then, specific values ​​of these behavioral indicators are extracted from the multimodal features, and a supervised classifier maps these indicators to preset skill nodes. Each skill node corresponds to an interpretable music skill concept, and its type is clarified based on Gagné learning results. Nodes for dimensions such as music aesthetic perception, style understanding, and expression interpretation are added. An implicit skill knowledge subgraph is constructed by calculating the weighted similarity between the embedding vectors of each implicit skill node based on the music skill development sequence. The association strength weight between implicit skill nodes is implemented as follows:

[0015] ,

[0016] In the formula, This represents the final association strength between latent skill nodes i and j, with a value range of [0, 1], and serves as the weight of the edge in the knowledge subgraph. This represents the modified line cell activation function. Negative similarity is truncated to 0. The feature embedding vectors representing latent skill nodes i and j are obtained through supervised learning from the mapping between multimodal features and expert behavior indicators. Representing vectors and L2 norm, Indicates the skill topology decay coefficient. , This represents the position encoding vectors of skills i and j within the preset music skill body. Represents the position encoding vector and Euclidean distance, The standard deviation of location distance is represented; a third-layer cognitive state node is created for each target user, including a forgetting rate parameter node and a cognitive load measurement node. Each cognitive state node is associated with the corresponding explicit knowledge node through a timestamp, forming a multimodal fusion knowledge graph for music teaching with interconnected three-layer node structures.

[0017] Preferably, in step S3, based on the multimodal fusion knowledge graph of music teaching, evolutionary attribute parameters are bound to each explicit knowledge node, implicit skill node, and cognitive state node in the graph. The evolutionary attribute parameters include node activity decay factor, knowledge decay rate parameter, and interaction influence coefficient. A global time axis index is constructed with a fixed time step as the unit, and the evolutionary attribute parameter value sequence of each node is recorded at different time steps. An incremental update strategy is adopted: the parameters of nodes in the local subgraph directly affected by the current interaction feedback are recalculated only. An evolutionary function model is pre-constructed. The evolutionary function model, based on the student interaction feedback data at the current time step, comprehensively considers the node activity decay factor, knowledge decay rate parameter, interaction influence coefficient, and topological association weight between nodes. It performs time-series modeling of the historical parameter sequence through a gated cyclic unit to achieve nonlinear weighted fusion, and dynamically updates the evolutionary attribute parameters of each node in the time axis index to generate a dynamic evolutionary knowledge graph with time dimension evolution capability.

[0018] Preferably, in step S4, based on the dynamically evolving knowledge graph and the multimodal teaching feature vector set, a subset of features related to implicit skills is extracted from the multimodal teaching feature vectors, and adaptively fused with the current embedding vector of the implicit skill node in the dynamically evolving knowledge graph through a cross-modal attention fusion mechanism to form the input features of the graph attention network (GAT); a graph attention network containing a multi-head attention layer, a graph convolutional layer, and an edge prediction layer is initialized, wherein the edge prediction layer uses a bilinear form to calculate the predicted value of the association strength between implicit skill nodes; the fused features are input into the GAT, and the neighbor node information is aggregated by weighted attention coefficients to obtain the updated representation of each implicit skill node, and then the association strength prediction matrix between node pairs is generated using a bilinear decoder.

[0019] Preferably, in step S4, the prediction matrix is ​​optimized for edge structure according to a preset dynamic threshold strategy: the upper and lower quartiles are calculated based on the weight distribution of the edges of the implicit skill nodes in the current knowledge graph as adaptive thresholds, implemented as follows:

[0020] ,

[0021] ,

[0022] ,

[0023] In the formula, This represents the set of fusion weights at time t. This represents the time smoothing coefficient, with a value range of [0, 1]. This represents the actual affinity strength of edge (k, l) at the previous time t-1. This represents the predicted association strength of edge (k, l) at the current time t. Let the set of edges at time t be represented. This represents the threshold for determining candidate edges. This represents the first quartile at time t. This represents the third quartile at time t. This represents the position parameter of the first quantile. This represents the position parameter of the third quantile. This represents the interquartile range at time t;

[0024] When the predicted value exceeds the upper quartile boundary When the predicted value is below the lower quartile boundary, establish a new strong correlation edge; At that time, the edge weights are weakened, where and The asymmetric threshold sensitivity coefficient is determined adaptively based on historical data using an optimization method. The strictness of strong association enhancement and weak association decay is controlled separately. The updated edge weights are exponentially smoothed using a time decay factor. The optimized graph structure is written back to the dynamically evolving knowledge graph to complete the dynamic reasoning and evolution of implicit skill associations.

[0025] Preferably, in step S5, based on the updated implicit skill node structure and the generated dynamic evolution knowledge graph, a cognitive state vector is initialized for each target user; for each explicit knowledge node, according to the preset knowledge decay rate parameter bound to that node in the dynamic evolution knowledge graph, combined with the difference between the most recent review timestamp and the current time, a preset improved forgetting curve model is loaded to calculate the current memory retention rate, which is implemented as follows:

[0026] ,

[0027] In the formula, This represents the memory retention rate of explicit knowledge node i at time t. This represents the preset knowledge decay rate parameter, with higher values ​​for declarative knowledge nodes and lower values ​​for procedural knowledge nodes. This represents the difference between the current time and the most recent review timestamp. Indicates the reference time constant. Indicates the individual difference adjustment coefficient. Indicates the degree of prior knowledge;

[0028] The mastery score is updated based on memory retention rate; the practice duration distribution and fingering error frequency of touch interaction data, rhythm stability deviation and pitch deviation sequence of audio data, and heart rate variability and skin conductance response characteristics of physiological signals are extracted from the multimodal teaching feature vector set; combined with the questionnaire results of the subjective cognitive load scale, a comprehensive cognitive load coefficient is generated by weighted fusion through a multi-task learning network; the updated mastery score sequence, activation status identifier and comprehensive cognitive load coefficient are integrated into the cognitive state perception result at the current time step.

[0029] Preferably, in step S6, three recommendation modules run in parallel based on the cognitive state vector and the updated implicit skill node structure:

[0030] The forgetting compensation module, based on the memory retention rate of each explicit knowledge node in the cognitive state vector and combined with the staggered practice effect, marks nodes that are below a preset threshold and whose time since the last practice exceeds the optimal interval as nodes that need to be reviewed, and retrieves the corresponding teaching track nodes and practice resource nodes from the dynamically evolving knowledge graph.

[0031] The skill advancement module, based on the set of currently active implicit skill nodes in the implicit skill node structure, calculates the association strength between the nodes and inactive nodes, introduces the zone of proximal development constraint to limit the difficulty span, selects the candidate implicit skill nodes with the greatest advancement potential, and expands their associated explicit knowledge nodes into a skill advancement recommendation list.

[0032] The cognitive load adjustment module dynamically adjusts the upper limit of the number of recommended resources and the difficulty span coefficient based on the comprehensive cognitive load coefficient in the cognitive state vector;

[0033] The recommendation lists output by the three modules are weighted and merged according to preset initial weights to generate an initial recommendation candidate set. Then, based on the actual adoption rate and completion rate of various recommendation resources recorded in the target user's historical interaction feedback data, as well as the learning gain and skill transfer test results, the weight coefficients of the three modules are adaptively fine-tuned through Bayesian personalized ranking, and finally, personalized recommendation results are output.

[0034] Preferably, in step S7, based on the personalized recommendation results, a push scheduling scheme is generated according to the resource type, difficulty level, and estimated learning time of each recommended resource, combined with the cognitive load comprehensive coefficient in the cognitive state vector.

[0035] The push scheduling scheme includes a resource push sequence arranged in chronological order, suggested learning intervals between resources, and push priority weights for each resource. Recommended resources are pushed to target users according to the push scheduling scheme, and feedback data is collected in real time during the push execution process. Feedback data includes resource adoption identifiers, resource completion progress, practice accuracy sequences, multimodal interaction feedback scores, and knowledge test scores before and after learning. The collected feedback data is used as a graph update trigger signal, input into the dynamic evolution knowledge graph, and the evolution attribute parameters of the affected nodes are updated in real time according to the update rules defined in the evolution function model. The updated node state is written to the corresponding position of the timeline index, completing the dynamic evolution push and closed-loop update.

[0036] The music teaching intelligent recommendation system based on knowledge graphs, implemented using the above methods, includes a multimodal feature extraction module, a knowledge graph construction module based on evidence center design and music skill ontology, an incremental dynamic evolution graph construction module, a graph attention network implicit skill reasoning module, a multimodal cognitive state perception module, a three-way fusion recommendation generation module, and a dynamic push closed-loop update module.

[0037] Preferably, the music teaching intelligent recommendation system based on knowledge graph includes a multimodal feature extraction module, a knowledge graph construction module based on evidence center design and music skill ontology, an incremental dynamic evolution graph construction module, a graph attention network implicit skill reasoning module, a multimodal cognitive state perception module, a three-way fusion recommendation generation module, and a dynamic push closed-loop update module.

[0038] Implicit skills, procedural knowledge, and perceptual-motor abilities refer to the abilities in musical performance that are difficult to teach directly through text or images and require long-term practice and internalization. These include the gradual control of dynamic changes in fingering, the fine-tuning of timbre, and the intrinsic driving force of rhythm and movement. The skills ontology is obtained through supervised mapping based on observable behavioral indicators predefined by music education experts according to the evidence center design framework. Each node corresponds to a clear teaching concept, and the skills ontology is established with reference to the NAFME skills standards and my country's compulsory education art curriculum standards.

[0039] Evolutionary attribute parameters include: node activity decay factor (controlling the rate at which a node's influence in recommendation decisions decreases over time), knowledge decay rate parameter (distinguishing between declarative knowledge nodes and procedural knowledge nodes, with the former having a higher decay rate and the latter a lower one), and interaction influence coefficient (recording the collaborative learning effect between nodes). These parameters are dynamically updated using incremental gated recurrent units, and are only affected by local interactions, avoiding global recalculation.

[0040] Cognitive Load Composite Coefficient: This is a scalar obtained by integrating subjective scales, physiological indicators (heart rate variability HRV, skin conductance response GSR), and behavioral performance (practice duration distribution, error frequency, rhythm deviation) through a multi-task learning network. It is used to quantify the student's current level of cognitive resource utilization.

[0041] Music Skills Ontology: Referring to Bloom's taxonomy of cognitive objectives, Gagné's taxonomy of learning outcomes, and the NAfME skill standards in music education, a systematic music skills ontology is constructed, clarifying the pre-requisite dependencies, developmental sequences, and positive / negative transfer relationships between skills, providing a disciplinary foundation for knowledge graphs.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. In order to address the problems of existing music teaching recommendation systems being unable to accurately capture the development patterns of users' implicit skills, lacking dynamic adaptability and cognitive science support, and improving the personalization level and learning efficiency of music teaching, this invention constructs a dynamic evolutionary knowledge graph with a three-layer node structure, integrates multimodal interactive data, and combines cognitive state perception and a three-way parallel recommendation mechanism to achieve accurate modeling and adaptive recommendation of music skill development.

[0044] 2. This invention constructs a three-layer knowledge graph structure comprising explicit knowledge nodes, implicit skill nodes based on supervised mapping within the ECD framework, and cognitive state nodes. It establishes a music skill ontology based on the NAFME skill standards and art curriculum standards. Through unsupervised clustering, multimodal features are abstracted into implicit skill nodes, and topological relationships between skills are established. This enables systematic modeling of implicit abilities in music performance, such as hand gestures, rhythm perception, and pitch control, which are difficult to articulate. This allows the teaching system to understand the skill gaps students expose in actual performances, significantly improving the relevance of music teaching.

[0045] 3. This invention introduces a dynamic evolutionary knowledge graph mechanism, and through incremental update strategies and GRU temporal modeling, realizes the autonomous evolution capability of the knowledge graph as the learning process progresses; the system can dynamically adjust node attribute parameters and edge association strength based on students' real-time interactive feedback, accurately capturing the nonlinear characteristics of skill development; combined with the predictive ability of graph attention networks for implicit skill associations, the system can proactively discover potential skill transfer paths and lay out advanced content in advance before students are even aware of it; making the teaching system adaptable like a teacher, able to perceive changes in learning status and adjust teaching strategies in a timely manner;

[0046] 4. This invention utilizes a three-pronged parallel recommendation module—forgetting compensation, skill advancement, and cognitive load adjustment—to deeply integrate cognitive science principles, distinguish the forgetting patterns of declarative and procedural memory, and introduce the spacing effect and staggered practice effect to achieve intelligent recommendations that conform to human learning patterns. The system accurately predicts knowledge retention status through an improved forgetting curve model and dynamically adjusts recommendation intensity based on a comprehensive cognitive load coefficient. The three modules work collaboratively through a weighted fusion mechanism, ensuring the consolidation of basic knowledge, promoting the step-by-step improvement of skills, and maintaining an appropriate level of cognitive challenge. The multi-dimensional recommendation strategy effectively balances the relationship between review and new learning, basic and advanced learning, and intensity and load, significantly improving learning efficiency and sustainability. Attached Figure Description

[0047] Figure 1 The following is the operation flow of the knowledge graph-based intelligent recommendation method for music teaching in this invention. Figure 1 ;

[0048] Figure 2 The following is the operation flow of the knowledge graph-based intelligent recommendation method for music teaching in this invention. Figure 2 ;

[0049] Figure 3 The following is the operation flow of the knowledge graph-based intelligent recommendation method for music teaching in this invention. Figure 3 ;

[0050] Figure 4 The following is the operation flow of the knowledge graph-based intelligent recommendation method for music teaching in this invention. Figure 4 ;

[0051] Figure 5 This is a schematic diagram of the structure of the knowledge graph-based intelligent recommendation system for music teaching according to the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example

[0054] Please see Figures 1-5 As shown, the present invention provides a technical solution comprising the following steps:

[0055] S1. Based on the multimodal interaction data collected from the target users, perform semantic alignment and feature extraction to form a set of multimodal teaching feature vectors;

[0056] S2. Based on the set of multimodal teaching feature vectors, construct a knowledge graph with a three-layer node structure to generate a multimodal fusion knowledge graph for music teaching;

[0057] S3. Based on the multimodal fusion knowledge graph of music teaching, bind evolution attribute parameters to each node and establish a time axis index to construct a dynamic evolution knowledge graph;

[0058] S4. Based on the dynamically evolving knowledge graph and the multimodal teaching feature vector set, perform association strength prediction and edge update through graph attention network to complete implicit skill association reasoning;

[0059] S5. Based on the updated implicit skill node structure and dynamically evolving knowledge graph, load an improved forgetting curve model that distinguishes between declarative memory and procedural memory, and integrate physiological signals and subjective scale data to calculate the cognitive load comprehensive coefficient, and perform cognitive state perception and cognitive load measurement.

[0060] S6. Based on the generated cognitive state vector and the updated implicit skill node structure, personalized recommendation results are generated through the weighted fusion of three parallel modules: forgetting compensation, skill advancement, and cognitive load adjustment.

[0061] S7. Based on the generated personalized recommendation results, generate a push scheduling plan and push resources to the target users. At the same time, use the real-time collected feedback data as a graph update trigger signal to input the dynamic evolution knowledge graph and execute dynamic evolution push and closed-loop update.

[0062] In this embodiment, in step S1, the target user's performance audio data, performance posture video data, and touch interaction log data are simultaneously collected through a microphone array, camera equipment, and touch screen interactive terminal, respectively. The audio data is subjected to short-time Fourier transform to extract Mel frequency cepstral coefficients, pitch deviation sequences, and rhythm stability features. The video data is subjected to posture key point detection to extract hand shape and sitting posture standard features. The touch interaction data is subjected to temporal pattern recognition to extract practice duration distribution, fingering error frequency, and accuracy sequences. The above three types of features are then timestamped and semantically mapped according to preset teaching label fields, and the multimodal features within the same time window are spliced ​​and merged into a unified multimodal teaching feature vector set.

[0063] In this embodiment, in step S2, based on the generated multimodal teaching feature vector set, explicit music theory knowledge point labels, teaching repertoire labels, and performance technique labels are extracted from the feature vectors. These labels are used as the first layer of explicit knowledge nodes, and explicit semantic association edges are established according to the preset music teaching ontology relationships, such as inclusion in the preceding dependency class. Unsupervised clustering analysis is performed on the skill features that cannot be explicitly expressed in the multimodal teaching feature vectors, and the cluster centers are abstracted as the second layer of implicit skill nodes. An implicit skill knowledge subgraph is constructed by calculating the cosine similarity between the embedding vectors of each implicit skill node, wherein the association strength weight between implicit skill nodes is implemented as follows:

[0064] ,

[0065] In the formula, This represents the final association strength between latent skill nodes i and j, with a value range of [0, 1], and serves as the weight of the edge in the knowledge subgraph. This represents the modified line cell activation function. Negative similarity is truncated to 0. Let represent the feature embedding vectors of latent skill nodes i and j, and the cluster center vectors extracted from multimodal features through unsupervised clustering. Representing vectors and L2 norm, Indicates the skill topology decay coefficient. , This represents the position encoding vectors of skills i and j within the preset music skill body. Represents the position encoding vector and Euclidean distance, The standard deviation of location distance is represented; a third-layer cognitive state node is created for each target user, including a forgetting rate parameter node and a cognitive load measurement node. Each cognitive state node is associated with the corresponding explicit knowledge node through a timestamp, forming a multimodal fusion knowledge graph for music teaching with interconnected three-layer node structures.

[0066] In this embodiment, in step S3, based on the multimodal fusion knowledge graph of music teaching, evolutionary attribute parameters are bound to each explicit knowledge node, implicit skill node, and cognitive state node in the graph. The evolutionary attribute parameters include node activity decay factor, knowledge decay rate parameter, and interaction influence coefficient. A global time axis index is constructed with a fixed time step as the unit, and the evolutionary attribute parameter value sequence of each node at different time steps is recorded. An evolutionary function model is pre-constructed. Based on the student interaction feedback data at the current time step, the evolutionary function model comprehensively considers the node activity decay factor, knowledge decay rate parameter, interaction influence coefficient, and topological association weight between nodes, and dynamically updates the evolutionary attribute parameters of each node in the time axis index through a nonlinear weighted fusion mechanism to generate a dynamic evolutionary knowledge graph with time dimension evolution capability.

[0067] Specifically, an evolutionary function model is pre-constructed: a three-layer node structure evolutionary attribute parameter system is determined, including node activity decay factor, knowledge decay rate parameter, and interaction influence coefficient; a global time axis index is established with a fixed time step, recording the historical parameter sequence for each node; a non-linear weighted fusion mechanism is preset to dynamically combine the node activity decay factor, knowledge decay rate parameter, interaction influence coefficient, and topological association weights between nodes; real-time student interaction feedback data is used as input signals, and the evolutionary attribute parameters of each node in the time axis index are iteratively updated through the fusion mechanism to generate a dynamic evolutionary knowledge graph with time-dimensional evolution capabilities.

[0068] In this embodiment, in step S4, based on the dynamically evolving knowledge graph and the multimodal teaching feature vector set, a subset of features related to implicit skills is extracted from the multimodal teaching feature vectors. This subset is then adaptively fused with the current embedding vector of the implicit skill nodes in the dynamically evolving knowledge graph through a cross-modal attention fusion mechanism to form the input features of the graph attention network. A graph attention network containing a multi-head attention layer, a graph convolutional layer, and an edge prediction layer is initialized. The edge prediction layer uses a bilinear form to calculate the predicted association strength between implicit skill nodes, implemented as follows:

[0069] ,

[0070] In the formula, This represents the predicted correlation strength between explicit knowledge node i and implicit skill node j at time t. This indicates the Sigmoid activation function, which maps the output to the interval [0, 1]. This represents the transpose of the weight vector w. This represents a multilayer perceptron. This represents the gate vector for node pair (i, j). This represents the temporal feature vector of explicit knowledge node i at time t. Represents the static skill feature vector of latent skill node j;

[0071] The fused features are propagated forward layer by layer, and the features are transformed by the nonlinear activation function of the hidden layer. The predicted values ​​are then mapped to the [0, 1] interval by the normalization function in the output layer to obtain the prediction matrix of the association strength between the latent skill nodes.

[0072] In this embodiment, in step S4, the edge structure of the prediction matrix is ​​optimized according to a preset dynamic threshold strategy: the upper quartile and lower quartile are calculated based on the weight distribution of the edges of the implicit skill nodes in the current knowledge graph as adaptive thresholds, which is implemented as follows:

[0073] ,

[0074] ,

[0075] ,

[0076] In the formula, This represents the set of fusion weights at time t. This represents the time smoothing coefficient, with a value range of [0, 1]. This represents the actual affinity strength of edge (k, l) at the previous time t-1. This represents the predicted association strength of edge (k, l) at the current time t. Let the set of edges at time t be represented. This represents the threshold for determining candidate edges. This represents the first quartile at time t. This represents the third quartile at time t. This represents the position parameter of the first quantile. This represents the position parameter of the third quantile. This represents the interquartile range at time t;

[0077] When the predicted value exceeds the upper quartile boundary When the predicted value is below the lower quartile boundary, new strong correlation edges are established or the weights of existing edges are increased proportionally to the predicted strength. At this time, weaken the edge weight or remove redundant edges, where and The asymmetric threshold sensitivity coefficient controls the strictness of strong association enhancement and weak association decay, respectively. The updated edge weights are exponentially smoothed by the time decay factor, and the optimized graph structure is written back to the dynamic evolution knowledge graph to complete the dynamic reasoning and evolution of implicit skill associations.

[0078] In this embodiment, in step S5, based on the updated implicit skill node structure and the generated dynamic evolution knowledge graph, a cognitive state vector is initialized for each target user. The cognitive state vector includes the mastery score of each explicit knowledge node, the recent review timestamp of each explicit knowledge node, and the activation status identifier of each implicit skill node. For each explicit knowledge node, according to the preset knowledge decay rate parameter bound to the node in the dynamic evolution knowledge graph, combined with the difference between the recent review timestamp and the current time, a preset improved forgetting curve model is loaded to calculate the current memory retention rate, which is implemented as follows:

[0079] ,

[0080] In the formula, This represents the memory retention rate of explicit knowledge node i at time t. This represents the preset knowledge decay rate parameter. This represents the difference between the current time and the most recent review timestamp. Indicates the reference time constant. Indicates the individual difference adjustment coefficient. Indicates the degree of prior knowledge;

[0081] The mastery score is updated based on the memory retention rate; the practice duration distribution and fingering error frequency of the touch interaction data, as well as the rhythm stability deviation and pitch deviation sequence of the audio data, are extracted from the multimodal teaching feature vector set and weighted and fused to generate a cognitive load comprehensive coefficient; the updated mastery score sequence, activation status identifier and cognitive load comprehensive coefficient are integrated into the cognitive state perception result of the current time step.

[0082] In this embodiment, in step S6, three recommendation modules run in parallel based on the cognitive state vector and the updated implicit skill node structure:

[0083] The forgetting compensation module marks nodes that are below a preset threshold and whose time since the last practice exceeds the optimal interval as nodes that need to be reviewed, based on the memory retention rate of each explicit knowledge node in the cognitive state vector. It also retrieves the corresponding teaching track nodes and practice resource nodes from the dynamically evolving knowledge graph.

[0084] The skill advancement module, based on the set of currently active hidden skill nodes in the hidden skill node structure, calculates the association strength between these nodes and inactive nodes, and achieves the following:

[0085] ,

[0086] In the formula, This represents the advanced potential score of the unactivated hidden skill node u. This represents the set of currently active hidden skill nodes. This represents the association strength between active node i and inactive node u. This represents the sensitivity coefficient for memory retention. This represents the skill distance decay coefficient. This represents the shortest path distance between nodes i and u in a dynamically evolving knowledge graph. This represents the exponential decay term, which filters out the candidate implicit skill nodes with the greatest potential for advancement, and expands their associated explicit knowledge nodes into a skill advancement recommendation list.

[0087] The cognitive load adjustment module dynamically adjusts the upper limit of the number of recommended resources and the difficulty span coefficient based on the comprehensive cognitive load coefficient in the cognitive state vector;

[0088] The recommendation lists output by the three modules are weighted and merged according to preset initial weights to generate an initial recommendation candidate set. Then, based on the actual adoption rate and completion rate of various recommendation resources recorded in the historical interaction feedback data of the target user, the weight coefficients of the three modules are adaptively fine-tuned through Bayesian personalized ranking to finally output personalized recommendation results.

[0089] In this embodiment, in step S7, based on the personalized recommendation results, a push scheduling scheme is generated according to the resource type, difficulty level and estimated learning time of each recommended resource, combined with the cognitive load comprehensive coefficient in the cognitive state vector.

[0090] The push scheduling scheme includes a resource push sequence arranged in chronological order, suggested learning intervals between resources, and push priority weights for each resource. Recommended resources are pushed to target users according to the push scheduling scheme, and feedback data is collected in real time during the push execution process. This feedback data includes resource adoption identifiers, resource completion progress, practice accuracy sequences, and multimodal interaction feedback scores. The collected feedback data is used as a graph update trigger signal, input into the dynamic evolution knowledge graph, and the evolution attribute parameters of the affected nodes are updated in real time according to the update rules defined in the evolution function model. The updated node state is then written to the corresponding position on the timeline index, completing the dynamic evolution push and closed-loop update.

[0091] In this embodiment, the knowledge graph-based intelligent recommendation system for music teaching includes a multimodal feature extraction module, a knowledge graph construction module based on evidence center design and music skill ontology, an incremental dynamic evolution graph construction module, a graph attention network implicit skill reasoning module, a multimodal cognitive state perception module, a three-way fusion recommendation generation module, and a dynamic push closed-loop update module.

[0092] Example: Parameter setting and control experiment in a piano beginner teaching scenario:

[0093] Parameter settings: Taking piano beginner teaching as an example, the time step is set to an appropriate value. Knowledge decay rate parameter: a higher value for music theory nodes, a lower value for fingering nodes, and a middle value for repertoire nodes. The interval effect factor is set to an appropriate constant, and the number of recent reviews is the number of reviews in the past few days. The physiological signal sampling rate is set to a value that meets the requirements of heart rate variability analysis. Heart rate variability features are extracted using time-domain and frequency-domain indicators. Skin conductance response is low-pass filtered to extract tetanic and phase components. The graph attention network has an appropriate number of attention heads, each with moderate dimensions. The learning rate is determined through optimization, and early stopping is used to prevent overfitting. The asymmetric coefficients in the dynamic threshold are determined on the validation set through Bayesian optimization. The recommendation module has equal initial weights, and is updated once after a certain number of interaction steps using Bayesian personalized ranking. The NASA-TLX simplified version collects data via a terminal pop-up after each practice session, using several dimensions and a Likert scale.

[0094] Dataset Description: To verify the effectiveness of this invention, a certain number of piano beginners were recruited and randomly divided into several groups. The control group used traditional collaborative filtering recommendation; another group used static knowledge graph recommendation (no dynamic evolution, no physiological signals, no interval effect); the experimental group used the method of this invention. The experimental period was set for several weeks, with a fixed number of practice sessions per week and a fixed duration for each session. The teaching content was basic piano skills. The collected data included: periodic skill tests, recommendation adoption rate, cognitive load self-assessment, and retention tests.

[0095] The experimental results showed that the experimental group had a significantly higher overall score improvement rate in the skills test than the control group; the experimental group had the highest adoption rate of recommended resources; the experimental group had the smallest fluctuation in the self-rated cognitive load score, indicating that the load adjustment was effective; and the experimental group had the highest retention rate of test scores.

[0096] Example Flow Description: Upon system startup, a knowledge graph constructed based on the evidence center design framework and music skill ontology is preloaded. During the user's first practice session, multimodal sensors collect data, and feature vectors are generated after cross-modal attention fusion. The graph attention network predicts implicit skill associations and updates graph edges. Node parameters are updated incrementally. The forgetting curve is improved to calculate memory retention rate. Cognitive load is calculated by combining physiological signals and scales. A three-way recommendation module generates a recommendation list. Bayesian personalized ranking optimizes weights. Push scheduling and feedback collection form a closed loop.

[0097] The calculation method for the association strength weight between latent skill nodes is as follows: The final association strength between latent skill nodes i and j is obtained through two steps. First, the cosine similarity of the feature embedding vectors of the two nodes is calculated, and negative values ​​are truncated to zero. Second, the Euclidean distance between the position encoding vectors of the two nodes in the music skill ontology is calculated, and then exponentially decayed after adjustment by the skill topology decay coefficient. The cosine similarity obtained in the first step is multiplied by the exponential decay value obtained in the second step to obtain the final association strength, which ranges from 0 to 1 and is used as the weight of the edge in the knowledge subgraph.

[0098] The prediction method for the correlation strength between explicit knowledge nodes and implicit skill nodes is as follows: At time t, for explicit knowledge node i and implicit skill node j, the system adaptively weights and fuses the temporal feature vector of the explicit knowledge node and the static skill feature vector of the implicit skill node through a gating vector. The fused features are input into a multilayer perceptron, then undergo a linear transformation of the weight vector, and finally the output value is mapped to the interval between 0 and 1 through the Sigmoid activation function to obtain the predicted correlation strength.

[0099] The fusion weight set is updated as follows: The fusion weight set at time t consists of two parts: first, the existing edges from the previous time t-1, whose actual association strength and the predicted association strength at the current time are weighted averaged using a time smoothing coefficient; second, newly emerging candidate edges, i.e., node pairs whose predicted association strength at the current time exceeds the candidate edge judgment threshold. The weights of all edges constitute the fusion weight set.

[0100] Calculation of quartiles and interquartile range: Sort the weight values ​​in the fusion weight set, and extract the first quartile and the third quartile according to the preset first quartile position parameters and third quartile position parameters respectively. The difference between the two is the interquartile range.

[0101] The decision rules for edge structure optimization are as follows: For each node pair, if its predicted association strength is greater than the third quartile plus the positive sensitivity coefficient multiplied by the interquartile range, then a new strong association edge is established or the weight of the existing edge is increased proportionally to the predicted strength; if the predicted association strength is less than the first quartile minus the negative sensitivity coefficient multiplied by the interquartile range, then the edge weight is weakened or redundant edges are removed. The positive and negative sensitivity coefficients are adaptively determined based on historical data using an optimization method. The updated edge weights are exponentially smoothed using a time decay factor.

[0102] The memory retention rate is calculated as follows: the memory retention rate of explicit knowledge node i at time t adopts an exponential decay form, and its decay exponent is obtained by multiplying multiple factors: the knowledge decay rate parameter multiplied by the ratio of the time difference to the baseline time constant, then multiplied by the individual difference adjustment term (plus the individual difference adjustment coefficient multiplied by the complement of the previous mastery level), and finally multiplied by the interval effect factor. The interval effect factor represents the exponential decay of recent review times and is used to realize the interval effect. Among them, a higher knowledge decay rate parameter is used for declarative knowledge nodes (such as music theory concepts), and a lower knowledge decay rate parameter is used for procedural knowledge nodes (such as fingering skills).

[0103] The advancement potential score for inactive latent skill nodes is calculated as follows: the advancement potential score of inactive node u is equal to the weighted average of the contributions of all currently active nodes i to u; the contribution value of each active node is: the association strength between active node i and inactive node u multiplied by a moderating factor. This moderating factor is calculated by adding the memory retention rate sensitivity coefficient multiplied by the memory retention rate of node i, and then multiplying by the exponential decay term of skill distance, where the skill distance decay coefficient is multiplied by the shortest path distance between the two nodes in the graph; the denominator of the weighted average is the sum of the moderating factors of all active nodes. Furthermore, a zone of proximal development constraint is introduced, meaning that if the difficulty level of node u exceeds the maximum acceptable difficulty range dynamically set based on the learner's current ability, its advancement potential score is zero, i.e., it is not recommended.

[0104] Other embodiments: The same method can be applied to the teaching of instruments such as violin and guitar. It is only necessary to replace the specific content of the skill ontology and the expert-defined behavioral indicators, and adjust the audio features (such as multi-pitch detection for polyphonic instruments) and video posture detection (for holding posture) in cross-modal feature extraction.

[0105] This invention can be applied to scenarios such as online music education platforms and intelligent musical instrument teaching systems. Through the above solution, significant improvements are made in the theoretical foundation of music education, the rigor of cognitive science applications, the feasibility of technical implementation, and the verifiability of effect evaluation, demonstrating high industrial application value.

[0106] Working Principle: The system simultaneously collects user performance audio, posture video, and touch interaction data using multiple sensor devices. Different modalities of data undergo specialized feature extraction processing, converting audio into acoustic features, video into posture features, and touch data into behavioral features. These are then aligned and semantically mapped according to a unified time dimension. Finally, the multi-source features are fused to form a comprehensive set of teaching feature vectors. Based on the extracted multimodal feature vectors, the system constructs a three-layer knowledge graph. The first layer extracts explicit music theory knowledge points, teaching pieces, and performance techniques as basic nodes. The second layer uses unsupervised learning methods to identify implicit skill features that cannot be directly expressed and abstracts them into skills. The third layer creates personalized cognitive state nodes for each user. Semantic connections between nodes at each layer are established through preset ontology relationships and similarity calculations, forming a teaching knowledge graph that integrates explicit knowledge and implicit skills. On the basic knowledge graph, each node is configured with evolvable attribute parameters, including dimensions such as activity decay, knowledge decay, and interaction influence. A unified timeline index system is established to record the historical changes of node parameters. An evolution function model is designed to dynamically update the parameter values ​​of each node based on students' real-time learning feedback, taking into account the node's own attributes and network topology. This is achieved through a non-linear fusion mechanism, enabling the knowledge graph to evolve autonomously over time and reflect the dynamic characteristics of the learning process.

[0107] The system utilizes a multi-layer perceptual inference network to adaptively fuse node information and multimodal features in a dynamically evolving knowledge graph. It predicts the potential association strength between latent skill nodes, dynamically calculates adaptive thresholds based on the current graph edge weight distribution, and optimizes the prediction results. Predictions exceeding the threshold establish or strengthen associated edges, while those below the threshold weaken or remove redundant edges. Temporal smoothing ensures the stability of the graph evolution, achieving accurate inference and dynamic updates of latent skill associations. Based on the updated skill node structure and dynamic knowledge graph, it constructs a cognitive state vector for the user, including knowledge mastery, review time, and skill activation status. Combining node-bound knowledge decay parameters and time differences, it applies a forgetting curve model to calculate the current memory retention level. Simultaneously, it extracts behavioral indicators from multimodal features and fuses them to generate a cognitive load metric, comprehensively forming a complete perception of the user's current cognitive state. Three recommendation modules run in parallel, and a forgetting compensation module identifies... The system retrieves relevant resources for reviewing knowledge points. The skills advancement module assesses advancement potential based on the correlation strength between currently activated and inactive skills. The cognitive load adjustment module dynamically adjusts recommendation intensity according to load levels. The outputs of these three modules are weighted and fused to generate initial recommendation candidates. Then, the weights of each module are adaptively adjusted based on historical user feedback data, ultimately outputting personalized recommendation results that match the user's current cognitive state and skill development needs. Based on the recommendation results and the user's cognitive state, and taking into account resource type, difficulty, and learning time, a scheduling scheme is generated that includes push sequence, learning interval, and priority weights. Teaching resources are pushed to the user according to the scheme. At the same time, user adoption, completion status, and multimodal feedback data are collected in real time. Feedback is used as a trigger signal to dynamically evolve the knowledge graph, updating the attribute parameters of relevant nodes according to preset rules, completing a closed-loop process from recommendation to feedback to graph update, and realizing continuous optimization and dynamic adaptation of the teaching system.

[0108] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0109] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A knowledge graph-based intelligent recommendation method for music teaching, characterized in that, Includes the following steps: S1. Based on the multimodal interaction data collected from the target users, perform semantic alignment and feature extraction to form a set of multimodal teaching feature vectors; S2. Based on the set of multimodal teaching feature vectors, construct a knowledge graph with a three-layer node structure to generate a multimodal fusion knowledge graph for music teaching; S3. Based on the multimodal fusion knowledge graph of music teaching, bind evolution attribute parameters to each node and establish a time axis index to construct a dynamic evolution knowledge graph; S4. Based on the dynamically evolving knowledge graph and the multimodal teaching feature vector set, perform association strength prediction and edge update through graph attention network to complete implicit skill association reasoning; S5. Based on the updated implicit skill node structure and dynamically evolving knowledge graph, load an improved forgetting curve model that distinguishes between declarative memory and procedural memory, and integrate physiological signals and subjective scale data to calculate the cognitive load comprehensive coefficient, and perform cognitive state perception and cognitive load measurement. S6. Based on the generated cognitive state vector and the updated implicit skill node structure, personalized recommendation results are generated through the weighted fusion of three parallel modules: forgetting compensation, skill advancement, and cognitive load adjustment. S7. Based on the generated personalized recommendation results, generate a push scheduling plan and push resources to the target users. At the same time, use the real-time collected feedback data as a graph update trigger signal to input the dynamic evolution knowledge graph and execute dynamic evolution push and closed-loop update.

2. The intelligent music teaching recommendation method based on knowledge graphs according to claim 1, characterized in that: In step S1, the target user's performance audio data, performance posture video data, touch interaction log data, and physiological signal data are simultaneously collected through a microphone array, camera equipment, touch screen interactive terminal, and wearable physiological sensor, respectively. The audio data is subjected to short-time Fourier transform to extract Mel frequency cepstral coefficients, pitch deviation sequence, and rhythm stability features. The video data is subjected to posture key point detection to extract hand shape standardization and sitting posture standardization features. The touch interaction data is subjected to time sequence pattern recognition method to extract practice duration distribution, fingering error frequency, and accuracy sequence. The three types of features are timestamped and semantically mapped according to the preset teaching label fields. Multimodal features within the same time window are spliced ​​and fused into a unified set of multimodal teaching feature vectors through a cross-modal attention mechanism.

3. The intelligent music teaching recommendation method based on knowledge graphs according to claim 1, characterized in that: In step S2, based on the generated multimodal teaching feature vector set, explicit music theory knowledge point labels, teaching repertoire labels, and performance technique labels are extracted from the feature vectors. These labels are used as the first layer of explicit knowledge nodes, and explicit semantic association edges are established according to the preset music teaching ontology relationship. Unsupervised clustering analysis is performed on the skill features that cannot be explicitly expressed in the multimodal teaching feature vectors. The cluster centers are abstracted as the second layer of implicit skill nodes, and an implicit skill knowledge subgraph is constructed by calculating the weighted similarity between the embedding vectors of each implicit skill node based on the music skill development sequence. A third-layer cognitive state node is created for each target user, including a forgetting rate parameter node and a cognitive load measurement node. Each cognitive state node is associated with a corresponding explicit knowledge node through a timestamp, forming a multimodal fusion knowledge graph for music teaching with interconnected three-layer node structures.

4. The intelligent music teaching recommendation method based on knowledge graphs according to claim 1, characterized in that: In S3, based on the multimodal fusion knowledge graph of music teaching, evolution attribute parameters are bound to each explicit knowledge node, implicit skill node and cognitive state node in the graph. The evolution attribute parameters include node activity decay factor, knowledge decay rate parameter and interaction influence coefficient. A global timeline index is constructed using fixed time steps, and each node records the sequence of evolution attribute parameter values ​​at different time steps. An evolution function model is pre-built. Based on the student interaction feedback data at the current time step, the evolution function model dynamically updates the evolution attribute parameters of each node in the time axis index through a non-linear weighted fusion mechanism, generating a dynamic evolution knowledge graph with time dimension evolution capability.

5. The intelligent recommendation method for music teaching based on knowledge graphs according to claim 1, characterized in that: In S4, based on the dynamic evolution knowledge graph and the multimodal teaching feature vector set, a subset of features related to implicit skills is extracted from the multimodal teaching feature vector, and adaptively fused with the current embedding vector of the implicit skill node in the dynamic evolution knowledge graph through a cross-modal attention fusion mechanism to form the input features of the graph attention network. Initialize a graph attention network containing a multi-head attention layer, a graph convolutional layer, and an edge prediction layer, where the edge prediction layer uses a bilinear form to compute the predicted association strength between latent skill nodes; The fused features are propagated forward layer by layer, and the neighbor node information is aggregated by attention coefficient weighting to obtain the updated representation of each latent skill node. Then, a bilinear decoder is used to generate a prediction matrix of the association strength between node pairs.

6. The intelligent music teaching recommendation method based on knowledge graphs according to claim 5, characterized in that: In step S4, the upper and lower quartiles are calculated as adaptive thresholds based on the weight distribution of the edges of the implicit skill nodes in the current knowledge graph, as follows: ， ， ， In the formula, This represents the set of fusion weights at time t. This represents the time smoothing coefficient, with a value range of [0, 1]. This represents the actual affinity strength of edge (k, l) at the previous time t-1. This represents the predicted association strength of edge (k, l) at the current time t. Let the set of edges at time t be represented. This represents the threshold for determining candidate edges. This represents the first quartile at time t. This represents the third quartile at time t. This represents the position parameter of the first quantile. This represents the position parameter of the third quantile. This represents the interquartile range at time t; When the predicted value exceeds the upper quartile boundary When the predicted value is below the lower quartile boundary, establish a new strong correlation edge; At that time, the edge weights are weakened, where and The asymmetric threshold sensitivity coefficient is determined adaptively based on historical data using an optimization method. The strictness of strong association enhancement and weak association decay is controlled separately. The updated edge weights are exponentially smoothed using a time decay factor. The optimized graph structure is written back to the dynamically evolving knowledge graph to complete the dynamic reasoning and evolution of implicit skill associations.

7. The intelligent music teaching recommendation method based on knowledge graphs according to claim 1, characterized in that: In step S5, based on the updated implicit skill node structure and the generated dynamic evolution knowledge graph, a cognitive state vector is initialized for each target user; for each explicit knowledge node, according to the preset knowledge decay rate parameter bound to the node in the dynamic evolution knowledge graph, combined with the difference between the most recent review timestamp and the current time, a preset improved forgetting curve model is loaded to calculate the current memory retention rate. The mastery score is updated based on memory retention rate; the practice duration distribution and fingering error frequency of touch interaction data, as well as the rhythm stability deviation, pitch deviation sequence, and heart rate variability and skin conductance response characteristics of audio data are extracted from the formed multimodal teaching feature vector set. Combined with the questionnaire results of the subjective cognitive load scale, a comprehensive cognitive load coefficient is generated through weighted fusion of multi-task learning network; the updated mastery score sequence, activation status identifier, and comprehensive cognitive load coefficient are integrated into the cognitive state perception result at the current time step.

8. The intelligent recommendation method for music teaching based on knowledge graphs according to claim 1, characterized in that: In S6, three recommendation modules run in parallel based on the cognitive state vector and the updated implicit skill node structure: The forgetting compensation module, based on the memory retention rate of each explicit knowledge node in the cognitive state vector and combined with the staggered practice effect, marks nodes that are below a preset threshold and whose time since the last practice exceeds the optimal interval as nodes that need to be reviewed, and retrieves the corresponding teaching track nodes and practice resource nodes from the dynamically evolving knowledge graph. The skill advancement module selects the candidate hidden skill nodes with the greatest potential for advancement by calculating the association strength between them and the set of currently active hidden skill nodes in the hidden skill node structure, and expands their associated explicit knowledge nodes into a skill advancement recommendation list. The cognitive load adjustment module dynamically adjusts the upper limit of the number of recommended resources and the difficulty span coefficient based on the comprehensive cognitive load coefficient in the cognitive state vector; The recommendation lists output by the three modules are weighted and merged according to preset initial weights to generate an initial recommendation candidate set. Then, based on the actual adoption rate and completion rate of various recommendation resources, as well as the learning gain and skill transfer test results recorded in the target user's historical interaction feedback data, the weight coefficients of the three modules are adaptively fine-tuned through Bayesian personalized ranking, and finally, personalized recommendation results are output.

9. The intelligent music teaching recommendation method based on knowledge graphs according to claim 1, characterized in that: In S7, based on the personalized recommendation results, a push scheduling scheme is generated according to the resource type, difficulty level and estimated learning time of each recommended resource, combined with the cognitive load comprehensive coefficient in the cognitive state vector. The push scheduling scheme includes a resource push sequence arranged in chronological order, a suggested learning interval between resources, and a push priority weight for each resource. Recommended resources are pushed to target users according to the push scheduling scheme, and feedback data is collected in real time during the push execution process. The collected feedback data is used as a graph update trigger signal, input into the dynamic evolution knowledge graph, and the evolution attribute parameters of the affected nodes are updated in real time according to the update rules defined in the evolution function model. The updated node status is written to the corresponding position of the time axis index, thus completing the dynamic evolution push and closed-loop update.

10. The knowledge graph-based intelligent recommendation system for music teaching implemented according to the knowledge graph-based intelligent recommendation method of claim 1 includes a multimodal feature extraction module, a knowledge graph construction module based on evidence center design and music skill ontology, an incremental dynamic evolution graph construction module, a graph attention network implicit skill reasoning module, a multimodal cognitive state perception module, a three-way fusion recommendation generation module, and a dynamic push closed-loop update module.

Images