Online education resource recommendation method and system based on knowledge graph

Abstract

The invention discloses an online education resource recommendation method and system based on a knowledge graph, and relates to the field of online education resource recommendation. The method comprises the following steps: S1, mapping an interaction record of a learner to a knowledge graph node space to generate an interaction matrix; s2, on the basis of the interaction matrix, calculating the mastery degree of each node in combination with interaction strength and knowledge graph path dependence; s3, correcting the mastery degree in combination with the forgetting curve to generate a corrected mastery degree; s4, calculating the energy mastering degree of the target node by adopting a path energy attenuation model; s5, calculating enhanced correlation based on the features of the resource and the target node and the space-time state; s6, filtering resources having conflicts with mastered knowledge to obtain security correlation; and S7, calculating a recommendation priority score in combination with the energy mastery degree and the safety correlation. Through personalized recommendation and cognitive conflict filtering, the accuracy and correlation of educational resource recommendation are improved, so that the learning efficiency is optimized, and the cognitive burden of learners is relieved.

Description

Technical Field

[0001] This invention relates to the field of online educational resource recommendation, specifically to a knowledge graph-based method and system for recommending online educational resources. Background Technology

[0002] Traditional educational resource recommendation methods mostly rely on content-based recommendation or collaborative filtering algorithms. However, these methods often fail to effectively consider learners' knowledge structures, cognitive states, and personalized learning progress, resulting in a lack of targeted and accurate resource recommendations. To address these issues, knowledge graph-based recommendation methods, by constructing relationships between knowledge nodes and learners, can fully leverage learners' learning history, knowledge point mastery, and path relationships within the knowledge graph to provide more accurate and personalized recommendation services.

[0003] However, existing knowledge graph-based recommendation methods still face challenges such as how to dynamically adjust learners' knowledge mastery, how to calculate the energy transfer effect between knowledge points, and how to filter resources that conflict with learners' existing knowledge. Therefore, designing a more intelligent, accurate, and efficient educational resource recommendation method based on learners' cognitive state, the relationships between knowledge points, and the spatiotemporal characteristics of educational resources has become a key research issue.

[0004] This invention provides a knowledge graph-based method for recommending online educational resources. It aims to comprehensively consider factors such as learners' knowledge mastery, cognitive conflict, resource relevance, and fatigue levels, thereby overcoming the shortcomings of traditional recommendation methods and improving the accuracy and effectiveness of the recommendation system. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide an online education resource recommendation method and system based on knowledge graphs to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online educational resource recommendation method based on knowledge graphs, comprising:

[0007] S1: Map learners’ interaction records within a preset time period to the knowledge graph node space to generate a multi-dimensional knowledge node interaction matrix;

[0008] S2: Based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph, calculate the mastery of each node;

[0009] S3: Based on the mastery level and combined with the forgetting curve dynamic correction mechanism, the mastery level is corrected to generate a corrected mastery level;

[0010] S4: Based on the corrected mastery level, the path energy decay model is used to calculate the energy propagation value of the learner from the mastered node to the target node, and the energy mastery level of the target node is obtained.

[0011] S5: Calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes;

[0012] S6: Based on enhanced relevance, filter educational resources that have cognitive conflicts with the learners' existing knowledge to obtain safety relevance;

[0013] S7: Calculate the recommendation priority score by combining energy control and safety relevance.

[0014] The present invention is further configured such that S1 includes:

[0015] Map learners’ original interaction records within a preset time period to the node space in the knowledge graph;

[0016] Based on learners' interactive behavior, an interaction feature matrix with time slice granularity is constructed, which represents the interaction strength between learners and each knowledge node;

[0017] The interaction feature matrix is ​​processed, and an embedding layer is used to map the interaction data to a high-dimensional semantic space. Then, a semantic vector representation of each knowledge node is generated through linear transformation.

[0018] The Hadamard product operation is used to fuse learners’ interaction data with the semantic vectors of knowledge nodes at the element level.

[0019] Based on the Frobenius norm of knowledge graph nodes, the interaction intensity is regularized to generate a knowledge node interaction matrix.

[0020] The present invention is further configured such that S2 includes:

[0021] Based on the knowledge node interaction matrix, combined with the path dependency weights derived from the knowledge graph and the current interaction intensity, a nonlinear normalization transformation is performed, and a time decay function is applied to adjust the normalization result. At the same time, a Gaussian process perturbation term is injected to simulate cognitive random fluctuations.

[0022] By applying a quadratic exponential decay mechanism, the knowledge propagation effect of long paths is suppressed, and the mastery of output nodes is improved.

[0023] The present invention is further configured such that S3 includes:

[0024] Based on the difference between the most recent interaction time and the current time of the nodes, and combined with path length modulation, the forgetting degree of each node is calculated. The path length modulation adjusts the attenuation effect of information propagation according to the path length between nodes.

[0025] Calculate the corrected mastery of nodes based on their forgetting and mastery levels.

[0026] The present invention is further configured such that S4 includes:

[0027] Based on the node's corrected mastery and combined with the shortest path distance between nodes, an exponential decay model is used to adjust the intensity of the path propagation effect.

[0028] Based on the semantic vector distance between nodes, a hyperbolic modulation mechanism is applied to adjust the path propagation effect;

[0029] The attenuation model adjustment results and hyperbolic modulation results are combined, and the weights of the path levels are adjusted using a random walk model.

[0030] By using a dynamic integration mechanism, the energy propagation effect of multi-level paths is aggregated to calculate the energy mastery of the target node.

[0031] The present invention is further configured such that S5 includes:

[0032] Calculate the absolute deviation between the resource update time difference and the target node time difference, and construct a time window suppression function based on this deviation.

[0033] Based on the frequency domain characteristics of resources and target nodes, the spectral correlation coefficient between resources and target nodes is calculated;

[0034] Analyze the relative positions and adjacency relationships between resources and target nodes in the knowledge graph, and evaluate the strength of the topological association between resources and target nodes;

[0035] By combining the spectral correlation coefficient and topological association strength under the constraint of the time window suppression function, the enhanced correlation between resources and target nodes is obtained.

[0036] The present invention is further configured such that S6 includes:

[0037] Identify the set of nodes that have cognitive conflict relationships with the target node, and calculate the cognitive topological distance from the target node to the conflicting nodes;

[0038] Extract the feature vectors of the target node and the conflict node, and calculate the adversarial loss value through element-wise multiplication and vector inner product operations;

[0039] Based on the enhanced correlation between resources and target nodes, cognitive topological distance, and adversarial loss value, the security correlation after conflict filtering is calculated.

[0040] The present invention is further configured such that S7 includes:

[0041] Tensor product operations are performed based on energy control and security relevance, and the results of the tensor product operations are transformed by a logarithmic smoothing function.

[0042] The resource fatigue factor is calculated based on the historical exposure frequency of the resource, and the historical exposure time decay value is obtained and multiplied with the resource fatigue factor.

[0043] The recommended priority score for a resource is calculated by combining the results of the logarithmic smoothing function transformation and the product operation.

[0044] The present invention is further configured to sort resources from high to low priority based on their recommended priority scores;

[0045] The top N resources in the recommendation list are selected from the sorting results and recommended as educational resources to learners.

[0046] This invention also provides an online educational resource recommendation system based on knowledge graphs, the system comprising:

[0047] Interaction Record Mapping Module: Used to map learners' interaction records within a preset time period to the knowledge graph node space, generating a multi-dimensional knowledge node interaction matrix;

[0048] Mastery calculation module: used to calculate the mastery of each node based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph;

[0049] Mastery Correction Module: This module is used to correct the mastery based on the mastery level and the forgetting curve dynamic correction mechanism, generating a corrected mastery level.

[0050] Energy Propagation Calculation Module: Based on the corrected mastery level, using the path energy decay model, calculates the energy propagation value from the learner's mastered nodes to the target node, and obtains the energy mastery level of the target node;

[0051] Correlation calculation module: used to calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes;

[0052] Cognitive conflict filtering module: Used to filter educational resources that have cognitive conflicts with the learner's existing knowledge based on enhanced relevance, thus obtaining safe relevance;

[0053] Priority Calculation Module: Used to calculate the recommended priority score by combining energy availability and security relevance.

[0054] This invention provides a method and system for recommending online educational resources based on knowledge graphs. The method comprises the following steps: S1: Mapping learners' interaction records within a preset time period to the knowledge graph node space to generate a multi-dimensional knowledge node interaction matrix; S2: Calculating the mastery level of each node based on the knowledge node interaction matrix, combined with the interaction intensity over time segments and the path dependency relationship of the knowledge graph; S3: Correcting the mastery level based on the mastery level and using a dynamic correction mechanism based on the forgetting curve to generate a corrected mastery level; S4: Calculating the energy propagation value from the learner's mastered nodes to the target node based on the corrected mastery level using a path energy decay model to obtain the energy mastery level of the target node; S5: Calculating the enhanced correlation between resources and target nodes based on their characteristics and spatiotemporal states; S6: Filtering educational resources that have cognitive conflicts with the learner's already mastered knowledge based on the enhanced correlation to obtain a safety correlation; S7: Calculating a recommendation priority score by combining the energy mastery level and the safety correlation. The resulting beneficial effects include:

[0055] 1. Personalized recommendations improve learning efficiency: The knowledge graph-based online education resource recommendation method provides personalized education resource recommendations by deeply analyzing learners' interaction history, mastery level, and learning status. By calculating learners' mastery of each knowledge node and combining it with a dynamic correction mechanism, it can recommend the most suitable resources for learners at different learning stages, thereby effectively improving learning efficiency.

[0056] 2. Avoiding Ineffective Recommendations Due to Cognitive Conflict: By using a cognitive conflict resource filtering mechanism, this invention avoids recommending resources that conflict with the learner's existing knowledge. By identifying cognitive gaps or conflict points in the learner's current knowledge system, this invention can more effectively filter out unsuitable or ineffective educational resources, thereby ensuring the high quality and high relevance of recommended resources and reducing the learner's cognitive burden.

[0057] 3. Over-recommendation controlled by fatigue factor: By introducing a resource fatigue factor, the over-recommendation of educational resources is effectively avoided, preventing learners from experiencing learning fatigue due to frequently receiving the same type of recommendations. This mechanism dynamically adjusts the priority of recommended resources by combining historical exposure frequency and fatigue factor, ensuring the diversity and efficiency of resource recommendations.

[0058] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0060] Figure 1 A flowchart illustrating an exemplary embodiment of the present invention is shown below, illustrating a method for recommending online educational resources based on a knowledge graph.

[0061] Figure 2 This is a schematic diagram illustrating the structure of an online educational resource recommendation system based on a knowledge graph, as an exemplary embodiment of the present invention. Detailed Implementation

[0062] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0063] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0064] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0065] Example 1

[0066] A knowledge graph-based method for recommending online educational resources, such as... Figure 1 As shown, it includes:

[0067] S1: Map learners’ interaction records within a preset time period to the knowledge graph node space to generate a multi-dimensional knowledge node interaction matrix;

[0068] S2: Based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph, calculate the mastery of each node;

[0069] S3: Based on the mastery level and combined with the forgetting curve dynamic correction mechanism, the mastery level is corrected to generate a corrected mastery level;

[0070] S4: Based on the corrected mastery level, the path energy decay model is used to calculate the energy propagation value of the learner from the mastered node to the target node, and the energy mastery level of the target node is obtained.

[0071] S5: Calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes;

[0072] S6: Based on enhanced relevance, filter educational resources that have cognitive conflicts with the learners' existing knowledge to obtain safety relevance;

[0073] S7: Calculate the recommendation priority score by combining energy control and safety relevance.

[0074] The present invention is further configured such that S1 includes:

[0075] Map learners’ original interaction records within a preset time period to the node space in the knowledge graph;

[0076] Based on learners' interactive behavior, an interaction feature matrix with time slice granularity is constructed, which represents the interaction strength between learners and each knowledge node;

[0077] The interaction feature matrix is ​​processed, and an embedding layer is used to map the interaction data to a high-dimensional semantic space. Then, a semantic vector representation of each knowledge node is generated through linear transformation.

[0078] The Hadamard product operation is used to fuse learners’ interaction data with the semantic vectors of knowledge nodes at the element level.

[0079] Based on the Frobenius norm of knowledge graph nodes, the interaction intensity is regularized to generate a knowledge node interaction matrix. Specifically, a set of original interaction records within a preset time window is obtained. This set of original interaction records contains time-series data such as learners' answer records, resource browsing time, and number of interactions. Each record reflects the interaction between learners and educational resources. By defining a time slice granularity parameter, such as 15 minutes, the continuous time axis is divided into discrete time segments according to this time slice granularity parameter to segment the learners' interaction behavior time sequence. A mapping relationship between interaction events and knowledge graph nodes is established, mapping the original interaction records to the corresponding node space of the knowledge graph, and mapping the original learning events to specific knowledge nodes. For example, a "video viewing event" is mapped to the knowledge node "Newton's Laws," while a "question answering event" is mapped to the knowledge node "probability distribution." This allows for more accurate learning resource recommendations based on the learner's interaction history with the knowledge nodes. Learner interaction behaviors are feature-encoded, including interaction intensity coding and interaction type coding. Interaction intensity coding represents the intensity of interaction between the learner and a specific knowledge point within a given time segment, considering both the duration of the time segment and the duration of the interaction. Interaction type coding categorizes learner participation in different interactive activities using predefined numerical codes; for example, in-depth reading, video viewing, and question testing each correspond to different coding values. A weight matrix W is used. Θ Linear transformation of interaction intensity, and through the embedding layer W Λ Encode the semantics of the interaction type, and the interaction strength Θ ij (t) through the weight matrix W Θ Perform a linear transformation to map to a higher-dimensional space, while the interaction type Λ ij (t) is then passed through the embedding layer W Λ The process of converting low-dimensional interaction data into a higher-dimensional semantic vector helps transform low-dimensional interaction data into a richer representation. The interaction strength and type are fused through the Hadamard product operation, and after a non-linear transformation using the tanh activation function, the fused feature vector is obtained. The calculation logic for the feature vector is as follows: h ij For feature vectors, For Hadamard product operations; to ensure the stability of the knowledge node interaction matrix and avoid overfitting, node structure regularization is introduced. This regularization is achieved by calculating the topological influence of node i. The calculation logic of the knowledge node interaction matrix is: M ij (t)=h ij +∈·||G i || F M ij (t) represents the knowledge node interaction matrix; ||Gi || F Let be the Frobenius norm of node i in the knowledge graph, used to measure the topological influence of the node; G represents the knowledge graph. i Let M be the matrix of adjacent nodes associated with node i; ∈ is the regularization coefficient used to control the complexity of the knowledge node interaction matrix, with a value range of [0.01, 0.1]; through the above steps, the knowledge node interaction matrix M is obtained. ij (t) Maps learners’ interaction information at various time segments to the node space of the knowledge graph, serving as the basis for subsequent learning state inference and recommendation.

[0080] The present invention is further configured such that S2 includes:

[0081] Based on the knowledge node interaction matrix, combined with the path dependency weights derived from the knowledge graph and the current interaction intensity, a nonlinear normalization transformation is performed, and a time decay function is applied to adjust the normalization result. At the same time, a Gaussian process perturbation term is injected to simulate cognitive random fluctuations.

[0082] A quadratic exponential decay mechanism is applied to suppress the knowledge propagation effect along long paths, thus increasing the mastery level of the output nodes. Specifically, this embodiment receives a knowledge node interaction matrix M constructed based on a knowledge graph. ij (t), the dimension of this matrix is ​​V×T, where V represents the number of knowledge nodes in the knowledge graph and T represents the total number of time slices. This interaction matrix contains information on the interaction intensity between learners and each knowledge node in different time slices; initialize the path hierarchy parameter τ∈{1,2,…,τ max}, where τ max It is the preset maximum number of path levels, for example, setting τ. max =5; Based on the adjacency matrix A of the knowledge graph, calculate the path dependency weight at each path level τ. The calculation logic for the path dependency weight is as follows: Γ iτ For path-dependent weights, Let M be the subset of the adjacency matrix of node i within a path length τ. For example, when τ = 2, the second-order adjacency relationship of node i is calculated, representing the influence of neighbors through two path lengths on node i; the knowledge node interaction matrix M is... ij (t) is combined with path-dependent weights and subjected to nonlinear normalization transformation. Feature fusion is then performed through tensor convolution. The computational logic for feature fusion is as follows: u iτ =||Γ iτ || ∞ ·M ij (t)⊙Γ iτ u iτ The feature fusion result value represents the nonlinear normalized interaction data of node i at path level τ; ||Γiτ || ∞ Given the infinite norm of the path-dependent weights, the path-dependent weight matrix Γ was calculated. iτ The maximum value within the path level τ represents the weight that has the greatest impact on the learning node during path propagation; ⊙ represents a feature fusion operation based on tensor convolution with a kernel size of 3×3. This operation effectively combines path-dependent weights and interaction strengths, capturing multi-dimensional learning features; a logarithmic time decay function is applied to modulate the temporal dependence of the interaction data. Considering that the influence of earlier interactions on the current learning state gradually decreases over time, the calculation logic of the time decay function is as follows: ζ(t) is the time decay function, used to dynamically adjust the weight of the learner's historical interaction data in the current learning state according to the changes in time; to simulate the random fluctuations in the cognitive process, a Gaussian perturbation term is added, with the following distribution: Ξ iτ (t) represents the Gaussian perturbation term; σ t The standard deviation is calculated using the following logic: σ t =0.1·τ -0.5 As the path hierarchy increases, the magnitude of the perturbation term gradually decreases. To further process the interaction intensity data, the Sigmoid function is used to compress the value range, making the output more suitable for subsequent processing. The compression logic is as follows: v iτ =sigmoid(u iτ +ζ(t)·Ξ iτ (t)), v iτ Let be the final value of node i at path level τ after compression, representing the processed interaction data. Compression using the Sigmoid function ensures the output is within a relatively stable and constrained range, avoiding values ​​that are too large or too small. To suppress the knowledge propagation effect along long paths, a path decay mechanism is employed to calculate the mastery of nodes. The logic for calculating mastery is as follows: K(i,m,τ) represents the mastery level, indicating the mastery level of node i in time segment m and path level τ; α is an adjustment factor used to control the scale of the output results, with a value range of [0.1,10]; β is a decay parameter used to control the influence of path length on the propagation effect, with a value range of [0.5,1.5]. By combining path dependency weights and interaction strength, the dependency relationships between nodes in the knowledge graph can be better reflected, making the calculation of learners' knowledge mastery level more accurate.

[0083] The present invention is further configured such that S3 includes:

[0084] Based on the difference between the most recent interaction time and the current time of the nodes, and combined with path length modulation, the forgetting degree of each node is calculated. The path length modulation adjusts the attenuation effect of information propagation according to the path length between nodes.

[0085] Based on the degree of forgetting and the degree of mastery of each node, a revised mastery score is calculated. Specifically, in this embodiment, to accurately depict the learner's forgetting process and to incorporate individual differences and structural features into the mastery score calculation, the mastery score of knowledge nodes is dynamically corrected based on an improved forgetting curve. The calculated three-dimensional mastery score is then received. For each node i, extract its τ-hop ​​neighbor subgraph N. i (τ), and calculate the structural information entropy of the neighboring subgraph to obtain the forgetting coefficient. The calculation logic of the forgetting coefficient is as follows: γ i Let be the forgetting coefficient of node i; λ is an adjustment factor used to control the overall scaling of the forgetting coefficient, with a value range of [0,1]; N i Let k be the τ-hop ​​neighbor subgraph of node i, containing the set of all nodes reachable within τ steps, and the connections between them; k is N. i (τ) is a neighboring node in the subgraph; p(k) is the normalized degree of node k, calculated as follows: deg(k) is the degree of node k, which is the number of other nodes connected to it. The larger the degree, the more centralized node k is in the knowledge graph; deg(m) is the degree of node m; m is N. i (τ) is a neighbor node in the subgraph; The information entropy formula reflects the structural complexity of a subgraph. A high entropy value indicates a uniform distribution of node degrees, meaning most nodes have similar degrees, resulting in a complex network with a dispersed center. Central nodes with higher information entropy have a larger forgetting coefficient and faster forgetting speed. A power-law function is used to achieve adaptive modulation of the path length, constructing a path modulator. The construction logic of the modulator is: D(τ)=1+δτ ν D(τ) is the attenuation factor, representing the value of the path modulator, used to dynamically adjust the transmission strength of information in the knowledge graph according to the path length τ. When D(τ) is closer to 1, it indicates that the path is shorter, the conductivity is stronger, and the degree of information attenuation is lower. δ is the path adjustment coefficient, used to control the sensitivity of the modulator to changes in path length, with a value range of [0, 0.5]. ν is the power law exponent, used to determine the nonlinear growth rate of the path length, with a value range of [1, 1.5]. The last interaction timestamp of node i is obtained, and the interval between the current time and the last time that node was learned is calculated, i.e., the second time difference. The calculation logic of the second time difference is: Δt 2 =(tt) ilast ) 2Δt represents the quadratic time difference. Squaring this quadratic time difference amplifies larger time intervals, causing the attenuation effect to nonlinearly increase as the time difference grows; t represents the current time. ilast Let K be the timestamp of the last interaction at node i. The aforementioned time difference enhancement mechanism can amplify the decay effect when the time interval is large, making the forgetting curve more effective in correcting knowledge that has not been reviewed for a long time. The forgetting coefficient, path modulator, and quadratic time difference squared are combined and applied to the mastery, and an exponential decay method is used for correction to obtain the corrected mastery. The calculation logic of the corrected mastery is: K ‘ (i,m,τ)=K(i,m,τ)·exp(-γ i ·D(τ)·Δt 2 ), K ‘ (i,m,τ) represents the corrected mastery level; γ i ·D(τ)·Δt 2 To comprehensively consider the forgetting factor, which represents the degree of forgetting of a node, a larger value indicates a more significant forgetting effect, requiring a stronger dynamic decay correction for the mastery of that node. This embodiment, based on the traditional forgetting curve model, integrates node structural features and path length factors, achieving differentiated dynamic correction for nodes at different positions and propagation levels, improving the accuracy of learner mastery calculation, and more closely reflecting the real cognitive process.

[0086] The present invention is further configured such that S4 includes:

[0087] Based on the node's corrected mastery and combined with the shortest path distance between nodes, an exponential decay model is used to adjust the intensity of the path propagation effect.

[0088] Based on the semantic vector distance between nodes, a hyperbolic modulation mechanism is applied to adjust the path propagation effect;

[0089] The attenuation model adjustment results and hyperbolic modulation results are combined, and the weights of the path levels are adjusted using a random walk model.

[0090] Through a dynamic integration mechanism, the energy propagation effect of multi-level paths is aggregated to calculate the energy mastery of the target node. Specifically, the spatial distance of knowledge propagation between different nodes is measured by the shortest path distance between node pairs in the knowledge graph structure. The Diikstra algorithm is used to calculate the shortest path distance between node pairs (i,k), where a greater distance indicates a longer knowledge propagation path. The calculation logic for the shortest path distance between nodes is as follows: d P (i,k)=Dijkstra(i,k;G,d P(i,k) represents the shortest path distance between node i and node k; G is the adjacency graph of the knowledge graph; the semantic similarity between nodes is measured by calculating the Manhattan distance between the embedding vectors of two nodes. The calculation logic for the semantic distance between node pairs is as follows: J ik v represents the semantic distance between node i and node k, used to measure the degree of semantic difference between the two nodes; i,l and v k,l Let be the embedding vector components of node i and node k in the l-th dimension, respectively; M is the number of embedding vector dimensions; l is the dimension index; J represents the Manhattan distance between two vectors, obtained by summing the absolute differences across all dimensions. ik The distance between two knowledge nodes in the semantic space is measured; the smaller the distance, the more semantically similar the nodes are. J... ik Introducing the modulation semantics influence weight of the hyperbolic modulation function, the calculation logic of the hyperbolic modulation function is as follows: H(J ik The function is a hyperbolic modulation function used to more accurately characterize the impact of semantic similarity between nodes on propagation weights. Based on the modified mastery, the shortest path distance between nodes, and the semantic distance between node pairs, the modified mastery is weighted and integrally calculated along the path hierarchy dimension to obtain the energy mastery. The calculation logic for the energy mastery is as follows: E k (t) represents the energy mastery of the target node; This is an exponential decay model used to control the strength of the path propagation effect, which decreases exponentially as the shortest path distance increases; α1 is the path decay coefficient, used to control the decay rate of the path length during knowledge propagation, with a value range of [0.1, 1]; Φ τ Φ is the diffusion adjustment coefficient at path level τ, designed based on the diffusion characteristics of the random walk process. Specifically, the random walk model describes the dynamic process of information or energy propagating along paths in a graph structure, and the propagation intensity weakens as the path level increases. τ The calculation logic is as follows: ρ is a constant used to control the decay rate of path propagation, with a value range of [0.1,1]. By combining the spatial distance between nodes, semantic similarity, and hierarchical weights, a comprehensive evaluation of the knowledge propagation energy is achieved, improving the accuracy of the knowledge propagation process and the personalization level of the recommendation results.

[0091] The present invention is further configured such that S5 includes:

[0092] Calculate the absolute deviation between the resource update time difference and the target node time difference, and construct a time window suppression function based on this deviation.

[0093] Based on the frequency domain characteristics of resources and target nodes, the spectral correlation coefficient between resources and target nodes is calculated;

[0094] Analyze the relative positions and adjacency relationships between resources and target nodes in the knowledge graph, and evaluate the strength of the topological association between resources and target nodes;

[0095] By fusing the spectral correlation coefficient and topological association strength under the constraint of the time window suppression function, the enhanced correlation between resources and target nodes is obtained. Specifically, the absolute deviation between the resource update time difference and the target node time difference is calculated. The calculation logic for the absolute deviation is: ΔT diff =|ΔT r (t)·-ΔT k (t)|,ΔT diff The absolute deviation is used to measure the time difference between resources and target nodes; ΔT r (t) represents the time difference between the recommendation time of resource r and the most recent relevant event of resource r, that is, the time difference between the most recent interaction or update time of the resource and the current time; ΔT k (t) represents the time difference between the recommended time of target node k and the most recent relevant event of target node k, i.e., the time difference between the target node and the current time; a time window suppression function is constructed to suppress the absolute value deviation ΔT between the resource and the target node. diff To adjust the correlation strength; when the absolute value deviation is large, the correlation between resources and target nodes is adjusted to avoid excessively high correlations from resources that have not interacted for a long time. The construction logic of the time window suppression function is as follows: W t λ1 is a time window suppression function used to reduce the correlation between resources and target nodes when the absolute value deviation is large; λ1 is an adjustment coefficient that controls the influence of the absolute value deviation on the time window suppression function, with a value range of [0,1]. To further extract multimodal features between resources and target nodes, spectral features of various resources are constructed, including text resources, video resources, and exercise resources. The TF-IDF method is used to construct the spectrum of text resources. The spectrum of video resources is constructed by combining the audio MFCC Mel-frequency cepstral coefficients and the visual HOG directional gradient histogram features. The spectrum of exercise resources is constructed through the knowledge point distribution histogram. Based on the obtained spectral features, the spectral correlation coefficient between the spectra is calculated to reflect the frequency domain correlation between resources and target nodes. The calculation logic of the spectral correlation coefficient is as follows: corr ω [S r ,S k [S] represents the spectral correlation coefficient, used to measure the similarity between resource r and target node k in the frequency domain. The larger the spectral correlation coefficient value, the more similar the spectral characteristics of resource r and target node k; r(ω) and S k (ω) represents the feature spectrum of resource r and target node k in the frequency domain, i.e., the frequency domain feature; ω is the frequency variable in the frequency domain; and Let S be the square integral of the characteristic spectra of resource r and target node k, respectively, representing the total energy of the resource and target node in the frequency domain; ∫S r (ω)S k (ω)dω is the integral of the product of the frequency domain feature spectra of resource r and target node k, representing the common similarity between the resource and the target node in the frequency domain. By analyzing the topological associations between resources and knowledge points, the reasoning process is further strengthened. Specifically, a resource-knowledge point subgraph matching is constructed to measure the structural similarity between resources and knowledge points. An adjacency matrix is ​​used to represent the subgraph of resources and knowledge points, and matrix operations are used to calculate their topological correlation. The construction logic of the subgraph matching is as follows: G(R r ,R k The value represents the topological association strength between the resource and the target node. A larger value indicates a stronger relationship and correlation between the resource and the target node in the graph structure, thus enabling more accurate recommendations of resources suitable for the learner's current needs. r Let A be the adjacency matrix of the resource r association subgraph, representing the connection relationships between resources and reflecting the topological structure of resources in the knowledge graph; k Let be the adjacency matrix of the subgraph associated with the target node k, representing the connection relationships between target nodes and describing the topological structure of different knowledge points in the knowledge graph; For matrix multiplication Perform trace operations to represent the correlation between resources and knowledge points; and By calculating the self-association degree of the resource subgraph and the knowledge point subgraph, the total relevance of resources and knowledge points in their respective graph structures is measured. Based on the obtained time window suppression function, spectral correlation coefficient, and topological association strength, the enhanced relevance between resources and target nodes is calculated. The calculation logic for enhanced relevance is as follows: C rk (t)=ζ·W t ·F+κ1·G(R r ,R k ), C rk(t) represents the enhanced correlation between resource r and target node k; ζ is the frequency domain correlation weight coefficient, used to adjust the relative importance of the spectral correlation coefficient in the enhanced correlation calculation, with a value range of [1,5]; κ1 is the topological correlation weight coefficient, used to adjust the importance of the topological structure association strength in the enhanced correlation calculation, with a value range of [0,1]; F is the frequency domain correlation measurement result, which is obtained by performing an integral transform operation on the spectral correlation coefficient, representing the correlation between the resource and target node at all frequencies, used to convert the frequency domain features into a time domain or frequency response representation, and the calculation logic is as follows: e -j2πft is a complex exponential function used to convert frequency domain signals into time domain or frequency characteristics, where j is the complex unit, f is the frequency, 2πf is the relationship between frequency and period, and df is a small change in frequency, representing the integral over all frequency components within the frequency range. By combining time difference, spectral correlation, and topological correlation, a multimodal enhanced correlation model is constructed, which improves the matching accuracy between resources and target nodes, thereby providing learners with more personalized and accurate recommendation results.

[0096] The present invention is further configured such that S6 includes:

[0097] Identify the set of nodes that have cognitive conflict relationships with the target node, and calculate the cognitive topological distance from the target node to the conflicting nodes;

[0098] Extract the feature vectors of the target node and the conflict node, and calculate the adversarial loss value through element-wise multiplication and vector inner product operations;

[0099] Based on the enhanced relevance between resources and target nodes, cognitive topological distance, and adversarial loss value, the security relevance after conflict filtering is calculated. Specifically, a set of nodes that conflict with the target node is generated. This set includes all nodes that have cognitive conflicts with the target node. The cognitive topological distance from the target node to the conflicting node is measured by calculating the degree of the target node in the knowledge graph and the shortest path distance between the target node and the conflicting node. The calculation logic for the cognitive topological distance is as follows: Z(k,q) represents the cognitive topological distance, indicating the cognitive difference between the target node k and the conflicting node q; deg(k) is the degree of the target node k; d G(k,q) represents the shortest path distance between the target node k and the conflicting node q in the knowledge graph, indicating the length of the shortest path between them. This distance can be found by traversing the graph structure, which minimizes the number of steps (or jumps) between them. G represents the knowledge graph. Feature vectors of the target node k and the conflicting node q are extracted using a graph neural network. These feature vectors represent the semantic information of the nodes in the knowledge graph. Adversarial analysis is performed on the feature vectors of the target node k and the conflicting node q, and the adversarial loss value is calculated to measure the degree of conflict between the two nodes and promote the inhibition of irrelevant nodes during the learning process. The logic for calculating the adversarial loss value is as follows: R(e q ,e k ) represents the adversarial loss value between the feature vectors of the target node k and the feature vectors of the conflicting node q; e m e is the feature vector of the conflict node q; k Let ||·||1 be the feature vector of the target node k; q ⊙e k Perform L1 norm; ⊙ represents element-wise product, i.e., product of two vectors e. q and e k Multiply the elements at corresponding positions in the vector one by one to obtain a new vector; Cosine similarity is the dot product of the feature vectors of conflicting node q and target node k. Based on the cognitive topological distance and adversarial loss value, the conflict influence factor of each node on the target node is aggregated. The calculation logic for the conflict influence factor is as follows: I conflict C is the conflict influence factor, representing the degree of conflict between conflicting node q and target node k; k Let φ be the conflict neighborhood of the target node k, that is, the set of all nodes that conflict with the target node k; kq The strength factor is used to determine the strength of the conflict between two nodes, and its value ranges from [0,1]. Based on the enhanced correlation between resources and the target node and the conflict impact factor, the security correlation is calculated. The calculation logic for the security correlation is as follows: C rk' (t) represents the security correlation, i.e., the correlation between resource r and target node k after considering the impact of conflict. When the impact of conflict is significant, i.e. The correlation will be attenuated, reducing the strength of the association between the resource and the target node; The results are obtained by nonlinearly compressing the conflict influence factor to ensure the stability and rationality of the calculation results. The calculation logic is as follows: By identifying cognitive conflict points, we can ensure that recommendations avoid suggesting resources that might cause cognitive conflict to learners, thereby improving learning outcomes.

[0100] The present invention is further configured such that S7 includes:

[0101] Tensor product operations are performed based on energy control and security relevance, and the results of the tensor product operations are transformed by a logarithmic smoothing function.

[0102] The resource fatigue factor is calculated based on the historical exposure frequency of the resource, and the historical exposure time decay value is obtained and multiplied with the resource fatigue factor.

[0103] The resource recommendation priority score is calculated by combining the results of the logarithmic smoothing function transformation and the product operation. Specifically, to optimize resource recommendations and enhance learners' learning outcomes, the resource recommendation priority score is calculated by combining energy mastery, safety relevance, and resource fatigue. The calculation logic for the recommendation priority score is as follows: P r (t) represents the recommendation priority score of resource r; For tensor product operations; L(·) is a logarithmic smoothing function used to logarithmically transform the input values, ensuring that extremely large values ​​do not have an excessive impact on the results, while also mapping small values ​​to a suitable range for calculation; η is an adjustment coefficient used to adjust the influence of energy mastery and safety relevance on the recommendation priority score, with a value range of [0,1]; Ω r The resource fatigue factor represents the degree of resource fatigue. This factor is used to consider the resource fatigue effect in recommendation ranking. When a resource is used frequently, its fatigue level increases, thus reducing its recommendation priority. The calculation logic is as follows: θ is an adjustment coefficient used to control the impact of fatigue on the recommendation priority score; freq(r) represents the frequency at which resource r is accessed or touched by users, i.e., the historical exposure frequency of the resource; β1 is an adjustment coefficient used to control the relationship between the fatigue factor and the resource access frequency, with a value range of [0.5,1]. The historical exposure time decay of resource r represents the time the resource has been exposed to the user. This time decay factor is introduced to prevent recommendations from relying too heavily on historical exposure, ensuring that more recent resources have a higher recommendation priority. Over time, historical exposure time reduces the resource's recommendation priority. By combining multiple factors such as energy mastery, resource security relevance, historical exposure frequency, and fatigue, the recommendation priority ranking method of this invention can dynamically adjust the resource recommendation priority. This method ensures that the recommendation results not only accurately reflect the learner's current needs but also effectively avoid resource fatigue effects, making recommendations more personalized and dynamic, ultimately improving the user experience and learning efficiency of the recommendation system.

[0104] The present invention is further configured to sort resources from high to low priority based on their recommended priority scores;

[0105] The top N resources in the recommendation list are selected from the ranking results and recommended as educational resources to learners. Specifically, the priority score of each resource is calculated, taking into account factors such as relevance, historical exposure time, and resource fatigue. All resources are then ranked according to their priority scores, with higher-scoring resources being recommended first. The top N resources are then selected from the ranked resources and ultimately recommended as educational resources to learners. The recommended resources are arranged from highest to lowest priority to ensure that learners first encounter the most relevant and valuable resources, thereby improving the learning experience and avoiding resource fatigue.

[0106] Example 2

[0107] Please see Figure 2 This exemplary knowledge graph-based online educational resource recommendation system includes:

[0108] Interaction Record Mapping Module: Used to map learners' interaction records within a preset time period to the knowledge graph node space, generating a multi-dimensional knowledge node interaction matrix;

[0109] Mastery calculation module: used to calculate the mastery of each node based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph;

[0110] Mastery Correction Module: This module is used to correct the mastery based on the mastery level and the forgetting curve dynamic correction mechanism, generating a corrected mastery level.

[0111] Energy Propagation Calculation Module: Based on the corrected mastery level, using the path energy decay model, calculates the energy propagation value from the learner's mastered nodes to the target node, and obtains the energy mastery level of the target node;

[0112] Correlation calculation module: used to calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes;

[0113] Cognitive conflict filtering module: Used to filter educational resources that have cognitive conflicts with the learner's existing knowledge based on enhanced relevance, thus obtaining safe relevance;

[0114] Priority Calculation Module: Used to calculate the recommended priority score by combining energy availability and security relevance.

[0115] It should be noted that the knowledge graph-based online education resource recommendation system and the knowledge graph-based online education resource recommendation method provided in the above embodiments belong to the same concept. The specific methods by which each module and unit performs its operations have been described in detail in the method embodiments and will not be repeated here. In practical applications, the knowledge graph-based online education resource recommendation system provided in the above embodiments can be configured to have different functional modules assigned to it as needed. That is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above, and this is not a limitation here.

[0116] 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 scope of the technology 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 knowledge graph-based method for recommending online educational resources, characterized in that, include: S1: Map learners’ interaction records within a preset time period to the knowledge graph node space to generate a multi-dimensional knowledge node interaction matrix; S2: Based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph, calculate the mastery of each node; S3: Based on the mastery level and combined with the forgetting curve dynamic correction mechanism, the mastery level is corrected to generate a corrected mastery level; S4: Based on the corrected mastery level, the path energy decay model is used to calculate the energy propagation value of the learner from the mastered node to the target node, and the energy mastery level of the target node is obtained. S5: Calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes; S6: Based on enhanced relevance, filter educational resources that have cognitive conflicts with the learners' existing knowledge to obtain safety relevance; S7: Calculate the recommendation priority score by combining energy control and safety relevance.

2. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S1 includes: Map learners’ original interaction records within a preset time period to the node space in the knowledge graph; Based on learners' interactive behavior, an interaction feature matrix with time slice granularity is constructed, which represents the interaction strength between learners and each knowledge node; The interaction feature matrix is ​​processed, and an embedding layer is used to map the interaction data to a high-dimensional semantic space. Then, a semantic vector representation of each knowledge node is generated through linear transformation. The Hadamard product operation is used to fuse learners’ interaction data with the semantic vectors of knowledge nodes at the element level. Based on the Frobenius norm of knowledge graph nodes, the interaction intensity is regularized to generate a knowledge node interaction matrix.

3. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S2 includes: Based on the knowledge node interaction matrix, combined with the path dependency weights derived from the knowledge graph and the current interaction intensity, a nonlinear normalization transformation is performed, and a time decay function is applied to adjust the normalization result. At the same time, a Gaussian process perturbation term is injected to simulate cognitive random fluctuations. By applying a quadratic exponential decay mechanism, the knowledge propagation effect of long paths is suppressed, and the mastery of output nodes is improved.

4. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S3 includes: Based on the difference between the most recent interaction time and the current time of the nodes, and combined with path length modulation, the forgetting degree of each node is calculated. The path length modulation adjusts the attenuation effect of information propagation according to the path length between nodes. Calculate the corrected mastery of nodes based on their forgetting and mastery levels.

5. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S4 includes: Based on the node's corrected mastery and combined with the shortest path distance between nodes, an exponential decay model is used to adjust the intensity of the path propagation effect. Based on the semantic vector distance between nodes, a hyperbolic modulation mechanism is applied to adjust the path propagation effect; The attenuation model adjustment results and hyperbolic modulation results are combined, and the weights of the path levels are adjusted using a random walk model. By using a dynamic integration mechanism, the energy propagation effect of multi-level paths is aggregated to calculate the energy mastery of the target node.

6. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S5 includes: Calculate the absolute deviation between the resource update time difference and the target node time difference, and construct a time window suppression function based on this deviation. Based on the frequency domain characteristics of resources and target nodes, the spectral correlation coefficient between resources and target nodes is calculated; Analyze the relative positions and adjacency relationships between resources and target nodes in the knowledge graph, and evaluate the strength of the topological association between resources and target nodes; By combining the spectral correlation coefficient and topological association strength under the constraint of the time window suppression function, the enhanced correlation between resources and target nodes is obtained.

7. The online educational resource recommendation method based on knowledge graphs according to claim 1, characterized in that, S6 includes: Identify the set of nodes that have cognitive conflict relationships with the target node, and calculate the cognitive topological distance from the target node to the conflicting nodes; Extract the feature vectors of the target node and the conflict node, and calculate the adversarial loss value through element-wise multiplication and vector inner product operations; Based on the enhanced correlation between resources and target nodes, cognitive topological distance, and adversarial loss value, the security correlation after conflict filtering is calculated.

8. The method for recommending online educational resources based on knowledge graphs according to claim 1, characterized in that, S7 includes: Tensor product operations are performed based on energy control and security relevance, and the results of the tensor product operations are transformed by a logarithmic smoothing function. The resource fatigue factor is calculated based on the historical exposure frequency of the resource, and the historical exposure time decay value is obtained and multiplied with the resource fatigue factor. The recommended priority score for a resource is calculated by combining the results of the logarithmic smoothing function transformation and the product operation.

9. The online educational resource recommendation method based on knowledge graphs according to claim 8, characterized in that, Based on the recommended priority score, resources are sorted from highest to lowest priority. The top N resources in the recommendation list are selected from the sorting results and recommended as educational resources to learners.

10. A knowledge graph-based online educational resource recommendation system, used to implement the knowledge graph-based online educational resource recommendation method according to any one of claims 1-9, characterized in that, include: Interaction Record Mapping Module: Used to map learners' interaction records within a preset time period to the knowledge graph node space, generating a multi-dimensional knowledge node interaction matrix; Mastery calculation module: used to calculate the mastery of each node based on the knowledge node interaction matrix, combined with the interaction intensity of time segments and the path dependency relationship of the knowledge graph; Mastery Correction Module: This module is used to correct the mastery based on the mastery level and the forgetting curve dynamic correction mechanism, generating a corrected mastery level. Energy Propagation Calculation Module: Based on the corrected mastery level, using the path energy decay model, calculates the energy propagation value from the learner's mastered nodes to the target node, and obtains the energy mastery level of the target node; Correlation calculation module: used to calculate the enhanced correlation between resources and target nodes based on the characteristics and spatiotemporal states of resources and target nodes; Cognitive conflict filtering module: Used to filter educational resources that have cognitive conflicts with the learner's existing knowledge based on enhanced relevance, thus obtaining safe relevance; Priority Calculation Module: Used to calculate the recommended priority score by combining energy availability and security relevance.

Images