A multi-knowledge sub-graph based electronic learning recommendation system design method

Abstract

The application discloses a kind of based on multiple knowledge subgraph's electronic learning recommendation system design method.Knowledge graph is used as auxiliary information source to carry out electronic learning recommendation system design, and based on multiple knowledge subgraph's electronic learning recommendation system includes learner module, learning goal module and scoring module.The application has the ability to solve the problem that learner faces overload learning goal, which leads to that the recommendation system cannot accurately recommend and new learner cold start etc.;The application uses knowledge graph as information auxiliary source, and adopts attention mechanism to distinguish the contribution degree of different relationship attributes to target entity;By aggregating neighbor entity, the application enriches the semantic representation of target entity, and by the mode of propagation, the potential contact between learning target entity is mined, and the application provides personalized learning target recommendation for learner, thereby improving the learning effect and the beneficial effect of satisfaction of learner.

Description

Technical Field

[0001] This invention relates to a design method for a learning recommendation system, and more particularly to a design method for an electronic learning recommendation system based on multiple knowledge subgraphs. Background Technology

[0002] With the availability of a vast amount of online course resources, learners face the challenge of choosing resources without being overwhelmed by a multitude of learning options—a problem known as "information overload." Most online courses are geared towards a broad audience, failing to consider learners' background knowledge and skill levels, resulting in consistently unsatisfactory online teaching effectiveness. Therefore, how to utilize recommendation systems to alleviate "information overload" in educational settings, improve the effectiveness of e-learning systems, and help learners find learning goals that align with their own knowledge levels has become a primary research objective for e-learning recommendation systems.

[0003] In recent years, with the rapid development of deep learning, recommender systems combined with deep learning have been extensively studied by researchers. By introducing knowledge graphs as an information aid into recommender systems, researchers have significantly improved their performance, providing a solution for personalized knowledge point recommendations to learners in educational settings. Knowledge graphs, as a structured form of human knowledge, have attracted widespread attention in academia and industry. They are based on entities, relations, and semantic descriptions. Entities represent objects and abstract concepts in the real world, relations represent connections between entities, and the semantic descriptions of entities and their relations encompass clearly defined meanings such as type and attributes. However, recommender systems that use knowledge graphs as an information aid often focus too much on the item modeling module and neglect the importance of the user module. In educational settings, however, it is essential to pay attention to the learner's own attribute characteristics.

[0004] To address the aforementioned issues and improve the shortcomings of recommendation systems in educational settings, this invention provides a design method for an e-learning recommendation system based on multiple knowledge subgraphs. Summary of the Invention

[0005] The purpose of this invention is to provide a design method for an e-learning recommendation system based on multiple knowledge subgraphs. This method addresses issues such as learners facing an overload of learning objectives, leading to inaccurate recommendations and cold start problems for new learners. The invention utilizes knowledge graphs as an information auxiliary source and employs an attention mechanism to differentiate the contribution of different relational attributes to target entities. By aggregating neighboring entities, the invention enriches the semantic representation of target entities and mines potential connections between learning target entities through propagation. This invention provides learners with personalized learning objective recommendations, thereby improving learners' learning effectiveness and satisfaction.

[0006] The technical solution of this invention is: a design method for an e-learning recommendation system based on multiple knowledge subgraphs. The method utilizes knowledge graphs as auxiliary information sources to design the e-learning recommendation system. The e-learning recommendation system based on multiple knowledge subgraphs includes a learner module, a learning objective module, and a rating module.

[0007] In the aforementioned design method of an e-learning recommendation system based on multiple knowledge subgraphs, the e-learning recommendation system based on multiple knowledge subgraphs provides individualized instruction to learners based on their individual attributes, the time they spend on learning objectives, and the learning clusters that have the same knowledge acquisition paradigm.

[0008] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learner module combines a learner-learning-target interaction subgraph, a learner cluster subgraph, and a learner's own attribute subgraph to create a learner profile; the learning target module combines a learning target feature subgraph and a learning target subgraph to mine semantic information between learning targets; and the rating module calculates each learner's rating for a learning target and recommends high-rated learning targets to the learner.

[0009] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learner-learning-goal interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph respectively obtain the feature representation of the learner in the three subgraphs. and By giving and Different weights are assigned to represent the importance of the three features to the final learner feature representation, with the aim of obtaining the learner feature representation L that performs best in the recommender system. i :

[0010]

[0011] Where σ is a nonlinear activation function, and ω1, ω2, and ω3 are respectively... and The weights are denoted by ⊕, which represents vector addition, and b is the bias. and These are the learner feature representations obtained from the learning objective interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph, respectively.

[0012] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learner features By aggregating learners i The aggregation function, derived from the set of learning objectives C(i) he has studied, is defined as follows:

[0013]

[0014] Where σ is a nonlinear activation function, w and b are weight matrices and biases, respectively; C(i) is the set of learning objectives learned by the learner, and p j These are the embedding vectors of these learning objectives; by assigning different weights to each learning objective to represent the importance of each historical learning objective to the learner, the aim is to better simulate the learner's learning preferences.

[0015]

[0016] Where σ is a nonlinear activation function, w and b are the weight matrix and bias, respectively; p j These are the embedding vectors of the learning objectives, α ij It refers to the attention weighting of learning objectives, which aims to distinguish the importance of different learning objectives to learners. ij Calculated by the neural attention network:

[0017]

[0018] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q j p is the learner's embedding vector. j These are the embedding vectors of these learning objectives, where ⊕ represents vector addition. To capture the contribution of each learning objective to the target user, the attention score is normalized using the softmax function.

[0019]

[0020] in, Indicates to To index.

[0021] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learner features By setting learner l i Let C(i) be the set of learning objectives learned, and let l be the learner. j The set of learned learning objectives is C(j). If l i and learners i If learners have shared a common learning objective, then their sets of learning objectives have an intersection, namely C(ij). Based on this concept, a knowledge graph of learning clusters is constructed, where the relationships between learners are represented by the number of elements in the intersection C(ij) of their shared learning objectives. S(i) is defined as learner l. i A set of similar learners, which is connected to learner l in the learning cluster subgraph. iThe learner composition, defined as the learner feature representation obtained by aggregating learning cluster subgraphs, is... Aggregate functions are defined as follows:

[0022]

[0023] Where σ is a nonlinear activation function, w and b are weight matrices and biases, and S(i) is the sum of the learning function and the bias of the learner. i A set of similar learners, q j These are the embedding vectors of these learners. By assigning different weights to each similar learner to represent the importance of each similar learner to the learner, the aim is to better simulate learners' school preferences.

[0024]

[0025] Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and β is the bias. ij It represents the attention weights of similar learners, q j These are the embedding vectors of similar learners;

[0026] In distinguishing the importance of different similar learners to learners, β ij Calculated by the neural attention network:

[0027]

[0028] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q i It is the embedding vector of the target learner, q j Here are the embedding vectors of similar learners, and ⊕ denotes vector addition. To capture the contribution of each similar learner to the target learner, the attention score is normalized using the softmax function.

[0029]

[0030] in, Indicates to To index.

[0031] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learner features O(i) represents learner l i The feature set is used to obtain learner feature representations by aggregating features such as learner age, knowledge level, and major. Aggregate functions are defined as follows:

[0032]

[0033] Where σ is a nonlinear activation function, w and b are weight matrices and biases, and O(i) is the sum of the learning thresholds and the learning thresholds. i Feature set, t j It is the embedding vector of the learner's own features; by assigning different weights to each feature to represent the importance of each feature to the learner, the aim is to better simulate the learner's school preferences:

[0034]

[0035] Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and λ is the weight matrix and bias. ij It is the learner's own characteristic attention weight, t j It is the embedding vector of the learner's own features, designed to distinguish the importance of different features to the target learner, λ ij Calculated by the neural attention network:

[0036]

[0037] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively, and q is a nonlinear activation function. i It is the embedding vector of the target learner, t j This is the embedding vector of the learner's own features, where ⊕ represents vector addition. To capture the contribution of each learner's own features to the target learner, the attention score is normalized using the softmax function.

[0038]

[0039] in, Indicates to To index.

[0040] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learning target feature subgraph and the learning target subgraph respectively obtain the feature representation of the learning target in the two subgraphs. and By aggregating the two, we obtain the final learning target embedding representation K. i :

[0041]

[0042] in, It is the feature representation of the learning objective obtained from the learning objective subgraph. The learning target features are obtained from the feature subgraph of the learning target, and ⊕ represents the vector addition operation.

[0043] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learning target features This is achieved by representing the learning target k as N(i). i The set of attributes, through aggregation, learns the target k. i The learning target k is obtained from the various attributes. i Feature representation Aggregate functions are defined as follows:

[0044]

[0045] Where σ is a nonlinear activation function, w and b are the weight matrix and bias, and N(i) is the learning target k. i The set of attributes, n j It is the embedding vector of the learning target attribute.

[0046] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the learning target features It is achieved by recursively aggregating and learning the target k. i The relevant k-hop entities are used to ultimately obtain the learning target k. i The embedding representation; specifically: the learning target k in the learning target subgraph i The set of k-hop entities is defined as:

[0047]

[0048] When k = 0 This represents the learning target node subgraph at this point, representing the learning target k. i The characteristic is its own embedding vector p. i H is the maximum number of hops k, G is the knowledge graph, and the set of related entities can be regarded as the learning target k in the knowledge node subgraph. i The extension;

[0049] Learning objective k in the knowledge node subgraph i The k-jump ripple set is defined as starting from... The initial set of learning objective triples:

[0050]

[0051] Each learning objective k is embedded in the learning objective p. j ∈R d The related terms are: d is the embedding dimension; H is the maximum number of hops k; and G is the knowledge graph.

[0052] Given learning objective p j And the set of triples affected by its k=1 jump. for Each triple (h) i ,r i ,t i), by calculating the learning objective p j The relation r in this triple i Chinese h i Similarity P between them:

[0053]

[0054] Among them, R i ∈R d×d It is a relation r i The embedding vector, h i ∈R d It is the embedding of the head entity in the ripple set; P represents the embedding in the relation space R. i The selected learning objective p j and entity h i Similarity between them; after obtaining the similarity P, by analyzing the similarity between them... The tail entity in the vector is obtained by weighting:

[0055]

[0056] Among them, t i ∈R d It is the embedding of the tail entity t, vector regarded as a learning objective p j First-order response in the ripple set; P is the learning target p j The relation r in this triple i Chinese h i The similarity P between them;

[0057] By using p in formula (18) j Replace with The learning objective p can be obtained through repeated propagation. j Second-order, third-order and H-order responses The learning target k can be obtained from the responses of all combinations. i The final embedding

[0058]

[0059] Where, vector regarded as a learning objective p j First-order response in the ripple set, vector regarded as a learning objective p j Second-order response in a ripple set, vector regarded as a learning objective p j H-order response in the ripple set.

[0060] In the aforementioned design method for an e-learning recommendation system based on multiple knowledge subgraphs, the scoring module uses learner features L from formula (1). i The learning objective feature K of formula (14) i The data is then concatenated, and the interaction score between the two is calculated through multiple fully connected layers:

[0061]

[0062] Among them, L i and K i These are the final feature representations of the learner and the learning objective, respectively.

[0063] The parameters involved in training are updated using the loss function (22):

[0064]

[0065] Where m is the number of samples and λ is the regularization parameter; The interaction score between the two is calculated through multiple fully connected layers. lk is the interaction score between the learner and the historical learning target, w is the model's weight parameter, and L is the number of fully connected layers. The optimal parameters are obtained by minimizing the loss function to calculate each learner's score for the learning target, and then the learning target with the highest score is recommended to the learner.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. The present invention discloses an electronic learning recommendation system design method. This method utilizes a knowledge graph as an auxiliary information source. First, a heterogeneous knowledge graph is constructed, where entities are learning objectives and their attributes, and learners and their attributes. Second, feature representations of each final learner and learning objective are obtained by modeling the learner module and the learning objective module respectively. Finally, a scoring module is used to calculate the probability of interaction between the two and recommend the learning objective with the highest score to the learner. The knowledge graph, as an information aid for the recommendation system, has the following advantages: First, it enriches the knowledge base of the recommendation system by providing a wide range of entities, attributes, and relationships, enabling the system to acquire in-depth and comprehensive domain knowledge. Second, the knowledge graph reveals the connections and similarities between knowledge points, allowing the recommendation system to provide more accurate and comprehensive recommendation results by analyzing these relationships. Third, by associating learners' personal information with the knowledge graph, the recommendation system can achieve personalized recommendations and better understand learners' interests and needs. Fourth, the knowledge graph spans multiple domains and topics, providing diverse knowledge coverage for the recommendation system, enabling learners to obtain comprehensive and diverse learning resources and path choices, and providing learners with a better learning recommendation experience.

[0068] 2. The system design method of this invention is based on multiple knowledge subgraphs. Through the combination and association of different subgraphs, it achieves the mining of learner profiles and learning objectives, thereby providing personalized learning recommendations. First, the learner module creates a learner profile by combining information from the learner-learning-objective interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph. Next, the learning objective module mines the semantic information between learning objectives by combining the learning objective feature subgraph and the learning objective subgraph. Finally, the scoring module calculates the learner's rating of the learning objectives and recommends high-scoring learning objectives to the learner. Through the combination and association of multiple knowledge subgraphs, a comprehensive understanding and mining of learners and learning objectives is achieved. By combining learner profiles and the semantic information of learning objectives, the system can provide personalized, accurate, and relevant learning recommendations, improving learning effectiveness and user satisfaction. Utilizing multiple knowledge subgraphs can integrate learners and learning objectives, providing more comprehensive, accurate, and personalized learning recommendations. By integrating the knowledge from different subgraphs, the recommendation system can better understand the relationship between learners and knowledge, providing learners with a better learning experience and learning outcomes.

[0069] 3. The system design method of this invention obtains the optimal learner feature representation by weighted combination of the feature representations of the learner-learning target interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph, for application in the recommendation system. First, the learner-learning target interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph capture the learner's features in three aspects, respectively. These subgraphs are then used to extract features using a graph neural network method to obtain the corresponding learner feature representations, which are as follows: and Secondly, to determine the importance of each feature to the final learner feature representation, weight parameters ω1, ω2, and ω3 are introduced to assign weights to different features. Finally, the three feature representations are weighted and combined using a non-linear activation function to obtain the final learner feature representation L. i By utilizing graph neural networks and assigning weights to the three elements (graph, graph, and graph), the expressive power and personalization of recommendation systems can be improved, capturing multi-dimensional features, flexibly adjusting the importance of features, and integrating information from different levels. These advantages help recommendation systems provide more accurate, personalized, and comprehensive learning recommendations, thereby improving learners' learning outcomes and user satisfaction.

[0070] 4. The learner feature representation obtained from the learner-learning-goal interaction subgraph described in this invention This design method primarily uses graph neural networks and attention mechanisms to model the relationship between learners and learning objectives, aiming to capture learners' learning characteristics and personalize learning objective recommendations. First, learner characteristics are obtained by aggregating the set of learning objectives C(i) that the learner has previously learned. Second, to better simulate learners' learning preferences, an attention mechanism is used to assign different weights to each historical learning objective, representing its importance to the learner. Finally, to capture the contribution of each learning objective to the target learner, a softmax function is used to normalize the attention score. The set of learning objectives already learned by the learner can reflect their learning preferences to a certain extent. The importance of each learning objective to the learner can be measured based on the learning time spent on each historical objective. Therefore, by introducing an attention mechanism into the learner-learning objective interaction subgraph, a more accurate representation of learner characteristics can be obtained, enabling personalized learning objective recommendations and better meeting learners' needs.

[0071] 5. The learner feature representation obtained from the learner cluster subgraph described in this invention This design approach primarily uses graph neural networks and attention mechanisms to model the relationships between learners and similar learners, capturing learner cluster characteristics and personalizing learning resource recommendations. First, the similarity between learners is determined based on their learning goal sets. Specifically, if two learners share common learning goals, their goal sets will intersect. Based on this concept, a learning cluster knowledge graph is constructed, where the relationship between learners is represented by the number of elements in the intersection of their shared learning goals. This representation quantifies the degree of similarity between learners. Then, based on the set of learners connected to the target learner in the learning cluster subgraph, a set of similar learners S(i) for the target learner is defined. This set consists of learners connected to the target learner in the learning cluster subgraph. Second, to better simulate learners' learning preferences, an attention mechanism is used to assign different weights to each similar learner, representing their importance to the learner. Finally, to capture the contribution of each similar learner to the target learner, a softmax function is used to normalize the attention scores. Learners with similar knowledge acquisition paradigms often exhibit a certain degree of similarity. By constructing clusters of similar learners, we can better understand the similarity and relationships among learners, consider their learning preferences, and provide personalized learning recommendations, thereby improving learners' learning outcomes and satisfaction.

[0072] 6. The learner feature representation obtained from the learner's own attribute subgraph as described in this invention This design method primarily uses graph neural networks and attention mechanisms to model the relationship between learners and their attributes, capturing learner attribute characteristics and personalizing learning resource recommendations. First, the learner's feature set is represented as O(i), including features such as grade level, knowledge level, and major. An aggregation function is introduced to weight and sum these features according to different weights, yielding the learner's feature representation. Second, to better simulate learners' learning preferences, an attention mechanism is used to assign different weights to each learning feature, representing its importance to the learner. Finally, to capture the contribution of each learner's feature to the target learner, a softmax function is used to normalize the attention score. Learners' own characteristics largely reflect their learning preferences and play a crucial role in solving the cold start problem. When a new learner chooses to use the system, by registering a personal account and providing personal information such as age, knowledge level, and major, the system can build a personalized profile and recommendation model for the new learner. Since newly registered learners have no historical learning records, the system provides targeted learning goal recommendations by matching the preferences of historical learners with similar characteristics. This method effectively solves the cold start problem faced by new learners in the system. By leveraging learners' individual characteristics and the behavioral patterns of similar learners, the system can better understand the needs of new learners and provide them with suitable learning resources and suggestions, thereby improving their learning outcomes and satisfaction.

[0073] 7. The learning target module of this invention, the design method obtains the feature representation of the learning target in two subgraphs respectively through a learning target feature subgraph and a learning target subgraph, and then obtains the final learning target embedding representation by aggregating these two feature representations. First, the learning target feature subgraph is a subgraph constructed based on the feature information of the learning target. This subgraph can contain attributes related to the learning target, such as subject area, difficulty level, application scenario, etc. By processing the learning target feature subgraph, the feature representation of the learning target can be obtained, denoted as K. i N Secondly, the learning objective subgraph is a subgraph constructed based on the relationships and connections between learning objectives. This subgraph can represent the similarity, correlation, or dependency among learning objectives. By processing the learning objective subgraph, we can obtain the feature representation of the learning objectives, denoted as K. i O Finally, in order to obtain the final learning objective, the embedding representation K... i This method aggregates the feature representations of the learning target feature subgraphs to obtain a more comprehensive and accurate embedding representation of the learning target. It comprehensively considers different features of the learning target, provides a more accurate measure of learning target similarity, and enhances the expressive power of the learning target.

[0074] 8. The learning target feature representation K obtained from the learning target feature subgraph as described in this invention i N This design method represents the features of the learning objective using a set O(i), which includes the subject area, difficulty level, and application scenario of the learning objective. Finally, an aggregation function is introduced to aggregate these features, resulting in a feature representation of the learning objective. By considering the features of the learning objective, its feature representation can be enriched, thereby improving the accuracy and personalization capabilities of the recommendation system.

[0075] 9. The learning target feature representation K obtained from the learning target feature subgraph as described in this invention i O This design method first obtains the embedding representation of the learning objective by recursively aggregating k-hop entities related to the learning objective. In this process, the set of k-hop entities of the learning objective in the learning objective subgraph is defined as a ripple set, where k represents the number of hops, H represents the maximum number of hops, and G represents the knowledge graph. When k=0, the feature of the learning objective in the learning objective node subgraph is its own embedding vector. The set of related entities can be considered as an extension of the learning objective in the knowledge node subgraph. Secondly, for each learning objective and the set of triples involved in its k=1 hops, the similarity between the learning objective and the relation in each triple is calculated. Here, the embedding vector of the relation and the embedding vector of the head entity in the ripple set are used to calculate the similarity. The similarity P represents the degree of similarity between the selected learning objective and the entity. Thirdly, based on the calculated similarity P, the tail entities in the ripple set are weighted and summed to obtain the vector of the first-order response. Here, the embedding vector of the tail entity is used to represent the first-order response. Finally, through multiple iterative calculations, the second and third orders of the learning objective can be obtained, until the maximum number of hops H is reached, yielding the final embedding of the learning objective. This method allows for a better understanding of the relationships and similarities between learning objectives, fully utilizing information from the knowledge graph to provide more accurate and personalized learning recommendations, thereby improving the learner's learning experience.

[0076] 10. The scoring module of this invention first concatenates the learner's features and the learning target's features to form a comprehensive feature representation. Second, it calculates the interaction score between the learner's features and the learning target's features using multiple fully connected layers. Third, it optimizes the loss function to obtain the optimal parameter configuration for calculating each learner's score for the learning target. Finally, high-scoring learning targets are recommended to learners as personalized learning target recommendations.

[0077] In summary, this invention addresses the problems of learners facing an overload of learning objectives, leading to inaccurate recommendations by the system and cold start for new learners. It utilizes knowledge graphs as an auxiliary information source and employs an attention mechanism to differentiate the contribution of different relational attributes to target entities. By aggregating neighboring entities, it enriches the semantic representation of target entities and mines potential connections between learning target entities through propagation. This invention provides learners with personalized learning objective recommendations, thereby improving learners' learning outcomes and satisfaction. Attached Figure Description

[0078] Figure 1 This is the overall architecture diagram of the electronic learning recommendation system of this invention;

[0079] Figure 2 This is a modeling diagram of the learner module in the e-learning recommendation system of this invention;

[0080] Figure 3 This is a modeling diagram of the learning objective module of the e-learning recommendation system of the present invention;

[0081] Figure 4 This is a modeling diagram of the learning objective subgraph module of the electronic learning recommendation system of this invention;

[0082] Figure 5 This is a graph showing the impact of different variant experiments of the electronic learning recommendation system of this invention on the system evaluation index on dataset one;

[0083] Figure 6 This is a graph showing the impact of different variant experiments of the electronic learning recommendation system of this invention on the system evaluation index on dataset 2;

[0084] Figure 7 This is a graph showing the impact of different variant experiments of the electronic learning recommendation system of this invention on the system evaluation index on dataset 3;

[0085] Figure 8 This is a graph showing the impact of whether or not learner attributes are considered on the system evaluation indicators of the electronic learning recommendation system of this invention.

[0086] Figure 9 This is a diagram showing the influence of different feature weights of the e-learning recommendation system of the present invention on the system evaluation index in the learner modeling module;

[0087] Figure 10 This diagram illustrates the impact of different regularization parameters on system evaluation metrics when optimizing the electronic learning recommendation system of this invention. Detailed Implementation

[0088] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0089] Example. A design method for an e-learning recommendation system based on multiple knowledge subgraphs. This method utilizes knowledge graphs as auxiliary information sources to design the e-learning recommendation system. The e-learning recommendation system based on multiple knowledge subgraphs includes a learner module, a learning objective module, and a rating module.

[0090] The learner module combines a learner-learning-target interaction subgraph, a learner cluster subgraph, and a learner's own attribute subgraph to create a learner profile. The learning target module combines a learning target feature subgraph and a learning target subgraph to mine semantic information between learning targets. The rating module calculates each learner's rating of the learning target and recommends high-rated learning targets to learners.

[0091] The learner-learning-objective interaction subgraph, learner cluster subgraph, and learner self-attribute subgraph respectively yield feature representations of the learner in the three subgraphs. and By giving and Different weights are assigned to represent the importance of the three features to the final learner feature representation, with the aim of obtaining the learner feature representation L that performs best in the recommender system. i :

[0092]

[0093] Where σ is a nonlinear activation function, and ω1, ω2, and ω3 are respectively... and The weights are denoted by ⊕, which represents vector addition, and b is the bias. and These are the learner feature representations obtained from the learning objective interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph, respectively.

[0094] learner characteristics By aggregating learners i The aggregation function, derived from the set of learning objectives C(i) he has studied, is defined as follows:

[0095]

[0096] Where σ is a nonlinear activation function, w and b are weight matrices and biases, respectively; C(i) is the set of learning objectives learned by the learner, and p j These are the embedding vectors of these learning objectives; by assigning different weights to each learning objective to represent the importance of each historical learning objective to the learner, the aim is to better simulate the learner's school preferences:

[0097]

[0098] Where σ is a nonlinear activation function, w and b are the weight matrix and bias, respectively; p j These are the embedding vectors of the learning objectives, α ij It refers to the attention weighting of learning objectives, which aims to distinguish the importance of different learning objectives to learners. ij Calculated by the neural attention network:

[0099]

[0100] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q j p is the learner's embedding vector. j These are the embedding vectors of these learning objectives, where ⊕ represents vector addition. To capture the contribution of each learning objective to the target user, the attention score is normalized using the softmax function.

[0101]

[0102] in, Indicates to To index.

[0103] learner characteristics By setting learner l i Let C(i) be the set of learning objectives learned, and let l be the learner. j The set of learned learning objectives is C(j). If l i and learners i If learners have shared a common learning objective, then their sets of learning objectives have an intersection, namely C(ij). Based on this concept, a knowledge graph of learning clusters is constructed, where the relationships between learners are represented by the number of elements in the intersection C(ij) of their shared learning objectives. S(i) is defined as learner l. i A set of similar learners, which is connected to learner l in the learning cluster subgraph. i The learner composition, defined as the learner feature representation obtained by aggregating learning cluster subgraphs, is... Aggregate functions are defined as follows:

[0104]

[0105] Where σ is a nonlinear activation function, w and b are weight matrices and biases, and S(i) is the sum of the learning function and the bias of the learner. i A set of similar learners, q jThese are the embedding vectors of these learners. By assigning different weights to each similar learner to represent the importance of each similar learner to the learner, the aim is to better simulate learners' school preferences.

[0106]

[0107] Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and β is the bias. ij It represents the attention weights of similar learners, q j These are the embedding vectors of similar learners;

[0108] In distinguishing the importance of different similar learners to learners, β ij Calculated by the neural attention network:

[0109]

[0110] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q i It is the embedding vector of the target learner, q j Here are the embedding vectors of similar learners, and ⊕ denotes vector addition. To capture the contribution of each similar learner to the target learner, the attention score is normalized using the softmax function.

[0111]

[0112] in, Indicates to To index.

[0113] learner characteristics O(i) represents learner l i The feature set is used to obtain learner feature representations by aggregating features such as learner age, knowledge level, and major. Aggregate functions are defined as follows:

[0114]

[0115] Where σ is a nonlinear activation function, w and b are weight matrices and biases, and O(i) is the sum of the learning thresholds and the learning thresholds. i Feature set, t j It is the embedding vector of the learner's own features; by assigning different weights to each feature to represent the importance of each feature to the learner, the aim is to better simulate the learner's school preferences:

[0116]

[0117] Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and λ is the weight matrix and bias.ij It is the learner's own characteristic attention weight, t j It is the embedding vector of the learner's own features, designed to distinguish the importance of different features to the target learner, λ ij Calculated by the neural attention network:

[0118]

[0119] Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively, and q is a nonlinear activation function. i It is the embedding vector of the target learner, t j This is the embedding vector of the learner's own features, where ⊕ represents vector addition. To capture the contribution of each learner's own features to the target learner, the attention score is normalized using the softmax function.

[0120]

[0121] in, Indicates to To index.

[0122] The e-learning recommendation system design method based on multi-knowledge subgraphs is characterized in that: the learning target feature subgraph and the learning target subgraph respectively obtain the feature representation of the learning target in the two subgraphs. and By aggregating the two, we obtain the final learning target embedding representation K. i :

[0123]

[0124] in, It is the feature representation of the learning objective obtained from the learning objective subgraph. The learning target features are obtained from the feature subgraph of the learning target, and ⊕ represents the vector addition operation.

[0125] The learning objective features This is achieved by representing the learning target k as N(i). i The set of attributes, through aggregation, learns the target k. i The learning target k is obtained from the various attributes. i Feature representation Aggregate functions are defined as follows:

[0126]

[0127] Where σ is a nonlinear activation function, w and b are the weight matrix and bias, and N(i) is the learning target k. i The set of attributes, n jIt is the embedding vector of the learning target attribute.

[0128] The learning objective features It is achieved by recursively aggregating and learning the target k. i The relevant k-hop entities are used to ultimately obtain the learning target k. i The embedding representation; specifically: the learning target k in the learning target subgraph i The set of k-hop entities is defined as:

[0129]

[0130] When k = 0 This represents the learning target node subgraph at this point, representing the learning target k. i The characteristic is its own embedding vector p. i H is the maximum number of hops k, G is the knowledge graph, and the set of related entities can be regarded as the learning target k in the knowledge node subgraph. i The extension;

[0131] Learning objective k in the knowledge node subgraph i The k-jump ripple set is defined as starting from... The initial set of learning objective triples:

[0132]

[0133] Each learning objective k is embedded in the learning objective p. j ∈R d The related terms are: d is the embedding dimension; H is the maximum number of hops k; and G is the knowledge graph.

[0134] Given learning objective p j And the set of triples affected by its k=1 jump. for Each triple (h) i ,r i ,t i ), by calculating the learning objective p j The relation r in this triple i Chinese h i Similarity P between them:

[0135]

[0136] Among them, R i ∈R d×d It is a relation r i The embedding vector, h i ∈R d It is the embedding of the head entity in the ripple set; P represents the embedding in the relation space R. i The selected learning objective pj and entity h i Similarity between them; after obtaining the similarity P, by analyzing the similarity between them... The tail entity in the vector is obtained by weighting:

[0137]

[0138] Among them, t i ∈R d It is the embedding of the tail entity t, vector regarded as a learning objective p j First-order response in the ripple set; P is the learning target p j The relation r in this triple i Chinese h i The similarity P between them;

[0139] By using p in formula (18) j Replace with The learning objective p can be obtained through repeated propagation. j Second-order, third-order and H-order responses The learning target k can be obtained from the responses of all combinations. i The final embedding

[0140]

[0141] Where, vector regarded as a learning objective p j First-order response in the ripple set, vector regarded as a learning objective p j Second-order response in a ripple set, vector regarded as a learning objective p j H-order response in the ripple set.

[0142] The scoring module uses learner features L from formula (1) i The learning objective feature K of formula (14) i The data is then concatenated, and the interaction score between the two is calculated through multiple fully connected layers:

[0143]

[0144] Among them, L i and K i These are the final feature representations of the learner and the learning objective, respectively.

[0145] The parameters involved in training are updated using the loss function (22):

[0146]

[0147] Where m is the number of samples and λ is the regularization parameter; The interaction score between the two is calculated through multiple fully connected layers. lk is the interaction score between the learner and the historical learning target, w is the model's weight parameter, and L is the number of fully connected layers. The optimal parameters are obtained by minimizing the loss function to calculate each learner's score for the learning target, and then the learning target with the highest score is recommended to the learner.

[0148] Experiments have shown that:

[0149] The training process of the e-learning recommendation system based on multiple knowledge subgraphs of this invention was carried out in the Ubuntu 18.04 and CUDA 10.2 environment, with GPU configuration of NVIDIA GeForce RTX 4090, 24GB of video memory, and the GPU was used for training.

[0150] All experimental training parameter settings: Training was conducted using three public datasets. The optimal hyperparameter selections for different datasets are shown in Table 1. The Adam optimizer was used, trained for 100 epochs with an early stopping strategy. Mean squared error was used to calculate the difference between the predicted and target values. Regularization was employed to prevent overfitting. Root mean squared error (RMSE) and mean absolute error (MAE) were selected as evaluation metrics for model performance.

[0151] Table 1

[0152] Dataset Dimension Number of jumps Regularization parameters Learning rate Dataset-1 32 2 <![CDATA[10 -5 ]]> 0.1 Dataset-2 16 1 <![CDATA[10 -3 ]]> 0.001 Dataset-3 64 1 <![CDATA[10 -7 ]]> 0.001

[0153] The overall architecture diagram of the e-learning recommendation system of this invention is as follows: Figure 1 As shown, it consists of three key parts: a learner module, a knowledge module, and a rating prediction module. In the learner module, this invention aggregates the learner-learning goal interaction subgraph, the learning cluster subgraph, and the learner's own attribute subgraph to obtain learner feature representations in the three subgraphs. and Furthermore, an attention mechanism is used during the aggregation process to distinguish the importance of different neighboring entities to the target learner. Finally, the three are aggregated and assigned different weights to obtain the learner's final feature representation L. i In the learning objective module, this invention aggregates the learning objective subgraph and the learning objective feature subgraph to obtain the feature representation of the learning objective in the two subgraphs. and Finally, the two are aggregated to obtain the final feature representation K of the learning objective. iIn the rating prediction module, this invention integrates learner feature representations and learning target feature representations and places them into a fully connected layer (MLP) to predict the probability of association between the two. The present invention compares the model with other baseline models, and the results are shown in Table 2:

[0154] Table 2

[0155]

[0156] As shown in the table, the model of this invention outperforms other baseline models in all three datasets. Compared to the best-performing baseline model, Ripple Net, the RMSE of the three datasets is reduced by approximately 4.23%, 4.46%, and 4.29%, respectively, and the MAE is reduced by approximately 6.15%, 4.23%, and 4.76%, respectively. This indicates that the recommendation effect of the model of this invention is superior to other baseline models.

[0157] The learner module modeling of the e-learning recommendation system of this invention is as follows: Figure 2 As shown, this invention aggregates the learner-learning target interaction subgraph, the learning cluster subgraph, and the learner's own attribute subgraph to obtain learner feature representations in the three subgraphs. and Furthermore, an attention mechanism is used during the aggregation process to distinguish the importance of different neighboring entities to the target learner. Finally, the three are aggregated and assigned different weights to obtain the learner's final feature representation L. i .

[0158] The learning objective module modeling of the e-learning recommendation system of this invention is as follows: Figure 3 As shown, in the learning target module, this invention aggregates the learning target subgraph and the learning target feature subgraph to obtain the feature representation of the learning target in the two subgraphs. and Finally, the two are aggregated to obtain the final feature representation K of the learning objective. i .

[0159] The learning objective subgraph module modeling of the e-learning recommendation system of this invention is as follows: Figure 4 As shown, the embedding representation of the learning target is obtained by recursively aggregating k-hop entities related to the learning target. Entities in the ripple set are considered as extensions of the learning target, and the similarity between the learning target and entities in the ripple set is calculated. A first-order corresponding vector is obtained through weighted summation, and second-order, third-order, and other embeddings of the learning target are obtained through multiple iterations until the maximum hop count H is reached, resulting in the final embedding of the learning target.

[0160] Figure 5This is a graph showing the impact of different variant experiments of the electronic learning recommendation system of this invention on the system evaluation index on dataset one; Figure 6 This is a graph showing the impact of different variant experiments of the electronic learning recommendation system of this invention on the system evaluation index on dataset 2; Figure 7 This is a graph showing the impact of different variant experiments of the e-learning recommendation system of this invention on the system evaluation index on dataset 3; the impact of different variant experiments of the e-learning recommendation system of this invention on the system evaluation index on the dataset is as follows. Figure 5 , Figure 6 and Figure 7 As shown, this invention discusses the impact of different variant experiments on Learning Target Recommendation (LORec). One variant, LOREC-KG, removes the knowledge graph as auxiliary information. This variant uses the original learner and learning target embedding vectors q. i and p i To predict the probability of interaction between the two. The variant LOREC-Cluster removes the learner cluster. Aggregation and Then calculate the feature representation obtained by aggregating the two and K. i The probability of interaction. The variant LOREC-T removes the factor of learner time invested in the learning objective, i.e., in aggregation. The contribution of each learning objective to the learner is treated equally. Variant experiments were conducted on three datasets, and the results are as follows: Figure 5 , Figure 6 and Figure 7 As shown, the horizontal axis represents the name of the variant experiment, and the vertical axis represents the evaluation index of the model. The left and right graphs are the evaluation indices RMSE and MAE, respectively. From Figure 5 , Figure 6 and Figure 7 As can be seen, LORec-KG performs the worst, indicating that using knowledge graphs as an auxiliary means can enhance the performance of recommendation systems. LORec consistently maintains the best performance among the three datasets, suggesting that considering learner clusters and the time learners spend on learning objectives helps improve model metrics.

[0161] Does the e-learning recommendation system of this invention consider the impact of learner attributes on the system evaluation indicators across three data points? Figure 8 As shown, this invention discusses the impact of whether and how learner attributes are considered on recommendation effectiveness in educational scenarios. The model of this invention is compared with previous research. Ripple Net does not consider user attributes. While Social-Ripple Net considers user attributes, it treats the impact of each attribute on the target user equally when aggregating users. Figure 8This diagram illustrates the impact of considering learner attributes and treating each attribute equally on recommendation performance in an educational setting. The horizontal axis represents the dataset name, and the vertical axis represents the evaluation metrics. The left and right graphs show the evaluation metrics RMSE and MAE, respectively. The diagram shows that RMSE and MAE decrease progressively when learner attributes are not considered, when learner attributes are treated equally, and when learner attributes are treated differently. This indicates that in educational settings, differentiating learner attributes is essential for recommendation performance and can significantly improve it. Furthermore, as the dataset size increases, the RMSE and MAE values ​​decrease further in all three cases. This demonstrates that expanding the dataset size provides the model with more information, thereby further improving recommendation accuracy.

[0162] The impact of different feature weights in the learner modeling module of the e-learning recommendation system of this invention on the system evaluation index is as follows: Figure 9 As shown, this invention discusses the impact of different subgraph feature weights on the model in the learner modeling module. The results are as follows... Figure 9 As shown in the figure, the horizontal axis represents different weight parameters, and the vertical axis represents the evaluation metrics of the model. The left and right figures are the evaluation metrics RMSE and MAE, respectively. The figures show that, firstly, different values ​​of ω1, ω2, and ω3 all outperform the values ​​of ω1, ω2, and ω3 (1, 1, 1), indicating that the feature weights distinguishing the three subgraphs affect the model. Secondly, the model reaches its optimal state when the values ​​of ω1, ω2, and ω3 are 0.5, 0.2, and 0.3, indicating that in educational scenarios, learner attribute features and learning clusters improve recommendation performance to some extent. Finally, minimizing the weights of the learner and learning target subgraphs reduces the model's performance, suggesting that while considering learner attribute features and learning clusters improves model performance, the interaction feature weights between learners and learning targets should not be too low. Furthermore, the figures also show that the metrics significantly improve with increasing dataset size, indicating that increasing the data sample size can improve recommendation performance.

[0163] The impact of different regularization parameters of the e-learning recommendation system of this invention on the system evaluation index during system optimization is as follows: Figure 10 As shown, this invention discusses the impact of different regularization parameters λ on the two indices RMSE and MAE, and the results are as follows. Figure 10As shown in the figure, the horizontal axis represents different regularization parameters, and the vertical axis represents the evaluation metrics of the model. The left and right plots are the evaluation metrics RMSE and MAE, respectively. The figures show that datasets Dataset-1, Dataset-2, and Dataset-3 achieved optimal performance at λ = 10^-5, 10^-3, and 10^-7, respectively. Theoretically, smaller datasets require larger regularization parameters to suppress overfitting. However, because the learning rate for Dataset-1 was set to 0.1, a larger learning rate would lead to larger changes in parameters with each update, requiring a smaller regularization parameter to reduce model complexity and the possibility of overfitting. This demonstrates that even small changes in parameters can significantly alter the model's evaluation metrics. Datasets Dataset-2 and Dataset-3, with learning rates set to 0.001, achieved optimal performance at λ = 10^-3 and λ = 10^-7, respectively. In fact, smaller regularization parameters may not provide sufficient regularization constraints, while setting excessively large regularization parameters may distort the model's loss function. Finding a balance between preventing overfitting and avoiding excessive penalty to model complexity is crucial. This balance is reflected in the selection of the optimal regularization parameter for achieving the best performance.

Claims

1. A design method for an e-learning recommendation system based on multiple knowledge subgraphs, characterized in that: This paper designs an e-learning recommendation system using knowledge graphs as an auxiliary information source. The e-learning recommendation system based on multiple knowledge subgraphs includes a learner module, a learning objective module, and a rating module. The learner module combines a learner-learning-goal interaction subgraph, a learner cluster subgraph, and a learner-attribute subgraph, with the aim of creating a profile of the learner. The learning objective module combines the learning objective feature subgraph and the learning objective subgraph to mine semantic information between learning objectives; the scoring module calculates each learner's score for the learning objective and recommends high-scoring learning objectives to learners. The learner-learning-objective interaction subgraph, learner cluster subgraph, and learner self-attribute subgraph respectively yield feature representations of the learner in the three subgraphs. and By giving and Different weights are assigned to represent the importance of the three features to the final learner feature representation, with the aim of obtaining the learner feature representation L that performs best in the recommender system. i : Where σ is a nonlinear activation function, and ω1, ω2, and ω3 are respectively... and weights, This represents vector addition, where b is the bias. and These are the learner feature representations obtained from the learning objective interaction subgraph, the learner cluster subgraph, and the learner's own attribute subgraph, respectively. learner characteristics By aggregating learners i The aggregation function, derived from the set of learning objectives C(i) he has studied, is defined as follows: Where σ is a nonlinear activation function, w and b are weight matrices and biases, respectively; C(i) is the set of learning objectives learned by the learner, and p j These are the embedding vectors of these learning objectives; by assigning different weights to each learning objective to represent the importance of each historical learning objective to the learner, the aim is to better simulate the learner's school preferences: Where σ is a nonlinear activation function, w and b are the weight matrix and bias, respectively; p j These are the embedding vectors of the learning objectives, α ij It refers to the attention weighting of learning objectives, which aims to distinguish the importance of different learning objectives to learners. ij Calculated by the neural attention network: Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q j p is the learner's embedding vector. j These are the embedding vectors of these learning objectives, where ⊕ represents vector addition. To capture the contribution of each learning objective to the target user, the attention score is normalized using the softmax function. in, Indicates to Indexation; The learning objective feature subgraph and the learning objective subgraph respectively obtain the feature representation of the learning objective in the two subgraphs. and By aggregating the two, we obtain the final learning target embedding representation K. i : in, It is the feature representation of the learning objective obtained from the learning objective subgraph. The learning objective features are obtained from the feature subgraph of the learning objective. Represents vector addition; The scoring module uses learner features L from formula (1) i The learning objective feature K of formula (14) i The data is then concatenated, and the interaction score between the two is calculated through multiple fully connected layers: Among them, L i and K i These are the final feature representations of the learner and the learning objective, respectively. The parameters involved in training are updated using the loss function (22): Where m is the number of samples and λ is the regularization parameter; The interaction score between the two is calculated through multiple fully connected layers. lk is the interaction score between the learner and the historical learning target, w is the model's weight parameter, and L is the number of fully connected layers. The optimal parameters are obtained by minimizing the loss function to calculate each learner's score for the learning target, and then the learning target with the highest score is recommended to the learner.

2. The design method for an e-learning recommendation system based on multiple knowledge subgraphs according to claim 1, characterized in that: learner characteristics By setting learner l i Let C(i) be the set of learning objectives learned, and let l be the learner. j Let C(j) be the set of learning objectives that have been learned; if l i and learners i If learners have shared a common learning objective, then their sets of learning objectives have an intersection, namely C(ij). Based on this concept, a knowledge graph of learning clusters is constructed, where the relationships between learners are represented by the number of elements in the intersection C(ij) of their shared learning objectives. S(i) is defined as learner l. i A set of similar learners, which is connected to learner l in the learning cluster subgraph. i The learner composition, defined as the learner feature representation obtained by aggregating learning cluster subgraphs, is... Aggregate functions are defined as follows: Where σ is a nonlinear activation function, w and b are weight matrices and biases, and S(i) is the sum of the learning function and the bias of the learner. i A set of similar learners, q j These are the embedding vectors of these learners; by assigning different weights to each similar learner to represent the importance of each similar learner to the learner, the aim is to better simulate learners' school preferences: Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and β is the bias. ij It represents the attention weights of similar learners, q j These are the embedding vectors of similar learners; In distinguishing the importance of different similar learners to learners, β ij Calculated by the neural attention network: Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively; q i It is the embedding vector of the target learner, q j These are the embedding vectors of similar learners. This represents vector addition; to capture the contribution of each similar learner to the target learner, the attention score is normalized using the softmax function. in, Indicates to To index.

3. The design method for an e-learning recommendation system based on multiple knowledge subgraphs according to claim 1, characterized in that: learner characteristics O(i) represents learner l i The feature set is used to obtain learner feature representations by aggregating learners' age, knowledge level, and professional characteristics. Aggregate functions are defined as follows: Where σ is a nonlinear activation function, w and b are weight matrices and biases, and O(i) is the sum of the learning thresholds and the learning thresholds. i Feature set, t j It is the embedding vector of the learner's own features; by assigning different weights to each feature to represent the importance of each feature to the learner, the aim is to better simulate the learner's school preferences: Where σ is the nonlinear activation function, w and b are the weight matrix and bias, respectively, and λ is the weight matrix and bias. ij It is the learner's own characteristic attention weight, t j It is the embedding vector of the learner's own features, designed to distinguish the importance of different features to the target learner, λ ij Calculated by the neural attention network: Where σ is a nonlinear activation function, w1, w2 and b1, b2 are weight matrices and biases, respectively, and q is a nonlinear activation function. i It is the embedding vector of the target learner, t j It is the embedding vector of the learner's own features. This represents vector addition; to capture the contribution of each learner's own features to the target learner, the softmax function is used to normalize the attention score. in, Indicates to To index.

4. The design method for an e-learning recommendation system based on multiple knowledge subgraphs according to claim 1, characterized in that: The learning objective features This is achieved by representing the learning target k as N(i). i The set of attributes, through aggregation, learns the target k. i The learning target k is obtained from the various attributes. i Feature representation Aggregate functions are defined as follows: Where σ is a nonlinear activation function, w and b are the weight matrix and bias, and N(i) is the learning target k. i The set of attributes, n j It is the embedding vector of the learning target attribute.

5. The design method for an e-learning recommendation system based on multiple knowledge subgraphs according to claim 1, characterized in that: The learning objective features It is achieved by recursively aggregating and learning the target k. i The relevant k-hop entities are used to ultimately obtain the learning target k. i The embedding representation; specifically: the learning target k in the learning target subgraph i The set of k-hop entities is defined as: When k = 0 This represents the learning target node subgraph at this point, representing the learning target k. i The characteristic is its own embedding vector p. i H is the maximum number of hops k, G is the knowledge graph, and the set of related entities can be regarded as the learning target k in the knowledge node subgraph. i The extension; Learning objective k in the knowledge node subgraph i The k-jump ripple set is defined as starting from... The initial set of learning objective triples: Each learning objective k is embedded in the learning objective p. j ∈R d The related terms are: d is the embedding dimension; H is the maximum number of hops k; and G is the knowledge graph. Given learning objective p j And the set of triples affected by its k=1 jump. for Each triple (h) i ,r i ,t i ), by calculating the learning objective p j The relation r in this triple i Chinese h i Similarity P between them: Among them, R i ∈R d×d It is a relation r i The embedding vector, h i ∈R d It is the embedding of the head entity in the ripple set; P represents the embedding in the relation space R. i The selected learning objective p j and entity h i Similarity between them; after obtaining the similarity P, by analyzing the similarity between them... The tail entity in the vector is obtained by weighting: Among them, t i ∈R d It is the embedding of the tail entity t, vector regarded as a learning objective p j First-order response in the ripple set; P is the learning target p j The relation r in this triple i Chinese h i The similarity P between them; By using p in formula (18) j Replace with The learning objective p can be obtained through repeated propagation. j Second-order, third-order and H-order responses The learning target k can be obtained from the responses of all combinations. i The final embedding Where, vector regarded as a learning objective p j First-order response in the ripple set, vector regarded as a learning objective p j Second-order response in a ripple set, vector regarded as a learning objective p j H-order response in the ripple set.

Images