Electronic education resource recommendation method and system based on knowledge graph feature extraction

Abstract

The invention discloses an electronic education resource recommendation method and system based on knowledge graph feature extraction. The method comprises the following steps: representing entity and relational data as a knowledge graph structure to construct an electronic education resource knowledge graph; wherein the nodes represent entities, and the edges represent the relationship between the entities; determining a node embedding expression and a relation embedding expression of the electronic education resource knowledge graph; inputting the node embedding representation and the relation embedding representation into a neural network model to output deep embedding layer representation and interaction features; and processing the deep embedded layer representation and the interaction features by using a neural network model to extract a feature mode, and predicting whether an edge of a learning relationship exists between two nodes of a learner and an educational resource in the electronic educational resource knowledge graph by using the feature mode so as to recommend the electronic educational resource based on a prediction result. According to the method, the importance weight between the educational resources is dynamically adjusted, and finally prediction and personalized recommendation of the educational resources possibly interested by the learner in the future are achieved.

Description

Technical Field

[0001] The present invention relates to technical fields such as machine learning, natural language processing, knowledge graphs, and in particular to an electronic education resource recommendation method and system based on knowledge graph feature extraction. Background Art

[0002] Electronic educational resource recommendation refers to the use of machine learning, natural language processing, knowledge graphs and other technical means to recommend suitable learning materials such as courses, textbooks, videos, articles, exercises, etc. according to learners' interests, learning history, learning goals, subject areas and other factors, so that learners can obtain learning resources that meet their needs more efficiently. At present, more and more online learning platforms are widely used. Learners will feel confused and lost when faced with a large number of educational resources. They don't know how to choose the most appropriate learning materials to achieve their learning goals. Therefore, there is an urgent need for an electronic educational resource recommendation system to solve this difficulty. The diversity and complexity of electronic educational resources make them present heterogeneous forms, and there are multiple relationships between them. These heterogeneous and multi-relational electronic educational resources are intertwined with other entities involved in the learning process and the learners themselves, forming an intricate heterogeneous network that contains rich implicit information. Therefore, the core problem faced by electronic educational resource recommendation is how to effectively mine and integrate a large amount of knowledge based on this heterogeneous knowledge graph to provide learners with personalized and accurate electronic educational resource recommendation services. In recent years, studies have shown that when dealing with graph data with complex relationships and structures, recommendation systems using graph neural networks can better capture the characteristics and associations of data, and improve the performance of recommendation systems compared to other methods [Shang C, 2019]. However, the recommendation models currently proposed do not capture the complex relationships between nodes in the knowledge graph of electronic educational resources very well. The neural network architecture currently used for recommendation systems usually includes simple fully connected layers and operations such as convolution, dot product, and outer product [Yang Y, 2020].

[0003] The core idea of ​​electronic educational resource recommendation is to analyze learners' historical learning behaviors and interest models, combine rich knowledge graphs, and use technical means such as feature extraction, prediction models, and recommendation generation to accurately predict and recommend educational resources that learners may be interested in in the future, thereby optimizing the learning experience and improving learning outcomes. Previous electronic educational resource recommendation methods, when faced with a large number of educational resources from different sources, in different formats and granularities, were unable to fully consider the complex associations and hierarchical structures between educational resources, as well as the impact of different entities and relationship diversity on learners when selecting educational resources. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the present invention proposes an electronic education resource recommendation method based on knowledge graph feature extraction. In order to better capture the implicit interaction behavior between users and recommended items, the present invention adopts an operation-aware network based on multi-scale features to generate recommendations, which can mine the multiple relationships of nodes in the electronic education resource knowledge graph, thereby achieving more accurate electronic education resource recommendations.

[0006] Another object of the present invention is to propose an electronic educational resource recommendation system based on knowledge graph feature extraction.

[0007] To achieve the above-mentioned purpose, the present invention proposes, on one hand, a method for recommending electronic educational resources based on knowledge graph feature extraction, comprising:

[0008] Representing entity and relationship data into a knowledge graph structure to construct an electronic education resource knowledge graph; where nodes represent entities and edges represent relationships between entities;

[0009] Determining a node embedding representation and a relationship embedding representation of the electronic educational resource knowledge graph;

[0010] Input the node embedding representation and the relationship embedding representation into the neural network model to output the deep embedding layer representation and the interaction features;

[0011] A neural network model is used to process the deep embedding layer representation and interaction features to extract feature patterns, and the feature patterns are used to predict whether there is an edge of learning relationship between the two nodes of learners and educational resources in the electronic educational resource knowledge graph, so as to recommend electronic educational resources based on the prediction results.

[0012] The electronic education resource recommendation method based on knowledge graph feature extraction in the embodiment of the present invention may also have the following additional technical features:

[0013] In one embodiment of the present invention, the entities in the electronic educational resource knowledge graph include learners, educational resources and other descriptive entities; the various association relationships in the electronic educational resource knowledge graph include the relationship between learners and educational resources, the relationship between educational resources and other descriptive entities, and the relationship between learners and other descriptive entities.

[0014] In one embodiment of the present invention, the electronic educational resource knowledge graph is defined as Among them, v represents the set of all entities in the graph, Represents the set of all relationships. The edge (u, v, r) in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, describing the nodes and has dimension d0 The vector features of Represents; describes a relationship and has a dimension of d 0 The vector features of Denote; define β' and To represent the reverse edge set and inverse relationship set of the two mutually reverse edges and relationships between the established entities; the reverse edge set β' is the union of the original edge set β, the set of inverse relationships of the edges, and any relationship set between the node u and itself; where the inverse relationship set of the edge means that if there is an edge (u, v, r), then the inverse relationship (u, v, r) is added -1 ) set; inverse relation set yes The inverse relation r of every relation r in -1 Join The union of the new set formed in the graph and the set of relationships between the nodes in the graph and the nodes themselves.

[0015] In one embodiment of the present invention, the d of the relationship and entity 0 dimensional vector features and Representation; the embedding vectors of the head node, tail node and relationship are represented as e v ,

[0016] e u and e r , non-parameter combination operator use Indicates that, through the corresponding position of the multiplication operation, the tail node embedding vector e u The calculation process is The set of direct neighbors of node v is Represents that the trainable parameters of relation-awareness are expressed as Represented as follows, where λ(r) is used to model the importance of different relations r; the attention weights of node and relation embeddings are represented by α u,v express: Among them, k represents the number of network layers, a is the weight vector of the dot product operation, and the linear parameters that can be learned by the model are denoted by W l Indicates that The characteristics of node u and relationship r are q u ,q r indicates; when When W λ(r) =W O ;when When W λ(r) =W I ; When the relation r belongs to the relation between nodes themselves, W λ(r) =Ws ; The embedding representation h of node v v and the relation embedding representation h r It is expressed as:

[0017]

[0018] h r =W'q r

[0019] The embedding representation of node v is updated by weighted summing of the feature representations of its neighbor nodes u and the relationship r between them, where the influence of each neighbor node is determined by the weight α u,v Decision: The transformation matrix W′ in the relation embedding representation maps the representation of the relation from the relation space to the node embedding space.

[0020] In one embodiment of the present invention, the method further includes:

[0021] use as well as They represent the output d of the collaborative graph convolutional network based on the attention mechanism. 0 dimensional embedding representation of learners, educational resources, and learning relationships; if the operation perception network based on multi-scale features as well as Each dimension of After performing l operations, the embedding matrix of the i-th feature is V i Represented as, then the deep embedding layer is represented as:

[0022]

[0023] Define the o operation as the nth operation performed on the ith feature as Generate interaction-based features for the network layer, representing the operation between the i-th and j-th features, the copy operation and the inner product operation, represented by o(c,i) and o(p,i,j) respectively, then the output of this network layer is It is composed of deep embedding-based features and interaction-based features; the deep embedding-based features and interaction-based features are expressed as:

[0024]

[0025] The output of the neural network model is represented as:

[0026]

[0027] in, and b n+1 represents the learnable parameters of the model, L nrepresents the output of the nth hidden layer, L n The value of is determined by first The value of is used as the parameter of the rectified linear unit function to calculate the value, and then the value is used as the parameter of the batch normalization function to calculate the value; when n = 1, at this time Represents the output of the previous module;

[0028] Loss Function The regularization term uses l 2 represents the 2-norm, and the loss function is defined as:

[0029]

[0030] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes an electronic education resource recommendation system based on knowledge graph feature extraction, comprising:

[0031] A knowledge graph construction module is used to represent entity and relationship data into a knowledge graph structure to construct an electronic education resource knowledge graph; wherein nodes represent entities and edges represent relationships between entities;

[0032] An embedded representation module, used to determine the node embedding representation and relationship embedding representation of the electronic educational resource knowledge graph;

[0033] The network model design module is used to input the node embedding representation and the relationship embedding representation into the neural network model to output the deep embedding layer representation and the interaction feature; and use the neural network model to process the deep embedding layer representation and the interaction feature to extract the feature pattern, and use the feature pattern to predict whether there is an edge of the learning relationship between the two nodes of the learner and the educational resource in the electronic educational resource knowledge graph, so as to recommend the electronic educational resources based on the prediction results.

[0034] The electronic educational resource recommendation method and system based on knowledge graph feature extraction of the embodiments of the present invention dynamically adjusts the importance weights between educational resources based on the learner's historical learning behavior, interest model and the association between educational resources in the knowledge graph on a graph neural network model, and ultimately achieves prediction and personalized recommendation of educational resources that the learner may be interested in in the future.

[0035] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0037] Figure 1It is a flowchart of an electronic educational resource recommendation method based on knowledge graph feature extraction according to an embodiment of the present invention.

[0038] Figure 2 It is a schematic diagram of various association relationships between entities in an educational resource graph according to an embodiment of the present invention.

[0039] Figure 3 This is an example diagram of some entity and relationship categories in the educational resource graph according to an embodiment of the present invention.

[0040] Figure 4 It is a structural diagram of an electronic educational resource recommendation system based on knowledge graph feature extraction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0043] The following describes an electronic education resource recommendation method and system based on knowledge graph feature extraction proposed according to an embodiment of the present invention with reference to the accompanying drawings.

[0044] Figure 1 It is a flowchart of an electronic educational resource recommendation method based on knowledge graph feature extraction according to an embodiment of the present invention.

[0045] First, the problem to be solved by the present invention can be formally defined as follows:

[0046] After constructing the educational resource map, it is defined as Where v represents the set of all entities in the graph, Represents the set of all relationships. The edge 9u, v, r in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, describing the nodes and has dimension d 0 The vector features of Describes the relationship and has a dimension of d 0 The vector features of express.

[0047] The problem of electronic educational resource prediction is to construct a method g, and then input two node vectors into g to let f predict whether there is an edge between the nodes corresponding to the two vectors, that is, there is a "learning" relationship between the learner and the electronic educational resource.

[0048] like Figure 1 As shown, the method includes:

[0049] S1, represents entity and relationship data into a knowledge graph structure to construct an electronic education resource knowledge graph; where nodes represent entities and edges represent relationships between entities.

[0050] For example, taking multiple associations between entities in an educational resource graph as an example, Figure 2 The left side of shows that the educational resource graph is composed of several nodes and edges. Each node represents a different entity, and the edge linking from one node to another node represents the relationship between entities. Figure 2 The right side of the diagram shows a zoomed-in representation of a part of the graph, which displays the relationships among learners, teachers, schools, videos, and documents.

[0051] Specifically, the entities in the electronic educational resource knowledge graph constructed by the present invention include learners (school students, platform users, etc.), educational resources (courses, videos, text materials, etc.) and other descriptive entities (schools, teachers, platforms, etc.). The various associations in the graph include not only the relationship between learners and educational resources, but also the relationship between educational resources and other descriptive entities and the relationship between learners and other descriptive entities, such as Figure 3 As shown. For example, there is an "opening" relationship between educational resources and other descriptive entities (the school has opened a course or produced a video); there is a "teaching" relationship between educational resources and other descriptive entities (the teacher has participated in the teaching of a course or the production of a video); there is a "collection" relationship between educational resources and learners (platform users have collected a course or video); there is an "interaction" relationship between learners and educational resources (learners have watched videos or learned courses and interacted with educational resources); there is a "studying" relationship between learners and other descriptive entities (learners are enrolled in a school and participate in the school's educational and teaching activities); there is a "teaching" relationship between other descriptive entities and learners (learners have participated in the study of the course taught by the teacher). Then, in order to more comprehensively describe the connectivity between entities, two mutually inverse relationships are established, which will be used in the next step to train the embedding expression model to better capture the semantic relationship and structural information between entities. Finally, the collected entity and relationship data are represented as a knowledge graph structure, in which nodes represent entities and edges represent relationships between entities.

[0052] It can be understood that according to the process of learners learning through electronic educational resources, the objectively existing entities and the multiple associations between entities are analyzed, the entire interactive system is described, and the knowledge graph of electronic educational resources is constructed.

[0053] S2, determine the node embedding representation and relationship embedding representation of the electronic education resource knowledge graph.

[0054] Specifically, according to the electronic education resource knowledge graph constructed in step S1, it is defined as Where V represents the set of all entities in the graph, Represents the set of all relationships. The edge (u, v, r) in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, describing the nodes and has dimension d 0 The vector features of Describes the relationship and has a dimension of d 0 The vector features of In order to cover the operation of any relationship, the present invention defines β' and To represent the reverse edge set and inverse relationship set of the two mutually reverse edges and relationships between the established entities. The reverse edge set β' is the union of the original edge set β, the set of inverse relationships of the edges, and any set of relationships between the node u and itself. The inverse relationship set of the edge means that if there is an edge (u, v, r), then the inverse relationship (u, v, r) is added. -1 ). Inverse relation set yes The inverse relation r of every relation r in -1 (the reverse relationship of the original relationship) join to The graph feature representation method proposed in the present invention can not only be used to train and learn the embedded representation of each node in the graph (i.e., represent each node as a low-dimensional vector), but also can be used to learn the representation of the relationship in the entire graph (i.e., represent each relationship in the graph as a vector), helping the model to effectively learn the association between nodes in the graph and the semantic association between different relationships in the graph. 0 dimensional vector features and By learning the embedding vectors of nodes and relationships, we can use them to perform various complex combination operations to generate new representations, so as to better understand the relationship between entities in the graph. Among them, the embedding vectors of the head node, tail node and relationship are represented as e v , e u and e r . Non-parameter combination operators use The present invention performs a positional multiplication operation at the corresponding position, and the tail node is embedded in the vector e u The calculation process is The present invention updates the node and relationship representation formula in the graph convolutional neural network, which can better consider the contribution of different neighbor nodes to the target node and solve the problem of updating node representation in the traditional graph convolutional neural network. Represented by . The trainable parameters of relation awareness are expressed as , where λ(r) is used to model the importance of different relations r, so that the model can better consider the differences of different relations when learning node representations. The attention weights of node and relationship embeddings are α u,v express: Where k represents the number of network layers, a is the weight vector of the dot product operation. The linear parameters that the model can learn are denoted by W l Indicates that The characteristics of node u and relationship r are q u ,q r Indicates. When W λ(r) =W O ;when When W λ(r) =W I ; When the relation r belongs to the relation between nodes themselves, W λ(r) =W S Therefore, the embedding of node v represents h v and the relation embedding representation h r It can be expressed as:

[0055]

[0056] h r =W'q r

[0057] The embedding representation of node v is updated by weighted summing of the feature representations of its neighbor nodes u and the relationship r between them, where the influence of each neighbor node is determined by the weight α u,v The transformation matrix W′ in the relation embedding representation can map the representation of the relation from the relation space to the node embedding space, so that the model can better learn the semantic association between nodes and relations. Thus, the embedded representation of nodes and relations output by any layer of the neural network can be derived.

[0058] It is understandable that the electronic educational resource graph and the nodes and edges it contains are defined and constructed by the present invention. The present invention considers that the relationship between nodes is a bidirectional relationship, so the forward and reverse edges and the forward and reverse relationships are defined respectively. The method proposed by the present invention is used to perform representation learning (node ​​embedding representation and relationship vector representation) on the electronic educational resource graph.

[0059] S3, input the node embedding representation and the relationship embedding representation into the neural network model to output the deep embedding layer representation and interaction features;

[0060] S4, uses a neural network model to process the deep embedding layer representation and interaction features to extract feature patterns, and uses the feature patterns to predict whether there is an edge of a learning relationship between the two nodes of learners and educational resources in the electronic educational resource knowledge graph, so as to recommend electronic educational resources based on the prediction results.

[0061] Specifically, the embedded representation of nodes and relationships obtained in the above step S2 is used as the input of the classification neural network. The classification neural network used in the present invention is a collaborative graph convolutional network based on the attention mechanism, which includes multiple modules, each of which is composed of one or more nonlinear layers. These nonlinear layers include fully connected layers, convolutional layers, pooling layers, etc., which are used to extract complex patterns and features in the input features, and finally use these features to predict the learner's preference or interest in the recommended educational resource items (predict whether there is an edge between the learner node and the educational resource node). The deep embedding layer used in step S3 has the characteristics of perceptual operation. It can recognize and adapt to different operations, better process input data, and thus have a deeper understanding of the learner's implicit behavior and preferences. The neural network model designed in step S3 can effectively process multi-scale features, capture the interactive behavior between users and recommended items in combination with different computing operations, and generate personalized recommendation results. Use as well as They represent the output d of the collaborative graph convolutional network based on the attention mechanism. 0 dimensional embedding representation of learners, educational resources, and “learning” relationships. as well as Each dimension of After performing l operations, the embedding matrix of the i-th feature is V i

[0062] Represented as, then the deep embedding layer can be expressed as:

[0063]

[0064] Define the o operation as the nth operation performed on the ith feature as Next, we generate interaction-based features for the network layer, which represent operations between the i-th and j-th features, such as “copy operation” and “inner product operation”, represented by o(c,i) and o(p,i,j) respectively. Then the output of this network layer is (consisting of deep embedding-based features and interaction-based features). Among them, the deep embedding-based features and interaction-based features are expressed as:

[0065]

[0066]

[0067] Then, the batch normalization mechanism is used in each layer of the neural network. By normalizing the input of each layer in each batch of data, and then scaling and translating it, the input distribution of each layer is made more stable, which helps alleviate the problems of gradient disappearance and gradient explosion, makes the network more stably trained, and also helps prevent overfitting and improve the generalization ability of the model.

[0068] The output of the entire network model can be expressed as:

[0069]

[0070] in, and b n+1 Represents the learnable parameters of the model. n represents the output of the nth hidden layer,

[0071] L n The value of is determined by first The value of is used as the parameter of the rectified linear unit function to calculate the value, and then the value is used as the parameter of the batch normalization function to calculate the value. When n = 1, at this time Represents the output of the previous module.

[0072] Finally, the model parameters are continuously adjusted through the optimization process, and the loss function is used to measure the gap between the model prediction result and the actual label (the smaller the loss function value, the more accurate the prediction result), so as to solve the loss function. The purpose of minimization. Loss function The regularization term uses l 2 Representing the 2-norm (also known as the Euclidean norm), it can constrain the parameters of the entire neural network (input layer, hidden layer, output layer, etc.), reduce the risk of the model overfitting the training data, and improve the generalization ability of the model. Therefore, the loss function is defined as:

[0073]

[0074] It can be understood that the present invention uses the embedded representation of nodes and relationships obtained above as the input of the classification neural network to obtain the representation of the deep embedding layer and the output of the interaction features, and then the recommendation network model uses a module composed of multiple nonlinear layers (ReLU, Sigmoid, Tanh, etc.) to process the input data (the representation and interaction features of the deep embedding layer) to extract complex feature patterns, and use these patterns to predict whether there is an edge (indicating a "learning" relationship) between the two nodes of the learner and the educational resource in the educational resource knowledge graph, thereby realizing the recommendation of electronic educational resources.

[0075] Furthermore, the experiment selected the MOOCCube dataset (5 entities and 6 relationships) collected by XuetangX and the dataset (7 entities and 8 relationships) in a university's online course platform to construct an electronic resource knowledge graph, including two types of educational resources (videos, documents). The knowledge graph used in the experiment contains 3,100 nodes and 42,200 edges. The experiments of the model proposed in the present invention and the comparative model were conducted on a server with an AMD Ryzen 9 5950X CPU, 64GB DDR4 RAM memory and 4 Nvidia GeForce RTX 3080. The positive sample in the experiment refers to the situation where the learner has learned specific learning materials, while the negative sample refers to the situation where the learner has not learned specific learning materials. Finally, the experiment randomly sampled 180,500 positive samples and 3,500 negative samples, which were randomly assigned to the test set based on the MOOCCube dataset and the test set of a certain university online course platform dataset (the ratio of positive and negative samples in the test set was kept at 1:1). For each iteration, the experiment randomly extracted the validation set from the training set at a ratio of 30%:70% for model training. The experimental results are shown in Table 1. The present invention achieved the best performance in both accuracy and recall in the two datasets. In the MOOCCube dataset, the present invention improved the accuracy and recall by 1.2% compared with the suboptimal model. In the dataset of a certain university online course platform, the present invention improved the accuracy and recall by 3.5% and 2.3% respectively compared with the suboptimal model.

[0076] Table 1

[0077]

[0078] According to the electronic education resource recommendation method based on knowledge graph feature extraction according to the embodiment of the present invention, the prediction of learners' demand for electronic education resources is realized. The basic idea is to construct a knowledge graph and model it, and then obtain a unified embedded representation of the knowledge graph through a graph neural network model, so that the unified embedded representation is used as the input of the classification model to obtain a deep embedded layer representation and interactive features, and then input into the recommendation network model, and finally realize the prediction of electronic education resources. In this way, the multi-relationships of nodes in the electronic education resource knowledge graph can be mined, so as to achieve more accurate electronic education resource recommendations.

[0079] In order to implement the above embodiment, Figure 4 As shown, this embodiment also provides an electronic educational resource recommendation system 10 based on knowledge graph feature extraction, including:

[0080] The knowledge graph construction module 100 is used to represent entity and relationship data into a knowledge graph structure to construct an electronic education resource knowledge graph; wherein nodes represent entities and edges represent relationships between entities;

[0081] The embedded representation module 200 is used to determine the node embedded representation and relationship embedded representation of the electronic educational resource knowledge graph;

[0082] The network model design module 300 is used to input the node embedding representation and the relationship embedding representation into the neural network model to output the deep embedding layer representation and the interaction features; and use the neural network model to process the deep embedding layer representation and the interaction features to extract feature patterns, and use the feature patterns to predict whether there is an edge of the learning relationship between the two nodes of the learner and the educational resource in the electronic educational resource knowledge graph, so as to recommend electronic educational resources based on the prediction results.

[0083] Furthermore, the entities in the electronic educational resource knowledge graph include learners, educational resources and other descriptive entities; the various association relationships in the electronic educational resource knowledge graph include the relationship between learners and educational resources, the relationship between educational resources and other descriptive entities, and the relationship between learners and other descriptive entities.

[0084] Furthermore, the embedded representation module 200 is also used for:

[0085] The electronic educational resource knowledge graph is defined as Among them, v represents the set of all entities in the graph, Represents the set of all relationships. The edge (u, v, r) in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, describing the nodes and has dimension d 0 The vector features of Represents; describes a relationship and has a dimension of d 0 The vector features of Denote; define β' and To represent the reverse edge set and inverse relationship set of the two mutually reverse edges and relationships between the established entities; the reverse edge set β' is the union of the original edge set β, the set of inverse relationships of the edges, and any relationship set between the node u and itself; where the inverse relationship set of the edge means that if there is an edge (u, v, r), then the inverse relationship (u, v, r) is added -1 ) set; inverse relation set yes The inverse relation r of every relation r in -1 Join The union of the new set formed in the graph and the set of relationships between the nodes in the graph and the nodes themselves.

[0086] Furthermore, the embedded representation module 200 is also used for:

[0087] d of relations and entities 0 dimensional vector features and Representation; the embedding vectors of the head node, tail node and relationship are represented as e v , e u and e r , non-parameter combination operator use Indicates that, through the corresponding position of the multiplication operation, the tail node embedding vector e u The calculation process is The set of direct neighbors of node v is Representation; relation-aware trainable parameters are expressed as Represented as follows, where λ(r) is used to model the importance of different relations r; the attention weights of node and relation embeddings are represented by α u,v express: Among them, k represents the number of network layers, a is the weight vector of the dot product operation, and the linear parameters that can be learned by the model are denoted by W l Indicates that The characteristics of node u and relationship r are q u ,q r indicates; when When W λ(r) =W O ;when When W λ(r) =W I ; When the relation r belongs to the relation between nodes themselves, W λ(r) =W S ; The embedding representation h of node v v and the relation embedding representation h r It is expressed as:

[0088]

[0089] h r =W'q r

[0090] The embedding representation of node v is updated by weighted summing of the feature representations of its neighbor nodes u and the relationship r between them, where the influence of each neighbor node is determined by the weight α u,v Decision: The transformation matrix W′ in the relation embedding representation maps the representation of the relation from the relation space to the node embedding space.

[0091] Furthermore, the network model design module 300 is also used for:

[0092] use as well as They represent the output d of the collaborative graph convolutional network based on the attention mechanism. 0 dimensional embedding representation of learners, educational resources, and learning relationships; if the operation perception network based on multi-scale features as well as Each dimension of After performing l operations, the embedding matrix of the i-th feature is V i Represented as, then the deep embedding layer is represented as:

[0093]

[0094] Define the o operation as the nth operation performed on the ith feature as Generate interaction-based features for the network layer, representing the operation between the i-th and j-th features, the copy operation and the inner product operation, represented by o(c,i) and o(p,i,j) respectively, then the output of this network layer is It is composed of deep embedding-based features and interaction-based features; the deep embedding-based features and interaction-based features are expressed as:

[0095]

[0096] The output of the neural network model is represented as:

[0097]

[0098] in, and b n+1 represents the learnable parameters of the model, L n represents the output of the nth hidden layer,

[0099] L n The value of is determined by first The value of is used as the parameter of the rectified linear unit function to calculate the value, and then the value is used as the parameter of the batch normalization function to calculate the value; when n = 1, at this time Represents the output of the previous module;

[0100] Loss Function The regularization term uses l 2 represents the 2-norm, and the loss function is defined as:

[0101]

[0102] According to the parameterless geometry-guided test-time adaptation system for point cloud scene segmentation according to an embodiment of the present invention, the present invention utilizes a parameterless geometry model to extract features. Since the geometry information can capture the underlying manifold information, it can represent the rough target domain distribution. The present invention uses it as an intermediate domain to make the source domain model features close to the geometric features through the feature distillation method, so as to reduce the difference with the target domain. Then, in order to further promote the migration of the source domain model to the target domain, the present invention proposes to adjust the pseudo-labels through soft voting perceived by geometric features and model features, and use pseudo-labels to supervise the learning of the model. Finally, compared with the classification task, point cloud scene segmentation has a larger class imbalance problem, which may lead to a decrease in model performance. In order to solve this problem, the present invention proposes an entropy minimization loss for class-balanced sampling, and selects a certain proportion of clean samples for different class balances to minimize entropy.

[0103] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0104] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

Claims

1. A method for recommending electronic educational resources based on knowledge graph feature extraction, characterized in that: include: Representing entity and relationship data into knowledge graph structures to construct the knowledge graph of electronic educational resources; The nodes represent entities and the edges represent the relationships between entities. Determining a node embedding representation and a relationship embedding representation of the electronic educational resource knowledge graph; Input the node embedding representation and the relationship embedding representation into the neural network model to output the deep embedding layer representation and the interaction features; A neural network model is used to process the deep embedding layer representation and interaction features to extract feature patterns, and the feature patterns are used to predict whether there is an edge of learning relationship between the two nodes of learners and educational resources in the electronic educational resource knowledge graph, so as to recommend electronic educational resources based on the prediction results.

2. The method according to claim 1, characterized in that The entities in the electronic educational resource knowledge graph include learners, educational resources and other descriptive entities; the various association relationships in the electronic educational resource knowledge graph include the relationship between learners and educational resources, the relationship between educational resources and other descriptive entities, and the relationship between learners and other descriptive entities.

3. The method according to claim 2, characterized in that The electronic educational resource knowledge graph is defined as in, represents the set of all entities in the graph, Represents the set of all relationships. The edge (u, v, r) in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, and the vector features describing the nodes with dimension d0 are denoted by Represents; the vector feature describing the relationship and with dimension d0 is used Representation; definition β ’ and To represent the reverse edge set and inverse relationship set of the two mutually reverse edges and relationships between the established entities; the reverse edge set β ’ is the union of the original edge set β, the set of inverse relations of the edges, and the set of any relations between node u and itself; where the set of inverse relations of the edges means that if there is an edge (u, v, r), then the inverse relation (u, v, r) is added. -1 ) set; inverse relation set yes The inverse relation r of every relation r in -1 Join The union of the new set formed in the graph and the set of relationships between the nodes in the graph and the nodes themselves.

4. The method according to claim 3, characterized in that The d0-dimensional vector features of relations and entities are and Representation; the embedding vectors of the head node, tail node and relationship are represented as e v , e u and e r , non-parameter combination operator use Indicates that, through the corresponding position of the multiplication operation, the tail node embedding vector e u The calculation process is The set of direct neighbors of node v is denoted by Representation; relation-aware trainable parameters are expressed as Represents, where the role of λ(r) is to model the importance of different relations r; The attention weights of node and relation embeddings are denoted by α u,v express: Among them, k represents the number of network layers, a is the weight vector of the dot product operation, and the linear parameters that can be learned by the model are denoted by W l Indicates that The characteristics of node u and relationship r are q u ,q r indicates; when When W λ(r) =W O ;when When W λ(r) =W I ; When the relation r belongs to the relation between nodes themselves, W λ(r) =W S ; The embedding representation h of node v v and the relation embedding representation h r It is expressed as: h r =W ′ q r The embedding representation of node v is updated by weighted summing of the feature representations of its neighbor nodes i and the relationship r between them, where the influence of each neighbor node is determined by the weight α u,v Determine; the transformation matrix W in the relation embedding representation ′ Map the representation of relations from the relation space to the node embedding space.

5. The method according to claim 4, characterized in that The method further comprises: use as well as They represent the embedding representation of learners, educational resources and learning relationships output by the collaborative graph convolutional network based on the attention mechanism in d0 dimension respectively; if the operation perception network based on multi-scale features is used to input as well as Each dimension of After performing l operations, the embedding matrix of the i-th feature is V i Represented as, then the deep embedding layer is represented as: Define the o operation as the nth operation performed on the ith feature as Generate interaction-based features for the network layer, representing the operation between the i-th and j-th features, the copy operation and the inner product operation, represented by o(c,i) and o(p,i,j) respectively, then the output of this network layer is It is composed of deep embedding-based features and interaction-based features; the deep embedding-based features and interaction-based features are expressed as: The output of the neural network model is represented as: in, and b n+1 represents the learnable parameters of the model, L n represents the output of the nth hidden layer, L n The value of is determined by first The value of is used as the parameter of the rectified linear unit function to calculate the value, and then the value is used as the parameter of the batch normalization function to calculate the value; when n = 1, at this time Represents the output of the previous module; Loss Function The regularization term uses l2 to represent the 2-norm, and the loss function is defined as:

6. An electronic educational resource recommendation system based on knowledge graph feature extraction, characterized in that: include: A knowledge graph building module, used to represent entity and relationship data into a knowledge graph structure to build an electronic education resource knowledge graph; The nodes represent entities and the edges represent the relationships between entities. An embedded representation module, used to determine the node embedding representation and relationship embedding representation of the electronic educational resource knowledge graph; A network model design module, which is used to input node embedding representation and relationship embedding representation into a neural network model to output deep embedding layer representation and interaction features; And use the neural network model to process the deep embedding layer representation and interaction features to extract feature patterns, and use the feature patterns to predict whether there is an edge of learning relationship between the two nodes of learner and educational resource in the electronic educational resource knowledge graph, so as to recommend electronic educational resources based on the prediction results.

7. The system according to claim 6, characterized in that The entities in the electronic educational resource knowledge graph include learners, educational resources and other descriptive entities; the various association relationships in the electronic educational resource knowledge graph include the relationship between learners and educational resources, the relationship between educational resources and other descriptive entities, and the relationship between learners and other descriptive entities.

8. The system according to claim 7, characterized in that The embedded representation module is further used for: The electronic educational resource knowledge graph is defined as in, represents the set of all entities in the graph, Represents the set of all relationships. The edge (u, v, r) in the graph indicates that there is a relationship from the first node u to the last node v. The set of edges is denoted by β, and the vector features describing the nodes with dimension d0 are denoted by Represents; the vector feature describing the relationship and with dimension d0 is used Representation; definition β ’ and To represent the reverse edge set and inverse relationship set of the two mutually reverse edges and relationships between the established entities; the reverse edge set β ’ is the union of the original edge set β, the set of inverse relations of the edges, and the set of any relations between node u and itself; where the set of inverse relations of the edges means that if there is an edge (u, v, r), then the inverse relation (u, v, r) is added. -1 ) set; inverse relation set yes The inverse relation r of every relation r in -1 Join The union of the new set formed in the graph and the set of relationships between the nodes in the graph and the nodes themselves.

9. The system according to claim 8, characterized in that The embedded representation module is also used for: The d0-dimensional vector features of relations and entities are and Representation; the embedding vectors of the head node, tail node and relationship are represented as e v , e u and e r , non-parameter combination operator use Indicates that, through the corresponding position of the multiplication operation, the tail node embedding vector e u The calculation process is The set of direct neighbors of node v is Representation; relation-aware trainable parameters are expressed as Represents, where the role of λ(r) is to model the importance of different relations r; The attention weights of node and relation embeddings are denoted by α u,v express: Among them, k represents the number of network layers, a is the weight vector of the dot product operation, and the linear parameters that can be learned by the model are denoted by W l Indicates that The characteristics of node u and relationship r are q u ,q r indicates; when When W λ(r) =W o ;when When W λ(r) =W I ; When the relation r belongs to the relation between nodes themselves, W λ(r) =W S ; The embedding representation h of node v v and the relation embedding representation h r It is expressed as: h r =W ′ q r The embedding representation of node v is updated by weighted summing of the feature representations of its neighbor nodes u and the relationship r between them, where the influence of each neighbor node is determined by the weight α u,v Determine; the transformation matrix W in the relation embedding representation ′ Map the representation of relations from the relation space to the node embedding space.

10. The system according to claim 9, characterized in that The network model design module is also used for: use as well as They represent the embedding representation of learners, educational resources and learning relationships output by the collaborative graph convolutional network based on the attention mechanism in d0 dimension respectively; if the operation perception network based on multi-scale features is used to input as well as Each dimension of After performing l operations, the embedding matrix of the i-th feature is V i Represented as, then the deep embedding layer is represented as: Define the o operation as the nth operation performed on the ith feature as Generate interaction-based features for the network layer, representing the operation between the i-th and j-th features, the copy operation and the inner product operation, represented by o(c,i) and o(p,i,j) respectively, then the output of this network layer is It is composed of deep embedding-based features and interaction-based features; the deep embedding-based features and interaction-based features are expressed as: The output of the neural network model is represented as: in, and b n+1 represents the learnable parameters of the model, L n represents the output of the nth hidden layer, L n The value of is determined by first The value of is used as the parameter of the rectified linear unit function to calculate the value, and then the value is used as the parameter of the batch normalization function to calculate the value; when n = 1, at this time Represents the output of the previous module; Loss Function The regularization term uses l2 to represent the 2-norm, and the loss function is defined as:

Images