Fusion education priori knowledge double hypergraph contrast learning course recommendation method, device and equipment and storage medium

Abstract

The application discloses a double hypergraph contrast learning course recommendation method and device fusing education prior knowledge, equipment and a storage medium, and the method comprises the steps of: performing hypergraph aggregation on a user hypergraph and a course hypergraph, performing education feature fusion based on a node feature matrix after aggregation, obtaining hypergraph embedding features, performing contrast learning on interactive graph embedding features and hypergraph embedding features from the user and course levels, constructing a multi-objective optimization loss function based on a user contrast loss term and a course contrast loss term, performing education course recommendation through the target optimization loss function, and encoding the interactive graph through multi-layer graph convolution, so that multi-dimensional features and time evolution rules of learner behavior can be accurately captured, the education rationality of recommendation can be ensured through double-level feature contrast learning, potential learning needs and knowledge development tracks of learners can be identified, personalized needs and education rules can be balanced through a multi-objective optimization framework, and more accurate and more education-essence-compliant course recommendation can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph, and in particular to a dual-hypergraph contrastive learning course recommendation method and device fusing education prior knowledge, equipment and storage medium. BACKGROUND

[0002] With the rapid development of online education platforms, the number of courses on online education platforms is showing explosive growth, and learners are facing a serious information overload problem. As an important part of online education platforms, the education course recommendation system can help learners better choose courses suitable for themselves. Among them, the education recommendation system needs to provide personalized services while ensuring that the recommended results meet the education learning rules and knowledge acquisition logic, and the current course recommendation in the education field still has many technical challenges in this regard.

[0003] At present, the education recommendation system has obvious deficiencies in dealing with the particularity of the education scene, and it is difficult to effectively fuse the knowledge in the education field (such as course sequence relationship, difficulty level structure, etc.), resulting in that the recommended results have a certain personalization, but often ignore the educational rationality and the coherence of the learning path. At the same time, new registered learners and newly online courses lack sufficient interaction history, and existing methods are difficult to provide accurate personalized recommendations. SUMMARY

[0004] The main purpose of the present application is to provide a dual-hypergraph contrastive learning course recommendation method, device, equipment and storage medium fusing education prior knowledge, which aims to solve the technical problems that the prior art is difficult to effectively fuse the knowledge in the education field, ignores the educational rationality and the coherence of the learning path, and leads to unreasonable education course recommendation.

[0005] To achieve the above purpose, the present application provides a dual-hypergraph contrastive learning course recommendation method fusing education prior knowledge, which comprises the following steps:

[0006] The education data is preprocessed, and the preprocessed education data is analyzed for features to obtain multi-dimensional education features, the education data including learner data, course data, video data and concept data, and the multi-dimensional education features including interaction relationship features, user features and course features;

[0007] Based on the multi-dimensional education features, a target user hypergraph and a target course hypergraph are determined to be constructed;

[0008] Based on the interaction relationship features, a user-course interaction graph structure is constructed, and multi-layer graph convolution coding is performed on the user-course interaction graph structure to obtain interaction graph embedding features;

[0009] perform supergraph aggregation on the target user supergraph and the target course supergraph, and perform education feature fusion based on a node feature matrix after the aggregation to obtain supergraph embedding features;

[0010] respectively perform contrast learning on the interaction graph embedding features and the supergraph embedding features from the user level and the course level to obtain a user contrast loss term and a course contrast loss term;

[0011] construct a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term, and construct a course recommendation model based on the multi-objective optimization loss function, and perform education course recommendation through the course recommendation model.

[0012] Optionally, the education data is preprocessed, and feature analysis is performed on the preprocessed education data to obtain multi-dimensional education features, including:

[0013] The education data is subjected to data cleaning and standardization processing to obtain preprocessed education data;

[0014] The learning behavior data of learners in the preprocessed education data is analyzed to construct a weighted interaction matrix, and the weighted interaction matrix is used to represent interaction relationship features;

[0015] Based on the preprocessed education data, the interaction relationships between each user and course data and video data are analyzed to determine learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity between users, and the learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity are used to represent user features;

[0016] Based on the preprocessed education data, course content information is extracted, and content features of each course are generated based on the course content information;

[0017] Based on the content features, content similarity, difficulty similarity, structure similarity, and correlation similarity between courses are determined, and the content similarity, difficulty similarity, structure similarity, and correlation similarity are used to represent course features.

[0018] Optionally, the target user supergraph and the target course supergraph are constructed based on the multi-dimensional education features, including:

[0019] Based on the learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity, user similarity between users is determined;

[0020] An initial user supergraph is constructed based on the user similarity between users;

[0021] determine a course similarity between each course based on the content similarity, the difficulty similarity, the structure similarity and the correlation similarity, and construct an initial course hypergraph based on the course similarity;

[0022] perform sparsification processing on the initial user hypergraph and the initial course hypergraph to obtain a candidate user hypergraph and a candidate course hypergraph;

[0023] perform hypergraph normalization processing on the candidate user hypergraph and the candidate course hypergraph to obtain a target user hypergraph and a target course hypergraph.

[0024] Optionally, the user-course interaction graph structure is constructed based on the interaction relationship feature, and multi-layer graph convolution coding is performed on the user-course interaction graph structure to obtain an interaction graph embedding feature, including:

[0025] The user-course interaction graph structure is constructed based on the interaction relationship feature, and multi-layer graph convolution operation is performed on the user-course interaction graph structure to aggregate information through a normalized adjacency matrix, and a mathematical expression of each layer of graph convolution operation is:

[0026]

[0027]

[0028] wherein, is a normalized adjacency matrix, denotes an adjacency matrix, denotes a degree matrix, denotes an embedding representation of the i-th layer; The results of the multi-layer graph convolution operation are averaged and aggregated to obtain the interaction graph embedding feature, as shown in the following formula:

[0029]

[0030]

[0031] wherein, denotes the number of layers of graph convolution, denotes the interaction graph embedding feature.

[0032] Optionally, the target user hypergraph and the target course hypergraph are aggregated by hypergraph, and education feature fusion is performed based on the aggregated node feature matrix to obtain a hypergraph embedding feature, including:

[0033] The target user hypergraph and the target course hypergraph are aggregated by hypergraph by multiplying the hypergraph correlation matrix and the node feature, as shown in the following formula:

[0034] ​

[0035] wherein, denotes the aggregated node feature matrix, denotes the hypergraph association matrix, denotes the original node feature matrix;

[0036] Based on the aggregated node feature matrix, the educational feature fusion is performed to obtain an enhanced node feature matrix. Refer to the following formula:

[0037]

[0038] wherein, denotes the enhanced node feature matrix, denotes the aggregated node feature matrix, denotes element-level multiplication, denotes a sigmoid activation function, denotes an educational feature projection matrix, denotes an educational feature vector;

[0039] The multi-head attention mechanism is adopted to perform deep processing on the aggregated node feature matrix to obtain a transformed feature matrix.

[0040] Based on the transformed feature matrix and the enhanced node feature matrix, a gating weight vector is calculated, and the transformed feature matrix and the enhanced node feature matrix are gate fused based on the gating weight vector to obtain a hypergraph embedding feature.

[0041] Optionally, the interaction graph embedding features and the hypergraph embedding features are respectively compared and learned from the user level and the course level to obtain a user comparison loss term and a course comparison loss term, including:

[0042] The interaction graph embedding features and the hypergraph embedding features of the same user are compared and learned from the user level to construct a positive and negative sample pair of the user.

[0043] The positive and negative sample pair of the course is constructed from the course level based on the educational priori knowledge information.

[0044] The positive and negative sample pair of the user and the positive and negative sample pair of the course are input into an InfoNCE loss function to obtain the user comparison loss term and the course comparison loss term. Refer to the following formula:

[0045]

[0046] wherein, and denote a positive sample pair, denote a negative sample, is a temperature parameter for controlling the degree of differentiation, is a similarity function, is a number of negative samples.

[0047] Optionally, constructing a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term comprises:

[0048] constructing a before-after repair loss term, the before-after repair loss term referring to the following formula:

[0049]

[0050] wherein, represents a before-after repair relationship loss, represents a before-after repair relationship set, represents a predicted score of a user to a course , represents a predicted score of a user to a course , is a tolerance parameter;

[0051] constructing a difficulty loss term, the difficulty loss term referring to the following formula:

[0052]

[0053] wherein, represents a difficulty matching loss, represents a user set, represents a course set, represents a predicted score of a user to a course , represents a learning stage feature of a user , represents a difficulty feature of a course , is a difficulty matching function;

[0054] constructing an education consistency loss term based on the before-after repair loss term and the difficulty loss term;

[0055] constructing a multi-objective optimization loss function based on the education consistency loss term, the user contrast loss term and the course contrast loss term, referring to the following formula:

[0056]

[0057] wherein, represents a multi-objective optimization loss function, represents a BPR loss term, used for optimizing course recommendation ranking, and respectively represent a user contrast loss term and a course contrast loss term, represents an education consistency loss term, used to ensure the education rationality of the recommendation result, 、 and respectively represent the weight coefficients of each loss term.

[0058] In addition, to achieve the above-mentioned purpose, the application also provides a dual supergraph contrast learning course recommendation device fusing education prior knowledge, which comprises:

[0059] An education data processing module is configured to pre-process education data and perform feature analysis on the pre-processed education data to obtain multi-dimensional education features, wherein the education data comprises learner data, course data, video data and concept data, and the multi-dimensional education features comprise interactive relationship features, user features and course features.

[0060] A supergraph construction module is configured to determine to construct a target user supergraph and a target course supergraph based on the multi-dimensional education features.

[0061] A multi-layer graph convolution module is configured to construct a user-course interaction graph structure based on the interactive relationship features and perform multi-layer graph convolution coding on the user-course interaction graph structure to obtain interactive graph embedding features.

[0062] A supergraph aggregation module is configured to aggregate the target user supergraph and the target course supergraph, and fuse education features based on the aggregated node feature matrix to obtain supergraph embedding features.

[0063] A contrast learning module is configured to perform contrast learning on the interactive graph embedding features and the supergraph embedding features from the user level and the course level respectively to obtain a user contrast loss term and a course contrast loss term.

[0064] A multi-objective optimization module is configured to construct a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term, and construct a course recommendation model based on the multi-objective optimization loss function, and perform education course recommendation through the course recommendation model.

[0065] In addition, to achieve the above-mentioned purpose, the application also provides a dual supergraph contrast learning course recommendation device fusing education prior knowledge, which comprises: a memory, a processor and a dual supergraph contrast learning course recommendation program fusing education prior knowledge stored in the memory, wherein the processor is configured to run the dual supergraph contrast learning course recommendation program fusing education prior knowledge, and the computer program is configured to implement the steps of the dual supergraph contrast learning course recommendation method fusing education prior knowledge as described above.

[0066] In addition, to achieve the above-mentioned object, the application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the dual-hypergraph contrast learning course recommendation method based on fused education prior knowledge as described above.

[0067] The application obtains multi-dimensional education features by preprocessing education data and performing feature analysis on the preprocessed education data, the education data including learner data, course data, video data and concept data, the multi-dimensional education features including interaction relationship features, user features and course features, determines to construct a target user hypergraph and a target course hypergraph based on the multi-dimensional education features, constructs a user-course interaction graph structure based on the interaction relationship features, performs multi-layer graph convolution coding on the user-course interaction graph structure to obtain interaction graph embedding features, performs hypergraph aggregation on the target user hypergraph and the target course hypergraph, and performs education feature fusion based on the aggregated node feature matrix to obtain hypergraph embedding features, respectively performs contrast learning on the interaction graph embedding features and the hypergraph embedding features from the user level and the course level to obtain user contrast loss terms and course contrast loss terms, constructs a multi-objective optimization loss function based on the user contrast loss terms and the course contrast loss terms, constructs a course recommendation model based on the multi-objective optimization loss function, and performs education course recommendation through the course recommendation model. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0069] Figure 1 is a structural schematic diagram of a fusion education prior knowledge dual-hypergraph contrast learning course recommendation device related to the hardware running environment of the embodiment scheme of the present application.

[0070] Figure 2A flowchart of a first embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 1.

[0071] Figure 3 A flowchart of a second embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 2.

[0072] Figure 4 A flowchart of a third embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 3.

[0073] Figure 5 A flowchart of a fourth embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 4.

[0074] Figure 6 A flowchart of a fifth embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 5.

[0075] Figure 7 A course recommendation flowchart in an embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation method of the present application is shown in Figure 6.

[0076] Figure 8 A structural block diagram of a first embodiment of the fusion education prior knowledge double hypergraph contrastive learning course recommendation device of the present application is shown in Figure 7.

[0077] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0078] It should be understood that the specific embodiments described herein are intended to be illustrative only and not limiting of the present application.

[0079] Reference Figure 1 , Figure 1 A fusion education prior knowledge double hypergraph contrastive learning course recommendation device structural diagram related to the hardware running environment of the embodiment scheme of the present application is shown in Figure 8.

[0080] As Figure 1As shown in the figure, the fusion education prior knowledge double hypergraph contrastive learning course recommendation device can include a processor 1001, for example, a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection and communication between the components. The user interface 1003 can include a display screen, an input unit such as a keyboard, and can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (WI-FI) interface). The memory 1005 can be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM) such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0081] Those skilled in the art can understand that Figure 1 The structure shown in the figure does not constitute a limitation on the fusion education prior knowledge double hypergraph contrastive learning course recommendation device, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0082] As Figure 1 As shown in the figure, the memory 1005 as a computer readable storage medium can include an operating system, a network communication module, a user interface module, and a fusion education prior knowledge double hypergraph contrastive learning course recommendation program.

[0083] In Figure 1 In the fusion education prior knowledge double hypergraph contrastive learning course recommendation device shown in the figure, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the fusion education prior knowledge double hypergraph contrastive learning course recommendation device can be arranged in the fusion education prior knowledge double hypergraph contrastive learning course recommendation device, and the fusion education prior knowledge double hypergraph contrastive learning course recommendation device calls the fusion education prior knowledge double hypergraph contrastive learning course recommendation program stored in the memory 1005 through the processor 1001, and executes the fusion education prior knowledge double hypergraph contrastive learning course recommendation method provided by the embodiment of the present application.

[0084] The embodiment of the present application provides a fusion education prior knowledge double hypergraph contrastive learning course recommendation method, which is described with reference to Figure 2 ,Figure 2 A flowchart of a first embodiment of a fusion education priori knowledge double hypergraph contrastive learning course recommendation method of the present application.

[0085] In this embodiment, the fusion education priori knowledge double hypergraph contrastive learning course recommendation method includes the following steps:

[0086] Step S10: Preprocess the education data and perform feature analysis on the preprocessed education data to obtain multi-dimensional education features.

[0087] It should be noted that the present embodiment is applied to online education course recommendation, aiming to solve the problems of serious data sparsity, insufficient complex relationship modeling, and lack of balance between individualization and education rationality in traditional education recommendation. The present embodiment models learning data based on a double-level education hypergraph structure, fuses learner and learner and course and course relationship data to construct a hypergraph, processes the interactive graph structure through a multi-layer hypergraph convolution network, accurately captures the multi-dimensional features and time evolution law of learner behavior, and simultaneously incorporates the course sequence relationship to ensure the rationality of the recommended education. The present embodiment can identify the potential learning needs and knowledge development trajectory of learners, recommend individualized courses that meet the scientific learning progression path according to the current knowledge level and historical learning records of learners, and balance individualized needs and education rules through a multi-objective optimization framework, achieving more accurate and more educationally essential intelligent recommendation services.

[0088] It should be understood that the execution subject of the present embodiment can be a computing service device with data processing, network communication, and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or a terminal electronic device capable of realizing the above functions. The following takes a fusion education priori knowledge double hypergraph contrastive learning course recommendation device (referred to as a recommendation device) as an example to illustrate the present embodiment and the following embodiments.

[0089] It should be noted that the education data includes learner data, course data, video data, and concept data. The learner data can be the basic information of the learner (age, grade), learning behavior (course selection record, video viewing time), and learning state (score, progress); the course data can be the attribute information of the course (name, difficulty, duration), education association (sequence relationship, knowledge field), video data can be the viewing log of the course supporting video (viewing times, pause / fast-forward nodes), and video tags (knowledge point association); the concept data can be the knowledge concept corresponding to the course (such as "matrix multiplication" under "linear algebra"), and the hierarchical relationship between concepts (such as "derivative" is a sub-concept of "calculus").

[0090] It should be noted that the multi-dimensional education features include interaction features, user features and course features. The interaction features can include interaction records of learners and courses / videos (such as “user A takes course B” and “user C watches video D”); the user features can be features describing learners (such as learning progress, preferred knowledge field, historical performance); and the course features can be features describing courses (such as difficulty level, contained knowledge concept, applicable learning stage).

[0091] In some embodiments, the recommendation device can perform data cleaning, data standardization and data structuring on the education data, such as eliminating missing values (such as invalid users without course selection records), deduplication (such as duplicate video watching logs), correcting abnormal values (such as learning time exceeding a reasonable range); mapping numerical features (such as performance, time) to the [0, 1] interval, unifying feature scales; and converting unstructured data (such as video tags) into structured format (such as “video-concept” association table).

[0092] In specific implementations, the recommendation device can count the interaction frequency and interaction time of users and courses / videos, generate a “user-course” interaction matrix; integrate the basic information and learning behavior of learners to generate a low-dimensional feature vector of the user (such as “learning progress 0.7 + preferred field ‘mathematics’”); and integrate the attributes and education associations of courses to generate a feature vector of the course (such as “difficulty 2 + contained concept ‘matrix’ + prerequisite course ‘linear algebra foundation’”).

[0093] Step S20: determining to construct a target user hypergraph and a target course hypergraph based on the multi-dimensional education features.

[0094] It should be noted that a hypergraph is an extension concept in graph theory, and an edge can connect any number of vertices. In the education recommendation scenario, a hypergraph can more flexibly model complex learner and course interaction relationships and multi-modal feature associations.

[0095] It should be noted that the target user hypergraph is a hypergraph with learners as nodes and “commonly associated education elements” as hyperedges (such as a hyperedge connecting “all users who have taken the ‘calculus’ concept”); and the target course hypergraph is a hypergraph with courses as nodes and “commonly associated education elements” as hyperedges (such as a hyperedge connecting “all courses containing the ‘matrix multiplication’ concept”).

[0096] In specific implementations, the recommendation device can take learners as nodes, connect users who have taken the same knowledge concept and are in the same learning stage through hyperedges, and generate an association matrix of the user hypergraph; and take courses as nodes, connect courses containing the same knowledge concept and belonging to the same learning path through hyperedges, and generate an association matrix of the course hypergraph.

[0097] Step S30: constructing a user-course interaction graph structure based on the interaction relationship features, and performing multi-layer graph convolution coding on the user-course interaction graph structure to obtain interaction graph embedding features.

[0098] It should be noted that the user-course interaction graph structure is an interaction graph structure with users and courses as two types of nodes and user course selection and course learning as edges (only edges exist between users and courses, and no edges exist between users and courses); the interaction graph embedding features refer to the low-dimensional vector representation of the user / course obtained after graph convolution, which contains the structural correlation information in the interaction graph.

[0099] In a specific implementation, the recommendation device takes users and courses as nodes, and if a user has a course selection / learning behavior, an edge is established between the corresponding user and course nodes to generate an adjacency matrix; a graph convolution model (such as GCN, GAT, Light GCN, or Graph Transformer architecture) is used to perform multi-layer graph convolution using a normalized adjacency matrix, and direct neighbor information of nodes is aggregated at each layer (such as direct course selection course information of users at the first layer and other course selection user information of courses at the second layer); the output results of multi-layer graph convolution are fused to obtain interaction graph embedding features, thereby capturing the interaction structure relationship between users and courses, converting discrete interaction behaviors into continuous low-dimensional embedding vectors, and enabling the embedding to reflect the course selection preferences of users and the audience characteristics of courses.

[0100] Further, in order to accurately mine the interaction features between the course graph, the above step S30 can include:

[0101] Step S301: constructing a user-course interaction graph structure based on the interaction relationship features, and performing multi-layer graph convolution operation on the user-course interaction graph structure to aggregate information through a normalized adjacency matrix.

[0102] It should be noted that the recommendation device can use a Light GCN model to perform multi-layer graph convolution operation on the user-course interaction graph structure. In the graph convolution operation link, the recommendation device can perform multi-layer graph convolution operation on the user-course interaction graph and aggregate information through a normalized adjacency matrix to ensure the stability of information propagation and enable each user and course node to effectively aggregate information of neighbor nodes, thereby learning a representation vector containing rich neighborhood information. The formula of the k-th layer graph convolution operation is as follows:

[0103]

[0104]

[0105] wherein, is a normalized adjacency matrix, denotes an adjacency matrix, representation matrix, representing the layer embedding representation.

[0106] Step S302: average aggregation is performed on the results of the multi-layer graph convolution operation to obtain the interaction graph embedding feature.

[0107] It can be understood that in the multi-layer aggregation link, the system performs average aggregation on the results of the multi-layer graph convolution, fuses the neighborhood information of different levels, effectively avoids the problem of excessive smoothing, and at the same time preserves the structure information of multiple levels, so that the system can balance the local neighborhood information and the global structure information, and provide high-quality initial representation for subsequent supergraph enhancement. The calculation formula of the final interaction graph embedding feature is as follows:

[0108]

[0109] wherein, represents the number of graph convolution layers, represents the interaction graph embedding feature.

[0110] Step S40: supergraph aggregation is performed on the target user supergraph and the target course supergraph, and education feature fusion is performed based on the aggregated node feature matrix to obtain a supergraph embedding feature.

[0111] In a specific implementation, the recommendation device can use a supergraph convolution method to aggregate all node features in a superedge through a supergraph association matrix (for example, in a user supergraph, the user features in the same superedge are averaged); education feature fusion: using an attention mechanism (such as a multi-head attention mechanism) to assign weights to the features aggregated by the supergraph and the original multi-dimensional education features, and performing weighted fusion to obtain a supergraph embedding feature, thereby integrating the high-order association information of the supergraph and the education attribute features, improving the richness of the embedding vector, and making the embedding not only contain structural relationships, but also reflect the knowledge association, learning stage, and other characteristics of the education scene.

[0112] Step S50: respectively performing contrastive learning on the interaction graph embedding feature and the supergraph embedding feature from the user level and the course level to obtain a user contrastive loss term and a course contrastive loss term.

[0113] It should be noted that contrastive learning is a self-supervised learning method that learns the representation of data by bringing similar samples closer and pushing different samples further apart. In course recommendation, contrastive learning can be used to learn more discriminative learner and course representations.

[0114] It should be noted that the user contrastive loss term refers to the contrastive loss for user embedding, which measures the similarity of different view embeddings of the same user; the course contrastive loss term refers to the contrastive loss for course embedding, which measures the similarity of the educational association course embedding.

[0115] In the specific implementation, user-level contrastive learning includes: taking the "interaction graph embedding features" and "hypergraph embedding features" of the same user as positive sample pairs, and the embeddings of other users as negative sample pairs, and using InfoNCE loss to calculate user contrastive loss;

[0116] Course-level contrastive learning includes embedding courses with educational relevance (such as prior or subsequent courses, or the same knowledge concepts) as positive sample pairs and embedding courses without relevance as negative sample pairs, and using InfoNCE loss to calculate the course contrastive loss.

[0117] Furthermore, to improve the rationality of course recommendations, step S50 above may include:

[0118] Step S501: Compare and learn the interaction graph embedding features and hypergraph embedding features of the same user at the user level to construct positive and negative sample pairs for the user.

[0119] It's important to note that the user-level contrastive learning phase focuses on improving the discriminativeness of user representations. The system performs contrastive learning on different view representations of the same user, specifically including interactive graph embeddings and hypergraph embeddings. In this way, recommendation devices can make the two representations of the same user closer in the embedding space, while simultaneously making the representations of different users more distinct. This contrastive learning strategy significantly improves the discriminativeness of user representations, providing a more accurate user profile for personalized recommendations.

[0120] Step S502: Construct positive and negative sample pairs for the curriculum based on prior educational knowledge information at the curriculum level.

[0121] It should be noted that the comparative learning at the course level constructs positive and negative sample pairs of courses based on prior educational knowledge. For positive sample construction, the recommendation device prioritizes course pairs with sequential relationships, adjacent courses within the same learning path, and courses with similar content and matching difficulty. This construction strategy ensures that courses with educational relevance are closer together in the embedding space, conforming to educational principles. For negative sample construction, the recommendation device actively excludes course pairs with excessively large differences in difficulty, course pairs that violate sequential relationships, and course pairs with completely unrelated content, preventing the model from learning incorrect relationships.

[0122] Step S503: Input the positive and negative sample pairs of the user and the positive and negative sample pairs of the course into the InfoNCE loss function to obtain the user contrast loss term and the course contrast loss term.

[0123] It can be understood that the embodiment adopts an InfoNCE loss function for the contrast loss calculation link, and optimizes the similarity distribution of positive and negative sample pairs. The parameter τ controls the sharpness of the contrast learning, and a smaller τ value makes the model pay more attention to the most similar samples, and a larger τ value makes the distribution smoother, so that the model can learn more discriminative and educational representations, and significantly improve the overall performance of the recommendation system. The mathematical expression of the InfoNCE loss function is as follows:

[0124]

[0125] wherein, and denote a positive sample pair, denote a negative sample, is a temperature parameter for controlling the discriminability, is a similarity function, is the number of negative samples.

[0126] Step S60: constructing a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term, and constructing a course recommendation model based on the multi-objective optimization loss function, and performing educational course recommendation through the course recommendation model.

[0127] It should be noted that the multi-objective optimization loss function can be a composite loss function integrating multiple losses (such as contrast loss, recommendation ranking loss, and education consistency loss); and the course recommendation model can be a model based on user / course embedding features, outputting a course recommendation list.

[0128] In a specific implementation, the recommendation device integrates the user contrast loss, the course contrast loss, the recommendation ranking loss (such as BPR loss), and the education consistency loss (such as the before-after loss) into a total loss function according to the weights; takes the total loss function as an optimization target, trains the model through gradient descent, and updates the embedding features and the model parameters; calculates the similarity (such as cosine similarity) between the embedding of the target user and the embedding of all courses, sorts according to the similarity, and combines the education rules (such as filtering the courses not yet taken), and outputs a recommendation list.

[0129] The embodiment pre-processes education data, and performs feature analysis on the pre-processed education data to obtain multi-dimensional education features, the education data including learner data, course data, video data and concept data, the multi-dimensional education features including interaction relationship features, user features and course features, target user supergraph and target course supergraph are determined based on the multi-dimensional education features, a user-course interaction graph structure is constructed based on the interaction relationship features, and multi-layer graph convolution coding is performed on the user-course interaction graph structure to obtain interaction graph embedding features, the target user supergraph and the target course supergraph are supergraph aggregated, and education feature fusion is performed based on the aggregated node feature matrix to obtain supergraph embedding features, the interaction graph embedding features and the supergraph embedding features are compared and learned from the user level and the course level respectively to obtain user comparison loss terms and course comparison loss terms, a multi-objective optimization loss function is constructed based on the user comparison loss terms and the course comparison loss terms, a course recommendation model is constructed based on the multi-objective optimization loss function, and education course recommendation is performed through the course recommendation model; since the embodiment models learning data based on a double-level education supergraph structure, learner and learner relationship data and course and course relationship data are fused to construct a supergraph, an interaction graph structure is processed through a multi-layer supergraph convolution network, multi-dimensional features and time evolution rules of learner behavior are accurately captured, course sequence relationship is integrated to ensure the rationality of recommended education, potential learning needs and knowledge development tracks of learners can be identified, personalized courses in line with scientific learning progression paths are recommended according to current knowledge levels and historical learning records of learners, and a multi-objective optimization framework is used to balance personalized needs and education rules, so that more accurate and more education-essence-compliant intelligent recommendation services are realized.

[0130] Reference Figure 3 , Figure 3 The flowchart of the second embodiment of the double supergraph comparison learning course recommendation method fusing education prior knowledge of the application is shown.

[0131] Based on the first embodiment, in the embodiment, the step S10 further includes:

[0132] Step S101: data cleaning and standardization processing are performed on the education data to obtain pre-processed education data.

[0133] In a specific implementation, the recommendation device can use an efficient data loading framework for education scenarios to load learner data (L), course data (C), video data (V), and concept data (K) in an education dataset (e.g., MOOC Cube) by type. By implementing an outlier detection algorithm, noise data deviating from the normal range is automatically identified and removed based on statistical distribution characteristics. For missing values, a domain knowledge-based interpolation method is used for processing to ensure data integrity and consistency. Specific processing includes: filtering abnormal learning duration and unreasonable interaction behavior in learner data; removing duplicate content in course data; and standardizing meta-information of video data.

[0134] Step S102: analyzing learning behavior data of learners in the preprocessed education data, and constructing a weighted interaction matrix, wherein the weighted interaction matrix is used to represent interaction relationship features.

[0135] It can be understood that the recommendation device can subdivide the interaction relationship into strong interaction and weak interaction by analyzing the learning behavior data of the learners, and construct a weighted interaction matrix to provide a richer representation of the relationship between users and courses for subsequent models, so that the model can more accurately learn user preferences and course features. The definition of the interaction intensity matrix R is as follows:

[0136]

[0137] wherein, represents the interaction intensity of user u to course c, (user set), (course set).

[0138] Step S103: analyzing the interaction relationship between each user and course data and video data based on the preprocessed education data, and determining learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity between users, wherein the learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity are used to represent user features.

[0139] It should be noted that in this embodiment, the user feature extraction includes:

[0140] The recommendation device first counts the number of courses that the user has learned, and uses the number as a basic indicator of the learning stage. By setting a stage division threshold, the learning stage level is determined according to the user's course learning interval, thereby obtaining the learning stage feature. The mathematical expression of the learning stage feature is as follows:

[0141]

[0142] wherein, represents the user The characteristics of the learning stages Indicates user The collection of courses already studied Indicates the size of the set.

[0143] Subsequently, the recommendation device generates learning sequences based on user-course relationship data, recording the order in which users learn courses, thereby constructing user learning path vectors. Jaccard is then used to calculate the similarity between different users on their learning paths. The formula for calculating learning path similarity is as follows:

[0144]

[0145] in, This represents the learning path similarity function. Indicates user A collection of learning course sequences, Indicates user A set of learning course sequences.

[0146] When extracting learning preference features, the system statistically analyzes the distribution of user interests across different types of courses based on user interaction data, constructs a feature vector of user preferences, and then calculates the similarity of course interests between users using Jaccard similarity, as shown in the following formula:

[0147]

[0148] in, This represents a function for determining the similarity of course interests. Indicates user A collection of courses already taken. Indicates user A collection of courses already taken.

[0149] Furthermore, the recommendation device uses user interaction data with video resources to count the number of videos watched by each user, normalizes the data to obtain learning speed features, and then calculates the learning speed similarity between users. The mathematical expression for learning speed similarity is as follows:

[0150]

[0151] in, The similarity function represents the learning speed. Indicates user The collection of videos to watch Indicates user A collection of videos to watch.

[0152] Step S104: Extract course content information based on the preprocessed educational data, and generate content features for each course based on the course content information;

[0153] Step S105: Determine the content similarity, difficulty similarity, structural similarity, and association similarity among the courses based on the content features. The content similarity, difficulty similarity, structural similarity, and association similarity are used to characterize the course features.

[0154] It should be noted that course feature extraction may include:

[0155] First, the recommendation device extracts course content information based on the text fields of course entities. The acquired text data, such as course titles and descriptions, is then converted into numerical vector representations using TF-IDF vectorization technology, thus forming the content features of the courses. Next, the system calculates the content similarity between courses to characterize the degree of relevance of different courses at the knowledge content level. The mathematical expression for content similarity calculation is as follows:

[0156]

[0157] in, Indicates course TF-IDF vector, Represents the vector dot product. Represents the magnitude of the vector. This represents the content similarity function. Indicates course TF-IDF feature vectors Indicates course TF-IDF feature vectors Represents the vector dot product. This represents the magnitude of the vector.

[0158] Subsequently, the recommendation device extracts chapter information based on course entity information and calculates course difficulty based on the number of chapters. The mathematical expression for difficulty feature extraction is as follows:

[0159]

[0160] in, Indicates course Chapter list, Indicates course A collection of chapters, Indicates the number of chapters.

[0161] In extracting the course structure features, the system extracts the video information contained in the course based on the course-video relationship data, and then calculates the structural complexity of the course through the number of videos, which is further used to calculate the structural similarity between courses. The mathematical expression of the structural similarity calculation is as follows:

[0162]

[0163] wherein, denotes the structural similarity function, denotes the course contains the video set, denotes the course contains the video set, denotes the size of the set.

[0164] Finally, the recommendation device obtains the prerequisite information of the course from the data set, and calculates the association similarity between courses through the Jaccard similarity. The mathematical expression of the association similarity calculation is as follows:

[0165]

[0166] wherein, denotes the prerequisite relationship similarity function, denotes the prerequisite course set of the course , denotes the prerequisite course set of the course , denotes the size of the set.

[0167] The present embodiment is aimed at the situation where the data sparsity and cold start problem are prominent. By adopting multi-dimensional data preprocessing technology, the strong interaction (course registration) and weak interaction (video watching) of learners are distinguished, multi-dimensional features such as learning stage, course difficulty and course content features are extracted, structured educational knowledge such as course prerequisite relationship and knowledge point association is integrated, multi-modal features (text, time, interaction) and cross-domain information fusion technology are used to construct comprehensive feature representation, and the data sparsity and cold start problem are effectively alleviated.

[0168] Reference Figure 4 , Figure 4 is a flowchart of the third embodiment of the dual hypergraph contrastive learning course recommendation method of the present application which fuses educational prior knowledge.

[0169] Based on the above embodiments, in the present embodiment, the step S20 further comprises:

[0170] Step S201: determining the user similarity between each user based on the learning stage similarity, learning path similarity, course interest similarity and learning speed similarity;

[0171] Step S202: constructing an initial user hypergraph based on user similarities between users.

[0172] It can be understood that the recommendation device can calculate the similarity between users based on course interest similarity, video watching similarity, learning stage similarity, learning path similarity and learning speed similarity, for constructing a user hypergraph, reflecting high-order relationships between users, enabling users with similar learning characteristics to influence each other, thereby constructing a user-user hypergraph, i.e., an initial user hypergraph. The user similarity calculation is referred to the following formula:

[0173]

[0174]

[0175] wherein, represents the user similarity between user and user , , , , and respectively represent weight coefficients, , , , and respectively represent course interest similarity, video watching similarity, learning stage similarity, learning path similarity and learning speed similarity.

[0176] Step S203: determining course similarity between courses based on the content similarity, difficulty similarity, structure similarity and association similarity, and constructing an initial course hypergraph based on the course similarity.

[0177] It should be understood that the recommendation device can construct a course-course hypergraph based on course association relationships, and the recommendation device calculates course similarity by comprehensively considering content feature similarity, video resource similarity and sequence relationship similarity of courses, thereby constructing a course-course hypergraph, i.e., an initial course hypergraph. The course similarity calculation is referred to the following formula:

[0178]

[0179]

[0180] wherein, represents the course similarity between course and course , represents the content similarity calculated based on TF-IDF and cosine similarity, representing the similarity of video resources, representing the similarity of the precedence relationship.

[0181] Step S204: Sparse processing is performed on the initial user hypergraph and the initial course hypergraph to obtain a candidate user hypergraph and a candidate course hypergraph.

[0182] It can be understood that, for the initial user hypergraph and the initial course hypergraph, the recommendation device respectively retains 30 neighbor nodes with the highest similarity for each node, and the remaining connection weights are set to 0. This sparse strategy can effectively improve the calculation efficiency and reduce the noise influence, while retaining the most important similarity relationship, ensuring the stability and efficiency of model training. Therefore, the candidate user hypergraph and the candidate course hypergraph are obtained. The mathematical expression of sparse processing is as follows:

[0183]

[0184] wherein, denotes the hypergraph matrix after sparsification, denotes the original hypergraph matrix, denotes the node set of similar neighbor nodes of .

[0185] Step S205: Hypergraph normalization processing is performed on the candidate user hypergraph and the candidate course hypergraph to obtain a target user hypergraph and a target course hypergraph.

[0186] It can be understood that, in order to improve the stability of model training, the recommendation device performs normalization processing on the sparse hypergraph to ensure that the sum of all connection weights of each node is 1, avoiding gradient explosion and numerical instability problems. The mathematical expression of hypergraph normalization is as follows:

[0187]

[0188] wherein, denotes the hypergraph matrix after normalization, denotes the hypergraph matrix after sparsification, is a small constant to prevent division by zero (usually set to ), is the total number of nodes.

[0189] ​The embodiment aims at the limitation of the traditional graph model in modeling high-order correlation. A double-level education hypergraph structure of learner and learner hypergraph and course and course hypergraph is constructed. The similarity between learners is calculated based on five dimensions of course interest, video watching, learning stage, learning path and learning speed, and a TopK sparsification strategy is used to optimize the calculation efficiency. The course and course hypergraph fuses the course text content similarity, video resource similarity and course sequence relationship to realize effective modeling of high-order relationship.

[0190] Reference Figure 5 , Figure 5 The flowchart of the fourth embodiment of the double hypergraph contrastive learning course recommendation method of the application fusing education prior knowledge is shown.

[0191] Based on the above embodiments, in the embodiment, the step S40 further includes:

[0192] Step S401: The hypergraph aggregation is performed on the target user hypergraph and the target course hypergraph by multiplying the hypergraph association matrix and the node feature.

[0193] It can be understood that in the embodiment, the hypergraph aggregation adopts the multiplication of the hypergraph association matrix and the node feature, so that each node can aggregate all the node information of the hyperedge where the node is located, thereby modeling the high-order relationship that cannot be expressed by the traditional graph structure, and enabling the users or courses with similar characteristics to influence each other and learn more rich and distinctive representations. The mathematical expression of the hypergraph aggregation is as follows:

[0194]

[0195] wherein, represents the aggregated node feature matrix, represents the hypergraph association matrix, represents the original node feature matrix.

[0196] Step S402: Education feature fusion is performed based on the aggregated node feature matrix to obtain an enhanced node feature matrix.

[0197] It should be noted that the education feature fusion step integrates the education field features into the hypergraph aggregation result, ensuring that the representation learned by the model contains rich education prior knowledge. For user nodes, the embodiment uses a sigmoid function to calculate the weight of the education feature, and adjusts the aggregated feature through a multiplication operation. For course nodes, the system adds the education feature to the aggregated feature with a weight of 0.2. The system ensures that the recommendation result is more in line with the education rules through the differentiated fusion strategy, and avoids the combination of courses that violate the education common sense. The mathematical expression of the education feature fusion is as follows:

[0198]

[0199] in, This represents the enhanced node feature matrix. This represents the aggregated node feature matrix. Represents element-wise multiplication. This represents the sigmoid activation function. Represents the projection matrix of educational features. This represents the educational feature vector.

[0200] Step S403: Use a multi-head attention mechanism to perform deep processing on the aggregated node feature matrix to obtain the transformed feature matrix.

[0201] Understandably, this embodiment uses multi-head attention to perform deep processing on the aggregated features, learning the complex dependencies in the feature space. Through self-attention calculation of the Query, Key, and Value matrices, the model can adaptively focus on important feature dimensions, significantly improving recommendation accuracy. Subsequently, the system employs layer normalization technology to stabilize the training process, ensuring the model converges to the optimal solution, further enhancing the model's expressive power. The calculation formula for the multi-head attention mechanism is as follows:

[0202]

[0203]

[0204] in, These are Query, Key, and Value matrices. It is the dimension of the key vector. It's the number of heads that attract attention. It is the output projection matrix.

[0205] Step S404: Calculate the gating weight vector based on the transformed feature matrix and the enhanced node feature matrix, and perform gating fusion of the transformed feature matrix and the enhanced node feature matrix based on the gating weight vector to obtain the hypergraph embedding features.

[0206] It should be noted that the gated fusion stage achieves adaptive fusion of original and enhanced features through a gating mechanism. The system uses the sigmoid function to calculate the gating weights and dynamically adjusts the contribution ratio of the two features based on these weights to balance the importance of different information sources, effectively avoiding information redundancy and improving the robustness and generalization ability of the model. The mathematical expression for gated fusion is as follows:

[0207]

[0208]

[0209] in, Represents the gating weight vector. This represents the final output feature matrix. Represents the transformed feature matrix. This represents the enhanced feature matrix. Represents the gate weight matrix. This indicates a feature concatenation operation. Represents element-wise multiplication. This represents the sigmoid activation function.

[0210] This embodiment addresses the problem that traditional hypergraph convolution uses a single feature extraction path, which cannot independently learn and effectively fuse representations in different subspaces. It constructs a multi-head attention mechanism to capture the associated features in different semantic spaces in parallel, adaptively focusing on key relationships and important features in educational scenarios. Furthermore, it addresses the issue that traditional recommendation models often ignore the temporal dependencies of learning paths, failing to effectively model the knowledge accumulation process and predict learners' long-term learning needs. This embodiment designs a temporal modeling sublayer within the hypergraph convolutional layer to specifically capture the temporal sequence characteristics of learners' behavior, modeling historical learning paths and knowledge development trajectories. This enables the recommendation system to predict future learning needs based on temporal information, providing more forward-looking and continuous educational recommendation services.

[0211] refer to Figure 6 , Figure 6 This is a flowchart illustrating the fifth embodiment of the dual-hypergraph comparative learning course recommendation method that integrates prior educational knowledge according to the present invention.

[0212] Based on the above embodiments, in this embodiment, step S60 further includes:

[0213] Step S601: Construct the pre- and post-repair loss terms.

[0214] It should be noted that the pre- and post-course relationship loss is a core component of the educational consistency constraint. The system calculates the loss due to the difference in user ratings between prerequisite and subsequent courses, forcing the model to follow the educational pre- and post-course logic. This ensures that users' ratings for prerequisite courses are no lower than those for subsequent courses, enabling the system to avoid recommending course combinations that violate the learning order and guaranteeing the educational rationality of the recommendations. The mathematical expression for the pre- and post-course relationship loss is as follows:

[0215]

[0216] in, This indicates the loss in the relationship between the repairers and the repairers. This represents a set of pre- and post-modification relationships. Indicates user For the course Predicted score representing a user predicted score of the user on a course, is a tolerance parameter; Step S602: constructing a difficulty loss term.

[0217] It should be noted that the difficulty matching loss further strengthens the educational rationality constraint. The system calculates the loss of the matching degree of the recommended course difficulty and the user's learning stage, ensures that the course difficulty recommended to the user matches the current learning stage, avoids recommending too difficult or too easy courses, enables the system to provide courses suitable for the ability level of users at different learning stages, and significantly improves the learning effect and user experience. The mathematical expression of the difficulty matching loss is as follows:

[0218]

[0219] wherein,

[0220] difficultymatchingloss represents the difficulty matching loss, user set represents a user set, course set represents a course set, predicted score of the user on a course, learning stage feature of the user, difficulty feature of the course, is a difficulty matching function; Step S603: constructing an educational consistency loss term based on the sequence loss term and the difficulty loss term. It should be noted that the educational consistency loss term is responsible for ensuring the educational rationality of the recommendation result. Step S604: constructing a multi-objective optimization loss function based on the educational consistency loss term, the user contrast loss term, and the course contrast loss term.

[0221] In specific implementation, the recommendation device combines the BPR loss, the contrast loss, and the educational consistency loss according to preset weights to form a comprehensive total loss function. The mathematical expression of the multi-objective optimization loss function is as follows:

[0222]

[0223]

[0224] wherein, multi-objective optimization loss function represents a multi-objective optimization loss function,

[0225] BPR loss term is used for optimizing course recommendation ranking, and

[0226] ​​​​​respectively represent the user contrast loss term and the curriculum contrast loss term, respectively represent the user contrast loss term and the curriculum contrast loss term, respectively represent the user contrast loss term and the curriculum contrast loss term,

[0227] The embodiment is aimed at the problem that the traditional education recommendation system generally ignores the knowledge dependency structure specific to the education field (such as the relationship between courses and the difficulty level structure), which leads to the problem that the recommended result may violate the scientific learning progression path and it is difficult to balance the recommendation accuracy and the education rationality. The embodiment designs a multi-objective optimization loss function, including a BPR recommendation loss (optimizing the recommendation accuracy), a contrast learning loss (enhancing the discriminativeness of the feature representation), and an education consistency loss (including a sequence compliance loss and a difficulty matching loss, ensuring that the recommended result conforms to the education law), so as to ensure that the recommended result conforms to the education law while ensuring the recommendation accuracy.

[0228] In some embodiments, reference is made to Figure 7 Figure 7 ​​​A course recommendation flowchart is provided, a multi-dimensional data preprocessing technology is used for recommending equipment, strong interaction (course registration) and weak interaction (video watching) of learners are distinguished, multi-dimensional features such as learning stage, course difficulty and course content characteristics of learners are extracted, structured education knowledge such as course sequence relationship and knowledge point association is integrated, multi-modal features (text, time, interaction) and cross-domain information fusion technology are used to construct comprehensive feature representation, and the data sparsity and cold start problems are effectively alleviated; a double-layer education hypergraph structure of learner and learner hypergraph and course and course hypergraph is constructed, the similarity between learners is calculated based on five dimensions of course interest, video watching, learning stage, learning path and learning speed, and a TopK sparsification strategy is used to optimize the calculation efficiency; the course and course hypergraph fuses course text content similarity, video resource similarity and course sequence relationship, and effectively models high-order relationships; multi-head attention is constructed to capture the associated features in different semantic spaces in parallel, and the key relationships and important features in the education scene are adaptively focused; a time sequence modeling sublayer is designed in the hypergraph convolution layer to specially capture the time sequence characteristics of learner behavior, model the historical learning path and knowledge development track, so that the recommendation system can predict future learning needs based on time sequence information and provide more forward-looking and continuous education recommendation services; a multi-objective optimization loss function is designed, including a BPR recommendation loss (optimizing recommendation accuracy), a contrastive learning loss (enhancing the discriminability of feature representation) and an education consistency loss (including a sequence compliance loss and a difficulty matching loss, ensuring that the recommended results conform to the education law), which ensures that the recommended results conform to the education law while ensuring the recommendation accuracy.

[0229] In addition, the embodiment of the present application also provides a computer readable storage medium, and the computer readable storage medium stores a double hypergraph contrastive learning course recommendation program fusing education prior knowledge, and the double hypergraph contrastive learning course recommendation program fusing education prior knowledge realizes the steps of the double hypergraph contrastive learning course recommendation method fusing education prior knowledge when being executed by a processor.

[0230] The computer readable storage medium provided in the application may be, for example, a U disk, but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination of the above. More specific examples of the computer readable storage medium may include, but are not limited to, an electrical connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the embodiment, the computer readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to an electrical wire, an optical cable, an RF (Radio Frequency), and the like, or any suitable combination of the above.

[0231] The above computer readable storage medium may be included in the dual hypergraph contrastive learning course recommendation device fusing educational prior knowledge, or may exist independently and not be assembled into the dual hypergraph contrastive learning course recommendation device fusing educational prior knowledge.

[0232] In addition, the embodiment of the application also provides a computer program product, including a dual hypergraph contrastive learning course recommendation program fusing educational prior knowledge, and the dual hypergraph contrastive learning course recommendation program fusing educational prior knowledge realizes the steps of the dual hypergraph contrastive learning course recommendation method fusing educational prior knowledge when executed by a processor.

[0233] The computer program product embodiment of the application is basically the same as the above-mentioned dual hypergraph contrastive learning course recommendation method fusing educational prior knowledge, and will not be repeated here.

[0234] Reference Figure 8 , Figure 8 The application provides a structure block diagram of a first embodiment of the dual hypergraph contrastive learning course recommendation device fusing educational prior knowledge.

[0235] As shown in Figure 8 , the dual hypergraph contrastive learning course recommendation device fusing educational prior knowledge provided by the embodiment of the application includes:

[0236] An education data processing module 10 is configured to preprocess education data and perform feature analysis on the preprocessed education data to obtain multi-dimensional education features, the education data including learner data, course data, video data and concept data, and the multi-dimensional education features including interaction relationship features, user features and course features;

[0237] A hypergraph construction module 20 is configured to determine a target user hypergraph and a target course hypergraph based on the multi-dimensional education features;

[0238] A multi-layer graph convolution module 30 is configured to construct a user-course interaction graph structure based on the interaction relationship features and perform multi-layer graph convolution coding on the user-course interaction graph structure to obtain interaction graph embedding features;

[0239] A hypergraph aggregation module 40 is configured to aggregate the target user hypergraph and the target course hypergraph and perform education feature fusion based on the aggregated node feature matrix to obtain hypergraph embedding features;

[0240] A contrast learning module 50 is configured to perform contrast learning on the interaction graph embedding features and the hypergraph embedding features from the user level and the course level respectively to obtain a user contrast loss term and a course contrast loss term;

[0241] A multi-objective optimization module 60 is configured to construct a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term and construct a course recommendation model based on the multi-objective optimization loss function to perform education course recommendation through the course recommendation model.

[0242] The embodiment pre-processes education data, and performs feature analysis on the pre-processed education data to obtain multi-dimensional education features, the education data including learner data, course data, video data and concept data, the multi-dimensional education features including interaction relationship features, user features and course features, target user supergraph and target course supergraph are determined based on the multi-dimensional education features, a user-course interaction graph structure is constructed based on the interaction relationship features, and multi-layer graph convolution coding is performed on the user-course interaction graph structure to obtain interaction graph embedding features, the target user supergraph and the target course supergraph are aggregated, and education feature fusion is performed based on the aggregated node feature matrix to obtain supergraph embedding features, the interaction graph embedding features and the supergraph embedding features are learned from the user level and the course level respectively to obtain user contrast loss terms and course contrast loss terms, a multi-objective optimization loss function is constructed based on the user contrast loss terms and the course contrast loss terms, a course recommendation model is constructed based on the multi-objective optimization loss function, and education course recommendation is performed through the course recommendation model; since the embodiment models learning data based on a double-level education supergraph structure, the supergraph is constructed by fusing learner and learner relationship data and course and course relationship data, the multi-layer supergraph convolution network processes the interaction graph structure, accurately captures multi-dimensional features and time evolution rules of learner behavior, and simultaneously integrates course sequence relationship to ensure the rationality of recommended education, can identify potential learning needs and knowledge development trajectories of learners, recommends personalized courses in line with scientific learning progression paths according to current knowledge levels and historical learning records of learners, and balances personalized needs and education rules through a multi-objective optimization framework to realize more accurate and more education-essence-compliant intelligent recommendation services.

[0243] The double supergraph contrast learning course recommendation device fusing education prior knowledge provided in the application adopts the double supergraph contrast learning course recommendation method fusing education prior knowledge in the above embodiment, and can solve the technical problem of the double supergraph contrast learning course recommendation fusing education prior knowledge. Compared with the prior art, the double supergraph contrast learning course recommendation device fusing education prior knowledge provided in the application has the same beneficial effects as the double supergraph contrast learning course recommendation method fusing education prior knowledge provided in the above embodiment, and other technical features in the double supergraph contrast learning course recommendation device fusing education prior knowledge are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0244] It should be understood that the above is only for illustration, and does not constitute any limitation on the technical solutions of the application. In specific applications, those skilled in the art can set it up as needed, and the application does not limit this.

[0245] It should be noted that the above-described workflow is merely illustrative and does not limit the scope of protection of the present application. In actual application, a person skilled in the art can select part or all of the above-described workflow to achieve the purpose of the embodiment according to actual needs, which is not limited herein.

[0246] In addition, technical details not described in detail in the present embodiment can be found in the fusion education priori knowledge of the dual hypergraph contrast learning course recommendation method provided by any embodiment of the present application, which will not be repeated here.

[0247] It should be noted that in this document, the terms "comprise", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or systems including a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or systems. Without more limitations, the element defined by the sentence "comprises a" does not exclude the presence of additional identical elements in the process, method, article or system comprising the element.

[0248] The above-mentioned embodiment numbers of the present application are only for description, not representing the advantages and disadvantages of the embodiments.

[0249] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platform, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of software product, which is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk), including a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) execute the methods described in various embodiments of the present application.

[0250] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for recommending courses based on a dual-hypergraph comparative learning approach that integrates prior educational knowledge, characterized in that: The recommended method for integrating prior knowledge in education through a dual-hypergraph comparative learning courses includes: Educational data is preprocessed, and feature analysis is performed on the preprocessed educational data to obtain multidimensional educational features. The educational data includes learner data, course data, video data, and concept data. The multidimensional educational features include interaction relationship features, user features, and course features. Based on the aforementioned multidimensional educational characteristics, a target user hypergraph and a target course hypergraph are determined and constructed. Based on the interaction relationship features, a user-course interaction graph structure is constructed, and multi-layer graph convolutional encoding is performed on the user-course interaction graph structure to obtain the interaction graph embedding features. Hypergraph aggregation is performed on the target user hypergraph and the target course hypergraph, and educational feature fusion is performed based on the aggregated node feature matrix to obtain hypergraph embedding features; The interaction graph embedding features and the hypergraph embedding features are compared and learned from the user level and the course level, respectively, to obtain the user comparison loss term and the course comparison loss term. A multi-objective optimization loss function is constructed based on the user comparison loss term and the course comparison loss term, and a course recommendation model is constructed based on the multi-objective optimization loss function. Educational courses are recommended through the course recommendation model. The construction of a multi-objective optimization loss function based on the user contrast loss term and the course contrast loss term includes: Construct the pre- and post-repair loss terms, which are based on the following formula: in, This indicates the loss in the relationship between the repairers and the repairers. This represents a set of pre- and post-modification relationships. Indicates user For the course Predicted score Indicates user For the course Predicted score For tolerance parameters; Construct a difficulty loss term, which is based on the following formula: in, This represents the difficulty matching loss. Represents a set of users. Represents a set of courses. Indicates user For the course Predicted score Indicates user The characteristics of the learning stages Indicates course The difficulty characteristics, It is a difficulty matching function; Construct an education consistency loss term based on the aforementioned prior repair loss term and the aforementioned difficulty loss term; A multi-objective optimization loss function is constructed based on the education consistency loss term, the user contrast loss term, and the course contrast loss term, as shown in the following formula: in, This represents the loss function for multi-objective optimization. This represents the BPR loss term, used to optimize the ranking of recommended courses. and These represent the user comparison loss and the course comparison loss, respectively. This represents the educational consistency loss term, used to ensure the educational rationality of the recommendation results. , and These represent the weighting coefficients of each loss term.

2. The method for recommending courses based on the dual-hypergraph comparison of prior knowledge in integrated education as described in claim 1, characterized in that, The process of preprocessing educational data and performing feature analysis on the preprocessed data to obtain multidimensional educational features includes: The educational data is cleaned and standardized to obtain preprocessed educational data. Analyze learners' learning behavior data in the preprocessed educational data, and construct a weighted interaction matrix, which is used to characterize the interaction relationship features; Based on the preprocessed educational data analysis, the interaction relationships between each user and the course data and video data are analyzed to determine the similarity of learning stage, learning path, course interest, and learning speed among each user. The similarity of learning stage, learning path, course interest, and learning speed are used to characterize user features. Based on the preprocessed educational data, course content information is extracted, and content features of each course are generated based on the course content information. Based on the aforementioned content features, the content similarity, difficulty similarity, structural similarity, and association similarity among the courses are determined. These similarities are used to characterize the course features.

3. The method for recommending courses based on the dual-hypergraph comparison of prior knowledge in integrated education as described in claim 2, characterized in that, The process of determining and constructing the target user hypergraph and the target course hypergraph based on the multidimensional educational features includes: The user similarity between users is determined based on the learning stage similarity, learning path similarity, course interest similarity, and learning speed similarity. Construct an initial user hypergraph based on user similarity among users; The course similarity between courses is determined based on the content similarity, difficulty similarity, structural similarity, and association similarity, and an initial course hypergraph is constructed based on the course similarity. The initial user hypergraph and the initial course hypergraph are sparsified to obtain candidate user hypergraphs and candidate course hypergraphs. Hypergraph normalization is performed on the candidate user hypergraph and the candidate course hypergraph to obtain the target user hypergraph and the target course hypergraph.

4. The method for recommending courses based on the dual-hypergraph comparative learning of integrated educational prior knowledge as described in any one of claims 1 to 3, characterized in that, The step of constructing a user-course interaction graph structure based on the interaction relationship features, and performing multi-layer graph convolutional encoding on the user-course interaction graph structure to obtain interaction graph embedding features includes: Based on the aforementioned interaction relationship features, a user-course interaction graph structure is constructed, and multi-layer graph convolution operations are performed on the user-course interaction graph structure to aggregate information through a normalized adjacency matrix. The mathematical expressions for each layer of graph convolution operations are as follows: in, It is a normalized adjacency matrix. Represents the adjacency matrix. Degree matrix, Indicates the first Layer embedding representation; The results of multi-layer graph convolution operations are averaged and aggregated to obtain the interaction graph embedding features, as shown in the following formula: in, Indicates the number of convolutional layers in the graph. This represents the embedding features of the interaction graph.

5. The method for recommending courses based on the dual-hypergraph comparative learning of integrated educational prior knowledge as described in any one of claims 1 to 3, characterized in that, The process of hypergraph aggregation of the target user hypergraph and the target course hypergraph, followed by educational feature fusion based on the aggregated node feature matrix to obtain hypergraph embedding features, includes: Hypergraph aggregation is performed on the target user hypergraph and the target course hypergraph by multiplying the hypergraph association matrix with node features, as shown in the following formula: in, This represents the aggregated node feature matrix. Represents the hypergraph incidence matrix. Represents the original node feature matrix; Based on the aggregated node feature matrix, educational features are fused to obtain the enhanced node feature matrix, as shown in the following formula: in, This represents the enhanced node feature matrix. This represents the aggregated node feature matrix. Represents element-wise multiplication. This represents the sigmoid activation function. Represents the projection matrix of educational features. Represents the educational feature vector; A multi-head attention mechanism is used to perform deep processing on the aggregated node feature matrix to obtain the transformed feature matrix. Based on the transformed feature matrix and the enhanced node feature matrix, a gating weight vector is calculated, and based on the gating weight vector, the transformed feature matrix and the enhanced node feature matrix are gating fusion to obtain the hypergraph embedding features.

6. The method for recommending courses based on the dual-hypergraph comparative learning of integrated educational prior knowledge as described in any one of claims 1 to 3, characterized in that, The step of performing comparative learning on the interaction graph embedding features and the hypergraph embedding features from the user level and the course level respectively to obtain user contrast loss terms and course contrast loss terms includes: At the user level, we compare and learn the interaction graph embedding features and hypergraph embedding features of the same user to construct positive and negative sample pairs for the user. Construct positive and negative sample pairs for the curriculum based on prior educational knowledge at the curriculum level; Input the positive and negative sample pairs of users and courses into the InfoNCE loss function to obtain the user contrastive loss term and the course contrastive loss term, as shown in the following formula: in, and Indicates positive sample pairs. Indicates a negative sample. It is a temperature parameter that controls the discrimination ability. It is a similarity function. It represents the number of negative samples.

7. A dual-hypergraph comparative learning curriculum recommendation device that integrates prior educational knowledge, characterized in that, The device for recommending courses based on the dual hypergraph comparison learning of prior knowledge in integrated education includes: The education data processing module is used to preprocess education data and perform feature analysis on the preprocessed education data to obtain multidimensional education features. The education data includes learner data, course data, video data, and concept data. The multidimensional education features include interaction relationship features, user features, and course features. The hypergraph construction module is used to determine and construct the target user hypergraph and the target course hypergraph based on the multidimensional educational features. The multi-layer graph convolution module is used to construct a user-course interaction graph structure based on the interaction relationship features, and to perform multi-layer graph convolution encoding on the user-course interaction graph structure to obtain interaction graph embedding features; The hypergraph aggregation module is used to aggregate the target user hypergraph and the target course hypergraph, and perform educational feature fusion based on the aggregated node feature matrix to obtain hypergraph embedding features; The contrastive learning module is used to perform contrastive learning on the interaction graph embedding features and the hypergraph embedding features from the user level and the course level, respectively, to obtain the user contrastive loss term and the course contrastive loss term. The multi-objective optimization module is used to construct a multi-objective optimization loss function based on the user comparison loss term and the course comparison loss term, and to construct a course recommendation model based on the multi-objective optimization loss function, and to recommend educational courses through the course recommendation model; The multi-objective optimization module is also used to construct the pre- and post-repair loss terms, which are formulated according to the following formula: in, This indicates the loss in the relationship between the repairers and the repairers. This represents a set of pre- and post-modification relationships. Indicates user For the course Predicted score Indicates user For the course Predicted score For tolerance parameters; Construct a difficulty loss term, which is based on the following formula: in, This represents the difficulty matching loss. Represents a set of users. Represents a set of courses. Indicates user For the course Predicted score Indicates user The characteristics of the learning stages Indicates course The difficulty characteristics, It is a difficulty matching function; An educational consistency loss term is constructed based on the prior and difficulty loss terms; a multi-objective optimization loss function is constructed based on the educational consistency loss term, the user comparison loss term, and the course comparison loss term, referring to the following formula: in, This represents the loss function for multi-objective optimization. This represents the BPR loss term, used to optimize the ranking of recommended courses. and These represent the user comparison loss and the course comparison loss, respectively. This represents the educational consistency loss term, used to ensure the educational rationality of the recommendation results. , and These represent the weighting coefficients of each loss term.

8. A dual-hypergraph comparative learning curriculum recommendation device that integrates prior educational knowledge, characterized in that, The device for recommending courses based on the dual hypergraph comparison learning of prior knowledge in inclusive education includes: a memory, a processor, and a program for recommending courses based on the dual hypergraph comparison learning of prior knowledge in inclusive education stored in the memory. The processor is used to run the program for recommending courses based on the dual hypergraph comparison learning of prior knowledge in inclusive education. The program for recommending courses based on the dual hypergraph comparison learning of prior knowledge in inclusive education is configured to implement the method for recommending courses based on the dual hypergraph comparison learning of prior knowledge in inclusive education as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a course recommendation program for comparative learning of hypergraphs based on prior knowledge of inclusive education. When the course recommendation program for comparative learning of hypergraphs based on prior knowledge of inclusive education is executed by a processor, it implements the course recommendation method for comparative learning of hypergraphs based on prior knowledge of inclusive education as described in any one of claims 1 to 6.

Images