A method, system, and device for recommending learning resources
By combining retrieval enhancement generation, large language models, and graph convolutional neural networks to recommend learning resources, this method addresses the issues of insufficient personalized recommendations and inconsistent content quality in existing systems, achieving more accurate learning resource recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG NORMAL UNIV
- Filing Date
- 2025-05-30
- Publication Date
- 2026-07-17
Smart Images
Figure CN120578809B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of computer science, and more specifically, relates to a method, system, and device for recommending learning resources. Background Technology
[0002] With the rapid development of online education platforms, the emergence of massive amounts of learning resources has led to information overload for learners. Although recommendation systems alleviate the difficulty of resource selection through personalized recommendations, existing technologies still have significant limitations. Traditional recommendation models rely on historical interaction data, making it difficult to solve the cold start and data sparsity problems; deep learning models, while capable of capturing high-order interaction features, do not fully utilize semantic information; and large language models, while possessing semantic understanding capabilities, are susceptible to the "illusion" problem. Meanwhile, educational resource review still relies on manual annotation or rule-based screening, which is inefficient and struggles to identify hidden misleading content. These problems collectively result in insufficient accuracy of recommendation results and inconsistent resource quality, urgently requiring an innovative method that integrates multi-source information while also considering quality review.
[0003] Currently, the abundance of learning resources provides learners with unprecedented convenience, but it also raises a series of problems and challenges:
[0004] (1) Learners often struggle to distinguish which resources suit their learning needs and goals. Different learners have different learning styles, preferences, and objectives, but most online education platforms do not offer personalized recommendation systems, making it impossible to accurately recommend suitable resources based on learners' characteristics and needs. This forces learners to spend a significant amount of time and energy browsing, filtering, and trying to find learning resources that meet their needs, thus impacting learning efficiency and experience. Learners often spend more energy searching for and filtering resources than focusing on actual learning activities.
[0005] (2) Current online education platforms often suffer from inconsistent content quality. Faced with a large amount of learning resources, learners find it difficult to discern the credibility and authority of these resources, and are easily influenced by inaccurate or low-quality resources, thereby affecting their learning outcomes and experience. Error-laden and unverified resources can have disastrous consequences for learners, and inflammatory false information is more likely to spread rapidly on the platform, frequently interacting with learners and being learned by the recommendation model, causing the model to tend to recommend low-quality resources, resulting in a sharp drop in recommendation accuracy.
[0006] (3) Current online education platform resource recommendation systems often suffer from insufficient personalized recommendations and inaccurate recommendation algorithms. Since recommendation systems are usually based on learners' historical behavior and preferences, the accuracy and effectiveness of the recommendation system may be greatly reduced for new users or users with irregular behavior. In addition, most current recommendation systems have not yet made full use of learners' social networks and community information, and cannot provide learners with more accurate and effective resource recommendations by leveraging their social relationships and learning circles. Summary of the Invention
[0007] In view of the shortcomings of the prior art, the purpose of this application is to provide a learning resource recommendation method, system and device, which aims to solve the problem of insufficient accuracy of existing learning resource recommendation methods.
[0008] To achieve the above objectives, firstly, this application provides a learning resource recommendation method, comprising the following steps:
[0009] Obtain the user characteristics of the users to be recommended and the item characteristics of the learning resources to be recommended;
[0010] The user features and item features are input into a resource recommendation model to obtain learning resources recommended to the user. The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representations and item semantic representations combined with external knowledge. It then obtains user embedding representations and item embedding representations by combining the user-item interaction bipartite graph with a graph convolutional neural network. The corresponding embedding representations are then compared and aligned with the semantic representations combined with external knowledge. The user-item interaction bipartite graph is graph-enhanced based on subgraph contrastive learning. Finally, the user-item interaction bipartite graph is compared and aligned with the graph-enhanced graph, and a graph convolutional neural network is used to recommend items to the user.
[0011] In one possible implementation, the loss function of the resource recommendation model is the sum of the loss function of the graph convolutional neural network for recommendation, the loss function of subgraph contrastive learning, and the loss function of representation contrastive alignment.
[0012] In one possible implementation, the loss function of the resource recommendation model ;in, This represents the loss function used by the graph convolutional neural network for recommendation. This represents the loss function for subgraph contrastive learning. This represents the loss function characterizing the contrast alignment;
[0013]
[0014] in, It is the number of users. User The neighborhood, and User For items and items Predicted score It is the Sigmoid function. It is the regularization coefficient. It is the initial embedding matrix;
[0015]
[0016] in, It is a collection of users. and User The original graph structure and the graph structure after perturbation. and User The original graph structure and the graph structure after perturbation. It is a similarity function. It is a temperature parameter;
[0017]
[0018] in, For the expected value, It is a density ratio function. It is the set of all possible semantic representations of items or users. Represents items or users i semantic representation, Represents items or users j semantic representation, Represents items or users i Embedded representation.
[0019] In one possible implementation, obtaining user characteristics and item characteristics includes:
[0020] User characteristics are constructed based on the set of items that the user has interacted with, the titles of each item in the set, and the set of comments the user makes on all items.
[0021] Item features are constructed based on the item's title, original description, resource domain attributes, and user comments.
[0022] The user and item features are enhanced based on retrieval enhancement generation and a large language model, including:
[0023] The constructed user features and item features are expanded by combining preset teaching resources and knowledge graphs, and the expanded features are input into a large language model to generate user semantic representations and item semantic representations that combine external knowledge; the preset teaching resources and knowledge graphs both include information from high-quality learning resources.
[0024] In one possible implementation, user embeddings and item embeddings are obtained by combining the user-item interaction bipartite graph with a graph convolutional neural network, including:
[0025] The graph convolution is iteratively performed by a graph convolutional neural network to aggregate the features of each node's neighbors in the user-item interaction bipartite graph, thereby obtaining a new representation of the node. Then, the new representations of nodes in multiple aggregation layers are weighted and merged to obtain the embedded representation of the node; the node is a user or an item.
[0026] In one possible implementation, the representation contrast alignment includes:
[0027] By capturing similar features between the embedded representations of users or items and their corresponding semantic representations using a density ratio function, the embedding representations are aligned with their corresponding semantic representations to increase the similarity between the two.
[0028] In one possible implementation, graph augmentation is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including:
[0029] By randomly changing the graph structure of the user-item interaction bipartite graph, a subgraph structure is obtained;
[0030] By combining different subgraph structures, user embeddings and item embeddings corresponding to different subgraph structures are obtained through graph convolutional neural networks;
[0031] By comparing and learning the user embedding representations and item embedding representations corresponding to different subgraph structures, graph augmentation of the user-item interaction bipartite graph is achieved.
[0032] In one possible implementation, graph augmentation is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including:
[0033] By combining the user-item interaction bipartite graph, noise vectors are introduced into different aggregation layers of the graph convolutional neural network to obtain the corresponding user embedding representation and item embedding representation;
[0034] Different noise vectors are introduced into different aggregation layers to obtain different graph convolutional neural networks. The user embedding representations and item embedding representations corresponding to the different graph convolutional neural networks are compared and learned to achieve graph augmentation of the user-item interaction bipartite graph.
[0035] Secondly, this application provides a learning resource recommendation system, including:
[0036] The feature acquisition module is used to acquire the user features of the user to be recommended and the item features corresponding to the learning resources to be recommended;
[0037] The resource recommendation module is used to input the user features and item features into the resource recommendation model to obtain learning resources recommended for the user. The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representations and item semantic representations combined with external knowledge. It then obtains user embedding representations and item embedding representations by combining the user-item interaction bipartite graph with a graph convolutional neural network. After that, the corresponding embedding representations are compared and aligned with the semantic representations combined with external knowledge. The user-item interaction bipartite graph is graph-enhanced based on subgraph contrastive learning. Then, the user-item interaction bipartite graph is compared and aligned with the graph-enhanced graph, and a graph convolutional neural network is used to recommend items to the user.
[0038] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.
[0039] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.
[0040] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.
[0041] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art:
[0042] This application provides a learning resource recommendation method, system, and device. Based on retrieval enhancement generation technology and a large language model, it enhances the semantic features of users and items. It combines a user-item interaction bipartite graph with a graph convolutional neural network to obtain user and item embedding representations. Then, it compares and aligns the corresponding embedding representations with the semantic representations after incorporating external knowledge to reduce the impact of inaccurate or low-quality learning resources. Based on subgraph contrastive learning, it performs graph enhancement on the user-item interaction bipartite graph to fully consider the learning styles, preferences, and goals of different learners. Finally, combining the aligned and graph-enhanced user-item interaction bipartite graph, a graph convolutional neural network is used to recommend items to the user, achieving personalized learning resource recommendations and improving recommendation accuracy. Attached Figure Description
[0043] Figure 1 A flowchart of a learning resource recommendation method provided in this application embodiment;
[0044] Figure 2 This is an overall architecture diagram of the Graph RAG model provided in the embodiments of this application;
[0045] Figure 3 This is a visualization result of the knowledge graph provided in an embodiment of this application;
[0046] Figure 4 This is an overall architecture diagram of the RHCLRec model provided in the embodiments of this application;
[0047] Figure 5 This is a schematic diagram of on-screen comparison learning provided for an embodiment of this application;
[0048] Figure 6 A comparison diagram of edge reduction and random walk graph enhancement methods provided in the embodiments of this application;
[0049] Figure 7 This is a schematic diagram of a noise-adding image enhancement method provided in an embodiment of this application;
[0050] Figure 8 This is a diagram illustrating the architecture of a learning resource recommendation system provided in an embodiment of this application.
[0051] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0053] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.
[0054] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0055] The embodiments of this application are described below with reference to the accompanying drawings.
[0056] Figure 1 A flowchart of the learning resource recommendation method provided in the embodiments of this application is shown below. Figure 1 As shown, it includes the following steps:
[0057] Step S101: Obtain the user characteristics of the user to be recommended and the item characteristics corresponding to the learning resources to be recommended;
[0058] Step S102: Input the user features and item features into the resource recommendation model to obtain learning resources recommended for the user; The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representation and item semantic representation after combining external knowledge. Combine the user-item interaction bipartite graph and obtain user embedding representation and item embedding representation through graph convolutional neural network. Then, the corresponding embedding representation is compared and aligned with the semantic representation after combining external knowledge. Graph enhancement is performed on the user-item interaction bipartite graph based on subgraph contrastive learning. Then, combining the user-item interaction bipartite graph after representation comparison alignment and graph enhancement, a graph convolutional neural network is used to recommend items to the user to be recommended.
[0059] It should be noted that the resource recommendation model mentioned in this application can be abbreviated as RHCLRec (RAG GraphHybrid Contrastive Learning Recommendation) model. The RHCLRec model mainly includes: semantic feature enhancement based on large language models and retrieval-augmented generation (RAG) technology; recommendation based on graph neural networks through multi-layer graph convolution stacking; comparison and alignment of semantic features enhanced by large language models with user resource interaction features; and graph feature enhancement based on subgraph contrastive learning.
[0060] Among them, semantic feature enhancement is achieved based on knowledge graph and retrieval-augmented generation (RAG) techniques combined with a large language model (LLM), namely Graph-RAG, such as... Figure 2 As shown, it includes:
[0061] (1) Semantic Enhancement Prompt Construction
[0062] Constructing user feature input Q by combining user-item interaction graph u Input Q with item characteristics i .
[0063] Specifically, when constructing the input of item features, the item... Text information is divided into four types: title Original description Specific attributes of the resource field and a collection of user comments Based on this semantic information, Q is used to generate the item description input. i Describe using the following formula:
[0064]
[0065]
[0066] Used in data processing This process handles the textual features of all resources and integrates (concatenates) them into a string that comprehensively describes the item information. When the original item description is missing, a subset R' of comments is randomly sampled from the comment set and combined with other attributes as input to the retrieval layer. By combining item descriptions and user comments, accurate information is provided to the retrieval generation model, helping the model generate item semantic representations that adequately reflect resource characteristics.
[0067] To generate user feature input, given that item features have already been generated, the interaction information between the user and items is fully utilized. Specifically, the set of items that have interacted with user u is denoted as... Simultaneously, subsets are obtained through random sampling. .for For each item v in the list, its text attributes are combined and concatenated into a single string. ,in Comments representing user u The prompt word indicates item v. The final input is Q. u Describe using the following formula:
[0068]
[0069] in, Purpose and Both are similar in that they combine text content into a continuous string. Text attributes Including user reviews can provide a true reflection of user preferences. This construction method provides support for understanding users' true preferences.
[0070] (2) Enhanced search generation
[0071] Input user characteristics into Q u Input Q with item characteristics i The input is fed into the encoder, and the feature vector q(x) is obtained by retrieving the encoder. Simultaneously, for teaching resources... The feature vector d(z) is obtained through a resource encoder, and the Maximum Inner Product Search (MIPS) method is used to retrieve the top k relevant resources to obtain the top k teaching resources. subset On the other hand, we will use entity vectors from the knowledge graph. Similarity calculation is performed on the vector q(x) obtained by the encoder, which is based on the user and item feature inputs. Get similar entity set And search to obtain the set r of edges corresponding to the entity, and combine the entity description and edge set information with the teaching resource subset. The concatenated information is provided as context to the Large Language Model (LLM) generator. The LLM generator combines the knowledge in the model parameters with the educational resources and entity relationship information provided by the retrieval model to generate semantic representations of users and items, thereby enhancing the semantic features of items and users.
[0072] The aforementioned teaching resources and knowledge graphs serve as reference materials. Teaching resources include multiple text resources representing users and items, while the knowledge graph includes graphs representing user and item attributes and relationships. The users and items mentioned are commonly used in recommendation systems; for example, they may include high-quality user and item information, as well as users' social networks and community information, to serve as recommended learning resources, fully leveraging the ability of the large language model to distinguish low-quality resources. Here, "high-quality" refers to high credibility, high authority, or verification.
[0073] It's important to note that there are currently several ways to utilize large language models. One intuitive approach is to directly use the semantic representations of users and items, the recommendation task description, and output constraints as input. Leveraging the large model's ability to learn contextual information, the recommendation task is transformed into a text generation task. Then, the generated text is parsed to obtain the recommendation set. This approach, using a large language model as the core of the recommendation model, can be considered a content-based recommendation method. However, this method focuses on the similarity of user and item attribute features while completely ignoring the interaction information between users and items. The recommendations lack novelty and heavily rely on the quality of the item and user representations.
[0074] Optimizing the task description in the input of a large language model can improve this, but a more serious challenge is that large language models are prone to "illusion" and recommend non-existent items. Furthermore, the limitations imposed by large models on the input limit restrict the size of the input item set, while the actual item set in recommendation tasks far exceeds the input limit of large models. Another approach is to have the original recommendation model generate a candidate set of items, which is then fed back to the large model for re-ranking. In a large language model re-ranking experiment, as the number of candidates increased, several metrics after re-ranking, including Recall@10, Recall@20, and NDCG@10, showed a significant decline, far lower than traditional recommendation models. This may be due to the large language model's attention being diverted to the numerous item descriptions within long texts.
[0075] In this application, semantic enhancement is performed when constructing both item feature inputs and user feature inputs. The semantic enhancement fully utilizes the information of the items, and the user feature inputs fully utilize the interaction information between the user and the items to semantically enhance the features of the items and users. Then, when the semantically enhanced item and user features are input into the LLM, the semantic representations of the items and users are generated in conjunction with reference materials. This is to avoid the influence of inaccurate or low-quality resources through semantic enhancement and reference materials, improve the credibility and authority of the semantic representations, make the model more inclined to recommend high-quality resources, and improve the recommendation accuracy.
[0076] (3) Knowledge graph construction
[0077] The main steps in constructing the knowledge graph mentioned above include: building a large language model, text splitting, information extraction prompt construction, large model text generation, triple extraction, and finally visualizing the knowledge graph.
[0078] After splitting the knowledge text using a text segmentation model, the template extracted from the knowledge graph triples is combined with the split text and then fed into a large language model. First, a prompt describing the extraction task is constructed. The prompt text first describes the extraction task in detail and defines the entity relations in the triples. Then, three examples of text entity relation extraction are shown in a few-shot format, and a standard output format is clearly defined. Special format symbols are used to separate the triples to aid in subsequent text segmentation based on the large language model's response.
[0079] Finally, the triple extraction task is transformed into a large language model text generation task to obtain the generated extraction results. The entity relations set is then obtained through text parsing. In one example, the visualization result is as follows: Figure 3 As shown.
[0080] The RHCLRec model provided in this application embodiment is as follows: Figure 4 As shown, after semantic enhancement of users and items is achieved through the above Graph-RAG model, the following three operations are performed:
[0081] Firstly, recommendation based on graph neural networks, achieved through multi-layer graph convolution stacking, includes:
[0082] For bipartite graph-structured datasets such as user-item interactions, Graph Convolutional Networks (GCNs) are well-suited for learning user and item features. The fundamental idea of GCNs is to learn node representations from graph features. To achieve this, GCNs iteratively perform graph convolutions, aggregating the features of neighbors into a new representation of the target node. This neighborhood aggregation can be abstracted as:
[0083]
[0084] In graph convolution, AGG (aggregation function) is the core function, aiming to integrate the representations of the target node and its neighbors in the k-th layer. Many studies have improved models by refining the aggregation function, such as the weighted aggregation method in GIN and the LSTM aggregation method in GraphSAGE. In recommendation models, the weighted aggregation method is used to define the convolution operation as follows:
[0085]
[0086]
[0087] in, This represents the user's feature vector in the k-th layer graph convolution. Let be the vector representation of the items in the k-th layer graph convolution. Represents the neighborhood of user node u. For the neighborhood of item i, the normalization term The size of the suppressed node feature vector increases with the number of graph convolutional layers. In the aggregation layer, the lowest-level embeddings, including user embeddings, are trained. Then, higher-level representations are calculated using graph convolution. After k layers of graph convolution, the representations from all layers are merged, resulting in the final representation:
[0088]
[0089]
[0090] in, The model parameters representing the k-th layer are used to synthesize feature representations from multiple layers, allowing the model to learn high-order features captured by higher-level convolutions while suppressing excessive smoothing of high-order embeddings caused by stacked convolutional layers. The final user feature is obtained by weighted fusion of the representations from each layer. and item characteristics The user-item interaction matrix is as follows: Where M and N are the number of users and items, respectively, and if user u interacts with item i, then... If it is, then it is denoted as 1; otherwise, it is denoted as 0. This yields the user-item adjacency matrix:
[0091]
[0092] The lowest level embedding matrix is Where T is the embedding dimension, the matrix operation of graph convolution can be represented as:
[0093]
[0094] Where D is a (M+N)×(M+N) diagonal matrix, The number of non-zero terms in the i-th row vector of the adjacency matrix A is used to represent the final characteristic matrix as follows:
[0095]
[0096] Secondly, by comparing and aligning semantic features enhanced by a large language model with user resource interaction features, the following aspects are achieved:
[0097] User descriptions after data augmentation With item description Based on user preferences, semantic representations of user and item descriptions are obtained through text encoding. The text input is transformed into a fixed-length feature vector while preserving semantic and contextual information. This is done after obtaining the user embedding representation learned by the aggregation model. and item embedding representation Based on this, define the density ratio function. Capture semantic representation And graph aggregation representation Similar features between data help mitigate the impact of data quality and noise in representation learning. This is achieved using contrastive alignment. The definition is as follows:
[0098]
[0099] Function This represents the cosine similarity, while This indicates that the semantic representation will be... Mapped to A multilayer perceptron in the feature space. In contrast alignment, we will... and These are treated as positive sample pairs. During the learning process, these sample pairs influence each other to align their representations. The actual training model aims to reduce the distance between positive sample pairs while treating the remaining samples as negative samples. By combining ID-based graph aggregation learning representations with text-based behavioral semantics, a contrastive alignment method effectively aligns the semantic features of external knowledge with user-item interaction features.
[0100] Thirdly, graph feature enhancement methods based on subgraph contrastive learning include:
[0101] Besides feature augmentation of user and item semantic representations, graph features should also be further improved to enhance graph representation performance and model robustness. The distribution of node degree in the bipartite graph of user-item interactions obtained from historical interaction data is not uniform. For user or item nodes with low degree (i.e., fewer interactions), graph neural networks struggle to learn node features of less popular items. To address the impact of introducing active users and popular items, this study introduces graph data augmentation methods. The most widely used method for graph data augmentation is contrastive learning, which involves randomly changing the graph structure to form subgraphs and then augmenting the graph through subgraph comparison. Graph contrastive learning methods are as follows... Figure 5 As shown.
[0102] For ease of description, we simplify the operations of the aggregation layer as follows:
[0103]
[0104] in This represents the node of the next higher level. Represents the original graph structure. The representation vector representing this layer is used to prune edges and nodes during aggregation to modify the graph structure. Perform dropout operation to form a subgraph and Represented as:
[0105]
[0106]
[0107] function and Representing two completely independent dropout methods, the node deletion method probabilistically removes a node and the edges connected to that node, as follows:
[0108]
[0109] in The mask vectors are generated according to a Bernoulli distribution, and each vector is independently convolved to obtain a new node representation. and Edge removal uses a probabilistic method to remove some edges, similar to the node removal method.
[0110]
[0111] Adjusting the graph structure in each convolutional layer can achieve an effect similar to random walks, creating mask vectors that differentiate between layers. The difference between random walks and edge reduction is as follows: Figure 6 As shown.
[0112] These methods all achieve graph augmentation by modifying the graph structure. The effectiveness of graph augmentation depends on the interaction graph features used for training and the distribution of the generated subgraphs used for contrastive learning. Obtaining a more uniformly distributed representation space by modifying the graph structure is difficult and time-consuming; modifying the embedding space during aggregation is a more feasible approach. Another method is to achieve graph augmentation without changing the graph structure, but by adding small perturbations to the representation vectors during propagation. Specifically, random noise is directly added to the representation vectors to achieve efficient graph augmentation. For a given node... Its corresponding dimensional space representation Representational enhancements can be achieved in the following ways:
[0113]
[0114] in and For a noise vector, the dimensional space representation is... Adding a noise vector during each aggregation layer yields the following results. and The noise vector satisfies the following constraints:
[0115]
[0116]
[0117] in The first constraint is used to control the noise vector. Granularity, ensuring Numerically equivalent to Let be a point on the hypersphere with radius . The second constraint is used to avoid adding noise that could cause semantic features to deviate and reduce the number of effective positive samples. Adding a noise vector This can be viewed as rotating the vector space by a certain angle; controlling the magnitude of the rotation angle helps preserve the original feature information while simultaneously enhancing features. Furthermore, the random noise is different for each node. Graph augmentation methods that add noise include... Figure 7 As shown.
[0118] In the actual construction of the recommendation model, both graph augmentation methods that change the graph structure and augmentation methods that add noise improved the model performance in the experiments.
[0119] Furthermore, the loss function of the above RHCLRec model for:
[0120]
[0121] The model prediction part uses the Bayesian Personalized Ranking (BPR) loss function, expressed as follows: The BPR loss function is used to capture users' preferences for items, and it is calculated as follows:
[0122]
[0123] in, It is the number of users. User The neighborhood, and User For items and Predicted score It is the Sigmoid function. These are regularization coefficients used to prevent embedding into the matrix. Overfitting It is the initial embedding matrix.
[0124] The graph compares the loss function of the learning part. It is used to capture the structural information of users and items on the graph, and its calculation method is as follows:
[0125]
[0126] in, It is a collection of users. and User The original graph structure and the graph structure after perturbation. It is a similarity function. It is a temperature parameter used to control the contrast intensity during the contrast learning process.
[0127] Characterizing the contrast alignment loss function It is used to capture semantic information about users and items, and its calculation method is as follows:
[0128]
[0129] in, For the expected value, It is a density ratio function. It is the set of all possible semantic representations.
[0130] The three loss functions mentioned above jointly guide model learning, and through optimization... To improve the recommendation performance of the model.
[0131] Figure 8 An architecture diagram of a learning resource recommendation system provided in this application embodiment is shown below. Figure 8 As shown, it includes:
[0132] The feature acquisition module 810 is used to acquire the user features of the user to be recommended and the item features corresponding to the learning resources to be recommended.
[0133] The resource recommendation module 820 is used to input the user features and item features into the resource recommendation model to obtain learning resources recommended for the user. The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representations and item semantic representations combined with external knowledge. It then obtains user embedding representations and item embedding representations by combining the user-item interaction bipartite graph with a graph convolutional neural network. After that, it performs representation comparison and alignment with the semantic representations combined with external knowledge. It performs graph enhancement on the user-item interaction bipartite graph based on subgraph contrastive learning. Then, it uses a graph convolutional neural network to recommend items to the user after combining the representation comparison and alignment with the graph enhancement of the user-item interaction bipartite graph.
[0134] It should be understood that the above system is used to execute the methods in the above embodiments. The corresponding program modules in the system are similar in implementation principle and technical effect to those described in the above methods. The working process of the system can be referred to the corresponding process in the above methods, and will not be repeated here.
[0135] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute the methods in the above embodiments.
[0136] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.
[0137] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.
[0138] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.
[0139] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.
[0140] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.
[0141] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).
[0142] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.
[0143] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A learning resource recommendation method, characterized in that, Includes the following steps: Obtain the user characteristics of the users to be recommended and the item characteristics of the learning resources to be recommended; The user features and item features are input into the resource recommendation model to obtain learning resources recommended for the user. The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representation and item semantic representation after combining external knowledge. The user-item interaction bipartite graph is combined with graph convolutional neural network to obtain user embedding representation and item embedding representation. Then, the corresponding embedding representation is compared and aligned with the semantic representation after combining external knowledge. Graph augmentation is performed on the user-item interaction bipartite graph based on subgraph contrastive learning. Then, the user-item interaction bipartite graph is combined with the representation contrast-aligned and graph-augmented graph, and a graph convolutional neural network is used to recommend items to the user. Obtaining user and item characteristics includes: User characteristics are constructed based on the set of items that the user has interacted with, the titles of each item in the set, and the set of comments the user makes on all items. Item features are constructed based on the item's title, original description, resource domain attributes, and user comments. The user and item features are enhanced based on retrieval enhancement generation and a large language model, including: The constructed user and item features are expanded by combining preset teaching resources and knowledge graphs, and the expanded features are input into a large language model to generate user semantic representations and item semantic representations that incorporate external knowledge; the preset teaching resources and knowledge graphs both include information from high-quality learning resources. By combining the user-item interaction bipartite graph, a graph convolutional neural network is used to obtain user embedding representations and item embedding representations, including: The graph convolution is iteratively performed by a graph convolutional neural network to aggregate the features of the neighbors of each node in the user-item interaction bipartite graph to obtain a new representation of the node. Then, the new representations of the nodes in multiple aggregation layers are weighted and merged to obtain the embedded representation of the node; the node is a user or an item. The characterization contrast alignment includes: The embedding representation of a user or item is captured by a density ratio function to capture similar features between the embedding representation and the corresponding semantic representation. The goal is to increase the similarity between the embedding representation and the semantic representation and align the embedding representation with the corresponding semantic representation. Graph augmentation of the user-item interaction bipartite graph based on subgraph contrastive learning includes: By randomly changing the graph structure of the user-item interaction bipartite graph, a subgraph structure is obtained; By combining different subgraph structures, user embeddings and item embeddings corresponding to different subgraph structures are obtained through graph convolutional neural networks; By comparing and learning the user embedding representations and item embedding representations corresponding to different subgraph structures, graph augmentation of the user-item interaction bipartite graph is achieved.
2. The method according to claim 1, characterized in that, The loss function of the resource recommendation model is the sum of the loss function of the graph convolutional neural network for recommendation, the loss function of subgraph contrastive learning, and the loss function of representation contrastive alignment.
3. The method according to claim 1, characterized in that, Loss function of resource recommendation model ;in, This represents the loss function used by the graph convolutional neural network for recommendation. This represents the loss function for subgraph contrastive learning. This represents the loss function characterizing the contrast alignment; in, It is the number of users. User The neighborhood, and User For items and items Predicted score It is the Sigmoid function. It is the regularization coefficient. It is the initial embedding matrix; in, It is a collection of users. and User The original graph structure and the graph structure after perturbation. and User The original graph structure and the graph structure after perturbation. It is a similarity function. It is a temperature parameter; in, As the expected value, It is a density ratio function. It is the set of all possible semantic representations of items or users. Represents items or users i semantic representation, Represents items or users j semantic representation, Represents items or users i Embedded representation.
4. The method according to claim 1, characterized in that, Graph augmentation of the user-item interaction bipartite graph based on subgraph contrastive learning includes: By combining the user-item interaction bipartite graph, noise vectors are introduced into different aggregation layers of the graph convolutional neural network to obtain the corresponding user embedding representation and item embedding representation; Different noise vectors are introduced into different aggregation layers to obtain different graph convolutional neural networks. The user embedding representations and item embedding representations corresponding to the different graph convolutional neural networks are compared and learned to achieve graph augmentation of the user-item interaction bipartite graph.
5. A learning resource recommendation system, characterized in that, include: The feature acquisition module is used to acquire the user features of the user to be recommended and the item features corresponding to the learning resources to be recommended; The resource recommendation module is used to input the user features and item features into the resource recommendation model to obtain learning resources recommended for the user. The resource recommendation model enhances the user features and item features based on retrieval enhancement generation and a large language model to obtain user semantic representation and item semantic representation after combining external knowledge. It obtains user embedding representation and item embedding representation through graph convolutional neural network by combining user-item interaction bipartite graph. Then, the corresponding embedding representation is compared and aligned with the semantic representation after combining external knowledge. Graph augmentation is performed on the user-item interaction bipartite graph based on subgraph contrastive learning. Then, the user-item interaction bipartite graph is combined with the representation contrast-aligned and graph-augmented graph, and a graph convolutional neural network is used to recommend items to the user. Obtaining user and item characteristics includes: User characteristics are constructed based on the set of items that the user has interacted with, the titles of each item in the set, and the set of comments the user makes on all items. Item features are constructed based on the item's title, original description, resource domain attributes, and user comments. The user and item features are enhanced based on retrieval enhancement generation and a large language model, including: The constructed user and item features are expanded by combining preset teaching resources and knowledge graphs, and the expanded features are input into a large language model to generate user semantic representations and item semantic representations that incorporate external knowledge; the preset teaching resources and knowledge graphs both include information from high-quality learning resources. By combining the user-item interaction bipartite graph, a graph convolutional neural network is used to obtain user embedding representations and item embedding representations, including: The graph convolution is iteratively performed by a graph convolutional neural network to aggregate the features of the neighbors of each node in the user-item interaction bipartite graph to obtain a new representation of the node. Then, the new representations of the nodes in multiple aggregation layers are weighted and merged to obtain the embedded representation of the node; the node is a user or an item. The characterization contrast alignment includes: The embedding representation of a user or item is captured by a density ratio function to capture similar features between the embedding representation and the corresponding semantic representation. The goal is to increase the similarity between the embedding representation and the semantic representation and align the embedding representation with the corresponding semantic representation. Graph augmentation of the user-item interaction bipartite graph based on subgraph contrastive learning includes: By randomly changing the graph structure of the user-item interaction bipartite graph, a subgraph structure is obtained; By combining different subgraph structures, user embeddings and item embeddings corresponding to different subgraph structures are obtained through graph convolutional neural networks; By comparing and learning the user embedding representations and item embedding representations corresponding to different subgraph structures, graph augmentation of the user-item interaction bipartite graph is achieved.
6. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to perform the method as described in any one of claims 1-4.