Learning resource recommendation method, system and equipment

Abstract

The invention belongs to the field of computers, and particularly discloses a learning resource recommendation method, system and equipment, and the method comprises the steps: obtaining user features and article features corresponding to a to-be-recommended learning resource; inputting the user features and the article features into a resource recommendation model to obtain learning resources recommended for the user; the resource recommendation model enhances the user features and the article features based on retrieval enhancement generation and a large language model to obtain user semantic representation and article semantic representation, and obtains user embedded representation and article embedded representation through a graph convolutional neural network in combination with a user-article interaction bipartite graph; carrying out representation comparison and alignment on the embedded representation and the semantic representation; and performing graph enhancement on the user-article interaction bipartite graph based on subgraph comparative learning, and then performing article recommendation on a to-be-recommended user by adopting a graph convolutional neural network in combination with the user-article interaction bipartite graph which represents comparative alignment and is subjected to graph enhancement. Through the method and the device, more accurate learning resource recommendation is realized.

Description

Technical Field

[0001] The present application relates to the field of computers, and more specifically, to a method, system, and device for recommending learning resources. Background Art

[0002] With the rapid development of online education platforms, the emergence of massive learning resources has led to information overload for learners. Although recommendation systems have alleviated the difficulty of resource screening through personalized recommendations, existing technologies still have significant limitations. Traditional recommendation models rely on historical interaction data and have difficulty solving cold start and data sparsity problems; although deep learning models can capture high-order interaction features, they make insufficient use of semantic information; and although large language models have semantic understanding capabilities, they are susceptible to "hallucination" problems. At the same time, educational resource review still relies on manual labeling or rule-based screening, which is inefficient and difficult to identify hidden misleading content. These problems collectively lead to insufficient accuracy in recommendation results and uneven resource quality. There is an urgent need for an innovative method that integrates multi-source information and takes quality review into consideration.

[0003] Currently, the abundance of learning resources provides learners with unprecedented convenience, but it also raises a series of problems and challenges: (1) Learners often find it difficult to distinguish which resources are suitable for their learning needs and learning goals. Different learners have different learning styles, learning preferences, and learning goals, and most online education platforms do not provide personalized recommendation systems and are unable to accurately recommend suitable resources based on learners' characteristics and needs. This results in learners having to spend a lot of time and energy browsing, screening, and trial and error to find learning resources that meet their needs, which affects their learning efficiency and learning experience. Learners often spend more energy in the process of finding and screening resources rather than focusing on actual learning activities.

[0004] (2) Current online education platforms often have the problem of uneven content quality. When faced with a large number of learning resources, learners find it difficult to discern the credibility and authority of resources and are easily influenced by inaccurate or low-quality resources, which in turn affects learners' learning outcomes and experience. Errors in unreviewed resources can have disastrous consequences for learners. Inciting false information is more likely to spread rapidly on the platform, frequently interact with learners, and be learned by the recommendation model, causing the recommendation model to tend to recommend low-quality resources, and the recommendation accuracy rate to drop sharply.

[0005] (3) Current resource recommendation systems on online education platforms often suffer from issues such as insufficient personalized recommendations and inaccurate recommendation algorithms. Since recommendation systems are usually based on learners’ historical behavior and preferences, the accuracy and effectiveness of recommendation systems may be significantly reduced for new users or users with irregular behavior. Furthermore, most current recommendation systems fail to fully utilize learners’ social networks and community information, and are unable to leverage learners’ social relationships and learning circles to provide more accurate and effective resource recommendations. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this application is to provide a learning resource recommendation method, system and device, aiming to solve the problem of insufficient accuracy of existing learning resource recommendation methods.

[0007] To achieve the above objectives, in a first aspect, the present application provides a learning resource recommendation method, comprising the following steps: Obtain the user characteristics of the user to be recommended and the item characteristics corresponding to the learning resource to be recommended; The user features and item features are input into a 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, and obtains user embedding representations and item embedding representations through a graph convolutional neural network in combination with a user-item interaction bipartite graph, and then performs representation comparison and alignment on the corresponding embedding representations and the semantic representations combined with external knowledge; the user-item interaction bipartite graph is enhanced based on subgraph comparison learning, and then the user-item interaction bipartite graph after representation comparison and alignment and graph enhancement is combined with a graph convolutional neural network to recommend items to the recommended user.

[0008] 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 the subgraph comparative learning, and the loss function of the representation comparative alignment.

[0009] In one possible implementation, the loss function of the resource recommendation model is ;in, Represents the loss function of the graph convolutional neural network for recommendation, represents the loss function of subgraph contrastive learning, represents the loss function for characterizing contrastive alignment;

[0010] in, is the number of users, is a user Neighborhood, and is a user For items and items The prediction score of is the Sigmoid function, is the regularization coefficient, is the initial embedding matrix;

[0011] in, is a collection of users, and is a user The original graph structure and the perturbed graph structure, and is a user The original graph structure and the perturbed graph structure, is the similarity function, is the temperature parameter;

[0012] in, is the expected value, is the density ratio function, is the set of all possible item semantic representations or user semantic representations, Indicates an item or user i The semantic representation of Indicates an item or user j The semantic representation of Indicates an item or user i Embedding representation of .

[0013] In one possible implementation, obtaining user features and item features includes: Construct user features based on the set of items that the user interacts with, the titles of each item in the item set, and the set of user comments on all items; Construct item features based on the item title, original item description, resource domain attributes of the item, and user comments on the item; The user features and item features are enhanced based on retrieval enhancement generation and a large language model, including: The constructed user features and item features are expanded by combining them with 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 combined with external knowledge; the preset teaching resources and knowledge graphs both include information on high-quality learning resources.

[0014] In one possible implementation, a bipartite graph of user-item interactions is combined with a graph convolutional neural network to obtain user embedding representations and item embedding representations, including: Graph convolution is iteratively performed through a graph convolutional neural network to aggregate the features of each node's neighbors in the user-item interaction bipartite graph to obtain a new representation of the node. The new representations of the nodes in multiple aggregation layers are then weighted and merged to obtain an embedded representation of the node; the node is a user or an item.

[0015] In a possible implementation, the characterization comparison and alignment includes: The density ratio function is used to capture the similarities between the embedding representation of the user or item and the corresponding semantic representation, and the embedding representation is aligned with the corresponding semantic representation with the goal of increasing the similarity between the embedding representation and the semantic representation.

[0016] In one possible implementation, graph enhancement is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including: Obtaining a subgraph structure by randomly changing the graph structure of the user-item interaction bipartite graph; Combining different subgraph structures, we use graph convolutional neural networks to obtain user embedding representations and item embedding representations corresponding to different subgraph structures. The user embedding representations and item embedding representations corresponding to different subgraph structures are compared and learned to achieve graph enhancement of the user-item interaction bipartite graph.

[0017] In one possible implementation, graph enhancement is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including: Combined with the user-item interaction bipartite graph, noise vectors are introduced into different aggregation layers of the graph convolutional neural network to obtain corresponding user embedding representations and item embedding representations; 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 enhancement of the user-item interaction bipartite graph.

[0018] In a second aspect, the present application provides a learning resource recommendation system, comprising: A feature acquisition module is used to acquire user features of the user to be recommended and item features corresponding to the learning resource to be recommended; A resource recommendation module is used to input the user features and item features into a 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, and obtains user embedding representations and item embedding representations through a graph convolutional neural network in combination with a user-item interaction bipartite graph, and then performs representation comparison and alignment on the corresponding embedding representations and the semantic representations combined with external knowledge; the user-item interaction bipartite graph is enhanced based on subgraph comparison learning, and then the user-item interaction bipartite graph after representation comparison and alignment and graph enhancement is combined with a graph convolutional neural network to recommend items to the recommended user.

[0019] In a third aspect, the present application provides an electronic device comprising: at least one memory for storing programs; and at least one processor for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation of the first aspect.

[0020] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method described in the first aspect or any possible implementation of the first aspect.

[0021] In a fifth aspect, the present application provides a computer program product, which, when executed on a processor, enables the processor to execute the method described in the first aspect or any possible implementation of the first aspect.

[0022] In general, the above technical solutions conceived by this application have the following beneficial effects compared with the existing technologies: The present application provides a learning resource recommendation method, system and device, which enhance the semantic features of users and items based on retrieval enhancement generation technology and a large language model, obtain user embedding representations and item embedding representations through a graph convolutional neural network in combination with a user-item interaction bipartite graph, and then perform representation comparison and alignment of the corresponding embedding representations with the semantic representation after combining external knowledge to reduce the impact of inaccurate or low-quality learning resources; based on subgraph comparative learning, the user-item interaction bipartite graph is enhanced to fully consider the learning styles, learning preferences and learning goals of different learners, and then, combined with the user-item interaction bipartite graph after representation comparison and alignment and graph enhancement, a graph convolutional neural network is used to recommend items to the recommended users, thereby realizing personalized learning resource recommendations and improving recommendation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1Flowchart of the learning resource recommendation method provided in the embodiment of this application; Figure 2 This is the overall architecture diagram of the Graph RAG model provided in the embodiments of this application; Figure 3 A visualization result diagram of the knowledge graph provided in the embodiment of this application; Figure 4 The overall architecture diagram of the RHCLRec model provided in the embodiment of the present application; Figure 5 A schematic diagram of comparative learning provided in an embodiment of the present application; Figure 6 A comparison chart of the edge deletion and random walk graph enhancement methods provided in the embodiments of the present application; Figure 7 A schematic diagram of a method for image enhancement by adding noise provided in an embodiment of the present application; Figure 8 This is a diagram of the learning resource recommendation system architecture provided in the embodiment of the present application; Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0025] The term "and / or" as used herein describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. The symbol " / " as used herein indicates that the related objects are in an "or" relationship, for example, A / B means either A or B.

[0026] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0027] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application.

[0028] Figure 1 The flow chart of the learning resource recommendation method provided in the embodiment of this application is as follows: Figure 1 As shown, the following steps are included: Step S101, obtaining user characteristics of the user to be recommended and item characteristics corresponding to the learning resource to be recommended; In step S102, the user features and item features are input into a 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 combined with external knowledge, and obtains user embedding representation and item embedding representation through a graph convolutional neural network in combination with a user-item interaction bipartite graph, and then performs representation comparison and alignment on the corresponding embedding representation and the semantic representation combined with external knowledge; the user-item interaction bipartite graph is enhanced based on subgraph comparison learning, and then the user-item interaction bipartite graph after representation comparison and alignment and graph enhancement is combined with the user-item interaction bipartite graph, and a graph convolutional neural network is used to recommend items to the recommended user.

[0029] It should be noted that the resource recommendation model mentioned in this application can be abbreviated as the RHCLRec (RAG Graph Hybrid Contrastive Learning Recommendation) model. The RHCLRec model mainly includes: semantic feature enhancement based on a large language model and Retrieval-augmented Generation (RAG) technology, graph neural network-based recommendation through multi-layer graph convolution stacking, comparative alignment of semantic features enhanced by the large language model with user-resource interaction features, and on-graph feature enhancement based on subgraph comparative learning.

[0030] Among them, Graph-RAG is based on the knowledge graph and retrieval-augmented generation (RAG) technology and combined with the large language model (LLM) to achieve semantic feature enhancement. Figure 2 Shown, including: (1) Semantic Enhancement Prompt Construction Combine the user-item interaction graph to construct user feature input Q u and item feature input Q i .

[0031] Specifically, when constructing item feature input, the item The text information is divided into four types: title , original description , resource domain specific attributes and a collection of user comments Based on these semantic information, Q is used to generate the item description input iUse the following formula to describe it:

[0032]

[0033] Used in data processing The text features of all resources are processed and integrated (through text concatenation) into a string that comprehensively describes the item information. When the original item description is missing, a subset of reviews, R', is randomly sampled from the review set and combined with other attributes as input to the retrieval layer. By combining data such as item descriptions and user reviews, the retrieval generation model is provided with accurate information, ensuring that the semantic representation of the item generated by the model adequately reflects the characteristics of the resource.

[0034] In order to generate user feature input, we make full use of the interaction information between users and items, given that item features have been generated. Specifically, the set of items that user u has interacted with is denoted as At the same time, a subset is obtained by random sampling .for For each item v in , its text attributes are integrated and spliced ​​into ,in Represents the comment of user u, The prompt word for item v. The final input Q u Use the following formula to describe it:

[0035] in, The purpose and Similar to the above, both integrate text content into a continuous string. Including user comments can truly reflect user tendencies. This construction method provides support for understanding users' true preferences.

[0036] (2) Retrieval Enhancement Generation Input user features into Q u and item feature input Q i Input to the encoder, and obtain the feature vector q(x) by retrieving the encoder. At the same time, for teaching resources The feature vector d(z) is obtained through the resource encoder, and the maximum inner product search method (MIPS) is used to retrieve the top k relevant resources to obtain the top k teaching resources. Subset On the other hand, we will use the entity vector in the knowledge graph Calculate similarity with the vector q(x) obtained by the encoder with respect to the user and item features Get similar entity sets , and search for the set of edges corresponding to the entity r, and combine the entity description and edge set information with the teaching resource subset After concatenation, the context is provided 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 user semantic representations and item semantic representations, thereby enhancing the semantic features of items and users.

[0037] The aforementioned teaching resources and knowledge graph serve as reference materials. Teaching resources include multiple text resources representing users and items, and the knowledge graph includes a graph representing the attributes and relationships between users and items. The aforementioned users and items are those commonly involved in recommendations, and may include, for example, high-quality user and item information, as well as information about the user's social network and community, to serve as recommended learning resources, fully leveraging the ability of large language models to distinguish low-quality resources. High quality refers to high credibility, high authority, or audited content.

[0038] It's important to note that there are currently multiple ways to utilize large language models. One intuitive approach directly takes 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, this approach transforms the recommendation task into a text generation task. The generated text is then parsed to generate a set of recommendations. This approach, which uses a large language model as the core of the recommendation model, can be considered a content-based recommendation method. This approach focuses on the similarity between user and item attributes while completely ignoring the interaction between users and items. This leads to a lack of novelty in recommendations and a heavy reliance on the quality of both item and user representations.

[0039] This can be improved by optimizing the task descriptions in the large language model input. However, a more serious challenge is that large language models are prone to hallucinations and may recommend non-existent items. The input restrictions of large models further limit the size of the input item set, while the size of the item set in actual recommendation tasks far exceeds the input restrictions of large models. Another approach is to have the original recommendation model generate a candidate set of items and then re-rank them with the large model. In an experiment with re-ranking with a large language model, when the number of candidate sets increased, multiple metrics after re-ranking, including Recall@10, Recall@20, and NDCG@10, showed a significant decline, far below those of traditional recommendation models. This may be due to the large language model's distraction from the numerous item descriptions in long texts.

[0040] In this application, semantic enhancement is performed when constructing item feature input and user feature input. The semantic enhancement fully utilizes the information of the item, and the user feature input fully utilizes the interaction information between the user and the item to semantically enhance the features of the item and the user. Afterwards, when the semantically enhanced item and user features are input into the LLM, semantic representations of the item and the user are generated in combination 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 representation, make the model more inclined to recommend high-quality resources, and improve the recommendation accuracy.

[0041] (3) Knowledge graph construction The main steps of constructing the above knowledge graph include: building a large language model, text splitting, information extraction prompt construction, large model text generation, triple extraction, and finally visualizing the knowledge graph.

[0042] After the knowledge text is split using the text segmentation model, it is then integrated into the large language model by combining the knowledge graph triple extraction template with the split text. First, a prompt is constructed to extract the task description. The prompt first describes the extraction task in detail and defines the entity relationships within the triples. It then presents three examples of extracting entity relationships from text using a few-shot approach. A standard output format is specified, and special formatting symbols are used between triples to facilitate subsequent text segmentation based on the responses of the large language model.

[0043] Finally, the triple extraction task is transformed into a large language model text generation task to obtain the generated extraction results and obtain the set of entity relationships through text parsing. In an example, the visualization results are as follows: Figure 3 shown.

[0044] The RHCLRec model provided in the embodiment of the present application is as follows Figure 4 As shown in the figure, after achieving semantic enhancement of users and items through the above Graph-RAG model, the following three operations are also performed: First, recommendations based on graph neural networks are implemented through stacking multiple layers of graph convolution, including: For bipartite graph-structured datasets like user-item interactions, graph convolutional networks (GCNs) are well-suited for learning user and item features. The fundamental idea behind GCNs is to learn node representations from graph features. To achieve this, GCNs iteratively perform graph convolutions, aggregating neighbor features into a new representation for the target node. This neighborhood aggregation can be abstracted as:

[0045] AGG is the aggregation function at the core of graph convolution, aiming to integrate the representations of the target node and its neighboring nodes at the kth layer. Many studies have improved models by improving the aggregation function, such as the weighted sum aggregation method in GIN and the LSTM aggregation method in GraphSAGE. In the recommendation model, the weighted aggregation method is used. The convolution operation is defined as:

[0046]

[0047] in, is the feature vector representation of the user in the k-th layer of graph convolution, is the vector representation of the item in the k-th layer of graph convolution, represents the neighborhood of user node u, is the neighborhood of item i, the normalized term The size of the node feature vector is suppressed as the number of graph convolution layers increases. In the aggregation layer, the embedding of the bottom layer is trained including the user embedding. Then, the higher-level representation is calculated through graph convolution. After k layers of graph convolution, the representations of all layers are merged, and the final representation is:

[0048]

[0049] in, Represents the model parameters of the kth layer, integrating the feature representation of multiple layers, so that the model can learn the high-order features captured by high-level convolution while suppressing the excessive smoothing of high-order embeddings caused by stacked convolution layers. The final user features are obtained by weighted fusion of the representations of each layer. and item characteristics The user-item interaction matrix is , where M and N are the number of users and items respectively. If user u interacts with item i, then Then it is recorded as 1 otherwise it is recorded as 0, thus obtaining the user-item adjacency matrix:

[0050] The bottom embedding matrix is , where T is the dimension of embedding, the matrix operation of graph convolution can be expressed as:

[0051] Where D is a (M+N)×(M+N) diagonal matrix, It is used to represent the number of non-zero entries in the i-th row vector in the adjacency matrix A. The resulting characteristic matrix is ​​expressed as:

[0052] Secondly, we achieve comparative alignment between semantic features enhanced by the large language model and user resource interaction features, including: User description after data enhancement and item description Based on the characterization of user preferences, semantic representations of user descriptions and item descriptions are obtained by encoding with text. The text input is converted into a fixed-length feature vector while retaining semantic information and contextual information. After obtaining the user embedding representation learned by the aggregation model and item embedding representation Based on this, we define the density ratio function Capturing semantic representations and graph aggregation representation The similar features between them help to reduce the impact of data quality and noise in representation learning. is defined as follows:

[0053] Function Function represents the cosine similarity, and This means that the semantic representation Map to In contrastive alignment, we will and Considered as positive pairs. During the learning process, these pairs influence each other to align their representations. The goal during model training is to reduce the distance between positive pairs, while treating the remaining samples as negative. By combining ID-based graph aggregation representation learning with text-based behavioral semantics, a contrastive alignment method effectively aligns the semantic features of external knowledge and user-item interaction characteristics.

[0054] The third aspect is the graph feature enhancement method based on subgraph contrastive learning, including: In addition to feature enhancement in the semantic representation of users and items, graph features should also be used to further improve the model to enhance the graph representation effect and the robustness of the model. The distribution of node degrees in the bipartite graph of user-item interaction obtained through historical interaction data is not balanced. For users or item nodes with relatively small degrees, that is, those who participate in less interaction, it is difficult for graph neural networks to learn the node features of unpopular items. In response to the impact of introducing active users and popular items, this study introduces graph data enhancement methods. The most widely used method of graph data enhancement is contrastive learning, which is to form subgraphs by randomly changing the graph structure and achieve graph enhancement by comparing subgraphs. The graph contrast learning method is as follows Figure 5 shown.

[0055] For the convenience of description, we simplify the operation of the aggregation layer as follows:

[0056] in is the representation of the node in the previous layer, Represents the original graph structure, The representation vector representing this layer is used to aggregate the graph structure by removing edges and nodes. Perform dropout operation to form a subgraph and Expressed as:

[0057]

[0058] function and Representing two completely independent dropout methods, the probability of deleting a node and the edges connected to the node is expressed as:

[0059] in is the mask vector, which is generated by Bernoulli distribution, and the convolution operation is performed independently to obtain the new node representation: and The edge deletion method probabilistically deletes some edges, and the processing method is similar to the node deletion method:

[0060] Adjusting the graph structure in each convolution layer can achieve a similar effect to random walk, and realize the mask vector of layer difference. The difference between random walk and edge pruning is as follows Figure 6 shown.

[0061] These methods all achieve graph enhancement by changing the graph structure. The effect of graph enhancement depends on the interactive graph features used for training and the distribution of subgraphs generated for comparative learning. Changing the graph structure to obtain a more evenly distributed representation space is difficult and time-consuming. Changing the embedding space during the aggregation process is a more feasible approach. Instead of changing the graph structure, graph enhancement is achieved by adding small perturbations to the representation vector during the propagation process. The specific implementation method is to directly add random noise to the representation vector to achieve efficient graph enhancement. For a given node The corresponding dimensional space representation , the representation level can be enhanced by:

[0062] in and is the noise vector, and the dimension space is represented Adding noise vectors at each layer aggregation yields and , the noise vector satisfies the following constraints:

[0063]

[0064] in , the first constraint is used to control the noise vector Granularity, ensuring Numerically equivalent to The second constraint is used to avoid adding noise that causes semantic features to deviate and reduce the number of valid positive samples. It can be seen as rotating a certain angle in the vector space. Controlling the rotation angle can help preserve the original feature information while achieving feature enhancement. And the random noise is different for each node. The graph enhancement method of adding noise is as follows Figure 7 shown.

[0065] In the actual construction process of the recommendation model, the graph enhancement method of changing the graph structure and the enhancement method of adding noise both improved the model performance in experiments.

[0066] Furthermore, the loss function of the above RHCLRec model is for:

[0067] Among them, the model prediction part adopts the BPR (Bayesian Personalized Ranking) loss function, which is expressed as The BPR loss function is used to capture the user's preference for items and is calculated as follows:

[0068] in, is the number of users, is a user Neighborhood, and is a user For items and The prediction score of is the Sigmoid function, is the regularization coefficient, used to prevent the embedding matrix Overfitting, is the initial embedding matrix.

[0069] Partial loss function of graph contrastive learning , which is used to capture the structural information of users and items on the graph and is calculated as follows:

[0070] in, is a collection of users, and is a user The original graph structure and the perturbed graph structure, is the similarity function, is a temperature parameter used to control the contrast strength during contrastive learning.

[0071] Characterization comparison and alignment part loss function , used to capture the semantic information of users and items, and is calculated as follows:

[0072] in, is the expected value, is the density ratio function, is the set of all possible semantic representations.

[0073] The above three loss functions jointly guide model learning, through optimization To improve the recommendation performance of the model.

[0074] Figure 8 This is a diagram of a learning resource recommendation system architecture provided in an embodiment of the present application, such as Figure 8 Shown, including: A feature acquisition module 810 is used to acquire user features of the user to be recommended and item features corresponding to the learning resource to be recommended; 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 representation and item semantic representation combined with external knowledge, and obtains user embedding representation and item embedding representation through a graph convolutional neural network in combination with the user-item interaction bipartite graph, and then performs representation comparison and alignment on the corresponding embedding representation and the semantic representation combined with external knowledge; the user-item interaction bipartite graph is enhanced based on subgraph comparison learning, and then the user-item interaction bipartite graph after representation comparison and alignment and graph enhancement is combined with the user-item interaction bipartite graph, and a graph convolutional neural network is used to recommend items to the recommended user.

[0075] It should be understood that the above-mentioned system is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the system are similar to those described in the above-mentioned method. The working process of the system can refer to the corresponding process in the above-mentioned method and will not be repeated here.

[0076] Based on the method in the above embodiment, the embodiment of the present 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 may call the logic instructions in the memory 930 to execute the method in the above embodiment.

[0077] In addition, the logic instructions in the aforementioned memory 930 can be implemented in the form of a software functional unit and, when sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0078] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above embodiment.

[0079] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor executes the method in the above embodiment.

[0080] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0081] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed 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, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC.

[0082] The above embodiments can be implemented in whole or in part using software, hardware, firmware, or any combination thereof. When implemented using software, they can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions. When loaded and executed on a computer, the computer program instructions fully or partially produce the processes or functions described in the embodiments of this application. 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 via 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, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or semiconductor media (e.g., solid-state drive (SSD)).

[0083] It will be understood that the various numerical numbers involved in the embodiments of the present application are merely distinctions for the convenience of description and are not intended to limit the scope of the embodiments of the present application.

[0084] It is easy for those skilled in the art to understand that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A learning resource recommendation method, characterized in that: The following steps are involved: Obtain the user characteristics of the user to be recommended and the item characteristics corresponding to the learning resource to be recommended; Inputting the user features and item features into a resource recommendation model to obtain recommended learning resources 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, and then obtains user embedding representations and item embedding representations through a graph convolutional neural network combined with a bipartite graph of user-item interactions, and then performs representation comparison and alignment between the corresponding embedding representations and the semantic representations combined with external knowledge; The user-item interaction bipartite graph is enhanced based on subgraph contrastive learning, and then, a graph convolutional neural network is used to recommend items to the recommended user by combining the user-item interaction bipartite graph after representation contrastive alignment and graph enhancement.

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 comparative learning, and the loss function of representation comparative alignment.

3. The method according to claim 1, characterized in that Loss function of resource recommendation model ;in, Represents the loss function of the graph convolutional neural network for recommendation, represents the loss function of subgraph contrastive learning, represents the loss function for characterizing contrastive alignment; in, is the number of users, is a user Neighborhood, and is a user For items and items The prediction score of is the Sigmoid function, is the regularization coefficient, is the initial embedding matrix; in, is a collection of users, and is a user The original graph structure and the perturbed graph structure, and is a user The original graph structure and the perturbed graph structure, is the similarity function, is the temperature parameter; in, is the expected value, is the density ratio function, is the set of all possible item semantic representations or user semantic representations, Indicates an item or user i The semantic representation of Indicates an item or user j The semantic representation of Indicates an item or user i Embedding representation of .

4. The method according to claim 1, wherein Obtaining user features and item features includes: Construct user features based on the set of items that the user interacts with, the titles of each item in the item set, and the set of user comments on all items; Construct item features based on the item title, original item description, resource domain attributes of the item, and user comments on the item; The user features and item features are enhanced based on retrieval enhancement generation and a large language model, including: The constructed user features and item features are expanded by combining them with 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 combined with external knowledge; the preset teaching resources and knowledge graphs both include information on high-quality learning resources.

5. The method according to claim 1, wherein Combined with the bipartite graph of user-item interactions, the graph convolutional neural network is used to obtain user embedding representations and item embedding representations, including: Graph convolution is iteratively performed through a graph convolutional neural network to aggregate the features of each node's neighbors in the user-item interaction bipartite graph to obtain a new representation of the node. The new representations of the nodes in multiple aggregation layers are then weighted and merged to obtain an embedded representation of the node; the node is a user or an item.

6. The method according to claim 1, characterized in that The characterization comparison and alignment includes: The density ratio function is used to capture the similarities between the embedding representation of the user or item and the corresponding semantic representation, and the embedding representation is aligned with the corresponding semantic representation with the goal of increasing the similarity between the embedding representation and the semantic representation.

7. The method according to claim 1, characterized in that Graph enhancement is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including: Obtaining a subgraph structure by randomly changing the graph structure of the user-item interaction bipartite graph; Combining different subgraph structures, we use graph convolutional neural networks to obtain user embedding representations and item embedding representations corresponding to different subgraph structures. The user embedding representations and item embedding representations corresponding to different subgraph structures are compared and learned to achieve graph enhancement of the user-item interaction bipartite graph.

8. The method according to claim 1, characterized in that Graph enhancement is performed on the user-item interaction bipartite graph based on subgraph contrastive learning, including: Combined with the user-item interaction bipartite graph, noise vectors are introduced into different aggregation layers of the graph convolutional neural network to obtain corresponding user embedding representations and item embedding representations; 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 enhancement of the user-item interaction bipartite graph.

9. A learning resource recommendation system, characterized in that: include: A feature acquisition module is used to acquire user features of the user to be recommended and item features corresponding to the learning resource to be recommended; A resource recommendation module is configured to input the user and item features into a resource recommendation model to obtain recommended learning resources for the user. The resource recommendation model enhances the user and item features based on retrieval-enhanced generation and a large language model to obtain user and item semantic representations that are integrated with external knowledge. A graph convolutional neural network is used to obtain user and item embedding representations based on a bipartite graph of user-item interactions. The corresponding embedding representations are then compared and aligned with the semantic representations that are integrated with external knowledge. The user-item interaction bipartite graph is enhanced based on subgraph contrastive learning, and then, a graph convolutional neural network is used to recommend items to the recommended user by combining the user-item interaction bipartite graph after representation contrastive alignment and graph enhancement.

10. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 8.

Images