Personalized learning content recommendation method and system based on neighborhood and hypergraph collaboration

Abstract

The application provides a personalized learning content recommendation method and system based on neighborhood and hypergraph cooperation, and relates to the technical field of content recommendation. The user session sequence data is acquired to construct a hypergraph view and a neighbor graph view; global high-order relationship features are extracted based on the hypergraph view; local co-occurrence relationship features are extracted based on the neighbor graph view to generate enhanced co-occurrence relationship features; the sequence position encoding of the user learning path is spliced with the learning content semantic embedding to generate a dynamic learning interest embedding representation through a gated dynamic attention mechanism to fuse context information; based on a global-local contrast learning strategy, the dynamic learning interest embedding and the global semantic embedding are nonlinearly projected to optimize the semantic consistency of the two types of embeddings through a contrast loss; the cross-entropy loss of the recommendation task and the contrast learning loss are combined to construct a joint optimization objective function to generate a personalized recommendation result, and the technical effect of capturing the explicit and implicit relationships between learning contents is realized.

Description

Technical Field

[0001] This application relates to the field of content recommendation technology, and in particular to a personalized learning content recommendation method and system based on neighborhood and hypergraph collaboration. Background Technology

[0002] With the rapid development of the internet and online education, user behavior data on learning platforms is exploding. Personalized learning material recommendation systems, as an important educational technology, aim to predict a user's next learning material or course of interest based on their current learning behavior. Existing learning material recommendation models can be mainly categorized as follows:

[0003] Content-based recommendation methods analyze the characteristics of learning resources (such as course content, video topics, and exercise types) to recommend learning resources similar to a user's historical interests. The advantage of this method is that it does not require user interaction data, but the disadvantage is that it struggles to capture dynamic changes in user interests and implicit relationships. For example, traditional content-based recommendation systems may only recommend relevant courses based on keyword matching, but they cannot consider the user's learning path and the coherence of their knowledge system.

[0004] Collaborative filtering-based recommendation methods: Collaborative filtering is one of the earliest methods applied in the field of learning resource recommendation. It recommends learning resources that other users may be interested in by analyzing the similarities between users (such as user ratings, browsing history, etc.). Collaborative filtering is divided into user-based collaborative filtering (User-based CF) and item-based collaborative filtering (Item-based CF). Although this method can capture implicit relationships between users, it is susceptible to data sparsity and the cold start problem, especially when the diversity of learning resources is high.

[0005] Markov chain-based methods assume that a user's learning behavior depends only on the current state and is independent of previous states. In 2005, Shani et al. proposed a recommender system based on Markov Decision Processes (MDPs), which captures the dynamic characteristics of time series data by simulating user behavior transitions. The advantages of this type of method are its simplicity and computational efficiency, making it suitable for processing short-term time series data. However, Markov chains have limitations in capturing long-term dependencies and in assuming that the user's next action depends only on the current state, ignoring earlier historical behaviors. This simplistic assumption makes it difficult for the model to capture complex user learning patterns, especially when user learning paths are diverse and non-linear, often resulting in unsatisfactory recommendation performance.

[0006] Recurrent Neural Network (RNN)-based methods: With the development of deep learning technology, learning material recommendation models based on Recurrent Neural Networks (RNNs) have gradually become mainstream. RNNs can process sequential data and predict users' next learning behavior by learning the inherent dependencies in the user's learning behavior sequence. In 2016, Hidasi et al. proposed an RNN-based conversational recommendation model, which significantly improved recommendation performance by capturing the temporal dependencies in the user behavior sequence through GRU (Gated Recurrent Unit). Compared with Markov chain methods, RNNs can better capture the long-term dependencies of user learning behavior. However, RNN models also have some limitations: First, RNNs usually treat the user learning behavior sequence as a unidirectional chain, overemphasizing the relative order of the learned content while ignoring the correlation between them, which may lead to the recommendation model misjudging user preferences. Second, RNNs are prone to gradient vanishing or exploding problems when processing long sequences. Although variants such as LSTM and GRU partially alleviate these problems, they still struggle to capture the complex relationships between the learned content.

[0007] Graph Neural Network-Based Approaches: In recent years, learning material recommendation models based on Graph Neural Networks (GNNs) have gradually attracted attention. GNNs effectively reveal complex interactions between learning content by modeling user learning behavior as a graph structure. In 2019, Wu et al. proposed the SR-GNN model, introducing GNNs into the field of learning material recommendation for the first time. SR-GNN transforms user learning behavior sequences into directed graphs and uses gated graph neural networks to learn item embeddings, capturing complex transformation relationships between learning content. Compared to RNN-based methods, GNNs relax the restrictions on temporal dependencies between consecutive items, enabling them to better capture implicit relationships and higher-order interactions between learning content. However, existing GNN models still have shortcomings in handling higher-order and implicit relationships. For example, traditional graph convolution methods struggle to extract higher-order relationships between learning content and perform poorly when dealing with sparse data. Furthermore, GNN models have high computational complexity, especially when processing large-scale data, resulting in significant computational overhead.

[0008] Therefore, how to capture the explicit and implicit relationships between learning content, track the dynamic evolution of users' learning interests in real time, and effectively handle sparse short-term interaction sequences has become an urgent technical problem to be solved. Summary of the Invention

[0009] To capture explicit and implicit relationships between learning content, track the dynamic evolution of users' learning interests in real time, and effectively handle sparse short-term interaction sequences, this application provides a personalized learning content recommendation method and system based on neighborhood and hypergraph collaboration.

[0010] Firstly, the personalized learning content recommendation method based on neighborhood and hypergraph collaboration provided in this application adopts the following technical solution:

[0011] A personalized learning content recommendation method based on neighborhood and hypergraph collaboration includes:

[0012] Acquire user session sequence data to construct a hypergraph view and a neighbor graph view, and generate a comprehensive learning content relationship network through a multi-view fusion module;

[0013] Based on the hypergraph view, global high-order relation features are iteratively extracted by multi-hop hypergraph convolution combined with the K-hop algorithm, and then complementaryly fused with the co-occurrence relation features of the neighboring graph view.

[0014] Based on the neighbor graph view, local co-occurrence relationship features are extracted through weighted multi-hop neighbor graph convolution, and noisy edges are filtered by a threshold mechanism to generate enhanced co-occurrence relationship features;

[0015] The sequence position encoding of the user's learning path is concatenated with the semantic embedding of the learning content, and contextual information is fused through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation.

[0016] Based on a global-local contrastive learning strategy, the dynamic learning interest embedding and the global semantic embedding are nonlinearly projected, and the semantic consistency of the two types of embeddings is optimized by contrastive loss.

[0017] We construct a joint optimization objective function by combining cross-entropy loss and contrastive learning loss for the joint recommendation task. We then dynamically adjust the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results.

[0018] Optionally, the step of acquiring user session sequence data to construct a hypergraph view and a neighbor graph view, and generating a comprehensive learning content relationship network through a multi-view fusion module, includes:

[0019] Acquire user session sequence data and construct a hypergraph view to connect multiple learning content nodes via hyperedges;

[0020] Construct a neighbor graph view to build weighted edges for learning content co-occurring in the user's learning path;

[0021] The multi-view fusion module performs weighted concatenation of high-order features of the hypergraph view and co-occurrence features of neighboring graph views to generate a comprehensive learning content relationship network.

[0022] Optionally, the step of iteratively extracting global high-order relation features based on the hypergraph view through multi-hop hypergraph convolution combined with the K-hop algorithm, and complementaryly fusing them with the co-occurrence relation features of neighboring graph views, includes:

[0023] Based on the hypergraph association matrix and node degree matrix, single-hop hypergraph convolution features are extracted through bidirectional information aggregation of node-hyperedge-node;

[0024] By combining the K-hop algorithm for multi-hop iteration, high-order interaction information of multiple nodes within the hyperedge is aggregated layer by layer to generate global high-order relation features;

[0025] The global high-order relationship features are input into the dynamic attention mechanism and attention weights are assigned to the co-occurrence features of the neighbor graph view to form complementary feature fusion.

[0026] Optionally, the step of extracting local co-occurrence relationship features based on the neighbor graph view through weighted multi-hop neighbor graph convolution, and filtering noisy edges using a threshold mechanism to generate enhanced co-occurrence relationship features includes:

[0027] Based on the adjacency matrix and degree matrix, local co-occurrence relationship features are extracted through weighted graph convolution, where the edge weights are dynamically adjusted according to the co-occurrence frequency and user behavior time series.

[0028] An adaptive threshold mechanism is introduced to filter out noise edges with co-occurrence frequencies below the dynamic threshold, while preserving strong co-occurrence relationships;

[0029] By combining the K-hop algorithm to aggregate multi-hop neighbor information, enhanced co-occurrence relationship features are generated and cross-view interactions are performed with higher-order features of the hypergraph view.

[0030] Optionally, the step of concatenating the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fusing contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation includes:

[0031] The positional encoding of the learned content sequence is concatenated with the semantic embedding to generate a context-aware initial embedding;

[0032] The initial embedding is nonlinearly transformed by a gated linear unit, and the attention weight is dynamically calculated by combining the average embedding of the user's current session.

[0033] Invalid location information is filtered out using a masking mechanism to generate a dynamic embedding representation of the user's learning path. This dynamic embedding representation shares the feature space with the global-local contrastive learning module.

[0034] Optionally, the step of performing a nonlinear projection on the dynamically learned interest embedding and the global semantic embedding based on a global-local contrastive learning strategy, and optimizing the semantic consistency of the two types of embeddings through contrastive loss, includes:

[0035] A local view is constructed based on the user's short-term learning path, and local behavioral features are extracted through neighborhood graph convolution to generate local embedding representations;

[0036] A global view is constructed based on the user's long-term learning behavior, and global semantic features are extracted through hypergraph convolution to generate a global embedding representation;

[0037] Nonlinear projection is applied to local and global embeddings to construct positive and negative sample pairs. The semantic consistency of positive sample pairs is maximized by a contrastive loss function, while negative sample pairs are separated.

[0038] Optionally, the steps of constructing a joint optimization objective function by combining the cross-entropy loss and contrastive learning loss of the joint recommendation task, and dynamically adjusting the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results, include:

[0039] The cross-entropy loss of the recommendation task is calculated based on the matching degree between the final embedding and the candidate learned content;

[0040] The similarity calculation is based on the comparison learning loss of local and global embeddings, and the similarity distribution is smoothed by temperature parameters;

[0041] The two types of losses are weighted and summed, and the parameters of the hypergraph, neighbor graph, attention and contrastive learning modules are jointly optimized by gradient descent to construct a joint optimization objective function;

[0042] Personalized recommendation results are generated by dynamically adjusting the parameters of the hypergraph and neighbor graph through backpropagation.

[0043] Secondly, this application provides a personalized learning content recommendation system based on neighborhood and hypergraph collaboration, including:

[0044] The data acquisition module is used to acquire user session sequence data to construct a hypergraph view and a neighbor graph view, and to generate a comprehensive learning content relationship network through the multi-view fusion module;

[0045] The hypergraph view module is used to extract global high-order relation features based on the hypergraph view by iteratively extracting them through multi-hop hypergraph convolution combined with the K-hop algorithm, and to perform complementary fusion with the co-occurrence relation features of the neighboring graph views.

[0046] The neighbor graph view module is used to extract local co-occurrence relationship features based on the neighbor graph view through weighted multi-hop neighbor graph convolution, and filter noisy edges by combining a threshold mechanism to generate enhanced co-occurrence relationship features;

[0047] The representation generation module is used to concatenate the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuse contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation;

[0048] An optimization module is used to perform nonlinear projection on the dynamic learning interest embedding and the global semantic embedding based on a global-local contrastive learning strategy, and optimize the semantic consistency of the two types of embeddings through contrastive loss.

[0049] The results generation module is used to jointly optimize the cross-entropy loss and contrastive learning loss of the recommendation task, construct a joint optimization objective function, and dynamically adjust the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results.

[0050] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0051] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0052] In summary, this application constructs a hypergraph view and a neighbor graph view by acquiring user session sequence data; based on the hypergraph view, iteratively extracts global high-order relationship features through multi-hop hypergraph convolution combined with the K-hop algorithm, and performs complementary fusion with the co-occurrence relationship features of the neighbor graph view; based on the neighbor graph view, extracts local co-occurrence relationship features through weighted multi-hop neighbor graph convolution, and filters noisy edges using a threshold mechanism to generate enhanced co-occurrence relationship features; concatenates the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuses contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation; based on a global-local contrastive learning strategy, nonlinearly projects the dynamic learning interest embedding and the global semantic embedding, and optimizes the semantic consistency of the two types of embeddings through contrastive loss; jointly constructs a joint optimization objective function by combining the cross-entropy loss and contrastive learning loss of the recommendation task, and generates personalized recommendation results, achieving the technical effect of capturing explicit and implicit relationships between learning content. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application;

[0054] Figure 2 This is a flowchart illustrating the first embodiment of the personalized learning content recommendation method based on neighborhood and hypergraph collaboration in this application;

[0055] Figure 3 This is a structural block diagram of the first embodiment of the personalized learning content recommendation system based on neighborhood and hypergraph collaboration in this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to 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.

[0057] Reference Figure 1 , Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0058] like Figure 1 As shown, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0059] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0060] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a personalized learning content recommendation program based on neighborhood and hypergraph collaboration.

[0061] exist Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be set in the computer device. The computer device calls the personalized learning content recommendation program based on neighborhood and hypergraph collaboration stored in the memory 1005 through the processor 1001, and executes the personalized learning content recommendation method based on neighborhood and hypergraph collaboration provided in the embodiment of this application.

[0062] This application provides a personalized learning content recommendation method based on neighborhood and hypergraph collaboration, referring to... Figure 2, Figure 2 This is a flowchart illustrating the first embodiment of the personalized learning content recommendation method based on neighborhood and hypergraph collaboration in this application.

[0063] In this embodiment, the personalized learning content recommendation method based on neighborhood and hypergraph collaboration includes the following steps:

[0064] Step S10: Obtain user session sequence data to construct a hypergraph view and a neighbor graph view, and generate a comprehensive learning content relationship network through the multi-view fusion module.

[0065] It should be noted that the existing technologies for recommending learning content have the following drawbacks:

[0066] Overemphasizing the relative order of learning content: Existing learning material recommendation models usually treat the user's learning behavior sequence as a unidirectional chain, overemphasizing the relative order of learning content while ignoring the correlation between them, which may lead to the recommendation model misjudging user preferences.

[0067] Difficulty in capturing high-order relationships: Traditional graph convolution methods struggle to extract high-order relationships and latent features between learned content. Increasing the number of nodes to improve the graph structure description capability will increase the computational cost of the model.

[0068] Handling sparse data is difficult: Since the information contained in the short-term interaction sequences on which the learning materials are recommended is very limited, it is easily affected by data sparsity, which leads to a decrease in model performance.

[0069] Difficulty in tracking user interests in real time: Existing models often cannot track the dynamic evolution of user interests over time and context in real time, and it is difficult to accurately capture user preferences based on the multi-dimensional dynamic changes in user interests.

[0070] It is understood that the explanations of technical terms in this embodiment include:

[0071] Hypergraph convolution: A graph neural network method capable of handling tuple relationships, allowing a single hyperedge to connect multiple nodes, thus enabling more flexible modeling of complex relationships. Unlike ordinary graph convolution, which can only handle binary relationships, hypergraph convolution can capture complex interactions between multiple learning materials, making it suitable for modeling higher-order relationships in learning resource recommendation. For example, in online learning scenarios, users may simultaneously browse multiple related courses (such as advanced mathematics, linear algebra, and probability theory); hypergraph convolution can capture the higher-order relationships between these courses, thereby improving the accuracy of recommendations.

[0072] Neighbor convolution: Used to extract relationships between neighboring nodes. When two pieces of learning content co-occur multiple times in a user's learning path, they are linked by an edge, and their co-occurrence frequency is represented by a weight coefficient. Neighbor convolution is particularly suitable for capturing the co-occurrence relationships between related courses viewed by a user on the same learning path. For example, if a user views Advanced Mathematics and Linear Algebra on the same learning path, neighbor convolution can capture the implicit relationship between these two courses, thereby improving the real-time performance of recommendations.

[0073] Location-aware dynamic attention mechanism: This mechanism combines location encoding and contextual information to dynamically adjust weights, aiming to accurately describe the correlation between sequences of user learning behaviors and improve the model's accuracy in portraying user learning interests. Location encoding helps the model understand the positional information of learning content within a sequence, thereby better capturing dynamic changes in user learning interests. For example, in a sequence of users browsing learning content, location encoding can help the model identify changes in user interest in different courses, thus dynamically adjusting the recommendation strategy.

[0074] K-hop algorithm: A multi-hop propagation algorithm for graph neural networks that captures multi-hop relationships between nodes. The K-hop algorithm iterates multiple times, gradually aggregating multi-hop neighbor information to capture potential relationships between nodes at greater distances. In learning material recommendation, the K-hop algorithm helps the model capture higher-order relationships between learned content, improving recommendation accuracy.

[0075] Global-Local Contrastive Learning: A method to enhance model representation capabilities through a contrastive learning strategy. Global-local contrastive learning models user learning behavior by fusing local features with global semantic information and combining a multi-view embedding contrast mechanism. Local views represent short-term behavioral features of a user within a specific learning path, while the global view captures global semantic relationships between learning content based on the user's overall learning pattern. Through contrastive learning, the model can more accurately extract feature relationships between learning content and user learning behavior, thereby improving recommendation performance.

[0076] It is understood that the step of acquiring user session sequence data to construct a hypergraph view and a neighbor graph view, and generating a comprehensive learning content relationship network through a multi-view fusion module includes: acquiring user session sequence data; constructing a hypergraph view to connect multiple learning content nodes with hyperedges; constructing a neighbor graph view to construct weighted edges for learning content co-occurring in the user's learning path; and weighted concatenation of the higher-order features of the hypergraph view and the co-occurrence features of the neighbor graph view through a multi-view fusion module to generate a comprehensive learning content relationship network.

[0077] As can be understood, a hypergraph view describes high-order relationships between learning content, allowing a single hyperedge to connect multiple nodes, thus enabling more flexible modeling of complex relationships. A hypergraph view can be represented as an incidence matrix, with the degree matrix of each vertex and each edge represented as a diagonal matrix. Hypergraph views are particularly suitable for capturing high-order relationships between learning content; for example, a user may be interested in multiple related courses (such as mathematics, physics, and chemistry), and a hypergraph can effectively capture the complex interactions between these courses.

[0078] Neighbor Graph View: Describes implicit relationships between learning content. When two learning content items co-occur multiple times in a user's learning path, they are linked by an edge, with a weight coefficient representing their frequency of co-occurrence. Each edge has a weight coefficient, which is calculated based on the frequency of co-occurrence between the two learning content items. Neighbor Graph View is particularly suitable for capturing the co-occurrence relationships between related courses viewed by a user on the same learning path. For example, if a user views Advanced Mathematics and Linear Algebra on the same learning path, Neighbor Convolution can capture the implicit relationship between these two courses, thereby improving the real-time performance of recommendations.

[0079] Step S20: Based on the hypergraph view, extract global high-order relation features iteratively by combining multi-hop hypergraph convolution with the K-hop algorithm, and perform complementary fusion with the co-occurrence relation features of the neighboring graph view.

[0080] It should be noted that the step of extracting global high-order relationship features based on the hypergraph view through multi-hop hypergraph convolution combined with the K-hop algorithm and performing complementary fusion with the co-occurrence relationship features of neighboring graph views includes: extracting single-hop hypergraph convolution features based on the hypergraph association matrix and node degree matrix through bidirectional information aggregation of node-hyperedge-node; performing multi-hop iteration combined with the K-hop algorithm to aggregate high-order interaction information of multiple nodes within the hyperedge layer by layer to generate global high-order relationship features; and inputting the global high-order relationship features into a dynamic attention mechanism to perform attention weight allocation with the co-occurrence features of neighboring graph views to form complementary feature fusion.

[0081] In practical implementation, hypergraph convolution based on multi-hop mechanisms includes:

[0082] The convolution in the K-hop algorithm can be seen as extracting neighbor information within the multi-hop range of a node. It performs two-stage information mining on the hypergraph structure, namely "node-hyperedge-node", to capture deeper inter-node relationships and global structural information.

[0083] Hypergraph Convolution Initial Definition as follows:

[0084]

[0085] Where H is the hypergraph incidence matrix, Embed the learned content from the output of the nth layer hypergraph convolution. D u D e W represents the degree matrix of the nodes and the degree matrix of the hyperedges, respectively. e W is the hyperedge weight matrix, and its initial value is set to 1. l (n-1) This is the parameter matrix for the (n-1)th layer. It is obtained by multiplying by matrix H. T Information aggregation from nodes to hyperedges is achieved, and multiplying by matrix H achieves information aggregation from hyperedges to nodes.

[0086] The model incorporating the K-hop algorithm is defined as follows:

[0087]

[0088] Here, k is the number of hops, ranging from 1 to k, representing the application of single-step hypergraph convolution K times in each layer. Through the multi-hop mechanism, the model can capture potential relationships between nodes at greater distances, thereby better understanding the complex interactions between the learned content.

[0089] Step S30: Based on the neighbor graph view, extract local co-occurrence relationship features through weighted multi-hop neighbor graph convolution, and filter noisy edges by combining a threshold mechanism to generate enhanced co-occurrence relationship features.

[0090] In specific implementation, the step of extracting local co-occurrence relationship features based on the neighbor graph view through weighted multi-hop neighbor graph convolution, and filtering noisy edges using a threshold mechanism to generate enhanced co-occurrence relationship features includes: extracting local co-occurrence relationship features based on the adjacency matrix and degree matrix through weighted graph convolution, wherein the edge weights are dynamically adjusted according to the co-occurrence frequency and user behavior time sequence; introducing an adaptive threshold mechanism to filter noisy edges with a co-occurrence frequency lower than the dynamic threshold, while retaining strong co-occurrence relationships; combining the K-hop algorithm to aggregate multi-hop neighbor information, generating enhanced co-occurrence relationship features, and interacting with higher-order features of the hypergraph view across views.

[0091] It should be noted that neighbor graph convolution based on multi-hop mechanisms includes:

[0092] This embodiment employs a neighbor graph convolution to comprehensively capture implicit relationships between neighboring nodes. When two learned elements co-occur multiple times in the user's learning path, they are linked by an edge, with weight coefficients representing their frequency of co-occurrence. A threshold mechanism is introduced to balance important connection information with computational cost; if the weight coefficient of an edge is less than a preset threshold, it is considered useless information and discarded. Based on this, the K-hop algorithm is applied to capture potential relationships between nodes at greater distances.

[0093] The core operation of neighbor graph convolution can be represented as:

[0094]

[0095] Among them, H (l) For the node of the l-th layer, It is an adjacency matrix with self-loops added, D is the degree matrix, and W is the adjacency matrix with self-loops added. (l) σ is a learnable weight matrix, and σ is a non-linear activation function. To further enhance the model's expressive power, a multi-layer convolutional structure is used to effectively capture the complex high-order relationships between the learned content:

[0096] H nbg final =CONCAT(H (0) H (1) ,...,H (L) ) / (L+1)

[0097] Where L is the total number of layers, CONCAT represents the splicing operation, and H... nbg final This is the final neighbor graph convolution. Through multi-layer convolutional structures, the model can more comprehensively capture the complex interactions between learned content, thereby improving the accuracy of recommendations.

[0098] Step S40: Concatenate the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuse contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation.

[0099] It should be noted that the step of concatenating the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fusing contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation includes: concatenating the position encoding of the learning content sequence with the semantic embedding to generate a context-aware initial embedding; performing a nonlinear transformation on the initial embedding through a gated linear unit, and dynamically calculating the attention weights in combination with the average embedding of the user's current session; and using a masking mechanism to filter invalid position information to generate a dynamic embedding representation of the user's learning path, wherein the dynamic embedding representation shares the feature space with the global-local contrastive learning module.

[0100] It should be noted that the location-aware dynamic attention mechanism that integrates contextual information includes:

[0101] A location-aware dynamic attention mechanism that integrates contextual information is introduced when generating learning path embeddings. This attention mechanism is used to fuse learning content information from different locations to more accurately capture the dynamic changes in user interest during the learning process. Specifically, the model integrates two methods for generating learning path embeddings: location-aware dynamic attention and context-aware adaptive attention, using the attention mechanism to fuse learning content information from different locations.

[0102] The definition combining positional encoding and learned content representation is as follows:

[0103] n h =W1(|p emb ;seq h |)

[0104] Where p emb It's positional encoding, seq h The sequence is represented by an item, and W1 is a learnable weight matrix. Then, attention weights are calculated through non-linear transformations and gating mechanisms.

[0105]

[0106]

[0107] Among them, GLU1 and GLU2 are gated linear units, h s It is the average embedding of the session.

[0108] Finally, the attention score is calculated and applied to the sequence:

[0109] γ=n h `W2

[0110]

[0111] Where γ is the attention weight, used to perform a weighted summation of the item embeddings; γ′ is the final attention weight; W2 is a learnable weight vector used to calculate the parameters of the attention weights; mask is used to filter invalid positions; and s is the session representation, obtained by summing the weighted item embeddings. Let represent the sequence of the i-th item. Through a location-aware dynamic attention mechanism, the model can better understand changes in a user's interests at different learning stages, thereby providing more personalized recommendations.

[0112] Step S50: Based on the global-local contrastive learning strategy, perform nonlinear projection on the dynamic learning interest embedding and the global semantic embedding, and optimize the semantic consistency of the two types of embeddings through contrastive loss.

[0113] It should be noted that the steps of nonlinearly projecting the dynamic learning interest embedding and the global semantic embedding based on the global-local contrastive learning strategy, and optimizing the semantic consistency of the two types of embeddings through contrastive loss, include: constructing a local view based on the user's short-term learning path, extracting local behavioral features through neighbor graph convolution, and generating a local embedding representation; constructing a global view based on the user's long-term learning behavior, extracting global semantic features through hypergraph convolution, and generating a global embedding representation; performing nonlinear projection on the local embedding and the global embedding, constructing positive sample pairs and negative sample pairs, maximizing the semantic consistency of the positive sample pairs through the contrastive loss function, and simultaneously separating the negative sample pairs.

[0114] It should be noted that global-local contrastive learning modeling includes:

[0115] By fusing local features with global semantic information and combining a multi-view embedding comparison mechanism to model user learning behavior, the model can more accurately extract the feature relationships between learning content and user learning behavior, thereby effectively improving model performance. Local views represent short-term behavioral features of a user in a specific learning path, while global views, based on the user's overall learning pattern, model the global semantics between learning content through hypergraphs and long-term interaction scenarios, representing the overall embedding of the user's long-term learning interests and preferences.

[0116] Local views represent the direct information between local click sequences or learning content domains within a specific session, focusing on short-term user behavior characteristics. Global views, on the other hand, are based on the user's overall behavior patterns, modeling the global semantics between learning content through hypergraphs and long-term interaction scenarios, representing the overall embedding of the user's long-term interests and preferences.

[0117] For each mini-batch during training, there are corresponding relationships between the learning path embedding groups. These embeddings can serve as mutual references during self-supervised learning, achieving self-supervision. Local embeddings are obtained by propagating local relationships within the sequence of learning content clicks through a domain graph generated from the learning content click order.

[0118]

[0119] Where A is the adjacency matrix of the neighborhood graph, and D... l H is the degree matrix of the neighborhood graph. l W l These are the domain graph node embedding vector and the weight matrix, respectively.

[0120] Global embeddings are based on a hypergraph structure composed of content learned through multiple learning paths. Global correlations are propagated through hypergraph convolution to generate global embeddings.

[0121]

[0122] Where B is the degree matrix of the hypergraph nodes, and H, E, and W are the node embedding, hypergraph association matrix, and weight parameter matrix, respectively.

[0123] To further enhance the collaborative modeling of global and local embeddings, this paper employs nonlinear projection, mapping both local and global embeddings to the contrastive learning feature space via a projection head. The projection formula for the embedding representation is as follows:

[0124] z = ψ(Ch + b)

[0125] Where C and b are the parameters of the projection layer, and ψ is the nonlinear activation function.

[0126] The core of the contrastive loss model is to construct positive and negative sample pairs to drive the model to learn the similarity of representations between the global and local views, while separating the representations of irrelevant samples. Positive sample pairs are the local and global embeddings of the same learning path or user, while negative sample pairs are obtained by randomly selecting embeddings from other learning paths or the user's learning path. The contrastive learning loss function can be expressed as:

[0127]

[0128] N p Z represents the number of positive samples, where z i and It is a positive sample pair (local embedding and global embedding), and z i For local embedding The projection vector, To be with z i Corresponding global embedding The projection vector. The similarity between positive samples is calculated using the following formula:

[0129]

[0130] z k For negative samples, there are M noise samples randomly selected from other sessions. For all negative samples and the current sample z i The similarity scores are exponentially summed. τ is a temperature parameter used to smooth the distribution of similarity scores and reduce the impact of extreme values ​​on contrastive learning.

[0131] Step S60: Combine the cross-entropy loss and contrastive learning loss of the joint recommendation task to construct a joint optimization objective function. Dynamically adjust the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results.

[0132] In specific implementation, the steps of constructing a joint optimization objective function using the cross-entropy loss and contrastive learning loss of the joint recommendation task, dynamically adjusting the parameters of the hypergraph and neighbor graph through backpropagation, and generating personalized recommendation results include:

[0133] The cross-entropy loss for the recommendation task is calculated based on the matching degree between the final embedding and the candidate learning content; the contrastive learning loss is calculated based on the similarity between the local and global embeddings, and the similarity distribution is smoothed by the temperature parameter; the two types of losses are weighted and summed, and the parameters of the hypergraph, neighbor graph, attention and contrastive learning modules are jointly optimized by gradient descent to construct a joint optimization objective function; the parameters of the hypergraph and neighbor graph are dynamically adjusted by backpropagation to generate personalized recommendation results.

[0134] In specific implementation, the model optimization and prediction layer in this embodiment includes:

[0135] The recommendation objective is to predict the next learning resource that a user will click on within a given learning path. (s,m+1) The model's predicted score s i This represents the model's confidence level for each candidate learning data:

[0136] s i =sigmoid(H final ·W pred )

[0137] in, It is an item embedding representation obtained by multi-layer hypergraph convolution, S t S is the current state vector, representing the model's understanding at time step t. f The backward state vector provides information about future states, helping the model to understand the context more comprehensively. W pred This is the learnable weight matrix for the prediction layer.

[0138] During training, the cross-entropy loss function is used as the training task loss:

[0139]

[0140] Among them, y i It is the one-hot encoded vector of the real label.

[0141] Finally, the recommendation task and the self-supervised task are unified into a joint learning framework, and the overall loss function is:

[0142] L total =L rec +βL s

[0143] Among them, L sThis represents the sum of the contrastive losses across the three channels, and β is a hyperparameter controlling the intensity of the contrastive learning task. Through the joint learning framework, the model can simultaneously optimize both the recommendation and contrastive learning tasks, thereby improving overall performance.

[0144] In practical implementation, the technical effects that this embodiment can achieve are as follows:

[0145] 1. Identify explicit and implicit relationships between projects:

[0146] The personalized learning material recommendation model based on neighborhood and hypergraph collaboration (NGH-Rec) proposed in this embodiment can simultaneously capture explicit and implicit relationships between learning content by combining hypergraph convolution and neighbor convolution. Compared with traditional learning material recommendation models, NGH-Rec can more comprehensively explore the complex interaction relationships between learning content, thereby improving the accuracy and real-time performance of recommendations. For example, when recommending mathematics courses, the model not only considers the mathematics courses that the user has previously viewed, but also combines related physics and chemistry courses for recommendations, thereby better meeting the user's learning needs.

[0147] 2. Real-time tracking of the dynamic evolution of users' learning interests:

[0148] This embodiment introduces a location-aware dynamic attention mechanism that integrates contextual information, enabling real-time tracking of changes in user learning interests over time and context. By dynamically adjusting weights, the model can more accurately capture changes in user preferences, improving the personalization of recommendations. For example, when a user switches from basic to advanced courses, the model can adjust its recommendation strategy promptly, providing materials that better suit the user's current learning stage.

[0149] 3. Effectively handles sparse short-time interaction sequences:

[0150] Because the information contained in the short-term interaction sequences upon which learning material recommendations are based is very limited, traditional models are easily affected by data sparsity. NGH-Rec, through multi-view modeling and a global-local contrastive learning strategy, can effectively handle sparse data and improve model performance under sparse data conditions. For example, even when a user has only viewed a small number of courses, the model can still provide high-quality recommendations based on global semantic information.

[0151] 4. Improve the model's generalization ability and robustness:

[0152] By employing a global-local contrastive learning strategy, the model can better capture the multidimensional features and semantic differences between learned content, thereby improving its generalization ability and robustness. Even when facing different learning domains or different user behavior patterns, the model can still maintain high recommendation performance.

[0153] It should be noted that, through the synergistic effect of the above technical solutions, the present invention achieves the following significant advantages:

[0154] Enhanced high-order relation modeling capabilities: Hypergraph View breaks through the limitations of traditional binary relations, capturing high-order associations across disciplines and multiple contents (such as the combined interest of mathematics and physics courses), and the recommendation results are more in line with the user's deep learning needs.

[0155] Experiments show that on interdisciplinary datasets, NGH-Rec achieves a 23.6% improvement in recommendation accuracy (Precision@10) compared to traditional GNN models.

[0156] Dynamic interest real-time tracking: The dynamic update of neighbor graph edge weights and threshold filtering mechanism enable the model to quickly respond to changes in user interests (such as shifting from "basic mathematics" to "machine learning"), improving recommendation real-time performance by 37%.

[0157] The dynamic attention mechanism captures differences in learning stages (such as the beginner and advanced stages) through positional encoding, improving personalized recommendation coverage by 18.5%.

[0158] Enhanced robustness in sparse data scenarios: The multi-view fusion and contrastive learning strategy supplements local sparse information with global semantics, and the recommendation accuracy remains above 85% even when users only browse 3-5 items.

[0159] The contrastive learning module uncovers potential associations through self-supervised tasks, improving the model's AUC by 29.8% in cold-start scenarios.

[0160] Global-local feature collaborative optimization: Comparative learning of global hypergraph embedding and local neighborhood embedding effectively resolves the contradiction between short-term behavioral noise and long-term interest bias, improving user satisfaction (NDCG@10) by 31.2%.

[0161] Cross-domain recommendation scalability: The Hypergraph Unified Semantic Space supports cross-disciplinary associations (such as associating "programming courses" with "mathematical modeling"), increasing the click-through rate (CTR) of cross-domain recommendations by 42%.

[0162] It should be noted that this embodiment proposes a personalized learning material recommendation model based on neighborhood and hypergraph collaborative learning (NGH-Rec). This model combines hypergraph convolution, neighbor convolution, and location-aware dynamic attention mechanisms, enabling it to capture explicit and implicit relationships between learning content, track the dynamic evolution of user learning interests in real time, and effectively handle sparse short-term interaction sequences. Experimental results show that this model outperforms existing mainstream models on multiple benchmark datasets.

[0163] This embodiment constructs a hypergraph view and a neighbor graph view by acquiring user session sequence data; extracts global high-order relationship features based on the hypergraph view; extracts local co-occurrence relationship features based on the neighbor graph view to generate enhanced co-occurrence relationship features; concatenates the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuses contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation; based on a global-local contrastive learning strategy, performs nonlinear projection on the dynamic learning interest embedding and the global semantic embedding, and optimizes the semantic consistency of the two types of embeddings through contrastive loss; jointly constructs a joint optimization objective function by combining the cross-entropy loss and contrastive learning loss of the recommendation task, generates personalized recommendation results, and achieves the technical effect of capturing explicit and implicit relationships between learning content.

[0164] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a program for recommending personalized learning content based on neighborhood and hypergraph collaboration. When the program for recommending personalized learning content based on neighborhood and hypergraph collaboration is executed by a processor, it implements the steps of the method for recommending personalized learning content based on neighborhood and hypergraph collaboration as described above.

[0165] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the personalized learning content recommendation system based on neighborhood and hypergraph collaboration of this application.

[0166] like Figure 3 As shown in the embodiments of this application, the personalized learning content recommendation system based on neighborhood and hypergraph collaboration includes:

[0167] Data acquisition module 10 is used to acquire user session sequence data to construct a hypergraph view and a neighbor graph view, and to generate a comprehensive learning content relationship network through the multi-view fusion module;

[0168] The hypergraph view module 20 is used to extract global high-order relation features based on the hypergraph view by combining multi-hop hypergraph convolution with the K-hop algorithm, and to perform complementary fusion with the co-occurrence relation features of the neighboring graph view.

[0169] The neighbor graph view module 30 is used to extract local co-occurrence relationship features based on the neighbor graph view by weighted multi-hop neighbor graph convolution, and filter noisy edges by combining a threshold mechanism to generate enhanced co-occurrence relationship features;

[0170] The generation module 40 is used to concatenate the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and to generate a dynamic learning interest embedding representation by fusing contextual information through a gated dynamic attention mechanism.

[0171] The optimization module 50 is used to perform nonlinear projection on the dynamic learning interest embedding and the global semantic embedding based on a global-local contrastive learning strategy, and optimize the semantic consistency of the two types of embeddings through contrastive loss.

[0172] The result generation module 60 is used to jointly optimize the cross-entropy loss and contrastive learning loss of the recommendation task, construct a joint optimization objective function, and dynamically adjust the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results.

[0173] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0174] This embodiment constructs a hypergraph view and a neighbor graph view by acquiring user session sequence data; extracts global high-order relationship features based on the hypergraph view; extracts local co-occurrence relationship features based on the neighbor graph view to generate enhanced co-occurrence relationship features; concatenates the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuses contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation; based on a global-local contrastive learning strategy, performs nonlinear projection on the dynamic learning interest embedding and the global semantic embedding, and optimizes the semantic consistency of the two types of embeddings through contrastive loss; jointly constructs a joint optimization objective function by combining the cross-entropy loss and contrastive learning loss of the recommendation task, generates personalized recommendation results, and achieves the technical effect of capturing explicit and implicit relationships between learning content.

[0175] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0176] In addition, for technical details not described in detail in this embodiment, please refer to the method for personalized learning content recommendation based on neighborhood and hypergraph collaboration provided in any embodiment of this application, which will not be repeated here.

[0177] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0178] The sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A personalized learning content recommendation method based on neighborhood and hypergraph collaboration, characterized in that, include: Acquire user session sequence data to construct a hypergraph view and a neighbor graph view, and generate a comprehensive learning content relationship network through a multi-view fusion module; Based on the hypergraph view, global high-order relation features are iteratively extracted by multi-hop hypergraph convolution combined with the K-hop algorithm, and then complementaryly fused with the co-occurrence relation features of the neighboring graph view. Based on the neighbor graph view, local co-occurrence relationship features are extracted through weighted multi-hop neighbor graph convolution, and noisy edges are filtered by a threshold mechanism to generate enhanced co-occurrence relationship features; The sequence position encoding of the user's learning path is concatenated with the semantic embedding of the learning content, and contextual information is fused through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation. Based on a global-local contrastive learning strategy, the dynamic learning interest embedding and the global semantic embedding are nonlinearly projected, and the semantic consistency of the two types of embeddings is optimized by contrastive loss. The cross-entropy loss and contrastive learning loss of the joint recommendation task are used to construct a joint optimization objective function. The parameters of the hypergraph and neighbor graph are dynamically adjusted through backpropagation to generate personalized recommendation results. The step of performing a nonlinear projection on the dynamically learned interest embedding and the global semantic embedding based on a global-local contrastive learning strategy, and optimizing the semantic consistency of the two types of embeddings through contrastive loss, includes: A local view is constructed based on the user's short-term learning path, and local behavioral features are extracted through neighborhood graph convolution to generate local embedding representations; A global view is constructed based on the user's long-term learning behavior, and global semantic features are extracted through hypergraph convolution to generate a global embedding representation; Nonlinear projection is performed on local and global embeddings to construct positive and negative sample pairs. The semantic consistency of positive sample pairs is maximized by a contrastive loss function, while negative sample pairs are separated. The step of concatenating the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fusing contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation includes: The positional encoding of the learned content sequence is concatenated with the semantic embedding of the learned content to generate a context-aware initial embedding; The initial embedding is nonlinearly transformed by a gated linear unit, and the attention weight is dynamically calculated by combining the average embedding of the user's current session. Invalid location information is filtered out using a masking mechanism to generate a dynamic embedding representation of the user's learning path. This dynamic embedding representation shares the feature space with the global-local contrastive learning module.

2. The method according to claim 1, characterized in that, The steps of acquiring user session sequence data to construct a hypergraph view and a neighbor graph view, and generating a comprehensive learning content relationship network through a multi-view fusion module, include: Acquire user session sequence data and construct a hypergraph view to connect multiple learning content nodes via hyperedges; Construct a neighbor graph view to build weighted edges for learning content co-occurring in the user's learning path; The multi-view fusion module performs weighted concatenation of high-order features of the hypergraph view and co-occurrence features of neighboring graph views to generate a comprehensive learning content relationship network.

3. The method according to claim 1, characterized in that, The step of iteratively extracting global high-order relation features based on the hypergraph view through multi-hop hypergraph convolution combined with the K-hop algorithm, and then complementaryally fusing them with the co-occurrence relation features of neighboring graph views, includes: Based on the hypergraph association matrix and node degree matrix, single-hop hypergraph convolution features are extracted through bidirectional information aggregation of node-hyperedge-node; By combining the K-hop algorithm for multi-hop iteration, high-order interaction information of multiple nodes within the hyperedge is aggregated layer by layer to generate global high-order relation features; The global high-order relationship features are input into the dynamic attention mechanism and attention weights are assigned to the co-occurrence features of the neighbor graph view to form complementary feature fusion.

4. The method according to claim 1, characterized in that, The step of extracting local co-occurrence relationship features based on the neighbor graph view through weighted multi-hop neighbor graph convolution, and filtering noisy edges using a threshold mechanism to generate enhanced co-occurrence relationship features includes: Based on the adjacency matrix and degree matrix, local co-occurrence relationship features are extracted through weighted graph convolution, where the edge weights are dynamically adjusted according to the co-occurrence frequency and user behavior time series. An adaptive threshold mechanism is introduced to filter out noise edges with co-occurrence frequencies below the dynamic threshold, while preserving strong co-occurrence relationships; By combining the K-hop algorithm to aggregate multi-hop neighbor information, enhanced co-occurrence relationship features are generated and cross-view interactions are performed with higher-order features of the hypergraph view.

5. The method according to claim 1, characterized in that, The steps for constructing a joint optimization objective function using cross-entropy loss and contrastive learning loss in the joint recommendation task, and dynamically adjusting the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results include: The cross-entropy loss of the recommendation task is calculated based on the matching degree between the final embedding and the candidate learned content; The similarity calculation is based on the comparison learning loss of local and global embeddings, and the similarity distribution is smoothed by temperature parameters; The two types of losses are weighted and summed, and the parameters of the hypergraph, neighbor graph, attention and contrastive learning modules are jointly optimized by gradient descent to construct a joint optimization objective function; Personalized recommendation results are generated by dynamically adjusting the parameters of the hypergraph and neighbor graph through backpropagation.

6. A personalized learning content recommendation system based on neighborhood and hypergraph collaboration, characterized in that, Performing the method as described in claim 1 includes: The data acquisition module is used to acquire user session sequence data to construct a hypergraph view and a neighbor graph view, and to generate a comprehensive learning content relationship network through the multi-view fusion module; The hypergraph view module is used to extract global high-order relation features based on the hypergraph view by iteratively extracting them through multi-hop hypergraph convolution combined with the K-hop algorithm, and to perform complementary fusion with the co-occurrence relation features of the neighboring graph views. The neighbor graph view module is used to extract local co-occurrence relationship features based on the neighbor graph view through weighted multi-hop neighbor graph convolution, and filter noisy edges by combining a threshold mechanism to generate enhanced co-occurrence relationship features; The representation generation module is used to concatenate the sequence position encoding of the user's learning path with the semantic embedding of the learning content, and fuse contextual information through a gated dynamic attention mechanism to generate a dynamic learning interest embedding representation; An optimization module is used to perform nonlinear projection on the dynamic learning interest embedding and the global semantic embedding based on a global-local contrastive learning strategy, and optimize the semantic consistency of the two types of embeddings through contrastive loss. The results generation module is used to jointly optimize the cross-entropy loss and contrastive learning loss of the recommendation task, construct a joint optimization objective function, and dynamically adjust the parameters of the hypergraph and neighbor graph through backpropagation to generate personalized recommendation results.

7. A computer device, characterized in that, The device includes a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 5.

Images