Course session recommendation method and device, equipment, storage medium and program product

Abstract

The invention discloses a course session recommendation method and device, equipment, a storage medium and a program product, and the method comprises the steps: carrying out the channel splicing of a semantic embedding matrix of course text description data and a user behavior feature matrix, and generating a multi-dimensional feature matrix; constructing an adjacent graph and a hypergraph between courses according to the multi-dimensional feature matrix, performing graph convolution operation based on the adjacent graph and the hypergraph, and generating a target embedded representation vector of each course node; generating an attention weight of each course node, performing weighted aggregation on the target embedded representation vector based on the attention weight, and performing course session recommendation based on the target embedded representation vector after weighted aggregation; according to the method, the curriculum semantic features and the user behavior features are fused, so that the feature expression ability of curriculum nodes is effectively improved, local neighborhood features between curriculums and global features in a hypergraph structure are captured, dynamic changes of user interests are accurately captured, and accurate curriculum session recommendation is achieved.

Description

Technical Field

[0001] The present invention relates to the field of online education technology, and in particular to a course conversation recommendation method, device, equipment, storage medium and program product. Background Art

[0002] With the rapid development of the online education industry, a large number of complex and diverse course resources have emerged on education platforms. How to mine courses that users are interested in from a large amount of course resources has become a key issue in the research of online education recommendation systems. In recent years, recommendation systems have gradually been applied to the field of online education, and a variety of course recommendation methods for online learning environments have been proposed to generate more personalized course recommendations. As an important branch of the recommendation system, conversational recommendation is suitable for the characteristics of diverse and real-time changing user needs in online education scenarios.

[0003] However, the existing conversational recommendation models still have some limitations. They mainly rely on the interaction data between users and courses, ignoring the rich semantic information contained in course resources, resulting in the recommendation results lacking a deep understanding of the characteristics of course content and difficulty in capturing the potential semantic associations between courses. In addition, when dealing with complex course relationships, the existing conversational recommendation models only focus on local information and are unable to capture the interactive relationships between multiple nodes. Therefore, they fail to effectively capture the complex relationships between courses and are unable to mine the implicit feature relationships between nodes, resulting in inaccurate course recommendations. Summary of the invention

[0004] The main purpose of the present invention is to provide a course conversation recommendation method, device, equipment, storage medium and program product, aiming to solve the technical problem that the prior art insufficiently utilizes the semantic information of courses, fails to effectively capture the complex associations and implicit feature connections between courses, and leads to inaccurate course recommendations.

[0005] To achieve the above object, the present invention provides a course conversation recommendation method, which comprises the following steps: Inputting the text description data of the course into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector, and generating a semantic embedding matrix based on the semantic feature vector; Conduct user behavior analysis on the user's historical learning interaction data, and build a user behavior feature matrix based on the user behavior analysis results; Channel-joining the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix; Constructing an adjacency graph and a hypergraph between courses according to the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; Performing graph convolution operations based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; Generate an attention weight for each course node, perform weighted aggregation on the target embedding representation vector based on the attention weight, and perform course session recommendation based on the weighted aggregated target embedding representation vector.

[0006] Optionally, the step of inputting the text description data of the course into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector, and generating a semantic embedding matrix based on the semantic feature vector includes: Collecting text description data of the course, and preprocessing the text description data, wherein the preprocessing includes word segmentation, stop word screening, and special character screening; The preprocessed text description data is input into the pre-trained deep language model for semantic quantization to obtain a semantic feature vector; The semantic feature vectors of each course are semantically embedded and integrated to generate an initial semantic matrix; The semantic feature vectors of each dimension in the initial semantic matrix are normalized to generate a semantic embedding matrix.

[0007] Optionally, performing user behavior analysis on the user's historical learning interaction data and constructing a user behavior feature matrix according to the user behavior analysis results include: Modeling the user's learning behavior into a course learning session sequence in chronological order based on the user's historical learning interaction data, wherein the course learning session sequence includes a set of learning courses for the user; Acquire the course learning time sequence feature of the user according to the sequence position coding information of the course learning session sequence; Analyzing the interaction frequency between the user and each course according to the course learning session sequence, and generating an explicit relationship weight matrix based on the interaction frequency, wherein the elements in the explicit relationship weight matrix represent the occurrence frequency of different courses in the same session; Mapping the course ID corresponding to each course into an initial behavior embedding vector, and constructing an initial behavior feature matrix based on the initial behavior embedding vector; The initial behavior feature matrix is ​​weighted and aggregated based on the course learning time sequence feature and the explicit relationship weight matrix to generate enhanced behavior features: in, Indicates enhanced behavioral characteristics, Representation and Courses The set of neighbor courses that appear in the same session, Indicates the course The position code, Indicates the course With Courses Frequency of occurrence in the same session, Indicates the course The initial behavior embedding vector of Construct a user behavior feature matrix based on the enhanced behavior features of each course.

[0008] Optionally, performing a graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node includes: Performing neighborhood aggregation operation on the adjacency graph through a graph convolutional network to obtain short-range dependencies between courses; Acquire local feature information of each course according to the short-range dependency relationship; Performing a graph convolution operation on the hypergraph through a graph convolution network to obtain hyperedge information; Acquire the interaction relationship between the student and each course based on the hyperedge information, and acquire the potential feature information of the course in the learning path based on the interaction relationship; Fusing the local feature information and the potential feature information to obtain a fused feature; The fused features are activated and normalized to generate a target embedding representation vector for each course node.

[0009] Optionally, generating the attention weights of each course node, performing weighted aggregation on the target embedding representation vector based on the attention weights, and performing course session recommendation based on the weighted aggregated target embedding representation vector comprises: Determine the cosine similarity between course nodes based on the semantic feature vector of each course node: in, Representation Node i With Node j The cosine similarity between and Respectively represent nodes i and nodes j The semantic feature vector of Determine semantic difference information between course nodes according to the cosine similarity; The attention weight of each course node is generated based on the semantic difference information and the position encoding information of each course node: in, represents the attention weight, Indicates position encoding information; Generating a cluster center of the target embedding representation vector; Calculate the cluster similarity between each course node and the cluster center; The target embedding representation vector is weightedly aggregated based on the cluster similarity and the attention weight: in, Represents the target embedding representation vector after weighted aggregation; Generating a behavior feature vector of the user based on the user's historical learning interaction data; The weighted aggregated target embedding representation vector and the behavior feature vector are input into a pre-trained recommendation model for course session recommendation.

[0010] Optionally, the step of inputting the weighted aggregated target embedding representation vector and the behavior feature vector into a pre-trained recommendation model for course session recommendation includes: The weighted aggregated target embedding representation vector and the behavior feature vector are input into the pre-trained recommendation model to calculate the predicted score of each course node: in, represents the prediction score, represents the behavior feature vector, Represents the course embedding feature matrix constructed by the target embedding representation vector after weighted aggregation; The label difference information between the predicted score and the true label is determined based on the cross entropy loss function: in, Indicates label difference information. represents the true label, represents the prediction score; The model parameters of the recommendation model are updated through the back-propagation algorithm: in, η represents the learning rate, represents the gradient of the loss function, represents the model parameters before updating, represents the updated model parameters; The recommendation model is optimized according to the updated model parameters, and course session recommendations are performed based on the optimized recommendation model.

[0011] In addition, to achieve the above-mentioned purpose, the present invention also proposes a course conversation recommendation device, which includes: A semantic analysis module, used to input the text description data of the course into a pre-trained deep language model for semantic quantification, obtain a semantic feature vector, and generate a semantic embedding matrix based on the semantic feature vector; User behavior analysis module, used to perform user behavior analysis on the user's historical learning interaction data and construct a user behavior feature matrix based on the user behavior analysis results; A feature splicing module, used for performing channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix; A graph construction module, used to construct an adjacency graph and a hypergraph between courses according to the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; A graph convolution processing module, used to perform a graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; The course recommendation module is used to generate the attention weight of each course node, and perform weighted aggregation on the target embedding representation vector based on the attention weight, and recommend course sessions based on the weighted aggregated target embedding representation vector.

[0012] In addition, to achieve the above-mentioned purpose, the present application also proposes a course conversation recommendation device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program is configured to implement the steps of the course conversation recommendation method described above.

[0013] In addition, to achieve the above-mentioned purpose, the present application also proposes a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the course conversation recommendation method described above are implemented.

[0014] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the course conversation recommendation method described above.

[0015] The present invention inputs the text description data of the course into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector, and generates a semantic embedding matrix based on the semantic feature vector; performs user behavior analysis on the user's historical learning interaction data, and constructs a user behavior feature matrix according to the user behavior analysis results; performs channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multidimensional feature matrix; constructs an adjacency graph and a hypergraph between courses based on the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; performs graph convolution operations based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; generates an attention vector for each course node. Weight, and based on the attention weight, weighted aggregation is performed on the target embedding representation vector, and course conversation recommendation is performed based on the weighted aggregated target embedding representation vector; because the present invention integrates course semantic features and user behavior features, thereby improving feature expression capabilities by integrating multi-source information, and constructs an adjacency graph and a hypergraph between courses according to the multi-dimensional feature matrix, thereby effectively improving the feature expression capabilities of course nodes in the graph structure, effectively capturing the multi-node interaction relationship in the hypergraph structure and the neighborhood local feature interaction relationship of each course node in the adjacency graph, and performing graph convolution operations based on the adjacency graph and the hypergraph, thereby accurately mining the explicit feature relationships and implicit feature relationships between course nodes, and accurately capturing the dynamic changes of user interests, thereby accurately performing course conversation recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0017] Figure 1 It is a structural diagram of a course conversation recommendation device in a hardware operating environment involved in an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of a first embodiment of a course conversation recommendation method of the present invention; Figure 3 A schematic diagram of a flow chart of a second embodiment of the course conversation recommendation method of the present invention; Figure 4 A schematic diagram of a flow chart of a third embodiment of a course conversation recommendation method of the present invention; Figure 5 A schematic diagram of a flow chart of a fourth embodiment of a course conversation recommendation method of the present invention; Figure 6 A schematic diagram of a flow chart of a fifth embodiment of a course conversation recommendation method of the present invention; Figure 7 It is a structural block diagram of the first embodiment of the course conversation recommendation device of the present invention.

[0018] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0020] Reference Figure 1 , Figure 1 A schematic diagram of the course session recommendation device structure of the hardware operating environment involved in the embodiment of the present invention.

[0021] like Figure 1 As shown, the course conversation recommendation 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. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation of the course session recommendation device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0023] like Figure 1 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a course session recommendation program.

[0024] exist Figure 1In the course conversation recommendation 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 the course conversation recommendation device of the present invention can be set in the course conversation recommendation device, and the course conversation recommendation device calls the course conversation recommendation program stored in the memory 1005 through the processor 1001, and executes the course conversation recommendation method provided by the embodiment of the present invention.

[0025] The embodiment of the present invention provides a course conversation recommendation method, referring to Figure 2 , Figure 2 Schematic diagram of the flow chart of the first embodiment of the course conversation recommendation method of the present invention.

[0026] In the first embodiment of the course conversation recommendation method of the present invention, the course conversation recommendation method comprises the following steps: Step S10: Input the text description data of the course into the pre-trained deep language model for semantic quantization, obtain a semantic feature vector, and generate a semantic embedding matrix based on the semantic feature vector.

[0027] It should be understood that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or a terminal electronic device capable of realizing the above functions, etc. The following takes the course conversation recommendation device (referred to as the recommendation device) as an example to illustrate this embodiment and the following embodiments.

[0028] It can be understood that this embodiment can convert the text description of the course (including title, introduction and content) into a high-dimensional semantic vector through a pre-trained deep language model (such as a BERT model) to generate a semantic embedding matrix for the course.

[0029] In some embodiments, the recommendation device can extract semantic information from the text description of the course through the semantic generation module to provide high-quality semantic embedding representation for subsequent node feature generation. Specifically, the module uses a pre-trained deep language model (such as BERT) to process text information such as the title and introduction of the course, embed it into a high-dimensional semantic space, and generate a semantic embedding vector for each course. By encoding the text descriptions of all courses, a semantic embedding matrix E is generated, in which each row represents the semantic features of a course. In order to ensure the quality of the semantic embedding, the module also normalizes the embedding vector to make it consistent in the feature space. In addition, the semantic generation module dynamically adjusts the dimension of the semantic embedding to ensure that it can adapt to course data sets of different sizes. The introduction of the semantic generation module makes up for the neglect of course content characteristics in traditional recommendation methods, and provides rich semantic information support for subsequent node feature fusion and graph convolution operations.

[0030] Step S20: Performing user behavior analysis on the user's historical learning interaction data, and constructing a user behavior feature matrix based on the user behavior analysis results.

[0031] It is understandable that this embodiment can construct a user behavior sequence based on the historical interaction records of the user's course learning, and extract the time series information in the sequence and the explicit relationship between courses to generate a user behavior feature matrix.

[0032] In some embodiments, the recommendation device can extract behavioral features from the user's learning history through a student behavior sequence generation module to build an interactive relationship between the user and the course. The module first collects the user's learning records and generates a user's session sequence in chronological order. Each session sequence represents a set of courses that the user has learned in a specific time period. Subsequently, the module combines the timestamp information to extract the temporal features of the user's behavior and capture the dynamic changes in the user's interests. In addition, the module also generates a user behavior matrix by analyzing the frequency of interaction between the user and the course. , which is used to characterize the explicit relationship between users and courses. In order to enhance the robustness of the data, the module completes the sparse interaction data to ensure that the generated behavior sequence can fully reflect the user's learning path. The user behavior sequence generation module provides basic data support for subsequent graph construction and enhancement, and also provides dynamic features of user interests for the course recommendation model.

[0033] Step S30: performing channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix.

[0034] It can be understood that this embodiment can splice the semantic information of the course node with the user behavior characteristics to generate a node embedding representation that integrates semantics and relationship features for the construction of the course conversation graph.

[0035] In some embodiments, the recommendation device can generate a high-dimensional course node feature representation by fusing semantic information and user interaction relationships through a node feature generation module. Specifically, the module first embeds the course semantics E output by the semantic generation module with the interaction relationship features output by the user behavior sequence generation module. Through feature concatenation, the module generates node feature representations that integrate semantics and relationships. ,Then, the module performs nonlinear transformation on the node features to ,further improve the expressiveness of the features.,The design of the node feature generation module aims to ,make up for the problem of split modeling of course content and ,user behavior in traditional recommendation methods, and provide high-quality node input ,features for subsequent graph convolution operations.

[0036] In some embodiments, the recommendation device may embed the semantic feature vector in the semantic embedding matrix and the behavior feature vector in the user behavior feature matrix The channel splicing method is used to fuse and generate a fusion feature matrix containing multiple initial fusion feature vectors. : The feature scale is unified through layer normalization (LayerNorm, LN), and a multi-dimensional feature matrix is ​​generated based on the fused feature vector after layer normalization: in, is a dimension reduction matrix, representing the initial fusion feature vector, Represents the fused feature vector after normalization of the layer.

[0037] Step S40: construct an adjacency graph and a hypergraph between courses according to the multi-dimensional feature matrix.

[0038] It should be noted that the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node.

[0039] It should be noted that the recommendation device can construct a course adjacency graph based on the neighbor co-occurrence relationship between courses, and the adjacency graph is used to capture local aggregation features. A high-order course relationship graph based on a hypergraph is constructed to model the relationship between multiple nodes.

[0040] In some embodiments, the recommendation device may extract the global clustering center in the course embedding based on the mean shift clustering mechanism, and enhance the global distribution perception capability of the node features.

[0041] In some embodiments, the recommendation device can construct a relationship graph between courses through a graph construction and enhancement module, and enhance the graph structure through multiple mechanisms. First, the module generates an adjacency graph G based on the user behavior sequence to capture the local relationship between courses. Secondly, the module constructs a hypergraph H through the "student-course" interaction relationship, in which the nodes represent courses and the hyperedges represent all courses that the students have studied, which is used to model high-order associations between courses. In addition, the module introduces a mean shift clustering mechanism to dynamically cluster the embeddings of course nodes, generate global cluster centers, and enhance the global feature expression capabilities of nodes by calculating the similarity between nodes and cluster centers. By combining the structural information of the adjacency graph and the hypergraph, the graph construction and enhancement module can comprehensively capture the local and global relationships between courses, providing rich structural information for subsequent graph convolution operations.

[0042] In some embodiments, the recommendation device can capture high-order relationships between courses by modeling the student's learning behavior as a hypergraph. Hypergraph definition: The node set V represents the course, and the hyperedge set Represents the student's session data, and each hyperedge connects all the courses a student has studied. The adjacency matrix H of the hypergraph is defined as: Among them, if the course ,but If the course ,but .

[0043] By counting the co-occurrence of courses in user sessions, an adjacency graph is constructed to capture the local aggregation characteristics between courses. The elements of the adjacency matrix A are Indicates the course and Co-occurrence frequency in student conversations: Among them, if the course and courses If they are jointly selected by the same student, If the course and courses If they are not selected by the same student, .

[0044] In some embodiments, in order to enhance the global distribution perception capability of node embedding, the recommendation device can use the mean shift clustering method to extract the global cluster center in the course embedding and dynamically enhance the node features. According to the distribution characteristics of the node embedding, the bandwidth h of the mean shift is dynamically adjusted: in, is the mean of node embedding; Generate cluster center C and cluster label L through mean shift operation: Among them, MS represents the mean shift operation; By calculating the similarity between the node embedding X and the cluster center C, and generating the cluster feature F by weighting: in, is the learnable weight matrix, is the bias term.

[0045] Step S50: Perform graph convolution operations based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node.

[0046] It can be understood that this embodiment can respectively aggregate information on the adjacency graph and the hypergraph through the graph convolutional network, combine local features with global features, and generate the final embedded representation of the course node.

[0047] In some embodiments, the recommendation device can perform information aggregation and relationship modeling on the features of the course nodes through a graph convolution processing module. The module applies graph convolution operations to the adjacency graph and hypergraph respectively to extract local features and high-order associations between courses. For the adjacency graph, the module captures the short-range dependencies of courses through neighborhood aggregation operations; for the hypergraph, the module uses hyperedge information to model the interactions between courses participated by the same students, thereby capturing the potential connections between courses in the learning path. Subsequently, the module fuses the output features of the adjacency graph convolution and the hypergraph convolution to generate the final embedded representation of the course node. The design of the graph convolution processing module aims to make up for the deficiencies in the modeling of course relationships in traditional methods and provide a more comprehensive node feature representation for the recommendation model.

[0048] Step S60: Generate attention weights for each course node, perform weighted aggregation on the target embedding representation vector based on the attention weights, and perform course session recommendations based on the weighted aggregated target embedding representation vector.

[0049] It can be understood that this embodiment can use the dynamic attention mechanism to optimize the weight distribution between course nodes, combine node embedding to generate prediction results for course recommendations, and output a personalized course recommendation list.

[0050] In some embodiments, the recommendation device can further optimize the embedding representation of course nodes by combining the mean shift clustering mechanism and the dynamic attention mechanism through the node embedding enhancement module. First, the module uses mean shift clustering to generate the cluster center of the course embedding, and dynamically enhances the node embedding by calculating the similarity between the node and the cluster center, thereby improving the node's perception of global characteristics. Secondly, the module introduces a position-aware dynamic attention mechanism, which dynamically adjusts the weight distribution between nodes by combining the semantic features and position encoding information of the nodes to capture implicit relationships and semantic differences. Finally, the module uses the enhanced node embedding as the input of the recommendation model. The design of the node embedding enhancement module effectively improves the node expression ability in sparse data scenarios and provides a more accurate feature representation for course recommendations.

[0051] In some embodiments, the recommendation device can predict the score of the node embedding through the recommendation generation module to generate a personalized course recommendation list for the user. The module first uses the node embedding and user features to calculate the predicted score of the course, and converts the score into a probability distribution through the softmax function. Subsequently, the module selects the courses that best suit the user's interests based on the scoring results and generates a recommendation list. In addition, the module introduces a cross entropy loss function to optimize the model, minimize the divergence between the predicted value and the true label, and thus improve the accuracy of the recommendation results. The design of the recommendation generation module aims to convert the high-quality node embeddings generated by the aforementioned modules into actual recommendation results, providing users with accurate course recommendation services.

[0052] This embodiment obtains semantic feature vectors by inputting text description data of courses into a pre-trained deep language model for semantic quantization, and generates a semantic embedding matrix based on the semantic feature vectors; performs user behavior analysis on historical learning interaction data of users, and constructs a user behavior feature matrix based on the results of user behavior analysis; performs channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multidimensional feature matrix; constructs an adjacency graph and a hypergraph between courses based on the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; performs graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; generates an attention vector for each course node. Weight, and weighted aggregation is performed on the target embedding representation vector based on the attention weight, and course conversation recommendation is performed based on the target embedding representation vector after weighted aggregation; since this embodiment integrates course semantic features and user behavior features, thereby improving feature expression capabilities by integrating multi-source information, and constructing adjacency graphs and hypergraphs between courses according to the multi-dimensional feature matrix, the feature expression capabilities of course nodes in the graph structure are effectively improved, and the multi-node interaction relationships in the hypergraph structure and the neighborhood local feature interaction relationships of each course node in the adjacency graph are effectively captured. Graph convolution operations are performed based on the adjacency graph and the hypergraph, thereby accurately mining the explicit feature relationships and implicit feature relationships between course nodes, and accurately capturing the dynamic changes of user interests, thereby accurately performing course conversation recommendations.

[0053] refer to Figure 3 , Figure 3 Schematic diagram of the flow chart of the second embodiment of the course conversation recommendation method of the present invention.

[0054] Based on the above first embodiment, in the second embodiment of the course conversation recommendation method of the present invention, the step S10 further includes: Step S101: Collect text description data of the course and pre-process the text description data.

[0055] It should be noted that the preprocessing includes word segmentation, stop word screening and special character screening.

[0056] It is understandable that this embodiment can collect text description data of courses, including course titles, introductions, and other related content, and perform preprocessing operations on the text data, such as word segmentation, stop word removal, and special character removal, to ensure the standardization and consistency of the data and provide high-quality input for subsequent semantic embedding generation.

[0057] Step S102: input the preprocessed text description data into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector.

[0058] It should be understood that this embodiment can use a pre-trained deep language model (such as a BERT model) to encode the pre-processed course text data and embed it into a high-dimensional semantic space. The model processes the text description of each course and generates a semantic embedding vector , used to characterize the semantic features of the course.

[0059] Step S103: semantically embed and integrate the semantic feature vectors of each course to generate an initial semantic matrix.

[0060] It can be understood that in this embodiment, the semantic embedding vectors of all courses can be integrated into a semantic embedding matrix E, where each row represents the semantic features of a course. The dimension of this matrix is ,in m Indicates the total number of courses, d Represents the dimension of the semantic embedding vector.

[0061] Step S104: normalizing the semantic feature vectors of each dimension in the initial semantic matrix to generate a semantic embedding matrix.

[0062] It should be understood that this embodiment can normalize the generated initial semantic matrix E to ensure that the semantic features of different courses are consistent in the feature space. The normalization operation can improve the sensitivity of the model to the feature and avoid the degradation of the model performance due to the difference in feature scale.

[0063] This embodiment collects text description data of courses and preprocesses the text description data, wherein the preprocessing includes word segmentation, stop word screening and special character screening, and inputs the preprocessed text description data into a pre-trained deep language model for semantic quantization to obtain semantic feature vectors, and semantically embeds and integrates the semantic feature vectors of each course to generate an initial semantic matrix, and normalizes the semantic feature vectors of each dimension in the initial semantic matrix to generate a semantic embedding matrix, thereby accurately capturing the rich semantic features in the course content, providing rich semantic information support for node feature fusion and graph convolution operations, and effectively improving the accuracy of course recommendations.

[0064] refer to Figure 4 , Figure 4 This is a flow chart of the third embodiment of the course conversation recommendation method of the present invention.

[0065] Based on the above embodiments, in the third embodiment of the course conversation recommendation method of the present invention, the step S20 further includes: Step S201: Modeling the user's learning behavior into a course learning session sequence in chronological order based on the user's historical learning interaction data.

[0066] It should be noted that the course learning session sequence includes the user's learning course set.

[0067] It can be understood that, in this embodiment, the learning behavior of each student can be modeled as a course learning session sequence in chronological order according to the historical interaction record of the user's course learning: , each student The session is represented as ,in Indicates students Learn in chronological order i courses, each session contains the masked validation set , used to predict the subsequent recommendation target of the last course.

[0068] Step S202: Acquire the course learning time sequence characteristics of the user according to the sequence position coding information of the course learning session sequence.

[0069] It should be understood that the present embodiment can capture the time sequence characteristics of course learning through sequence position coding, integrate the time sequence information into the behavior sequence representation, and generate the time sequence characteristics of course learning.

[0070] Step S203: analyzing the interaction frequency between the user and each course according to the course learning session sequence, and generating an explicit relationship weight matrix based on the interaction frequency.

[0071] It should be noted that the elements in the explicit relationship weight matrix represent the frequency of occurrence of different courses in the same session.

[0072] It is understandable that this embodiment can count the co-occurrence frequency of courses in the sequence to generate an explicit relationship weight matrix ,in Indicates the course and How often it occurs in the same session.

[0073] Step S204: Map the course ID corresponding to each course to an initial behavior embedding vector, and construct an initial behavior feature matrix based on the initial behavior embedding vector.

[0074] It should be understood that this embodiment can map the course ID to the initial behavior embedding vector , forming a behavioral feature matrix (m is the total number of courses).

[0075] Step S205: performing weighted aggregation on the initial behavior feature matrix based on the course learning time sequence feature and the explicit relationship weight matrix to generate enhanced behavior features.

[0076] It can be understood that this embodiment can combine temporal coding and co-occurrence weights to generate enhanced behavior features through weighted aggregation: in, Indicates enhanced behavioral characteristics, Representation and Courses The collection of neighbor courses that appear in the same session, Indicates the course The position code, Indicates the course With Courses Frequency of occurrence in the same session, Indicates the course The initial behavior embedding vector of .

[0077] Step S206: Construct a user behavior feature matrix based on the enhanced behavior features of each course.

[0078] It should be understood that this embodiment can use a sliding window mechanism to extract local sequence patterns, capture short-term changes in user interests, and compress long sequences through a gating mechanism to generate a time-series-aware user behavior feature matrix. .

[0079] This embodiment models the user's learning behavior into a course learning session sequence in chronological order based on the user's historical learning interaction data, and the course learning session sequence includes the user's learning course set; The course learning time sequence characteristics of the user are obtained according to the sequence position coding information of the course learning session sequence, the interaction frequency between the user and each course is analyzed according to the course learning session sequence, and an explicit relationship weight matrix is ​​generated based on the interaction frequency, the course ID corresponding to each course is mapped to an initial behavior embedding vector, and an initial behavior feature matrix is ​​constructed based on the initial behavior embedding vector, the initial behavior feature matrix is ​​weighted and aggregated based on the course learning time sequence characteristics and the explicit relationship weight matrix to generate enhanced behavior features, and a user behavior feature matrix is ​​constructed based on the enhanced behavior features of each course, thereby providing basic data support for subsequent graph construction and enhancement, and providing the course recommendation model with dynamic characteristics of user interests, realizing accurate analysis of user dynamic behavior and interest characteristics, and improving the accuracy of course recommendations.

[0080] refer to Figure 5 , Figure 5 This is a flow chart of the fourth embodiment of the course conversation recommendation method of the present invention.

[0081] Based on the above embodiments, in the fourth embodiment of the course conversation recommendation method of the present invention, the step S50 further includes: Step S501: performing a neighborhood aggregation operation on the adjacency graph through a graph convolutional network to obtain short-range dependencies between courses; Step S502: Obtain local feature information of each course according to the short-range dependency relationship.

[0082] It should be noted that the adjacency graph is used to capture the local characteristics and short-range dependencies between courses. Through the adjacency graph convolution operation, the neighborhood information of the course nodes is aggregated and the node feature representation is updated. The adjacency graph is processed using a graph convolutional network, and the update rule is: in, is the degree matrix of the adjacency matrix, is the learnable weight matrix, is the activation function, Graph convolution output features representing the adjacency graph.

[0083] Step S503: performing a graph convolution operation on the hypergraph through a graph convolution network to obtain hyper-edge information; Step S504: acquiring the interaction relationship between the student and each course based on the hyperedge information, and acquiring the potential feature information of the course in the learning path based on the interaction relationship.

[0084] It should be noted that the hypergraph is used to capture high-order associations between courses and model potential connections in students’ learning paths. Through the hypergraph convolution operation, the hyperedge information is used to further enhance the feature expression ability of the course nodes. The hypergraph convolution updates the node features through the following formula: in, and are the degree matrices of hypergraph nodes and hyperedges respectively, Represents the features output by the graph convolution operation on the hypergraph.

[0085] Step S505: Fusing the local feature information and the potential feature information to obtain a fused feature.

[0086] It should be noted that after completing the convolution processing of the adjacency graph and the hypergraph, the output features of the two parts are fused to generate the final embedding representation of the course node. Adjacency graph convolution output Perform weighted fusion: in, αis an adjustable weight parameter used to balance the contribution of local features and global features. Represents fusion features.

[0087] Step S506: Activate and normalize the fused features to generate a target embedding representation vector for each course node.

[0088] It can be understood that this embodiment ensures the stability and robustness of the embedded representation by activating and normalizing the fused features.

[0089] In this embodiment, a neighborhood aggregation operation is performed on the adjacency graph through a graph convolution network to obtain short-range dependencies between courses, local feature information of each course is obtained according to the short-range dependencies, a graph convolution operation is performed on the hypergraph through a graph convolution network to obtain hyperedge information, an interaction relationship between students and courses is obtained based on the hyperedge information, and potential feature information of courses in the learning path is obtained based on the interaction relationship, the local feature information and the potential feature information are fused to obtain fused features, the fused features are activated and normalized, and a target embedding representation vector of each course node is generated; since this embodiment captures the short-range dependencies of courses through a neighborhood aggregation operation, the interaction between courses participated by the same students is modeled based on hyperedge information, thereby capturing the potential connection of courses in the learning path, the output features of the adjacency graph convolution and the hypergraph convolution are fused to generate the final embedding representation of the course node, thereby combining the local feature information and the global feature information of the course for embedded feature representation, providing a more comprehensive node feature representation for the recommendation model, and effectively improving the accuracy of course recommendation.

[0090] refer to Figure 6 , Figure 6 This is a flowchart of the fifth embodiment of the course conversation recommendation method of the present invention.

[0091] Based on the above embodiments, in the fifth embodiment of the course conversation recommendation method of the present invention, the step S60 further includes: Step S601: determining the cosine similarity between each course node according to the semantic feature vector of each course node; Step S602: Determine semantic difference information between course nodes according to the cosine similarity.

[0092] It should be noted that this embodiment can optimize the weight distribution between course nodes based on the dynamic attention mechanism through dynamic attention weight calculation, capture implicit relationships and semantic differences, and first calculate the nodes through cosine similarity. i and j The similarity between the course nodes is based on cosine similarity to capture the semantic difference information between the course nodes: in, Representation Node i With Node j The cosine similarity between and Respectively represent nodes i and nodes j The semantic feature vector of .

[0093] Step S603: Generate the attention weight of each course node based on the semantic difference information and the position encoding information of each course node.

[0094] It should be noted that this embodiment can dynamically adjust the weight distribution by introducing position coding and combining the semantic features of the nodes and the position coding information: in, represents the attention weight, Indicates position encoding information; Step S604: Generate the cluster center of the target embedding representation vector.

[0095] In some embodiments, the recommendation device may generate cluster centers C and cluster labels L through a mean shift operation: Here, MS represents the mean shift operation.

[0096] Step S605: Calculate the clustering similarity between each course node and the cluster center.

[0097] In some embodiments, the recommendation device may calculate the similarity between the node embedding X and the cluster center C, and generate the cluster feature F by weighting: in, is the learnable weight matrix, is the bias term.

[0098] Step S606: performing weighted aggregation on the target embedding representation vector based on the cluster similarity and the attention weight.

[0099] It can be understood that in this embodiment, the node embeddings can be weighted and aggregated by the attention weights to generate session embeddings and capture the dynamic changes of user interests. The weighted aggregation formula is as follows: in, Represents the target embedding representation vector after weighted aggregation.

[0100] Step S607: Generate a behavior feature vector of the user based on the user's historical learning interaction data.

[0101] In some embodiments, historical learning interaction data may include the course IDs that the user has studied, the length of time the user has studied each course, the user's speaking activity in the comment section of each course, the user's historical learning search records, the user's collection of notes, etc.

[0102] In some embodiments, the recommendation device may extract behavioral features based on the user's historical learning interaction data, and the behavioral features include: learning progress-related features, interactivity features, learning performance features, interest preference features, learning habit features, etc. The behavioral features are encoded and converted, and then normalized or standardized to ensure that data of different scales can be processed within a comparable range, thereby generating a behavioral feature vector.

[0103] Step S608: Input the weighted aggregated target embedding representation vector and the behavior feature vector into the pre-trained recommendation model for course session recommendation.

[0104] In some embodiments, the recommendation device may input the weighted aggregated target embedding representation vector and the behavioral feature vector into a pre-trained recommendation model for course prediction scoring, and convert the score into a probability distribution through a softmax function, and make course session recommendations based on the probability distribution information.

[0105] Furthermore, in order to improve the recommendation accuracy of the recommendation model and provide users with more accurate recommended course push, in some embodiments, the above step S608 may include: Step S6081: input the weighted aggregated target embedding representation vector and the behavior feature vector into the pre-trained recommendation model to calculate the predicted score of each course node; Step S6082: determining label difference information between the predicted score and the true label based on a cross entropy loss function; Step S6083: updating the model parameters of the recommendation model through a back propagation algorithm; Step S6084: Optimize the recommendation model according to the updated model parameters, and make course session recommendations based on the optimized recommendation model.

[0106] It should be noted that the recommendation device can predict the user's preference rating for the course based on the generated session embedding. Rating calculation is based on the user embedding feature matrix and the course embedding feature matrix The interaction generates a prediction score , refer to the following formula: in, represents the prediction score, represents the behavior feature vector, Represents the course embedding feature matrix constructed by the target embedding representation vector after weighted aggregation.

[0107] It is understandable that the recommendation device can optimize the model parameters through the loss function to generate a personalized recommendation list. The cross entropy loss function is used to measure the gap between the predicted score and the true label: in, Indicates label difference information. represents the true label, Represents the prediction score.

[0108] It should be understood that the recommendation device can adjust the model parameters through the back-propagation algorithm: in, η represents the learning rate, represents the gradient of the loss function, represents the model parameters before updating, Represents the updated model parameters.

[0109] This embodiment determines the cosine similarity between each course node according to the semantic feature vector of each course node, determines the semantic difference information between each course node according to the cosine similarity, generates the attention weight of each course node based on the semantic difference information and the position coding information of each course node, generates the cluster center of the target embedding representation vector, calculates the clustering similarity between each course node and the cluster center, performs weighted aggregation on the target embedding representation vector based on the clustering similarity and the attention weight, generates the user's behavior feature vector based on the user's historical learning interaction data, and inputs the weighted aggregated target embedding representation vector and the behavior feature vector into the pre-trained recommendation model for course session recommendation; because this embodiment dynamically adjusts the weight distribution between nodes and captures implicit relationships and semantic differences by combining the semantic features and position coding information of nodes, it effectively improves the node expression ability in sparse data scenarios and provides a more accurate feature representation for course recommendation.

[0110] In addition, an embodiment of the present invention further proposes a computer-readable storage medium, on which a course conversation recommendation program is stored. When the course conversation recommendation program is executed by a processor, the steps of the course conversation recommendation method described above are implemented.

[0111] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.

[0112] The computer-readable storage medium may be included in the course conversation recommendation device; or may exist independently without being assembled into the course conversation recommendation device.

[0113] In addition, an embodiment of the present invention further proposes a computer program product, including a course conversation recommendation program, which implements the steps of the course conversation recommendation method described above when executed by a processor.

[0114] The specific implementation of the computer program product of the present invention is basically the same as the various embodiments of the above-mentioned course conversation recommendation method, and will not be repeated here.

[0115] Reference Figure 7 , Figure 7 It is a structural block diagram of the first embodiment of the course conversation recommendation device of the present invention.

[0116] like Figure 7 As shown, the course conversation recommendation device proposed in the embodiment of the present invention includes: The semantic analysis module 10 is used to input the text description data of the course into the pre-trained deep language model for semantic quantification, obtain a semantic feature vector, and generate a semantic embedding matrix based on the semantic feature vector; A user behavior analysis module 20 is used to perform user behavior analysis on the user's historical learning interaction data and construct a user behavior feature matrix based on the user behavior analysis results; A feature splicing module 30, used for performing channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix; A graph construction module 40, for constructing an adjacency graph and a hypergraph between courses according to the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; A graph convolution processing module 50, configured to perform a graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; The course recommendation module 60 is used to generate the attention weight of each course node, and perform weighted aggregation on the target embedding representation vector based on the attention weight, and perform course session recommendation based on the weighted aggregated target embedding representation vector.

[0117] Furthermore, the semantic analysis module 10 is also used to collect text description data of the course and preprocess the text description data, wherein the preprocessing includes word segmentation, stop word screening and special character screening; the preprocessed text description data is input into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector; the semantic feature vectors of each course are semantically embedded and integrated to generate an initial semantic matrix; the semantic feature vectors of each dimension in the initial semantic matrix are normalized to generate a semantic embedding matrix.

[0118] Furthermore, the user behavior analysis module 20 is also used to model the user's learning behavior as a course learning session sequence in chronological order based on the user's historical learning interaction data, and the course learning session sequence includes the user's learning course set; obtain the user's course learning time sequence characteristics according to the sequence position encoding information of the course learning session sequence; analyze the interaction frequency between the user and each course according to the course learning session sequence, and generate an explicit relationship weight matrix based on the interaction frequency, and the elements in the explicit relationship weight matrix represent the frequency of occurrence of different courses in the same session; map the course ID corresponding to each course to an initial behavior embedding vector, and construct an initial behavior feature matrix based on the initial behavior embedding vector; perform weighted aggregation on the initial behavior feature matrix based on the course learning time sequence characteristics and the explicit relationship weight matrix to generate enhanced behavior features: in, Indicates enhanced behavioral characteristics, Representation and Courses The set of neighbor courses that appear in the same session, Indicates the course The position code, Indicates the course With Courses Frequency of occurrence in the same session, Indicates the course The initial behavior embedding vector of Construct a user behavior feature matrix based on the enhanced behavior features of each course.

[0119] Furthermore, the graph convolution processing module 50 is also used to perform neighborhood aggregation operations on the adjacency graph through a graph convolution network to obtain short-range dependencies between courses; obtain local feature information of each course based on the short-range dependencies; perform graph convolution operations on the hypergraph through a graph convolution network to obtain hyperedge information; obtain the interaction relationship between students and courses based on the hyperedge information, and obtain potential feature information of courses in the learning path based on the interaction relationship; fuse the local feature information and the potential feature information to obtain fused features; activate and normalize the fused features to generate a target embedding representation vector for each course node.

[0120] Furthermore, the course recommendation module 60 is also used to determine the cosine similarity between each course node according to the semantic feature vector of each course node: in, Representation Node i With Node j The cosine similarity between and Respectively represent nodes i and nodes j The semantic feature vector of Determine semantic difference information between course nodes according to the cosine similarity; The attention weight of each course node is generated based on the semantic difference information and the position encoding information of each course node: in, represents the attention weight, Indicates position encoding information; Generating a cluster center of the target embedding representation vector; Calculate the cluster similarity between each course node and the cluster center; The target embedding representation vector is weightedly aggregated based on the cluster similarity and the attention weight: in, Represents the target embedding representation vector after weighted aggregation; Generating a behavior feature vector of the user based on the user's historical learning interaction data; The weighted aggregated target embedding representation vector and the behavior feature vector are input into a pre-trained recommendation model for course session recommendation.

[0121] Furthermore, the course recommendation module 60 is also used to input the weighted aggregated target embedding representation vector and the behavior feature vector into the pre-trained recommendation model to calculate the predicted score of each course node: in, represents the prediction score, represents the behavior feature vector, Represents the course embedding feature matrix constructed by the target embedding representation vector after weighted aggregation; The label difference information between the predicted score and the true label is determined based on the cross entropy loss function: in, Indicates label difference information. represents the true label, represents the prediction score; The model parameters of the recommendation model are updated through the back-propagation algorithm: in, η represents the learning rate, represents the gradient of the loss function, represents the model parameters before updating, represents the updated model parameters; The recommendation model is optimized according to the updated model parameters, and course session recommendations are performed based on the optimized recommendation model.

[0122] This embodiment obtains semantic feature vectors by inputting text description data of courses into a pre-trained deep language model for semantic quantization, and generates a semantic embedding matrix based on the semantic feature vectors; performs user behavior analysis on historical learning interaction data of users, and constructs a user behavior feature matrix based on the results of user behavior analysis; performs channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multidimensional feature matrix; constructs an adjacency graph and a hypergraph between courses based on the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; performs graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; generates an attention vector for each course node. Weight, and weighted aggregation is performed on the target embedding representation vector based on the attention weight, and course conversation recommendation is performed based on the target embedding representation vector after weighted aggregation; since this embodiment integrates course semantic features and user behavior features, thereby improving feature expression capabilities by integrating multi-source information, and constructing adjacency graphs and hypergraphs between courses according to the multi-dimensional feature matrix, the feature expression capabilities of course nodes in the graph structure are effectively improved, and the multi-node interaction relationships in the hypergraph structure and the neighborhood local feature interaction relationships of each course node in the adjacency graph are effectively captured. Graph convolution operations are performed based on the adjacency graph and the hypergraph, thereby accurately mining the explicit feature relationships and implicit feature relationships between course nodes, and accurately capturing the dynamic changes of user interests, thereby accurately performing course conversation recommendations.

[0123] The course conversation recommendation device provided by the present application adopts the course conversation recommendation method in the above embodiment, which can solve the technical problem of course conversation recommendation. Compared with the prior art, the beneficial effects of the course conversation recommendation device provided by the present application are the same as the beneficial effects of the course conversation recommendation method provided by the above embodiment, and other technical features in the course conversation recommendation device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0124] It should be understood that the above is only an example and does not constitute any limitation on the technical solution of the present invention. In specific applications, technicians in this field can make settings as needed, and the present invention does not limit this.

[0125] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0126] In addition, for technical details not fully described in this embodiment, reference can be made to the course conversation recommendation method provided in any embodiment of the present invention, and will not be repeated here.

[0127] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0128] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0129] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0130] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A course conversation recommendation method, characterized in that: The course session recommendation method comprises: Inputting the text description data of the course into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector, and generating a semantic embedding matrix based on the semantic feature vector; Conduct user behavior analysis on the user's historical learning interaction data, and build a user behavior feature matrix based on the user behavior analysis results; Channel-joining the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix; Constructing an adjacency graph and a hypergraph between courses according to the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; Performing graph convolution operations based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; Generate an attention weight for each course node, perform weighted aggregation on the target embedding representation vector based on the attention weight, and perform course session recommendation based on the weighted aggregated target embedding representation vector.

2. The course conversation recommendation method according to claim 1, characterized in that: The step of inputting the text description data of the course into a pre-trained deep language model for semantic quantization to obtain a semantic feature vector, and generating a semantic embedding matrix based on the semantic feature vector includes: Collecting text description data of the course, and preprocessing the text description data, wherein the preprocessing includes word segmentation, stop word screening, and special character screening; The preprocessed text description data is input into the pre-trained deep language model for semantic quantization to obtain a semantic feature vector; The semantic feature vectors of each course are semantically embedded and integrated to generate an initial semantic matrix; The semantic feature vectors of each dimension in the initial semantic matrix are normalized to generate a semantic embedding matrix.

3. The course conversation recommendation method according to claim 1, characterized in that: The user behavior analysis is performed on the user's historical learning interaction data, and a user behavior feature matrix is ​​constructed according to the user behavior analysis results, including: Modeling the user's learning behavior into a course learning session sequence in chronological order based on the user's historical learning interaction data, wherein the course learning session sequence includes a set of learning courses for the user; Acquire the course learning time sequence feature of the user according to the sequence position coding information of the course learning session sequence; Analyzing the interaction frequency between the user and each course according to the course learning session sequence, and generating an explicit relationship weight matrix based on the interaction frequency, wherein the elements in the explicit relationship weight matrix represent the occurrence frequency of different courses in the same session; Mapping the course ID corresponding to each course into an initial behavior embedding vector, and constructing an initial behavior feature matrix based on the initial behavior embedding vector; The initial behavior feature matrix is ​​weighted and aggregated based on the course learning time sequence feature and the explicit relationship weight matrix to generate enhanced behavior features: in, Indicates enhanced behavioral characteristics, Representation and Courses The collection of neighbor courses that appear in the same session, Indicates the course The position code, Indicates the course With Courses Frequency of occurrence in the same session, Indicates the course The initial behavior embedding vector of Construct a user behavior feature matrix based on the enhanced behavior features of each course.

4. The course conversation recommendation method according to claim 1, characterized in that: The performing graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node includes: Performing neighborhood aggregation operation on the adjacency graph through a graph convolutional network to obtain short-range dependencies between courses; Acquire local feature information of each course according to the short-range dependency relationship; Performing a graph convolution operation on the hypergraph through a graph convolution network to obtain hyper-edge information; Acquire the interaction relationship between the student and each course based on the hyperedge information, and acquire the potential feature information of the course in the learning path based on the interaction relationship; Fusing the local feature information and the potential feature information to obtain a fused feature; The fused features are activated and normalized to generate a target embedding representation vector for each course node.

5. The course conversation recommendation method according to any one of claims 1 to 4, characterized in that: The generating of the attention weights of each course node, performing weighted aggregation on the target embedding representation vector based on the attention weights, and performing course session recommendation based on the weighted aggregated target embedding representation vector, comprises: Determine the cosine similarity between course nodes based on the semantic feature vector of each course node: in, Representation Node i With Node j The cosine similarity between and Respectively represent nodes i and nodes j The semantic feature vector of Determine semantic difference information between course nodes according to the cosine similarity; The attention weight of each course node is generated based on the semantic difference information and the position encoding information of each course node: in, represents the attention weight, Represents position encoding information; Generating a cluster center of the target embedding representation vector; Calculate the cluster similarity between each course node and the cluster center; The target embedding representation vector is weightedly aggregated based on the cluster similarity and the attention weight: in, Represents the target embedding representation vector after weighted aggregation; Generating a behavior feature vector of the user based on the user's historical learning interaction data; The weighted aggregated target embedding representation vector and the behavior feature vector are input into a pre-trained recommendation model for course session recommendation.

6. The course conversation recommendation method according to claim 5, characterized in that: The step of inputting the weighted aggregated target embedding representation vector and the behavior feature vector into a pre-trained recommendation model for course session recommendation includes: The weighted aggregated target embedding representation vector and the behavior feature vector are input into the pre-trained recommendation model to calculate the predicted score of each course node: in, represents the prediction score, represents the behavior feature vector, Represents the course embedding feature matrix constructed by the target embedding representation vector after weighted aggregation; The label difference information between the predicted score and the true label is determined based on the cross entropy loss function: in, Indicates label difference information. represents the true label, represents the prediction score; The model parameters of the recommendation model are updated through the back-propagation algorithm: in, η represents the learning rate, represents the gradient of the loss function, represents the model parameters before updating, represents the updated model parameters; The recommendation model is optimized according to the updated model parameters, and course session recommendations are performed based on the optimized recommendation model.

7. A course conversation recommendation device, characterized in that: The course conversation recommendation device comprises: A semantic analysis module, used to input the text description data of the course into a pre-trained deep language model for semantic quantification, obtain a semantic feature vector, and generate a semantic embedding matrix based on the semantic feature vector; User behavior analysis module, used to perform user behavior analysis on the user's historical learning interaction data and construct a user behavior feature matrix based on the user behavior analysis results; A feature splicing module, used for performing channel splicing on the semantic embedding matrix and the user behavior feature matrix to generate a multi-dimensional feature matrix; A graph construction module, used to construct an adjacency graph and a hypergraph between courses according to the multidimensional feature matrix, wherein the adjacency graph includes the adjacency relationship between each course node, and the hypergraph includes the course node and the interaction relationship between the user and each course node; A graph convolution processing module, used to perform a graph convolution operation based on the adjacency graph and the hypergraph to generate a target embedding representation vector for each course node; The course recommendation module is used to generate the attention weight of each course node, and perform weighted aggregation on the target embedding representation vector based on the attention weight, and recommend course sessions based on the weighted aggregated target embedding representation vector.

8. A course conversation recommendation device, characterized in that: The course conversation recommendation device includes: a memory, a processor, and a course conversation recommendation program stored in the memory and executable on the processor, wherein the course conversation recommendation program is configured to implement the course conversation recommendation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a course conversation recommendation program, which, when executed by a processor, implements the course conversation recommendation method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes a course session recommendation program, which implements the steps of the course session recommendation method according to any one of claims 1 to 6 when executed by a processor.

Images