Method, system and apparatus for online learning resource recommendation

Abstract

The application discloses a kind of online learning resource recommendation method, system and device, the method includes, S1, the dependent relationship between learning resources is acquired;S2, the dynamic preference of user to learning resource embodied in user behavior sequence is acquired, and the learned dependent relationship is fused into preference representation;S3, score is predicted and recommendation list is generated.The application can realize online learning resource recommendation.

Description

Technical Field

[0001] This invention relates to the field of online learning resource recommendation, and in particular to a method, system, and apparatus for recommending online learning resources. Background Technology

[0002] With the continuous development of online education, many Massive Open Online Course (MOOC) platforms, represented by Coursera and XuetangX, have rapidly emerged, attracting a large number of users by providing high-quality resources at low cost. At the same time, learners inevitably face the dilemma of information overload and knowledge disorientation due to the explosive growth of data volume. To accurately understand individual student differences and implement precise teaching based on individual needs, recommendation systems have been widely used as an effective technology for filtering information.

[0003] Learning resource recommendation is an effective technical means to solve the above problems. It aims to rely on massive educational data and use information technologies such as cloud platforms, big data, and artificial intelligence to conduct personalized analysis of learners. Based on the learner's personality characteristics and the context of the learning environment, it provides adaptive learning resource recommendation services to help learners efficiently filter and match resources, thereby guiding their online learning path.

[0004] Current research focuses on applying deep learning techniques to traditional recommendation models to capture the complex nonlinear relationships between users and items, but two shortcomings remain. First, learning resources and items naturally have complex dependencies, such as prerequisite dependencies, but existing work usually treats them as independent nodes, ignoring the static dependency characteristics between such items. Second, learners' learning abilities and goals may change dynamically over time, requiring more accurate characterization and modeling of user preferences, while existing content-based and collaborative filtering algorithms lack the ability to handle such dynamic changes. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and apparatus for recommending online learning resources, aiming to solve the problem of online learning resource recommendation.

[0006] This invention provides an online learning resource recommendation method, comprising:

[0007] S1. Obtain the dependencies between learning resources;

[0008] S2. Obtain the user's dynamic preferences for learning resources reflected in the user behavior sequence, and integrate the learned dependencies into the preference representation;

[0009] S3. Predict ratings and generate a recommendation list.

[0010] This invention also provides an online learning resource recommendation system, comprising:

[0011] Dependency learning module: Retrieves the dependencies between learning resources;

[0012] User interest mining module: Obtains the user's dynamic preferences for learning resources reflected in the user behavior sequence, and integrates the learned dependencies into the preference representation;

[0013] Prediction module: Predicts ratings and generates a recommendation list.

[0014] This invention also provides an online learning resource recommendation device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above method.

[0015] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method.

[0016] By employing embodiments of the present invention, the inherent dependencies between learning resources and the dynamic preferences implied in the learner's historical behavioral sequence are fully considered, providing a more personalized learning resource recommendation service that better meets the learner's needs and is more accurate.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an online learning resource recommendation method according to an embodiment of the present invention;

[0020] Figure 2 This is a training flowchart of the online learning resource recommendation method according to an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the model framework of the online learning resource recommendation method according to an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of resource and concept mapping for the online learning resource recommendation method according to an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram of the session graph construction of the online learning resource recommendation method according to an embodiment of the present invention;

[0024] Figure 6 This is a schematic diagram of the global graph construction of the online learning resource recommendation method according to an embodiment of the present invention;

[0025] Figure 7 This is a schematic diagram of an online learning resource recommendation system according to an embodiment of the present invention;

[0026] Figure 8 This is a schematic diagram of an online learning resource recommendation device according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Method Implementation Examples

[0029] According to embodiments of the present invention, a method for recommending online learning resources is provided. Figure 1 This is a flowchart of an online learning resource recommendation method according to an embodiment of the present invention, such as... Figure 1 As shown, it specifically includes:

[0030] S1. Obtain the dependencies between learning resources;

[0031] S2. Obtain the user's dynamic preferences for learning resources reflected in the user behavior sequence, and integrate the learned dependencies into the preference representation;

[0032] S3. Predict ratings and generate a recommendation list.

[0033] S1 specifically includes:

[0034] S11. Obtain the learning resource entity data of the MOOC platform, obtain the course prerequisite relationship based on the learning resource entity data, and use the course prerequisite relationship as a dependency relationship label.

[0035] S12. Using a word-based representation method, the subtitles and resource description texts in the learning resources on the MOOC platform are preprocessed using a standard language processing flow to obtain a preprocessed word set. The preprocessing includes word segmentation, stop word removal, and word frequency statistics. The preprocessed word set is used as the overall concept space.

[0036] S13. Based on the mapping relationship between learning resources and word sets, construct a resource representation scheme based on concept packages to obtain a unified vector representation of learning resources in the concept space;

[0037] S14. Recover the concept graph based on the visible resource-level dependency labels. Define concept graph learning as a concept-level link prediction problem. Use the dependency labels obtained in S11 as supervision information. Adopt a classification-based optimization method to optimize the model parameter matrix B according to the objective function and learn the directed link weights between all concept nodes in the concept graph.

[0038] S15. Based on parameter matrix B, predict the invisible dependencies between all resources in the test set and output dependency matrix D.

[0039] Obtaining the dynamic preferences for learning resources among learning users specifically includes:

[0040] S21. Construct a session graph and a global graph for a single user sequence and the entire sequence set, respectively. Use a graph neural network-based representation layer to learn the session-level and global-level embeddings of items, and connect the session-level and global-level to construct the final embedding representation of items. Introduce a sparse attention mechanism to calculate the importance weight of each item in the session, and aggregate the item representations according to the weight coefficients to obtain the user preference embedding representation of the current session.

[0041] The process of integrating the learned dependencies into the preference representation specifically includes:

[0042] S22. Construct a resource dependency graph based on the output dependency matrix D;

[0043] S23. Use the same graph neural network structure as the global representation learning layer to learn the static dependency representation of items;

[0044] S24. The learned static dependency representation of the project is fused with the final embedding representation of the project through linear transformation;

[0045] S25. Construct one-hot encoding of project dependencies based on dependency matrix D, and use an embedding layer to map the dependency encoding to the embedding space to obtain the project's dependency embedding.

[0046] S26. When calculating the importance weight of each item in a session, the dynamic sequence features and dependency embeddings of the items are fused together.

[0047] S3 specifically includes:

[0048] S31. Based on the fused user preference embedding representation, calculate its dot product similarity with the candidate item representation to obtain the score corresponding to the candidate item set;

[0049] S32. Sort the predicted scores in descending order;

[0050] S33. Select the K highest-rated items to form the final resource recommendation list.

[0051] Figure 2 This is a training flowchart of the online learning resource recommendation method according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the model framework of the online learning resource recommendation method according to an embodiment of the present invention;

[0053] Step 1: Learn the complex dependencies inherent in the resources;

[0054] Collect resource entity data from the MOOC platform, obtain available course prerequisite relationship relationships, and use them as dependency relationship tags;

[0055] We adopt a word-based representation method and use standard natural language processing workflow to preprocess text such as subtitles and resource descriptions in video resources on the MOOC platform, including steps such as word segmentation, stop word removal, and word frequency statistics. The preprocessed word set is used as the overall concept space.

[0056] Figure 4 This is a schematic diagram of resource and concept mapping for the online learning resource recommendation method according to an embodiment of the present invention;

[0057] Based on the mapping relationship between resources and concepts, a resource representation scheme based on concept packages is constructed to obtain a unified vector representation of resources in the concept space;

[0058] Concept graph learning is defined as a concept-level link prediction problem. A classification-based optimization method is employed to optimize the model parameter matrix B according to the objective function, learning the directed link weights between all concept nodes in the concept graph.

[0059] Based on parameter matrix B, predict the invisible dependencies between all resources in the test set and output dependency matrix D;

[0060] Step two: Learn dynamic preference representations in user sequences;

[0061] Figure 5 This is a schematic diagram of the session graph construction of the online learning resource recommendation method according to an embodiment of the present invention;

[0062] Figure 6 This is a schematic diagram of the global graph construction of the online learning resource recommendation method according to an embodiment of the present invention;

[0063] For a single user sequence and the entire sequence set, a session graph and a global graph are constructed respectively. A graph neural network-based representation layer is used to learn the session-level and global-level embeddings of items, and the two are connected to construct the final embedding representation of items. A sparse attention mechanism is introduced to calculate the importance weight of each item in the session, and the item representations are aggregated according to the weight coefficients to obtain the user preference embedding representation of the current session.

[0064] Local modules:

[0065] A session graph is constructed based on the user's historical learning behavior sequence on MOOC. Each user behavior sequence corresponds to a directed session graph, and each node in the graph corresponds to an interaction item in the session. Two adjacent interaction items in the sequence are connected by a directed edge.

[0066] Calculate the weight of each edge in the graph;

[0067] Construct an m-order graph adjacency matrix based on the in-degree and out-degree information of each node in the graph;

[0068] Each item is embedded into a unified latent space to obtain an initial representation;

[0069] For each session graph, the graph node representation is learned through a gated graph neural network;

[0070] Learn the m-order local embedding representation of nodes based on the m-order adjacency matrix;

[0071] The m-order local embedding representation of dynamically aggregated nodes yields a session-level node vector representation.

[0072] Global module:

[0073] For all sequences in the dataset, we construct the set of its epsilon-order neighbors for each node;

[0074] We construct a global graph based on the resource neighbor set, where each resource corresponds to a node in the graph, and each resource is connected to each of its neighbors by an undirected edge.

[0075] The weights of edges in the global graph are calculated based on the number of edges and the distance between items, and then the weights are normalized.

[0076] For each item, sort its neighboring nodes in descending order of weight, and retain the top N neighboring nodes by weight to construct a global graph.

[0077] By leveraging the conversation-aware attention mechanism, the importance of items in each node's neighbor set is distinguished, importance coefficients are calculated and normalized to determine which neighbor nodes should be assigned more attention.

[0078] The neighbor context representation of each node is obtained by linearly combining the representations of neighbor nodes based on their importance coefficients.

[0079] The node vector representation at the global level is obtained by aggregating the node's own representation and the context representation of its neighbors.

[0080] Attention-based session representation:

[0081] The session-level and global-level data learned by the local and global modules are combined to obtain the embedding representation of each item;

[0082] Embedding the project into the feedforward neural network represents the input to obtain more non-linear features.

[0083] The weighting coefficient for each item is calculated using an attention mechanism;

[0084] Alpha-entmax is used to replace the traditional transformation function for weight coefficient normalization, thereby reducing the weight of irrelevant terms.

[0085] The final session representation is obtained by linearly combining all item representations in the session based on this weight.

[0086] Step 3: Integrate resource dependencies;

[0087] Design a multi-faceted integration mechanism for dependencies.

[0088] Based on the dependency matrix D output in step one, construct a resource dependency graph in a manner similar to that of a global graph;

[0089] The same graph neural network structure as the global representation learning layer is used to learn the static dependency representation of items;

[0090] Then, the learned item dependency feature representation is fused with the item sequence feature embedding representation from step two through linear transformation;

[0091] The original dependency input of each node is constructed based on the dependency matrix, and the dependency features of the input are extracted using an embedding layer on the basis of one-hot encoding, and the original feature data is mapped to the embedding space.

[0092] The dependency embedding of the project is integrated into the attention coefficient calculation, and the dynamic sequence characteristics and static dependency characteristics of the project are considered at the same time when calculating the importance of the project in the session.

[0093] Step 4: Predict ratings and generate a recommendation list;

[0094] Based on the user preference embedding representation learned in step two, calculate its dot product similarity with the candidate item representation to obtain the score corresponding to the candidate item set;

[0095] Sort the predicted scores in descending order;

[0096] The K highest-rated items are selected to form the final resource recommendation list.

[0097] To address the recommendation problem of learning resources of different granularities for Massive Open Online Classroom (MOOC) platforms, this invention provides a solution that integrates dependency learning and sequence recommendation models. This solution fully considers the inherent dependencies between learning resources and the dynamic preferences implied in learners' historical behavioral sequences, providing more personalized learning resource recommendation services that better meet learners' needs and are more accurate.

[0098] The main components include a resource dependency learning module, which uses a sequence model based on graph neural networks and a multi-angle fusion mechanism. First, learning resources are mapped to a general concept space to obtain a unified representation. Then, visible dependencies between some resources are used to supervise the learning of the concept graph, thereby predicting resource-level dependencies. Next, for user behavior sequences, a graph neural network is used to mine the transformation relationships between items to obtain resource embedding representations. Item embeddings are then aggregated based on a sparse attention mechanism to represent user preferences in the current session. Finally, a multi-angle fusion mechanism is used to fully integrate the static dependency features and dynamic sequence features of items to predict recommendation scores, ultimately realizing a learning resource recommendation framework.

[0099] A multi-scale dependency reasoning framework consisting of a resource graph and a concept graph is designed to address the shortcomings of current work in modeling resource-level dependencies and provide an effective solution for automatically predicting and learning static complex dependencies between resources.

[0100] A sequence recommendation model based on graph neural networks was designed. Local modules are used to capture the features of items within a session, and a global module is further introduced to address the shortcomings of existing works in learning global information between sessions. Finally, a session representation method based on sparse attention mechanism is designed to reduce the negative impact of noise items, improve the model's ability to model user interests and preferences in the sequence, and thus improve the accuracy and relevance of recommendations.

[0101] A multi-angle fusion mechanism was designed, which considers both the static dependency characteristics between resource nodes and the dynamic preference characteristics in user behavior sequences. It makes full use of platform data to realize a recommendation framework applicable to various types of learning resources such as courses, videos, and exercises.

[0102] System Implementation Examples

[0103] According to embodiments of the present invention, a system for recommending online learning resources is provided. Figure 7 This is a schematic diagram of an online learning resource recommendation system according to an embodiment of the present invention, such as... Figure 7 As shown, it specifically includes:

[0104] Dependency learning module: Retrieves the dependencies between learning resources;

[0105] User interest mining module: Obtains the user's dynamic preferences for learning resources reflected in the user behavior sequence, and integrates the learned dependencies into the preference representation;

[0106] Prediction module: Predicts ratings and generates a recommendation list.

[0107] The dependency learning module is specifically used for:

[0108] Obtain the learning resource entity data from the MOOC platform, extract the course prerequisite relationship based on the learning resource entity data, and use the course prerequisite relationship as a dependency relationship tag.

[0109] A word-based representation method is adopted, and the subtitles and resource description texts in the learning resources on the MOOC platform are preprocessed using standard language processing procedures to obtain a preprocessed word set. The preprocessing includes word segmentation, stop word removal and word frequency statistics, and the preprocessed word set is used as the overall concept space.

[0110] Based on the mapping relationship between learning resources and word sets, a resource representation scheme based on concept packages is constructed to obtain a unified vector representation of learning resources in the concept space;

[0111] Based on the visible resource-level dependency labels, the concept graph is recovered. Concept graph learning is defined as a concept-level link prediction problem. The obtained dependency labels are used as supervision information. A classification-based optimization method is adopted to optimize the model parameter matrix B according to the objective function and learn the directed link weights between all concept nodes in the concept graph.

[0112] Based on parameter matrix B, predict the invisible dependencies between all resources in the test set and output dependency matrix D.

[0113] The user interest mining module is specifically used for:

[0114] For a single user sequence and the entire sequence set, a session graph and a global graph are constructed respectively. A graph neural network-based representation layer is used to learn the session-level and global-level embeddings of items. The session-level and global-level embeddings are then connected to construct the final embedding representation of items. A sparse attention mechanism is introduced to calculate the importance weights of each item in the session. Based on the weight coefficients, the item representations are aggregated to obtain the user preference embedding representation of the current session.

[0115] Construct a resource dependency graph based on the output dependency matrix D;

[0116] The same graph neural network structure as the global representation learning layer is used to learn the static dependency representation of items;

[0117] The learned static dependency representation of the project is fused with the final embedded representation of the project through linear transformation;

[0118] Construct one-hot encodings of project dependencies based on the dependency matrix D, and use an embedding layer to map the dependency encodings to the embedding space to obtain the project's dependency embeddings.

[0119] When calculating the importance weights of each item in a session, the dynamic sequence features and dependency embeddings of the items are fused together.

[0120] The prediction module is specifically used for:

[0121] Based on the fused user preference embedding representation and candidates, the dot product similarity between the fused representation and the candidate item representation is calculated to obtain the score corresponding to the candidate item set.

[0122] Sort the predicted scores in descending order;

[0123] The K highest-rated items are selected to form the final resource recommendation list.

[0124] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0125] Device Example 1

[0126] This invention provides an apparatus for recommending online learning resources, such as... Figure 8 As shown, it includes: a memory 80, a processor 82, and a computer program stored on the memory 80 and executable on the processor 82. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0127] Device Example 2

[0128] This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 82, implements the steps described in the above method embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions to the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of the present solution.

Claims

1. A method for recommending online learning resources, characterized in that, include, S1. Obtain the dependencies between learning resources; S11. Obtain the learning resource entity data of the MOOC platform, obtain the course prerequisite relationship based on the learning resource entity data, and use the course prerequisite relationship as a dependency relationship label. S12. Using a word-based representation method, the subtitles and resource description texts in the learning resources on the MOOC platform are preprocessed using a standard language processing flow to obtain a preprocessed word set. The preprocessing includes word segmentation, stop word removal, and word frequency statistics. The preprocessed word set is used as the overall concept space. S13. Based on the mapping relationship between learning resources and word sets, construct a resource representation scheme based on concept packages to obtain a unified vector representation of learning resources in the concept space; S14. Recover the concept graph based on the visible resource-level dependency labels. Define concept graph learning as a concept-level link prediction problem. Use the dependency labels obtained in S11 as supervision information. Adopt a classification-based optimization method to optimize the model parameter matrix B according to the objective function and learn the directed link weights between all concept nodes in the concept graph. S15. Based on parameter matrix B, predict the invisible dependencies between all resources in the test set and output dependency matrix D. S2. Obtain the user's dynamic preferences for learning resources reflected in the user behavior sequence, and integrate the learned dependencies into the preference representation; S21. Construct a session graph and a global graph for a single user sequence and the entire sequence set, respectively. Use a graph neural network-based representation layer to learn session-level and global-level item embeddings, and connect the session-level and global-level embeddings to construct the final item embedding representation. Introduce a sparse attention mechanism to calculate the importance weights of each item in the session, and aggregate the item representations based on the weight coefficients to obtain the user preference embedding representation for the current session. The specific steps of fusing the learned dependencies into the preference representation include: S22. Construct a resource dependency graph based on the output dependency matrix D; S23. Use the same graph neural network structure as the global representation learning layer to learn the static dependency representation of items; S24. The learned static dependency representation of the project is fused with the final embedding representation of the project through linear transformation; S25. Construct one-hot encoding of project dependencies based on dependency matrix D, and use an embedding layer to map the dependency encoding to the embedding space to obtain the project's dependency embedding. S26. When calculating the importance weight of each item in a session, the dynamic sequence features and dependency embeddings of the items are fused together. S3. Predict ratings and generate a recommendation list; S31. Based on the fused user preference embedding representation, calculate its dot product similarity with the candidate item representation to obtain the score corresponding to the candidate item set; S32. Sort the predicted scores in descending order; S33. Select the K highest-rated items to form the final resource recommendation list.

2. A system for recommending online learning resources, characterized in that, include, Dependency learning module: Retrieves the dependencies between learning resources; User interest mining module: This module acquires the user's dynamic preferences for learning resources reflected in the user's behavior sequence and integrates the learned dependencies into the preference representation. Specifically, the user interest mining module is used for: For a single user sequence and the entire sequence set, a session graph and a global graph are constructed respectively. A graph neural network-based representation layer is used to learn the session-level and global-level embeddings of items. The session-level and global-level embeddings are then connected to construct the final embedding representation of items. A sparse attention mechanism is introduced to calculate the importance weights of each item in the session. Based on the weight coefficients, the item representations are aggregated to obtain the user preference embedding representation of the current session. Construct a resource dependency graph based on the output dependency matrix D; The same graph neural network structure as the global representation learning layer is used to learn the static dependency representation of items; The learned static dependency representation of the project is fused with the final embedded representation of the project through linear transformation; Construct one-hot encodings of project dependencies based on the dependency matrix D, and use an embedding layer to map the dependency encodings to the embedding space to obtain the project's dependency embeddings. When calculating the importance weights of each item in a session, the dynamic sequence features and dependency embeddings of the items are fused together; Prediction module: predicts ratings and generates a recommendation list; specifically, the prediction module is used for: Based on the fused user preference embedding representation and candidates, the dot product similarity between the fused representation and the candidate item representation is calculated to obtain the score corresponding to the candidate item set. Sort the predicted scores in descending order; The K highest-rated items are selected to form the final resource recommendation list.

3. The online learning resource recommendation system according to claim 2, characterized in that, The dependency learning module is specifically used for: Obtain the learning resource entity data from the MOOC platform, extract the course prerequisite relationship based on the learning resource entity data, and use the course prerequisite relationship as a dependency relationship tag. A word-based representation method is adopted, and the subtitles and resource description texts in the learning resources on the MOOC platform are preprocessed using standard language processing procedures to obtain a preprocessed word set. The preprocessing includes word segmentation, stop word removal and word frequency statistics, and the preprocessed word set is used as the overall concept space. Based on the mapping relationship between learning resources and word sets, a resource representation scheme based on concept packages is constructed to obtain a unified vector representation of learning resources in the concept space; Based on the visible resource-level dependency labels, the concept graph is recovered. Concept graph learning is defined as a concept-level link prediction problem. The obtained dependency labels are used as supervision information. A classification-based optimization method is adopted to optimize the model parameter matrix B according to the objective function and learn the directed link weights between all concept nodes in the concept graph. Based on parameter matrix B, predict the invisible dependencies between all resources in the test set and output dependency matrix D.

4. A device for recommending online learning resources, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for recommending online learning resources as described in claim 1.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the online learning resource recommendation method as described in claim 1.

Images