A course recommendation method based on collaborative graph convolution and learning style

Abstract

This invention discloses a course recommendation method based on graph convolutional networks and learning styles, comprising the following steps: Step 1, predicting scores: encoding course embedding information connected to learners into learners' first-order embedding information, and obtaining higher-order embedding information through the first-order embedding information; aggregating each order embedding information in the higher-order embedding information into a single vector through an aggregation function to obtain an aggregated embedding representation; performing an inner product operation on the aggregated embedding representations of learners and courses to obtain the learner's predicted score for the course; Step 2, defining learner profiles and course profiles respectively, calculating learning style vectors, and obtaining course learning style similarity scores based on learner learning style vectors and course learning style vectors; Step 3, optimizing the predicted score list through learning style similarity scores to obtain collaborative predicted scores. This method solves the problem of graph convolutional network-based recommendation algorithms ignoring learners' own learning patterns.

Description

Technical Field

[0001] This invention relates to the fields of graph neural networks and personalized recommendation technology, specifically to a course recommendation method based on graph convolutional networks and learning styles. Background Technology

[0002] Today, massive open online courses (MOOCs) are attracting widespread interest from learners as an emerging educational model. This not only reduces learning costs but also provides easy access to high-quality courses from renowned universities and institutions worldwide. Working professionals also benefit from online education; however, these learners often rely on self-study, lack guidance, and experience low learning efficiency, finding it difficult to quickly identify suitable courses. Therefore, personalized course recommendations are particularly important for improving learning efficiency, enhancing the learning experience, and improving learning outcomes.

[0003] There are many methods for recommending learning resources, mainly based on deep learning and traditional algorithmic recommendation models. Traditional course recommendation models primarily employ collaborative filtering (CF) technology, which predicts content a user might be interested in by measuring the similarity between users or courses and then recommends it. This method relies on behavioral data between users and online courses and makes recommendations by calculating similarity. However, traditional CF methods may face challenges when handling large-scale data, thus requiring more efficient models. With the development of deep learning technology, its application in course recommendation is becoming increasingly prominent. Deep learning technology can automatically learn feature representations without requiring manual feature design. While the above methods are effective, they neglect information encoded in key collaborative information when creating embedding vectors for users and courses. The interaction behavior of users and courses hides this crucial collaborative information, which can be used to reveal behavioral similarities between users / items. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a course recommendation method based on collaborative graph convolutional networks and learning styles, thereby solving the problem that graph convolutional network-based recommendation algorithms ignore learners' own learning patterns.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A course recommendation method based on graph convolutional networks and learning styles includes the following steps:

[0007] Step 1: Predict Scores

[0008] Step 1.1: Encode the course embedding information from learners into first-order embedding information of learners, and obtain higher-order embedding information through the first-order embedding information;

[0009] Step 1.2: Aggregate each order of embedding information in the higher-order embedding information into a single vector using an aggregation function to obtain the aggregated embedding representation;

[0010] Step 1.3: Perform inner product operation on the aggregated embedding representations of learners and courses to obtain the learner's predicted rating for the course;

[0011] Step 2: Learning Style Similarity Score

[0012] Step 2.1: Define the learner profile and course profile respectively. Their learning style vector expressions are as follows:

[0013] LS u = (rea, tra, soc)

[0014] LS c = (rea, tra, soc)

[0015] Where rea, tra, and soc represent realistic, traditional, and social learning styles, respectively;

[0016] Step 2.2: Based on the course-concept matrix, concept-domain matrix, and domain-style matching metric obtained from real questionnaires and guidance from educational professionals, construct the concept-style matrix. From the correlation between course, concept, and learning style, obtain the course-style matrix, where each row represents the learning style vector for the corresponding course, and this matrix is ​​normalized. Simultaneously, based on learners' course subscription information and the correlation between course and learning style, calculate the learner's learning style attribute ratio, constructing a learner learning style matrix. Each row of this matrix represents the learning style vector for different learners, and this matrix is ​​normalized to obtain the final learner learning style vector.

[0017] Step 2.3: Based on the learner's learning style vector and the course's learning style vector, use cosine similarity to calculate the learning style similarity between the learner and the course, and obtain the course learning style similarity score.

[0018] Step 3: Obtain the learner's predicted rating list for the course based on the predicted rating, and optimize the predicted rating list by learning style similarity rating to obtain the collaborative predicted rating. Select the top-K courses as the user's final course recommendation result.

[0019] Preferably, the specific method of step 1.1 is as follows:

[0020] The embedding information of courses c connected to learner u is encoded as the first-order embedding information of learner u in the following form, defined as follows:

[0021]

[0022] in, Let e ​​be the first-order connectivity information of learner u. c Let α represent the initial embedding of course c. uc This represents the attenuation coefficient for each propagation on edge (u,c);

[0023] a uc Set as in and This represents the number of first-hop neighbors of learner u and course c;

[0024] Based on the propagation form of first-order embedded information, the l-order embedded information of learner u can be represented as follows:

[0025]

[0026] The expression form of the l-th order higher-order embedding information of course c is obtained as follows:

[0027]

[0028] Preferably, the specific method of step 1.2 is as follows:

[0029] After performing embedding propagation at layer L, various higher-order information about the learner node u was obtained:

[0030]

[0031] An aggregation function is used to aggregate the embedding information of each order into a single vector, as defined below:

[0032]

[0033] Where p l It is 1 / (L+1) to balance the impact of each layer of embedding propagation on the final embedding representation.

[0034] Preferably, step 1 is implemented using a graph convolutional network model.

[0035] As a preferred method, the optimization method for the graph convolutional network model is as follows:

[0036] The BPR loss function is chosen, which calculates the overall loss of the model by assuming that learners should have higher prediction values ​​for learned lessons than for unlearned lessons. The expression is as follows:

[0037]

[0038] in, Denotes the training set, where This represents the set of courses that the learner has taken. It is the set of courses that the learner has not studied, obtained through a random negative sampling strategy; σ is the sigmoid function; λ is used to control the strength of L2 regularization; and Θ represents the overall parameters of the model, i.e., the initial embedding vector.

[0039] Preferably, the graph convolutional network model uses the mini-batch Adam optimizer to optimize and update the model parameters.

[0040] Preferably, in step 2.3, the method for calculating learning style similarity is as follows:

[0041]

[0042] in This represents the i-th value in the learner's learning style vector. This represents the i-th value in the course learning style vector.

[0043] Preferably, the specific method of step 3 is as follows:

[0044] For each learner, a list of predicted ratings for the course is obtained through a graph convolutional network model.

[0045] Sim learns style similarity scoring cls (u, c) for predicted score The optimization is performed using the following calculation formula:

[0046]

[0047] Where θ and ε are the collaborative weights of the predicted score and θ+ε=1.

[0048] This invention has the following characteristics and beneficial effects:

[0049] 1. This invention effectively provides learners with personalized course recommendations, adapting well to learner and course sets of varying sizes while exhibiting robustness. This is because the invention employs graph convolutional networks, a technique capable of automatically learning and capturing the complex relationships between learners and courses. In this way, the embedding information of learners and courses is successfully represented as vectors, effectively reflecting the correlations and differences between them. This allows the invention to more accurately understand and capture user interests and needs, improving its sensitivity to learner interests and overall interpretability.

[0050] 2. To improve the effectiveness of recommended courses and take into account learners' personalized learning styles to better meet user needs, this invention introduces the concept of collaborative weighting. This method combines predicted course scores with learners' learning style similarity scores to provide personalized course resource recommendations. Suggestions based on learners' personalized learning styles help improve the learning experience, meet individual needs and preferences, and ultimately promote improved learning outcomes. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A model diagram of a course recommendation method based on collaborative graph convolution and learning style provided in an embodiment of the present invention;

[0053] Figure 2 This is a diagram illustrating the learning style vector calculation process in an embodiment of the present invention;

[0054] Figure 3 This is a diagram of higher-order interactive information; where node u1 is the learner to be recommended.

[0055] Figure 4 This is a schematic diagram of a graph convolutional model architecture provided in an embodiment of the present invention. Detailed Implementation

[0056] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0057] With the rapid development of information technology, people can easily access and store vast amounts of data and information, but this also raises challenges regarding learner privacy. Among this data, process behavior data specifically refers to data related to human-computer interaction, such as learners' behavior records when selecting courses or viewing learning resources. This process data can help protect learners' privacy and security. Therefore, combining learner modeling with process behavior data can successfully address students' personalized needs by recommending courses or learning materials tailored to their requirements.

[0058] The following detailed explanation of the course recommendation method based on collaborative graph convolutional networks and learning styles, using implementation examples, mainly comprises three parts: score prediction via a graph convolutional network model, course learning style similarity score, and collaborative prediction score.

[0059] I. Predicting scores using a graph convolutional network model:

[0060] In this embodiment, the graph convolutional network model consists of multiple substructures, including a higher-order embedding propagation layer, an embedding aggregation layer, and a prediction scoring layer. The higher-order embedding propagation layer learns feature representations of nodes by propagating the neighbor information of nodes through multi-level embeddings to capture more complex graph structure information. The embedding aggregation layer is responsible for aggregating the features learned from the higher-order embedding propagation layer to generate a global graph-level representation. The prediction scoring layer, as the model's output layer, is responsible for mapping the learned graph-level representation to the final prediction score.

[0061] Step 1: Construct a higher-order embedding propagation layer to encode the course embedding information from learners into the learners' first-order embedding information.

[0062] The specific process is as follows:

[0063] Step 1.1: Based on the concept of GCN, the embedding information of courses c connected to learner u is encoded into first-order embedding information of learner u in the following form to enhance learner u's own embedding information, defined as follows:

[0064]

[0065] in, Let e ​​be the first-order connectivity information of learner u. c Let α represent the initial embedding of course c. uc This represents the attenuation coefficient for each propagation on edge (u,c).

[0066] Step 1.2: In this invention, the same approach as GCN is adopted, and α... uc Set as in and This represents the number of first-hop neighbors of learner u and course c. α uc It not only fulfills the function of balancing the contribution of the history course c to the preferences of learner u, but also allows different embedded information to decay as the path length changes during the embedding propagation process.

[0067] Step 1.3: Based on the propagation form of first-order embedded information, the l-order embedded information of learner u is presented in the following form:

[0068]

[0069] Step 1.4: We can obtain the expression form of the l-th order higher-order embedding information of course c:

[0070]

[0071] Step 2: Embedding aggregation layer, using aggregation functions to aggregate the embedding information of each order into a single vector.

[0072] The specific process is as follows:

[0073] Step 2.1: After performing embedding propagation at layer L, various higher-order information of learner node u is obtained. The same applies to course milestones.

[0074] Step 2.2: The outputs of different layers emphasize connectivity information at different orders. Therefore, an aggregation function is used to aggregate the embedding information of each order into a single vector, defined as follows:

[0075]

[0076] Where P l The value is 1 / (L+1), used to balance the impact of each layer of embedding propagation on the final embedding representation. By aggregating the embedding information of each level in this way, the node embedding representation obtained contains different semantic information in the graph structure, and the final embedding representation will be more comprehensive.

[0077] Step 3: Predictive rating layer, which performs inner product operation on the aggregated embedding representations of learners and courses to predict learners' predicted ratings for the course.

[0078] The specific process is as follows:

[0079] Perform an inner product operation on the aggregated embedding representations of learner u and course c to predict learner u's predicted rating for course c:

[0080]

[0081] This will be integrated into the subsequent collaborative prediction score, used to generate the recommendation list.

[0082] Step 4: Construct the loss function and trainer. Choose the BPR loss function, which calculates the overall loss of the model by assuming that learners should have higher prediction values ​​for learned courses than for unlearned courses.

[0083] The specific process is as follows:

[0084] Step 4.1: The only parameters that need to be trained in the entire graph collaborative filtering model are the embedding lookup table mentioned earlier, i.e., the embedding representation at layer 0. To optimize the graph collaborative filtering model, the BPR loss function is chosen, which calculates the overall loss of the model by assuming that learners should have higher prediction values ​​for learned courses than for unlearned courses.

[0085]

[0086] in, Denotes the training set, where This represents the set of courses that the learner has taken. It is the set of courses that the learner has not studied, obtained through a random negative sampling strategy; σ is the sigmoid function; λ is used to control the strength of L2 regularization; and Θ represents the overall parameters of the model, i.e., the initial embedding vector.

[0087] Step 4.2: Use the mini-batch Adam optimizer to optimize and update the model parameters.

[0088] II. Course Learning Style Similarity Score

[0089] Step 1: Define the learning style representation.

[0090] Furthermore, the specific process of step 1 is as follows:

[0091] Step 1.1: Define the learner profile. Assume the learner's learning style is represented by the real-valued vector LS from 0 to 1 in Formula 1. u The expression indicates that rea, tra, and soc represent realistic, conventional, and social learning styles, respectively, and the formula is as follows:

[0092] LS u =(rea,tra,soc)

[0093] Step 1.2: Define the course summary. Online courses cover many theoretical concepts, which belong to different professional fields, and different professional fields have their own suitable learning styles. Therefore, based on the correlation between course, concept, and learning style, the course summary attribute can be represented by a learning style vector, as shown in Formula 2. This vector represents the learning style category suitable for the course.

[0094] LS c =(rea,tra,soc)

[0095] Step 2: Calculate the learning style vector. Based on the course-concept matrix, concept-domain matrix, and domain-style matching metric obtained from real questionnaires and guidance from educational professionals, construct the concept-style matrix. From the correlation between course, concept, and learning style, obtain the course-style matrix and normalize it. Each row of the matrix represents the learning style vector of the corresponding course, as shown in Table 1. Simultaneously, based on learners' course subscription information and the correlation between course and learning style, calculate the learner's learning style attribute ratio to construct a learner learning style matrix. Each row of this matrix represents the learning style vector of different learners and is normalized to obtain the final learner learning style vector, as shown in Table 2.

[0096] Table 1 Examples of Course Learning Style Vectors

[0097]

[0098] Table 2 Examples of Learner Learning Style Vectors

[0099]

[0100] The specific process is as follows:

[0101] Table 3 presents some results of the educational professional guidance and the actual questionnaire-based on-site learning style matching measures. Based on personality type, they are categorized as Realistic, Traditionalist, and Socialist. The Realistic, Traditionalist, and Socialist types mentioned here correspond to learners' learning styles "rea," "tra," and "soc" in Step 1. Based on statistical data from learners' responses in the questionnaire, according to... Figure 2 The method described in the text calculates the proportion of corresponding learning styles, and finally determines their learning styles.

[0102] Table 3 Examples of Professional Domain-Learning Style Type Matching Metrics

[0103]

[0104] Step 3: Calculate the course learning style similarity score.

[0105] The specific process is as follows:

[0106] Based on learner learning style vectors and course learning style vectors, cosine similarity is used to calculate the learning style similarity between learners and courses, resulting in a course learning style similarity score, Sim. cls (u, c), as shown in Formula 3:

[0107]

[0108] in This represents the i-th value in the learner's learning style vector. This represents the i-th value in the course learning style vector. Learning style similarity will be incorporated into the subsequent collaborative prediction score.

[0109] III. Collaborative Prediction Scoring

[0110] In this embodiment, the predicted score Predicted scores for courses using graph convolutional networks Similarity score with learning style Sim cls The structure, consisting of (u, c), mainly includes the following steps:

[0111] Step 1: For each learner, obtain a list of predicted ratings for the course using a graph convolutional network recommendation algorithm.

[0112] Step 2: Calculate the similarity of learning styles.

[0113] The specific process is as follows:

[0114] Calculate the learning style similarity between learner u and course c, and evaluate the predicted scores obtained by the graph convolutional network model. The optimization is performed using the following calculation formula:

[0115]

[0116] Where θ and ε are the collaborative weights of the predicted ratings, and θ + ε = 1. Based on the collaborative predicted ratings, the top-K courses are selected as the final course recommendation result for user u, R(u){c1, c2, ..., c...}. Top-K}

[0117] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A course recommendation method based on graph convolutional networks and learning styles, characterized in that, Includes the following steps: Step 1: Implement score prediction using a graph convolutional network model. Step 1.1: Encode the course embedding information from learners into first-order embedding information of learners, and obtain higher-order embedding information through the first-order embedding information; Step 1.2: Aggregate each order of embedding information in the higher-order embedding information into a single vector using an aggregation function to obtain the aggregated embedding representation; Step 1.3: Perform inner product operation on the aggregated embedding representations of learners and courses to obtain the learner's predicted rating for the course; The optimization method for the graph convolutional network model is as follows: The BPR loss function is chosen, which calculates the overall loss of the model by assuming that learners should have higher prediction values ​​for learned lessons than for unlearned lessons. The expression is as follows: in, Denotes the training set, where This represents the set of courses that the learner has taken. It is the set of courses that the learner has not studied, obtained through a random negative sampling strategy; σ is the sigmoid function; λ is used to control the strength of L2 regularization; Θ represents the parameters of the overall model, i.e., the initial embedding vector. Step 2: Learning Style Similarity Score Step 2.1: Define the learner profile and course profile respectively, and their learning style vectors, as shown in the following expression: LS u =(rea,tra,soc) LS c =(rea,tra,soc) Among them, LS u Represents the learner's learning style vector; LS c This represents a vector of learning styles; rea, tra, and soc represent realistic, conventional, and social learning styles, respectively. Step 2.2: Based on the course-concept matrix, concept-domain matrix, and domain-style matching metric obtained from real questionnaires and guidance from educational professionals, construct the concept-style matrix. Based on the correlation between course, concept, and learning style, obtain the course-style matrix. Each row of the matrix represents the learning style vector of the corresponding course and is normalized. At the same time, based on learner course subscription information and the correlation between course and learning style, calculate the learner learning style attribute ratio and construct the learner learning style matrix. Each row of the matrix represents the learning style vector of different learners and is normalized to obtain the final learner learning style vector. Step 2.3: Based on the learner's learning style vector and the course's learning style vector, use cosine similarity to calculate the learning style similarity between the learner and the course, and obtain the course learning style similarity score. Step 3: Obtain the learner's predicted rating list for the course based on the predicted rating, and optimize the predicted rating list by learning style similarity rating to obtain the collaborative predicted rating. Select the top-K courses as the user's final course recommendation result.

2. The course recommendation method based on graph convolutional networks and learning styles according to claim 1, characterized in that, The specific method for step 1.1 is as follows: The embedding information of courses c connected to learner u is encoded as the first-order embedding information of learner u in the following form, defined as follows: in, Let e ​​be the first-order connectivity information of learner u. c Let α represent the initial embedding of course c. uc This represents the attenuation coefficient for each propagation on edge (u,c); α uc Set as in and This represents the number of first-hop neighbors of learner u and course c; Based on the propagation form of first-order embedded information, the l-order embedded information of learner u can be represented as follows: The l-order embedding information of course c is expressed in the following form:

3. The course recommendation method based on graph convolutional networks and learning styles according to claim 2, characterized in that, The specific method for step 1.2 is as follows: After performing embedding propagation at layer L, various higher-order information about the learner node u was obtained: An aggregation function is used to aggregate the embedding information of each order into a single vector, as defined below: Where p l It is 1 / (L+1) to balance the impact of each layer of embedding propagation on the final embedding representation.

4. The course recommendation method based on graph convolutional networks and learning styles according to claim 3, characterized in that, The graph convolutional network model uses the mini-batch Adam optimizer to optimize and update the model parameters.

5. The course recommendation method based on graph convolutional networks and learning styles according to claim 1, characterized in that, In step 2.3, the method for calculating learning style similarity is as follows: in This represents the i-th value in the learner's learning style vector. This represents the i-th value in the course learning style vector.

6. The course recommendation method based on graph convolutional networks and learning styles according to claim 5, characterized in that, The specific method for step 3 is as follows: For each learner, a list of predicted ratings for the course is obtained through a graph convolutional network model. Sim learns style similarity scoring cls (u, c) for predicted score The optimization is performed using the following calculation formula: Where θ and ε are the collaborative weights of the predicted score and θ+ε=1.

Images