A knowledge concept recommendation method based on user learning process

Abstract

The application discloses a kind of knowledge concept recommendation methods based on user learning process, including the following steps: define node type and relationship type in heterogeneous information network;Define meta-path, obtain based on meta-path isomorphism network and adjacent matrix;The name of concept is used to carry out feature embedding to concept, and the historical behavior in user learning process is used to carry out feature embedding to user;By graph neural network and self-attention mechanism, the concept embedding and user embedding obtained in step 3) are carried out feature learning, and concept embedding matrix and user embedding are obtained;Concept embedding matrix and user embedding obtained in step 4) are sent to matrix decomposition model, and the predicted user's score matrix to concept is obtained, and the concept with high predicted score is recommended to user.The application introduces a time attention mechanism in user feature embedding process, which can effectively distinguish the user behavior at different time points, thereby effectively solving the problem that user feature embedding capability is limited.

Description

Technical Field

[0001] This invention relates to the field of learning resource recommendation, and in particular to a knowledge concept recommendation method based on the user's learning process. Background Technology

[0002] In recent years, Massive Open Online Courses (MOOCs) platforms have developed rapidly worldwide, greatly promoting the popularization of education and benefiting millions of users. However, despite the abundance of learning resources these platforms provide, information overload remains a prominent issue, often leaving users struggling to sift through the vast amount of resources to find the courses and learning materials that best suit their needs. To address this problem, knowledge concept recommendation has gradually become a research hotspot. Knowledge concepts, as terms extracted from the subtitle text of online courses, are closely related to the course content. By recommending knowledge concepts, users can be helped to more accurately locate learning content that interests them, thereby improving their learning experience.

[0003] Existing knowledge concept recommendation methods often model various entities and their relationships on MOOC platforms using Heterogeneous Information Networks (HINs), focusing on mining effective information between users and knowledge concepts to make recommendations. However, these methods focus on the static interaction between users and knowledge concepts, neglecting the temporal dynamics of user learning preferences—that is, user learning preferences change over time, with the deepening of learning content and the improvement of knowledge level. It is precisely because of the lack of effective consideration of this temporal evolution that the feature embedding ability of existing methods is limited, making it difficult to accurately capture changes in user learning preferences, thus leading to reduced recommendation accuracy. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a knowledge concept recommendation method based on the user's learning process that is highly efficient and accurate.

[0005] The technical solution of this invention to solve the above problems is: a knowledge concept recommendation method based on the user's learning process, comprising the following steps:

[0006] 1) Define the node types and relationship types in heterogeneous information networks;

[0007] 2) Define meta-paths and obtain isomorphic networks and adjacency matrices based on meta-paths;

[0008] 3) Embed features of concepts using their names and embed features of users using their historical behavior during the learning process;

[0009] 4) Using graph neural networks and self-attention mechanisms, feature learning is performed on the concept embeddings and user embeddings obtained in step 3) to obtain the concept embedding matrix and user embeddings;

[0010] 5) Input the concept embedding matrix and user embedding obtained in step 4) into the matrix factorization model to obtain the predicted user rating matrix for the concept, and recommend the concepts with high predicted ratings to the user.

[0011] The knowledge concept recommendation method based on the user's learning process described above, specifically includes the following steps in step 1):

[0012] Step 1-1) For the MOOCCube large-scale open online course dataset, select historical records over a period of time as the model's dataset, and define five node types and five relationship types in the heterogeneous information network. The five node types include: user (an individual learning on the platform); teacher (a person teaching courses on the platform); video (an educational resource covering specific knowledge concepts); course (composed of a series of videos); and knowledge concept (the basic unit that makes up the course content). The five relationship types include: user → video (indicating a user watched a specific video); video → knowledge concept (indicating a video contains a specific knowledge concept); user → knowledge concept (indicating a user learned a specific knowledge concept); video → course (indicating a video is part of a course); and course → teacher (indicating a course is taught by a specific teacher).

[0013] The knowledge concept recommendation method based on the user's learning process described above, specifically includes the following steps in step 2):

[0014] Step 2-1) Define meta-paths. Meta-paths are used to describe the composite relationships between nodes in the heterogeneous information network described in Step 1-1). Let v i ∈v、ε i Let ∈ε represent the type of the i-th node and the type of the i-th relation, respectively. Given a meta-path MP, it follows the definition:

[0015]

[0016] Where v1 is the starting node, v l It is one of the neighbors of v1 based on meta-path MP, which describes the relationship between nodes v1 and v in a heterogeneous information network. l The l-1 order composite neighbor relationship ε between them:

[0017] ε=ε1∧ε2∧...∧ε l-1 (2)

[0018] Where ∧ represents a compound relation operation, ε l-1 This indicates the relation type of the (l-1)th edge.

[0019] Step 2-2) Obtain the isomorphic network and adjacency matrix based on meta-paths. When nodes v1 and v2 in the meta-path MP are... lWhen nodes belong to the same type, find the starting node v1 and node v in the heterogeneous information network along the meta-path. l The connections between nodes are identified, ignoring other types of nodes in between and retaining only the association relationships of that type of node. This transforms the original heterogeneous information network containing multiple node types into a homogeneous information network containing only type v1 nodes. Multiplying the association matrix of this homogeneous network by its transpose yields the adjacency matrix A based on the meta-path MP. MP .

[0020] The knowledge concept recommendation method based on the user learning process described above, specifically step 3) includes the following steps:

[0021] Step 3-1) Perform feature embedding on the concepts, embedding each concept c in the dataset... i The name is fed into the word2vec model to obtain concept c. i Embedded

[0022] Step 3-2) Perform feature embedding on users. For each user u in the dataset... i Given user u i Ordered sequence of historical learning behaviors For user u i The τth concept learned For user u i The last concept in learning, also known as the goal concept, is the user u i The specific process of feature embedding is as follows:

[0023] Step 3-2-1) Adjusting the concept and target concept The time distance between them, d(t,τ), is calculated as follows:

[0024]

[0025] Where γ t,t' The concept was computed in [0,1]. With the concept of goal The similarity was calculated, and the similarity values ​​were normalized. For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation, D q For the dimension of the query vector, The concept of time index t' The key vector obtained from the corresponding embedding mapping, The concept of time index τ' The corresponding key vector, D, obtained from the embedding mapping k Let exp() be the dimension of the key vector, e be the exponential function with the natural number e as the base, and d(t,τ) be the expression that appears in the concept. The subsequent concept (i.e., concept) ) and target concept The similarity was used to adjust the concepts accordingly. and target concept The simple time distance between them is |t-τ|, where τ is a concept. Time index, t is the target concept Time index (τ) <t)。

[0026] Step 3-2-2) Using the time distance d(t,τ) as a concept Time allocation attention coefficient s t,τ s t,τ The calculation formula is as follows:

[0027]

[0028] Where s t,τ To comprehensively consider the concept With the concept of goal The concept is formed after considering the time distance and contextual information between them. The time attention coefficient allocated, For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation, D q For the dimension of the query vector, For the concept The corresponding key vector obtained from the embedding mapping, D k Let be the dimension of the key vector. `exp()` is an exponential function with base e. The coefficients are s. t,τ The size reflects the concept during the user's learning process. With the concept of goal The degree of relevance, specifically, if it appears in the concept Subsequent concepts and target concepts If the relevance is higher, then the concept is assigned. A smaller coefficient is assigned to a larger coefficient, and vice versa.

[0029] Step 3-2-3) User u i The user u is obtained by weighting and summing the historical behaviors in the learning behavior sequence with their corresponding attention weights. i Embedded The calculation process is as follows:

[0030]

[0031] in To make s t,τ The result obtained after normalization using the Softmax function is that exp() is an exponential function with the natural number e as its base. To make user u i The user embedding is obtained by weighting and summing the historical behavior and its corresponding attention weight. For the concept The corresponding value vector obtained by the embedding mapping, D v is the dimension of the value vector.

[0032] The knowledge concept recommendation method based on the user learning process described above, specifically includes the following steps in step 4):

[0033] Step 4-1) The embeddings of all concepts obtained in Step 3-1) are fed into a multi-layer GCN graph neural network for feature learning, and the embeddings generated under different meta-paths are fused through a self-attention mechanism. The specific process of feature learning is as follows:

[0034] Step 4-1-1) In the metapath MP i Next, all the concept embeddings obtained in step 3-1) are fed into a multi-layer GCN graph neural network for feature learning, resulting in an embedding matrix composed of all concept embeddings. The calculation formula is as follows:

[0035]

[0036] in For the metapath MP i The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. The embedding matrix consists of all concept embeddings learned by the l-layer GCN graph neural network. Let D be the embedding matrix composed of all concept embeddings obtained in step 3-1), and let D be the dimension of each learned concept embedding. v Let n be the dimension of each concept embedding obtained in step 3-1), n ​​be the number of concepts, and l be an indicator indicating the number of layers in the GCN. For meta-path based MP i Let I be the adjacency matrix of , and let I be the identity matrix. for The angle matrix, W l Let σ be the weight matrix to be trained in the l-th layer of the GCN graph neural network, where σ is the ReLU nonlinear activation function.

[0037] Step 4-1-2) The embedding matrices learned under different meta-paths are fused through a self-attention mechanism to obtain the final embedding matrix e. c e c The calculation process is as follows:

[0038]

[0039] in For the metapath MP i Attention coefficients of the embedding matrix obtained through learning. For the metapath MP i The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. The final embedding matrix consists of all concept embeddings, where D is the dimension of each learned concept embedding, n is the number of concepts, and a is the vector to be learned. For the defined set of metapaths, express The number of metapaths contained therein, where σ is the ReLU nonlinear activation function.

[0040] Step 4-2) Feed all the user embeddings obtained in Step 3-2) into a multi-layer GAT graph neural network for feature learning, and fuse the embeddings generated under different meta-paths through a self-attention mechanism. The specific process of feature learning is as follows:

[0041] Step 4-2-1) In the metapath MP x Next, all user embeddings obtained in step 3-2) are fed into a multi-layer GAT graph neural network for feature learning to obtain user embeddings. The calculation process is as follows:

[0042]

[0043] in For the metapath MP x Below, user u i With its neighboring user u j The relationship coefficients For user u obtained in step 3-2) i Embedded, For user u in the l-th layer of the GAT graph neural network i Embedded, For user u in the l-th layer of the GAT graph neural network j Embedded, For the metapath MP x User u i The set of neighboring nodes, Let be the vector to be learned in the l-th layer of the GAT graph neural network. W is the result of performing a transpose operation on a. l Let be the weight matrix to be trained in the l-th layer of the GAT graph neural network, LeakyReLU be the activation function, l be an indicator indicating the layer number of the GAT, || denote the concatenation operation, and exp() be an exponential function with the natural number e as the base. User u i In the metapath MP x The embeddedness obtained from learning For user u after learning through l-layer GAT graph neural network i The embedding is denoted by D, where D is the dimension of the learned embedding, and σ is the ReLU nonlinear activation function.

[0044] Step 4-2-2) The user embeddings learned under different meta-paths are fused through a self-attention mechanism to obtain the final user embedding. The calculation process is as follows:

[0045]

[0046] in For the metapath MP x Learn more and get user u i Attention coefficient For the metapath MP x The user u obtained from the following learning i The embedding is given by , where b is the vector to be learned, σ is the ReLU nonlinear activation function, and exp() is an exponential function with base e. For the end user u i The embedding, where D is the dimension of the learned user embedding. For the defined set of metapaths, express The number of metapaths contained therein.

[0047] The knowledge concept recommendation method based on the user learning process described above, specifically step 5) includes the following steps:

[0048] Step 5-1) Input all the concept embeddings and user embeddings obtained in Steps 4-1) and 4-2) into the extended matrix factorization model to obtain the user rating matrix for the concepts, and recommend the concepts with high predicted ratings to the users. The formula for calculating the rating matrix is ​​as follows:

[0049]

[0050] in For predicting user ratings of concepts, For the user's latent vector, Let F be the latent vector of the user, and F be the dimension of the latent vector. The embedding matrix consists of the embeddings of all users learned in step 4-2). For the concept embedding matrix learned in step 4-1), t c and t u It is to make e u and e c The parameter β introduced while existing in the same vector dimension u and β c To optimize the parameters, For x u The result after performing the transpose operation. For e u The result of performing the transpose operation, (t) u ) T For t u The result obtained after performing the transpose operation.

[0051] Step 5-2) Based on the objective function and gradient descent algorithm, optimize the predicted user rating matrix for the concept. The formula for calculating the objective function is as follows:

[0052]

[0053] in This is the actual user rating matrix for concepts, where m is the number of users, n is the number of concepts, λ is the regularization parameter, and ||x| is the number of users. u ||2 is x u The L2 norm, ||y c ||2 is y c L2 norm, ||t u ||2 is t u L2 norm, ||t c ||2 is t c The L2 norm, along with the other two, are used as regularization terms to control the complexity of the model.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention introduces a temporal attention mechanism into the user embedding part to process temporal information during the user's learning process. This mechanism differentiates user behavior based on the time of occurrence and assigns different temporal weights, thereby solving the problems of existing methods' inability to accurately capture changes in user learning preferences and limited feature embedding capabilities. Simultaneously, this invention connects users on the MOOC platform with other entities through a heterogeneous information network and combines meta-paths and graph neural networks to achieve embedding aggregation, resulting in embedded embeddings containing richer information. This invention demonstrates excellent accuracy and robustness in recommendations and has certain market value in fields such as learning resource recommendation and smart education. Attached Figure Description

[0056] Figure 1 This is an overall flowchart of the present invention;

[0057] Figure 2 This is basic information about a portion of the MOOCCube dataset used in this invention;

[0058] Figure 3 This is the meta-path defined for users and knowledge concepts in this invention; Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings and examples.

[0060] like Figure 1 As shown, a knowledge concept recommendation method based on the user's learning process includes the following steps:

[0061] 1) Define the node types and relationship types in heterogeneous information networks.

[0062] Step 1) specifically includes the following steps:

[0063] Step 1-1) For the MOOCCube large-scale open online course dataset, select historical records over a period of time as the model's dataset, and define five node types and five relationship types in the heterogeneous information network. The five node types include: user (an individual learning on the platform); teacher (a person teaching courses on the platform); video (an educational resource covering specific knowledge concepts); course (composed of a series of videos); and knowledge concept (the basic unit that makes up the course content). The five relationship types include: user → video (indicating a user watched a specific video); video → knowledge concept (indicating a video contains a specific knowledge concept); user → knowledge concept (indicating a user learned a specific knowledge concept); video → course (indicating a video is part of a course); and course → teacher (indicating a course is taught by a specific teacher).

[0064] 2) Define meta-paths and obtain isomorphic networks and adjacency matrices based on meta-paths.

[0065] Step 2) specifically includes the following steps:

[0066] Step 2-1) Define meta-paths. Meta-paths are used to describe the composite relationships between nodes in the heterogeneous information network described in Step 1-1). Let v i ∈v、ε i Let ∈ε represent the type of the i-th node and the type of the i-th relation, respectively. Given a meta-path MP, it follows the definition:

[0067]

[0068] Where v1 is the starting node, v l It is one of the neighbors of v1 based on meta-path MP, which describes the relationship between nodes v1 and v in a heterogeneous information network. l The l-1 order composite neighbor relationship ε between them:

[0069] ε=ε1∧ε2∧...∧ε l-1 (2)

[0070] Where ∧ represents a compound relation operation, ε l-1 This indicates the relation type of the (l-1)th edge.

[0071] Step 2-2) Obtain the isomorphic network and adjacency matrix based on meta-paths. When nodes v1 and v2 in the meta-path MP are... l When nodes belong to the same type, find the starting node v1 and node v in the heterogeneous information network along the meta-path. l The connections between nodes are identified, ignoring other types of nodes in between and retaining only the association relationships of that type of node. This transforms the original heterogeneous information network containing multiple node types into a homogeneous information network containing only type v1 nodes. Multiplying the association matrix of this homogeneous network by its transpose yields the adjacency matrix A based on the meta-path MP. MP .

[0072] 3) Embed features of concepts using their names and embed features of users using their historical behavior during the learning process.

[0073] Step 3) specifically includes the following steps:

[0074] Step 3-1) Perform feature embedding on the concepts, embedding each concept c in the dataset... i The name is fed into the word2vec model to obtain concept c. i Embedded

[0075] Step 3-2) Perform feature embedding on users. For each user u in the dataset... i Given user ui Ordered sequence of historical learning behaviors For user u i The τth concept learned For user u i The last concept in learning, also known as the goal concept, is the user u i The specific process of feature embedding is as follows:

[0076] Step 3-2-1) Adjusting the concept and target concept The time distance between them, d(t,τ), is calculated as follows:

[0077]

[0078] Where γ t,t' The concept was computed in [0,1]. With the concept of goal The similarity was calculated, and the similarity values ​​were normalized. For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation, D q For the dimension of the query vector, The concept of time index t' The key vector obtained from the corresponding embedding mapping, The concept of time index τ' The corresponding key vector obtained from the embedding mapping, D k Let exp() be the dimension of the key vector, e be the exponential function with the natural number e as the base, and d(t,τ) be the expression that appears in the concept. The subsequent concept (i.e., concept) ) and target concept The similarity was used to adjust the concepts accordingly. and target concept The simple time distance between them is |t-τ|, where τ is a concept. Time index, t is the target concept Time index (τ) <t)。

[0079] Step 3-2-2) Using the time distance d(t,τ) as a concept Time allocation attention coefficient s t,τ s t,τ The calculation formula is as follows:

[0080]

[0081] Where s t,τ To comprehensively consider the concept With the concept of goal The concept is formed after considering the time distance and contextual information between them. The time attention coefficient allocated, For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation, D q For the dimension of the query vector, For the concept The corresponding key vector obtained from the embedding mapping, D k Let be the dimension of the key vector. `exp()` is an exponential function with base e. The coefficients are s. t,τ The size reflects the concept during the user's learning process. With the concept of goal The degree of relevance, specifically, if it appears in the concept Subsequent concepts and target concepts If the relevance is higher, then the concept is assigned. A smaller coefficient is assigned to a larger coefficient, and vice versa.

[0082] Step 3-2-3) User u i The user u is obtained by weighting and summing the historical behaviors in the learning behavior sequence with their corresponding attention weights. i Embedded The calculation process is as follows:

[0083]

[0084] in To make s t,τ The result obtained after normalization using the Softmax function is that exp() is an exponential function with the natural number e as its base. To make user u i The user embedding is obtained by weighting and summing the historical behavior and its corresponding attention weight. For the concept The corresponding value vector obtained by the embedding mapping, D v is the dimension of the value vector.

[0085] 4) Using graph neural networks and self-attention mechanisms, feature learning is performed on the concept embeddings and user embeddings obtained in step 3) to obtain the concept embedding matrix and user embeddings.

[0086] Step 4) specifically includes the following steps:

[0087] Step 4-1) The embeddings of all concepts obtained in Step 3-1) are fed into a multi-layer GCN graph neural network for feature learning, and the embeddings generated under different meta-paths are fused through a self-attention mechanism. The specific process of feature learning is as follows:

[0088] Step 4-1-1) In the metapath MP i Next, all the concept embeddings obtained in step 3-1) are fed into a multi-layer GCN graph neural network for feature learning, resulting in an embedding matrix composed of all concept embeddings. The calculation formula is as follows:

[0089]

[0090] in For the metapath MP i The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. The embedding matrix consists of all concept embeddings learned by the l-layer GCN graph neural network. Let D be the embedding matrix composed of all concept embeddings obtained in step 3-1), and let D be the dimension of each learned concept embedding. v Let n be the dimension of each concept embedding obtained in step 3-1), n ​​be the number of concepts, and l be an indicator indicating the number of layers in the GCN. For meta-path based MP i Let I be the adjacency matrix of , and let I be the identity matrix. for The angle matrix, W l Let σ be the weight matrix to be trained in the l-th layer of the GCN graph neural network, where σ is the ReLU nonlinear activation function.

[0091] Step 4-1-2) The embedding matrices learned under different meta-paths are fused through a self-attention mechanism to obtain the final embedding matrix e. c e c The calculation process is as follows:

[0092]

[0093] in For the metapath MP i Attention coefficients of the embedding matrix obtained through learning. For the metapath MP i The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. The final embedding matrix consists of all concept embeddings, where D is the dimension of each learned concept embedding, n is the number of concepts, and a is the vector to be learned. For the defined set of metapaths, express The number of metapaths contained therein, where σ is the ReLU nonlinear activation function.

[0094] Step 4-2) Feed all the user embeddings obtained in Step 3-2) into a multi-layer GAT graph neural network for feature learning, and fuse the embeddings generated under different meta-paths through a self-attention mechanism. The specific process of feature learning is as follows:

[0095] Step 4-2-1) In the metapath MP x Next, all user embeddings obtained in step 3-2) are fed into a multi-layer GAT graph neural network for feature learning to obtain user embeddings. The calculation process is as follows:

[0096]

[0097] in For the metapath MP x Below, user u i With its neighboring user u j The relationship coefficients For user u obtained in step 3-2) i Embedded, For user u in the l-th layer of the GAT graph neural network i Embedded, For user u in the l-th layer of the GAT graph neural network j Embedded, For the metapath MP x User u i The set of neighboring nodes, Let be the vector to be learned in the l-th layer of the GAT graph neural network. W is the result of performing a transpose operation on a. l Let be the weight matrix to be trained in the l-th layer of the GAT graph neural network, LeakyReLU be the activation function, l be an indicator indicating the layer number of the GAT, || denote the concatenation operation, and exp() be an exponential function with the natural number e as the base. User u i In the metapath MP x The embeddedness obtained from learning For user u after learning through l-layer GAT graph neural network i The embedding is denoted by D, where D is the dimension of the learned embedding, and σ is the ReLU nonlinear activation function.

[0098] Step 4-2-2) The user embeddings learned under different meta-paths are fused through a self-attention mechanism to obtain the final user embedding. The calculation process is as follows:

[0099]

[0100] in For the metapath MP x Learn more and get user u i Attention coefficient For the metapath MP x The user u obtained from the following learning i The embedding is given by , where b is the vector to be learned, σ is the ReLU nonlinear activation function, and exp() is an exponential function with base e. For the end user u i The embedding, where D is the dimension of the learned user embedding. For the defined set of metapaths, express The number of metapaths contained therein.

[0101] 5) Input the concept embedding matrix and user embedding obtained in step 4) into the matrix factorization model to obtain the predicted user rating matrix for the concept.

[0102] Step 5) specifically includes the following steps:

[0103] Step 5-1) Input all the concept embeddings and user embeddings obtained in Steps 4-1) and 4-2) into the extended matrix factorization model to obtain the user rating matrix for the concepts, and recommend the concepts with high predicted ratings to the users. The formula for calculating the rating matrix is ​​as follows:

[0104]

[0105] in For predicting user ratings of concepts, For the user's latent vector, Let F be the latent vector of the user, and F be the dimension of the latent vector. The embedding matrix consists of the embeddings of all users learned in step 4-2). For the concept embedding matrix learned in step 4-1), t c and t u It is to make e u and e c The parameter β introduced while existing in the same vector dimension u and β c To optimize the parameters, For xu The result after performing the transpose operation. For e u The result after performing the transpose operation. For t u The result obtained after performing the transpose operation.

[0106] Step 5-2) Based on the objective function and gradient descent algorithm, optimize the predicted user rating matrix for the concept. The formula for calculating the objective function is as follows:

[0107]

[0108] in This is the actual user rating matrix for concepts, where m is the number of users, n is the number of concepts, λ is the regularization parameter, and ||x| is the number of users. u ||2 is x u The L2 norm, ||y c ||2 is y c L2 norm, ||t u ||2 is t u L2 norm, ||t c ||2 is t c The L2 norm, along with the other two, are used as regularization terms to control the complexity of the model.

[0109] Example

[0110] [1] Rendle S, Freudenthaler C, Gantner Z, et al. BPR: Bayesian personalized ranking from implicit feedback [J]. arXiv preprint arXiv:1205.2618, 2012.

[0111] Reference [1] discloses a recommendation method based on Bayesian ranking optimization. This method optimizes the ranking objective of user preferences by defining the order of user preferences for concepts and adopting a pairwise sampling strategy. Finally, it generates a personalized recommendation list by modeling the latent factors of users and concepts through matrix factorization and combining it with Bayesian inference.

[0112] [2]Kabbur S, Ning

[0113] Reference [2] discloses a recommendation method based on factor decomposition. This method models concept similarity, uses matrix factorization to represent the implicit relationship between each concept and other concepts, and predicts user preferences for new concepts by aggregating the concept embeddings of user interactions. Finally, it combines regularization strategies to avoid overfitting and generate a personalized recommendation list.

[0114] [3]He

[0115] Reference [3] discloses a recommendation method based on neural attention mechanism. Based on the similarity between concepts, this method introduces an attention mechanism to aggregate weighted user historical behavior embeddings, thereby predicting the user's preference for the target concept, and uses a regularization strategy to optimize the model and generate a personalized recommendation list.

[0116] [4] Dong Y, Chawla NV, Swami A. metapath2vec: Scalable representation learning for heterogeneous networks, in: Proceedings of the 23rd ACMSIGKDDInternational Conference on Knowledge Discovery and DataMining, 2017, pp.135-144.

[0117] Reference [4] discloses a recommendation method based on heterogeneous information networks. This method captures the complex relationship between users and concepts by learning the low-dimensional embeddings of nodes. First, it generates node sequences by randomly walking through different types of nodes such as users and concepts through meta-path design. Then, it trains these sequences using the Skip-Gram model to learn the embedding representations of users and concepts. Finally, based on these embeddings, it combines with downstream recommendation models to generate personalized recommendation lists.

[0118] [5]Gong J, Wan Y, Liu Y, et al. Attentional graph convolutional networks for knowledge concept recommendation in MOOCs in a heterogeneous view, in: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2020, pp.79-88.

[0119] Reference [5] proposes a recommendation method based on heterogeneous information networks and graph neural networks. This method constructs a heterogeneous information network to represent the relationships between entity nodes such as users, courses and knowledge concepts, and uses graph convolutional networks to embed and aggregate the nodes in the network. At the same time, it adjusts the weights between nodes through an attention mechanism to obtain the final node embedding. Finally, the obtained embedding is fed into a matrix factorization model to generate a personalized recommendation list.

[0120] To verify the recommendation accuracy of the method of this invention, the performance of the method was evaluated on four indicators: top K hit rate (HR@K), top K normalized loss cumulative gain (NDCG@K), mean reciprocal rank (MRR), and area under the receiver operating characteristic curve (AUC). The results were compared with the five references mentioned above, and the comparison results are shown in Table 1.

[0121] Table 1 compares the performance indicators with other methods.

[0122]

[0123] Conclusion: Compared with other methods for knowledge concept recommendation, the method of this invention achieves the best performance across all indicators, and has advantages such as high accuracy and high reliability.

Claims

1. A knowledge concept recommendation method based on the user's learning process, characterized in that, Includes the following steps: Step 1) Define the node types and relationship types in the heterogeneous information network; Step 2) Define meta-paths and obtain the isomorphic network and adjacency matrix based on meta-paths; Step 3) Embed features of concepts using their names and embed features of users using their historical behavior during the learning process; Step 4) Using a graph neural network and self-attention mechanism, perform feature learning on the concept embeddings and user embeddings obtained in Step 3) to obtain the concept embedding matrix and user embeddings. This specifically includes the following steps: Step 4-1) Feed the embeddings of all the concepts obtained in Step 3) into a multi-layered... In a graph neural network, feature learning is performed on the feature learning process, and embeddings generated from different meta-paths are fused through a self-attention mechanism. The specific process of feature learning is as follows: Step 4-1-1) In the metapath Next, all the concepts obtained in step 3) will be embedded into a multi-layered... Feature learning is performed on the graph neural network to obtain an embedding matrix composed of all concept embeddings. , The calculation formula is as follows: (8) in For the metapath The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. For the process layer An embedding matrix is ​​composed of all concept embeddings learned by a graph neural network. The embedding matrix is ​​composed of all the concept embeddings obtained in step 3). For each concept learned, the embedded dimension, For each concept embedding dimension obtained in step 3), For the number of concepts, For indicator, indicating The number of floors, , , For meta-path based The adjacency matrix, It is the identity matrix. for The angle matrix, For the first layer The weight matrix to be trained Here Nonlinear activation function; Step 4-1-2) The embedding matrices learned under different meta-paths are fused through a self-attention mechanism to obtain the final embedding matrix. , The calculation process is as follows: (9) (10) in For the metapath Attention coefficients of the embedding matrix obtained through learning. For the metapath The embedding matrix is ​​composed of all the concept embeddings obtained from the learning process. The final embedding matrix consists of all the concept embeddings. For each concept learned, the embedded dimension, For the number of concepts, Let be the vector to be learned. For the defined set of metapaths, express The number of metapaths included. Here Nonlinear activation function; Step 4-2) Feed all the user embeddings obtained in Step 3) into a multi-layered... In a graph neural network, feature learning is performed on the feature learning process, and embeddings generated from different meta-paths are fused through a self-attention mechanism. The specific process of feature learning is as follows: Step 4-2-1) In the metapath Next, all user embeddings obtained in step 3) are fed into a multi-layered... Feature learning is performed on the graph neural network to obtain the user embedding. , The calculation process is as follows: (11) (12) in For the metapath Below, user and its neighboring users The relationship coefficients For the user obtained in step 3) Embedded, In the first layer Users in graph neural networks Embedded, In the first layer Users in graph neural networks Embedded, For the metapath Next user The set of neighboring nodes, For the first layer The vector to be learned in a graph neural network, To The result after performing the transpose operation. For the first layer The weight matrix to be trained For activation function, For indicator, indicating The number of floors, This indicates a splicing operation. It is natural number An exponential function with base 0. user In metapath The embeddedness obtained from learning For the process layer Users after learning from graph neural networks Embedded, The dimensions of the learned embedding. Here Nonlinear activation function; Step 4-2-2) The user embeddings learned under different meta-paths are fused through a self-attention mechanism to obtain the final user embedding. , The calculation process is as follows: (13) (14) in For the metapath Learn more users Attention coefficient For the metapath Users who learn from it Embedded, Let be the vector to be learned. Here Nonlinear activation function It is natural number An exponential function with base 0. For end users Embedded, The dimensions of user embedding obtained through learning. For the defined set of metapaths, express The number of metapaths included; Step 5) Input the concept embedding matrix and user embedding obtained in Step 4) into the matrix factorization model to obtain the predicted user rating matrix for the concept, and recommend the concepts with high predicted ratings to the user.

2. The knowledge concept recommendation method based on the user learning process according to claim 1, characterized in that, Step 1) specifically includes the following steps: Step 1-1) For large-scale open online course datasets Historical records over a period of time are selected as the dataset for the model, and five node types and five relationship types are defined in the heterogeneous information network. The five node types include: users (individuals learning on the platform); teachers (people teaching courses on the platform); videos (educational resources covering specific knowledge concepts); courses (composed of a series of videos); and knowledge concepts (the basic units that make up the course content). The five relationship types include: users... Video indicates that a user has watched a specific video; video Knowledge concepts indicate that the video contains specific knowledge concepts; users Knowledge concepts refer to specific knowledge concepts that users have learned; videos A course indicates that the video is part of the course; a course Teacher, indicating that the course is taught by a specific teacher.

3. The knowledge concept recommendation method based on the user learning process according to claim 1, characterized in that, Step 2) specifically includes the following steps: Step 2-1) Define meta-paths. Meta-paths are used to describe the composite relationships between nodes in the heterogeneous information network described in Step 1-1). Let... , They represent the first The node type and the first Given a relation type and a metapath It follows the definition: (1) in As the starting node, for Meta-path based One of the neighbors, metapath Describes nodes in heterogeneous information networks and between Composite Neighbor Relationships of Rank : (2) in This indicates a compound relation operation. Indicates the first The relationship type of the edge; Step 2-2) Obtain the isomorphic network and adjacency matrix based on metapaths. Nodes in and When nodes belong to the same type, find the starting node in the heterogeneous information network along the meta-path. and nodes The connections between nodes are simplified, ignoring other types of nodes in between and retaining only the relationships between nodes of that type. This transforms a heterogeneous information network that originally contained multiple node types into one that only contains... For isomorphic information networks of type nodes, multiplying the correlation matrix of the isomorphic network with its transpose yields the meta-path-based information network. adjacency matrix .

4. The knowledge concept recommendation method based on the user learning process according to claim 1, characterized in that, Step 3) specifically includes the following steps: Step 3-1) Perform feature embedding on the concepts, embedding each concept in the dataset... Send the name Model, to obtain concept Embedded ; Step 3-2) Perform feature embedding for each user in the dataset. Given a user Ordered sequence of historical learning behaviors , For users The first day of learning A concept, For users The last concept in learning, also known as the goal concept, is the user. The specific process of feature embedding is as follows: Step 3-2-1) Adjusting the concept and target concept The time distance between them The calculation process is as follows: (3) (4) in The concept was calculated. With the concept of goal The similarity was calculated, and the similarity values ​​were normalized. For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation. For the dimension of the query vector, For time index Concept The key vector obtained from the corresponding embedding mapping, For time index Concept The key vector obtained from the corresponding embedding mapping, Let be the dimension of the key vector. It is natural number An exponential function with base 0. Calculations were made in the concept The subsequent concept (i.e., concept) ) and target concept The similarity was used to adjust the concepts accordingly. and target concept The simple time distance between them τ is a concept Time index, For the concept of target Time Index ; Step 3-2-2) Utilizing time distance For the concept Time allocation attention coefficient , The calculation formula is as follows: (5) in To comprehensively consider the concept With the concept of goal The concept is formed after considering the time distance and contextual information between them. The time attention coefficient allocated, For the concept of target The query vector obtained from the corresponding embedding mapping, To The result after performing the transpose operation. For the dimension of the query vector, For the concept The key vector obtained from the corresponding embedding mapping, Let be the dimension of the key vector. It is natural number An exponential function with base 0; coefficient The size reflects the concept during the user's learning process. With the concept of goal The degree of relevance, specifically, if it appears in the concept Subsequent concepts and target concepts If the relevance is higher, then the concept is assigned. A smaller coefficient is assigned, and vice versa; Step 3-2-3) User The user's learning behavior sequence is obtained by weighting and summing the historical behaviors with their corresponding attention weights. Embedded , The calculation process is as follows: (6) (7) in To be Send in The result obtained after normalizing the function It is natural number An exponential function with base 0. To make users The user embedding is obtained by weighting and summing the historical behavior and its corresponding attention weight. For the concept The corresponding value vector obtained from the embedding mapping, is the dimension of the value vector.

5. The knowledge concept recommendation method based on the user learning process according to claim 1, characterized in that, Step 5) specifically includes the following steps: Step 5-1) Input all the concept embeddings and user embeddings obtained in Step 4-1) and Step 4-2) into the extended matrix factorization model to obtain the user rating matrix for the concepts, and recommend the concepts with high predicted ratings to the users. The formula for calculating the rating matrix is ​​as follows: (15) in For predicting user ratings of concepts, For the user's latent vector, For the user's latent vector, Let be the dimension of the latent vector. The embedding matrix is ​​composed of the embeddings of all users learned in step 4-2). The concept embedding matrix learned in step 4-1), and It is to make and Parameters introduced while existing within the same vector dimension and To optimize the parameters, To The result after performing the transpose operation. To The result after performing the transpose operation. To The result obtained after performing the transpose operation; Step 5-2) Based on the objective function and gradient descent algorithm, optimize the predicted user rating matrix for the concept. The formula for calculating the objective function is as follows: (16) in This is a rating matrix of actual users on the concept. For the number of users, For the number of concepts, For regularization parameters, for of Norm, for of Norm, for of Norm, for of Norms, along with the other three, are used as regularization terms to control the complexity of the model.

Images