Knowledge concept recommendation method based on user learning process

Abstract

The invention discloses a knowledge concept recommendation method based on a user learning process. The knowledge concept recommendation method comprises the following steps: defining a node type and a relationship type in a heterogeneous information network; defining a meta-path, and obtaining an isomorphic network and an adjacent matrix based on the meta-path; carrying out feature embedding on the concepts by utilizing names of the concepts, and carrying out feature embedding on the users by utilizing historical behaviors in the learning process of the users; performing feature learning on the concept embedding and the user embedding obtained in the step 3) through a graph neural network and a self-attention mechanism to obtain a concept embedding matrix and the user embedding; and 4) inputting the concept embedding matrix obtained in the step 4) and user embedding into a matrix decomposition model to obtain a predicted score matrix of the user to the concepts, and recommending the concepts with high predicted scores to the user. According to the method, a time attention mechanism is introduced in the user feature embedding process, and the mechanism can effectively distinguish user behaviors at different time points, so that the problem that the user feature embedding capability is limited is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of learning resource recommendation, and in particular to a knowledge concept recommendation method based on user learning process. Background Art

[0002] In recent years, Massive Open Online Course (MOOC) platforms have developed rapidly around the world, greatly promoting the popularization of education and benefiting millions of users. However, although these platforms provide users with rich learning resources, the problem of information overload is still prominent, and users often find it difficult to filter out the courses and learning materials that best meet their needs from massive learning resources. 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 more accurately locate the learning content they are interested in, thereby improving their learning experience.

[0003] Existing knowledge concept recommendation methods often use heterogeneous information networks (HINs) to model various entities and their relationships on MOOC platforms, and focus on mining effective information between users and knowledge concepts to recommend concepts. However, these methods focus on the static interaction between users and knowledge concepts, ignoring the temporal dynamics of user learning preferences, that is, users' learning preferences will 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 time evolution characteristic that the feature embedding ability of existing methods is limited, making it difficult to accurately capture the changes in user learning preferences, which in turn leads to a decrease in the accuracy of recommendations. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a knowledge concept recommendation method based on the user learning process with efficient algorithm and high recommendation accuracy.

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

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

[0007] 2) Define the meta-path and obtain the isomorphic network and adjacency matrix based on the meta-path;

[0008] 3) Use the names of concepts to embed concepts, and use the historical behaviors of users in the learning process to embed users;

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

[0010] 5) The concept embedding matrix and user embedding obtained in step 4) are fed into the matrix decomposition model to obtain the predicted user rating matrix for the concepts, and the concepts with high predicted ratings are recommended to the users;

[0011] In the above-mentioned knowledge concept recommendation method based on the user learning process, the step 1) specifically includes the following steps:

[0012] Step 1-1) For the massive open online course dataset MOOCCube, select historical records within a period of time as the dataset of the model, and define five node types and five relationship types in the heterogeneous information network. The five node types include: users, that is, individuals who learn on the platform; teachers, that is, people who teach courses on the platform; videos, that is, educational resources covering specific knowledge concepts; courses, which are composed of a series of videos; knowledge concepts, that is, the basic units that make up the course content. The five relationship types include: user→video, indicating that the user has watched a specific video; video→knowledge concept, indicating that the video contains a specific knowledge concept; user→knowledge concept, indicating that the user has learned a specific knowledge concept; video→course, indicating that the video is part of the course; course→teacher, indicating that the course is taught by a specific teacher.

[0013] In the above-mentioned knowledge concept recommendation method based on the user learning process, the step 2) specifically includes the following steps:

[0014] Step 2-1) Define the meta-path, which is used to describe the composite relationship between nodes in the heterogeneous information network described in step 1-1). Let v i ∈v, ε i ∈ε denote the i-th node type and i-th relationship type respectively. Given a meta-path MP, it follows from the definition:

[0015]

[0016] Where v1 is the starting node, v l is one of v1’s neighbors based on metapath MP. Metapath MP describes the relationship between nodes v1 and v in heterogeneous information networks. l The l-1 order composite neighbor relationship ε between them is:

[0017] ε=ε1∧ε2∧...∧ε l-1 (2)

[0018] Where ∧ represents a composite relational operation, ε l-1 Indicates the relationship type of the l-1th edge.

[0019] Step 2-2) Obtain the isomorphic network and adjacency matrix based on the meta-path. When the nodes v1 and v lWhen they belong to the same node type, find the starting node v1 and node v along the meta-path in the heterogeneous information network. l The connection between them is ignored, and other types of nodes in the middle are ignored, and only the association relationship of this type of nodes is retained, thereby converting the heterogeneous information network containing multiple node types into a homogeneous information network containing only v1 type nodes. The association matrix of the homogeneous network is multiplied by its transposed matrix to obtain the adjacency matrix A based on the meta-path MP. MP .

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

[0021] Step 3-1) embed the concept features and embed each concept c in the dataset i The name is fed into the word2vec model to get the concept c i Embed

[0022] Step 3-2) Embed the user features. For each user u in the dataset i , given user u i Ordered historical learning behavior sequence For user u i The τth concept learned, For user u i The last concept learned, also called the target concept, is the user u i The specific process of feature embedding is as follows:

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

[0024]

[0025] where γ t,t' ∈[0,1] computes the concept With target concept The similarity is normalized. Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, The concept with time index t' The corresponding key vector obtained by embedding mapping, The concept of time index τ' The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, exp() is the exponential function with the natural number e as the base, and d(t,τ) calculates the number of times it appears in the concept The concept after that (i.e. the concept ) and target concept and adjusted the concept accordingly. and target concept The simple time distance between |t-τ|, τ is the concept The time index of t is the target concept The time index (τ <t)。

[0026] Step 3-2-2) Using the concept of time distance d(t,τ) Allocate time 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 target concept The temporal distance and context information between them are then followed by the concept The time-attention coefficient allocated, Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, For concept The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, and exp() is the exponential function with the natural number e as the base. The coefficient s t,τ The size of the concept reflects the user's learning process With target concept Specifically, if it appears in the concept The concept of follow-up and the concept of target The higher the relevance, the more concepts are assigned The smaller the coefficient, the larger the coefficient.

[0029] Step 3-2-3) Set user u i The historical behaviors in the learning behavior sequence and their corresponding attention weights are weighted and summed to get user u i Embed The calculation process is as follows:

[0030]

[0031] in To t,τ The result obtained after normalization by the Softmax function. exp() is an exponential function with the natural number e as the base. To set user u i The user embedding is obtained by weighted summing of the historical behaviors and their corresponding attention weights. For concept The corresponding embedding mapping value vector, D v is the dimension of the value vector.

[0032] In the above-mentioned knowledge concept recommendation method based on the user learning process, the step 4) specifically includes the following steps:

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

[0034] Step 4-1-1) In the metapath MP i Next, all concept embeddings obtained in step 3-1) are sent to a multi-layer GCN graph neural network for feature learning to obtain 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 composed of all concept embeddings learned below, is the embedding matrix composed of all concept embeddings after learning through the l-layer GCN graph neural network, is the embedding matrix composed of all concept embeddings obtained in step 3-1), D is the dimension of each concept embedding obtained by learning, and D v is the dimension of each concept embedding obtained in step 3-1), n ​​is the number of concepts, l is an indicator indicating the number of GCN layers, Based on metapath MP i The adjacency matrix of , I is the identity matrix, for The diagonal matrix, W l is the weight matrix to be trained in the l-th layer of the GCN graph neural network, and σ here is the ReLU nonlinear activation function.

[0037] Step 4-1-2) The embedding matrices learned under different meta-paths are fused through the 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 The attention coefficient of the embedding matrix learned below, For the metapath MP i The embedding matrix composed of all concept embeddings learned below, is the embedding matrix composed of all the final concept embeddings, D is the dimension of each concept embedding learned, n is the number of concepts, a is the vector to be learned, is a set of defined meta paths, express The number of meta-paths contained in , where σ is the ReLU nonlinear activation function.

[0040] Step 4-2) Send all 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 the 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 sent to 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 Next, user u i Its neighbor user u j The relationship coefficient between is the user u obtained in step 3-2) i The embedding is the user u in the l-th layer of the GAT graph neural network i The embedding is the user u in the l-th layer of the GAT graph neural network j The embedding For the metapath MP x Next user u i The set of neighbor nodes of is the vector to be learned in the l-th layer of the GAT graph neural network, is the result of performing the transpose operation on a, W l is the weight matrix to be trained in the l-th layer of the GAT graph neural network, LeakyReLU is the activation function, l is an indicator indicating the number of layers of GAT, || represents the concatenation operation, exp() is an exponential function with the natural number e as the base, User i In metapath MP x The learned embedding is is the user u after learning through the l-layer GAT graph neural network i The embedding of , D is the dimension of the learned embedding, and σ here is the ReLU nonlinear activation function.

[0044] Step 4-2-2) Use the self-attention mechanism to fuse the user embeddings learned in different meta-paths to obtain the final user embedding The calculation process is as follows:

[0045]

[0046] in For the metapath MP x Next, we learn to get user u i The attention coefficient, For the metapath MP x The user u learned below i The embedding of , b is the vector to be learned, σ is the ReLU nonlinear activation function, exp() is the exponential function with the natural number e as the base, For the end user i The embedding of , D is the dimension of the learned user embedding, is a set of defined meta paths, express The number of meta-paths contained in .

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

[0048] Step 5-1) Send all concept embeddings and user embeddings obtained in step 4-1) and step 4-2) to the extended matrix decomposition model to obtain the user's rating matrix for the concepts, and recommend concepts with high predicted scores to users. The calculation formula of the rating matrix is ​​as follows:

[0049]

[0050] in is the predicted user-to-concept rating matrix, is the user’s latent vector, is the user's hidden vector, F is the dimension of the hidden vector, is the embedding matrix composed of all users’ embeddings learned in step 4-2), is the concept embedding matrix learned in step 4-1), t c and t u In order to make u and e c The parameter introduced by being in the same vector dimension, β u and β c To tune the parameters, For x u The result after performing the transpose operation is: For e u The result after performing the transpose operation, (t u ) T For t u The result of performing the transpose operation.

[0051] Step 5-2) According to the objective function and the gradient descent algorithm, the predicted user rating matrix for the concept is optimized. The calculation formula of the objective function is as follows:

[0052]

[0053] in is the actual user-to-concept rating matrix, m is the number of users, n is the number of concepts, λ is the regularization parameter, ||x u ||2 is x u The L2 norm of ||y c ||2 is y c The L2 norm of ||t u ||2 is t u The L2 norm of ||t c ||2 is t c The four of them are used as regularization terms to control the complexity of the model.

[0054] The beneficial effects of the present invention are:

[0055] The present invention introduces a temporal attention mechanism in the user embedding part to process the time information in the user learning process. The mechanism differentiates the user behavior according to the time when it occurs and assigns different time weights, thereby solving the problem that the existing methods are difficult to accurately capture the changes in user learning preferences and the limited feature embedding capabilities; at the same time, the present invention connects users and other entities on the MOOC platform through a heterogeneous information network, and combines meta-paths and graph neural networks to achieve embedding aggregation, so that the generated embedding contains richer information. The present invention performs well in the accuracy and robustness of recommendations, and has certain market value in the fields of learning resource recommendation and smart education. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention;

[0057] Figure 2 Basic information of some MOOCCube datasets used in this invention;

[0058] Figure 3 is the meta-path defined for the user and knowledge concepts in the present invention; DETAILED DESCRIPTION

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

[0060] like Figure 1 As shown, a knowledge concept recommendation method based on user 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 massive open online course dataset MOOCCube, select historical records within a period of time as the dataset of the model, and define five node types and five relationship types in the heterogeneous information network. The five node types include: users, that is, individuals who learn on the platform; teachers, that is, people who teach courses on the platform; videos, that is, educational resources covering specific knowledge concepts; courses, which are composed of a series of videos; knowledge concepts, that is, the basic units that make up the course content. The five relationship types include: user→video, indicating that the user has watched a specific video; video→knowledge concept, indicating that the video contains a specific knowledge concept; user→knowledge concept, indicating that the user has learned a specific knowledge concept; video→course, indicating that the video is part of the course; course→teacher, indicating that the course is taught by a specific teacher.

[0064] 2) Define the meta-path and obtain the isomorphic network and adjacency matrix based on the meta-path.

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

[0066] Step 2-1) Define the meta-path, which is used to describe the composite relationship between nodes in the heterogeneous information network described in step 1-1). Let v i ∈v, ε i ∈ε denote the i-th node type and i-th relationship type respectively. Given a meta-path MP, it follows from the definition:

[0067]

[0068] Where v1 is the starting node, v l is one of v1’s neighbors based on metapath MP. Metapath MP describes the relationship between nodes v1 and v in heterogeneous information networks. l The l-1 order composite neighbor relationship ε between them is:

[0069] ε=ε1∧ε2∧...∧ε l-1 (2)

[0070] Where ∧ represents a composite relational operation, ε l-1 Indicates the relationship type of the l-1th edge.

[0071] Step 2-2) Obtain the isomorphic network and adjacency matrix based on the meta-path. When the nodes v1 and v l When they belong to the same node type, find the starting node v1 and node v along the meta-path in the heterogeneous information network. l The connection between them is ignored, and other types of nodes in the middle are ignored, and only the association relationship of this type of nodes is retained, thereby converting the heterogeneous information network containing multiple node types into a homogeneous information network containing only v1 type nodes. The association matrix of the homogeneous network is multiplied by its transposed matrix to obtain the adjacency matrix A based on the meta-path MP. MP .

[0072] 3) Use the names of concepts to embed concepts, and use the historical behaviors of users in the learning process to embed users.

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

[0074] Step 3-1) embed the concept features and embed each concept c in the dataset i The name is fed into the word2vec model to get the concept c i Embed

[0075] Step 3-2) Embed the user features. For each user u in the dataset i , given user ui Ordered historical learning behavior sequence For user u i The τth concept learned, For user u i The last concept learned, also called the target concept, is the user u i The specific process of feature embedding is as follows:

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

[0077]

[0078] where γ t,t' ∈[0,1] computes the concept With target concept The similarity is normalized. Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, The concept with time index t' The corresponding key vector obtained by embedding mapping, The concept of time index τ' The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, exp() is the exponential function with the natural number e as the base, and d(t,τ) calculates the number of times it appears in the concept The concept after that (i.e. the concept ) and target concept and adjusted the concept accordingly. and target concept The simple time distance between |t-τ|, τ is the concept The time index of t is the target concept The time index (τ <t)。

[0079] Step 3-2-2) Using the concept of time distance d(t,τ) Allocate time 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 target concept The temporal distance and context information between them are then followed by the concept The time-attention coefficient allocated, Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, For concept The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, and exp() is the exponential function with the natural number e as the base. The coefficient s t,τ The size of the concept reflects the user's learning process With target concept Specifically, if it appears in the concept The concept of follow-up and the concept of target The higher the correlation, the more concepts are assigned The smaller the coefficient, the larger the coefficient.

[0082] Step 3-2-3) Set user u i The historical behaviors in the learning behavior sequence and their corresponding attention weights are weighted and summed to get user u i Embed The calculation process is as follows:

[0083]

[0084] in To t,τ The result obtained after normalization by the Softmax function. exp() is an exponential function with the natural number e as the base. To set user u i The user embedding is obtained by weighted summing of the historical behaviors and their corresponding attention weights. For concept The corresponding embedding mapping value vector, D v is the dimension of the value vector.

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

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

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

[0088] Step 4-1-1) In the metapath MP i Next, all concept embeddings obtained in step 3-1) are sent to a multi-layer GCN graph neural network for feature learning to obtain 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 composed of all concept embeddings learned below, is the embedding matrix composed of all concept embeddings after learning through the l-layer GCN graph neural network, is the embedding matrix composed of all concept embeddings obtained in step 3-1), D is the dimension of each concept embedding obtained by learning, and D v is the dimension of each concept embedding obtained in step 3-1), n ​​is the number of concepts, l is an indicator indicating the number of GCN layers, Based on metapath MP i The adjacency matrix of , I is the identity matrix, for The diagonal matrix, W l is the weight matrix to be trained in the l-th layer of the GCN graph neural network, and σ here is the ReLU nonlinear activation function.

[0091] Step 4-1-2) The embedding matrices learned under different meta-paths are fused through the 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 The attention coefficient of the embedding matrix learned below, For the metapath MP i The embedding matrix composed of all concept embeddings learned below, is the embedding matrix composed of all the final concept embeddings, D is the dimension of each concept embedding learned, n is the number of concepts, a is the vector to be learned, is a set of defined meta paths, express The number of meta-paths contained in , where σ is the ReLU nonlinear activation function.

[0094] Step 4-2) Send all 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 the 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 sent to 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 Next, user u i Its neighbor user u j The relationship coefficient between is the user u obtained in step 3-2) i Embedded, is the user u in the l-th layer of the GAT graph neural network i The embedding is the user u in the l-th layer of the GAT graph neural network j The embedding For the metapath MP x Next user u i The set of neighbor nodes of is the vector to be learned in the l-th layer of the GAT graph neural network, is the result of performing the transpose operation on a, W l is the weight matrix to be trained in the l-th layer of the GAT graph neural network, LeakyReLU is the activation function, l is an indicator indicating the number of layers of GAT, || represents the concatenation operation, exp() is an exponential function with the natural number e as the base, User i In metapath MP x The learned embedding is is the user u after learning through the l-layer GAT graph neural network i The embedding of , D is the dimension of the learned embedding, and σ here is the ReLU nonlinear activation function.

[0098] Step 4-2-2) Use the self-attention mechanism to fuse the user embeddings learned in different meta-paths to obtain the final user embedding The calculation process is as follows:

[0099]

[0100] in For the metapath MP x Next, we learn to get user u i The attention coefficient, For the metapath MP x The user u learned from i The embedding of , b is the vector to be learned, σ is the ReLU nonlinear activation function, exp() is the exponential function with the natural number e as the base, For the end user i The embedding of , D is the dimension of the learned user embedding, is a set of defined meta paths, express The number of meta-paths contained in .

[0101] 5) The concept embedding matrix and user embedding obtained in step 4) are fed into the matrix decomposition model to obtain the predicted user rating matrix for the concepts.

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

[0103] Step 5-1) Send all concept embeddings and user embeddings obtained in step 4-1) and step 4-2) to the extended matrix decomposition model to obtain the user's rating matrix for the concepts, and recommend concepts with high predicted scores to users. The calculation formula of the rating matrix is ​​as follows:

[0104]

[0105] in is the predicted user-to-concept rating matrix, is the user’s latent vector, is the user's hidden vector, F is the dimension of the hidden vector, is the embedding matrix composed of all users’ embeddings learned in step 4-2), is the concept embedding matrix learned in step 4-1), t c and t u Is to make e u and e c The parameter introduced by being in the same vector dimension, β u and β c To tune the parameters, For xu The result after performing the transpose operation is: For e u The result after performing the transpose operation is: For t u The result of performing the transpose operation.

[0106] Step 5-2) According to the objective function and the gradient descent algorithm, the predicted user rating matrix for the concept is optimized. The calculation formula of the objective function is as follows:

[0107]

[0108] in is the actual user-to-concept rating matrix, m is the number of users, n is the number of concepts, λ is the regularization parameter, ||x u ||2 is x u The L2 norm of ||y c ||2 is y c The L2 norm of ||t u ||2 is t u The L2 norm of ||t c ||2 is t c The four of them are used as regularization terms to control the complexity of the model.

[0109] Examples

[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 target of user preferences by defining the user's preference order for concepts and adopting a paired sampling strategy. Finally, the latent factors of users and concepts are modeled through matrix decomposition, and combined with Bayesian inference, a personalized recommendation list is generated.

[0112] [2]Kabbur S, Ning

[0113] Reference [2] discloses a recommendation method based on factor decomposition. This method models concept similarity and uses matrix decomposition technology to represent the implicit relationship between each concept and other concepts. It also predicts user preferences for new concepts by aggregating 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 a neural attention mechanism. Based on the similarity between concepts, this method introduces an attention mechanism and aggregates weighted user historical behavior embedding to predict the user's preference for the target concept. It also 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 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, combined with downstream recommendation models, a personalized recommendation list is generated.

[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] proposed a recommendation method based on heterogeneous information networks and graph neural networks. This method constructs a heterogeneous information network to represent the relationship between entity nodes such as users, courses, and knowledge concepts, and uses a graph convolutional network to embed and aggregate the nodes in the network. At the same time, the weights between nodes are adjusted through the attention mechanism to obtain the final node embedding. Finally, the obtained embedding is sent to the matrix decomposition model to generate a personalized recommendation list.

[0120] In order to verify the recommendation accuracy of the method of the present invention, the performance of the method of the present invention is evaluated on four indicators: the hit rate of top K items (HR@K), the normalized discounted cumulative gain of top K items (NDCG@K), the mean reciprocal rank (MRR), and the area under the receiver operating characteristic curve (AUC). The method is compared with the above five documents. The comparison results are shown in Table 1:

[0121] Table 1 Comparison of performance indicators with other methods

[0122]

[0123] Conclusion: Compared with other related methods for knowledge concept recommendation, the method of the present invention achieves the best performance in all performance indicators and has the advantages of high accuracy and strong reliability.

Claims

1. A knowledge concept recommendation method based on user learning process, characterized in that: The following steps are involved: Step 1) Define the node types and relationship types in the heterogeneous information network. Step 2) Define the meta-path and obtain the isomorphic network and adjacency matrix based on the meta-path. Step 3) Use the names of concepts to embed concepts, and use the historical behaviors of users in the learning process to embed users. Step 4) Through the graph neural network and self-attention mechanism, feature learning is performed on the concept embedding and user embedding obtained in step 3) to obtain the concept embedding matrix and user embedding. Step 5) The concept embedding matrix and user embedding obtained in step 4) are fed into the matrix decomposition model to obtain the predicted user rating matrix for the concepts, and concepts with high predicted ratings are recommended to the users.

2. A knowledge concept recommendation method based on user learning process according to claim 1, characterized in that: The step 1) specifically comprises the following steps: Step 1-1) For the massive open online course dataset MOOCCube, select historical records within a period of time as the dataset of the model, and define five node types and five relationship types in the heterogeneous information network. The five node types include: users, that is, individuals who learn on the platform; teachers, that is, people who teach courses on the platform; videos, that is, educational resources covering specific knowledge concepts; courses, which are composed of a series of videos; knowledge concepts, that is, the basic units that make up the course content. The five relationship types include: user→video, indicating that the user has watched a specific video; video→knowledge concept, indicating that the video contains a specific knowledge concept; user→knowledge concept, indicating that the user has learned a specific knowledge concept; video→course, indicating that the video is part of the course; course→teacher, indicating that the course is taught by a specific teacher.

3. The method for recommending knowledge concepts based on user learning process according to claim 1, characterized in that: The step 2) specifically comprises the following steps: Step 2-1) Define the meta-path, which is used to describe the composite relationship between nodes in the heterogeneous information network described in step 1-1). Let v i ∈v, ε i ∈ε denote the i-th node type and i-th relationship type respectively. Given a meta-path MP, it follows from the definition: Where v1 is the starting node, v l is one of v1’s neighbors based on metapath MP. Metapath MP describes the relationship between nodes v1 and v in heterogeneous information networks. l The l-1 order composite neighbor relationship ε between them is: ε=ε1∧ε2∧…∧ε l-1 (2) Where ∧ represents a composite relational operation, ε l-1 Indicates the relationship type of the l-1th edge. Step 2-2) Obtain the isomorphic network and adjacency matrix based on the meta-path. When the nodes v1 and v l When they belong to the same node type, find the starting node v1 and node v along the meta-path in the heterogeneous information network. l The connection between them is ignored, and other types of nodes in the middle are ignored, and only the association relationship of this type of nodes is retained, thereby converting the heterogeneous information network containing multiple node types into a homogeneous information network containing only v1 type nodes. The association matrix of the homogeneous network is multiplied by its transposed matrix to obtain the adjacency matrix A based on the meta-path MP. MP .

4. The method for recommending knowledge concepts based on user learning process according to claim 1, characterized in that: The step 3) specifically comprises the following steps: Step 3-1) embed the concept features and embed each concept c in the dataset i The name is fed into the word2vec model to get the concept c i Embed Step 3-2) Embed the user features. For each user u in the dataset i , given user u i Ordered historical learning behavior sequence For user u i The τth concept learned, For user u i The last concept learned, also called the target concept, is the user u i The specific process of feature embedding is as follows: Step 3-2-1) Adjust the concept and target concept The calculation process of the time distance between d(t, τ) is as follows: where γ t，t’ ∈[0, 1] computes the concept With target concept The similarity is normalized. Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, is the concept with time index t′ The corresponding key vector obtained by embedding mapping, is the concept with time index τ′ The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, exp() is the exponential function with the natural number e as the base, and d(t, τ) calculates the number of times it appears in the concept The concept after that (i.e. the concept ) and target concept and adjusted the concept accordingly. and target concept The simple time distance between |t-τ|, τ is the concept The time index of t is the target concept The time index (τ <t)。 Step 3-2-2) Use the time distance d(t, τ) as the concept Allocate time attention coefficient s t，τ ,s t，τ The calculation formula is as follows: where s t，τ To comprehensively consider the concept With target concept The temporal distance and context information between them are then followed by the concept The time-attention coefficient allocated, Target concept The corresponding embedding mapping results in the query vector, For The result after performing the transpose operation, D q is the dimension of the query vector, For concept The corresponding key vector obtained by embedding mapping, D k is the dimension of the key vector, and exp() is the exponential function with the natural number e as the base. The coefficient s t，τ The size of the concept reflects the user's learning process With target concept Specifically, if it appears in the concept The concept of follow-up and the concept of target The higher the correlation, the more concepts are assigned The smaller the coefficient, the larger the coefficient. Step 3-2-3) Set user u i The historical behaviors in the learning behavior sequence and their corresponding attention weights are weighted and summed to get user u i Embed The calculation process is as follows: in To t,τ The result obtained after normalization by the Softmax function. exp() is an exponential function with the natural number e as the base. To set user u i The user embedding is obtained by weighted summing of the historical behaviors and their corresponding attention weights. For concept The corresponding embedding mapping value vector, D v is the dimension of the value vector.

5. The method for recommending knowledge concepts based on user learning process according to claim 1, characterized in that: The step 4) specifically comprises the following steps: Step 4-1) sends the embeddings of all concepts obtained in step 3-1) to a multi-layer GCN graph neural network for feature learning, and fuses the embeddings generated under different meta-paths through the self-attention mechanism. The specific process of feature learning is as follows: Step 4-1-1) In the metapath MP i Next, all concept embeddings obtained in step 3-1) are sent to a multi-layer GCN graph neural network for feature learning to obtain an embedding matrix composed of all concept embeddings The calculation formula is as follows: in For the metapath MP i The embedding matrix composed of all concept embeddings learned below, is the embedding matrix composed of all concept embeddings after learning through the l-layer GCN graph neural network, is the embedding matrix composed of all concept embeddings obtained in step 3-1), D is the dimension of each concept embedding obtained by learning, and D v is the dimension of each concept embedding obtained in step 3-1), n ​​is the number of concepts, l is an indicator indicating the number of GCN layers, Based on metapath MP i The adjacency matrix of , I is the identity matrix, for The diagonal matrix, W l is the weight matrix to be trained in the l-th layer of the GCN graph neural network, and σ here is the ReLU nonlinear activation function. Step 4-1-2) The embedding matrices learned under different meta-paths are fused through the self-attention mechanism to obtain the final embedding matrix e c , e c The calculation process is as follows: in For the metapath MP i The attention coefficient of the embedding matrix learned below, For the metapath MP i The embedding matrix composed of all concept embeddings learned below, is the embedding matrix composed of all the final concept embeddings, D is the dimension of each concept embedding learned, n is the number of concepts, a is the vector to be learned, is a set of defined meta paths, express The number of meta-paths contained in , where σ is the ReLU nonlinear activation function. Step 4-2) Send all 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 the self-attention mechanism. The specific process of feature learning is as follows: Step 4-2-1) In the metapath MP x Next, all user embeddings obtained in step 3-2) are sent to a multi-layer GAT graph neural network for feature learning to obtain user embeddings. The calculation process is as follows: in For the metapath MP x Next, user u i Its neighbor user u j The relationship coefficient between is the user u obtained in step 3-2) i The embedding is the user u in the l-th layer of the GAT graph neural network i The embedding is the user u in the l-th layer of the GAT graph neural network j The embedding For the metapath MP x Next user u i The neighbor node set of T ) l is the vector to be learned in the l-th layer of the GAT graph neural network, is the result of performing the transpose operation on a, W l is the weight matrix to be trained in the l-th layer of the GAT graph neural network, LeakyReLU is the activation function, l is an indicator indicating the number of layers of GAT, || represents the concatenation operation, exp() is an exponential function with the natural number e as the base, User i In metapath MP x The learned embedding is is the user u after learning through the l-layer GAT graph neural network i The embedding of , D is the dimension of the learned embedding, and σ here is the ReLU nonlinear activation function. Step 4-2-2) Use the self-attention mechanism to fuse the user embeddings learned in different meta-paths to obtain the final user embedding The calculation process is as follows: in For the metapath MP x Next, we learn to get user u i The attention coefficient, For the metapath MP x The user u learned below i The embedding of , b is the vector to be learned, σ is the ReLU nonlinear activation function, exp() is the exponential function with the natural number e as the base, For the end user i The embedding of , D is the dimension of the learned user embedding, is a set of defined meta paths, express The number of meta-paths contained in .

6. The method for recommending knowledge concepts based on user learning process according to claim 1, characterized in that: The step 5) specifically comprises the following steps: Step 5-1) Send all concept embeddings and user embeddings obtained in step 4-1) and step 4-2) to the extended matrix decomposition model to obtain the user's rating matrix for the concepts, and recommend concepts with high predicted scores to users. The calculation formula of the rating matrix is ​​as follows: in is the predicted user-to-concept rating matrix, is the user’s latent vector, is the user's hidden vector, F is the dimension of the hidden vector, is the embedding matrix composed of all users’ embeddings learned in step 4-2), is the concept embedding matrix learned in step 4-1), t c and t u In order to make u and e c The parameter introduced by being in the same vector dimension, β u and β c To tune the parameters, For x u The result after performing the transpose operation is: For e u The result after performing the transpose operation is: For t u The result of performing the transpose operation. Step 5-2) According to the objective function and the gradient descent algorithm, the predicted user rating matrix for the concept is optimized. The calculation formula of the objective function is as follows: in is the actual user-to-concept rating matrix, m is the number of users, n is the number of concepts, λ is the regularization parameter, and ||xu||2 is x u The L2 norm of ||y c ||2 is y c The L2 norm of ||t u ||2 is the L2 norm of tu, ||t c ||2 is t c The four of them are used as regularization terms to control the complexity of the model.

Images