A student book recommendation system and method based on an improved temporal graph neural network

Abstract

This invention proposes a student book recommendation system and method based on an improved temporal graph neural network, comprising a student borrowing graph representation model generation module, a graph node embedding matrix generation module, a graph-level embedding vector generation module, a book borrowing relationship prediction matrix generation module, and a borrowing relationship decoding and reconstruction module. It is implemented through five processes: determining the student book borrowing graph representation model, obtaining graph-level embedding vectors, obtaining a book borrowing relationship prediction matrix based on an improved gated recurrent unit model, and obtaining the borrowing relationships after decoding and reconstructing the prediction matrix. This invention considers the changes in students' borrowing interests over time, thus influencing their borrowing behavior and preferences. Compared to the previous recommendation algorithm, this invention reduces computation time by 50% and doubles efficiency.

Description

Technical Field

[0001] This invention relates to the field of higher education, and in particular to a student book recommendation system and method based on an improved temporal graph neural network. Background Technology

[0002] Book recommendation is an important research area in the field of higher education development, and in recent years, the use of big data technology for book recommendation has become a research hotspot.

[0003] The search revealed that most studies use book recommendation methods based on student-book interaction information, such as Hao Ningning's "A Book Recommendation Method and System Based on Weighted Hybrid k-Nearest Neighbor Algorithm" (Hao Ningning, Li Yuanming, Wang Chuan, Chen Mengyao, Shi Bingjie, Liu Erbao, Zhu Xiaoxue, Gao Jing. A Book Recommendation Method and System Based on Weighted Hybrid k-Nearest Neighbor Algorithm [P]. Shandong Province: CN108415928B, 2022-05-31.). Such research methods only utilize user-book interaction information, without refining the time factor, and without combining the development pattern of target users' reading habits and reading preference trends to make targeted recommendations.

[0004] Furthermore, the borrowing relationship between students and books is a complex network structure, encompassing social relationships between students, borrowing relationships between students and books, and category relationships between books. Previous research techniques have been unable to adequately represent and extract this rich information. Graph neural network technology is a perfect solution for this scenario. Currently, no research has been found applying graph neural network technology to the field of university book recommendation. Most studies use collaborative filtering and neural network techniques, such as Wang Xinhua's "A Personalized Book Recommendation Method and System Based on Reinforcement Learning" (Wang Xinhua, Wang Yuchen, Guo Lei, Jiang Haoran, Xu Changdi, Liu Fangai. A Personalized Book Recommendation Method and System Based on Reinforcement Learning [P]. Shandong Province: CN112765339A, 2021-05-07.).

[0005] In conclusion, existing technologies do not adequately consider the impact of time factors on users' reading habits and preferences, nor do they fully express and extract the rich interactive information between users and books. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide a personalized book recommendation system and method for students based on an improved temporal dynamic graph neural network. This system considers the influence of time factors on students' borrowing preferences, fully explores the correlation between students and books using graph network technology, and takes into account the impact of the dynamic changes in students' borrowing interests over time, thereby improving the effectiveness of personalized book prediction for students.

[0007] The technical solution of this invention is as follows:

[0008] In a first aspect, the present invention provides a student book recommendation system based on an improved temporal graph neural network, comprising: a student borrowing graph representation model generation module, a graph node embedding matrix generation module, a graph-level embedding vector generation module, a book borrowing relationship prediction matrix generation module, and a borrowing relationship decoding and reconstruction module, wherein:

[0009] Student borrowing graph representation model generation module: Represents the borrowing relationship between students and books as a graph, with students and books as nodes, and the links between students and books with borrowing relationships as edges;

[0010] Graph node embedding matrix generation module: Calculates the node embedding matrix in the graph based on the obtained graph structure;

[0011] Graph-level embedding vector generation module: Takes the node embedding matrix of the graph as input and calculates the graph-level embedding vector;

[0012] The book borrowing relationship prediction matrix generation module divides the student's historical borrowing data according to the borrowing date, records the total number of dates as T, constructs a student borrowing relationship graph for each of the T days, captures the graph-level embedding vector for each day using graph convolution, and inputs the graph-level embedding vector into the improved temporal graph neural network to obtain the borrowing relationship prediction matrix.

[0013] The improved temporal graph neural network is as follows:

[0014]

[0015] in:

[0016]

[0017]

[0018] in:

[0019]

[0020]

[0021] Among them, Z (t) f is the graph-level embedding vector for day t, where t∈[1,...,T]; t It's an update gate, h (t-1) It is the hidden state of period t-1. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation, W fU f W c U c b c These are the parameters that need to be learned;

[0022] The above steps are used to train and obtain the graph-level embedding vector for the predicted date.

[0023]

[0024] h (T) for h (t) The value at t = T;

[0025] Borrowing relationship decoding and reconstruction module: decodes the prediction results and reconstructs the books that students will borrow next, realizing personalized book recommendations.

[0026] Furthermore, in the student borrowing graph representation model generation module, the graph is shown below. in It is a set of nodes of size N. N is the number of nodes in the node set. It is a set of edges of size M. M is the number of edges in the edge set.

[0027] Furthermore, the nodes of the graph are embedded in the matrix generation module, and the graph... The node embedding matrix F':

[0028]

[0029] Where θ is the parameter of the matrix, which is obtained during training in the improved temporal graph neural network. for The degree matrix, This represents an adjacency matrix that takes into account its own information. It is a picture , where d is the feature matrix of the node;

[0030]

[0031]

[0032] Where A is the adjacency matrix, and I is the identity matrix of A. express The values ​​in the i-th row and j-th column, yes The value in the i-th row and i-th column.

[0033] Furthermore, in the graph-level embedding vector generation module, the node embedding matrix of the graph is used as input, and the graph-level embedding vector is calculated using the maximum graph pooling layer method. The formula for calculating the graph-level embedding vector Z is:

[0034]

[0035] F' represents the node embedding matrix, and N represents the number of nodes in the set.

[0036] Furthermore, the specific implementation of the borrowing relationship decoding and reconstruction module is as follows:

[0037] Embed the prediction map into a vector Substituting into the following decoding formula, we obtain the graph matrix of the predicted graph.

[0038]

[0039] Where σ is the sigmoid activation function, Z (T) Let T be the graph-level embedding vector. It includes node information and edge weight information;

[0040] The loss function and training are as follows: cross-entropy is used as the loss function.

[0041]

[0042] Where y represents the value of an element in the adjacency matrix A, which is either 0 or 1. Represents the reconstructed adjacency matrix The values ​​of the corresponding elements are numbers between 0 and 1. Training is stopped when the loss function is reduced to a minimum or when the loss function is within an acceptable range after a certain number of iterations. The parameters at this point are the optimal parameters.

[0043] The adjacency matrix of the borrowing graph for the predicted date If the edge represents the student's borrowing relationship, then element in row i and column j This represents the book node corresponding to the student. The borrowing probability is used to set a threshold P for the number of books recommended by students. The top P books are selected as the personalized book recommendation results for students, sorted in descending order of borrowing probability.

[0044] Secondly, this invention provides a student book recommendation method based on an improved temporal graph neural network, the implementation steps of which are as follows:

[0045] (1) Represent the borrowing relationship between students and books as a graph, with students and books as nodes, and the links between students and books with borrowing relationships as edges;

[0046] (2) Calculate the node embedding matrix in the graph based on the obtained graph structure;

[0047] (3) Using the node embedding matrix of the graph as input, calculate the graph-level embedding vector;

[0048] (4) Divide the student’s historical borrowing data according to the borrowing date, record the total number of dates as T, construct the student borrowing relationship graph for T days, capture the graph-level embedding vector for each day using graph convolution, and input the graph-level embedding vector into the improved temporal graph neural network to obtain the borrowing relationship prediction matrix.

[0049] The improved temporal graph neural network is as follows:

[0050]

[0051] in:

[0052]

[0053]

[0054] in:

[0055]

[0056]

[0057] Among them, Z (t) Let f be the graph-level embedding vector for day t, t∈[1,...,T], and f t It's an update gate, h (t) It is the hidden state in period t, h (t-1) It is the hidden state of period t-1. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation, W f U f b f W c U c b c These are the parameters that need to be learned.

[0058] The above steps are used to train and obtain the graph-level embedding vector for the predicted date.

[0059]

[0060] h (T) for h (t) The value at t = T;

[0061] (5) Decode the prediction results and reconstruct the books that students will borrow next time to realize personalized book recommendations.

[0062] The advantages of this invention compared to existing technologies are:

[0063] The method proposed in this invention takes into account the impact of time factors on students' borrowing behavior and preferences, making the recommendation results more representative, increasing the likelihood that users will accept the recommendation results, and achieving the purpose of recommendation.

[0064] This method employs an improved temporal graph neural network algorithm. In existing similar technologies, the large number of model parameters leads to excessively long model training time and low efficiency. Compared with the previous recommendation algorithm, the computation time of this invention is reduced by 50%, and the efficiency is doubled.

[0065] The graph neural network technology used in this method has a very strong learning ability in the representation and learning of graph data. It can make full use of the rich non-linear structural relationships between students, between books, and between students and books contained in the data, which can greatly improve the book recommendation effect. Attached Figure Description

[0066] Figure 1 This is a block diagram of the system of the present invention;

[0067] Figure 2 This is the borrowing relationship representation diagram in the student borrowing diagram representation model generation module of the present invention;

[0068] Figure 3 This is a diagram illustrating the graph model feature extraction process of the method of the present invention;

[0069] Figure 4 The flowchart of the improved gated recurrent neural network for the book borrowing relationship prediction matrix generation module of this invention is shown below.

[0070] Figure 5 This is a flowchart of the method of the present invention;

[0071] Figure 6 This is a rendering of the invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0073] like Figure 1As shown, this invention discloses a personalized book recommendation system for students based on a temporal dynamic graph neural network. The application scenario is to recommend books for university students' book borrowing. Specifically, it is described as follows:

[0074] 1. Data Acquisition and Preprocessing. Obtain anonymized student library borrowing data from a university, including student ID, date, book title, book attribute information, borrowing time, and return time. Perform data cleaning and remove missing data from the above data.

[0075] 2. Using students and books as nodes, and links between students and books with borrowing relationships as edges, the student-book borrowing relationship is represented as a graph structure. in It is a set of nodes of size N. N is the number of elements in the node set. It is a set of edges of size M. M is the number of elements in the edge set. Student borrowing data is divided according to date, and student-book borrowing relationship graphs are constructed for each of the T dates, as shown below. Figure 2 As shown, Figure 2 A schematic diagram of a graphical representation model for a single date.

[0076] 3. Feature extraction is divided into two steps: graph node embedding vector calculation and graph-level embedding vector calculation, such as... Figure 3 As shown: By analyzing Figure 2 The graph representation model in the code performs node representation to capture the co-occurrence relationships between nodes in the graph and expresses the nodes in the graph as low-dimensional dense vectors, ultimately obtaining the graph node embedding matrix F'. Using the graph node embedding matrix F' as input for graph representation, further feature extraction is performed. By fusing the structural information of the graph and the feature information of the nodes, a probability distribution is used to describe the information of this graph data. This probability distribution is the graph-level embedding vector Z. The specific operations are as follows:

[0077] Compute the embedding matrix F' of the graph nodes:

[0078]

[0079] Where θ is the parameter of the matrix. for The degree matrix, This represents an adjacency matrix that takes into account its own information. is the node feature matrix of the graph, where d is the feature dimension of the node;

[0080]

[0081] Where A is the adjacency matrix and I is the identity matrix of A.

[0082] Using the node representations of the graph as input, the graph-level embedding vector is calculated using the maximum graph pooling layer method. The formula for calculating the graph-level embedding vector Z is:

[0083]

[0084] F' is the embedding matrix of the graph nodes.

[0085] 4. After completing the graph-level embedding vector calculation, the graph-level embedding vector [Z] is... (1) ,...Z (T-1) The input is fed into the improved gated recurrent neural network model for model parameter updates, such as... Figure 4 As shown, Figure 4 h (t-1) It's the state transmitted from the previous node, containing relevant information about the previous node, Z. (t) It is the current input, combined with h (t-1) and Z (t) Get the output of the current hidden node and the hidden state h passed to the next node (t) After training, the feature vector for predicting the semester is obtained. Decoding the eigenvectors of the predicted semester yields a matrix representation of the graph for the predicted semester. Finally, the graph is reconstructed to obtain the predicted student book borrowing relationship graph, such as... Figure 5 As shown: First, the student-book borrowing relationship is represented as a graph structure. Node representation is then performed on the graph structure to capture the co-occurrence relationships between nodes, resulting in an embedding matrix for the graph nodes. By fusing the structural information of the graph and the feature information of the nodes, the embedding matrix is ​​further represented as a graph, yielding graph-level embedding vectors. These graph-level embedding vectors are then input into an improved recurrent neural network to obtain the graph representation for the predicted semester. After decoding, the graph-level embedding vectors for the predicted semester are obtained. Finally, the student-book borrowing relationship graph for the predicted semester is reconstructed. The specific calculations for each step are as follows:

[0086] Model Training: First, student borrowing data is divided according to date, and the number of borrowing dates is denoted as T. Student-book borrowing relationship graphs are constructed for each of the T days. Graph convolution is used to capture the features of the network topology for each day. The learned features are then input into an improved gated recurrent neural network. The specific calculation of the gated recurrent neural network model is as follows:

[0087]

[0088] in:

[0089]

[0090]

[0091] in:

[0092]

[0093]

[0094] in:

[0095]

[0096]

[0097] Among them, h (t) It is the hidden state in period t, h (t-1) It is the hidden state in period t-1, f t It's an update gate. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation, r t It's a door reset. It's the reset door before activation, Z (t) Let W be the graph-level embedding vector for day t, where t∈[1,...,T] f U f b f W c U c b c W r U r b r These are the parameters that need to be learned.

[0098] Because the current method has high computational complexity and may take a long time to process large-scale data, this invention improves the gated recurrent neural network by weakening the model's short-term memory while enhancing its long-term memory, and uses students' long-term reading habits as the main basis for recommendations.

[0099] The advantage of the improved gated recurrent neural network is that it reduces the parameters required for model learning and greatly improves the efficiency of model training. The improved temporal graph neural network is as follows:

[0100]

[0101] in:

[0102]

[0103]

[0104] in:

[0105]

[0106]

[0107] Among them, Z (t) Let f be the graph-level embedding vector for day t, t∈[1,...,T], and f t It's an update gate, h (t) It is the hidden state in period t, h (t-1) It is the hidden state of period t-1. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation, W f U f b f W c U c b c These are the parameters that need to be learned.

[0108] The above steps are used to train and obtain the graph-level embedding vector for the predicted date.

[0109]

[0110] h (T) for h (t) The value at t = T;

[0111] Feature vector decoding: The features obtained from training the GRU model are decoded to obtain a matrix representation of the graph. Then, the graph is reconstructed to recover the predicted student book borrowing relationship graph, such as... Figure 5 As shown. The specific steps are as follows:

[0112] Graph-level embedding vectors for predicting the semester Substituting into the following decoding formula, we obtain the graph matrix for the predicted semester:

[0113]

[0114] Where σ is the sigmoid activation function, Z (T) Let be the feature vector obtained by graph embedding of the graph on day T. It includes node information and edge weight information.

[0115] The loss function is defined and the model is trained as follows: Cross-entropy is used as the loss function.

[0116]

[0117] Where y represents the value (0 or 1) of an element in the adjacency matrix A. Represents the reconstructed adjacency matrix The value of the corresponding element (a number between 0 and 1) N represents the number of nodes.

[0118] Once the loss function is minimized through training the model, or once it reaches an acceptable range after a certain number of iterations, model training is stopped. The parameters at this point are the optimal parameters.

[0119] Graph reconstruction for predicted dates:

[0120] The adjacency matrix of the borrowing graph for the predicted date If the edge represents the student's borrowing relationship, then element in row i and column j This represents the book node corresponding to the student. The borrowing probability is used to set a threshold P for the number of books recommended by students. The top P books are selected as the personalized book recommendation results for students, sorted in descending order of probability value.

[0121] Model evaluation:

[0122] Efficiency calculations for book recommendation were performed using both the previous recommendation algorithm and this invention. The computation time of this invention was reduced by 50%, and efficiency was improved by 50%. Figure 6 As shown, the training time of the model gradually increases with the number of rounds. At the 6000th round, the training time of the algorithm before the improvement was 500 seconds, while the training time of the improved algorithm after the present invention is shortened to 250 seconds at the same number of rounds.

Claims

1. A student book recommendation system based on an improved temporal graph neural network, characterized in that, include: The system includes a student borrowing graph representation model generation module, a graph node embedding matrix generation module, a graph-level embedding vector generation module, a book borrowing relationship prediction matrix generation module, and a borrowing relationship decoding and reconstruction module, among which: Student borrowing graph representation model generation module: Represents the borrowing relationship between students and books as a graph, with students and books as nodes, and the links between students and books with borrowing relationships as edges; Graph node embedding matrix generation module: Calculates the node embedding matrix in the graph based on the obtained graph structure; Graph-level embedding vector generation module: Takes the node embedding matrix of the graph as input and calculates the graph-level embedding vector; The library borrowing relationship prediction matrix generation module divides student historical borrowing data according to borrowing dates, and records the total number of dates as follows: Construct respectively The student borrowing relationship graph for each day is obtained by capturing the graph-level embedding vector for each day using graph convolution, and then inputting the graph-level embedding vector into the improved temporal graph neural network to obtain the borrowing relationship prediction matrix. The improved temporal graph neural network is as follows: in: in: = in, Let be the graph-level embedding vector for day t. ; It's an update gate. It is the hidden state in period t. It is the hidden state of period t-1. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation. , , , , These are the parameters that need to be learned; Training yields graph-level embedding vectors for predicted dates. , for The value at t=T; Borrowing relationship decoding and reconstruction module: decodes the prediction results and reconstructs the books that students will borrow next, realizing personalized book recommendations.

2. The student book recommendation system based on an improved temporal graph neural network according to claim 1, characterized in that: In the student borrowing graph representation model generation module, the graph is shown below. ,in It is the size of The set of nodes, It is a set of edges of size M. .

3. The student book recommendation system based on an improved temporal graph neural network according to claim 1, characterized in that: The nodes of the graph are embedded in the matrix generation module. Node embedding matrix : , in, The parameters of the matrix are obtained through training in the improved temporal graphical neural network. for The degree matrix, This represents an adjacency matrix that takes into account its own information. It is a picture The node feature matrix, where The feature dimension of the node; in, It is an adjacency matrix. yes The identity matrix, value, yes The value in the i-th row and i-th column.

4. The student book recommendation system based on an improved temporal graph neural network according to claim 1, characterized in that: In the graph-level node embedding vector generation module, the graph's node embedding matrix is ​​used as input, and the graph-level embedding vector is calculated using the maximized graph pooling layer method. The calculation formula is: It is a node embedding matrix. It is the number of nodes in the set.

5. The student book recommendation system based on an improved temporal graph neural network according to claim 1, characterized in that: The specific implementation of the borrowing relationship decoding and reconstruction module is as follows: Graph-level embedding vectors for predicting dates Substituting into the following decoding formula, we obtain the graph matrix of the predicted graph. : in, It is the sigmoid activation function. Let T be the graph-level embedding vector. It includes node information and edge weight information; The loss function and training are as follows: cross-entropy is used as the loss function. : in, This represents the value of an element in the adjacency matrix A, which is either 0 or 1. Represents the reconstructed adjacency matrix The value of the corresponding element is a number between 0 and 1. It is the number of nodes in the node set; the loss function is improved through training. When the loss function is reduced to the minimum or reaches an acceptable range after a certain number of iterations, training is stopped, and the corresponding parameters at this point are the optimal parameters. The adjacency matrix of the borrowing graph for the predicted date If the edge represents the student's borrowing relationship, then element in row i and column j This represents the book node corresponding to the student. The probability of borrowing is used to set a threshold for the number of books recommended by students. Select the top borrowers in descending order of borrowing probability. These books serve as personalized book recommendations for students.

6. A student book recommendation method based on an improved temporal graph neural network, characterized in that... The implementation steps are as follows: (1) Represent the borrowing relationship between students and books as a graph, with students and books as nodes, and the links between students and books with borrowing relationships as edges; (2) Calculate the node embedding matrix in the graph based on the obtained graph structure; (3) Using the node embedding matrix of the graph as input, calculate the graph-level embedding vector; (4) Divide the student's historical borrowing data according to the borrowing date, and record the total number of dates as follows: Construct respectively The student borrowing relationship graph for each day is obtained by capturing the graph-level embedding vector for each day using graph convolution, and then inputting the graph-level embedding vector into the improved temporal graph neural network to obtain the borrowing relationship prediction matrix. The improved temporal graph neural network is as follows: in: in: = in, Let be the graph-level embedding vector for day t. , It's an update gate. It is the hidden state in period t. It is the hidden state of period t-1. It's the update portal before activation. It is the predicted value of the eigenvector in period t. It is the hidden state before activation. , , , , These are the parameters that need to be learned; Training yields graph-level embedding vectors for predicted dates. : for The value at t=T; (5) Decode the prediction results and reconstruct the books that students will borrow next time to realize personalized book recommendations.

Images